USB: xhci: Set stream ID to 0 after cleaning up stalls.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slab.c
blobbac0f4fcc216548fae98d352b0787cf509c0f262
1 /*
2 * linux/mm/slab.c
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
30 * kmem_cache_free.
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
38 * partial slabs
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
66 * his patch.
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 'cache_chain_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>
90 #include <linux/mm.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/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
119 #include <asm/cacheflush.h>
120 #include <asm/tlbflush.h>
121 #include <asm/page.h>
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
133 #ifdef CONFIG_DEBUG_SLAB
134 #define DEBUG 1
135 #define STATS 1
136 #define FORCED_DEBUG 1
137 #else
138 #define DEBUG 0
139 #define STATS 0
140 #define FORCED_DEBUG 0
141 #endif
143 /* Shouldn't this be in a header file somewhere? */
144 #define BYTES_PER_WORD sizeof(void *)
145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147 #ifndef ARCH_KMALLOC_MINALIGN
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
157 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #endif
160 #ifndef ARCH_SLAB_MINALIGN
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
168 #define ARCH_SLAB_MINALIGN 0
169 #endif
171 #ifndef ARCH_KMALLOC_FLAGS
172 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 #endif
175 /* Legal flag mask for kmem_cache_create(). */
176 #if DEBUG
177 # define CREATE_MASK (SLAB_RED_ZONE | \
178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_CACHE_DMA | \
180 SLAB_STORE_USER | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
184 #else
185 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
186 SLAB_CACHE_DMA | \
187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
190 #endif
193 * kmem_bufctl_t:
195 * Bufctl's are used for linking objs within a slab
196 * linked offsets.
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
211 typedef unsigned int kmem_bufctl_t;
212 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
214 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
218 * struct slab
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
224 struct slab {
225 struct list_head list;
226 unsigned long colouroff;
227 void *s_mem; /* including colour offset */
228 unsigned int inuse; /* num of objs active in slab */
229 kmem_bufctl_t free;
230 unsigned short nodeid;
234 * struct slab_rcu
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
247 * We assume struct slab_rcu can overlay struct slab when destroying.
249 struct slab_rcu {
250 struct rcu_head head;
251 struct kmem_cache *cachep;
252 void *addr;
256 * struct array_cache
258 * Purpose:
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
263 * The limit is stored in the per-cpu structure to reduce the data cache
264 * footprint.
267 struct array_cache {
268 unsigned int avail;
269 unsigned int limit;
270 unsigned int batchcount;
271 unsigned int touched;
272 spinlock_t lock;
273 void *entry[]; /*
274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
276 * the entries.
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
284 #define BOOT_CPUCACHE_ENTRIES 1
285 struct arraycache_init {
286 struct array_cache cache;
287 void *entries[BOOT_CPUCACHE_ENTRIES];
291 * The slab lists for all objects.
293 struct kmem_list3 {
294 struct list_head slabs_partial; /* partial list first, better asm code */
295 struct list_head slabs_full;
296 struct list_head slabs_free;
297 unsigned long free_objects;
298 unsigned int free_limit;
299 unsigned int colour_next; /* Per-node cache coloring */
300 spinlock_t list_lock;
301 struct array_cache *shared; /* shared per node */
302 struct array_cache **alien; /* on other nodes */
303 unsigned long next_reap; /* updated without locking */
304 int free_touched; /* updated without locking */
308 * Need this for bootstrapping a per node allocator.
310 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
311 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
312 #define CACHE_CACHE 0
313 #define SIZE_AC MAX_NUMNODES
314 #define SIZE_L3 (2 * MAX_NUMNODES)
316 static int drain_freelist(struct kmem_cache *cache,
317 struct kmem_list3 *l3, int tofree);
318 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
319 int node);
320 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
321 static void cache_reap(struct work_struct *unused);
324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
327 static __always_inline int index_of(const size_t size)
329 extern void __bad_size(void);
331 if (__builtin_constant_p(size)) {
332 int i = 0;
334 #define CACHE(x) \
335 if (size <=x) \
336 return i; \
337 else \
338 i++;
339 #include <linux/kmalloc_sizes.h>
340 #undef CACHE
341 __bad_size();
342 } else
343 __bad_size();
344 return 0;
347 static int slab_early_init = 1;
349 #define INDEX_AC index_of(sizeof(struct arraycache_init))
350 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
352 static void kmem_list3_init(struct kmem_list3 *parent)
354 INIT_LIST_HEAD(&parent->slabs_full);
355 INIT_LIST_HEAD(&parent->slabs_partial);
356 INIT_LIST_HEAD(&parent->slabs_free);
357 parent->shared = NULL;
358 parent->alien = NULL;
359 parent->colour_next = 0;
360 spin_lock_init(&parent->list_lock);
361 parent->free_objects = 0;
362 parent->free_touched = 0;
365 #define MAKE_LIST(cachep, listp, slab, nodeid) \
366 do { \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
369 } while (0)
371 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
372 do { \
373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
376 } while (0)
378 #define CFLGS_OFF_SLAB (0x80000000UL)
379 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
381 #define BATCHREFILL_LIMIT 16
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
386 * OTOH the cpuarrays can contain lots of objects,
387 * which could lock up otherwise freeable slabs.
389 #define REAPTIMEOUT_CPUC (2*HZ)
390 #define REAPTIMEOUT_LIST3 (4*HZ)
392 #if STATS
393 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
394 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396 #define STATS_INC_GROWN(x) ((x)->grown++)
397 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
398 #define STATS_SET_HIGH(x) \
399 do { \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
402 } while (0)
403 #define STATS_INC_ERR(x) ((x)->errors++)
404 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
405 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
406 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
407 #define STATS_SET_FREEABLE(x, i) \
408 do { \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
411 } while (0)
412 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
416 #else
417 #define STATS_INC_ACTIVE(x) do { } while (0)
418 #define STATS_DEC_ACTIVE(x) do { } while (0)
419 #define STATS_INC_ALLOCED(x) do { } while (0)
420 #define STATS_INC_GROWN(x) do { } while (0)
421 #define STATS_ADD_REAPED(x,y) do { } while (0)
422 #define STATS_SET_HIGH(x) do { } while (0)
423 #define STATS_INC_ERR(x) do { } while (0)
424 #define STATS_INC_NODEALLOCS(x) do { } while (0)
425 #define STATS_INC_NODEFREES(x) do { } while (0)
426 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
427 #define STATS_SET_FREEABLE(x, i) do { } while (0)
428 #define STATS_INC_ALLOCHIT(x) do { } while (0)
429 #define STATS_INC_ALLOCMISS(x) do { } while (0)
430 #define STATS_INC_FREEHIT(x) do { } while (0)
431 #define STATS_INC_FREEMISS(x) do { } while (0)
432 #endif
434 #if DEBUG
437 * memory layout of objects:
438 * 0 : objp
439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
443 * redzone word.
444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
449 static int obj_offset(struct kmem_cache *cachep)
451 return cachep->obj_offset;
454 static int obj_size(struct kmem_cache *cachep)
456 return cachep->obj_size;
459 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
462 return (unsigned long long*) (objp + obj_offset(cachep) -
463 sizeof(unsigned long long));
466 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
468 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
469 if (cachep->flags & SLAB_STORE_USER)
470 return (unsigned long long *)(objp + cachep->buffer_size -
471 sizeof(unsigned long long) -
472 REDZONE_ALIGN);
473 return (unsigned long long *) (objp + cachep->buffer_size -
474 sizeof(unsigned long long));
477 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
479 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
480 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
483 #else
485 #define obj_offset(x) 0
486 #define obj_size(cachep) (cachep->buffer_size)
487 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
489 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
491 #endif
493 #ifdef CONFIG_TRACING
494 size_t slab_buffer_size(struct kmem_cache *cachep)
496 return cachep->buffer_size;
498 EXPORT_SYMBOL(slab_buffer_size);
499 #endif
502 * Do not go above this order unless 0 objects fit into the slab.
504 #define BREAK_GFP_ORDER_HI 1
505 #define BREAK_GFP_ORDER_LO 0
506 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
513 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
515 page->lru.next = (struct list_head *)cache;
518 static inline struct kmem_cache *page_get_cache(struct page *page)
520 page = compound_head(page);
521 BUG_ON(!PageSlab(page));
522 return (struct kmem_cache *)page->lru.next;
525 static inline void page_set_slab(struct page *page, struct slab *slab)
527 page->lru.prev = (struct list_head *)slab;
530 static inline struct slab *page_get_slab(struct page *page)
532 BUG_ON(!PageSlab(page));
533 return (struct slab *)page->lru.prev;
536 static inline struct kmem_cache *virt_to_cache(const void *obj)
538 struct page *page = virt_to_head_page(obj);
539 return page_get_cache(page);
542 static inline struct slab *virt_to_slab(const void *obj)
544 struct page *page = virt_to_head_page(obj);
545 return page_get_slab(page);
548 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
549 unsigned int idx)
551 return slab->s_mem + cache->buffer_size * idx;
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
560 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
561 const struct slab *slab, void *obj)
563 u32 offset = (obj - slab->s_mem);
564 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
570 struct cache_sizes malloc_sizes[] = {
571 #define CACHE(x) { .cs_size = (x) },
572 #include <linux/kmalloc_sizes.h>
573 CACHE(ULONG_MAX)
574 #undef CACHE
576 EXPORT_SYMBOL(malloc_sizes);
578 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
579 struct cache_names {
580 char *name;
581 char *name_dma;
584 static struct cache_names __initdata cache_names[] = {
585 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586 #include <linux/kmalloc_sizes.h>
587 {NULL,}
588 #undef CACHE
591 static struct arraycache_init initarray_cache __initdata =
592 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
593 static struct arraycache_init initarray_generic =
594 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
596 /* internal cache of cache description objs */
597 static struct kmem_cache cache_cache = {
598 .batchcount = 1,
599 .limit = BOOT_CPUCACHE_ENTRIES,
600 .shared = 1,
601 .buffer_size = sizeof(struct kmem_cache),
602 .name = "kmem_cache",
605 #define BAD_ALIEN_MAGIC 0x01020304ul
608 * chicken and egg problem: delay the per-cpu array allocation
609 * until the general caches are up.
611 static enum {
612 NONE,
613 PARTIAL_AC,
614 PARTIAL_L3,
615 EARLY,
616 FULL
617 } g_cpucache_up;
620 * used by boot code to determine if it can use slab based allocator
622 int slab_is_available(void)
624 return g_cpucache_up >= EARLY;
627 #ifdef CONFIG_LOCKDEP
630 * Slab sometimes uses the kmalloc slabs to store the slab headers
631 * for other slabs "off slab".
632 * The locking for this is tricky in that it nests within the locks
633 * of all other slabs in a few places; to deal with this special
634 * locking we put on-slab caches into a separate lock-class.
636 * We set lock class for alien array caches which are up during init.
637 * The lock annotation will be lost if all cpus of a node goes down and
638 * then comes back up during hotplug
640 static struct lock_class_key on_slab_l3_key;
641 static struct lock_class_key on_slab_alc_key;
643 static void init_node_lock_keys(int q)
645 struct cache_sizes *s = malloc_sizes;
647 if (g_cpucache_up != FULL)
648 return;
650 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
651 struct array_cache **alc;
652 struct kmem_list3 *l3;
653 int r;
655 l3 = s->cs_cachep->nodelists[q];
656 if (!l3 || OFF_SLAB(s->cs_cachep))
657 continue;
658 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
659 alc = l3->alien;
661 * FIXME: This check for BAD_ALIEN_MAGIC
662 * should go away when common slab code is taught to
663 * work even without alien caches.
664 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
665 * for alloc_alien_cache,
667 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
668 continue;
669 for_each_node(r) {
670 if (alc[r])
671 lockdep_set_class(&alc[r]->lock,
672 &on_slab_alc_key);
677 static inline void init_lock_keys(void)
679 int node;
681 for_each_node(node)
682 init_node_lock_keys(node);
684 #else
685 static void init_node_lock_keys(int q)
689 static inline void init_lock_keys(void)
692 #endif
695 * Guard access to the cache-chain.
697 static DEFINE_MUTEX(cache_chain_mutex);
698 static struct list_head cache_chain;
700 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
702 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
704 return cachep->array[smp_processor_id()];
707 static inline struct kmem_cache *__find_general_cachep(size_t size,
708 gfp_t gfpflags)
710 struct cache_sizes *csizep = malloc_sizes;
712 #if DEBUG
713 /* This happens if someone tries to call
714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
717 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
718 #endif
719 if (!size)
720 return ZERO_SIZE_PTR;
722 while (size > csizep->cs_size)
723 csizep++;
726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
730 #ifdef CONFIG_ZONE_DMA
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
733 #endif
734 return csizep->cs_cachep;
737 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
739 return __find_general_cachep(size, gfpflags);
742 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
748 * Calculate the number of objects and left-over bytes for a given buffer size.
750 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
752 unsigned int *num)
754 int nr_objs;
755 size_t mgmt_size;
756 size_t slab_size = PAGE_SIZE << gfporder;
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
761 * slab is used for:
763 * - The struct slab
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
773 if (flags & CFLGS_OFF_SLAB) {
774 mgmt_size = 0;
775 nr_objs = slab_size / buffer_size;
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
779 } else {
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
786 * into account.
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
796 > slab_size)
797 nr_objs--;
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
802 mgmt_size = slab_mgmt_size(nr_objs, align);
804 *num = nr_objs;
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
808 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
810 static void __slab_error(const char *function, struct kmem_cache *cachep,
811 char *msg)
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
814 function, cachep->name, msg);
815 dump_stack();
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
823 * line
826 static int use_alien_caches __read_mostly = 1;
827 static int __init noaliencache_setup(char *s)
829 use_alien_caches = 0;
830 return 1;
832 __setup("noaliencache", noaliencache_setup);
834 #ifdef CONFIG_NUMA
836 * Special reaping functions for NUMA systems called from cache_reap().
837 * These take care of doing round robin flushing of alien caches (containing
838 * objects freed on different nodes from which they were allocated) and the
839 * flushing of remote pcps by calling drain_node_pages.
