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;
170 * - LIFO ordering, to hand out cache-warm objects from _alloc
171 * - reduce the number of linked list operations
172 * - reduce spinlock operations
174 * The limit is stored in the per-cpu structure to reduce the data cache
181 unsigned int batchcount;
182 unsigned int touched;
185 * Must have this definition in here for the proper
186 * alignment of array_cache. Also simplifies accessing
189 * Entries should not be directly dereferenced as
190 * entries belonging to slabs marked pfmemalloc will
191 * have the lower bits set SLAB_OBJ_PFMEMALLOC
195 #define SLAB_OBJ_PFMEMALLOC 1
196 static inline bool is_obj_pfmemalloc(void *objp)
198 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
201 static inline void set_obj_pfmemalloc(void **objp)
203 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
207 static inline void clear_obj_pfmemalloc(void **objp)
209 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
213 * bootstrap: The caches do not work without cpuarrays anymore, but the
214 * cpuarrays are allocated from the generic caches...
216 #define BOOT_CPUCACHE_ENTRIES 1
217 struct arraycache_init {
218 struct array_cache cache;
219 void *entries[BOOT_CPUCACHE_ENTRIES];
223 * Need this for bootstrapping a per node allocator.
225 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
226 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
227 #define CACHE_CACHE 0
228 #define SIZE_AC MAX_NUMNODES
229 #define SIZE_NODE (2 * MAX_NUMNODES)
231 static int drain_freelist(struct kmem_cache *cache,
232 struct kmem_cache_node *n, int tofree);
233 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
235 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
236 static void cache_reap(struct work_struct *unused);
238 static int slab_early_init = 1;
240 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
241 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
243 static void kmem_cache_node_init(struct kmem_cache_node *parent)
245 INIT_LIST_HEAD(&parent->slabs_full);
246 INIT_LIST_HEAD(&parent->slabs_partial);
247 INIT_LIST_HEAD(&parent->slabs_free);
248 parent->shared = NULL;
249 parent->alien = NULL;
250 parent->colour_next = 0;
251 spin_lock_init(&parent->list_lock);
252 parent->free_objects = 0;
253 parent->free_touched = 0;
256 #define MAKE_LIST(cachep, listp, slab, nodeid) \
258 INIT_LIST_HEAD(listp); \
259 list_splice(&(cachep->node[nodeid]->slab), listp); \
262 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
264 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
265 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
266 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
269 #define CFLGS_OFF_SLAB (0x80000000UL)
270 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
272 #define BATCHREFILL_LIMIT 16
274 * Optimization question: fewer reaps means less probability for unnessary
275 * cpucache drain/refill cycles.
277 * OTOH the cpuarrays can contain lots of objects,
278 * which could lock up otherwise freeable slabs.
280 #define REAPTIMEOUT_CPUC (2*HZ)
281 #define REAPTIMEOUT_LIST3 (4*HZ)
284 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
285 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
286 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
287 #define STATS_INC_GROWN(x) ((x)->grown++)
288 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
289 #define STATS_SET_HIGH(x) \
291 if ((x)->num_active > (x)->high_mark) \
292 (x)->high_mark = (x)->num_active; \
294 #define STATS_INC_ERR(x) ((x)->errors++)
295 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
296 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
297 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
298 #define STATS_SET_FREEABLE(x, i) \
300 if ((x)->max_freeable < i) \
301 (x)->max_freeable = i; \
303 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
304 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
305 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
306 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
308 #define STATS_INC_ACTIVE(x) do { } while (0)
309 #define STATS_DEC_ACTIVE(x) do { } while (0)
310 #define STATS_INC_ALLOCED(x) do { } while (0)
311 #define STATS_INC_GROWN(x) do { } while (0)
312 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
313 #define STATS_SET_HIGH(x) do { } while (0)
314 #define STATS_INC_ERR(x) do { } while (0)
315 #define STATS_INC_NODEALLOCS(x) do { } while (0)
316 #define STATS_INC_NODEFREES(x) do { } while (0)
317 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
318 #define STATS_SET_FREEABLE(x, i) do { } while (0)
319 #define STATS_INC_ALLOCHIT(x) do { } while (0)
320 #define STATS_INC_ALLOCMISS(x) do { } while (0)
321 #define STATS_INC_FREEHIT(x) do { } while (0)
322 #define STATS_INC_FREEMISS(x) do { } while (0)
328 * memory layout of objects:
330 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
331 * the end of an object is aligned with the end of the real
332 * allocation. Catches writes behind the end of the allocation.
333 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
335 * cachep->obj_offset: The real object.
336 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
337 * cachep->size - 1* BYTES_PER_WORD: last caller address
338 * [BYTES_PER_WORD long]
340 static int obj_offset(struct kmem_cache *cachep)
342 return cachep->obj_offset;
345 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
347 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
348 return (unsigned long long*) (objp + obj_offset(cachep) -
349 sizeof(unsigned long long));
352 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
354 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
355 if (cachep->flags & SLAB_STORE_USER)
356 return (unsigned long long *)(objp + cachep->size -
357 sizeof(unsigned long long) -
359 return (unsigned long long *) (objp + cachep->size -
360 sizeof(unsigned long long));
363 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
365 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
366 return (void **)(objp + cachep->size - BYTES_PER_WORD);
371 #define obj_offset(x) 0
372 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
373 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
374 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
379 * Do not go above this order unless 0 objects fit into the slab or
380 * overridden on the command line.
382 #define SLAB_MAX_ORDER_HI 1
383 #define SLAB_MAX_ORDER_LO 0
384 static int slab_max_order = SLAB_MAX_ORDER_LO;
385 static bool slab_max_order_set __initdata;
387 static inline struct kmem_cache *virt_to_cache(const void *obj)
389 struct page *page = virt_to_head_page(obj);
390 return page->slab_cache;
393 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
396 return page->s_mem + cache->size * idx;
400 * We want to avoid an expensive divide : (offset / cache->size)
401 * Using the fact that size is a constant for a particular cache,
402 * we can replace (offset / cache->size) by
403 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
405 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
406 const struct page *page, void *obj)
408 u32 offset = (obj - page->s_mem);
409 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
412 static struct arraycache_init initarray_generic =
413 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
415 /* internal cache of cache description objs */
416 static struct kmem_cache kmem_cache_boot = {
418 .limit = BOOT_CPUCACHE_ENTRIES,
420 .size = sizeof(struct kmem_cache),
421 .name = "kmem_cache",
424 #define BAD_ALIEN_MAGIC 0x01020304ul
426 #ifdef CONFIG_LOCKDEP
429 * Slab sometimes uses the kmalloc slabs to store the slab headers
430 * for other slabs "off slab".
431 * The locking for this is tricky in that it nests within the locks
432 * of all other slabs in a few places; to deal with this special
433 * locking we put on-slab caches into a separate lock-class.
435 * We set lock class for alien array caches which are up during init.
436 * The lock annotation will be lost if all cpus of a node goes down and
437 * then comes back up during hotplug
439 static struct lock_class_key on_slab_l3_key;
440 static struct lock_class_key on_slab_alc_key;
442 static struct lock_class_key debugobj_l3_key;
443 static struct lock_class_key debugobj_alc_key;
445 static void slab_set_lock_classes(struct kmem_cache *cachep,
446 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
449 struct array_cache **alc;
450 struct kmem_cache_node *n;
457 lockdep_set_class(&n->list_lock, l3_key);
460 * FIXME: This check for BAD_ALIEN_MAGIC
461 * should go away when common slab code is taught to
462 * work even without alien caches.
463 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
464 * for alloc_alien_cache,
466 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
470 lockdep_set_class(&alc[r]->lock, alc_key);
474 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
476 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
479 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
483 for_each_online_node(node)
484 slab_set_debugobj_lock_classes_node(cachep, node);
487 static void init_node_lock_keys(int q)
494 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
495 struct kmem_cache_node *n;
496 struct kmem_cache *cache = kmalloc_caches[i];
502 if (!n || OFF_SLAB(cache))
505 slab_set_lock_classes(cache, &on_slab_l3_key,
506 &on_slab_alc_key, q);
510 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
512 if (!cachep->node[q])
515 slab_set_lock_classes(cachep, &on_slab_l3_key,
516 &on_slab_alc_key, q);
519 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
523 VM_BUG_ON(OFF_SLAB(cachep));
525 on_slab_lock_classes_node(cachep, node);
528 static inline void init_lock_keys(void)
533 init_node_lock_keys(node);
536 static void init_node_lock_keys(int q)
540 static inline void init_lock_keys(void)
544 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
548 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
552 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
556 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
561 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
563 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
565 return cachep->array[smp_processor_id()];
568 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
570 return ALIGN(nr_objs * sizeof(unsigned int), align);
574 * Calculate the number of objects and left-over bytes for a given buffer size.
576 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
577 size_t align, int flags, size_t *left_over,
582 size_t slab_size = PAGE_SIZE << gfporder;
585 * The slab management structure can be either off the slab or
586 * on it. For the latter case, the memory allocated for a
589 * - One unsigned int for each object
590 * - Padding to respect alignment of @align
591 * - @buffer_size bytes for each object
593 * If the slab management structure is off the slab, then the
594 * alignment will already be calculated into the size. Because
595 * the slabs are all pages aligned, the objects will be at the
596 * correct alignment when allocated.
598 if (flags & CFLGS_OFF_SLAB) {
600 nr_objs = slab_size / buffer_size;
604 * Ignore padding for the initial guess. The padding
605 * is at most @align-1 bytes, and @buffer_size is at
606 * least @align. In the worst case, this result will
607 * be one greater than the number of objects that fit
608 * into the memory allocation when taking the padding
611 nr_objs = (slab_size) / (buffer_size + sizeof(unsigned int));
614 * This calculated number will be either the right
615 * amount, or one greater than what we want.
617 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
621 mgmt_size = slab_mgmt_size(nr_objs, align);
624 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
628 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
630 static void __slab_error(const char *function, struct kmem_cache *cachep,
633 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
634 function, cachep->name, msg);
636 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
641 * By default on NUMA we use alien caches to stage the freeing of
642 * objects allocated from other nodes. This causes massive memory
643 * inefficiencies when using fake NUMA setup to split memory into a
644 * large number of small nodes, so it can be disabled on the command
648 static int use_alien_caches __read_mostly = 1;
649 static int __init noaliencache_setup(char *s)
651 use_alien_caches = 0;
654 __setup("noaliencache", noaliencache_setup);
656 static int __init slab_max_order_setup(char *str)
658 get_option(&str, &slab_max_order);
659 slab_max_order = slab_max_order < 0 ? 0 :
660 min(slab_max_order, MAX_ORDER - 1);
661 slab_max_order_set = true;
665 __setup("slab_max_order=", slab_max_order_setup);
669 * Special reaping functions for NUMA systems called from cache_reap().
670 * These take care of doing round robin flushing of alien caches (containing
671 * objects freed on different nodes from which they were allocated) and the
672 * flushing of remote pcps by calling drain_node_pages.
674 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
676 static void init_reap_node(int cpu)
680 node = next_node(cpu_to_mem(cpu), node_online_map);
681 if (node == MAX_NUMNODES)
682 node = first_node(node_online_map);
684 per_cpu(slab_reap_node, cpu) = node;
687 static void next_reap_node(void)
689 int node = __this_cpu_read(slab_reap_node);
691 node = next_node(node, node_online_map);
692 if (unlikely(node >= MAX_NUMNODES))
693 node = first_node(node_online_map);
694 __this_cpu_write(slab_reap_node, node);
698 #define init_reap_node(cpu) do { } while (0)
699 #define next_reap_node(void) do { } while (0)
703 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
704 * via the workqueue/eventd.
