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>
95 #include <asm/uaccess.h>
96 #include <asm/cacheflush.h>
97 #include <asm/tlbflush.h>
100 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
101 * SLAB_RED_ZONE & SLAB_POISON.
102 * 0 for faster, smaller code (especially in the critical paths).
104 * STATS - 1 to collect stats for /proc/slabinfo.
105 * 0 for faster, smaller code (especially in the critical paths).
107 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
110 #ifdef CONFIG_DEBUG_SLAB
113 #define FORCED_DEBUG 1
117 #define FORCED_DEBUG 0
121 /* Shouldn't this be in a header file somewhere? */
122 #define BYTES_PER_WORD sizeof(void *)
124 #ifndef cache_line_size
125 #define cache_line_size() L1_CACHE_BYTES
128 #ifndef ARCH_KMALLOC_MINALIGN
129 #define ARCH_KMALLOC_MINALIGN 0
132 /* Legal flag mask for kmem_cache_create(). */
134 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
135 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
136 SLAB_NO_REAP | SLAB_CACHE_DMA | \
137 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
138 SLAB_RECLAIM_ACCOUNT )
140 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
141 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
142 SLAB_RECLAIM_ACCOUNT)
148 * Bufctl's are used for linking objs within a slab
151 * This implementation relies on "struct page" for locating the cache &
152 * slab an object belongs to.
153 * This allows the bufctl structure to be small (one int), but limits
154 * the number of objects a slab (not a cache) can contain when off-slab
155 * bufctls are used. The limit is the size of the largest general cache
156 * that does not use off-slab slabs.
157 * For 32bit archs with 4 kB pages, is this 56.
158 * This is not serious, as it is only for large objects, when it is unwise
159 * to have too many per slab.
160 * Note: This limit can be raised by introducing a general cache whose size
161 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
164 #define BUFCTL_END 0xffffFFFF
165 #define BUFCTL_FREE 0xffffFFFE
166 #define SLAB_LIMIT 0xffffFFFD
167 typedef unsigned int kmem_bufctl_t;
169 /* Max number of objs-per-slab for caches which use off-slab slabs.
170 * Needed to avoid a possible looping condition in cache_grow().
172 static unsigned long offslab_limit;
177 * Manages the objs in a slab. Placed either at the beginning of mem allocated
178 * for a slab, or allocated from an general cache.
179 * Slabs are chained into three list: fully used, partial, fully free slabs.
182 struct list_head list;
183 unsigned long colouroff;
184 void *s_mem; /* including colour offset */
185 unsigned int inuse; /* num of objs active in slab */
194 * - LIFO ordering, to hand out cache-warm objects from _alloc
195 * - reduce the number of linked list operations
196 * - reduce spinlock operations
198 * The limit is stored in the per-cpu structure to reduce the data cache
205 unsigned int batchcount;
206 unsigned int touched;
209 /* bootstrap: The caches do not work without cpuarrays anymore,
210 * but the cpuarrays are allocated from the generic caches...
212 #define BOOT_CPUCACHE_ENTRIES 1
213 struct arraycache_init {
214 struct array_cache cache;
215 void * entries[BOOT_CPUCACHE_ENTRIES];
219 * The slab lists of all objects.
220 * Hopefully reduce the internal fragmentation
221 * NUMA: The spinlock could be moved from the kmem_cache_t
222 * into this structure, too. Figure out what causes
223 * fewer cross-node spinlock operations.
