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 intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in kmem_cache_t and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the semaphore 'cache_chain_sem'.
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.
80 #include <linux/config.h>
81 #include <linux/slab.h>
83 #include <linux/swap.h>
84 #include <linux/cache.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compiler.h>
88 #include <linux/seq_file.h>
89 #include <linux/notifier.h>
90 #include <linux/kallsyms.h>
91 #include <linux/cpu.h>
92 #include <linux/sysctl.h>
93 #include <linux/module.h>
94 #include <linux/rcupdate.h>
96 #include <asm/uaccess.h>
97 #include <asm/cacheflush.h>
98 #include <asm/tlbflush.h>
102 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
103 * SLAB_RED_ZONE & SLAB_POISON.
104 * 0 for faster, smaller code (especially in the critical paths).
106 * STATS - 1 to collect stats for /proc/slabinfo.
107 * 0 for faster, smaller code (especially in the critical paths).
109 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
112 #ifdef CONFIG_DEBUG_SLAB
115 #define FORCED_DEBUG 1
119 #define FORCED_DEBUG 0
123 /* Shouldn't this be in a header file somewhere? */
124 #define BYTES_PER_WORD sizeof(void *)
126 #ifndef cache_line_size
127 #define cache_line_size() L1_CACHE_BYTES
130 #ifndef ARCH_KMALLOC_MINALIGN
132 * Enforce a minimum alignment for the kmalloc caches.
133 * Usually, the kmalloc caches are cache_line_size() aligned, except when
134 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
135 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
136 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
137 * Note that this flag disables some debug features.
139 #define ARCH_KMALLOC_MINALIGN 0
142 #ifndef ARCH_SLAB_MINALIGN
144 * Enforce a minimum alignment for all caches.
145 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
146 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
147 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
148 * some debug features.
150 #define ARCH_SLAB_MINALIGN 0
153 #ifndef ARCH_KMALLOC_FLAGS
154 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
157 /* Legal flag mask for kmem_cache_create(). */
159 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
160 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
161 SLAB_NO_REAP | SLAB_CACHE_DMA | \
162 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
166 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
167 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
168 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
175 * Bufctl's are used for linking objs within a slab
178 * This implementation relies on "struct page" for locating the cache &
179 * slab an object belongs to.
180 * This allows the bufctl structure to be small (one int), but limits
181 * the number of objects a slab (not a cache) can contain when off-slab
182 * bufctls are used. The limit is the size of the largest general cache
183 * that does not use off-slab slabs.
184 * For 32bit archs with 4 kB pages, is this 56.
185 * This is not serious, as it is only for large objects, when it is unwise
186 * to have too many per slab.
187 * Note: This limit can be raised by introducing a general cache whose size
188 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
191 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
192 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
193 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
195 /* Max number of objs-per-slab for caches which use off-slab slabs.
196 * Needed to avoid a possible looping condition in cache_grow().
198 static unsigned long offslab_limit;
203 * Manages the objs in a slab. Placed either at the beginning of mem allocated
204 * for a slab, or allocated from an general cache.
205 * Slabs are chained into three list: fully used, partial, fully free slabs.
208 struct list_head list;
209 unsigned long colouroff;
210 void *s_mem; /* including colour offset */
211 unsigned int inuse; /* num of objs active in slab */
218 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
219 * arrange for kmem_freepages to be called via RCU. This is useful if
220 * we need to approach a kernel structure obliquely, from its address
221 * obtained without the usual locking. We can lock the structure to
222 * stabilize it and check it's still at the given address, only if we
223 * can be sure that the memory has not been meanwhile reused for some
224 * other kind of object (which our subsystem's lock might corrupt).
226 * rcu_read_lock before reading the address, then rcu_read_unlock after
227 * taking the spinlock within the structure expected at that address.
229 * We assume struct slab_rcu can overlay struct slab when destroying.
232 struct rcu_head head;
233 kmem_cache_t *cachep;
242 * - LIFO ordering, to hand out cache-warm objects from _alloc
243 * - reduce the number of linked list operations
244 * - reduce spinlock operations
246 * The limit is stored in the per-cpu structure to reduce the data cache
253 unsigned int batchcount;
254 unsigned int touched;
257 /* bootstrap: The caches do not work without cpuarrays anymore,
258 * but the cpuarrays are allocated from the generic caches...
260 #define BOOT_CPUCACHE_ENTRIES 1
261 struct arraycache_init {
262 struct array_cache cache;
263 void * entries[BOOT_CPUCACHE_ENTRIES];
267 * The slab lists of all objects.
268 * Hopefully reduce the internal fragmentation
269 * NUMA: The spinlock could be moved from the kmem_cache_t
270 * into this structure, too. Figure out what causes
271 * fewer cross-node spinlock operations.
274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
279 unsigned long next_reap;
280 struct array_cache *shared;
283 #define LIST3_INIT(parent) \
285 .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \
286 .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \
287 .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \
289 #define list3_data(cachep) \
293 #define list3_data_ptr(cachep, ptr) \
302 struct kmem_cache_s {
303 /* 1) per-cpu data, touched during every alloc/free */
304 struct array_cache *array[NR_CPUS];
305 unsigned int batchcount;
307 /* 2) touched by every alloc & free from the backend */
308 struct kmem_list3 lists;
309 /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */
310 unsigned int objsize;
311 unsigned int flags; /* constant flags */
312 unsigned int num; /* # of objs per slab */
313 unsigned int free_limit; /* upper limit of objects in the lists */
316 /* 3) cache_grow/shrink */
317 /* order of pgs per slab (2^n) */
318 unsigned int gfporder;
320 /* force GFP flags, e.g. GFP_DMA */
321 unsigned int gfpflags;
323 size_t colour; /* cache colouring range */
324 unsigned int colour_off; /* colour offset */
325 unsigned int colour_next; /* cache colouring */
326 kmem_cache_t *slabp_cache;
327 unsigned int slab_size;
328 unsigned int dflags; /* dynamic flags */
330 /* constructor func */
331 void (*ctor)(void *, kmem_cache_t *, unsigned long);
333 /* de-constructor func */
334 void (*dtor)(void *, kmem_cache_t *, unsigned long);
336 /* 4) cache creation/removal */
338 struct list_head next;
342 unsigned long num_active;
343 unsigned long num_allocations;
344 unsigned long high_mark;
346 unsigned long reaped;
347 unsigned long errors;
348 unsigned long max_freeable;
349 unsigned long node_allocs;
361 #define CFLGS_OFF_SLAB (0x80000000UL)
362 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
364 #define BATCHREFILL_LIMIT 16
365 /* Optimization question: fewer reaps means less
366 * probability for unnessary cpucache drain/refill cycles.
368 * OTHO the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
371 #define REAPTIMEOUT_CPUC (2*HZ)
372 #define REAPTIMEOUT_LIST3 (4*HZ)
375 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
376 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378 #define STATS_INC_GROWN(x) ((x)->grown++)
379 #define STATS_INC_REAPED(x) ((x)->reaped++)
380 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
383 #define STATS_INC_ERR(x) ((x)->errors++)
384 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385 #define STATS_SET_FREEABLE(x, i) \
386 do { if ((x)->max_freeable < i) \
387 (x)->max_freeable = i; \
390 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
391 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
392 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
393 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
395 #define STATS_INC_ACTIVE(x) do { } while (0)
396 #define STATS_DEC_ACTIVE(x) do { } while (0)
397 #define STATS_INC_ALLOCED(x) do { } while (0)
398 #define STATS_INC_GROWN(x) do { } while (0)
399 #define STATS_INC_REAPED(x) do { } while (0)
400 #define STATS_SET_HIGH(x) do { } while (0)
401 #define STATS_INC_ERR(x) do { } while (0)
402 #define STATS_INC_NODEALLOCS(x) do { } while (0)
403 #define STATS_SET_FREEABLE(x, i) \
406 #define STATS_INC_ALLOCHIT(x) do { } while (0)
407 #define STATS_INC_ALLOCMISS(x) do { } while (0)
408 #define STATS_INC_FREEHIT(x) do { } while (0)
409 #define STATS_INC_FREEMISS(x) do { } while (0)
413 /* Magic nums for obj red zoning.
414 * Placed in the first word before and the first word after an obj.
416 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
417 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
419 /* ...and for poisoning */
420 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
421 #define POISON_FREE 0x6b /* for use-after-free poisoning */
422 #define POISON_END 0xa5 /* end-byte of poisoning */
424 /* memory layout of objects:
426 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
427 * the end of an object is aligned with the end of the real
428 * allocation. Catches writes behind the end of the allocation.
429 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
431 * cachep->dbghead: The real object.