705 * Add the CPU number into the expiration time to minimize the possibility of
706 * the CPUs getting into lockstep and contending for the global cache chain
709 static void start_cpu_timer(int cpu)
711 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
714 * When this gets called from do_initcalls via cpucache_init(),
715 * init_workqueues() has already run, so keventd will be setup
718 if (keventd_up() && reap_work->work.func == NULL) {
720 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
721 schedule_delayed_work_on(cpu, reap_work,
722 __round_jiffies_relative(HZ, cpu));
726 static struct array_cache *alloc_arraycache(int node, int entries,
727 int batchcount, gfp_t gfp)
729 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
730 struct array_cache *nc = NULL;
732 nc = kmalloc_node(memsize, gfp, node);
734 * The array_cache structures contain pointers to free object.
735 * However, when such objects are allocated or transferred to another
736 * cache the pointers are not cleared and they could be counted as
737 * valid references during a kmemleak scan. Therefore, kmemleak must
738 * not scan such objects.
740 kmemleak_no_scan(nc);
744 nc->batchcount = batchcount;
746 spin_lock_init(&nc->lock);
751 static inline bool is_slab_pfmemalloc(struct page *page)
753 struct page *mem_page = virt_to_page(page->s_mem);
755 return PageSlabPfmemalloc(mem_page);
758 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
759 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
760 struct array_cache *ac)
762 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
766 if (!pfmemalloc_active)
769 spin_lock_irqsave(&n->list_lock, flags);
770 list_for_each_entry(page, &n->slabs_full, lru)
771 if (is_slab_pfmemalloc(page))
774 list_for_each_entry(page, &n->slabs_partial, lru)
775 if (is_slab_pfmemalloc(page))
778 list_for_each_entry(page, &n->slabs_free, lru)
779 if (is_slab_pfmemalloc(page))
782 pfmemalloc_active = false;
784 spin_unlock_irqrestore(&n->list_lock, flags);
787 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
788 gfp_t flags, bool force_refill)
791 void *objp = ac->entry[--ac->avail];
793 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
794 if (unlikely(is_obj_pfmemalloc(objp))) {
795 struct kmem_cache_node *n;
797 if (gfp_pfmemalloc_allowed(flags)) {
798 clear_obj_pfmemalloc(&objp);
802 /* The caller cannot use PFMEMALLOC objects, find another one */
803 for (i = 0; i < ac->avail; i++) {
804 /* If a !PFMEMALLOC object is found, swap them */
805 if (!is_obj_pfmemalloc(ac->entry[i])) {
807 ac->entry[i] = ac->entry[ac->avail];
808 ac->entry[ac->avail] = objp;
814 * If there are empty slabs on the slabs_free list and we are
815 * being forced to refill the cache, mark this one !pfmemalloc.
817 n = cachep->node[numa_mem_id()];
818 if (!list_empty(&n->slabs_free) && force_refill) {
819 struct page *page = virt_to_head_page(objp);
820 ClearPageSlabPfmemalloc(virt_to_head_page(page->s_mem));
821 clear_obj_pfmemalloc(&objp);
822 recheck_pfmemalloc_active(cachep, ac);
826 /* No !PFMEMALLOC objects available */
834 static inline void *ac_get_obj(struct kmem_cache *cachep,
835 struct array_cache *ac, gfp_t flags, bool force_refill)
839 if (unlikely(sk_memalloc_socks()))
840 objp = __ac_get_obj(cachep, ac, flags, force_refill);
842 objp = ac->entry[--ac->avail];
847 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
850 if (unlikely(pfmemalloc_active)) {
851 /* Some pfmemalloc slabs exist, check if this is one */
852 struct page *page = virt_to_head_page(objp);
853 struct page *mem_page = virt_to_head_page(page->s_mem);
854 if (PageSlabPfmemalloc(mem_page))
855 set_obj_pfmemalloc(&objp);
861 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
864 if (unlikely(sk_memalloc_socks()))
865 objp = __ac_put_obj(cachep, ac, objp);
867 ac->entry[ac->avail++] = objp;
871 * Transfer objects in one arraycache to another.
872 * Locking must be handled by the caller.
874 * Return the number of entries transferred.
876 static int transfer_objects(struct array_cache *to,
877 struct array_cache *from, unsigned int max)
879 /* Figure out how many entries to transfer */
880 int nr = min3(from->avail, max, to->limit - to->avail);
885 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
895 #define drain_alien_cache(cachep, alien) do { } while (0)
896 #define reap_alien(cachep, n) do { } while (0)
898 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
900 return (struct array_cache **)BAD_ALIEN_MAGIC;
903 static inline void free_alien_cache(struct array_cache **ac_ptr)
907 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
912 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
918 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
919 gfp_t flags, int nodeid)
924 #else /* CONFIG_NUMA */
926 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
927 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
929 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
931 struct array_cache **ac_ptr;
932 int memsize = sizeof(void *) * nr_node_ids;
937 ac_ptr = kzalloc_node(memsize, gfp, node);
940 if (i == node || !node_online(i))
942 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
944 for (i--; i >= 0; i--)
954 static void free_alien_cache(struct array_cache **ac_ptr)
965 static void __drain_alien_cache(struct kmem_cache *cachep,
966 struct array_cache *ac, int node)
968 struct kmem_cache_node *n = cachep->node[node];
971 spin_lock(&n->list_lock);
973 * Stuff objects into the remote nodes shared array first.
974 * That way we could avoid the overhead of putting the objects
975 * into the free lists and getting them back later.
978 transfer_objects(n->shared, ac, ac->limit);
980 free_block(cachep, ac->entry, ac->avail, node);
982 spin_unlock(&n->list_lock);
987 * Called from cache_reap() to regularly drain alien caches round robin.
989 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
991 int node = __this_cpu_read(slab_reap_node);
994 struct array_cache *ac = n->alien[node];
996 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
997 __drain_alien_cache(cachep, ac, node);
998 spin_unlock_irq(&ac->lock);
1003 static void drain_alien_cache(struct kmem_cache *cachep,
1004 struct array_cache **alien)
1007 struct array_cache *ac;
1008 unsigned long flags;
1010 for_each_online_node(i) {
1013 spin_lock_irqsave(&ac->lock, flags);
1014 __drain_alien_cache(cachep, ac, i);
1015 spin_unlock_irqrestore(&ac->lock, flags);
1020 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1022 int nodeid = page_to_nid(virt_to_page(objp));
1023 struct kmem_cache_node *n;
1024 struct array_cache *alien = NULL;
1027 node = numa_mem_id();
1030 * Make sure we are not freeing a object from another node to the array
1031 * cache on this cpu.
1033 if (likely(nodeid == node))
1036 n = cachep->node[node];
1037 STATS_INC_NODEFREES(cachep);
1038 if (n->alien && n->alien[nodeid]) {
1039 alien = n->alien[nodeid];
1040 spin_lock(&alien->lock);
1041 if (unlikely(alien->avail == alien->limit)) {
1042 STATS_INC_ACOVERFLOW(cachep);
1043 __drain_alien_cache(cachep, alien, nodeid);
1045 ac_put_obj(cachep, alien, objp);
1046 spin_unlock(&alien->lock);
1048 spin_lock(&(cachep->node[nodeid])->list_lock);
1049 free_block(cachep, &objp, 1, nodeid);
1050 spin_unlock(&(cachep->node[nodeid])->list_lock);
1057 * Allocates and initializes node for a node on each slab cache, used for
1058 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1059 * will be allocated off-node since memory is not yet online for the new node.
1060 * When hotplugging memory or a cpu, existing node are not replaced if
1063 * Must hold slab_mutex.
1065 static int init_cache_node_node(int node)
1067 struct kmem_cache *cachep;
1068 struct kmem_cache_node *n;
1069 const int memsize = sizeof(struct kmem_cache_node);
1071 list_for_each_entry(cachep, &slab_caches, list) {
1073 * Set up the size64 kmemlist for cpu before we can
1074 * begin anything. Make sure some other cpu on this
1075 * node has not already allocated this
1077 if (!cachep->node[node]) {
1078 n = kmalloc_node(memsize, GFP_KERNEL, node);
1081 kmem_cache_node_init(n);
1082 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1086 * The l3s don't come and go as CPUs come and
1087 * go. slab_mutex is sufficient
1090 cachep->node[node] = n;
1093 spin_lock_irq(&cachep->node[node]->list_lock);
1094 cachep->node[node]->free_limit =
1095 (1 + nr_cpus_node(node)) *
1096 cachep->batchcount + cachep->num;
1097 spin_unlock_irq(&cachep->node[node]->list_lock);
1102 static inline int slabs_tofree(struct kmem_cache *cachep,
1103 struct kmem_cache_node *n)
1105 return (n->free_objects + cachep->num - 1) / cachep->num;
1108 static void cpuup_canceled(long cpu)
1110 struct kmem_cache *cachep;
1111 struct kmem_cache_node *n = NULL;
1112 int node = cpu_to_mem(cpu);
1113 const struct cpumask *mask = cpumask_of_node(node);
1115 list_for_each_entry(cachep, &slab_caches, list) {
1116 struct array_cache *nc;
1117 struct array_cache *shared;
1118 struct array_cache **alien;
1120 /* cpu is dead; no one can alloc from it. */
1121 nc = cachep->array[cpu];
1122 cachep->array[cpu] = NULL;
1123 n = cachep->node[node];
1126 goto free_array_cache;
1128 spin_lock_irq(&n->list_lock);
1130 /* Free limit for this kmem_cache_node */
1131 n->free_limit -= cachep->batchcount;
1133 free_block(cachep, nc->entry, nc->avail, node);
1135 if (!cpumask_empty(mask)) {
1136 spin_unlock_irq(&n->list_lock);
1137 goto free_array_cache;
1142 free_block(cachep, shared->entry,
1143 shared->avail, node);
1150 spin_unlock_irq(&n->list_lock);
1154 drain_alien_cache(cachep, alien);
1155 free_alien_cache(alien);
1161 * In the previous loop, all the objects were freed to
1162 * the respective cache's slabs, now we can go ahead and
1163 * shrink each nodelist to its limit.
1165 list_for_each_entry(cachep, &slab_caches, list) {
1166 n = cachep->node[node];
1169 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1173 static int cpuup_prepare(long cpu)
1175 struct kmem_cache *cachep;
1176 struct kmem_cache_node *n = NULL;
1177 int node = cpu_to_mem(cpu);
1181 * We need to do this right in the beginning since
1182 * alloc_arraycache's are going to use this list.