841 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
843 static void init_reap_node(int cpu)
845 int node;
847 node = next_node(cpu_to_node(cpu), node_online_map);
848 if (node == MAX_NUMNODES)
849 node = first_node(node_online_map);
851 per_cpu(slab_reap_node, cpu) = node;
854 static void next_reap_node(void)
856 int node = __get_cpu_var(slab_reap_node);
858 node = next_node(node, node_online_map);
859 if (unlikely(node >= MAX_NUMNODES))
860 node = first_node(node_online_map);
861 __get_cpu_var(slab_reap_node) = node;
864 #else
865 #define init_reap_node(cpu) do { } while (0)
866 #define next_reap_node(void) do { } while (0)
867 #endif
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
874 * lock.
876 static void __cpuinit start_cpu_timer(int cpu)
878 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
883 * at that time.
885 if (keventd_up() && reap_work->work.func == NULL) {
886 init_reap_node(cpu);
887 INIT_DELAYED_WORK(reap_work, cache_reap);
888 schedule_delayed_work_on(cpu, reap_work,
889 __round_jiffies_relative(HZ, cpu));
893 static struct array_cache *alloc_arraycache(int node, int entries,
894 int batchcount, gfp_t gfp)
896 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
897 struct array_cache *nc = NULL;
899 nc = kmalloc_node(memsize, gfp, node);
901 * The array_cache structures contain pointers to free object.
902 * However, when such objects are allocated or transfered to another
903 * cache the pointers are not cleared and they could be counted as
904 * valid references during a kmemleak scan. Therefore, kmemleak must
905 * not scan such objects.
907 kmemleak_no_scan(nc);
908 if (nc) {
909 nc->avail = 0;
910 nc->limit = entries;
911 nc->batchcount = batchcount;
912 nc->touched = 0;
913 spin_lock_init(&nc->lock);
915 return nc;
919 * Transfer objects in one arraycache to another.
920 * Locking must be handled by the caller.
922 * Return the number of entries transferred.
924 static int transfer_objects(struct array_cache *to,
925 struct array_cache *from, unsigned int max)
927 /* Figure out how many entries to transfer */
928 int nr = min(min(from->avail, max), to->limit - to->avail);
930 if (!nr)
931 return 0;
933 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
934 sizeof(void *) *nr);
936 from->avail -= nr;
937 to->avail += nr;
938 return nr;
941 #ifndef CONFIG_NUMA
943 #define drain_alien_cache(cachep, alien) do { } while (0)
944 #define reap_alien(cachep, l3) do { } while (0)
946 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
948 return (struct array_cache **)BAD_ALIEN_MAGIC;
951 static inline void free_alien_cache(struct array_cache **ac_ptr)
955 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
957 return 0;
960 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
961 gfp_t flags)
963 return NULL;
966 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
967 gfp_t flags, int nodeid)
969 return NULL;
972 #else /* CONFIG_NUMA */
974 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
975 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
977 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
979 struct array_cache **ac_ptr;
980 int memsize = sizeof(void *) * nr_node_ids;
981 int i;
983 if (limit > 1)
984 limit = 12;
985 ac_ptr = kzalloc_node(memsize, gfp, node);
986 if (ac_ptr) {
987 for_each_node(i) {
988 if (i == node || !node_online(i))
989 continue;
990 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
991 if (!ac_ptr[i]) {
992 for (i--; i >= 0; i--)
993 kfree(ac_ptr[i]);
994 kfree(ac_ptr);
995 return NULL;
999 return ac_ptr;
1002 static void free_alien_cache(struct array_cache **ac_ptr)
1004 int i;
1006 if (!ac_ptr)
1007 return;
1008 for_each_node(i)
1009 kfree(ac_ptr[i]);
1010 kfree(ac_ptr);
1013 static void __drain_alien_cache(struct kmem_cache *cachep,
1014 struct array_cache *ac, int node)
1016 struct kmem_list3 *rl3 = cachep->nodelists[node];
1018 if (ac->avail) {
1019 spin_lock(&rl3->list_lock);
1021 * Stuff objects into the remote nodes shared array first.
1022 * That way we could avoid the overhead of putting the objects
1023 * into the free lists and getting them back later.
1025 if (rl3->shared)
1026 transfer_objects(rl3->shared, ac, ac->limit);
1028 free_block(cachep, ac->entry, ac->avail, node);
1029 ac->avail = 0;
1030 spin_unlock(&rl3->list_lock);
1035 * Called from cache_reap() to regularly drain alien caches round robin.
1037 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1039 int node = __get_cpu_var(slab_reap_node);
1041 if (l3->alien) {
1042 struct array_cache *ac = l3->alien[node];
1044 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1045 __drain_alien_cache(cachep, ac, node);
1046 spin_unlock_irq(&ac->lock);
1051 static void drain_alien_cache(struct kmem_cache *cachep,
1052 struct array_cache **alien)
1054 int i = 0;
1055 struct array_cache *ac;
1056 unsigned long flags;
1058 for_each_online_node(i) {
1059 ac = alien[i];
1060 if (ac) {
1061 spin_lock_irqsave(&ac->lock, flags);
1062 __drain_alien_cache(cachep, ac, i);
1063 spin_unlock_irqrestore(&ac->lock, flags);
1068 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1070 struct slab *slabp = virt_to_slab(objp);
1071 int nodeid = slabp->nodeid;
1072 struct kmem_list3 *l3;
1073 struct array_cache *alien = NULL;
1074 int node;
1076 node = numa_node_id();
1079 * Make sure we are not freeing a object from another node to the array
1080 * cache on this cpu.
1082 if (likely(slabp->nodeid == node))
1083 return 0;
1085 l3 = cachep->nodelists[node];
1086 STATS_INC_NODEFREES(cachep);
1087 if (l3->alien && l3->alien[nodeid]) {
1088 alien = l3->alien[nodeid];
1089 spin_lock(&alien->lock);
1090 if (unlikely(alien->avail == alien->limit)) {
1091 STATS_INC_ACOVERFLOW(cachep);
1092 __drain_alien_cache(cachep, alien, nodeid);
1094 alien->entry[alien->avail++] = objp;
1095 spin_unlock(&alien->lock);
1096 } else {
1097 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1098 free_block(cachep, &objp, 1, nodeid);
1099 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1101 return 1;
1103 #endif
1105 static void __cpuinit cpuup_canceled(long cpu)
1107 struct kmem_cache *cachep;
1108 struct kmem_list3 *l3 = NULL;
1109 int node = cpu_to_node(cpu);
1110 const struct cpumask *mask = cpumask_of_node(node);
1112 list_for_each_entry(cachep, &cache_chain, next) {
1113 struct array_cache *nc;
1114 struct array_cache *shared;
1115 struct array_cache **alien;
1117 /* cpu is dead; no one can alloc from it. */
1118 nc = cachep->array[cpu];
1119 cachep->array[cpu] = NULL;
1120 l3 = cachep->nodelists[node];
1122 if (!l3)
1123 goto free_array_cache;
1125 spin_lock_irq(&l3->list_lock);
1127 /* Free limit for this kmem_list3 */
1128 l3->free_limit -= cachep->batchcount;
1129 if (nc)
1130 free_block(cachep, nc->entry, nc->avail, node);
1132 if (!cpumask_empty(mask)) {
1133 spin_unlock_irq(&l3->list_lock);
1134 goto free_array_cache;
1137 shared = l3->shared;
1138 if (shared) {
1139 free_block(cachep, shared->entry,
1140 shared->avail, node);
1141 l3->shared = NULL;
1144 alien = l3->alien;
1145 l3->alien = NULL;
1147 spin_unlock_irq(&l3->list_lock);
1149 kfree(shared);
1150 if (alien) {
1151 drain_alien_cache(cachep, alien);
1152 free_alien_cache(alien);
1154 free_array_cache:
1155 kfree(nc);
1158 * In the previous loop, all the objects were freed to
1159 * the respective cache's slabs, now we can go ahead and
1160 * shrink each nodelist to its limit.
1162 list_for_each_entry(cachep, &cache_chain, next) {
1163 l3 = cachep->nodelists[node];
1164 if (!l3)
1165 continue;
1166 drain_freelist(cachep, l3, l3->free_objects);
1170 static int __cpuinit cpuup_prepare(long cpu)
1172 struct kmem_cache *cachep;
1173 struct kmem_list3 *l3 = NULL;
1174 int node = cpu_to_node(cpu);
1175 const int memsize = sizeof(struct kmem_list3);
1178 * We need to do this right in the beginning since
1179 * alloc_arraycache's are going to use this list.
1180 * kmalloc_node allows us to add the slab to the right
1181 * kmem_list3 and not this cpu's kmem_list3
1184 list_for_each_entry(cachep, &cache_chain, next) {
1186 * Set up the size64 kmemlist for cpu before we can
1187 * begin anything. Make sure some other cpu on this
1188 * node has not already allocated this
1190 if (!cachep->nodelists[node]) {
1191 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1192 if (!l3)
1193 goto bad;
1194 kmem_list3_init(l3);
1195 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1196 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1199 * The l3s don't come and go as CPUs come and
1200 * go. cache_chain_mutex is sufficient
1201 * protection here.
1203 cachep->nodelists[node] = l3;
1206 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1207 cachep->nodelists[node]->free_limit =
1208 (1 + nr_cpus_node(node)) *
1209 cachep->batchcount + cachep->num;
1210 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1214 * Now we can go ahead with allocating the shared arrays and
1215 * array caches
1217 list_for_each_entry(cachep, &cache_chain, next) {
1218 struct array_cache *nc;
1219 struct array_cache *shared = NULL;
1220 struct array_cache **alien = NULL;
1222 nc = alloc_arraycache(node, cachep->limit,
1223 cachep->batchcount, GFP_KERNEL);
1224 if (!nc)
1225 goto bad;
1226 if (cachep->shared) {
1227 shared = alloc_arraycache(node,
1228 cachep->shared * cachep->batchcount,
1229 0xbaadf00d, GFP_KERNEL);
1230 if (!shared) {
1231 kfree(nc);
1232 goto bad;
1235 if (use_alien_caches) {
1236 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1237 if (!alien) {
1238 kfree(shared);
1239 kfree(nc);
1240 goto bad;
1243 cachep->array[cpu] = nc;
1244 l3 = cachep->nodelists[node];
1245 BUG_ON(!l3);
1247 spin_lock_irq(&l3->list_lock);
1248 if (!l3->shared) {
1250 * We are serialised from CPU_DEAD or
1251 * CPU_UP_CANCELLED by the cpucontrol lock
1253 l3->shared = shared;
1254 shared = NULL;
1256 #ifdef CONFIG_NUMA
1257 if (!l3->alien) {
1258 l3->alien = alien;
1259 alien = NULL;
1261 #endif
1262 spin_unlock_irq(&l3->list_lock);
1263 kfree(shared);
1264 free_alien_cache(alien);
1266 init_node_lock_keys(node);
1268 return 0;
1269 bad:
1270 cpuup_canceled(cpu);
1271 return -ENOMEM;
1274 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1275 unsigned long action, void *hcpu)
1277 long cpu = (long)hcpu;
1278 int err = 0;
1280 switch (action) {
1281 case CPU_UP_PREPARE:
1282 case CPU_UP_PREPARE_FROZEN:
1283 mutex_lock(&cache_chain_mutex);
1284 err = cpuup_prepare(cpu);
1285 mutex_unlock(&cache_chain_mutex);
1286 break;
1287 case CPU_ONLINE:
1288 case CPU_ONLINE_FROZEN:
1289 start_cpu_timer(cpu);
1290 break;
1291 #ifdef CONFIG_HOTPLUG_CPU
1292 case CPU_DOWN_PREPARE:
1293 case CPU_DOWN_PREPARE_FROZEN:
1295 * Shutdown cache reaper. Note that the cache_chain_mutex is
1296 * held so that if cache_reap() is invoked it cannot do
1297 * anything expensive but will only modify reap_work
1298 * and reschedule the timer.
1300 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
1301 /* Now the cache_reaper is guaranteed to be not running. */
1302 per_cpu(slab_reap_work, cpu).work.func = NULL;
1303 break;
1304 case CPU_DOWN_FAILED:
1305 case CPU_DOWN_FAILED_FROZEN:
1306 start_cpu_timer(cpu);
1307 break;
1308 case CPU_DEAD:
1309 case CPU_DEAD_FROZEN:
1311 * Even if all the cpus of a node are down, we don't free the
1312 * kmem_list3 of any cache. This to avoid a race between
1313 * cpu_down, and a kmalloc allocation from another cpu for
1314 * memory from the node of the cpu going down. The list3
1315 * structure is usually allocated from kmem_cache_create() and
1316 * gets destroyed at kmem_cache_destroy().
1318 /* fall through */
1319 #endif
1320 case CPU_UP_CANCELED:
1321 case CPU_UP_CANCELED_FROZEN:
1322 mutex_lock(&cache_chain_mutex);
1323 cpuup_canceled(cpu);
1324 mutex_unlock(&cache_chain_mutex);
1325 break;
1327 return err ? NOTIFY_BAD : NOTIFY_OK;
1330 static struct notifier_block __cpuinitdata cpucache_notifier = {
1331 &cpuup_callback, NULL, 0
1335 * swap the static kmem_list3 with kmalloced memory
1337 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1338 int nodeid)
1340 struct kmem_list3 *ptr;
1342 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1343 BUG_ON(!ptr);
1345 memcpy(ptr, list, sizeof(struct kmem_list3));
1347 * Do not assume that spinlocks can be initialized via memcpy:
1349 spin_lock_init(&ptr->list_lock);
1351 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1352 cachep->nodelists[nodeid] = ptr;
1356 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1357 * size of kmem_list3.
1359 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1361 int node;
1363 for_each_online_node(node) {
1364 cachep->nodelists[node] = &initkmem_list3[index + node];
1365 cachep->nodelists[node]->next_reap = jiffies +
1366 REAPTIMEOUT_LIST3 +
1367 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1372 * Initialisation. Called after the page allocator have been initialised and
1373 * before smp_init().