432 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
433 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
435 static int obj_dbghead(kmem_cache_t *cachep)
437 return cachep->dbghead;
440 static int obj_reallen(kmem_cache_t *cachep)
442 return cachep->reallen;
445 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
447 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
448 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
451 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
453 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
454 if (cachep->flags & SLAB_STORE_USER)
455 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
456 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
459 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
461 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
462 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
467 #define obj_dbghead(x) 0
468 #define obj_reallen(cachep) (cachep->objsize)
469 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
470 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
471 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
476 * Maximum size of an obj (in 2^order pages)
477 * and absolute limit for the gfp order.
479 #if defined(CONFIG_LARGE_ALLOCS)
480 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
481 #define MAX_GFP_ORDER 13 /* up to 32Mb */
482 #elif defined(CONFIG_MMU)
483 #define MAX_OBJ_ORDER 5 /* 32 pages */
484 #define MAX_GFP_ORDER 5 /* 32 pages */
486 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
487 #define MAX_GFP_ORDER 8 /* up to 1Mb */
491 * Do not go above this order unless 0 objects fit into the slab.
493 #define BREAK_GFP_ORDER_HI 1
494 #define BREAK_GFP_ORDER_LO 0
495 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
497 /* Macros for storing/retrieving the cachep and or slab from the
498 * global 'mem_map'. These are used to find the slab an obj belongs to.
499 * With kfree(), these are used to find the cache which an obj belongs to.
501 #define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
502 #define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
503 #define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
504 #define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
506 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
507 struct cache_sizes malloc_sizes[] = {
508 #define CACHE(x) { .cs_size = (x) },
509 #include <linux/kmalloc_sizes.h>
514 EXPORT_SYMBOL(malloc_sizes);
516 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
522 static struct cache_names __initdata cache_names[] = {
523 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
524 #include <linux/kmalloc_sizes.h>
529 static struct arraycache_init initarray_cache __initdata =
530 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
531 static struct arraycache_init initarray_generic =
532 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
534 /* internal cache of cache description objs */
535 static kmem_cache_t cache_cache = {
536 .lists = LIST3_INIT(cache_cache.lists),
538 .limit = BOOT_CPUCACHE_ENTRIES,
539 .objsize = sizeof(kmem_cache_t),
540 .flags = SLAB_NO_REAP,
541 .spinlock = SPIN_LOCK_UNLOCKED,
542 .name = "kmem_cache",
544 .reallen = sizeof(kmem_cache_t),
548 /* Guard access to the cache-chain. */
549 static struct semaphore cache_chain_sem;
550 static struct list_head cache_chain;
553 * vm_enough_memory() looks at this to determine how many
554 * slab-allocated pages are possibly freeable under pressure
556 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
558 atomic_t slab_reclaim_pages;
559 EXPORT_SYMBOL(slab_reclaim_pages);
562 * chicken and egg problem: delay the per-cpu array allocation
563 * until the general caches are up.
571 static DEFINE_PER_CPU(struct work_struct, reap_work);
573 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
574 static void enable_cpucache (kmem_cache_t *cachep);
575 static void cache_reap (void *unused);
577 static inline void ** ac_entry(struct array_cache *ac)
579 return (void**)(ac+1);
582 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
584 return cachep->array[smp_processor_id()];
587 static kmem_cache_t * kmem_find_general_cachep (size_t size, int gfpflags)
589 struct cache_sizes *csizep = malloc_sizes;
591 /* This function could be moved to the header file, and
592 * made inline so consumers can quickly determine what
593 * cache pointer they require.
595 for ( ; csizep->cs_size; csizep++) {
596 if (size > csizep->cs_size)
600 return (gfpflags & GFP_DMA) ? csizep->cs_dmacachep : csizep->cs_cachep;
603 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
604 static void cache_estimate (unsigned long gfporder, size_t size, size_t align,
605 int flags, size_t *left_over, unsigned int *num)
608 size_t wastage = PAGE_SIZE<<gfporder;
612 if (!(flags & CFLGS_OFF_SLAB)) {
613 base = sizeof(struct slab);
614 extra = sizeof(kmem_bufctl_t);
617 while (i*size + ALIGN(base+i*extra, align) <= wastage)
627 wastage -= ALIGN(base+i*extra, align);
628 *left_over = wastage;
631 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
633 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
635 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
636 function, cachep->name, msg);
641 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
642 * via the workqueue/eventd.
643 * Add the CPU number into the expiration time to minimize the possibility of
644 * the CPUs getting into lockstep and contending for the global cache chain
647 static void __devinit start_cpu_timer(int cpu)
649 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
652 * When this gets called from do_initcalls via cpucache_init(),
653 * init_workqueues() has already run, so keventd will be setup
656 if (keventd_up() && reap_work->func == NULL) {
657 INIT_WORK(reap_work, cache_reap, NULL);
658 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
662 static struct array_cache *alloc_arraycache(int cpu, int entries, int batchcount)
664 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
665 struct array_cache *nc = NULL;
668 nc = kmem_cache_alloc_node(kmem_find_general_cachep(memsize,
669 GFP_KERNEL), cpu_to_node(cpu));
672 nc = kmalloc(memsize, GFP_KERNEL);
676 nc->batchcount = batchcount;
682 static int __devinit cpuup_callback(struct notifier_block *nfb,
683 unsigned long action,
686 long cpu = (long)hcpu;
687 kmem_cache_t* cachep;
691 down(&cache_chain_sem);
692 list_for_each_entry(cachep, &cache_chain, next) {
693 struct array_cache *nc;
695 nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
699 spin_lock_irq(&cachep->spinlock);
700 cachep->array[cpu] = nc;
701 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
703 spin_unlock_irq(&cachep->spinlock);
706 up(&cache_chain_sem);
709 start_cpu_timer(cpu);
711 #ifdef CONFIG_HOTPLUG_CPU
714 case CPU_UP_CANCELED:
715 down(&cache_chain_sem);
717 list_for_each_entry(cachep, &cache_chain, next) {
718 struct array_cache *nc;
720 spin_lock_irq(&cachep->spinlock);
721 /* cpu is dead; no one can alloc from it. */
722 nc = cachep->array[cpu];
723 cachep->array[cpu] = NULL;
724 cachep->free_limit -= cachep->batchcount;
725 free_block(cachep, ac_entry(nc), nc->avail);
726 spin_unlock_irq(&cachep->spinlock);
729 up(&cache_chain_sem);
735 up(&cache_chain_sem);
739 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
742 * Called after the gfp() functions have been enabled, and before smp_init().
744 void __init kmem_cache_init(void)
747 struct cache_sizes *sizes;
748 struct cache_names *names;
751 * Fragmentation resistance on low memory - only use bigger
752 * page orders on machines with more than 32MB of memory.
754 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
755 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
758 /* Bootstrap is tricky, because several objects are allocated
759 * from caches that do not exist yet:
760 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
761 * structures of all caches, except cache_cache itself: cache_cache
762 * is statically allocated.
763 * Initially an __init data area is used for the head array, it's
764 * replaced with a kmalloc allocated array at the end of the bootstrap.
765 * 2) Create the first kmalloc cache.
766 * The kmem_cache_t for the new cache is allocated normally. An __init
767 * data area is used for the head array.
768 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
769 * 4) Replace the __init data head arrays for cache_cache and the first
770 * kmalloc cache with kmalloc allocated arrays.
771 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
774 /* 1) create the cache_cache */
775 init_MUTEX(&cache_chain_sem);
776 INIT_LIST_HEAD(&cache_chain);
777 list_add(&cache_cache.next, &cache_chain);
778 cache_cache.colour_off = cache_line_size();
779 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
781 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
783 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
784 &left_over, &cache_cache.num);
785 if (!cache_cache.num)
788 cache_cache.colour = left_over/cache_cache.colour_off;
789 cache_cache.colour_next = 0;
790 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
791 sizeof(struct slab), cache_line_size());
793 /* 2+3) create the kmalloc caches */
794 sizes = malloc_sizes;
797 while (sizes->cs_size) {
798 /* For performance, all the general caches are L1 aligned.
799 * This should be particularly beneficial on SMP boxes, as it
800 * eliminates "false sharing".
801 * Note for systems short on memory removing the alignment will
802 * allow tighter packing of the smaller caches. */
803 sizes->cs_cachep = kmem_cache_create(names->name,
804 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
805 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
807 /* Inc off-slab bufctl limit until the ceiling is hit. */
808 if (!(OFF_SLAB(sizes->cs_cachep))) {
809 offslab_limit = sizes->cs_size-sizeof(struct slab);
810 offslab_limit /= sizeof(kmem_bufctl_t);
813 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
814 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
815 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
821 /* 4) Replace the bootstrap head arrays */
825 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
827 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
828 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
829 cache_cache.array[smp_processor_id()] = ptr;
832 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
834 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
835 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
836 sizeof(struct arraycache_init));
837 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
841 /* 5) resize the head arrays to their final sizes */
843 kmem_cache_t *cachep;
844 down(&cache_chain_sem);
845 list_for_each_entry(cachep, &cache_chain, next)
846 enable_cpucache(cachep);
847 up(&cache_chain_sem);
851 g_cpucache_up = FULL;
853 /* Register a cpu startup notifier callback
854 * that initializes ac_data for all new cpus
856 register_cpu_notifier(&cpucache_notifier);
859 /* The reap timers are started later, with a module init call:
860 * That part of the kernel is not yet operational.