1183 * kmalloc_node allows us to add the slab to the right
1184 * kmem_cache_node and not this cpu's kmem_cache_node
1186 err = init_cache_node_node(node);
1191 * Now we can go ahead with allocating the shared arrays and
1194 list_for_each_entry(cachep, &slab_caches, list) {
1195 struct array_cache *nc;
1196 struct array_cache *shared = NULL;
1197 struct array_cache **alien = NULL;
1199 nc = alloc_arraycache(node, cachep->limit,
1200 cachep->batchcount, GFP_KERNEL);
1203 if (cachep->shared) {
1204 shared = alloc_arraycache(node,
1205 cachep->shared * cachep->batchcount,
1206 0xbaadf00d, GFP_KERNEL);
1212 if (use_alien_caches) {
1213 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1220 cachep->array[cpu] = nc;
1221 n = cachep->node[node];
1224 spin_lock_irq(&n->list_lock);
1227 * We are serialised from CPU_DEAD or
1228 * CPU_UP_CANCELLED by the cpucontrol lock
1239 spin_unlock_irq(&n->list_lock);
1241 free_alien_cache(alien);
1242 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1243 slab_set_debugobj_lock_classes_node(cachep, node);
1244 else if (!OFF_SLAB(cachep) &&
1245 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1246 on_slab_lock_classes_node(cachep, node);
1248 init_node_lock_keys(node);
1252 cpuup_canceled(cpu);
1256 static int cpuup_callback(struct notifier_block *nfb,
1257 unsigned long action, void *hcpu)
1259 long cpu = (long)hcpu;
1263 case CPU_UP_PREPARE:
1264 case CPU_UP_PREPARE_FROZEN:
1265 mutex_lock(&slab_mutex);
1266 err = cpuup_prepare(cpu);
1267 mutex_unlock(&slab_mutex);
1270 case CPU_ONLINE_FROZEN:
1271 start_cpu_timer(cpu);
1273 #ifdef CONFIG_HOTPLUG_CPU
1274 case CPU_DOWN_PREPARE:
1275 case CPU_DOWN_PREPARE_FROZEN:
1277 * Shutdown cache reaper. Note that the slab_mutex is
1278 * held so that if cache_reap() is invoked it cannot do
1279 * anything expensive but will only modify reap_work
1280 * and reschedule the timer.
1282 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1283 /* Now the cache_reaper is guaranteed to be not running. */
1284 per_cpu(slab_reap_work, cpu).work.func = NULL;
1286 case CPU_DOWN_FAILED:
1287 case CPU_DOWN_FAILED_FROZEN:
1288 start_cpu_timer(cpu);
1291 case CPU_DEAD_FROZEN:
1293 * Even if all the cpus of a node are down, we don't free the
1294 * kmem_cache_node of any cache. This to avoid a race between
1295 * cpu_down, and a kmalloc allocation from another cpu for
1296 * memory from the node of the cpu going down. The node
1297 * structure is usually allocated from kmem_cache_create() and
1298 * gets destroyed at kmem_cache_destroy().
1302 case CPU_UP_CANCELED:
1303 case CPU_UP_CANCELED_FROZEN:
1304 mutex_lock(&slab_mutex);
1305 cpuup_canceled(cpu);
1306 mutex_unlock(&slab_mutex);
1309 return notifier_from_errno(err);
1312 static struct notifier_block cpucache_notifier = {
1313 &cpuup_callback, NULL, 0
1316 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1318 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1319 * Returns -EBUSY if all objects cannot be drained so that the node is not
1322 * Must hold slab_mutex.
1324 static int __meminit drain_cache_node_node(int node)
1326 struct kmem_cache *cachep;
1329 list_for_each_entry(cachep, &slab_caches, list) {
1330 struct kmem_cache_node *n;
1332 n = cachep->node[node];
1336 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1338 if (!list_empty(&n->slabs_full) ||
1339 !list_empty(&n->slabs_partial)) {
1347 static int __meminit slab_memory_callback(struct notifier_block *self,
1348 unsigned long action, void *arg)
1350 struct memory_notify *mnb = arg;
1354 nid = mnb->status_change_nid;
1359 case MEM_GOING_ONLINE:
1360 mutex_lock(&slab_mutex);
1361 ret = init_cache_node_node(nid);
1362 mutex_unlock(&slab_mutex);
1364 case MEM_GOING_OFFLINE:
1365 mutex_lock(&slab_mutex);
1366 ret = drain_cache_node_node(nid);
1367 mutex_unlock(&slab_mutex);
1371 case MEM_CANCEL_ONLINE:
1372 case MEM_CANCEL_OFFLINE:
1376 return notifier_from_errno(ret);
1378 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1381 * swap the static kmem_cache_node with kmalloced memory
1383 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1386 struct kmem_cache_node *ptr;
1388 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1391 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1393 * Do not assume that spinlocks can be initialized via memcpy:
1395 spin_lock_init(&ptr->list_lock);
1397 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1398 cachep->node[nodeid] = ptr;
1402 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1403 * size of kmem_cache_node.
1405 static void __init set_up_node(struct kmem_cache *cachep, int index)
1409 for_each_online_node(node) {
1410 cachep->node[node] = &init_kmem_cache_node[index + node];
1411 cachep->node[node]->next_reap = jiffies +
1413 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1418 * The memory after the last cpu cache pointer is used for the
1421 static void setup_node_pointer(struct kmem_cache *cachep)
1423 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1427 * Initialisation. Called after the page allocator have been initialised and
1428 * before smp_init().
1430 void __init kmem_cache_init(void)
1434 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1435 sizeof(struct rcu_head));
1436 kmem_cache = &kmem_cache_boot;
1437 setup_node_pointer(kmem_cache);
1439 if (num_possible_nodes() == 1)
1440 use_alien_caches = 0;
1442 for (i = 0; i < NUM_INIT_LISTS; i++)
1443 kmem_cache_node_init(&init_kmem_cache_node[i]);
1445 set_up_node(kmem_cache, CACHE_CACHE);
1448 * Fragmentation resistance on low memory - only use bigger
1449 * page orders on machines with more than 32MB of memory if
1450 * not overridden on the command line.
1452 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1453 slab_max_order = SLAB_MAX_ORDER_HI;
1455 /* Bootstrap is tricky, because several objects are allocated
1456 * from caches that do not exist yet:
1457 * 1) initialize the kmem_cache cache: it contains the struct
1458 * kmem_cache structures of all caches, except kmem_cache itself:
1459 * kmem_cache is statically allocated.
1460 * Initially an __init data area is used for the head array and the
1461 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1462 * array at the end of the bootstrap.
1463 * 2) Create the first kmalloc cache.
1464 * The struct kmem_cache for the new cache is allocated normally.
1465 * An __init data area is used for the head array.
1466 * 3) Create the remaining kmalloc caches, with minimally sized
1468 * 4) Replace the __init data head arrays for kmem_cache and the first
1469 * kmalloc cache with kmalloc allocated arrays.
1470 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1471 * the other cache's with kmalloc allocated memory.
1472 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1475 /* 1) create the kmem_cache */
1478 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1480 create_boot_cache(kmem_cache, "kmem_cache",
1481 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1482 nr_node_ids * sizeof(struct kmem_cache_node *),
1483 SLAB_HWCACHE_ALIGN);
1484 list_add(&kmem_cache->list, &slab_caches);
1486 /* 2+3) create the kmalloc caches */
1489 * Initialize the caches that provide memory for the array cache and the
1490 * kmem_cache_node structures first. Without this, further allocations will
1494 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1495 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1497 if (INDEX_AC != INDEX_NODE)
1498 kmalloc_caches[INDEX_NODE] =
1499 create_kmalloc_cache("kmalloc-node",
1500 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1502 slab_early_init = 0;
1504 /* 4) Replace the bootstrap head arrays */
1506 struct array_cache *ptr;
1508 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1510 memcpy(ptr, cpu_cache_get(kmem_cache),
1511 sizeof(struct arraycache_init));
1513 * Do not assume that spinlocks can be initialized via memcpy:
1515 spin_lock_init(&ptr->lock);
1517 kmem_cache->array[smp_processor_id()] = ptr;
1519 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1521 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1522 != &initarray_generic.cache);
1523 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1524 sizeof(struct arraycache_init));
1526 * Do not assume that spinlocks can be initialized via memcpy:
1528 spin_lock_init(&ptr->lock);
1530 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1532 /* 5) Replace the bootstrap kmem_cache_node */
1536 for_each_online_node(nid) {
1537 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1539 init_list(kmalloc_caches[INDEX_AC],
1540 &init_kmem_cache_node[SIZE_AC + nid], nid);
1542 if (INDEX_AC != INDEX_NODE) {
1543 init_list(kmalloc_caches[INDEX_NODE],
1544 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1549 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1552 void __init kmem_cache_init_late(void)
1554 struct kmem_cache *cachep;
1558 /* 6) resize the head arrays to their final sizes */
1559 mutex_lock(&slab_mutex);
1560 list_for_each_entry(cachep, &slab_caches, list)
1561 if (enable_cpucache(cachep, GFP_NOWAIT))
1563 mutex_unlock(&slab_mutex);
1565 /* Annotate slab for lockdep -- annotate the malloc caches */
1572 * Register a cpu startup notifier callback that initializes
1573 * cpu_cache_get for all new cpus
1575 register_cpu_notifier(&cpucache_notifier);
1579 * Register a memory hotplug callback that initializes and frees
1582 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1586 * The reap timers are started later, with a module init call: That part
1587 * of the kernel is not yet operational.
1591 static int __init cpucache_init(void)
1596 * Register the timers that return unneeded pages to the page allocator
1598 for_each_online_cpu(cpu)
1599 start_cpu_timer(cpu);
1605 __initcall(cpucache_init);
1607 static noinline void
1608 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1610 struct kmem_cache_node *n;
1612 unsigned long flags;
1616 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1618 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1619 cachep->name, cachep->size, cachep->gfporder);
1621 for_each_online_node(node) {
1622 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1623 unsigned long active_slabs = 0, num_slabs = 0;
1625 n = cachep->node[node];
1629 spin_lock_irqsave(&n->list_lock, flags);
1630 list_for_each_entry(page, &n->slabs_full, lru) {
1631 active_objs += cachep->num;
1634 list_for_each_entry(page, &n->slabs_partial, lru) {
1635 active_objs += page->active;
1638 list_for_each_entry(page, &n->slabs_free, lru)
1641 free_objects += n->free_objects;
1642 spin_unlock_irqrestore(&n->list_lock, flags);
1644 num_slabs += active_slabs;
1645 num_objs = num_slabs * cachep->num;
1647 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1648 node, active_slabs, num_slabs, active_objs, num_objs,
1654 * Interface to system's page allocator. No need to hold the cache-lock.
1656 * If we requested dmaable memory, we will get it. Even if we
1657 * did not request dmaable memory, we might get it, but that
1658 * would be relatively rare and ignorable.
1660 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1666 flags |= cachep->allocflags;
1667 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1668 flags |= __GFP_RECLAIMABLE;
1670 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1672 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1673 slab_out_of_memory(cachep, flags, nodeid);
1677 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1678 if (unlikely(page->pfmemalloc))
1679 pfmemalloc_active = true;
1681 nr_pages = (1 << cachep->gfporder);
1682 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1683 add_zone_page_state(page_zone(page),
1684 NR_SLAB_RECLAIMABLE, nr_pages);
1686 add_zone_page_state(page_zone(page),
1687 NR_SLAB_UNRECLAIMABLE, nr_pages);
1688 __SetPageSlab(page);
1689 if (page->pfmemalloc)
1690 SetPageSlabPfmemalloc(page);
1691 memcg_bind_pages(cachep, cachep->gfporder);
1693 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1694 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1697 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1699 kmemcheck_mark_unallocated_pages(page, nr_pages);
1706 * Interface to system's page release.