1375 void __init kmem_cache_init(void)
1377 size_t left_over;
1378 struct cache_sizes *sizes;
1379 struct cache_names *names;
1380 int i;
1381 int order;
1382 int node;
1384 if (num_possible_nodes() == 1)
1385 use_alien_caches = 0;
1387 for (i = 0; i < NUM_INIT_LISTS; i++) {
1388 kmem_list3_init(&initkmem_list3[i]);
1389 if (i < MAX_NUMNODES)
1390 cache_cache.nodelists[i] = NULL;
1392 set_up_list3s(&cache_cache, CACHE_CACHE);
1395 * Fragmentation resistance on low memory - only use bigger
1396 * page orders on machines with more than 32MB of memory.
1398 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1399 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1401 /* Bootstrap is tricky, because several objects are allocated
1402 * from caches that do not exist yet:
1403 * 1) initialize the cache_cache cache: it contains the struct
1404 * kmem_cache structures of all caches, except cache_cache itself:
1405 * cache_cache is statically allocated.
1406 * Initially an __init data area is used for the head array and the
1407 * kmem_list3 structures, it's replaced with a kmalloc allocated
1408 * array at the end of the bootstrap.
1409 * 2) Create the first kmalloc cache.
1410 * The struct kmem_cache for the new cache is allocated normally.
1411 * An __init data area is used for the head array.
1412 * 3) Create the remaining kmalloc caches, with minimally sized
1413 * head arrays.
1414 * 4) Replace the __init data head arrays for cache_cache and the first
1415 * kmalloc cache with kmalloc allocated arrays.
1416 * 5) Replace the __init data for kmem_list3 for cache_cache and
1417 * the other cache's with kmalloc allocated memory.
1418 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1421 node = numa_node_id();
1423 /* 1) create the cache_cache */
1424 INIT_LIST_HEAD(&cache_chain);
1425 list_add(&cache_cache.next, &cache_chain);
1426 cache_cache.colour_off = cache_line_size();
1427 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1428 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1431 * struct kmem_cache size depends on nr_node_ids, which
1432 * can be less than MAX_NUMNODES.
1434 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1435 nr_node_ids * sizeof(struct kmem_list3 *);
1436 #if DEBUG
1437 cache_cache.obj_size = cache_cache.buffer_size;
1438 #endif
1439 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1440 cache_line_size());
1441 cache_cache.reciprocal_buffer_size =
1442 reciprocal_value(cache_cache.buffer_size);
1444 for (order = 0; order < MAX_ORDER; order++) {
1445 cache_estimate(order, cache_cache.buffer_size,
1446 cache_line_size(), 0, &left_over, &cache_cache.num);
1447 if (cache_cache.num)
1448 break;
1450 BUG_ON(!cache_cache.num);
1451 cache_cache.gfporder = order;
1452 cache_cache.colour = left_over / cache_cache.colour_off;
1453 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1454 sizeof(struct slab), cache_line_size());
1456 /* 2+3) create the kmalloc caches */
1457 sizes = malloc_sizes;
1458 names = cache_names;
1461 * Initialize the caches that provide memory for the array cache and the
1462 * kmem_list3 structures first. Without this, further allocations will
1463 * bug.
1466 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1467 sizes[INDEX_AC].cs_size,
1468 ARCH_KMALLOC_MINALIGN,
1469 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1470 NULL);
1472 if (INDEX_AC != INDEX_L3) {
1473 sizes[INDEX_L3].cs_cachep =
1474 kmem_cache_create(names[INDEX_L3].name,
1475 sizes[INDEX_L3].cs_size,
1476 ARCH_KMALLOC_MINALIGN,
1477 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1478 NULL);
1481 slab_early_init = 0;
1483 while (sizes->cs_size != ULONG_MAX) {
1485 * For performance, all the general caches are L1 aligned.
1486 * This should be particularly beneficial on SMP boxes, as it
1487 * eliminates "false sharing".
1488 * Note for systems short on memory removing the alignment will
1489 * allow tighter packing of the smaller caches.
1491 if (!sizes->cs_cachep) {
1492 sizes->cs_cachep = kmem_cache_create(names->name,
1493 sizes->cs_size,
1494 ARCH_KMALLOC_MINALIGN,
1495 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1496 NULL);
1498 #ifdef CONFIG_ZONE_DMA
1499 sizes->cs_dmacachep = kmem_cache_create(
1500 names->name_dma,
1501 sizes->cs_size,
1502 ARCH_KMALLOC_MINALIGN,
1503 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1504 SLAB_PANIC,
1505 NULL);
1506 #endif
1507 sizes++;
1508 names++;
1510 /* 4) Replace the bootstrap head arrays */
1512 struct array_cache *ptr;
1514 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1516 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1517 memcpy(ptr, cpu_cache_get(&cache_cache),
1518 sizeof(struct arraycache_init));
1520 * Do not assume that spinlocks can be initialized via memcpy:
1522 spin_lock_init(&ptr->lock);
1524 cache_cache.array[smp_processor_id()] = ptr;
1526 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1528 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1529 != &initarray_generic.cache);
1530 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1531 sizeof(struct arraycache_init));
1533 * Do not assume that spinlocks can be initialized via memcpy:
1535 spin_lock_init(&ptr->lock);
1537 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1538 ptr;
1540 /* 5) Replace the bootstrap kmem_list3's */
1542 int nid;
1544 for_each_online_node(nid) {
1545 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1547 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1548 &initkmem_list3[SIZE_AC + nid], nid);
1550 if (INDEX_AC != INDEX_L3) {
1551 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1552 &initkmem_list3[SIZE_L3 + nid], nid);
1557 g_cpucache_up = EARLY;
1560 void __init kmem_cache_init_late(void)
1562 struct kmem_cache *cachep;
1564 /* 6) resize the head arrays to their final sizes */
1565 mutex_lock(&cache_chain_mutex);
1566 list_for_each_entry(cachep, &cache_chain, next)
1567 if (enable_cpucache(cachep, GFP_NOWAIT))
1568 BUG();
1569 mutex_unlock(&cache_chain_mutex);
1571 /* Done! */
1572 g_cpucache_up = FULL;
1574 /* Annotate slab for lockdep -- annotate the malloc caches */
1575 init_lock_keys();
1578 * Register a cpu startup notifier callback that initializes
1579 * cpu_cache_get for all new cpus
1581 register_cpu_notifier(&cpucache_notifier);
1584 * The reap timers are started later, with a module init call: That part
1585 * of the kernel is not yet operational.
1589 static int __init cpucache_init(void)
1591 int cpu;
1594 * Register the timers that return unneeded pages to the page allocator
1596 for_each_online_cpu(cpu)
1597 start_cpu_timer(cpu);
1598 return 0;
1600 __initcall(cpucache_init);
1603 * Interface to system's page allocator. No need to hold the cache-lock.
1605 * If we requested dmaable memory, we will get it. Even if we
1606 * did not request dmaable memory, we might get it, but that
1607 * would be relatively rare and ignorable.
1609 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1611 struct page *page;
1612 int nr_pages;
1613 int i;
1615 #ifndef CONFIG_MMU
1617 * Nommu uses slab's for process anonymous memory allocations, and thus
1618 * requires __GFP_COMP to properly refcount higher order allocations
1620 flags |= __GFP_COMP;
1621 #endif
1623 flags |= cachep->gfpflags;
1624 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1625 flags |= __GFP_RECLAIMABLE;
1627 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1628 if (!page)
1629 return NULL;
1631 nr_pages = (1 << cachep->gfporder);
1632 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1633 add_zone_page_state(page_zone(page),
1634 NR_SLAB_RECLAIMABLE, nr_pages);
1635 else
1636 add_zone_page_state(page_zone(page),
1637 NR_SLAB_UNRECLAIMABLE, nr_pages);
1638 for (i = 0; i < nr_pages; i++)
1639 __SetPageSlab(page + i);
1641 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1642 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1644 if (cachep->ctor)
1645 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1646 else
1647 kmemcheck_mark_unallocated_pages(page, nr_pages);
1650 return page_address(page);
1654 * Interface to system's page release.
1656 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1658 unsigned long i = (1 << cachep->gfporder);
1659 struct page *page = virt_to_page(addr);
1660 const unsigned long nr_freed = i;
1662 kmemcheck_free_shadow(page, cachep->gfporder);
1664 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1665 sub_zone_page_state(page_zone(page),
1666 NR_SLAB_RECLAIMABLE, nr_freed);
1667 else
1668 sub_zone_page_state(page_zone(page),
1669 NR_SLAB_UNRECLAIMABLE, nr_freed);
1670 while (i--) {
1671 BUG_ON(!PageSlab(page));
1672 __ClearPageSlab(page);
1673 page++;
1675 if (current->reclaim_state)
1676 current->reclaim_state->reclaimed_slab += nr_freed;
1677 free_pages((unsigned long)addr, cachep->gfporder);
1680 static void kmem_rcu_free(struct rcu_head *head)
1682 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1683 struct kmem_cache *cachep = slab_rcu->cachep;
1685 kmem_freepages(cachep, slab_rcu->addr);
1686 if (OFF_SLAB(cachep))
1687 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1690 #if DEBUG
1692 #ifdef CONFIG_DEBUG_PAGEALLOC
1693 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1694 unsigned long caller)
1696 int size = obj_size(cachep);
1698 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1700 if (size < 5 * sizeof(unsigned long))
1701 return;
1703 *addr++ = 0x12345678;
1704 *addr++ = caller;
1705 *addr++ = smp_processor_id();
1706 size -= 3 * sizeof(unsigned long);
1708 unsigned long *sptr = &caller;
1709 unsigned long svalue;
1711 while (!kstack_end(sptr)) {
1712 svalue = *sptr++;
1713 if (kernel_text_address(svalue)) {
1714 *addr++ = svalue;
1715 size -= sizeof(unsigned long);
1716 if (size <= sizeof(unsigned long))
1717 break;
1722 *addr++ = 0x87654321;
1724 #endif
1726 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1728 int size = obj_size(cachep);
1729 addr = &((char *)addr)[obj_offset(cachep)];
1731 memset(addr, val, size);
1732 *(unsigned char *)(addr + size - 1) = POISON_END;
1735 static void dump_line(char *data, int offset, int limit)
1737 int i;
1738 unsigned char error = 0;
1739 int bad_count = 0;
1741 printk(KERN_ERR "%03x:", offset);
1742 for (i = 0; i < limit; i++) {
1743 if (data[offset + i] != POISON_FREE) {
1744 error = data[offset + i];
1745 bad_count++;
1747 printk(" %02x", (unsigned char)data[offset + i]);
1749 printk("\n");
1751 if (bad_count == 1) {
1752 error ^= POISON_FREE;
1753 if (!(error & (error - 1))) {
1754 printk(KERN_ERR "Single bit error detected. Probably "
1755 "bad RAM.\n");
1756 #ifdef CONFIG_X86
1757 printk(KERN_ERR "Run memtest86+ or a similar memory "
1758 "test tool.\n");
1759 #else
1760 printk(KERN_ERR "Run a memory test tool.\n");
1761 #endif
1765 #endif
1767 #if DEBUG
1769 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1771 int i, size;
1772 char *realobj;
1774 if (cachep->flags & SLAB_RED_ZONE) {
1775 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1776 *dbg_redzone1(cachep, objp),
1777 *dbg_redzone2(cachep, objp));
1780 if (cachep->flags & SLAB_STORE_USER) {
1781 printk(KERN_ERR "Last user: [<%p>]",
1782 *dbg_userword(cachep, objp));
1783 print_symbol("(%s)",
1784 (unsigned long)*dbg_userword(cachep, objp));
1785 printk("\n");
1787 realobj = (char *)objp + obj_offset(cachep);
1788 size = obj_size(cachep);
1789 for (i = 0; i < size && lines; i += 16, lines--) {
1790 int limit;
1791 limit = 16;
1792 if (i + limit > size)
1793 limit = size - i;
1794 dump_line(realobj, i, limit);
1798 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1800 char *realobj;
1801 int size, i;
1802 int lines = 0;
1804 realobj = (char *)objp + obj_offset(cachep);
1805 size = obj_size(cachep);
1807 for (i = 0; i < size; i++) {
1808 char exp = POISON_FREE;
1809 if (i == size - 1)
1810 exp = POISON_END;
1811 if (realobj[i] != exp) {
1812 int limit;
1813 /* Mismatch ! */
1814 /* Print header */
1815 if (lines == 0) {
1816 printk(KERN_ERR
1817 "Slab corruption: %s start=%p, len=%d\n",
1818 cachep->name, realobj, size);
1819 print_objinfo(cachep, objp, 0);
1821 /* Hexdump the affected line */
1822 i = (i / 16) * 16;
1823 limit = 16;
1824 if (i + limit > size)
1825 limit = size - i;
1826 dump_line(realobj, i, limit);
1827 i += 16;
1828 lines++;
1829 /* Limit to 5 lines */
1830 if (lines > 5)
1831 break;
1834 if (lines != 0) {
1835 /* Print some data about the neighboring objects, if they
1836 * exist:
1838 struct slab *slabp = virt_to_slab(objp);
1839 unsigned int objnr;
1841 objnr = obj_to_index(cachep, slabp, objp);
1842 if (objnr) {
1843 objp = index_to_obj(cachep, slabp, objnr - 1);
1844 realobj = (char *)objp + obj_offset(cachep);
1845 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1846 realobj, size);
1847 print_objinfo(cachep, objp, 2);
1849 if (objnr + 1 < cachep->num) {
1850 objp = index_to_obj(cachep, slabp, objnr + 1);
1851 realobj = (char *)objp + obj_offset(cachep);
1852 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1853 realobj, size);
1854 print_objinfo(cachep, objp, 2);
1858 #endif
1860 #if DEBUG
1861 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1863 int i;
1864 for (i = 0; i < cachep->num; i++) {
1865 void *objp = index_to_obj(cachep, slabp, i);
1867 if (cachep->flags & SLAB_POISON) {
1868 #ifdef CONFIG_DEBUG_PAGEALLOC
1869 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1870 OFF_SLAB(cachep))
1871 kernel_map_pages(virt_to_page(objp),
1872 cachep->buffer_size / PAGE_SIZE, 1);
1873 else
1874 check_poison_obj(cachep, objp);
1875 #else
1876 check_poison_obj(cachep, objp);
1877 #endif
1879 if (cachep->flags & SLAB_RED_ZONE) {
1880 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1881 slab_error(cachep, "start of a freed object "
1882 "was overwritten");
1883 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1884 slab_error(cachep, "end of a freed object "
1885 "was overwritten");
1889 #else
1890 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1893 #endif
1896 * slab_destroy - destroy and release all objects in a slab
1897 * @cachep: cache pointer being destroyed
1898 * @slabp: slab pointer being destroyed
1900 * Destroy all the objs in a slab, and release the mem back to the system.