864 static int __init cpucache_init(void)
869 * Register the timers that return unneeded
872 for (cpu = 0; cpu < NR_CPUS; cpu++) {
874 start_cpu_timer(cpu);
880 __initcall(cpucache_init);
883 * Interface to system's page allocator. No need to hold the cache-lock.
885 * If we requested dmaable memory, we will get it. Even if we
886 * did not request dmaable memory, we might get it, but that
887 * would be relatively rare and ignorable.
889 static void *kmem_getpages(kmem_cache_t *cachep, int flags, int nodeid)
895 flags |= cachep->gfpflags;
896 if (likely(nodeid == -1)) {
897 page = alloc_pages(flags, cachep->gfporder);
899 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
903 addr = page_address(page);
905 i = (1 << cachep->gfporder);
906 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
907 atomic_add(i, &slab_reclaim_pages);
908 add_page_state(nr_slab, i);
917 * Interface to system's page release.
919 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
921 unsigned long i = (1<<cachep->gfporder);
922 struct page *page = virt_to_page(addr);
923 const unsigned long nr_freed = i;
926 if (!TestClearPageSlab(page))
930 sub_page_state(nr_slab, nr_freed);
931 if (current->reclaim_state)
932 current->reclaim_state->reclaimed_slab += nr_freed;
933 free_pages((unsigned long)addr, cachep->gfporder);
934 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
935 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
938 static void kmem_rcu_free(struct rcu_head *head)
940 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
941 kmem_cache_t *cachep = slab_rcu->cachep;
943 kmem_freepages(cachep, slab_rcu->addr);
944 if (OFF_SLAB(cachep))
945 kmem_cache_free(cachep->slabp_cache, slab_rcu);
950 #ifdef CONFIG_DEBUG_PAGEALLOC
951 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, unsigned long caller)
953 int size = obj_reallen(cachep);
955 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
957 if (size < 5*sizeof(unsigned long))
962 *addr++=smp_processor_id();
963 size -= 3*sizeof(unsigned long);
965 unsigned long *sptr = &caller;
966 unsigned long svalue;
968 while (!kstack_end(sptr)) {
970 if (kernel_text_address(svalue)) {
972 size -= sizeof(unsigned long);
973 if (size <= sizeof(unsigned long))
983 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
985 int size = obj_reallen(cachep);
986 addr = &((char*)addr)[obj_dbghead(cachep)];
988 memset(addr, val, size);
989 *(unsigned char *)(addr+size-1) = POISON_END;
992 static void dump_line(char *data, int offset, int limit)
995 printk(KERN_ERR "%03x:", offset);
996 for (i=0;i<limit;i++) {
997 printk(" %02x", (unsigned char)data[offset+i]);
1005 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1010 if (cachep->flags & SLAB_RED_ZONE) {
1011 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1012 *dbg_redzone1(cachep, objp),
1013 *dbg_redzone2(cachep, objp));
1016 if (cachep->flags & SLAB_STORE_USER) {
1017 printk(KERN_ERR "Last user: [<%p>]",
1018 *dbg_userword(cachep, objp));
1019 print_symbol("(%s)",
1020 (unsigned long)*dbg_userword(cachep, objp));
1023 realobj = (char*)objp+obj_dbghead(cachep);
1024 size = obj_reallen(cachep);
1025 for (i=0; i<size && lines;i+=16, lines--) {
1030 dump_line(realobj, i, limit);
1034 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1040 realobj = (char*)objp+obj_dbghead(cachep);
1041 size = obj_reallen(cachep);
1043 for (i=0;i<size;i++) {
1044 char exp = POISON_FREE;
1047 if (realobj[i] != exp) {
1052 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1054 print_objinfo(cachep, objp, 0);
1056 /* Hexdump the affected line */
1061 dump_line(realobj, i, limit);
1064 /* Limit to 5 lines */
1070 /* Print some data about the neighboring objects, if they
1073 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1076 objnr = (objp-slabp->s_mem)/cachep->objsize;
1078 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1079 realobj = (char*)objp+obj_dbghead(cachep);
1080 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1082 print_objinfo(cachep, objp, 2);
1084 if (objnr+1 < cachep->num) {
1085 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1086 realobj = (char*)objp+obj_dbghead(cachep);
1087 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1089 print_objinfo(cachep, objp, 2);
1095 /* Destroy all the objs in a slab, and release the mem back to the system.
1096 * Before calling the slab must have been unlinked from the cache.
1097 * The cache-lock is not held/needed.
1099 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1101 void *addr = slabp->s_mem - slabp->colouroff;
1105 for (i = 0; i < cachep->num; i++) {
1106 void *objp = slabp->s_mem + cachep->objsize * i;
1108 if (cachep->flags & SLAB_POISON) {
1109 #ifdef CONFIG_DEBUG_PAGEALLOC
1110 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1111 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1113 check_poison_obj(cachep, objp);
1115 check_poison_obj(cachep, objp);
1118 if (cachep->flags & SLAB_RED_ZONE) {
1119 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1120 slab_error(cachep, "start of a freed object "
1122 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1123 slab_error(cachep, "end of a freed object "
1126 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1127 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1132 for (i = 0; i < cachep->num; i++) {
1133 void* objp = slabp->s_mem+cachep->objsize*i;
1134 (cachep->dtor)(objp, cachep, 0);
1139 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1140 struct slab_rcu *slab_rcu;
1142 slab_rcu = (struct slab_rcu *) slabp;
1143 slab_rcu->cachep = cachep;
1144 slab_rcu->addr = addr;
1145 call_rcu(&slab_rcu->head, kmem_rcu_free);
1147 kmem_freepages(cachep, addr);
1148 if (OFF_SLAB(cachep))
1149 kmem_cache_free(cachep->slabp_cache, slabp);
1154 * kmem_cache_create - Create a cache.
1155 * @name: A string which is used in /proc/slabinfo to identify this cache.
1156 * @size: The size of objects to be created in this cache.
1157 * @align: The required alignment for the objects.
1158 * @flags: SLAB flags
1159 * @ctor: A constructor for the objects.
1160 * @dtor: A destructor for the objects.
1162 * Returns a ptr to the cache on success, NULL on failure.
1163 * Cannot be called within a int, but can be interrupted.
1164 * The @ctor is run when new pages are allocated by the cache
1165 * and the @dtor is run before the pages are handed back.
1167 * @name must be valid until the cache is destroyed. This implies that
1168 * the module calling this has to destroy the cache before getting
1173 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1174 * to catch references to uninitialised memory.
1176 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1177 * for buffer overruns.
1179 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1182 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1183 * cacheline. This can be beneficial if you're counting cycles as closely
1187 kmem_cache_create (const char *name, size_t size, size_t align,
1188 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1189 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1191 size_t left_over, slab_size, ralign;
1192 kmem_cache_t *cachep = NULL;
1195 * Sanity checks... these are all serious usage bugs.
1199 (size < BYTES_PER_WORD) ||
1200 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1202 printk(KERN_ERR "%s: Early error in slab %s\n",
1203 __FUNCTION__, name);
1208 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1209 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1210 /* No constructor, but inital state check requested */
1211 printk(KERN_ERR "%s: No con, but init state check "
1212 "requested - %s\n", __FUNCTION__, name);
1213 flags &= ~SLAB_DEBUG_INITIAL;
1218 * Enable redzoning and last user accounting, except for caches with
1219 * large objects, if the increased size would increase the object size
1220 * above the next power of two: caches with object sizes just above a
1221 * power of two have a significant amount of internal fragmentation.
1223 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1224 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1225 if (!(flags & SLAB_DESTROY_BY_RCU))
1226 flags |= SLAB_POISON;
1228 if (flags & SLAB_DESTROY_BY_RCU)
1229 BUG_ON(flags & SLAB_POISON);
1231 if (flags & SLAB_DESTROY_BY_RCU)
1235 * Always checks flags, a caller might be expecting debug
1236 * support which isn't available.
1238 if (flags & ~CREATE_MASK)
1241 /* Check that size is in terms of words. This is needed to avoid
1242 * unaligned accesses for some archs when redzoning is used, and makes
1243 * sure any on-slab bufctl's are also correctly aligned.