1708 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1710 const unsigned long nr_freed = (1 << cachep->gfporder);
1712 kmemcheck_free_shadow(page, cachep->gfporder);
1714 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1715 sub_zone_page_state(page_zone(page),
1716 NR_SLAB_RECLAIMABLE, nr_freed);
1718 sub_zone_page_state(page_zone(page),
1719 NR_SLAB_UNRECLAIMABLE, nr_freed);
1721 BUG_ON(!PageSlab(page));
1722 __ClearPageSlabPfmemalloc(page);
1723 __ClearPageSlab(page);
1724 page_mapcount_reset(page);
1725 page->mapping = NULL;
1727 memcg_release_pages(cachep, cachep->gfporder);
1728 if (current->reclaim_state)
1729 current->reclaim_state->reclaimed_slab += nr_freed;
1730 __free_memcg_kmem_pages(page, cachep->gfporder);
1733 static void kmem_rcu_free(struct rcu_head *head)
1735 struct kmem_cache *cachep;
1738 page = container_of(head, struct page, rcu_head);
1739 cachep = page->slab_cache;
1741 kmem_freepages(cachep, page);
1746 #ifdef CONFIG_DEBUG_PAGEALLOC
1747 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1748 unsigned long caller)
1750 int size = cachep->object_size;
1752 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1754 if (size < 5 * sizeof(unsigned long))
1757 *addr++ = 0x12345678;
1759 *addr++ = smp_processor_id();
1760 size -= 3 * sizeof(unsigned long);
1762 unsigned long *sptr = &caller;
1763 unsigned long svalue;
1765 while (!kstack_end(sptr)) {
1767 if (kernel_text_address(svalue)) {
1769 size -= sizeof(unsigned long);
1770 if (size <= sizeof(unsigned long))
1776 *addr++ = 0x87654321;
1780 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1782 int size = cachep->object_size;
1783 addr = &((char *)addr)[obj_offset(cachep)];
1785 memset(addr, val, size);
1786 *(unsigned char *)(addr + size - 1) = POISON_END;
1789 static void dump_line(char *data, int offset, int limit)
1792 unsigned char error = 0;
1795 printk(KERN_ERR "%03x: ", offset);
1796 for (i = 0; i < limit; i++) {
1797 if (data[offset + i] != POISON_FREE) {
1798 error = data[offset + i];
1802 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1803 &data[offset], limit, 1);
1805 if (bad_count == 1) {
1806 error ^= POISON_FREE;
1807 if (!(error & (error - 1))) {
1808 printk(KERN_ERR "Single bit error detected. Probably "
1811 printk(KERN_ERR "Run memtest86+ or a similar memory "
1814 printk(KERN_ERR "Run a memory test tool.\n");
1823 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1828 if (cachep->flags & SLAB_RED_ZONE) {
1829 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1830 *dbg_redzone1(cachep, objp),
1831 *dbg_redzone2(cachep, objp));
1834 if (cachep->flags & SLAB_STORE_USER) {
1835 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1836 *dbg_userword(cachep, objp),
1837 *dbg_userword(cachep, objp));
1839 realobj = (char *)objp + obj_offset(cachep);
1840 size = cachep->object_size;
1841 for (i = 0; i < size && lines; i += 16, lines--) {
1844 if (i + limit > size)
1846 dump_line(realobj, i, limit);
1850 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1856 realobj = (char *)objp + obj_offset(cachep);
1857 size = cachep->object_size;
1859 for (i = 0; i < size; i++) {
1860 char exp = POISON_FREE;
1863 if (realobj[i] != exp) {
1869 "Slab corruption (%s): %s start=%p, len=%d\n",
1870 print_tainted(), cachep->name, realobj, size);
1871 print_objinfo(cachep, objp, 0);
1873 /* Hexdump the affected line */
1876 if (i + limit > size)
1878 dump_line(realobj, i, limit);
1881 /* Limit to 5 lines */
1887 /* Print some data about the neighboring objects, if they
1890 struct page *page = virt_to_head_page(objp);
1893 objnr = obj_to_index(cachep, page, objp);
1895 objp = index_to_obj(cachep, page, objnr - 1);
1896 realobj = (char *)objp + obj_offset(cachep);
1897 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1899 print_objinfo(cachep, objp, 2);
1901 if (objnr + 1 < cachep->num) {
1902 objp = index_to_obj(cachep, page, objnr + 1);
1903 realobj = (char *)objp + obj_offset(cachep);
1904 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1906 print_objinfo(cachep, objp, 2);
1913 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1917 for (i = 0; i < cachep->num; i++) {
1918 void *objp = index_to_obj(cachep, page, i);
1920 if (cachep->flags & SLAB_POISON) {
1921 #ifdef CONFIG_DEBUG_PAGEALLOC
1922 if (cachep->size % PAGE_SIZE == 0 &&
1924 kernel_map_pages(virt_to_page(objp),
1925 cachep->size / PAGE_SIZE, 1);
1927 check_poison_obj(cachep, objp);
1929 check_poison_obj(cachep, objp);
1932 if (cachep->flags & SLAB_RED_ZONE) {
1933 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "start of a freed object "
1936 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "end of a freed object "
1943 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1950 * slab_destroy - destroy and release all objects in a slab
1951 * @cachep: cache pointer being destroyed
1952 * @slabp: slab pointer being destroyed
1954 * Destroy all the objs in a slab, and release the mem back to the system.
1955 * Before calling the slab must have been unlinked from the cache. The
1956 * cache-lock is not held/needed.
1958 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1960 struct freelist *freelist;
1962 freelist = page->freelist;
1963 slab_destroy_debugcheck(cachep, page);
1964 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1965 struct rcu_head *head;
1968 * RCU free overloads the RCU head over the LRU.
1969 * slab_page has been overloeaded over the LRU,
1970 * however it is not used from now on so that
1971 * we can use it safely.
1973 head = (void *)&page->rcu_head;
1974 call_rcu(head, kmem_rcu_free);
1977 kmem_freepages(cachep, page);
1981 * From now on, we don't use freelist
1982 * although actual page can be freed in rcu context
1984 if (OFF_SLAB(cachep))
1985 kmem_cache_free(cachep->freelist_cache, freelist);
1989 * calculate_slab_order - calculate size (page order) of slabs
1990 * @cachep: pointer to the cache that is being created
1991 * @size: size of objects to be created in this cache.
1992 * @align: required alignment for the objects.
1993 * @flags: slab allocation flags
1995 * Also calculates the number of objects per slab.
1997 * This could be made much more intelligent. For now, try to avoid using
1998 * high order pages for slabs. When the gfp() functions are more friendly
1999 * towards high-order requests, this should be changed.
2001 static size_t calculate_slab_order(struct kmem_cache *cachep,
2002 size_t size, size_t align, unsigned long flags)
2004 unsigned long offslab_limit;
2005 size_t left_over = 0;
2008 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2012 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2016 if (flags & CFLGS_OFF_SLAB) {
2018 * Max number of objs-per-slab for caches which
2019 * use off-slab slabs. Needed to avoid a possible
2020 * looping condition in cache_grow().
2022 offslab_limit = size;
2023 offslab_limit /= sizeof(unsigned int);
2025 if (num > offslab_limit)
2029 /* Found something acceptable - save it away */
2031 cachep->gfporder = gfporder;
2032 left_over = remainder;
2035 * A VFS-reclaimable slab tends to have most allocations
2036 * as GFP_NOFS and we really don't want to have to be allocating
2037 * higher-order pages when we are unable to shrink dcache.
2039 if (flags & SLAB_RECLAIM_ACCOUNT)
2043 * Large number of objects is good, but very large slabs are
2044 * currently bad for the gfp()s.
2046 if (gfporder >= slab_max_order)
2050 * Acceptable internal fragmentation?
2052 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2058 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2060 if (slab_state >= FULL)
2061 return enable_cpucache(cachep, gfp);
2063 if (slab_state == DOWN) {
2065 * Note: Creation of first cache (kmem_cache).
2066 * The setup_node is taken care
2067 * of by the caller of __kmem_cache_create
2069 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2070 slab_state = PARTIAL;
2071 } else if (slab_state == PARTIAL) {
2073 * Note: the second kmem_cache_create must create the cache
2074 * that's used by kmalloc(24), otherwise the creation of
2075 * further caches will BUG().
2077 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2080 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2081 * the second cache, then we need to set up all its node/,
2082 * otherwise the creation of further caches will BUG().
2084 set_up_node(cachep, SIZE_AC);
2085 if (INDEX_AC == INDEX_NODE)
2086 slab_state = PARTIAL_NODE;
2088 slab_state = PARTIAL_ARRAYCACHE;
2090 /* Remaining boot caches */
2091 cachep->array[smp_processor_id()] =
2092 kmalloc(sizeof(struct arraycache_init), gfp);
2094 if (slab_state == PARTIAL_ARRAYCACHE) {
2095 set_up_node(cachep, SIZE_NODE);
2096 slab_state = PARTIAL_NODE;
2099 for_each_online_node(node) {
2100 cachep->node[node] =
2101 kmalloc_node(sizeof(struct kmem_cache_node),
2103 BUG_ON(!cachep->node[node]);
2104 kmem_cache_node_init(cachep->node[node]);
2108 cachep->node[numa_mem_id()]->next_reap =
2109 jiffies + REAPTIMEOUT_LIST3 +
2110 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2112 cpu_cache_get(cachep)->avail = 0;
2113 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2114 cpu_cache_get(cachep)->batchcount = 1;
2115 cpu_cache_get(cachep)->touched = 0;
2116 cachep->batchcount = 1;
2117 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2122 * __kmem_cache_create - Create a cache.
2123 * @cachep: cache management descriptor
2124 * @flags: SLAB flags
2126 * Returns a ptr to the cache on success, NULL on failure.
2127 * Cannot be called within a int, but can be interrupted.
2128 * The @ctor is run when new pages are allocated by the cache.
2132 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2133 * to catch references to uninitialised memory.
2135 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2136 * for buffer overruns.
2138 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2139 * cacheline. This can be beneficial if you're counting cycles as closely
2143 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2145 size_t left_over, freelist_size, ralign;
2148 size_t size = cachep->size;
2153 * Enable redzoning and last user accounting, except for caches with
2154 * large objects, if the increased size would increase the object size
2155 * above the next power of two: caches with object sizes just above a
2156 * power of two have a significant amount of internal fragmentation.
2158 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2159 2 * sizeof(unsigned long long)))
2160 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2161 if (!(flags & SLAB_DESTROY_BY_RCU))
2162 flags |= SLAB_POISON;
2164 if (flags & SLAB_DESTROY_BY_RCU)
2165 BUG_ON(flags & SLAB_POISON);
2169 * Check that size is in terms of words. This is needed to avoid
2170 * unaligned accesses for some archs when redzoning is used, and makes
2171 * sure any on-slab bufctl's are also correctly aligned.
2173 if (size & (BYTES_PER_WORD - 1)) {
2174 size += (BYTES_PER_WORD - 1);
2175 size &= ~(BYTES_PER_WORD - 1);
2179 * Redzoning and user store require word alignment or possibly larger.
2180 * Note this will be overridden by architecture or caller mandated
2181 * alignment if either is greater than BYTES_PER_WORD.
2183 if (flags & SLAB_STORE_USER)
2184 ralign = BYTES_PER_WORD;
2186 if (flags & SLAB_RED_ZONE) {
2187 ralign = REDZONE_ALIGN;
2188 /* If redzoning, ensure that the second redzone is suitably
2189 * aligned, by adjusting the object size accordingly. */
2190 size += REDZONE_ALIGN - 1;
2191 size &= ~(REDZONE_ALIGN - 1);
2194 /* 3) caller mandated alignment */
2195 if (ralign < cachep->align) {
2196 ralign = cachep->align;
2198 /* disable debug if necessary */
2199 if (ralign > __alignof__(unsigned long long))
2200 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2204 cachep->align = ralign;
2206 if (slab_is_available())
2211 setup_node_pointer(cachep);
2215 * Both debugging options require word-alignment which is calculated
2218 if (flags & SLAB_RED_ZONE) {
2219 /* add space for red zone words */
2220 cachep->obj_offset += sizeof(unsigned long long);
2221 size += 2 * sizeof(unsigned long long);
2223 if (flags & SLAB_STORE_USER) {
2224 /* user store requires one word storage behind the end of
2225 * the real object. But if the second red zone needs to be
2226 * aligned to 64 bits, we must allow that much space.