1901 * Before calling the slab must have been unlinked from the cache. The
1902 * cache-lock is not held/needed.
1904 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1906 void *addr = slabp->s_mem - slabp->colouroff;
1908 slab_destroy_debugcheck(cachep, slabp);
1909 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1910 struct slab_rcu *slab_rcu;
1912 slab_rcu = (struct slab_rcu *)slabp;
1913 slab_rcu->cachep = cachep;
1914 slab_rcu->addr = addr;
1915 call_rcu(&slab_rcu->head, kmem_rcu_free);
1916 } else {
1917 kmem_freepages(cachep, addr);
1918 if (OFF_SLAB(cachep))
1919 kmem_cache_free(cachep->slabp_cache, slabp);
1923 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1925 int i;
1926 struct kmem_list3 *l3;
1928 for_each_online_cpu(i)
1929 kfree(cachep->array[i]);
1931 /* NUMA: free the list3 structures */
1932 for_each_online_node(i) {
1933 l3 = cachep->nodelists[i];
1934 if (l3) {
1935 kfree(l3->shared);
1936 free_alien_cache(l3->alien);
1937 kfree(l3);
1940 kmem_cache_free(&cache_cache, cachep);
1945 * calculate_slab_order - calculate size (page order) of slabs
1946 * @cachep: pointer to the cache that is being created
1947 * @size: size of objects to be created in this cache.
1948 * @align: required alignment for the objects.
1949 * @flags: slab allocation flags
1951 * Also calculates the number of objects per slab.
1953 * This could be made much more intelligent. For now, try to avoid using
1954 * high order pages for slabs. When the gfp() functions are more friendly
1955 * towards high-order requests, this should be changed.
1957 static size_t calculate_slab_order(struct kmem_cache *cachep,
1958 size_t size, size_t align, unsigned long flags)
1960 unsigned long offslab_limit;
1961 size_t left_over = 0;
1962 int gfporder;
1964 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1965 unsigned int num;
1966 size_t remainder;
1968 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1969 if (!num)
1970 continue;
1972 if (flags & CFLGS_OFF_SLAB) {
1974 * Max number of objs-per-slab for caches which
1975 * use off-slab slabs. Needed to avoid a possible
1976 * looping condition in cache_grow().
1978 offslab_limit = size - sizeof(struct slab);
1979 offslab_limit /= sizeof(kmem_bufctl_t);
1981 if (num > offslab_limit)
1982 break;
1985 /* Found something acceptable - save it away */
1986 cachep->num = num;
1987 cachep->gfporder = gfporder;
1988 left_over = remainder;
1991 * A VFS-reclaimable slab tends to have most allocations
1992 * as GFP_NOFS and we really don't want to have to be allocating
1993 * higher-order pages when we are unable to shrink dcache.
1995 if (flags & SLAB_RECLAIM_ACCOUNT)
1996 break;
1999 * Large number of objects is good, but very large slabs are
2000 * currently bad for the gfp()s.
2002 if (gfporder >= slab_break_gfp_order)
2003 break;
2006 * Acceptable internal fragmentation?
2008 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2009 break;
2011 return left_over;
2014 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2016 if (g_cpucache_up == FULL)
2017 return enable_cpucache(cachep, gfp);
2019 if (g_cpucache_up == NONE) {
2021 * Note: the first kmem_cache_create must create the cache
2022 * that's used by kmalloc(24), otherwise the creation of
2023 * further caches will BUG().
2025 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2028 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2029 * the first cache, then we need to set up all its list3s,
2030 * otherwise the creation of further caches will BUG().
2032 set_up_list3s(cachep, SIZE_AC);
2033 if (INDEX_AC == INDEX_L3)
2034 g_cpucache_up = PARTIAL_L3;
2035 else
2036 g_cpucache_up = PARTIAL_AC;
2037 } else {
2038 cachep->array[smp_processor_id()] =
2039 kmalloc(sizeof(struct arraycache_init), gfp);
2041 if (g_cpucache_up == PARTIAL_AC) {
2042 set_up_list3s(cachep, SIZE_L3);
2043 g_cpucache_up = PARTIAL_L3;
2044 } else {
2045 int node;
2046 for_each_online_node(node) {
2047 cachep->nodelists[node] =
2048 kmalloc_node(sizeof(struct kmem_list3),
2049 gfp, node);
2050 BUG_ON(!cachep->nodelists[node]);
2051 kmem_list3_init(cachep->nodelists[node]);
2055 cachep->nodelists[numa_node_id()]->next_reap =
2056 jiffies + REAPTIMEOUT_LIST3 +
2057 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2059 cpu_cache_get(cachep)->avail = 0;
2060 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2061 cpu_cache_get(cachep)->batchcount = 1;
2062 cpu_cache_get(cachep)->touched = 0;
2063 cachep->batchcount = 1;
2064 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2065 return 0;
2069 * kmem_cache_create - Create a cache.
2070 * @name: A string which is used in /proc/slabinfo to identify this cache.
2071 * @size: The size of objects to be created in this cache.
2072 * @align: The required alignment for the objects.
2073 * @flags: SLAB flags
2074 * @ctor: A constructor for the objects.
2076 * Returns a ptr to the cache on success, NULL on failure.
2077 * Cannot be called within a int, but can be interrupted.
2078 * The @ctor is run when new pages are allocated by the cache.
2080 * @name must be valid until the cache is destroyed. This implies that
2081 * the module calling this has to destroy the cache before getting unloaded.
2082 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2083 * therefore applications must manage it themselves.
2085 * The flags are
2087 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2088 * to catch references to uninitialised memory.
2090 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2091 * for buffer overruns.
2093 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2094 * cacheline. This can be beneficial if you're counting cycles as closely
2095 * as davem.
2097 struct kmem_cache *
2098 kmem_cache_create (const char *name, size_t size, size_t align,
2099 unsigned long flags, void (*ctor)(void *))
2101 size_t left_over, slab_size, ralign;
2102 struct kmem_cache *cachep = NULL, *pc;
2103 gfp_t gfp;
2106 * Sanity checks... these are all serious usage bugs.
2108 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2109 size > KMALLOC_MAX_SIZE) {
2110 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2111 name);
2112 BUG();
2116 * We use cache_chain_mutex to ensure a consistent view of
2117 * cpu_online_mask as well. Please see cpuup_callback
2119 if (slab_is_available()) {
2120 get_online_cpus();
2121 mutex_lock(&cache_chain_mutex);
2124 list_for_each_entry(pc, &cache_chain, next) {
2125 char tmp;
2126 int res;
2129 * This happens when the module gets unloaded and doesn't
2130 * destroy its slab cache and no-one else reuses the vmalloc
2131 * area of the module. Print a warning.
2133 res = probe_kernel_address(pc->name, tmp);
2134 if (res) {
2135 printk(KERN_ERR
2136 "SLAB: cache with size %d has lost its name\n",
2137 pc->buffer_size);
2138 continue;
2141 if (!strcmp(pc->name, name)) {
2142 printk(KERN_ERR
2143 "kmem_cache_create: duplicate cache %s\n", name);
2144 dump_stack();
2145 goto oops;
2149 #if DEBUG
2150 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2151 #if FORCED_DEBUG
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;
2163 #endif
2164 if (flags & SLAB_DESTROY_BY_RCU)
2165 BUG_ON(flags & SLAB_POISON);
2166 #endif
2168 * Always checks flags, a caller might be expecting debug support which
2169 * isn't available.
2171 BUG_ON(flags & ~CREATE_MASK);
2174 * Check that size is in terms of words. This is needed to avoid
2175 * unaligned accesses for some archs when redzoning is used, and makes
2176 * sure any on-slab bufctl's are also correctly aligned.
2178 if (size & (BYTES_PER_WORD - 1)) {
2179 size += (BYTES_PER_WORD - 1);
2180 size &= ~(BYTES_PER_WORD - 1);
2183 /* calculate the final buffer alignment: */
2185 /* 1) arch recommendation: can be overridden for debug */
2186 if (flags & SLAB_HWCACHE_ALIGN) {
2188 * Default alignment: as specified by the arch code. Except if
2189 * an object is really small, then squeeze multiple objects into
2190 * one cacheline.
2192 ralign = cache_line_size();
2193 while (size <= ralign / 2)
2194 ralign /= 2;
2195 } else {
2196 ralign = BYTES_PER_WORD;
2200 * Redzoning and user store require word alignment or possibly larger.
2201 * Note this will be overridden by architecture or caller mandated
2202 * alignment if either is greater than BYTES_PER_WORD.
2204 if (flags & SLAB_STORE_USER)
2205 ralign = BYTES_PER_WORD;
2207 if (flags & SLAB_RED_ZONE) {
2208 ralign = REDZONE_ALIGN;
2209 /* If redzoning, ensure that the second redzone is suitably
2210 * aligned, by adjusting the object size accordingly. */
2211 size += REDZONE_ALIGN - 1;
2212 size &= ~(REDZONE_ALIGN - 1);
2215 /* 2) arch mandated alignment */
2216 if (ralign < ARCH_SLAB_MINALIGN) {
2217 ralign = ARCH_SLAB_MINALIGN;
2219 /* 3) caller mandated alignment */
2220 if (ralign < align) {
2221 ralign = align;
2223 /* disable debug if necessary */
2224 if (ralign > __alignof__(unsigned long long))
2225 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2227 * 4) Store it.
2229 align = ralign;
2231 if (slab_is_available())
2232 gfp = GFP_KERNEL;
2233 else
2234 gfp = GFP_NOWAIT;
2236 /* Get cache's description obj. */
2237 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2238 if (!cachep)
2239 goto oops;
2241 #if DEBUG
2242 cachep->obj_size = size;
2245 * Both debugging options require word-alignment which is calculated
2246 * into align above.
2248 if (flags & SLAB_RED_ZONE) {
2249 /* add space for red zone words */
2250 cachep->obj_offset += sizeof(unsigned long long);
2251 size += 2 * sizeof(unsigned long long);
2253 if (flags & SLAB_STORE_USER) {
2254 /* user store requires one word storage behind the end of
2255 * the real object. But if the second red zone needs to be
2256 * aligned to 64 bits, we must allow that much space.
2258 if (flags & SLAB_RED_ZONE)
2259 size += REDZONE_ALIGN;
2260 else
2261 size += BYTES_PER_WORD;
2263 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2264 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2265 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2266 cachep->obj_offset += PAGE_SIZE - size;
2267 size = PAGE_SIZE;
2269 #endif
2270 #endif
2273 * Determine if the slab management is 'on' or 'off' slab.
2274 * (bootstrapping cannot cope with offslab caches so don't do
2275 * it too early on. Always use on-slab management when
2276 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2278 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2279 !(flags & SLAB_NOLEAKTRACE))
2281 * Size is large, assume best to place the slab management obj
2282 * off-slab (should allow better packing of objs).
2284 flags |= CFLGS_OFF_SLAB;
2286 size = ALIGN(size, align);
2288 left_over = calculate_slab_order(cachep, size, align, flags);
2290 if (!cachep->num) {
2291 printk(KERN_ERR
2292 "kmem_cache_create: couldn't create cache %s.\n", name);
2293 kmem_cache_free(&cache_cache, cachep);
2294 cachep = NULL;
2295 goto oops;
2297 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2298 + sizeof(struct slab), align);
2301 * If the slab has been placed off-slab, and we have enough space then
2302 * move it on-slab. This is at the expense of any extra colouring.
2304 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2305 flags &= ~CFLGS_OFF_SLAB;
2306 left_over -= slab_size;
2309 if (flags & CFLGS_OFF_SLAB) {
2310 /* really off slab. No need for manual alignment */
2311 slab_size =
2312 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2314 #ifdef CONFIG_PAGE_POISONING
2315 /* If we're going to use the generic kernel_map_pages()
2316 * poisoning, then it's going to smash the contents of
2317 * the redzone and userword anyhow, so switch them off.
2319 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2320 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2321 #endif
2324 cachep->colour_off = cache_line_size();
2325 /* Offset must be a multiple of the alignment. */
2326 if (cachep->colour_off < align)
2327 cachep->colour_off = align;
2328 cachep->colour = left_over / cachep->colour_off;
2329 cachep->slab_size = slab_size;
2330 cachep->flags = flags;
2331 cachep->gfpflags = 0;
2332 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2333 cachep->gfpflags |= GFP_DMA;
2334 cachep->buffer_size = size;
2335 cachep->reciprocal_buffer_size = reciprocal_value(size);
2337 if (flags & CFLGS_OFF_SLAB) {
2338 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2340 * This is a possibility for one of the malloc_sizes caches.
2341 * But since we go off slab only for object size greater than
2342 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2343 * this should not happen at all.
2344 * But leave a BUG_ON for some lucky dude.