226 struct list_head slabs_partial; /* partial list first, better asm code */
227 struct list_head slabs_full;
228 struct list_head slabs_free;
229 unsigned long free_objects;
231 unsigned long next_reap;
232 struct array_cache *shared;
235 #define LIST3_INIT(parent) \
237 .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \
238 .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \
239 .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \
241 #define list3_data(cachep) \
245 #define list3_data_ptr(cachep, ptr) \
254 struct kmem_cache_s {
255 /* 1) per-cpu data, touched during every alloc/free */
256 struct array_cache *array[NR_CPUS];
257 unsigned int batchcount;
259 /* 2) touched by every alloc & free from the backend */
260 struct kmem_list3 lists;
261 /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */
262 unsigned int objsize;
263 unsigned int flags; /* constant flags */
264 unsigned int num; /* # of objs per slab */
265 unsigned int free_limit; /* upper limit of objects in the lists */
268 /* 3) cache_grow/shrink */
269 /* order of pgs per slab (2^n) */
270 unsigned int gfporder;
272 /* force GFP flags, e.g. GFP_DMA */
273 unsigned int gfpflags;
275 size_t colour; /* cache colouring range */
276 unsigned int colour_off; /* colour offset */
277 unsigned int colour_next; /* cache colouring */
278 kmem_cache_t *slabp_cache;
279 unsigned int slab_size;
280 unsigned int dflags; /* dynamic flags */
282 /* constructor func */
283 void (*ctor)(void *, kmem_cache_t *, unsigned long);
285 /* de-constructor func */
286 void (*dtor)(void *, kmem_cache_t *, unsigned long);
288 /* 4) cache creation/removal */
290 struct list_head next;
294 unsigned long num_active;
295 unsigned long num_allocations;
296 unsigned long high_mark;
298 unsigned long reaped;
299 unsigned long errors;
300 unsigned long max_freeable;
312 #define CFLGS_OFF_SLAB (0x80000000UL)
313 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
315 #define BATCHREFILL_LIMIT 16
316 /* Optimization question: fewer reaps means less
317 * probability for unnessary cpucache drain/refill cycles.
319 * OTHO the cpuarrays can contain lots of objects,
320 * which could lock up otherwise freeable slabs.
322 #define REAPTIMEOUT_CPUC (2*HZ)
323 #define REAPTIMEOUT_LIST3 (4*HZ)
326 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
327 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
328 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
329 #define STATS_INC_GROWN(x) ((x)->grown++)
330 #define STATS_INC_REAPED(x) ((x)->reaped++)
331 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
332 (x)->high_mark = (x)->num_active; \
334 #define STATS_INC_ERR(x) ((x)->errors++)
335 #define STATS_SET_FREEABLE(x, i) \
336 do { if ((x)->max_freeable < i) \
337 (x)->max_freeable = i; \
340 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
341 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
342 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
343 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
345 #define STATS_INC_ACTIVE(x) do { } while (0)
346 #define STATS_DEC_ACTIVE(x) do { } while (0)
347 #define STATS_INC_ALLOCED(x) do { } while (0)
348 #define STATS_INC_GROWN(x) do { } while (0)
349 #define STATS_INC_REAPED(x) do { } while (0)
350 #define STATS_SET_HIGH(x) do { } while (0)
351 #define STATS_INC_ERR(x) do { } while (0)
352 #define STATS_SET_FREEABLE(x, i) \
355 #define STATS_INC_ALLOCHIT(x) do { } while (0)
356 #define STATS_INC_ALLOCMISS(x) do { } while (0)
357 #define STATS_INC_FREEHIT(x) do { } while (0)
358 #define STATS_INC_FREEMISS(x) do { } while (0)
362 /* Magic nums for obj red zoning.
363 * Placed in the first word before and the first word after an obj.
365 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
366 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
368 /* ...and for poisoning */
369 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
370 #define POISON_FREE 0x6b /* for use-after-free poisoning */
371 #define POISON_END 0xa5 /* end-byte of poisoning */
373 /* memory layout of objects:
375 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
376 * the end of an object is aligned with the end of the real
377 * allocation. Catches writes behind the end of the allocation.
378 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
380 * cachep->dbghead: The real object.
381 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
382 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
384 static inline int obj_dbghead(kmem_cache_t *cachep)
386 return cachep->dbghead;
389 static inline int obj_reallen(kmem_cache_t *cachep)
391 return cachep->reallen;
394 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
396 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
397 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
400 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
402 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
403 if (cachep->flags & SLAB_STORE_USER)
404 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
405 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
408 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
410 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
411 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
414 static inline int obj_dbghead(kmem_cache_t *cachep)
418 static inline int obj_reallen(kmem_cache_t *cachep)
420 return cachep->objsize;
422 static inline unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
427 static inline unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
432 static inline void **dbg_userword(kmem_cache_t *cachep, void *objp)
440 * Maximum size of an obj (in 2^order pages)
441 * and absolute limit for the gfp order.