2228 if (flags & SLAB_RED_ZONE)
2229 size += REDZONE_ALIGN;
2231 size += BYTES_PER_WORD;
2233 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2234 if (size >= kmalloc_size(INDEX_NODE + 1)
2235 && cachep->object_size > cache_line_size()
2236 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2237 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2244 * Determine if the slab management is 'on' or 'off' slab.
2245 * (bootstrapping cannot cope with offslab caches so don't do
2246 * it too early on. Always use on-slab management when
2247 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2249 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2250 !(flags & SLAB_NOLEAKTRACE))
2252 * Size is large, assume best to place the slab management obj
2253 * off-slab (should allow better packing of objs).
2255 flags |= CFLGS_OFF_SLAB;
2257 size = ALIGN(size, cachep->align);
2259 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2265 ALIGN(cachep->num * sizeof(unsigned int), cachep->align);
2268 * If the slab has been placed off-slab, and we have enough space then
2269 * move it on-slab. This is at the expense of any extra colouring.
2271 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2272 flags &= ~CFLGS_OFF_SLAB;
2273 left_over -= freelist_size;
2276 if (flags & CFLGS_OFF_SLAB) {
2277 /* really off slab. No need for manual alignment */
2278 freelist_size = cachep->num * sizeof(unsigned int);
2280 #ifdef CONFIG_PAGE_POISONING
2281 /* If we're going to use the generic kernel_map_pages()
2282 * poisoning, then it's going to smash the contents of
2283 * the redzone and userword anyhow, so switch them off.
2285 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2286 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2290 cachep->colour_off = cache_line_size();
2291 /* Offset must be a multiple of the alignment. */
2292 if (cachep->colour_off < cachep->align)
2293 cachep->colour_off = cachep->align;
2294 cachep->colour = left_over / cachep->colour_off;
2295 cachep->freelist_size = freelist_size;
2296 cachep->flags = flags;
2297 cachep->allocflags = __GFP_COMP;
2298 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2299 cachep->allocflags |= GFP_DMA;
2300 cachep->size = size;
2301 cachep->reciprocal_buffer_size = reciprocal_value(size);
2303 if (flags & CFLGS_OFF_SLAB) {
2304 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2306 * This is a possibility for one of the malloc_sizes caches.
2307 * But since we go off slab only for object size greater than
2308 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2309 * this should not happen at all.
2310 * But leave a BUG_ON for some lucky dude.
2312 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2315 err = setup_cpu_cache(cachep, gfp);
2317 __kmem_cache_shutdown(cachep);
2321 if (flags & SLAB_DEBUG_OBJECTS) {
2323 * Would deadlock through slab_destroy()->call_rcu()->
2324 * debug_object_activate()->kmem_cache_alloc().
2326 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2328 slab_set_debugobj_lock_classes(cachep);
2329 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2330 on_slab_lock_classes(cachep);
2336 static void check_irq_off(void)
2338 BUG_ON(!irqs_disabled());
2341 static void check_irq_on(void)
2343 BUG_ON(irqs_disabled());
2346 static void check_spinlock_acquired(struct kmem_cache *cachep)
2350 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2354 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2358 assert_spin_locked(&cachep->node[node]->list_lock);
2363 #define check_irq_off() do { } while(0)
2364 #define check_irq_on() do { } while(0)
2365 #define check_spinlock_acquired(x) do { } while(0)
2366 #define check_spinlock_acquired_node(x, y) do { } while(0)
2369 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2370 struct array_cache *ac,
2371 int force, int node);
2373 static void do_drain(void *arg)
2375 struct kmem_cache *cachep = arg;
2376 struct array_cache *ac;
2377 int node = numa_mem_id();
2380 ac = cpu_cache_get(cachep);
2381 spin_lock(&cachep->node[node]->list_lock);
2382 free_block(cachep, ac->entry, ac->avail, node);
2383 spin_unlock(&cachep->node[node]->list_lock);
2387 static void drain_cpu_caches(struct kmem_cache *cachep)
2389 struct kmem_cache_node *n;
2392 on_each_cpu(do_drain, cachep, 1);
2394 for_each_online_node(node) {
2395 n = cachep->node[node];
2397 drain_alien_cache(cachep, n->alien);
2400 for_each_online_node(node) {
2401 n = cachep->node[node];
2403 drain_array(cachep, n, n->shared, 1, node);
2408 * Remove slabs from the list of free slabs.
2409 * Specify the number of slabs to drain in tofree.
2411 * Returns the actual number of slabs released.
2413 static int drain_freelist(struct kmem_cache *cache,
2414 struct kmem_cache_node *n, int tofree)
2416 struct list_head *p;
2421 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2423 spin_lock_irq(&n->list_lock);
2424 p = n->slabs_free.prev;
2425 if (p == &n->slabs_free) {
2426 spin_unlock_irq(&n->list_lock);
2430 page = list_entry(p, struct page, lru);
2432 BUG_ON(page->active);
2434 list_del(&page->lru);
2436 * Safe to drop the lock. The slab is no longer linked
2439 n->free_objects -= cache->num;
2440 spin_unlock_irq(&n->list_lock);
2441 slab_destroy(cache, page);
2448 /* Called with slab_mutex held to protect against cpu hotplug */
2449 static int __cache_shrink(struct kmem_cache *cachep)
2452 struct kmem_cache_node *n;
2454 drain_cpu_caches(cachep);
2457 for_each_online_node(i) {
2458 n = cachep->node[i];
2462 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2464 ret += !list_empty(&n->slabs_full) ||
2465 !list_empty(&n->slabs_partial);
2467 return (ret ? 1 : 0);
2471 * kmem_cache_shrink - Shrink a cache.
2472 * @cachep: The cache to shrink.
2474 * Releases as many slabs as possible for a cache.
2475 * To help debugging, a zero exit status indicates all slabs were released.
2477 int kmem_cache_shrink(struct kmem_cache *cachep)
2480 BUG_ON(!cachep || in_interrupt());
2483 mutex_lock(&slab_mutex);
2484 ret = __cache_shrink(cachep);
2485 mutex_unlock(&slab_mutex);
2489 EXPORT_SYMBOL(kmem_cache_shrink);
2491 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2494 struct kmem_cache_node *n;
2495 int rc = __cache_shrink(cachep);
2500 for_each_online_cpu(i)
2501 kfree(cachep->array[i]);
2503 /* NUMA: free the node structures */
2504 for_each_online_node(i) {
2505 n = cachep->node[i];
2508 free_alien_cache(n->alien);
2516 * Get the memory for a slab management obj.
2517 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2518 * always come from malloc_sizes caches. The slab descriptor cannot
2519 * come from the same cache which is getting created because,
2520 * when we are searching for an appropriate cache for these
2521 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2522 * If we are creating a malloc_sizes cache here it would not be visible to
2523 * kmem_find_general_cachep till the initialization is complete.
2524 * Hence we cannot have freelist_cache same as the original cache.
2526 static struct freelist *alloc_slabmgmt(struct kmem_cache *cachep,
2527 struct page *page, int colour_off,
2528 gfp_t local_flags, int nodeid)
2530 struct freelist *freelist;
2531 void *addr = page_address(page);
2533 if (OFF_SLAB(cachep)) {
2534 /* Slab management obj is off-slab. */
2535 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2536 local_flags, nodeid);
2538 * If the first object in the slab is leaked (it's allocated
2539 * but no one has a reference to it), we want to make sure
2540 * kmemleak does not treat the ->s_mem pointer as a reference
2541 * to the object. Otherwise we will not report the leak.
2543 kmemleak_scan_area(&page->lru, sizeof(struct list_head),
2548 freelist = addr + colour_off;
2549 colour_off += cachep->freelist_size;
2552 page->s_mem = addr + colour_off;
2556 static inline unsigned int *slab_bufctl(struct page *page)
2558 return (unsigned int *)(page->freelist);
2561 static void cache_init_objs(struct kmem_cache *cachep,
2566 for (i = 0; i < cachep->num; i++) {
2567 void *objp = index_to_obj(cachep, page, i);
2569 /* need to poison the objs? */
2570 if (cachep->flags & SLAB_POISON)
2571 poison_obj(cachep, objp, POISON_FREE);
2572 if (cachep->flags & SLAB_STORE_USER)
2573 *dbg_userword(cachep, objp) = NULL;
2575 if (cachep->flags & SLAB_RED_ZONE) {
2576 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2577 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2580 * Constructors are not allowed to allocate memory from the same
2581 * cache which they are a constructor for. Otherwise, deadlock.
2582 * They must also be threaded.
2584 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2585 cachep->ctor(objp + obj_offset(cachep));
2587 if (cachep->flags & SLAB_RED_ZONE) {
2588 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2589 slab_error(cachep, "constructor overwrote the"
2590 " end of an object");
2591 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2592 slab_error(cachep, "constructor overwrote the"
2593 " start of an object");
2595 if ((cachep->size % PAGE_SIZE) == 0 &&
2596 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2597 kernel_map_pages(virt_to_page(objp),
2598 cachep->size / PAGE_SIZE, 0);
2603 slab_bufctl(page)[i] = i;
2607 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2609 if (CONFIG_ZONE_DMA_FLAG) {
2610 if (flags & GFP_DMA)
2611 BUG_ON(!(cachep->allocflags & GFP_DMA));
2613 BUG_ON(cachep->allocflags & GFP_DMA);
2617 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2622 objp = index_to_obj(cachep, page, slab_bufctl(page)[page->active]);
2625 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2631 static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2632 void *objp, int nodeid)
2634 unsigned int objnr = obj_to_index(cachep, page, objp);
2638 /* Verify that the slab belongs to the intended node */
2639 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2641 /* Verify double free bug */
2642 for (i = page->active; i < cachep->num; i++) {
2643 if (slab_bufctl(page)[i] == objnr) {
2644 printk(KERN_ERR "slab: double free detected in cache "
2645 "'%s', objp %p\n", cachep->name, objp);
2651 slab_bufctl(page)[page->active] = objnr;
2655 * Map pages beginning at addr to the given cache and slab. This is required
2656 * for the slab allocator to be able to lookup the cache and slab of a
2657 * virtual address for kfree, ksize, and slab debugging.
2659 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2660 struct freelist *freelist)
2662 page->slab_cache = cache;
2663 page->freelist = freelist;
2667 * Grow (by 1) the number of slabs within a cache. This is called by
2668 * kmem_cache_alloc() when there are no active objs left in a cache.
2670 static int cache_grow(struct kmem_cache *cachep,
2671 gfp_t flags, int nodeid, struct page *page)
2673 struct freelist *freelist;
2676 struct kmem_cache_node *n;
2679 * Be lazy and only check for valid flags here, keeping it out of the
2680 * critical path in kmem_cache_alloc().
2682 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2683 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2685 /* Take the node list lock to change the colour_next on this node */
2687 n = cachep->node[nodeid];
2688 spin_lock(&n->list_lock);
2690 /* Get colour for the slab, and cal the next value. */
2691 offset = n->colour_next;
2693 if (n->colour_next >= cachep->colour)
2695 spin_unlock(&n->list_lock);
2697 offset *= cachep->colour_off;
2699 if (local_flags & __GFP_WAIT)
2703 * The test for missing atomic flag is performed here, rather than
2704 * the more obvious place, simply to reduce the critical path length
2705 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2706 * will eventually be caught here (where it matters).