1245 if (size & (BYTES_PER_WORD-1)) {
1246 size += (BYTES_PER_WORD-1);
1247 size &= ~(BYTES_PER_WORD-1);
1250 /* calculate out the final buffer alignment: */
1251 /* 1) arch recommendation: can be overridden for debug */
1252 if (flags & SLAB_HWCACHE_ALIGN) {
1253 /* Default alignment: as specified by the arch code.
1254 * Except if an object is really small, then squeeze multiple
1255 * objects into one cacheline.
1257 ralign = cache_line_size();
1258 while (size <= ralign/2)
1261 ralign = BYTES_PER_WORD;
1263 /* 2) arch mandated alignment: disables debug if necessary */
1264 if (ralign < ARCH_SLAB_MINALIGN) {
1265 ralign = ARCH_SLAB_MINALIGN;
1266 if (ralign > BYTES_PER_WORD)
1267 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1269 /* 3) caller mandated alignment: disables debug if necessary */
1270 if (ralign < align) {
1272 if (ralign > BYTES_PER_WORD)
1273 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1275 /* 4) Store it. Note that the debug code below can reduce
1276 * the alignment to BYTES_PER_WORD.
1280 /* Get cache's description obj. */
1281 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1284 memset(cachep, 0, sizeof(kmem_cache_t));
1287 cachep->reallen = size;
1289 if (flags & SLAB_RED_ZONE) {
1290 /* redzoning only works with word aligned caches */
1291 align = BYTES_PER_WORD;
1293 /* add space for red zone words */
1294 cachep->dbghead += BYTES_PER_WORD;
1295 size += 2*BYTES_PER_WORD;
1297 if (flags & SLAB_STORE_USER) {
1298 /* user store requires word alignment and
1299 * one word storage behind the end of the real
1302 align = BYTES_PER_WORD;
1303 size += BYTES_PER_WORD;
1305 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1306 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1307 cachep->dbghead += PAGE_SIZE - size;
1313 /* Determine if the slab management is 'on' or 'off' slab. */
1314 if (size >= (PAGE_SIZE>>3))
1316 * Size is large, assume best to place the slab management obj
1317 * off-slab (should allow better packing of objs).
1319 flags |= CFLGS_OFF_SLAB;
1321 size = ALIGN(size, align);
1323 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1325 * A VFS-reclaimable slab tends to have most allocations
1326 * as GFP_NOFS and we really don't want to have to be allocating
1327 * higher-order pages when we are unable to shrink dcache.
1329 cachep->gfporder = 0;
1330 cache_estimate(cachep->gfporder, size, align, flags,
1331 &left_over, &cachep->num);
1334 * Calculate size (in pages) of slabs, and the num of objs per
1335 * slab. This could be made much more intelligent. For now,
1336 * try to avoid using high page-orders for slabs. When the
1337 * gfp() funcs are more friendly towards high-order requests,
1338 * this should be changed.
1341 unsigned int break_flag = 0;
1343 cache_estimate(cachep->gfporder, size, align, flags,
1344 &left_over, &cachep->num);
1347 if (cachep->gfporder >= MAX_GFP_ORDER)
1351 if (flags & CFLGS_OFF_SLAB &&
1352 cachep->num > offslab_limit) {
1353 /* This num of objs will cause problems. */
1360 * Large num of objs is good, but v. large slabs are
1361 * currently bad for the gfp()s.
1363 if (cachep->gfporder >= slab_break_gfp_order)
1366 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1367 break; /* Acceptable internal fragmentation. */
1374 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1375 kmem_cache_free(&cache_cache, cachep);
1379 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1380 + sizeof(struct slab), align);
1383 * If the slab has been placed off-slab, and we have enough space then
1384 * move it on-slab. This is at the expense of any extra colouring.
1386 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1387 flags &= ~CFLGS_OFF_SLAB;
1388 left_over -= slab_size;
1391 if (flags & CFLGS_OFF_SLAB) {
1392 /* really off slab. No need for manual alignment */
1393 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1396 cachep->colour_off = cache_line_size();
1397 /* Offset must be a multiple of the alignment. */
1398 if (cachep->colour_off < align)
1399 cachep->colour_off = align;
1400 cachep->colour = left_over/cachep->colour_off;
1401 cachep->slab_size = slab_size;
1402 cachep->flags = flags;
1403 cachep->gfpflags = 0;
1404 if (flags & SLAB_CACHE_DMA)
1405 cachep->gfpflags |= GFP_DMA;
1406 spin_lock_init(&cachep->spinlock);
1407 cachep->objsize = size;
1409 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1410 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1411 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1413 if (flags & CFLGS_OFF_SLAB)
1414 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1415 cachep->ctor = ctor;
1416 cachep->dtor = dtor;
1417 cachep->name = name;
1419 /* Don't let CPUs to come and go */
1422 if (g_cpucache_up == FULL) {
1423 enable_cpucache(cachep);
1425 if (g_cpucache_up == NONE) {
1426 /* Note: the first kmem_cache_create must create
1427 * the cache that's used by kmalloc(24), otherwise
1428 * the creation of further caches will BUG().
1430 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1431 g_cpucache_up = PARTIAL;
1433 cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL);
1435 BUG_ON(!ac_data(cachep));
1436 ac_data(cachep)->avail = 0;
1437 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1438 ac_data(cachep)->batchcount = 1;
1439 ac_data(cachep)->touched = 0;
1440 cachep->batchcount = 1;
1441 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1442 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1446 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1447 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1449 /* Need the semaphore to access the chain. */
1450 down(&cache_chain_sem);
1452 struct list_head *p;
1453 mm_segment_t old_fs;
1457 list_for_each(p, &cache_chain) {
1458 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1460 /* This happens when the module gets unloaded and doesn't
1461 destroy its slab cache and noone else reuses the vmalloc
1462 area of the module. Print a warning. */
1463 if (__get_user(tmp,pc->name)) {
1464 printk("SLAB: cache with size %d has lost its name\n",
1468 if (!strcmp(pc->name,name)) {
1469 printk("kmem_cache_create: duplicate cache %s\n",name);
1470 up(&cache_chain_sem);
1471 unlock_cpu_hotplug();
1478 /* cache setup completed, link it into the list */
1479 list_add(&cachep->next, &cache_chain);
1480 up(&cache_chain_sem);
1481 unlock_cpu_hotplug();
1483 if (!cachep && (flags & SLAB_PANIC))
1484 panic("kmem_cache_create(): failed to create slab `%s'\n",
1488 EXPORT_SYMBOL(kmem_cache_create);
1491 static void check_irq_off(void)
1493 BUG_ON(!irqs_disabled());
1496 static void check_irq_on(void)
1498 BUG_ON(irqs_disabled());
1501 static void check_spinlock_acquired(kmem_cache_t *cachep)
1505 BUG_ON(spin_trylock(&cachep->spinlock));
1509 #define check_irq_off() do { } while(0)
1510 #define check_irq_on() do { } while(0)
1511 #define check_spinlock_acquired(x) do { } while(0)
1515 * Waits for all CPUs to execute func().
1517 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1522 local_irq_disable();
1526 if (smp_call_function(func, arg, 1, 1))
1532 static void drain_array_locked(kmem_cache_t* cachep,
1533 struct array_cache *ac, int force);
1535 static void do_drain(void *arg)
1537 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1538 struct array_cache *ac;
1541 ac = ac_data(cachep);
1542 spin_lock(&cachep->spinlock);
1543 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1544 spin_unlock(&cachep->spinlock);
1548 static void drain_cpu_caches(kmem_cache_t *cachep)
1550 smp_call_function_all_cpus(do_drain, cachep);
1552 spin_lock_irq(&cachep->spinlock);
1553 if (cachep->lists.shared)
1554 drain_array_locked(cachep, cachep->lists.shared, 1);
1555 spin_unlock_irq(&cachep->spinlock);
1559 /* NUMA shrink all list3s */
1560 static int __cache_shrink(kmem_cache_t *cachep)
1565 drain_cpu_caches(cachep);
1568 spin_lock_irq(&cachep->spinlock);
1571 struct list_head *p;
1573 p = cachep->lists.slabs_free.prev;
1574 if (p == &cachep->lists.slabs_free)
1577 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1582 list_del(&slabp->list);
1584 cachep->lists.free_objects -= cachep->num;
1585 spin_unlock_irq(&cachep->spinlock);
1586 slab_destroy(cachep, slabp);
1587 spin_lock_irq(&cachep->spinlock);
1589 ret = !list_empty(&cachep->lists.slabs_full) ||
1590 !list_empty(&cachep->lists.slabs_partial);
1591 spin_unlock_irq(&cachep->spinlock);
1596 * kmem_cache_shrink - Shrink a cache.
1597 * @cachep: The cache to shrink.