2346 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2348 cachep->ctor = ctor;
2349 cachep->name = name;
2351 if (setup_cpu_cache(cachep, gfp)) {
2352 __kmem_cache_destroy(cachep);
2353 cachep = NULL;
2354 goto oops;
2357 /* cache setup completed, link it into the list */
2358 list_add(&cachep->next, &cache_chain);
2359 oops:
2360 if (!cachep && (flags & SLAB_PANIC))
2361 panic("kmem_cache_create(): failed to create slab `%s'\n",
2362 name);
2363 if (slab_is_available()) {
2364 mutex_unlock(&cache_chain_mutex);
2365 put_online_cpus();
2367 return cachep;
2369 EXPORT_SYMBOL(kmem_cache_create);
2371 #if DEBUG
2372 static void check_irq_off(void)
2374 BUG_ON(!irqs_disabled());
2377 static void check_irq_on(void)
2379 BUG_ON(irqs_disabled());
2382 static void check_spinlock_acquired(struct kmem_cache *cachep)
2384 #ifdef CONFIG_SMP
2385 check_irq_off();
2386 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2387 #endif
2390 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2392 #ifdef CONFIG_SMP
2393 check_irq_off();
2394 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2395 #endif
2398 #else
2399 #define check_irq_off() do { } while(0)
2400 #define check_irq_on() do { } while(0)
2401 #define check_spinlock_acquired(x) do { } while(0)
2402 #define check_spinlock_acquired_node(x, y) do { } while(0)
2403 #endif
2405 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2406 struct array_cache *ac,
2407 int force, int node);
2409 static void do_drain(void *arg)
2411 struct kmem_cache *cachep = arg;
2412 struct array_cache *ac;
2413 int node = numa_node_id();
2415 check_irq_off();
2416 ac = cpu_cache_get(cachep);
2417 spin_lock(&cachep->nodelists[node]->list_lock);
2418 free_block(cachep, ac->entry, ac->avail, node);
2419 spin_unlock(&cachep->nodelists[node]->list_lock);
2420 ac->avail = 0;
2423 static void drain_cpu_caches(struct kmem_cache *cachep)
2425 struct kmem_list3 *l3;
2426 int node;
2428 on_each_cpu(do_drain, cachep, 1);
2429 check_irq_on();
2430 for_each_online_node(node) {
2431 l3 = cachep->nodelists[node];
2432 if (l3 && l3->alien)
2433 drain_alien_cache(cachep, l3->alien);
2436 for_each_online_node(node) {
2437 l3 = cachep->nodelists[node];
2438 if (l3)
2439 drain_array(cachep, l3, l3->shared, 1, node);
2444 * Remove slabs from the list of free slabs.
2445 * Specify the number of slabs to drain in tofree.
2447 * Returns the actual number of slabs released.
2449 static int drain_freelist(struct kmem_cache *cache,
2450 struct kmem_list3 *l3, int tofree)
2452 struct list_head *p;
2453 int nr_freed;
2454 struct slab *slabp;
2456 nr_freed = 0;
2457 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2459 spin_lock_irq(&l3->list_lock);
2460 p = l3->slabs_free.prev;
2461 if (p == &l3->slabs_free) {
2462 spin_unlock_irq(&l3->list_lock);
2463 goto out;
2466 slabp = list_entry(p, struct slab, list);
2467 #if DEBUG
2468 BUG_ON(slabp->inuse);
2469 #endif
2470 list_del(&slabp->list);
2472 * Safe to drop the lock. The slab is no longer linked
2473 * to the cache.
2475 l3->free_objects -= cache->num;
2476 spin_unlock_irq(&l3->list_lock);
2477 slab_destroy(cache, slabp);
2478 nr_freed++;
2480 out:
2481 return nr_freed;
2484 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2485 static int __cache_shrink(struct kmem_cache *cachep)
2487 int ret = 0, i = 0;
2488 struct kmem_list3 *l3;
2490 drain_cpu_caches(cachep);
2492 check_irq_on();
2493 for_each_online_node(i) {
2494 l3 = cachep->nodelists[i];
2495 if (!l3)
2496 continue;
2498 drain_freelist(cachep, l3, l3->free_objects);
2500 ret += !list_empty(&l3->slabs_full) ||
2501 !list_empty(&l3->slabs_partial);
2503 return (ret ? 1 : 0);
2507 * kmem_cache_shrink - Shrink a cache.
2508 * @cachep: The cache to shrink.
2510 * Releases as many slabs as possible for a cache.
2511 * To help debugging, a zero exit status indicates all slabs were released.
2513 int kmem_cache_shrink(struct kmem_cache *cachep)
2515 int ret;
2516 BUG_ON(!cachep || in_interrupt());
2518 get_online_cpus();
2519 mutex_lock(&cache_chain_mutex);
2520 ret = __cache_shrink(cachep);
2521 mutex_unlock(&cache_chain_mutex);
2522 put_online_cpus();
2523 return ret;
2525 EXPORT_SYMBOL(kmem_cache_shrink);
2528 * kmem_cache_destroy - delete a cache
2529 * @cachep: the cache to destroy
2531 * Remove a &struct kmem_cache object from the slab cache.
2533 * It is expected this function will be called by a module when it is
2534 * unloaded. This will remove the cache completely, and avoid a duplicate
2535 * cache being allocated each time a module is loaded and unloaded, if the
2536 * module doesn't have persistent in-kernel storage across loads and unloads.
2538 * The cache must be empty before calling this function.
2540 * The caller must guarantee that noone will allocate memory from the cache
2541 * during the kmem_cache_destroy().
2543 void kmem_cache_destroy(struct kmem_cache *cachep)
2545 BUG_ON(!cachep || in_interrupt());
2547 /* Find the cache in the chain of caches. */
2548 get_online_cpus();
2549 mutex_lock(&cache_chain_mutex);
2551 * the chain is never empty, cache_cache is never destroyed
2553 list_del(&cachep->next);
2554 if (__cache_shrink(cachep)) {
2555 slab_error(cachep, "Can't free all objects");
2556 list_add(&cachep->next, &cache_chain);
2557 mutex_unlock(&cache_chain_mutex);
2558 put_online_cpus();
2559 return;
2562 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2563 rcu_barrier();
2565 __kmem_cache_destroy(cachep);
2566 mutex_unlock(&cache_chain_mutex);
2567 put_online_cpus();
2569 EXPORT_SYMBOL(kmem_cache_destroy);
2572 * Get the memory for a slab management obj.
2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2574 * always come from malloc_sizes caches. The slab descriptor cannot
2575 * come from the same cache which is getting created because,
2576 * when we are searching for an appropriate cache for these
2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2578 * If we are creating a malloc_sizes cache here it would not be visible to
2579 * kmem_find_general_cachep till the initialization is complete.
2580 * Hence we cannot have slabp_cache same as the original cache.
2582 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2583 int colour_off, gfp_t local_flags,
2584 int nodeid)
2586 struct slab *slabp;
2588 if (OFF_SLAB(cachep)) {
2589 /* Slab management obj is off-slab. */
2590 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2591 local_flags, nodeid);
2593 * If the first object in the slab is leaked (it's allocated
2594 * but no one has a reference to it), we want to make sure
2595 * kmemleak does not treat the ->s_mem pointer as a reference
2596 * to the object. Otherwise we will not report the leak.
2598 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2599 local_flags);
2600 if (!slabp)
2601 return NULL;
2602 } else {
2603 slabp = objp + colour_off;
2604 colour_off += cachep->slab_size;
2606 slabp->inuse = 0;
2607 slabp->colouroff = colour_off;
2608 slabp->s_mem = objp + colour_off;
2609 slabp->nodeid = nodeid;
2610 slabp->free = 0;
2611 return slabp;
2614 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2616 return (kmem_bufctl_t *) (slabp + 1);
2619 static void cache_init_objs(struct kmem_cache *cachep,
2620 struct slab *slabp)
2622 int i;
2624 for (i = 0; i < cachep->num; i++) {
2625 void *objp = index_to_obj(cachep, slabp, i);
2626 #if DEBUG
2627 /* need to poison the objs? */
2628 if (cachep->flags & SLAB_POISON)
2629 poison_obj(cachep, objp, POISON_FREE);
2630 if (cachep->flags & SLAB_STORE_USER)
2631 *dbg_userword(cachep, objp) = NULL;
2633 if (cachep->flags & SLAB_RED_ZONE) {
2634 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2635 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2638 * Constructors are not allowed to allocate memory from the same
2639 * cache which they are a constructor for. Otherwise, deadlock.
2640 * They must also be threaded.
2642 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2643 cachep->ctor(objp + obj_offset(cachep));
2645 if (cachep->flags & SLAB_RED_ZONE) {
2646 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2647 slab_error(cachep, "constructor overwrote the"
2648 " end of an object");
2649 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2650 slab_error(cachep, "constructor overwrote the"
2651 " start of an object");
2653 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2654 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2655 kernel_map_pages(virt_to_page(objp),
2656 cachep->buffer_size / PAGE_SIZE, 0);
2657 #else
2658 if (cachep->ctor)
2659 cachep->ctor(objp);
2660 #endif
2661 slab_bufctl(slabp)[i] = i + 1;
2663 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2666 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2668 if (CONFIG_ZONE_DMA_FLAG) {
2669 if (flags & GFP_DMA)
2670 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2671 else
2672 BUG_ON(cachep->gfpflags & GFP_DMA);
2676 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2677 int nodeid)
2679 void *objp = index_to_obj(cachep, slabp, slabp->free);
2680 kmem_bufctl_t next;
2682 slabp->inuse++;
2683 next = slab_bufctl(slabp)[slabp->free];
2684 #if DEBUG
2685 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2686 WARN_ON(slabp->nodeid != nodeid);
2687 #endif
2688 slabp->free = next;
2690 return objp;
2693 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2694 void *objp, int nodeid)
2696 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2698 #if DEBUG
2699 /* Verify that the slab belongs to the intended node */
2700 WARN_ON(slabp->nodeid != nodeid);
2702 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2703 printk(KERN_ERR "slab: double free detected in cache "
2704 "'%s', objp %p\n", cachep->name, objp);
2705 BUG();
2707 #endif
2708 slab_bufctl(slabp)[objnr] = slabp->free;
2709 slabp->free = objnr;
2710 slabp->inuse--;
2714 * Map pages beginning at addr to the given cache and slab. This is required
2715 * for the slab allocator to be able to lookup the cache and slab of a
2716 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2718 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2719 void *addr)
2721 int nr_pages;
2722 struct page *page;
2724 page = virt_to_page(addr);
2726 nr_pages = 1;
2727 if (likely(!PageCompound(page)))
2728 nr_pages <<= cache->gfporder;
2730 do {
2731 page_set_cache(page, cache);
2732 page_set_slab(page, slab);
2733 page++;
2734 } while (--nr_pages);
2738 * Grow (by 1) the number of slabs within a cache. This is called by
2739 * kmem_cache_alloc() when there are no active objs left in a cache.
2741 static int cache_grow(struct kmem_cache *cachep,
2742 gfp_t flags, int nodeid, void *objp)
2744 struct slab *slabp;
2745 size_t offset;
2746 gfp_t local_flags;
2747 struct kmem_list3 *l3;
2750 * Be lazy and only check for valid flags here, keeping it out of the
2751 * critical path in kmem_cache_alloc().
2753 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2754 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2756 /* Take the l3 list lock to change the colour_next on this node */
2757 check_irq_off();
2758 l3 = cachep->nodelists[nodeid];
2759 spin_lock(&l3->list_lock);
2761 /* Get colour for the slab, and cal the next value. */
2762 offset = l3->colour_next;
2763 l3->colour_next++;
2764 if (l3->colour_next >= cachep->colour)
2765 l3->colour_next = 0;
2766 spin_unlock(&l3->list_lock);
2768 offset *= cachep->colour_off;
2770 if (local_flags & __GFP_WAIT)
2771 local_irq_enable();
2774 * The test for missing atomic flag is performed here, rather than
2775 * the more obvious place, simply to reduce the critical path length
2776 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2777 * will eventually be caught here (where it matters).
2779 kmem_flagcheck(cachep, flags);
2782 * Get mem for the objs. Attempt to allocate a physical page from
2783 * 'nodeid'.
2785 if (!objp)
2786 objp = kmem_getpages(cachep, local_flags, nodeid);
2787 if (!objp)
2788 goto failed;
2790 /* Get slab management. */
2791 slabp = alloc_slabmgmt(cachep, objp, offset,
2792 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2793 if (!slabp)
2794 goto opps1;
2796 slab_map_pages(cachep, slabp, objp);
2798 cache_init_objs(cachep, slabp);
2800 if (local_flags & __GFP_WAIT)
2801 local_irq_disable();
2802 check_irq_off();
2803 spin_lock(&l3->list_lock);
2805 /* Make slab active. */
2806 list_add_tail(&slabp->list, &(l3->slabs_free));
2807 STATS_INC_GROWN(cachep);
2808 l3->free_objects += cachep->num;
2809 spin_unlock(&l3->list_lock);
2810 return 1;
2811 opps1:
2812 kmem_freepages(cachep, objp);
2813 failed:
2814 if (local_flags & __GFP_WAIT)
2815 local_irq_disable();
2816 return 0;
2819 #if DEBUG
2822 * Perform extra freeing checks:
2823 * - detect bad pointers.
2824 * - POISON/RED_ZONE checking
2826 static void kfree_debugcheck(const void *objp)
2828 if (!virt_addr_valid(objp)) {
2829 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2830 (unsigned long)objp);
2831 BUG();
2835 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2837 unsigned long long redzone1, redzone2;
2839 redzone1 = *dbg_redzone1(cache, obj);
2840 redzone2 = *dbg_redzone2(cache, obj);
2843 * Redzone is ok.
2845 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2846 return;
2848 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2849 slab_error(cache, "double free detected");
2850 else
2851 slab_error(cache, "memory outside object was overwritten");
2853 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2854 obj, redzone1, redzone2);
2857 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2858 void *caller)
2860 struct page *page;
2861 unsigned int objnr;
2862 struct slab *slabp;
2864 BUG_ON(virt_to_cache(objp) != cachep);
2866 objp -= obj_offset(cachep);
2867 kfree_debugcheck(objp);
2868 page = virt_to_head_page(objp);
2870 slabp = page_get_slab(page);
2872 if (cachep->flags & SLAB_RED_ZONE) {
2873 verify_redzone_free(cachep, objp);
2874 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2875 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2877 if (cachep->flags & SLAB_STORE_USER)
2878 *dbg_userword(cachep, objp) = caller;
2880 objnr = obj_to_index(cachep, slabp, objp);
2882 BUG_ON(objnr >= cachep->num);
2883 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2885 #ifdef CONFIG_DEBUG_SLAB_LEAK
2886 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2887 #endif
2888 if (cachep->flags & SLAB_POISON) {
2889 #ifdef CONFIG_DEBUG_PAGEALLOC
2890 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2891 store_stackinfo(cachep, objp, (unsigned long)caller);
2892 kernel_map_pages(virt_to_page(objp),
2893 cachep->buffer_size / PAGE_SIZE, 0);
2894 } else {
2895 poison_obj(cachep, objp, POISON_FREE);
2897 #else
2898 poison_obj(cachep, objp, POISON_FREE);
2899 #endif
2901 return objp;
2904 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2906 kmem_bufctl_t i;
2907 int entries = 0;
2909 /* Check slab's freelist to see if this obj is there. */
2910 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2911 entries++;
2912 if (entries > cachep->num || i >= cachep->num)
2913 goto bad;
2915 if (entries != cachep->num - slabp->inuse) {
2916 bad:
2917 printk(KERN_ERR "slab: Internal list corruption detected in "
2918 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2919 cachep->name, cachep->num, slabp, slabp->inuse);
2920 for (i = 0;
2921 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2922 i++) {
2923 if (i % 16 == 0)
2924 printk("\n%03x:", i);
2925 printk(" %02x", ((unsigned char *)slabp)[i]);
2927 printk("\n");
2928 BUG();
2931 #else
2932 #define kfree_debugcheck(x) do { } while(0)
2933 #define cache_free_debugcheck(x,objp,z) (objp)
2934 #define check_slabp(x,y) do { } while(0)
2935 #endif
2937 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2939 int batchcount;
2940 struct kmem_list3 *l3;
2941 struct array_cache *ac;
2942 int node;
2944 retry:
2945 check_irq_off();
2946 node = numa_node_id();
2947 ac = cpu_cache_get(cachep);
2948 batchcount = ac->batchcount;
2949 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2951 * If there was little recent activity on this cache, then
2952 * perform only a partial refill. Otherwise we could generate
2953 * refill bouncing.