2708 kmem_flagcheck(cachep, flags);
2711 * Get mem for the objs. Attempt to allocate a physical page from
2715 page = kmem_getpages(cachep, local_flags, nodeid);
2719 /* Get slab management. */
2720 freelist = alloc_slabmgmt(cachep, page, offset,
2721 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2725 slab_map_pages(cachep, page, freelist);
2727 cache_init_objs(cachep, page);
2729 if (local_flags & __GFP_WAIT)
2730 local_irq_disable();
2732 spin_lock(&n->list_lock);
2734 /* Make slab active. */
2735 list_add_tail(&page->lru, &(n->slabs_free));
2736 STATS_INC_GROWN(cachep);
2737 n->free_objects += cachep->num;
2738 spin_unlock(&n->list_lock);
2741 kmem_freepages(cachep, page);
2743 if (local_flags & __GFP_WAIT)
2744 local_irq_disable();
2751 * Perform extra freeing checks:
2752 * - detect bad pointers.
2753 * - POISON/RED_ZONE checking
2755 static void kfree_debugcheck(const void *objp)
2757 if (!virt_addr_valid(objp)) {
2758 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2759 (unsigned long)objp);
2764 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2766 unsigned long long redzone1, redzone2;
2768 redzone1 = *dbg_redzone1(cache, obj);
2769 redzone2 = *dbg_redzone2(cache, obj);
2774 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2777 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2778 slab_error(cache, "double free detected");
2780 slab_error(cache, "memory outside object was overwritten");
2782 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2783 obj, redzone1, redzone2);
2786 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2787 unsigned long caller)
2792 BUG_ON(virt_to_cache(objp) != cachep);
2794 objp -= obj_offset(cachep);
2795 kfree_debugcheck(objp);
2796 page = virt_to_head_page(objp);
2798 if (cachep->flags & SLAB_RED_ZONE) {
2799 verify_redzone_free(cachep, objp);
2800 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2801 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2803 if (cachep->flags & SLAB_STORE_USER)
2804 *dbg_userword(cachep, objp) = (void *)caller;
2806 objnr = obj_to_index(cachep, page, objp);
2808 BUG_ON(objnr >= cachep->num);
2809 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2811 if (cachep->flags & SLAB_POISON) {
2812 #ifdef CONFIG_DEBUG_PAGEALLOC
2813 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2814 store_stackinfo(cachep, objp, caller);
2815 kernel_map_pages(virt_to_page(objp),
2816 cachep->size / PAGE_SIZE, 0);
2818 poison_obj(cachep, objp, POISON_FREE);
2821 poison_obj(cachep, objp, POISON_FREE);
2828 #define kfree_debugcheck(x) do { } while(0)
2829 #define cache_free_debugcheck(x,objp,z) (objp)
2832 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2836 struct kmem_cache_node *n;
2837 struct array_cache *ac;
2841 node = numa_mem_id();
2842 if (unlikely(force_refill))
2845 ac = cpu_cache_get(cachep);
2846 batchcount = ac->batchcount;
2847 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2849 * If there was little recent activity on this cache, then
2850 * perform only a partial refill. Otherwise we could generate
2853 batchcount = BATCHREFILL_LIMIT;
2855 n = cachep->node[node];
2857 BUG_ON(ac->avail > 0 || !n);
2858 spin_lock(&n->list_lock);
2860 /* See if we can refill from the shared array */
2861 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2862 n->shared->touched = 1;
2866 while (batchcount > 0) {
2867 struct list_head *entry;
2869 /* Get slab alloc is to come from. */
2870 entry = n->slabs_partial.next;
2871 if (entry == &n->slabs_partial) {
2872 n->free_touched = 1;
2873 entry = n->slabs_free.next;
2874 if (entry == &n->slabs_free)
2878 page = list_entry(entry, struct page, lru);
2879 check_spinlock_acquired(cachep);
2882 * The slab was either on partial or free list so
2883 * there must be at least one object available for
2886 BUG_ON(page->active >= cachep->num);
2888 while (page->active < cachep->num && batchcount--) {
2889 STATS_INC_ALLOCED(cachep);
2890 STATS_INC_ACTIVE(cachep);
2891 STATS_SET_HIGH(cachep);
2893 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2897 /* move slabp to correct slabp list: */
2898 list_del(&page->lru);
2899 if (page->active == cachep->num)
2900 list_add(&page->list, &n->slabs_full);
2902 list_add(&page->list, &n->slabs_partial);
2906 n->free_objects -= ac->avail;
2908 spin_unlock(&n->list_lock);
2910 if (unlikely(!ac->avail)) {
2913 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
2915 /* cache_grow can reenable interrupts, then ac could change. */
2916 ac = cpu_cache_get(cachep);
2917 node = numa_mem_id();
2919 /* no objects in sight? abort */
2920 if (!x && (ac->avail == 0 || force_refill))
2923 if (!ac->avail) /* objects refilled by interrupt? */
2928 return ac_get_obj(cachep, ac, flags, force_refill);
2931 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2934 might_sleep_if(flags & __GFP_WAIT);
2936 kmem_flagcheck(cachep, flags);
2941 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2942 gfp_t flags, void *objp, unsigned long caller)
2946 if (cachep->flags & SLAB_POISON) {
2947 #ifdef CONFIG_DEBUG_PAGEALLOC
2948 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2949 kernel_map_pages(virt_to_page(objp),
2950 cachep->size / PAGE_SIZE, 1);
2952 check_poison_obj(cachep, objp);
2954 check_poison_obj(cachep, objp);
2956 poison_obj(cachep, objp, POISON_INUSE);
2958 if (cachep->flags & SLAB_STORE_USER)
2959 *dbg_userword(cachep, objp) = (void *)caller;
2961 if (cachep->flags & SLAB_RED_ZONE) {
2962 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2963 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2964 slab_error(cachep, "double free, or memory outside"
2965 " object was overwritten");
2967 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2968 objp, *dbg_redzone1(cachep, objp),
2969 *dbg_redzone2(cachep, objp));
2971 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2972 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2974 objp += obj_offset(cachep);
2975 if (cachep->ctor && cachep->flags & SLAB_POISON)
2977 if (ARCH_SLAB_MINALIGN &&
2978 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2979 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2980 objp, (int)ARCH_SLAB_MINALIGN);
2985 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2988 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
2990 if (cachep == kmem_cache)
2993 return should_failslab(cachep->object_size, flags, cachep->flags);
2996 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2999 struct array_cache *ac;
3000 bool force_refill = false;
3004 ac = cpu_cache_get(cachep);
3005 if (likely(ac->avail)) {
3007 objp = ac_get_obj(cachep, ac, flags, false);
3010 * Allow for the possibility all avail objects are not allowed
3011 * by the current flags
3014 STATS_INC_ALLOCHIT(cachep);
3017 force_refill = true;
3020 STATS_INC_ALLOCMISS(cachep);
3021 objp = cache_alloc_refill(cachep, flags, force_refill);
3023 * the 'ac' may be updated by cache_alloc_refill(),
3024 * and kmemleak_erase() requires its correct value.
3026 ac = cpu_cache_get(cachep);
3030 * To avoid a false negative, if an object that is in one of the
3031 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3032 * treat the array pointers as a reference to the object.
3035 kmemleak_erase(&ac->entry[ac->avail]);
3041 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3043 * If we are in_interrupt, then process context, including cpusets and
3044 * mempolicy, may not apply and should not be used for allocation policy.
3046 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3048 int nid_alloc, nid_here;
3050 if (in_interrupt() || (flags & __GFP_THISNODE))
3052 nid_alloc = nid_here = numa_mem_id();
3053 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3054 nid_alloc = cpuset_slab_spread_node();
3055 else if (current->mempolicy)
3056 nid_alloc = slab_node();
3057 if (nid_alloc != nid_here)
3058 return ____cache_alloc_node(cachep, flags, nid_alloc);
3063 * Fallback function if there was no memory available and no objects on a
3064 * certain node and fall back is permitted. First we scan all the
3065 * available node for available objects. If that fails then we
3066 * perform an allocation without specifying a node. This allows the page
3067 * allocator to do its reclaim / fallback magic. We then insert the
3068 * slab into the proper nodelist and then allocate from it.
3070 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3072 struct zonelist *zonelist;
3076 enum zone_type high_zoneidx = gfp_zone(flags);
3079 unsigned int cpuset_mems_cookie;
3081 if (flags & __GFP_THISNODE)
3084 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3087 cpuset_mems_cookie = get_mems_allowed();
3088 zonelist = node_zonelist(slab_node(), flags);
3092 * Look through allowed nodes for objects available
3093 * from existing per node queues.
3095 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3096 nid = zone_to_nid(zone);
3098 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3100 cache->node[nid]->free_objects) {
3101 obj = ____cache_alloc_node(cache,
3102 flags | GFP_THISNODE, nid);
3110 * This allocation will be performed within the constraints
3111 * of the current cpuset / memory policy requirements.
3112 * We may trigger various forms of reclaim on the allowed
3113 * set and go into memory reserves if necessary.
3117 if (local_flags & __GFP_WAIT)
3119 kmem_flagcheck(cache, flags);
3120 page = kmem_getpages(cache, local_flags, numa_mem_id());
3121 if (local_flags & __GFP_WAIT)
3122 local_irq_disable();
3125 * Insert into the appropriate per node queues
3127 nid = page_to_nid(page);
3128 if (cache_grow(cache, flags, nid, page)) {
3129 obj = ____cache_alloc_node(cache,
3130 flags | GFP_THISNODE, nid);
3133 * Another processor may allocate the
3134 * objects in the slab since we are
3135 * not holding any locks.
3139 /* cache_grow already freed obj */
3145 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3151 * A interface to enable slab creation on nodeid
3153 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3156 struct list_head *entry;
3158 struct kmem_cache_node *n;
3162 VM_BUG_ON(nodeid > num_online_nodes());
3163 n = cachep->node[nodeid];
3168 spin_lock(&n->list_lock);
3169 entry = n->slabs_partial.next;
3170 if (entry == &n->slabs_partial) {
3171 n->free_touched = 1;
3172 entry = n->slabs_free.next;
3173 if (entry == &n->slabs_free)
3177 page = list_entry(entry, struct page, lru);
3178 check_spinlock_acquired_node(cachep, nodeid);
3180 STATS_INC_NODEALLOCS(cachep);
3181 STATS_INC_ACTIVE(cachep);
3182 STATS_SET_HIGH(cachep);
3184 BUG_ON(page->active == cachep->num);
3186 obj = slab_get_obj(cachep, page, nodeid);
3188 /* move slabp to correct slabp list: */
3189 list_del(&page->lru);
3191 if (page->active == cachep->num)
3192 list_add(&page->lru, &n->slabs_full);
3194 list_add(&page->lru, &n->slabs_partial);
3196 spin_unlock(&n->list_lock);
3200 spin_unlock(&n->list_lock);
3201 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3205 return fallback_alloc(cachep, flags);
3211 static __always_inline void *
3212 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3213 unsigned long caller)
3215 unsigned long save_flags;
3217 int slab_node = numa_mem_id();
3219 flags &= gfp_allowed_mask;
3221 lockdep_trace_alloc(flags);
3223 if (slab_should_failslab(cachep, flags))
3226 cachep = memcg_kmem_get_cache(cachep, flags);
3228 cache_alloc_debugcheck_before(cachep, flags);
3229 local_irq_save(save_flags);
3231 if (nodeid == NUMA_NO_NODE)
3234 if (unlikely(!cachep->node[nodeid])) {
3235 /* Node not bootstrapped yet */
3236 ptr = fallback_alloc(cachep, flags);
3240 if (nodeid == slab_node) {
3242 * Use the locally cached objects if possible.