1599 * Releases as many slabs as possible for a cache.
1600 * To help debugging, a zero exit status indicates all slabs were released.
1602 int kmem_cache_shrink(kmem_cache_t *cachep)
1604 if (!cachep || in_interrupt())
1607 return __cache_shrink(cachep);
1610 EXPORT_SYMBOL(kmem_cache_shrink);
1613 * kmem_cache_destroy - delete a cache
1614 * @cachep: the cache to destroy
1616 * Remove a kmem_cache_t object from the slab cache.
1617 * Returns 0 on success.
1619 * It is expected this function will be called by a module when it is
1620 * unloaded. This will remove the cache completely, and avoid a duplicate
1621 * cache being allocated each time a module is loaded and unloaded, if the
1622 * module doesn't have persistent in-kernel storage across loads and unloads.
1624 * The cache must be empty before calling this function.
1626 * The caller must guarantee that noone will allocate memory from the cache
1627 * during the kmem_cache_destroy().
1629 int kmem_cache_destroy (kmem_cache_t * cachep)
1633 if (!cachep || in_interrupt())
1636 /* Don't let CPUs to come and go */
1639 /* Find the cache in the chain of caches. */
1640 down(&cache_chain_sem);
1642 * the chain is never empty, cache_cache is never destroyed
1644 list_del(&cachep->next);
1645 up(&cache_chain_sem);
1647 if (__cache_shrink(cachep)) {
1648 slab_error(cachep, "Can't free all objects");
1649 down(&cache_chain_sem);
1650 list_add(&cachep->next,&cache_chain);
1651 up(&cache_chain_sem);
1652 unlock_cpu_hotplug();
1656 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1657 synchronize_kernel();
1659 /* no cpu_online check required here since we clear the percpu
1660 * array on cpu offline and set this to NULL.
1662 for (i = 0; i < NR_CPUS; i++)
1663 kfree(cachep->array[i]);
1665 /* NUMA: free the list3 structures */
1666 kfree(cachep->lists.shared);
1667 cachep->lists.shared = NULL;
1668 kmem_cache_free(&cache_cache, cachep);
1670 unlock_cpu_hotplug();
1675 EXPORT_SYMBOL(kmem_cache_destroy);
1677 /* Get the memory for a slab management obj. */
1678 static struct slab* alloc_slabmgmt (kmem_cache_t *cachep,
1679 void *objp, int colour_off, int local_flags)
1683 if (OFF_SLAB(cachep)) {
1684 /* Slab management obj is off-slab. */
1685 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1689 slabp = objp+colour_off;
1690 colour_off += cachep->slab_size;
1693 slabp->colouroff = colour_off;
1694 slabp->s_mem = objp+colour_off;
1699 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1701 return (kmem_bufctl_t *)(slabp+1);
1704 static void cache_init_objs (kmem_cache_t * cachep,
1705 struct slab * slabp, unsigned long ctor_flags)
1709 for (i = 0; i < cachep->num; i++) {
1710 void* objp = slabp->s_mem+cachep->objsize*i;
1712 /* need to poison the objs? */
1713 if (cachep->flags & SLAB_POISON)
1714 poison_obj(cachep, objp, POISON_FREE);
1715 if (cachep->flags & SLAB_STORE_USER)
1716 *dbg_userword(cachep, objp) = NULL;
1718 if (cachep->flags & SLAB_RED_ZONE) {
1719 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1720 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1723 * Constructors are not allowed to allocate memory from
1724 * the same cache which they are a constructor for.
1725 * Otherwise, deadlock. They must also be threaded.
1727 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1728 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1730 if (cachep->flags & SLAB_RED_ZONE) {
1731 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1732 slab_error(cachep, "constructor overwrote the"
1733 " end of an object");
1734 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1735 slab_error(cachep, "constructor overwrote the"
1736 " start of an object");
1738 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1739 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1742 cachep->ctor(objp, cachep, ctor_flags);
1744 slab_bufctl(slabp)[i] = i+1;
1746 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1750 static void kmem_flagcheck(kmem_cache_t *cachep, int flags)
1752 if (flags & SLAB_DMA) {
1753 if (!(cachep->gfpflags & GFP_DMA))
1756 if (cachep->gfpflags & GFP_DMA)
1761 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
1766 /* Nasty!!!!!! I hope this is OK. */
1767 i = 1 << cachep->gfporder;
1768 page = virt_to_page(objp);
1770 SET_PAGE_CACHE(page, cachep);
1771 SET_PAGE_SLAB(page, slabp);
1777 * Grow (by 1) the number of slabs within a cache. This is called by
1778 * kmem_cache_alloc() when there are no active objs left in a cache.
1780 static int cache_grow (kmem_cache_t * cachep, int flags, int nodeid)
1786 unsigned long ctor_flags;
1788 /* Be lazy and only check for valid flags here,
1789 * keeping it out of the critical path in kmem_cache_alloc().
1791 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1793 if (flags & SLAB_NO_GROW)
1796 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1797 local_flags = (flags & SLAB_LEVEL_MASK);
1798 if (!(local_flags & __GFP_WAIT))
1800 * Not allowed to sleep. Need to tell a constructor about
1801 * this - it might need to know...
1803 ctor_flags |= SLAB_CTOR_ATOMIC;
1805 /* About to mess with non-constant members - lock. */
1807 spin_lock(&cachep->spinlock);
1809 /* Get colour for the slab, and cal the next value. */
1810 offset = cachep->colour_next;
1811 cachep->colour_next++;
1812 if (cachep->colour_next >= cachep->colour)
1813 cachep->colour_next = 0;
1814 offset *= cachep->colour_off;
1816 spin_unlock(&cachep->spinlock);
1818 if (local_flags & __GFP_WAIT)
1822 * The test for missing atomic flag is performed here, rather than
1823 * the more obvious place, simply to reduce the critical path length
1824 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1825 * will eventually be caught here (where it matters).
1827 kmem_flagcheck(cachep, flags);
1830 /* Get mem for the objs. */
1831 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
1834 /* Get slab management. */
1835 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1838 set_slab_attr(cachep, slabp, objp);
1840 cache_init_objs(cachep, slabp, ctor_flags);
1842 if (local_flags & __GFP_WAIT)
1843 local_irq_disable();
1845 spin_lock(&cachep->spinlock);
1847 /* Make slab active. */
1848 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1849 STATS_INC_GROWN(cachep);
1850 list3_data(cachep)->free_objects += cachep->num;
1851 spin_unlock(&cachep->spinlock);
1854 kmem_freepages(cachep, objp);
1856 if (local_flags & __GFP_WAIT)
1857 local_irq_disable();
1864 * Perform extra freeing checks:
1865 * - detect bad pointers.
1866 * - POISON/RED_ZONE checking
1867 * - destructor calls, for caches with POISON+dtor
1869 static void kfree_debugcheck(const void *objp)
1873 if (!virt_addr_valid(objp)) {
1874 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1875 (unsigned long)objp);
1878 page = virt_to_page(objp);
1879 if (!PageSlab(page)) {
1880 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1885 static void *cache_free_debugcheck (kmem_cache_t * cachep, void * objp, void *caller)
1891 objp -= obj_dbghead(cachep);
1892 kfree_debugcheck(objp);
1893 page = virt_to_page(objp);
1895 if (GET_PAGE_CACHE(page) != cachep) {
1896 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1897 GET_PAGE_CACHE(page),cachep);
1898 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1899 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1902 slabp = GET_PAGE_SLAB(page);
1904 if (cachep->flags & SLAB_RED_ZONE) {
1905 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1906 slab_error(cachep, "double free, or memory outside"
1907 " object was overwritten");
1908 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1909 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1911 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1912 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1914 if (cachep->flags & SLAB_STORE_USER)
1915 *dbg_userword(cachep, objp) = caller;
1917 objnr = (objp-slabp->s_mem)/cachep->objsize;
1919 BUG_ON(objnr >= cachep->num);
1920 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1922 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1923 /* Need to call the slab's constructor so the
1924 * caller can perform a verify of its state (debugging).
1925 * Called without the cache-lock held.