443 #if defined(CONFIG_LARGE_ALLOCS)
444 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
445 #define MAX_GFP_ORDER 13 /* up to 32Mb */
446 #elif defined(CONFIG_MMU)
447 #define MAX_OBJ_ORDER 5 /* 32 pages */
448 #define MAX_GFP_ORDER 5 /* 32 pages */
450 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
451 #define MAX_GFP_ORDER 8 /* up to 1Mb */
455 * Do not go above this order unless 0 objects fit into the slab.
457 #define BREAK_GFP_ORDER_HI 1
458 #define BREAK_GFP_ORDER_LO 0
459 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
461 /* Macros for storing/retrieving the cachep and or slab from the
462 * global 'mem_map'. These are used to find the slab an obj belongs to.
463 * With kfree(), these are used to find the cache which an obj belongs to.
465 #define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
466 #define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
467 #define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
468 #define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
470 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
471 struct cache_sizes malloc_sizes[] = {
472 #define CACHE(x) { .cs_size = (x) },
473 #include <linux/kmalloc_sizes.h>
478 EXPORT_SYMBOL(malloc_sizes);
480 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
481 static struct cache_names {
485 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
486 #include <linux/kmalloc_sizes.h>
491 struct arraycache_init initarray_cache __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
492 struct arraycache_init initarray_generic __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
494 /* internal cache of cache description objs */
495 static kmem_cache_t cache_cache = {
496 .lists = LIST3_INIT(cache_cache.lists),
498 .limit = BOOT_CPUCACHE_ENTRIES,
499 .objsize = sizeof(kmem_cache_t),
500 .flags = SLAB_NO_REAP,
501 .spinlock = SPIN_LOCK_UNLOCKED,
502 .name = "kmem_cache",
504 .reallen = sizeof(kmem_cache_t),
508 /* Guard access to the cache-chain. */
509 static struct semaphore cache_chain_sem;
511 struct list_head cache_chain;
514 * vm_enough_memory() looks at this to determine how many
515 * slab-allocated pages are possibly freeable under pressure
517 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
519 atomic_t slab_reclaim_pages;
520 EXPORT_SYMBOL(slab_reclaim_pages);
523 * chicken and egg problem: delay the per-cpu array allocation
524 * until the general caches are up.
532 static DEFINE_PER_CPU(struct timer_list, reap_timers);
534 static void reap_timer_fnc(unsigned long data);
535 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
536 static void enable_cpucache (kmem_cache_t *cachep);
538 static inline void ** ac_entry(struct array_cache *ac)
540 return (void**)(ac+1);
543 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
545 return cachep->array[smp_processor_id()];
548 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
549 static void cache_estimate (unsigned long gfporder, size_t size, size_t align,
550 int flags, size_t *left_over, unsigned int *num)
553 size_t wastage = PAGE_SIZE<<gfporder;
557 if (!(flags & CFLGS_OFF_SLAB)) {
558 base = sizeof(struct slab);
559 extra = sizeof(kmem_bufctl_t);
562 while (i*size + ALIGN(base+i*extra, align) <= wastage)
572 wastage -= ALIGN(base+i*extra, align);
573 *left_over = wastage;
576 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
578 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
580 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
581 function, cachep->name, msg);
586 * Start the reap timer running on the target CPU. We run at around 1 to 2Hz.
587 * Add the CPU number into the expiry time to minimize the possibility of the
588 * CPUs getting into lockstep and contending for the global cache chain lock.