3243 * However ____cache_alloc does not allow fallback
3244 * to other nodes. It may fail while we still have
3245 * objects on other nodes available.
3247 ptr = ____cache_alloc(cachep, flags);
3251 /* ___cache_alloc_node can fall back to other nodes */
3252 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3254 local_irq_restore(save_flags);
3255 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3256 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3260 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3262 if (unlikely((flags & __GFP_ZERO) && ptr))
3263 memset(ptr, 0, cachep->object_size);
3268 static __always_inline void *
3269 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3273 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3274 objp = alternate_node_alloc(cache, flags);
3278 objp = ____cache_alloc(cache, flags);
3281 * We may just have run out of memory on the local node.
3282 * ____cache_alloc_node() knows how to locate memory on other nodes
3285 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3292 static __always_inline void *
3293 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3295 return ____cache_alloc(cachep, flags);
3298 #endif /* CONFIG_NUMA */
3300 static __always_inline void *
3301 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3303 unsigned long save_flags;
3306 flags &= gfp_allowed_mask;
3308 lockdep_trace_alloc(flags);
3310 if (slab_should_failslab(cachep, flags))
3313 cachep = memcg_kmem_get_cache(cachep, flags);
3315 cache_alloc_debugcheck_before(cachep, flags);
3316 local_irq_save(save_flags);
3317 objp = __do_cache_alloc(cachep, flags);
3318 local_irq_restore(save_flags);
3319 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3320 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3325 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3327 if (unlikely((flags & __GFP_ZERO) && objp))
3328 memset(objp, 0, cachep->object_size);
3334 * Caller needs to acquire correct kmem_list's list_lock
3336 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3340 struct kmem_cache_node *n;
3342 for (i = 0; i < nr_objects; i++) {
3346 clear_obj_pfmemalloc(&objpp[i]);
3349 page = virt_to_head_page(objp);
3350 n = cachep->node[node];
3351 list_del(&page->lru);
3352 check_spinlock_acquired_node(cachep, node);
3353 slab_put_obj(cachep, page, objp, node);
3354 STATS_DEC_ACTIVE(cachep);
3357 /* fixup slab chains */
3358 if (page->active == 0) {
3359 if (n->free_objects > n->free_limit) {
3360 n->free_objects -= cachep->num;
3361 /* No need to drop any previously held
3362 * lock here, even if we have a off-slab slab
3363 * descriptor it is guaranteed to come from
3364 * a different cache, refer to comments before
3367 slab_destroy(cachep, page);
3369 list_add(&page->lru, &n->slabs_free);
3372 /* Unconditionally move a slab to the end of the
3373 * partial list on free - maximum time for the
3374 * other objects to be freed, too.
3376 list_add_tail(&page->lru, &n->slabs_partial);
3381 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3384 struct kmem_cache_node *n;
3385 int node = numa_mem_id();
3387 batchcount = ac->batchcount;
3389 BUG_ON(!batchcount || batchcount > ac->avail);
3392 n = cachep->node[node];
3393 spin_lock(&n->list_lock);
3395 struct array_cache *shared_array = n->shared;
3396 int max = shared_array->limit - shared_array->avail;
3398 if (batchcount > max)
3400 memcpy(&(shared_array->entry[shared_array->avail]),
3401 ac->entry, sizeof(void *) * batchcount);
3402 shared_array->avail += batchcount;
3407 free_block(cachep, ac->entry, batchcount, node);
3412 struct list_head *p;
3414 p = n->slabs_free.next;
3415 while (p != &(n->slabs_free)) {
3418 page = list_entry(p, struct page, lru);
3419 BUG_ON(page->active);
3424 STATS_SET_FREEABLE(cachep, i);
3427 spin_unlock(&n->list_lock);
3428 ac->avail -= batchcount;
3429 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3433 * Release an obj back to its cache. If the obj has a constructed state, it must
3434 * be in this state _before_ it is released. Called with disabled ints.
3436 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3437 unsigned long caller)
3439 struct array_cache *ac = cpu_cache_get(cachep);
3442 kmemleak_free_recursive(objp, cachep->flags);
3443 objp = cache_free_debugcheck(cachep, objp, caller);
3445 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3448 * Skip calling cache_free_alien() when the platform is not numa.
3449 * This will avoid cache misses that happen while accessing slabp (which
3450 * is per page memory reference) to get nodeid. Instead use a global
3451 * variable to skip the call, which is mostly likely to be present in
3454 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3457 if (likely(ac->avail < ac->limit)) {
3458 STATS_INC_FREEHIT(cachep);
3460 STATS_INC_FREEMISS(cachep);
3461 cache_flusharray(cachep, ac);
3464 ac_put_obj(cachep, ac, objp);
3468 * kmem_cache_alloc - Allocate an object
3469 * @cachep: The cache to allocate from.
3470 * @flags: See kmalloc().
3472 * Allocate an object from this cache. The flags are only relevant
3473 * if the cache has no available objects.
3475 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3477 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3479 trace_kmem_cache_alloc(_RET_IP_, ret,
3480 cachep->object_size, cachep->size, flags);
3484 EXPORT_SYMBOL(kmem_cache_alloc);
3486 #ifdef CONFIG_TRACING
3488 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3492 ret = slab_alloc(cachep, flags, _RET_IP_);
3494 trace_kmalloc(_RET_IP_, ret,
3495 size, cachep->size, flags);
3498 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3503 * kmem_cache_alloc_node - Allocate an object on the specified node
3504 * @cachep: The cache to allocate from.
3505 * @flags: See kmalloc().
3506 * @nodeid: node number of the target node.
3508 * Identical to kmem_cache_alloc but it will allocate memory on the given
3509 * node, which can improve the performance for cpu bound structures.
3511 * Fallback to other node is possible if __GFP_THISNODE is not set.
3513 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3515 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3517 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3518 cachep->object_size, cachep->size,
3523 EXPORT_SYMBOL(kmem_cache_alloc_node);
3525 #ifdef CONFIG_TRACING
3526 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3533 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3535 trace_kmalloc_node(_RET_IP_, ret,
3540 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3543 static __always_inline void *
3544 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3546 struct kmem_cache *cachep;
3548 cachep = kmalloc_slab(size, flags);
3549 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3551 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3554 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3555 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3557 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3559 EXPORT_SYMBOL(__kmalloc_node);
3561 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3562 int node, unsigned long caller)
3564 return __do_kmalloc_node(size, flags, node, caller);
3566 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3568 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3570 return __do_kmalloc_node(size, flags, node, 0);
3572 EXPORT_SYMBOL(__kmalloc_node);
3573 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3574 #endif /* CONFIG_NUMA */
3577 * __do_kmalloc - allocate memory
3578 * @size: how many bytes of memory are required.
3579 * @flags: the type of memory to allocate (see kmalloc).
3580 * @caller: function caller for debug tracking of the caller
3582 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3583 unsigned long caller)
3585 struct kmem_cache *cachep;
3588 /* If you want to save a few bytes .text space: replace
3590 * Then kmalloc uses the uninlined functions instead of the inline
3593 cachep = kmalloc_slab(size, flags);
3594 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3596 ret = slab_alloc(cachep, flags, caller);
3598 trace_kmalloc(caller, ret,
3599 size, cachep->size, flags);
3605 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3606 void *__kmalloc(size_t size, gfp_t flags)
3608 return __do_kmalloc(size, flags, _RET_IP_);
3610 EXPORT_SYMBOL(__kmalloc);
3612 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3614 return __do_kmalloc(size, flags, caller);
3616 EXPORT_SYMBOL(__kmalloc_track_caller);
3619 void *__kmalloc(size_t size, gfp_t flags)
3621 return __do_kmalloc(size, flags, 0);
3623 EXPORT_SYMBOL(__kmalloc);
3627 * kmem_cache_free - Deallocate an object
3628 * @cachep: The cache the allocation was from.
3629 * @objp: The previously allocated object.
3631 * Free an object which was previously allocated from this
3634 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3636 unsigned long flags;
3637 cachep = cache_from_obj(cachep, objp);
3641 local_irq_save(flags);
3642 debug_check_no_locks_freed(objp, cachep->object_size);
3643 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3644 debug_check_no_obj_freed(objp, cachep->object_size);
3645 __cache_free(cachep, objp, _RET_IP_);
3646 local_irq_restore(flags);
3648 trace_kmem_cache_free(_RET_IP_, objp);
3650 EXPORT_SYMBOL(kmem_cache_free);
3653 * kfree - free previously allocated memory
3654 * @objp: pointer returned by kmalloc.
3656 * If @objp is NULL, no operation is performed.
3658 * Don't free memory not originally allocated by kmalloc()
3659 * or you will run into trouble.
3661 void kfree(const void *objp)
3663 struct kmem_cache *c;
3664 unsigned long flags;
3666 trace_kfree(_RET_IP_, objp);
3668 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3670 local_irq_save(flags);
3671 kfree_debugcheck(objp);
3672 c = virt_to_cache(objp);
3673 debug_check_no_locks_freed(objp, c->object_size);
3675 debug_check_no_obj_freed(objp, c->object_size);
3676 __cache_free(c, (void *)objp, _RET_IP_);
3677 local_irq_restore(flags);
3679 EXPORT_SYMBOL(kfree);
3682 * This initializes kmem_cache_node or resizes various caches for all nodes.
3684 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3687 struct kmem_cache_node *n;
3688 struct array_cache *new_shared;
3689 struct array_cache **new_alien = NULL;
3691 for_each_online_node(node) {
3693 if (use_alien_caches) {
3694 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3700 if (cachep->shared) {
3701 new_shared = alloc_arraycache(node,
3702 cachep->shared*cachep->batchcount,
3705 free_alien_cache(new_alien);
3710 n = cachep->node[node];
3712 struct array_cache *shared = n->shared;
3714 spin_lock_irq(&n->list_lock);
3717 free_block(cachep, shared->entry,
3718 shared->avail, node);
3720 n->shared = new_shared;
3722 n->alien = new_alien;
3725 n->free_limit = (1 + nr_cpus_node(node)) *
3726 cachep->batchcount + cachep->num;
3727 spin_unlock_irq(&n->list_lock);
3729 free_alien_cache(new_alien);
3732 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3734 free_alien_cache(new_alien);
3739 kmem_cache_node_init(n);
3740 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3741 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3742 n->shared = new_shared;
3743 n->alien = new_alien;
3744 n->free_limit = (1 + nr_cpus_node(node)) *
3745 cachep->batchcount + cachep->num;
3746 cachep->node[node] = n;
3751 if (!cachep->list.next) {
3752 /* Cache is not active yet. Roll back what we did */
3755 if (cachep->node[node]) {
3756 n = cachep->node[node];
3759 free_alien_cache(n->alien);
3761 cachep->node[node] = NULL;
3769 struct ccupdate_struct {
3770 struct kmem_cache *cachep;
3771 struct array_cache *new[0];
3774 static void do_ccupdate_local(void *info)
3776 struct ccupdate_struct *new = info;
3777 struct array_cache *old;
3780 old = cpu_cache_get(new->cachep);
3782 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3783 new->new[smp_processor_id()] = old;
3786 /* Always called with the slab_mutex held */
3787 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3788 int batchcount, int shared, gfp_t gfp)
3790 struct ccupdate_struct *new;
3793 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3798 for_each_online_cpu(i) {
3799 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3802 for (i--; i >= 0; i--)
3808 new->cachep = cachep;
3810 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3813 cachep->batchcount = batchcount;
3814 cachep->limit = limit;
3815 cachep->shared = shared;
3817 for_each_online_cpu(i) {
3818 struct array_cache *ccold = new->new[i];
3821 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3822 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3823 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3827 return alloc_kmemlist(cachep, gfp);
3830 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3831 int batchcount, int shared, gfp_t gfp)
3834 struct kmem_cache *c = NULL;
3837 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3839 if (slab_state < FULL)
3842 if ((ret < 0) || !is_root_cache(cachep))
3845 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3846 for_each_memcg_cache_index(i) {
3847 c = cache_from_memcg(cachep, i);
3849 /* return value determined by the parent cache only */
3850 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3856 /* Called with slab_mutex held always */
3857 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3864 if (!is_root_cache(cachep)) {
3865 struct kmem_cache *root = memcg_root_cache(cachep);
3866 limit = root->limit;
3867 shared = root->shared;
3868 batchcount = root->batchcount;
3871 if (limit && shared && batchcount)
3874 * The head array serves three purposes:
3875 * - create a LIFO ordering, i.e. return objects that are cache-warm
3876 * - reduce the number of spinlock operations.