2955 batchcount = BATCHREFILL_LIMIT;
2957 l3 = cachep->nodelists[node];
2959 BUG_ON(ac->avail > 0 || !l3);
2960 spin_lock(&l3->list_lock);
2962 /* See if we can refill from the shared array */
2963 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
2964 l3->shared->touched = 1;
2965 goto alloc_done;
2968 while (batchcount > 0) {
2969 struct list_head *entry;
2970 struct slab *slabp;
2971 /* Get slab alloc is to come from. */
2972 entry = l3->slabs_partial.next;
2973 if (entry == &l3->slabs_partial) {
2974 l3->free_touched = 1;
2975 entry = l3->slabs_free.next;
2976 if (entry == &l3->slabs_free)
2977 goto must_grow;
2980 slabp = list_entry(entry, struct slab, list);
2981 check_slabp(cachep, slabp);
2982 check_spinlock_acquired(cachep);
2985 * The slab was either on partial or free list so
2986 * there must be at least one object available for
2987 * allocation.
2989 BUG_ON(slabp->inuse >= cachep->num);
2991 while (slabp->inuse < cachep->num && batchcount--) {
2992 STATS_INC_ALLOCED(cachep);
2993 STATS_INC_ACTIVE(cachep);
2994 STATS_SET_HIGH(cachep);
2996 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2997 node);
2999 check_slabp(cachep, slabp);
3001 /* move slabp to correct slabp list: */
3002 list_del(&slabp->list);
3003 if (slabp->free == BUFCTL_END)
3004 list_add(&slabp->list, &l3->slabs_full);
3005 else
3006 list_add(&slabp->list, &l3->slabs_partial);
3009 must_grow:
3010 l3->free_objects -= ac->avail;
3011 alloc_done:
3012 spin_unlock(&l3->list_lock);
3014 if (unlikely(!ac->avail)) {
3015 int x;
3016 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3018 /* cache_grow can reenable interrupts, then ac could change. */
3019 ac = cpu_cache_get(cachep);
3020 if (!x && ac->avail == 0) /* no objects in sight? abort */
3021 return NULL;
3023 if (!ac->avail) /* objects refilled by interrupt? */
3024 goto retry;
3026 ac->touched = 1;
3027 return ac->entry[--ac->avail];
3030 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3031 gfp_t flags)
3033 might_sleep_if(flags & __GFP_WAIT);
3034 #if DEBUG
3035 kmem_flagcheck(cachep, flags);
3036 #endif
3039 #if DEBUG
3040 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3041 gfp_t flags, void *objp, void *caller)
3043 if (!objp)
3044 return objp;
3045 if (cachep->flags & SLAB_POISON) {
3046 #ifdef CONFIG_DEBUG_PAGEALLOC
3047 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3048 kernel_map_pages(virt_to_page(objp),
3049 cachep->buffer_size / PAGE_SIZE, 1);
3050 else
3051 check_poison_obj(cachep, objp);
3052 #else
3053 check_poison_obj(cachep, objp);
3054 #endif
3055 poison_obj(cachep, objp, POISON_INUSE);
3057 if (cachep->flags & SLAB_STORE_USER)
3058 *dbg_userword(cachep, objp) = caller;
3060 if (cachep->flags & SLAB_RED_ZONE) {
3061 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3062 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3063 slab_error(cachep, "double free, or memory outside"
3064 " object was overwritten");
3065 printk(KERN_ERR
3066 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3067 objp, *dbg_redzone1(cachep, objp),
3068 *dbg_redzone2(cachep, objp));
3070 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3071 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3073 #ifdef CONFIG_DEBUG_SLAB_LEAK
3075 struct slab *slabp;
3076 unsigned objnr;
3078 slabp = page_get_slab(virt_to_head_page(objp));
3079 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3080 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3082 #endif
3083 objp += obj_offset(cachep);
3084 if (cachep->ctor && cachep->flags & SLAB_POISON)
3085 cachep->ctor(objp);
3086 #if ARCH_SLAB_MINALIGN
3087 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3088 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3089 objp, ARCH_SLAB_MINALIGN);
3091 #endif
3092 return objp;
3094 #else
3095 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3096 #endif
3098 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3100 if (cachep == &cache_cache)
3101 return false;
3103 return should_failslab(obj_size(cachep), flags, cachep->flags);
3106 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3108 void *objp;
3109 struct array_cache *ac;
3111 check_irq_off();
3113 ac = cpu_cache_get(cachep);
3114 if (likely(ac->avail)) {
3115 STATS_INC_ALLOCHIT(cachep);
3116 ac->touched = 1;
3117 objp = ac->entry[--ac->avail];
3118 } else {
3119 STATS_INC_ALLOCMISS(cachep);
3120 objp = cache_alloc_refill(cachep, flags);
3122 * the 'ac' may be updated by cache_alloc_refill(),
3123 * and kmemleak_erase() requires its correct value.
3125 ac = cpu_cache_get(cachep);
3128 * To avoid a false negative, if an object that is in one of the
3129 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3130 * treat the array pointers as a reference to the object.
3132 if (objp)
3133 kmemleak_erase(&ac->entry[ac->avail]);
3134 return objp;
3137 #ifdef CONFIG_NUMA
3139 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3141 * If we are in_interrupt, then process context, including cpusets and
3142 * mempolicy, may not apply and should not be used for allocation policy.
3144 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3146 int nid_alloc, nid_here;
3148 if (in_interrupt() || (flags & __GFP_THISNODE))
3149 return NULL;
3150 nid_alloc = nid_here = numa_node_id();
3151 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3152 nid_alloc = cpuset_mem_spread_node();
3153 else if (current->mempolicy)
3154 nid_alloc = slab_node(current->mempolicy);
3155 if (nid_alloc != nid_here)
3156 return ____cache_alloc_node(cachep, flags, nid_alloc);
3157 return NULL;
3161 * Fallback function if there was no memory available and no objects on a
3162 * certain node and fall back is permitted. First we scan all the
3163 * available nodelists for available objects. If that fails then we
3164 * perform an allocation without specifying a node. This allows the page
3165 * allocator to do its reclaim / fallback magic. We then insert the
3166 * slab into the proper nodelist and then allocate from it.
3168 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3170 struct zonelist *zonelist;
3171 gfp_t local_flags;
3172 struct zoneref *z;
3173 struct zone *zone;
3174 enum zone_type high_zoneidx = gfp_zone(flags);
3175 void *obj = NULL;
3176 int nid;
3178 if (flags & __GFP_THISNODE)
3179 return NULL;
3181 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3182 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3184 retry:
3186 * Look through allowed nodes for objects available
3187 * from existing per node queues.
3189 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3190 nid = zone_to_nid(zone);
3192 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3193 cache->nodelists[nid] &&
3194 cache->nodelists[nid]->free_objects) {
3195 obj = ____cache_alloc_node(cache,
3196 flags | GFP_THISNODE, nid);
3197 if (obj)
3198 break;
3202 if (!obj) {
3204 * This allocation will be performed within the constraints
3205 * of the current cpuset / memory policy requirements.
3206 * We may trigger various forms of reclaim on the allowed
3207 * set and go into memory reserves if necessary.
3209 if (local_flags & __GFP_WAIT)
3210 local_irq_enable();
3211 kmem_flagcheck(cache, flags);
3212 obj = kmem_getpages(cache, local_flags, numa_node_id());
3213 if (local_flags & __GFP_WAIT)
3214 local_irq_disable();
3215 if (obj) {
3217 * Insert into the appropriate per node queues
3219 nid = page_to_nid(virt_to_page(obj));
3220 if (cache_grow(cache, flags, nid, obj)) {
3221 obj = ____cache_alloc_node(cache,
3222 flags | GFP_THISNODE, nid);
3223 if (!obj)
3225 * Another processor may allocate the
3226 * objects in the slab since we are
3227 * not holding any locks.
3229 goto retry;
3230 } else {
3231 /* cache_grow already freed obj */
3232 obj = NULL;
3236 return obj;
3240 * A interface to enable slab creation on nodeid
3242 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3243 int nodeid)
3245 struct list_head *entry;
3246 struct slab *slabp;
3247 struct kmem_list3 *l3;
3248 void *obj;
3249 int x;
3251 l3 = cachep->nodelists[nodeid];
3252 BUG_ON(!l3);
3254 retry:
3255 check_irq_off();
3256 spin_lock(&l3->list_lock);
3257 entry = l3->slabs_partial.next;
3258 if (entry == &l3->slabs_partial) {
3259 l3->free_touched = 1;
3260 entry = l3->slabs_free.next;
3261 if (entry == &l3->slabs_free)
3262 goto must_grow;
3265 slabp = list_entry(entry, struct slab, list);
3266 check_spinlock_acquired_node(cachep, nodeid);
3267 check_slabp(cachep, slabp);
3269 STATS_INC_NODEALLOCS(cachep);
3270 STATS_INC_ACTIVE(cachep);
3271 STATS_SET_HIGH(cachep);
3273 BUG_ON(slabp->inuse == cachep->num);
3275 obj = slab_get_obj(cachep, slabp, nodeid);
3276 check_slabp(cachep, slabp);
3277 l3->free_objects--;
3278 /* move slabp to correct slabp list: */
3279 list_del(&slabp->list);
3281 if (slabp->free == BUFCTL_END)
3282 list_add(&slabp->list, &l3->slabs_full);
3283 else
3284 list_add(&slabp->list, &l3->slabs_partial);
3286 spin_unlock(&l3->list_lock);
3287 goto done;
3289 must_grow:
3290 spin_unlock(&l3->list_lock);
3291 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3292 if (x)
3293 goto retry;
3295 return fallback_alloc(cachep, flags);
3297 done:
3298 return obj;
3302 * kmem_cache_alloc_node - Allocate an object on the specified node
3303 * @cachep: The cache to allocate from.
3304 * @flags: See kmalloc().
3305 * @nodeid: node number of the target node.
3306 * @caller: return address of caller, used for debug information
3308 * Identical to kmem_cache_alloc but it will allocate memory on the given
3309 * node, which can improve the performance for cpu bound structures.
3311 * Fallback to other node is possible if __GFP_THISNODE is not set.
3313 static __always_inline void *
3314 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3315 void *caller)
3317 unsigned long save_flags;
3318 void *ptr;
3320 flags &= gfp_allowed_mask;
3322 lockdep_trace_alloc(flags);
3324 if (slab_should_failslab(cachep, flags))
3325 return NULL;
3327 cache_alloc_debugcheck_before(cachep, flags);
3328 local_irq_save(save_flags);
3330 if (nodeid == -1)
3331 nodeid = numa_node_id();
3333 if (unlikely(!cachep->nodelists[nodeid])) {
3334 /* Node not bootstrapped yet */
3335 ptr = fallback_alloc(cachep, flags);
3336 goto out;
3339 if (nodeid == numa_node_id()) {
3341 * Use the locally cached objects if possible.
3342 * However ____cache_alloc does not allow fallback
3343 * to other nodes. It may fail while we still have
3344 * objects on other nodes available.
3346 ptr = ____cache_alloc(cachep, flags);
3347 if (ptr)
3348 goto out;
3350 /* ___cache_alloc_node can fall back to other nodes */
3351 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3352 out:
3353 local_irq_restore(save_flags);
3354 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3355 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3356 flags);
3358 if (likely(ptr))
3359 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3361 if (unlikely((flags & __GFP_ZERO) && ptr))
3362 memset(ptr, 0, obj_size(cachep));
3364 return ptr;
3367 static __always_inline void *
3368 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3370 void *objp;
3372 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3373 objp = alternate_node_alloc(cache, flags);
3374 if (objp)
3375 goto out;
3377 objp = ____cache_alloc(cache, flags);
3380 * We may just have run out of memory on the local node.
3381 * ____cache_alloc_node() knows how to locate memory on other nodes
3383 if (!objp)
3384 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3386 out:
3387 return objp;
3389 #else
3391 static __always_inline void *
3392 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3394 return ____cache_alloc(cachep, flags);
3397 #endif /* CONFIG_NUMA */
3399 static __always_inline void *
3400 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3402 unsigned long save_flags;
3403 void *objp;
3405 flags &= gfp_allowed_mask;
3407 lockdep_trace_alloc(flags);
3409 if (slab_should_failslab(cachep, flags))
3410 return NULL;
3412 cache_alloc_debugcheck_before(cachep, flags);
3413 local_irq_save(save_flags);
3414 objp = __do_cache_alloc(cachep, flags);
3415 local_irq_restore(save_flags);
3416 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3417 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3418 flags);
3419 prefetchw(objp);
3421 if (likely(objp))
3422 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3424 if (unlikely((flags & __GFP_ZERO) && objp))
3425 memset(objp, 0, obj_size(cachep));
3427 return objp;
3431 * Caller needs to acquire correct kmem_list's list_lock
3433 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3434 int node)
3436 int i;
3437 struct kmem_list3 *l3;
3439 for (i = 0; i < nr_objects; i++) {
3440 void *objp = objpp[i];
3441 struct slab *slabp;
3443 slabp = virt_to_slab(objp);
3444 l3 = cachep->nodelists[node];
3445 list_del(&slabp->list);
3446 check_spinlock_acquired_node(cachep, node);
3447 check_slabp(cachep, slabp);
3448 slab_put_obj(cachep, slabp, objp, node);
3449 STATS_DEC_ACTIVE(cachep);
3450 l3->free_objects++;
3451 check_slabp(cachep, slabp);
3453 /* fixup slab chains */
3454 if (slabp->inuse == 0) {
3455 if (l3->free_objects > l3->free_limit) {
3456 l3->free_objects -= cachep->num;
3457 /* No need to drop any previously held
3458 * lock here, even if we have a off-slab slab
3459 * descriptor it is guaranteed to come from
3460 * a different cache, refer to comments before
3461 * alloc_slabmgmt.