590 static void __devinit start_cpu_timer(int cpu)
592 struct timer_list *rt = &per_cpu(reap_timers, cpu);
594 if (rt->function == NULL) {
596 rt->expires = jiffies + HZ + 3*cpu;
598 rt->function = reap_timer_fnc;
599 add_timer_on(rt, cpu);
603 #ifdef CONFIG_HOTPLUG_CPU
604 static void stop_cpu_timer(int cpu)
606 struct timer_list *rt = &per_cpu(reap_timers, cpu);
610 WARN_ON(timer_pending(rt));
616 static int __devinit cpuup_callback(struct notifier_block *nfb,
617 unsigned long action,
620 long cpu = (long)hcpu;
621 kmem_cache_t* cachep;
625 down(&cache_chain_sem);
626 list_for_each_entry(cachep, &cache_chain, next) {
628 struct array_cache *nc;
630 memsize = sizeof(void*)*cachep->limit+sizeof(struct array_cache);
631 nc = kmalloc(memsize, GFP_KERNEL);
635 nc->limit = cachep->limit;
636 nc->batchcount = cachep->batchcount;
639 spin_lock_irq(&cachep->spinlock);
640 cachep->array[cpu] = nc;
641 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
643 spin_unlock_irq(&cachep->spinlock);
646 up(&cache_chain_sem);
649 start_cpu_timer(cpu);
651 #ifdef CONFIG_HOTPLUG_CPU
655 case CPU_UP_CANCELED:
656 down(&cache_chain_sem);
658 list_for_each_entry(cachep, &cache_chain, next) {
659 struct array_cache *nc;
661 spin_lock_irq(&cachep->spinlock);
662 /* cpu is dead; no one can alloc from it. */
663 nc = cachep->array[cpu];
664 cachep->array[cpu] = NULL;
665 cachep->free_limit -= cachep->batchcount;
666 free_block(cachep, ac_entry(nc), nc->avail);
667 spin_unlock_irq(&cachep->spinlock);
670 up(&cache_chain_sem);
676 up(&cache_chain_sem);
680 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
683 * Called after the gfp() functions have been enabled, and before smp_init().
685 void __init kmem_cache_init(void)
688 struct cache_sizes *sizes;
689 struct cache_names *names;
692 * Fragmentation resistance on low memory - only use bigger
693 * page orders on machines with more than 32MB of memory.
695 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
696 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
699 /* Bootstrap is tricky, because several objects are allocated
700 * from caches that do not exist yet:
701 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
702 * structures of all caches, except cache_cache itself: cache_cache
703 * is statically allocated.
704 * Initially an __init data area is used for the head array, it's
705 * replaced with a kmalloc allocated array at the end of the bootstrap.
706 * 2) Create the first kmalloc cache.
707 * The kmem_cache_t for the new cache is allocated normally. An __init
708 * data area is used for the head array.
709 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
710 * 4) Replace the __init data head arrays for cache_cache and the first
711 * kmalloc cache with kmalloc allocated arrays.
712 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
715 /* 1) create the cache_cache */
716 init_MUTEX(&cache_chain_sem);
717 INIT_LIST_HEAD(&cache_chain);
718 list_add(&cache_cache.next, &cache_chain);
719 cache_cache.colour_off = cache_line_size();
720 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
722 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
724 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
725 &left_over, &cache_cache.num);
726 if (!cache_cache.num)
729 cache_cache.colour = left_over/cache_cache.colour_off;
730 cache_cache.colour_next = 0;
731 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
732 sizeof(struct slab), cache_line_size());
734 /* 2+3) create the kmalloc caches */
735 sizes = malloc_sizes;
738 while (sizes->cs_size) {
739 /* For performance, all the general caches are L1 aligned.
740 * This should be particularly beneficial on SMP boxes, as it
741 * eliminates "false sharing".