3877 * - reduce the number of linked list operations on the slab and
3878 * bufctl chains: array operations are cheaper.
3879 * The numbers are guessed, we should auto-tune as described by
3882 if (cachep->size > 131072)
3884 else if (cachep->size > PAGE_SIZE)
3886 else if (cachep->size > 1024)
3888 else if (cachep->size > 256)
3894 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3895 * allocation behaviour: Most allocs on one cpu, most free operations
3896 * on another cpu. For these cases, an efficient object passing between
3897 * cpus is necessary. This is provided by a shared array. The array
3898 * replaces Bonwick's magazine layer.
3899 * On uniprocessor, it's functionally equivalent (but less efficient)
3900 * to a larger limit. Thus disabled by default.
3903 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3908 * With debugging enabled, large batchcount lead to excessively long
3909 * periods with disabled local interrupts. Limit the batchcount
3914 batchcount = (limit + 1) / 2;
3916 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3918 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3919 cachep->name, -err);
3924 * Drain an array if it contains any elements taking the node lock only if
3925 * necessary. Note that the node listlock also protects the array_cache
3926 * if drain_array() is used on the shared array.
3928 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3929 struct array_cache *ac, int force, int node)
3933 if (!ac || !ac->avail)
3935 if (ac->touched && !force) {
3938 spin_lock_irq(&n->list_lock);
3940 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3941 if (tofree > ac->avail)
3942 tofree = (ac->avail + 1) / 2;
3943 free_block(cachep, ac->entry, tofree, node);
3944 ac->avail -= tofree;
3945 memmove(ac->entry, &(ac->entry[tofree]),
3946 sizeof(void *) * ac->avail);
3948 spin_unlock_irq(&n->list_lock);
3953 * cache_reap - Reclaim memory from caches.
3954 * @w: work descriptor
3956 * Called from workqueue/eventd every few seconds.
3958 * - clear the per-cpu caches for this CPU.
3959 * - return freeable pages to the main free memory pool.
3961 * If we cannot acquire the cache chain mutex then just give up - we'll try
3962 * again on the next iteration.
3964 static void cache_reap(struct work_struct *w)
3966 struct kmem_cache *searchp;
3967 struct kmem_cache_node *n;
3968 int node = numa_mem_id();
3969 struct delayed_work *work = to_delayed_work(w);
3971 if (!mutex_trylock(&slab_mutex))
3972 /* Give up. Setup the next iteration. */
3975 list_for_each_entry(searchp, &slab_caches, list) {
3979 * We only take the node lock if absolutely necessary and we
3980 * have established with reasonable certainty that
3981 * we can do some work if the lock was obtained.
3983 n = searchp->node[node];
3985 reap_alien(searchp, n);
3987 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3990 * These are racy checks but it does not matter
3991 * if we skip one check or scan twice.
3993 if (time_after(n->next_reap, jiffies))
3996 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
3998 drain_array(searchp, n, n->shared, 0, node);
4000 if (n->free_touched)
4001 n->free_touched = 0;
4005 freed = drain_freelist(searchp, n, (n->free_limit +
4006 5 * searchp->num - 1) / (5 * searchp->num));
4007 STATS_ADD_REAPED(searchp, freed);
4013 mutex_unlock(&slab_mutex);
4016 /* Set up the next iteration */
4017 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4020 #ifdef CONFIG_SLABINFO
4021 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4024 unsigned long active_objs;
4025 unsigned long num_objs;
4026 unsigned long active_slabs = 0;
4027 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4031 struct kmem_cache_node *n;
4035 for_each_online_node(node) {
4036 n = cachep->node[node];
4041 spin_lock_irq(&n->list_lock);
4043 list_for_each_entry(page, &n->slabs_full, lru) {
4044 if (page->active != cachep->num && !error)
4045 error = "slabs_full accounting error";
4046 active_objs += cachep->num;
4049 list_for_each_entry(page, &n->slabs_partial, lru) {
4050 if (page->active == cachep->num && !error)
4051 error = "slabs_partial accounting error";
4052 if (!page->active && !error)
4053 error = "slabs_partial accounting error";
4054 active_objs += page->active;
4057 list_for_each_entry(page, &n->slabs_free, lru) {
4058 if (page->active && !error)
4059 error = "slabs_free accounting error";
4062 free_objects += n->free_objects;
4064 shared_avail += n->shared->avail;
4066 spin_unlock_irq(&n->list_lock);
4068 num_slabs += active_slabs;
4069 num_objs = num_slabs * cachep->num;
4070 if (num_objs - active_objs != free_objects && !error)
4071 error = "free_objects accounting error";
4073 name = cachep->name;
4075 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4077 sinfo->active_objs = active_objs;
4078 sinfo->num_objs = num_objs;
4079 sinfo->active_slabs = active_slabs;
4080 sinfo->num_slabs = num_slabs;
4081 sinfo->shared_avail = shared_avail;
4082 sinfo->limit = cachep->limit;
4083 sinfo->batchcount = cachep->batchcount;
4084 sinfo->shared = cachep->shared;
4085 sinfo->objects_per_slab = cachep->num;
4086 sinfo->cache_order = cachep->gfporder;
4089 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4093 unsigned long high = cachep->high_mark;
4094 unsigned long allocs = cachep->num_allocations;
4095 unsigned long grown = cachep->grown;
4096 unsigned long reaped = cachep->reaped;
4097 unsigned long errors = cachep->errors;
4098 unsigned long max_freeable = cachep->max_freeable;
4099 unsigned long node_allocs = cachep->node_allocs;
4100 unsigned long node_frees = cachep->node_frees;
4101 unsigned long overflows = cachep->node_overflow;
4103 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4104 "%4lu %4lu %4lu %4lu %4lu",
4105 allocs, high, grown,
4106 reaped, errors, max_freeable, node_allocs,
4107 node_frees, overflows);
4111 unsigned long allochit = atomic_read(&cachep->allochit);
4112 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4113 unsigned long freehit = atomic_read(&cachep->freehit);
4114 unsigned long freemiss = atomic_read(&cachep->freemiss);
4116 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4117 allochit, allocmiss, freehit, freemiss);
4122 #define MAX_SLABINFO_WRITE 128
4124 * slabinfo_write - Tuning for the slab allocator
4126 * @buffer: user buffer
4127 * @count: data length
4130 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4131 size_t count, loff_t *ppos)
4133 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4134 int limit, batchcount, shared, res;
4135 struct kmem_cache *cachep;
4137 if (count > MAX_SLABINFO_WRITE)
4139 if (copy_from_user(&kbuf, buffer, count))
4141 kbuf[MAX_SLABINFO_WRITE] = '\0';
4143 tmp = strchr(kbuf, ' ');
4148 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4151 /* Find the cache in the chain of caches. */
4152 mutex_lock(&slab_mutex);
4154 list_for_each_entry(cachep, &slab_caches, list) {
4155 if (!strcmp(cachep->name, kbuf)) {
4156 if (limit < 1 || batchcount < 1 ||
4157 batchcount > limit || shared < 0) {
4160 res = do_tune_cpucache(cachep, limit,
4167 mutex_unlock(&slab_mutex);
4173 #ifdef CONFIG_DEBUG_SLAB_LEAK
4175 static void *leaks_start(struct seq_file *m, loff_t *pos)
4177 mutex_lock(&slab_mutex);
4178 return seq_list_start(&slab_caches, *pos);
4181 static inline int add_caller(unsigned long *n, unsigned long v)
4191 unsigned long *q = p + 2 * i;
4205 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4211 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4219 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4222 for (j = page->active; j < c->num; j++) {
4223 /* Skip freed item */
4224 if (slab_bufctl(page)[j] == i) {
4232 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4237 static void show_symbol(struct seq_file *m, unsigned long address)
4239 #ifdef CONFIG_KALLSYMS
4240 unsigned long offset, size;
4241 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4243 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4244 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4246 seq_printf(m, " [%s]", modname);
4250 seq_printf(m, "%p", (void *)address);
4253 static int leaks_show(struct seq_file *m, void *p)
4255 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4257 struct kmem_cache_node *n;
4259 unsigned long *x = m->private;
4263 if (!(cachep->flags & SLAB_STORE_USER))
4265 if (!(cachep->flags & SLAB_RED_ZONE))
4268 /* OK, we can do it */
4272 for_each_online_node(node) {
4273 n = cachep->node[node];
4278 spin_lock_irq(&n->list_lock);
4280 list_for_each_entry(page, &n->slabs_full, lru)
4281 handle_slab(x, cachep, page);
4282 list_for_each_entry(page, &n->slabs_partial, lru)
4283 handle_slab(x, cachep, page);
4284 spin_unlock_irq(&n->list_lock);
4286 name = cachep->name;
4288 /* Increase the buffer size */
4289 mutex_unlock(&slab_mutex);
4290 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4292 /* Too bad, we are really out */
4294 mutex_lock(&slab_mutex);
4297 *(unsigned long *)m->private = x[0] * 2;
4299 mutex_lock(&slab_mutex);
4300 /* Now make sure this entry will be retried */
4304 for (i = 0; i < x[1]; i++) {
4305 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4306 show_symbol(m, x[2*i+2]);
4313 static const struct seq_operations slabstats_op = {
4314 .start = leaks_start,
4320 static int slabstats_open(struct inode *inode, struct file *file)
4322 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4325 ret = seq_open(file, &slabstats_op);
4327 struct seq_file *m = file->private_data;
4328 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4337 static const struct file_operations proc_slabstats_operations = {
4338 .open = slabstats_open,
4340 .llseek = seq_lseek,
4341 .release = seq_release_private,
4345 static int __init slab_proc_init(void)
4347 #ifdef CONFIG_DEBUG_SLAB_LEAK
4348 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4352 module_init(slab_proc_init);
4356 * ksize - get the actual amount of memory allocated for a given object
4357 * @objp: Pointer to the object
4359 * kmalloc may internally round up allocations and return more memory
4360 * than requested. ksize() can be used to determine the actual amount of
4361 * memory allocated. The caller may use this additional memory, even though
4362 * a smaller amount of memory was initially specified with the kmalloc call.
4363 * The caller must guarantee that objp points to a valid object previously
4364 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4365 * must not be freed during the duration of the call.
4367 size_t ksize(const void *objp)
4370 if (unlikely(objp == ZERO_SIZE_PTR))
4373 return virt_to_cache(objp)->object_size;
4375 EXPORT_SYMBOL(ksize);