3463 slab_destroy(cachep, slabp);
3464 } else {
3465 list_add(&slabp->list, &l3->slabs_free);
3467 } else {
3468 /* Unconditionally move a slab to the end of the
3469 * partial list on free - maximum time for the
3470 * other objects to be freed, too.
3472 list_add_tail(&slabp->list, &l3->slabs_partial);
3477 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3479 int batchcount;
3480 struct kmem_list3 *l3;
3481 int node = numa_node_id();
3483 batchcount = ac->batchcount;
3484 #if DEBUG
3485 BUG_ON(!batchcount || batchcount > ac->avail);
3486 #endif
3487 check_irq_off();
3488 l3 = cachep->nodelists[node];
3489 spin_lock(&l3->list_lock);
3490 if (l3->shared) {
3491 struct array_cache *shared_array = l3->shared;
3492 int max = shared_array->limit - shared_array->avail;
3493 if (max) {
3494 if (batchcount > max)
3495 batchcount = max;
3496 memcpy(&(shared_array->entry[shared_array->avail]),
3497 ac->entry, sizeof(void *) * batchcount);
3498 shared_array->avail += batchcount;
3499 goto free_done;
3503 free_block(cachep, ac->entry, batchcount, node);
3504 free_done:
3505 #if STATS
3507 int i = 0;
3508 struct list_head *p;
3510 p = l3->slabs_free.next;
3511 while (p != &(l3->slabs_free)) {
3512 struct slab *slabp;
3514 slabp = list_entry(p, struct slab, list);
3515 BUG_ON(slabp->inuse);
3517 i++;
3518 p = p->next;
3520 STATS_SET_FREEABLE(cachep, i);
3522 #endif
3523 spin_unlock(&l3->list_lock);
3524 ac->avail -= batchcount;
3525 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3529 * Release an obj back to its cache. If the obj has a constructed state, it must
3530 * be in this state _before_ it is released. Called with disabled ints.
3532 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3534 struct array_cache *ac = cpu_cache_get(cachep);
3536 check_irq_off();
3537 kmemleak_free_recursive(objp, cachep->flags);
3538 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3540 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3543 * Skip calling cache_free_alien() when the platform is not numa.
3544 * This will avoid cache misses that happen while accessing slabp (which
3545 * is per page memory reference) to get nodeid. Instead use a global
3546 * variable to skip the call, which is mostly likely to be present in
3547 * the cache.
3549 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3550 return;
3552 if (likely(ac->avail < ac->limit)) {
3553 STATS_INC_FREEHIT(cachep);
3554 ac->entry[ac->avail++] = objp;
3555 return;
3556 } else {
3557 STATS_INC_FREEMISS(cachep);
3558 cache_flusharray(cachep, ac);
3559 ac->entry[ac->avail++] = objp;
3564 * kmem_cache_alloc - Allocate an object
3565 * @cachep: The cache to allocate from.
3566 * @flags: See kmalloc().
3568 * Allocate an object from this cache. The flags are only relevant
3569 * if the cache has no available objects.
3571 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3573 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3575 trace_kmem_cache_alloc(_RET_IP_, ret,
3576 obj_size(cachep), cachep->buffer_size, flags);
3578 return ret;
3580 EXPORT_SYMBOL(kmem_cache_alloc);
3582 #ifdef CONFIG_TRACING
3583 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3585 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3587 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3588 #endif
3591 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3592 * @cachep: the cache we're checking against
3593 * @ptr: pointer to validate
3595 * This verifies that the untrusted pointer looks sane;
3596 * it is _not_ a guarantee that the pointer is actually
3597 * part of the slab cache in question, but it at least
3598 * validates that the pointer can be dereferenced and
3599 * looks half-way sane.
3601 * Currently only used for dentry validation.
3603 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3605 unsigned long size = cachep->buffer_size;
3606 struct page *page;
3608 if (unlikely(!kern_ptr_validate(ptr, size)))
3609 goto out;
3610 page = virt_to_page(ptr);
3611 if (unlikely(!PageSlab(page)))
3612 goto out;
3613 if (unlikely(page_get_cache(page) != cachep))
3614 goto out;
3615 return 1;
3616 out:
3617 return 0;
3620 #ifdef CONFIG_NUMA
3621 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3623 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3624 __builtin_return_address(0));
3626 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3627 obj_size(cachep), cachep->buffer_size,
3628 flags, nodeid);
3630 return ret;
3632 EXPORT_SYMBOL(kmem_cache_alloc_node);
3634 #ifdef CONFIG_TRACING
3635 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3636 gfp_t flags,
3637 int nodeid)
3639 return __cache_alloc_node(cachep, flags, nodeid,
3640 __builtin_return_address(0));
3642 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3643 #endif
3645 static __always_inline void *
3646 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3648 struct kmem_cache *cachep;
3649 void *ret;
3651 cachep = kmem_find_general_cachep(size, flags);
3652 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3653 return cachep;
3654 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3656 trace_kmalloc_node((unsigned long) caller, ret,
3657 size, cachep->buffer_size, flags, node);
3659 return ret;
3662 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3663 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3665 return __do_kmalloc_node(size, flags, node,
3666 __builtin_return_address(0));
3668 EXPORT_SYMBOL(__kmalloc_node);
3670 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3671 int node, unsigned long caller)
3673 return __do_kmalloc_node(size, flags, node, (void *)caller);
3675 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3676 #else
3677 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3679 return __do_kmalloc_node(size, flags, node, NULL);
3681 EXPORT_SYMBOL(__kmalloc_node);
3682 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3683 #endif /* CONFIG_NUMA */
3686 * __do_kmalloc - allocate memory
3687 * @size: how many bytes of memory are required.
3688 * @flags: the type of memory to allocate (see kmalloc).
3689 * @caller: function caller for debug tracking of the caller
3691 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3692 void *caller)
3694 struct kmem_cache *cachep;
3695 void *ret;
3697 /* If you want to save a few bytes .text space: replace
3698 * __ with kmem_.
3699 * Then kmalloc uses the uninlined functions instead of the inline
3700 * functions.
3702 cachep = __find_general_cachep(size, flags);
3703 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3704 return cachep;
3705 ret = __cache_alloc(cachep, flags, caller);
3707 trace_kmalloc((unsigned long) caller, ret,
3708 size, cachep->buffer_size, flags);
3710 return ret;
3714 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3715 void *__kmalloc(size_t size, gfp_t flags)
3717 return __do_kmalloc(size, flags, __builtin_return_address(0));
3719 EXPORT_SYMBOL(__kmalloc);
3721 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3723 return __do_kmalloc(size, flags, (void *)caller);
3725 EXPORT_SYMBOL(__kmalloc_track_caller);
3727 #else
3728 void *__kmalloc(size_t size, gfp_t flags)
3730 return __do_kmalloc(size, flags, NULL);
3732 EXPORT_SYMBOL(__kmalloc);
3733 #endif
3736 * kmem_cache_free - Deallocate an object
3737 * @cachep: The cache the allocation was from.
3738 * @objp: The previously allocated object.
3740 * Free an object which was previously allocated from this
3741 * cache.
3743 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3745 unsigned long flags;
3747 local_irq_save(flags);
3748 debug_check_no_locks_freed(objp, obj_size(cachep));
3749 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3750 debug_check_no_obj_freed(objp, obj_size(cachep));
3751 __cache_free(cachep, objp);
3752 local_irq_restore(flags);
3754 trace_kmem_cache_free(_RET_IP_, objp);
3756 EXPORT_SYMBOL(kmem_cache_free);
3759 * kfree - free previously allocated memory
3760 * @objp: pointer returned by kmalloc.
3762 * If @objp is NULL, no operation is performed.
3764 * Don't free memory not originally allocated by kmalloc()
3765 * or you will run into trouble.
3767 void kfree(const void *objp)
3769 struct kmem_cache *c;
3770 unsigned long flags;
3772 trace_kfree(_RET_IP_, objp);
3774 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3775 return;
3776 local_irq_save(flags);
3777 kfree_debugcheck(objp);
3778 c = virt_to_cache(objp);
3779 debug_check_no_locks_freed(objp, obj_size(c));
3780 debug_check_no_obj_freed(objp, obj_size(c));
3781 __cache_free(c, (void *)objp);
3782 local_irq_restore(flags);
3784 EXPORT_SYMBOL(kfree);
3786 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3788 return obj_size(cachep);
3790 EXPORT_SYMBOL(kmem_cache_size);
3792 const char *kmem_cache_name(struct kmem_cache *cachep)
3794 return cachep->name;
3796 EXPORT_SYMBOL_GPL(kmem_cache_name);
3799 * This initializes kmem_list3 or resizes various caches for all nodes.
3801 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3803 int node;
3804 struct kmem_list3 *l3;
3805 struct array_cache *new_shared;
3806 struct array_cache **new_alien = NULL;
3808 for_each_online_node(node) {
3810 if (use_alien_caches) {
3811 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3812 if (!new_alien)
3813 goto fail;
3816 new_shared = NULL;
3817 if (cachep->shared) {
3818 new_shared = alloc_arraycache(node,
3819 cachep->shared*cachep->batchcount,
3820 0xbaadf00d, gfp);
3821 if (!new_shared) {
3822 free_alien_cache(new_alien);
3823 goto fail;
3827 l3 = cachep->nodelists[node];
3828 if (l3) {
3829 struct array_cache *shared = l3->shared;
3831 spin_lock_irq(&l3->list_lock);
3833 if (shared)
3834 free_block(cachep, shared->entry,
3835 shared->avail, node);
3837 l3->shared = new_shared;
3838 if (!l3->alien) {
3839 l3->alien = new_alien;
3840 new_alien = NULL;
3842 l3->free_limit = (1 + nr_cpus_node(node)) *
3843 cachep->batchcount + cachep->num;
3844 spin_unlock_irq(&l3->list_lock);
3845 kfree(shared);
3846 free_alien_cache(new_alien);
3847 continue;
3849 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3850 if (!l3) {
3851 free_alien_cache(new_alien);
3852 kfree(new_shared);
3853 goto fail;
3856 kmem_list3_init(l3);
3857 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3858 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3859 l3->shared = new_shared;
3860 l3->alien = new_alien;
3861 l3->free_limit = (1 + nr_cpus_node(node)) *
3862 cachep->batchcount + cachep->num;
3863 cachep->nodelists[node] = l3;
3865 return 0;
3867 fail:
3868 if (!cachep->next.next) {
3869 /* Cache is not active yet. Roll back what we did */
3870 node--;
3871 while (node >= 0) {
3872 if (cachep->nodelists[node]) {
3873 l3 = cachep->nodelists[node];
3875 kfree(l3->shared);
3876 free_alien_cache(l3->alien);
3877 kfree(l3);
3878 cachep->nodelists[node] = NULL;
3880 node--;
3883 return -ENOMEM;
3886 struct ccupdate_struct {
3887 struct kmem_cache *cachep;
3888 struct array_cache *new[NR_CPUS];
3891 static void do_ccupdate_local(void *info)
3893 struct ccupdate_struct *new = info;
3894 struct array_cache *old;
3896 check_irq_off();
3897 old = cpu_cache_get(new->cachep);
3899 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3900 new->new[smp_processor_id()] = old;
3903 /* Always called with the cache_chain_mutex held */
3904 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3905 int batchcount, int shared, gfp_t gfp)
3907 struct ccupdate_struct *new;
3908 int i;
3910 new = kzalloc(sizeof(*new), gfp);
3911 if (!new)
3912 return -ENOMEM;
3914 for_each_online_cpu(i) {
3915 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3916 batchcount, gfp);
3917 if (!new->new[i]) {
3918 for (i--; i >= 0; i--)
3919 kfree(new->new[i]);
3920 kfree(new);
3921 return -ENOMEM;
3924 new->cachep = cachep;
3926 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3928 check_irq_on();
3929 cachep->batchcount = batchcount;
3930 cachep->limit = limit;
3931 cachep->shared = shared;
3933 for_each_online_cpu(i) {
3934 struct array_cache *ccold = new->new[i];
3935 if (!ccold)
3936 continue;
3937 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3938 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3939 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3940 kfree(ccold);
3942 kfree(new);
3943 return alloc_kmemlist(cachep, gfp);
3946 /* Called with cache_chain_mutex held always */
3947 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3949 int err;
3950 int limit, shared;
3953 * The head array serves three purposes:
3954 * - create a LIFO ordering, i.e. return objects that are cache-warm
3955 * - reduce the number of spinlock operations.
3956 * - reduce the number of linked list operations on the slab and
3957 * bufctl chains: array operations are cheaper.
3958 * The numbers are guessed, we should auto-tune as described by
3959 * Bonwick.
3961 if (cachep->buffer_size > 131072)
3962 limit = 1;
3963 else if (cachep->buffer_size > PAGE_SIZE)
3964 limit = 8;
3965 else if (cachep->buffer_size > 1024)
3966 limit = 24;
3967 else if (cachep->buffer_size > 256)
3968 limit = 54;
3969 else
3970 limit = 120;
3973 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3974 * allocation behaviour: Most allocs on one cpu, most free operations
3975 * on another cpu. For these cases, an efficient object passing between
3976 * cpus is necessary. This is provided by a shared array. The array
3977 * replaces Bonwick's magazine layer.
3978 * On uniprocessor, it's functionally equivalent (but less efficient)
3979 * to a larger limit. Thus disabled by default.