742 * Note for systems short on memory removing the alignment will
743 * allow tighter packing of the smaller caches. */
744 sizes->cs_cachep = kmem_cache_create(
745 names->name, sizes->cs_size,
746 ARCH_KMALLOC_MINALIGN, 0, NULL, NULL);
747 if (!sizes->cs_cachep)
750 /* Inc off-slab bufctl limit until the ceiling is hit. */
751 if (!(OFF_SLAB(sizes->cs_cachep))) {
752 offslab_limit = sizes->cs_size-sizeof(struct slab);
753 offslab_limit /= sizeof(kmem_bufctl_t);
756 sizes->cs_dmacachep = kmem_cache_create(
757 names->name_dma, sizes->cs_size,
758 ARCH_KMALLOC_MINALIGN, SLAB_CACHE_DMA, NULL, NULL);
759 if (!sizes->cs_dmacachep)
765 /* 4) Replace the bootstrap head arrays */
769 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
771 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
772 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
773 cache_cache.array[smp_processor_id()] = ptr;
776 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
778 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
779 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
780 sizeof(struct arraycache_init));
781 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
785 /* 5) resize the head arrays to their final sizes */
787 kmem_cache_t *cachep;
788 down(&cache_chain_sem);
789 list_for_each_entry(cachep, &cache_chain, next)
790 enable_cpucache(cachep);
791 up(&cache_chain_sem);
795 g_cpucache_up = FULL;
797 /* Register a cpu startup notifier callback
798 * that initializes ac_data for all new cpus
800 register_cpu_notifier(&cpucache_notifier);
803 /* The reap timers are started later, with a module init call:
804 * That part of the kernel is not yet operational.
808 int __init cpucache_init(void)
813 * Register the timers that return unneeded
816 for (cpu = 0; cpu < NR_CPUS; cpu++) {
818 start_cpu_timer(cpu);
824 __initcall(cpucache_init);
827 * Interface to system's page allocator. No need to hold the cache-lock.
829 * If we requested dmaable memory, we will get it. Even if we
830 * did not request dmaable memory, we might get it, but that
831 * would be relatively rare and ignorable.
833 static inline void *kmem_getpages(kmem_cache_t *cachep, unsigned long flags)
837 flags |= cachep->gfpflags;
838 addr = (void*)__get_free_pages(flags, cachep->gfporder);
840 int i = (1 << cachep->gfporder);
841 struct page *page = virt_to_page(addr);
843 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
844 atomic_add(i, &slab_reclaim_pages);
845 add_page_state(nr_slab, i);
855 * Interface to system's page release.
857 static inline void kmem_freepages(kmem_cache_t *cachep, void *addr)
859 unsigned long i = (1<<cachep->gfporder);
860 struct page *page = virt_to_page(addr);
861 const unsigned long nr_freed = i;
864 if (!TestClearPageSlab(page))
868 sub_page_state(nr_slab, nr_freed);
869 if (current->reclaim_state)
870 current->reclaim_state->reclaimed_slab += nr_freed;
871 free_pages((unsigned long)addr, cachep->gfporder);
872 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
873 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
878 #ifdef CONFIG_DEBUG_PAGEALLOC
879 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, unsigned long caller)
881 int size = obj_reallen(cachep);
883 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
885 if (size < 5*sizeof(unsigned long))
890 *addr++=smp_processor_id();
891 size -= 3*sizeof(unsigned long);
893 unsigned long *sptr = &caller;
894 unsigned long svalue;
896 while (!kstack_end(sptr)) {
898 if (kernel_text_address(svalue)) {
900 size -= sizeof(unsigned long);
901 if (size <= sizeof(unsigned long))
911 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
913 int size = obj_reallen(cachep);
914 addr = &((char*)addr)[obj_dbghead(cachep)];
916 memset(addr, val, size);
917 *(unsigned char *)(addr+size-1) = POISON_END;
920 static void dump_line(char *data, int offset, int limit)
923 printk(KERN_ERR "%03x:", offset);
924 for (i=0;i<limit;i++) {
925 printk(" %02x", (unsigned char)data[offset+i]);
931 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
937 if (cachep->flags & SLAB_RED_ZONE) {
938 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
939 *dbg_redzone1(cachep, objp),
940 *dbg_redzone2(cachep, objp));
943 if (cachep->flags & SLAB_STORE_USER) {
944 printk(KERN_ERR "Last user: [<%p>]", *dbg_userword(cachep, objp));
945 print_symbol("(%s)", (unsigned long)*dbg_userword(cachep, objp));
948 realobj = (char*)objp+obj_dbghead(cachep);
949 size = obj_reallen(cachep);
950 for (i=0; i<size && lines;i+=16, lines--) {
955 dump_line(realobj, i, limit);
962 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
968 realobj = (char*)objp+obj_dbghead(cachep);
969 size = obj_reallen(cachep);
971 for (i=0;i<size;i++) {
972 char exp = POISON_FREE;
975 if (realobj[i] != exp) {
980 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
982 print_objinfo(cachep, objp, 0);
984 /* Hexdump the affected line */
989 dump_line(realobj, i, limit);
992 /* Limit to 5 lines */
998 /* Print some data about the neighboring objects, if they
1001 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1004 objnr = (objp-slabp->s_mem)/cachep->objsize;
1006 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1007 realobj = (char*)objp+obj_dbghead(cachep);
1008 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1010 print_objinfo(cachep, objp, 2);
1012 if (objnr+1 < cachep->num) {
1013 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1014 realobj = (char*)objp+obj_dbghead(cachep);
1015 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1017 print_objinfo(cachep, objp, 2);
1023 /* Destroy all the objs in a slab, and release the mem back to the system.