1927 cachep->ctor(objp+obj_dbghead(cachep),
1928 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1930 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1931 /* we want to cache poison the object,
1932 * call the destruction callback
1934 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1936 if (cachep->flags & SLAB_POISON) {
1937 #ifdef CONFIG_DEBUG_PAGEALLOC
1938 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1939 store_stackinfo(cachep, objp, (unsigned long)caller);
1940 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1942 poison_obj(cachep, objp, POISON_FREE);
1945 poison_obj(cachep, objp, POISON_FREE);
1951 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
1956 check_spinlock_acquired(cachep);
1957 /* Check slab's freelist to see if this obj is there. */
1958 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
1960 if (entries > cachep->num || i < 0 || i >= cachep->num)
1963 if (entries != cachep->num - slabp->inuse) {
1966 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
1967 cachep->name, cachep->num, slabp, slabp->inuse);
1968 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
1970 printk("\n%03x:", i);
1971 printk(" %02x", ((unsigned char*)slabp)[i]);
1978 #define kfree_debugcheck(x) do { } while(0)
1979 #define cache_free_debugcheck(x,objp,z) (objp)
1980 #define check_slabp(x,y) do { } while(0)
1983 static void* cache_alloc_refill(kmem_cache_t* cachep, int flags)
1986 struct kmem_list3 *l3;
1987 struct array_cache *ac;
1990 ac = ac_data(cachep);
1992 batchcount = ac->batchcount;
1993 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
1994 /* if there was little recent activity on this
1995 * cache, then perform only a partial refill.
1996 * Otherwise we could generate refill bouncing.
1998 batchcount = BATCHREFILL_LIMIT;
2000 l3 = list3_data(cachep);
2002 BUG_ON(ac->avail > 0);
2003 spin_lock(&cachep->spinlock);
2005 struct array_cache *shared_array = l3->shared;
2006 if (shared_array->avail) {
2007 if (batchcount > shared_array->avail)
2008 batchcount = shared_array->avail;
2009 shared_array->avail -= batchcount;
2010 ac->avail = batchcount;
2011 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
2012 sizeof(void*)*batchcount);
2013 shared_array->touched = 1;
2017 while (batchcount > 0) {
2018 struct list_head *entry;
2020 /* Get slab alloc is to come from. */
2021 entry = l3->slabs_partial.next;
2022 if (entry == &l3->slabs_partial) {
2023 l3->free_touched = 1;
2024 entry = l3->slabs_free.next;
2025 if (entry == &l3->slabs_free)
2029 slabp = list_entry(entry, struct slab, list);
2030 check_slabp(cachep, slabp);
2031 check_spinlock_acquired(cachep);
2032 while (slabp->inuse < cachep->num && batchcount--) {
2034 STATS_INC_ALLOCED(cachep);
2035 STATS_INC_ACTIVE(cachep);
2036 STATS_SET_HIGH(cachep);
2038 /* get obj pointer */
2039 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
2042 next = slab_bufctl(slabp)[slabp->free];
2044 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2048 check_slabp(cachep, slabp);
2050 /* move slabp to correct slabp list: */
2051 list_del(&slabp->list);
2052 if (slabp->free == BUFCTL_END)
2053 list_add(&slabp->list, &l3->slabs_full);
2055 list_add(&slabp->list, &l3->slabs_partial);
2059 l3->free_objects -= ac->avail;
2061 spin_unlock(&cachep->spinlock);
2063 if (unlikely(!ac->avail)) {
2065 x = cache_grow(cachep, flags, -1);
2067 // cache_grow can reenable interrupts, then ac could change.
2068 ac = ac_data(cachep);
2069 if (!x && ac->avail == 0) // no objects in sight? abort
2072 if (!ac->avail) // objects refilled by interrupt?
2076 return ac_entry(ac)[--ac->avail];
2080 cache_alloc_debugcheck_before(kmem_cache_t *cachep, int flags)
2082 might_sleep_if(flags & __GFP_WAIT);
2084 kmem_flagcheck(cachep, flags);
2090 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
2091 unsigned long flags, void *objp, void *caller)
2095 if (cachep->flags & SLAB_POISON) {
2096 #ifdef CONFIG_DEBUG_PAGEALLOC
2097 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2098 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2100 check_poison_obj(cachep, objp);
2102 check_poison_obj(cachep, objp);
2104 poison_obj(cachep, objp, POISON_INUSE);
2106 if (cachep->flags & SLAB_STORE_USER)
2107 *dbg_userword(cachep, objp) = caller;
2109 if (cachep->flags & SLAB_RED_ZONE) {
2110 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2111 slab_error(cachep, "double free, or memory outside"
2112 " object was overwritten");
2113 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2114 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2116 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2117 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2119 objp += obj_dbghead(cachep);
2120 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2121 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2123 if (!(flags & __GFP_WAIT))
2124 ctor_flags |= SLAB_CTOR_ATOMIC;
2126 cachep->ctor(objp, cachep, ctor_flags);
2131 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2135 static inline void * __cache_alloc (kmem_cache_t *cachep, int flags)
2137 unsigned long save_flags;
2139 struct array_cache *ac;
2141 cache_alloc_debugcheck_before(cachep, flags);
2143 local_irq_save(save_flags);
2144 ac = ac_data(cachep);
2145 if (likely(ac->avail)) {
2146 STATS_INC_ALLOCHIT(cachep);
2148 objp = ac_entry(ac)[--ac->avail];
2150 STATS_INC_ALLOCMISS(cachep);
2151 objp = cache_alloc_refill(cachep, flags);
2153 local_irq_restore(save_flags);
2154 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2159 * NUMA: different approach needed if the spinlock is moved into
2163 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2167 check_spinlock_acquired(cachep);
2169 /* NUMA: move add into loop */
2170 cachep->lists.free_objects += nr_objects;
2172 for (i = 0; i < nr_objects; i++) {
2173 void *objp = objpp[i];
2177 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2178 list_del(&slabp->list);
2179 objnr = (objp - slabp->s_mem) / cachep->objsize;
2180 check_slabp(cachep, slabp);
2182 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2183 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2184 cachep->name, objp);
2188 slab_bufctl(slabp)[objnr] = slabp->free;
2189 slabp->free = objnr;
2190 STATS_DEC_ACTIVE(cachep);
2192 check_slabp(cachep, slabp);
2194 /* fixup slab chains */
2195 if (slabp->inuse == 0) {
2196 if (cachep->lists.free_objects > cachep->free_limit) {
2197 cachep->lists.free_objects -= cachep->num;
2198 slab_destroy(cachep, slabp);
2200 list_add(&slabp->list,
2201 &list3_data_ptr(cachep, objp)->slabs_free);
2204 /* Unconditionally move a slab to the end of the
2205 * partial list on free - maximum time for the
2206 * other objects to be freed, too.
2208 list_add_tail(&slabp->list,
2209 &list3_data_ptr(cachep, objp)->slabs_partial);
2214 static void cache_flusharray (kmem_cache_t* cachep, struct array_cache *ac)
2218 batchcount = ac->batchcount;
2220 BUG_ON(!batchcount || batchcount > ac->avail);
2223 spin_lock(&cachep->spinlock);
2224 if (cachep->lists.shared) {
2225 struct array_cache *shared_array = cachep->lists.shared;
2226 int max = shared_array->limit-shared_array->avail;
2228 if (batchcount > max)
2230 memcpy(&ac_entry(shared_array)[shared_array->avail],
2232 sizeof(void*)*batchcount);
2233 shared_array->avail += batchcount;
2238 free_block(cachep, &ac_entry(ac)[0], batchcount);
2243 struct list_head *p;
2245 p = list3_data(cachep)->slabs_free.next;
2246 while (p != &(list3_data(cachep)->slabs_free)) {
2249 slabp = list_entry(p, struct slab, list);
2250 BUG_ON(slabp->inuse);
2255 STATS_SET_FREEABLE(cachep, i);
2258 spin_unlock(&cachep->spinlock);
2259 ac->avail -= batchcount;
2260 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2261 sizeof(void*)*ac->avail);
2266 * Release an obj back to its cache. If the obj has a constructed
2267 * state, it must be in this state _before_ it is released.
2269 * Called with disabled ints.
2271 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
2273 struct array_cache *ac = ac_data(cachep);
2276 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2278 if (likely(ac->avail < ac->limit)) {
2279 STATS_INC_FREEHIT(cachep);
2280 ac_entry(ac)[ac->avail++] = objp;
2283 STATS_INC_FREEMISS(cachep);
2284 cache_flusharray(cachep, ac);
2285 ac_entry(ac)[ac->avail++] = objp;
2290 * kmem_cache_alloc - Allocate an object
2291 * @cachep: The cache to allocate from.
2292 * @flags: See kmalloc().
2294 * Allocate an object from this cache. The flags are only relevant
2295 * if the cache has no available objects.
2297 void * kmem_cache_alloc (kmem_cache_t *cachep, int flags)
2299 return __cache_alloc(cachep, flags);
2302 EXPORT_SYMBOL(kmem_cache_alloc);
2305 * kmem_ptr_validate - check if an untrusted pointer might
2307 * @cachep: the cache we're checking against
2308 * @ptr: pointer to validate
2310 * This verifies that the untrusted pointer looks sane:
2311 * it is _not_ a guarantee that the pointer is actually
2312 * part of the slab cache in question, but it at least
2313 * validates that the pointer can be dereferenced and
2314 * looks half-way sane.