3981 shared = 0;
3982 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3983 shared = 8;
3985 #if DEBUG
3987 * With debugging enabled, large batchcount lead to excessively long
3988 * periods with disabled local interrupts. Limit the batchcount
3990 if (limit > 32)
3991 limit = 32;
3992 #endif
3993 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
3994 if (err)
3995 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3996 cachep->name, -err);
3997 return err;
4001 * Drain an array if it contains any elements taking the l3 lock only if
4002 * necessary. Note that the l3 listlock also protects the array_cache
4003 * if drain_array() is used on the shared array.
4005 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4006 struct array_cache *ac, int force, int node)
4008 int tofree;
4010 if (!ac || !ac->avail)
4011 return;
4012 if (ac->touched && !force) {
4013 ac->touched = 0;
4014 } else {
4015 spin_lock_irq(&l3->list_lock);
4016 if (ac->avail) {
4017 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4018 if (tofree > ac->avail)
4019 tofree = (ac->avail + 1) / 2;
4020 free_block(cachep, ac->entry, tofree, node);
4021 ac->avail -= tofree;
4022 memmove(ac->entry, &(ac->entry[tofree]),
4023 sizeof(void *) * ac->avail);
4025 spin_unlock_irq(&l3->list_lock);
4030 * cache_reap - Reclaim memory from caches.
4031 * @w: work descriptor
4033 * Called from workqueue/eventd every few seconds.
4034 * Purpose:
4035 * - clear the per-cpu caches for this CPU.
4036 * - return freeable pages to the main free memory pool.
4038 * If we cannot acquire the cache chain mutex then just give up - we'll try
4039 * again on the next iteration.
4041 static void cache_reap(struct work_struct *w)
4043 struct kmem_cache *searchp;
4044 struct kmem_list3 *l3;
4045 int node = numa_node_id();
4046 struct delayed_work *work = to_delayed_work(w);
4048 if (!mutex_trylock(&cache_chain_mutex))
4049 /* Give up. Setup the next iteration. */
4050 goto out;
4052 list_for_each_entry(searchp, &cache_chain, next) {
4053 check_irq_on();
4056 * We only take the l3 lock if absolutely necessary and we
4057 * have established with reasonable certainty that
4058 * we can do some work if the lock was obtained.
4060 l3 = searchp->nodelists[node];
4062 reap_alien(searchp, l3);
4064 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4067 * These are racy checks but it does not matter
4068 * if we skip one check or scan twice.
4070 if (time_after(l3->next_reap, jiffies))
4071 goto next;
4073 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4075 drain_array(searchp, l3, l3->shared, 0, node);
4077 if (l3->free_touched)
4078 l3->free_touched = 0;
4079 else {
4080 int freed;
4082 freed = drain_freelist(searchp, l3, (l3->free_limit +
4083 5 * searchp->num - 1) / (5 * searchp->num));
4084 STATS_ADD_REAPED(searchp, freed);
4086 next:
4087 cond_resched();
4089 check_irq_on();
4090 mutex_unlock(&cache_chain_mutex);
4091 next_reap_node();
4092 out:
4093 /* Set up the next iteration */
4094 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4097 #ifdef CONFIG_SLABINFO
4099 static void print_slabinfo_header(struct seq_file *m)
4102 * Output format version, so at least we can change it
4103 * without _too_ many complaints.
4105 #if STATS
4106 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4107 #else
4108 seq_puts(m, "slabinfo - version: 2.1\n");
4109 #endif
4110 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4111 "<objperslab> <pagesperslab>");
4112 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4113 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4114 #if STATS
4115 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4116 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4117 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4118 #endif
4119 seq_putc(m, '\n');
4122 static void *s_start(struct seq_file *m, loff_t *pos)
4124 loff_t n = *pos;
4126 mutex_lock(&cache_chain_mutex);
4127 if (!n)
4128 print_slabinfo_header(m);
4130 return seq_list_start(&cache_chain, *pos);
4133 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4135 return seq_list_next(p, &cache_chain, pos);
4138 static void s_stop(struct seq_file *m, void *p)
4140 mutex_unlock(&cache_chain_mutex);
4143 static int s_show(struct seq_file *m, void *p)
4145 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4146 struct slab *slabp;
4147 unsigned long active_objs;
4148 unsigned long num_objs;
4149 unsigned long active_slabs = 0;
4150 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4151 const char *name;
4152 char *error = NULL;
4153 int node;
4154 struct kmem_list3 *l3;
4156 active_objs = 0;
4157 num_slabs = 0;
4158 for_each_online_node(node) {
4159 l3 = cachep->nodelists[node];
4160 if (!l3)
4161 continue;
4163 check_irq_on();
4164 spin_lock_irq(&l3->list_lock);
4166 list_for_each_entry(slabp, &l3->slabs_full, list) {
4167 if (slabp->inuse != cachep->num && !error)
4168 error = "slabs_full accounting error";
4169 active_objs += cachep->num;
4170 active_slabs++;
4172 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4173 if (slabp->inuse == cachep->num && !error)
4174 error = "slabs_partial inuse accounting error";
4175 if (!slabp->inuse && !error)
4176 error = "slabs_partial/inuse accounting error";
4177 active_objs += slabp->inuse;
4178 active_slabs++;
4180 list_for_each_entry(slabp, &l3->slabs_free, list) {
4181 if (slabp->inuse && !error)
4182 error = "slabs_free/inuse accounting error";
4183 num_slabs++;
4185 free_objects += l3->free_objects;
4186 if (l3->shared)
4187 shared_avail += l3->shared->avail;
4189 spin_unlock_irq(&l3->list_lock);
4191 num_slabs += active_slabs;
4192 num_objs = num_slabs * cachep->num;
4193 if (num_objs - active_objs != free_objects && !error)
4194 error = "free_objects accounting error";
4196 name = cachep->name;
4197 if (error)
4198 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4200 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4201 name, active_objs, num_objs, cachep->buffer_size,
4202 cachep->num, (1 << cachep->gfporder));
4203 seq_printf(m, " : tunables %4u %4u %4u",
4204 cachep->limit, cachep->batchcount, cachep->shared);
4205 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4206 active_slabs, num_slabs, shared_avail);
4207 #if STATS
4208 { /* list3 stats */
4209 unsigned long high = cachep->high_mark;
4210 unsigned long allocs = cachep->num_allocations;
4211 unsigned long grown = cachep->grown;
4212 unsigned long reaped = cachep->reaped;
4213 unsigned long errors = cachep->errors;
4214 unsigned long max_freeable = cachep->max_freeable;
4215 unsigned long node_allocs = cachep->node_allocs;
4216 unsigned long node_frees = cachep->node_frees;
4217 unsigned long overflows = cachep->node_overflow;
4219 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4220 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4221 reaped, errors, max_freeable, node_allocs,
4222 node_frees, overflows);
4224 /* cpu stats */
4226 unsigned long allochit = atomic_read(&cachep->allochit);
4227 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4228 unsigned long freehit = atomic_read(&cachep->freehit);
4229 unsigned long freemiss = atomic_read(&cachep->freemiss);
4231 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4232 allochit, allocmiss, freehit, freemiss);
4234 #endif
4235 seq_putc(m, '\n');
4236 return 0;
4240 * slabinfo_op - iterator that generates /proc/slabinfo
4242 * Output layout:
4243 * cache-name
4244 * num-active-objs
4245 * total-objs
4246 * object size
4247 * num-active-slabs
4248 * total-slabs
4249 * num-pages-per-slab
4250 * + further values on SMP and with statistics enabled
4253 static const struct seq_operations slabinfo_op = {
4254 .start = s_start,
4255 .next = s_next,
4256 .stop = s_stop,
4257 .show = s_show,
4260 #define MAX_SLABINFO_WRITE 128
4262 * slabinfo_write - Tuning for the slab allocator
4263 * @file: unused
4264 * @buffer: user buffer
4265 * @count: data length
4266 * @ppos: unused
4268 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4269 size_t count, loff_t *ppos)
4271 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4272 int limit, batchcount, shared, res;
4273 struct kmem_cache *cachep;
4275 if (count > MAX_SLABINFO_WRITE)
4276 return -EINVAL;
4277 if (copy_from_user(&kbuf, buffer, count))
4278 return -EFAULT;
4279 kbuf[MAX_SLABINFO_WRITE] = '\0';
4281 tmp = strchr(kbuf, ' ');
4282 if (!tmp)
4283 return -EINVAL;
4284 *tmp = '\0';
4285 tmp++;
4286 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4287 return -EINVAL;
4289 /* Find the cache in the chain of caches. */
4290 mutex_lock(&cache_chain_mutex);
4291 res = -EINVAL;
4292 list_for_each_entry(cachep, &cache_chain, next) {
4293 if (!strcmp(cachep->name, kbuf)) {
4294 if (limit < 1 || batchcount < 1 ||
4295 batchcount > limit || shared < 0) {
4296 res = 0;
4297 } else {
4298 res = do_tune_cpucache(cachep, limit,
4299 batchcount, shared,
4300 GFP_KERNEL);
4302 break;
4305 mutex_unlock(&cache_chain_mutex);
4306 if (res >= 0)
4307 res = count;
4308 return res;
4311 static int slabinfo_open(struct inode *inode, struct file *file)
4313 return seq_open(file, &slabinfo_op);
4316 static const struct file_operations proc_slabinfo_operations = {
4317 .open = slabinfo_open,
4318 .read = seq_read,
4319 .write = slabinfo_write,
4320 .llseek = seq_lseek,
4321 .release = seq_release,
4324 #ifdef CONFIG_DEBUG_SLAB_LEAK
4326 static void *leaks_start(struct seq_file *m, loff_t *pos)
4328 mutex_lock(&cache_chain_mutex);
4329 return seq_list_start(&cache_chain, *pos);
4332 static inline int add_caller(unsigned long *n, unsigned long v)
4334 unsigned long *p;
4335 int l;
4336 if (!v)
4337 return 1;
4338 l = n[1];
4339 p = n + 2;
4340 while (l) {
4341 int i = l/2;
4342 unsigned long *q = p + 2 * i;
4343 if (*q == v) {
4344 q[1]++;
4345 return 1;
4347 if (*q > v) {
4348 l = i;
4349 } else {
4350 p = q + 2;
4351 l -= i + 1;
4354 if (++n[1] == n[0])
4355 return 0;
4356 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4357 p[0] = v;
4358 p[1] = 1;
4359 return 1;
4362 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4364 void *p;
4365 int i;
4366 if (n[0] == n[1])
4367 return;
4368 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4369 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4370 continue;
4371 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4372 return;
4376 static void show_symbol(struct seq_file *m, unsigned long address)
4378 #ifdef CONFIG_KALLSYMS
4379 unsigned long offset, size;
4380 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4382 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4383 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4384 if (modname[0])
4385 seq_printf(m, " [%s]", modname);
4386 return;
4388 #endif
4389 seq_printf(m, "%p", (void *)address);
4392 static int leaks_show(struct seq_file *m, void *p)
4394 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4395 struct slab *slabp;
4396 struct kmem_list3 *l3;
4397 const char *name;
4398 unsigned long *n = m->private;
4399 int node;
4400 int i;
4402 if (!(cachep->flags & SLAB_STORE_USER))
4403 return 0;
4404 if (!(cachep->flags & SLAB_RED_ZONE))
4405 return 0;
4407 /* OK, we can do it */
4409 n[1] = 0;
4411 for_each_online_node(node) {
4412 l3 = cachep->nodelists[node];
4413 if (!l3)
4414 continue;
4416 check_irq_on();
4417 spin_lock_irq(&l3->list_lock);
4419 list_for_each_entry(slabp, &l3->slabs_full, list)
4420 handle_slab(n, cachep, slabp);
4421 list_for_each_entry(slabp, &l3->slabs_partial, list)
4422 handle_slab(n, cachep, slabp);
4423 spin_unlock_irq(&l3->list_lock);
4425 name = cachep->name;
4426 if (n[0] == n[1]) {
4427 /* Increase the buffer size */
4428 mutex_unlock(&cache_chain_mutex);
4429 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4430 if (!m->private) {
4431 /* Too bad, we are really out */
4432 m->private = n;
4433 mutex_lock(&cache_chain_mutex);
4434 return -ENOMEM;
4436 *(unsigned long *)m->private = n[0] * 2;
4437 kfree(n);
4438 mutex_lock(&cache_chain_mutex);
4439 /* Now make sure this entry will be retried */
4440 m->count = m->size;
4441 return 0;
4443 for (i = 0; i < n[1]; i++) {
4444 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4445 show_symbol(m, n[2*i+2]);
4446 seq_putc(m, '\n');
4449 return 0;
4452 static const struct seq_operations slabstats_op = {
4453 .start = leaks_start,
4454 .next = s_next,
4455 .stop = s_stop,
4456 .show = leaks_show,
4459 static int slabstats_open(struct inode *inode, struct file *file)
4461 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4462 int ret = -ENOMEM;
4463 if (n) {
4464 ret = seq_open(file, &slabstats_op);
4465 if (!ret) {
4466 struct seq_file *m = file->private_data;
4467 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4468 m->private = n;
4469 n = NULL;
4471 kfree(n);
4473 return ret;
4476 static const struct file_operations proc_slabstats_operations = {
4477 .open = slabstats_open,
4478 .read = seq_read,
4479 .llseek = seq_lseek,
4480 .release = seq_release_private,
4482 #endif
4484 static int __init slab_proc_init(void)
4486 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4487 #ifdef CONFIG_DEBUG_SLAB_LEAK
4488 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4489 #endif
4490 return 0;
4492 module_init(slab_proc_init);
4493 #endif
4496 * ksize - get the actual amount of memory allocated for a given object
4497 * @objp: Pointer to the object
4499 * kmalloc may internally round up allocations and return more memory
4500 * than requested. ksize() can be used to determine the actual amount of
4501 * memory allocated. The caller may use this additional memory, even though
4502 * a smaller amount of memory was initially specified with the kmalloc call.
4503 * The caller must guarantee that objp points to a valid object previously
4504 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4505 * must not be freed during the duration of the call.
4507 size_t ksize(const void *objp)
4509 BUG_ON(!objp);
4510 if (unlikely(objp == ZERO_SIZE_PTR))
4511 return 0;
4513 return obj_size(virt_to_cache(objp));
4515 EXPORT_SYMBOL(ksize);