1024 * Before calling the slab must have been unlinked from the cache.
1025 * The cache-lock is not held/needed.
1027 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1031 for (i = 0; i < cachep->num; i++) {
1032 void *objp = slabp->s_mem + cachep->objsize * i;
1034 if (cachep->flags & SLAB_POISON) {
1035 #ifdef CONFIG_DEBUG_PAGEALLOC
1036 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1037 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1039 check_poison_obj(cachep, objp);
1041 check_poison_obj(cachep, objp);
1044 if (cachep->flags & SLAB_RED_ZONE) {
1045 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1046 slab_error(cachep, "start of a freed object "
1048 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1049 slab_error(cachep, "end of a freed object "
1052 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1053 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1058 for (i = 0; i < cachep->num; i++) {
1059 void* objp = slabp->s_mem+cachep->objsize*i;
1060 (cachep->dtor)(objp, cachep, 0);
1065 kmem_freepages(cachep, slabp->s_mem-slabp->colouroff);
1066 if (OFF_SLAB(cachep))
1067 kmem_cache_free(cachep->slabp_cache, slabp);
1071 * kmem_cache_create - Create a cache.
1072 * @name: A string which is used in /proc/slabinfo to identify this cache.
1073 * @size: The size of objects to be created in this cache.
1074 * @align: The required alignment for the objects.
1075 * @flags: SLAB flags
1076 * @ctor: A constructor for the objects.
1077 * @dtor: A destructor for the objects.
1079 * Returns a ptr to the cache on success, NULL on failure.
1080 * Cannot be called within a int, but can be interrupted.
1081 * The @ctor is run when new pages are allocated by the cache
1082 * and the @dtor is run before the pages are handed back.
1084 * @name must be valid until the cache is destroyed. This implies that
1085 * the module calling this has to destroy the cache before getting
1090 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1091 * to catch references to uninitialised memory.
1093 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1094 * for buffer overruns.
1096 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1099 * %SLAB_HWCACHE_ALIGN - This flag has no effect and will be removed soon.
1103 kmem_cache_create (const char *name, size_t size, size_t align,
1104 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1105 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1107 size_t left_over, slab_size;
1108 kmem_cache_t *cachep = NULL;
1111 * Sanity checks... these are all serious usage bugs.
1115 (size < BYTES_PER_WORD) ||
1116 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1119 printk(KERN_ERR "%s: Early error in slab %s\n",
1120 __FUNCTION__, name);
1125 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1126 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1127 /* No constructor, but inital state check requested */
1128 printk(KERN_ERR "%s: No con, but init state check "
1129 "requested - %s\n", __FUNCTION__, name);
1130 flags &= ~SLAB_DEBUG_INITIAL;
1135 * Enable redzoning and last user accounting, except for caches with
1136 * large objects, if the increased size would increase the object size
1137 * above the next power of two: caches with object sizes just above a
1138 * power of two have a significant amount of internal fragmentation.
1140 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1141 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1142 flags |= SLAB_POISON;
1146 * Always checks flags, a caller might be expecting debug
1147 * support which isn't available.
1149 if (flags & ~CREATE_MASK)
1153 /* combinations of forced alignment and advanced debugging is
1154 * not yet implemented.
1156 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1158 if (flags & SLAB_HWCACHE_ALIGN) {
1159 /* Default alignment: as specified by the arch code.