2316 * Currently only used for dentry validation.
2318 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2320 unsigned long addr = (unsigned long) ptr;
2321 unsigned long min_addr = PAGE_OFFSET;
2322 unsigned long align_mask = BYTES_PER_WORD-1;
2323 unsigned long size = cachep->objsize;
2326 if (unlikely(addr < min_addr))
2328 if (unlikely(addr > (unsigned long)high_memory - size))
2330 if (unlikely(addr & align_mask))
2332 if (unlikely(!kern_addr_valid(addr)))
2334 if (unlikely(!kern_addr_valid(addr + size - 1)))
2336 page = virt_to_page(ptr);
2337 if (unlikely(!PageSlab(page)))
2339 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2348 * kmem_cache_alloc_node - Allocate an object on the specified node
2349 * @cachep: The cache to allocate from.
2350 * @flags: See kmalloc().
2351 * @nodeid: node number of the target node.
2353 * Identical to kmem_cache_alloc, except that this function is slow
2354 * and can sleep. And it will allocate memory on the given node, which
2355 * can improve the performance for cpu bound structures.
2357 void *kmem_cache_alloc_node(kmem_cache_t *cachep, int nodeid)
2364 for (loop = 0;;loop++) {
2365 struct list_head *q;
2369 spin_lock_irq(&cachep->spinlock);
2370 /* walk through all partial and empty slab and find one
2371 * from the right node */
2372 list_for_each(q,&cachep->lists.slabs_partial) {
2373 slabp = list_entry(q, struct slab, list);
2375 if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
2379 list_for_each(q, &cachep->lists.slabs_free) {
2380 slabp = list_entry(q, struct slab, list);
2382 if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
2386 spin_unlock_irq(&cachep->spinlock);
2388 local_irq_disable();
2389 if (!cache_grow(cachep, GFP_KERNEL, nodeid)) {
2396 /* found one: allocate object */
2397 check_slabp(cachep, slabp);
2398 check_spinlock_acquired(cachep);
2400 STATS_INC_ALLOCED(cachep);
2401 STATS_INC_ACTIVE(cachep);
2402 STATS_SET_HIGH(cachep);
2403 STATS_INC_NODEALLOCS(cachep);
2405 objp = slabp->s_mem + slabp->free*cachep->objsize;
2408 next = slab_bufctl(slabp)[slabp->free];
2410 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2413 check_slabp(cachep, slabp);
2415 /* move slabp to correct slabp list: */
2416 list_del(&slabp->list);
2417 if (slabp->free == BUFCTL_END)
2418 list_add(&slabp->list, &cachep->lists.slabs_full);
2420 list_add(&slabp->list, &cachep->lists.slabs_partial);
2422 list3_data(cachep)->free_objects--;
2423 spin_unlock_irq(&cachep->spinlock);
2425 objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
2426 __builtin_return_address(0));
2429 EXPORT_SYMBOL(kmem_cache_alloc_node);
2434 * kmalloc - allocate memory
2435 * @size: how many bytes of memory are required.
2436 * @flags: the type of memory to allocate.
2438 * kmalloc is the normal method of allocating memory
2441 * The @flags argument may be one of:
2443 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2445 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2447 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2449 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2450 * must be suitable for DMA. This can mean different things on different
2451 * platforms. For example, on i386, it means that the memory must come
2452 * from the first 16MB.
2454 void * __kmalloc (size_t size, int flags)
2456 struct cache_sizes *csizep = malloc_sizes;
2458 for (; csizep->cs_size; csizep++) {
2459 if (size > csizep->cs_size)
2462 /* This happens if someone tries to call
2463 * kmem_cache_create(), or kmalloc(), before
2464 * the generic caches are initialized.
2466 BUG_ON(csizep->cs_cachep == NULL);
2468 return __cache_alloc(flags & GFP_DMA ?
2469 csizep->cs_dmacachep : csizep->cs_cachep, flags);
2474 EXPORT_SYMBOL(__kmalloc);
2478 * __alloc_percpu - allocate one copy of the object for every present
2479 * cpu in the system, zeroing them.
2480 * Objects should be dereferenced using the per_cpu_ptr macro only.
2482 * @size: how many bytes of memory are required.
2483 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2485 void *__alloc_percpu(size_t size, size_t align)
2488 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2493 for (i = 0; i < NR_CPUS; i++) {
2494 if (!cpu_possible(i))
2496 pdata->ptrs[i] = kmem_cache_alloc_node(
2497 kmem_find_general_cachep(size, GFP_KERNEL),
2500 if (!pdata->ptrs[i])
2502 memset(pdata->ptrs[i], 0, size);
2505 /* Catch derefs w/o wrappers */
2506 return (void *) (~(unsigned long) pdata);
2510 if (!cpu_possible(i))
2512 kfree(pdata->ptrs[i]);
2518 EXPORT_SYMBOL(__alloc_percpu);
2522 * kmem_cache_free - Deallocate an object
2523 * @cachep: The cache the allocation was from.
2524 * @objp: The previously allocated object.
2526 * Free an object which was previously allocated from this
2529 void kmem_cache_free (kmem_cache_t *cachep, void *objp)
2531 unsigned long flags;
2533 local_irq_save(flags);
2534 __cache_free(cachep, objp);
2535 local_irq_restore(flags);
2538 EXPORT_SYMBOL(kmem_cache_free);
2541 * kcalloc - allocate memory for an array. The memory is set to zero.
2542 * @n: number of elements.
2543 * @size: element size.
2544 * @flags: the type of memory to allocate.
2546 void *kcalloc(size_t n, size_t size, int flags)
2550 if (n != 0 && size > INT_MAX / n)
2553 ret = kmalloc(n * size, flags);
2555 memset(ret, 0, n * size);
2559 EXPORT_SYMBOL(kcalloc);
2562 * kfree - free previously allocated memory
2563 * @objp: pointer returned by kmalloc.
2565 * Don't free memory not originally allocated by kmalloc()
2566 * or you will run into trouble.
2568 void kfree (const void *objp)
2571 unsigned long flags;
2575 local_irq_save(flags);
2576 kfree_debugcheck(objp);
2577 c = GET_PAGE_CACHE(virt_to_page(objp));
2578 __cache_free(c, (void*)objp);
2579 local_irq_restore(flags);
2582 EXPORT_SYMBOL(kfree);
2586 * free_percpu - free previously allocated percpu memory
2587 * @objp: pointer returned by alloc_percpu.
2589 * Don't free memory not originally allocated by alloc_percpu()
2590 * The complemented objp is to check for that.
2593 free_percpu(const void *objp)
2596 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2598 for (i = 0; i < NR_CPUS; i++) {
2599 if (!cpu_possible(i))
2606 EXPORT_SYMBOL(free_percpu);
2609 unsigned int kmem_cache_size(kmem_cache_t *cachep)
2611 return obj_reallen(cachep);
2614 EXPORT_SYMBOL(kmem_cache_size);
2616 struct ccupdate_struct {
2617 kmem_cache_t *cachep;
2618 struct array_cache *new[NR_CPUS];
2621 static void do_ccupdate_local(void *info)
2623 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2624 struct array_cache *old;
2627 old = ac_data(new->cachep);
2629 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2630 new->new[smp_processor_id()] = old;
2634 static int do_tune_cpucache (kmem_cache_t* cachep, int limit, int batchcount, int shared)
2636 struct ccupdate_struct new;
2637 struct array_cache *new_shared;
2640 memset(&new.new,0,sizeof(new.new));
2641 for (i = 0; i < NR_CPUS; i++) {
2642 if (cpu_online(i)) {
2643 new.new[i] = alloc_arraycache(i, limit, batchcount);
2645 for (i--; i >= 0; i--) kfree(new.new[i]);
2652 new.cachep = cachep;
2654 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2657 spin_lock_irq(&cachep->spinlock);
2658 cachep->batchcount = batchcount;
2659 cachep->limit = limit;
2660 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2661 spin_unlock_irq(&cachep->spinlock);
2663 for (i = 0; i < NR_CPUS; i++) {
2664 struct array_cache *ccold = new.new[i];
2667 spin_lock_irq(&cachep->spinlock);
2668 free_block(cachep, ac_entry(ccold), ccold->avail);
2669 spin_unlock_irq(&cachep->spinlock);
2672 new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
2674 struct array_cache *old;
2676 spin_lock_irq(&cachep->spinlock);
2677 old = cachep->lists.shared;
2678 cachep->lists.shared = new_shared;
2680 free_block(cachep, ac_entry(old), old->avail);
2681 spin_unlock_irq(&cachep->spinlock);
2689 static void enable_cpucache (kmem_cache_t *cachep)
2694 /* The head array serves three purposes:
2695 * - create a LIFO ordering, i.e. return objects that are cache-warm
2696 * - reduce the number of spinlock operations.
2697 * - reduce the number of linked list operations on the slab and
2698 * bufctl chains: array operations are cheaper.
2699 * The numbers are guessed, we should auto-tune as described by
2702 if (cachep->objsize > 131072)
2704 else if (cachep->objsize > PAGE_SIZE)
2706 else if (cachep->objsize > 1024)
2708 else if (cachep->objsize > 256)
2713 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2714 * allocation behaviour: Most allocs on one cpu, most free operations
2715 * on another cpu. For these cases, an efficient object passing between
2716 * cpus is necessary. This is provided by a shared array. The array
2717 * replaces Bonwick's magazine layer.
2718 * On uniprocessor, it's functionally equivalent (but less efficient)
2719 * to a larger limit. Thus disabled by default.
2723 if (cachep->objsize <= PAGE_SIZE)
2728 /* With debugging enabled, large batchcount lead to excessively
2729 * long periods with disabled local interrupts. Limit the
2735 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2737 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2738 cachep->name, -err);
2741 static void drain_array_locked(kmem_cache_t *cachep,
2742 struct array_cache *ac, int force)
2746 check_spinlock_acquired(cachep);
2747 if (ac->touched && !force) {
2749 } else if (ac->avail) {
2750 tofree = force ? ac->avail : (ac->limit+4)/5;
2751 if (tofree > ac->avail) {
2752 tofree = (ac->avail+1)/2;
2754 free_block(cachep, ac_entry(ac), tofree);
2755 ac->avail -= tofree;
2756 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2757 sizeof(void*)*ac->avail);
2762 * cache_reap - Reclaim memory from caches.
2764 * Called from workqueue/eventd every few seconds.
2766 * - clear the per-cpu caches for this CPU.
2767 * - return freeable pages to the main free memory pool.
2769 * If we cannot acquire the cache chain semaphore then just give up - we'll
2770 * try again on the next iteration.
2772 static void cache_reap(void *unused)
2774 struct list_head *walk;
2776 if (down_trylock(&cache_chain_sem)) {
2777 /* Give up. Setup the next iteration. */
2778 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
2782 list_for_each(walk, &cache_chain) {
2783 kmem_cache_t *searchp;
2784 struct list_head* p;
2788 searchp = list_entry(walk, kmem_cache_t, next);
2790 if (searchp->flags & SLAB_NO_REAP)
2795 spin_lock_irq(&searchp->spinlock);
2797 drain_array_locked(searchp, ac_data(searchp), 0);
2799 if(time_after(searchp->lists.next_reap, jiffies))
2802 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2804 if (searchp->lists.shared)
2805 drain_array_locked(searchp, searchp->lists.shared, 0);
2807 if (searchp->lists.free_touched) {
2808 searchp->lists.free_touched = 0;
2812 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2814 p = list3_data(searchp)->slabs_free.next;
2815 if (p == &(list3_data(searchp)->slabs_free))
2818 slabp = list_entry(p, struct slab, list);
2819 BUG_ON(slabp->inuse);
2820 list_del(&slabp->list);
2821 STATS_INC_REAPED(searchp);
2823 /* Safe to drop the lock. The slab is no longer
2824 * linked to the cache.
2825 * searchp cannot disappear, we hold
2828 searchp->lists.free_objects -= searchp->num;
2829 spin_unlock_irq(&searchp->spinlock);
2830 slab_destroy(searchp, slabp);
2831 spin_lock_irq(&searchp->spinlock);
2832 } while(--tofree > 0);
2834 spin_unlock_irq(&searchp->spinlock);
2839 up(&cache_chain_sem);
2840 /* Setup the next iteration */
2841 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
2844 #ifdef CONFIG_PROC_FS
2846 static void *s_start(struct seq_file *m, loff_t *pos)
2849 struct list_head *p;
2851 down(&cache_chain_sem);
2854 * Output format version, so at least we can change it
2855 * without _too_ many complaints.
2858 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
2860 seq_puts(m, "slabinfo - version: 2.1\n");
2862 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2863 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
2864 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2866 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
2867 " <error> <maxfreeable> <freelimit> <nodeallocs>");
2868 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2872 p = cache_chain.next;
2875 if (p == &cache_chain)
2878 return list_entry(p, kmem_cache_t, next);
2881 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2883 kmem_cache_t *cachep = p;
2885 return cachep->next.next == &cache_chain ? NULL
2886 : list_entry(cachep->next.next, kmem_cache_t, next);
2889 static void s_stop(struct seq_file *m, void *p)
2891 up(&cache_chain_sem);
2894 static int s_show(struct seq_file *m, void *p)
2896 kmem_cache_t *cachep = p;
2897 struct list_head *q;
2899 unsigned long active_objs;
2900 unsigned long num_objs;
2901 unsigned long active_slabs = 0;
2902 unsigned long num_slabs;
2907 spin_lock_irq(&cachep->spinlock);
2910 list_for_each(q,&cachep->lists.slabs_full) {
2911 slabp = list_entry(q, struct slab, list);
2912 if (slabp->inuse != cachep->num && !error)
2913 error = "slabs_full accounting error";
2914 active_objs += cachep->num;
2917 list_for_each(q,&cachep->lists.slabs_partial) {
2918 slabp = list_entry(q, struct slab, list);
2919 if (slabp->inuse == cachep->num && !error)
2920 error = "slabs_partial inuse accounting error";
2921 if (!slabp->inuse && !error)
2922 error = "slabs_partial/inuse accounting error";
2923 active_objs += slabp->inuse;
2926 list_for_each(q,&cachep->lists.slabs_free) {
2927 slabp = list_entry(q, struct slab, list);
2928 if (slabp->inuse && !error)
2929 error = "slabs_free/inuse accounting error";
2932 num_slabs+=active_slabs;
2933 num_objs = num_slabs*cachep->num;
2934 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2935 error = "free_objects accounting error";
2937 name = cachep->name;
2939 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2941 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2942 name, active_objs, num_objs, cachep->objsize,
2943 cachep->num, (1<<cachep->gfporder));
2944 seq_printf(m, " : tunables %4u %4u %4u",
2945 cachep->limit, cachep->batchcount,
2946 cachep->lists.shared->limit/cachep->batchcount);
2947 seq_printf(m, " : slabdata %6lu %6lu %6u",
2948 active_slabs, num_slabs, cachep->lists.shared->avail);
2951 unsigned long high = cachep->high_mark;
2952 unsigned long allocs = cachep->num_allocations;
2953 unsigned long grown = cachep->grown;
2954 unsigned long reaped = cachep->reaped;
2955 unsigned long errors = cachep->errors;
2956 unsigned long max_freeable = cachep->max_freeable;
2957 unsigned long free_limit = cachep->free_limit;
2958 unsigned long node_allocs = cachep->node_allocs;
2960 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu",
2961 allocs, high, grown, reaped, errors,
2962 max_freeable, free_limit, node_allocs);
2966 unsigned long allochit = atomic_read(&cachep->allochit);
2967 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2968 unsigned long freehit = atomic_read(&cachep->freehit);
2969 unsigned long freemiss = atomic_read(&cachep->freemiss);
2971 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2972 allochit, allocmiss, freehit, freemiss);
2976 spin_unlock_irq(&cachep->spinlock);
2981 * slabinfo_op - iterator that generates /proc/slabinfo
2990 * num-pages-per-slab
2991 * + further values on SMP and with statistics enabled
2994 struct seq_operations slabinfo_op = {
3001 #define MAX_SLABINFO_WRITE 128
3003 * slabinfo_write - Tuning for the slab allocator
3005 * @buffer: user buffer
3006 * @count: data length
3009 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3010 size_t count, loff_t *ppos)
3012 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3013 int limit, batchcount, shared, res;
3014 struct list_head *p;
3016 if (count > MAX_SLABINFO_WRITE)
3018 if (copy_from_user(&kbuf, buffer, count))
3020 kbuf[MAX_SLABINFO_WRITE] = '\0';
3022 tmp = strchr(kbuf, ' ');
3027 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3030 /* Find the cache in the chain of caches. */
3031 down(&cache_chain_sem);
3033 list_for_each(p,&cache_chain) {
3034 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3036 if (!strcmp(cachep->name, kbuf)) {
3039 batchcount > limit ||
3043 res = do_tune_cpucache(cachep, limit, batchcount, shared);
3048 up(&cache_chain_sem);
3055 unsigned int ksize(const void *objp)
3058 unsigned long flags;
3059 unsigned int size = 0;
3061 if (likely(objp != NULL)) {
3062 local_irq_save(flags);
3063 c = GET_PAGE_CACHE(virt_to_page(objp));
3064 size = kmem_cache_size(c);
3065 local_irq_restore(flags);