1160 * Except if an object is really small, then squeeze multiple
1161 * into one cacheline.
1163 align = cache_line_size();
1164 while (size <= align/2)
1167 align = BYTES_PER_WORD;
1171 /* Get cache's description obj. */
1172 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1175 memset(cachep, 0, sizeof(kmem_cache_t));
1177 /* Check that size is in terms of words. This is needed to avoid
1178 * unaligned accesses for some archs when redzoning is used, and makes
1179 * sure any on-slab bufctl's are also correctly aligned.
1181 if (size & (BYTES_PER_WORD-1)) {
1182 size += (BYTES_PER_WORD-1);
1183 size &= ~(BYTES_PER_WORD-1);
1187 cachep->reallen = size;
1189 if (flags & SLAB_RED_ZONE) {
1190 /* redzoning only works with word aligned caches */
1191 align = BYTES_PER_WORD;
1193 /* add space for red zone words */
1194 cachep->dbghead += BYTES_PER_WORD;
1195 size += 2*BYTES_PER_WORD;
1197 if (flags & SLAB_STORE_USER) {
1198 /* user store requires word alignment and
1199 * one word storage behind the end of the real
1202 align = BYTES_PER_WORD;
1203 size += BYTES_PER_WORD;
1205 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1206 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1207 cachep->dbghead += PAGE_SIZE - size;
1213 /* Determine if the slab management is 'on' or 'off' slab. */
1214 if (size >= (PAGE_SIZE>>3))
1216 * Size is large, assume best to place the slab management obj
1217 * off-slab (should allow better packing of objs).
1219 flags |= CFLGS_OFF_SLAB;
1221 size = ALIGN(size, align);
1223 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1225 * A VFS-reclaimable slab tends to have most allocations
1226 * as GFP_NOFS and we really don't want to have to be allocating
1227 * higher-order pages when we are unable to shrink dcache.
1229 cachep->gfporder = 0;
1230 cache_estimate(cachep->gfporder, size, align, flags,
1231 &left_over, &cachep->num);
1234 * Calculate size (in pages) of slabs, and the num of objs per
1235 * slab. This could be made much more intelligent. For now,
1236 * try to avoid using high page-orders for slabs. When the
1237 * gfp() funcs are more friendly towards high-order requests,
1238 * this should be changed.
1241 unsigned int break_flag = 0;
1243 cache_estimate(cachep->gfporder, size, align, flags,
1244 &left_over, &cachep->num);
1247 if (cachep->gfporder >= MAX_GFP_ORDER)
1251 if (flags & CFLGS_OFF_SLAB &&
1252 cachep->num > offslab_limit) {
1253 /* This num of objs will cause problems. */
1260 * Large num of objs is good, but v. large slabs are
1261 * currently bad for the gfp()s.
1263 if (cachep->gfporder >= slab_break_gfp_order)
1266 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1267 break; /* Acceptable internal fragmentation. */
1274 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1275 kmem_cache_free(&cache_cache, cachep);
1279 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1280 + sizeof(struct slab), align);
1283 * If the slab has been placed off-slab, and we have enough space then
1284 * move it on-slab. This is at the expense of any extra colouring.
1286 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1287 flags &= ~CFLGS_OFF_SLAB;
1288 left_over -= slab_size;
1291 if (flags & CFLGS_OFF_SLAB) {
1292 /* really off slab. No need for manual alignment */
1293 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1296 cachep->colour_off = cache_line_size();
1297 /* Offset must be a multiple of the alignment. */
1298 if (cachep->colour_off < align)
1299 cachep->colour_off = align;
1300 cachep->colour = left_over/cachep->colour_off;
1301 cachep->slab_size = slab_size;
1302 cachep->flags = flags;
1303 cachep->gfpflags = 0;
1304 if (flags & SLAB_CACHE_DMA)
1305 cachep->gfpflags |= GFP_DMA;
1306 spin_lock_init(&cachep->spinlock);
1307 cachep->objsize = size;
1309 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1310 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1311 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1313 if (flags & CFLGS_OFF_SLAB)
1314 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1315 cachep->ctor = ctor;
1316 cachep->dtor = dtor;
1317 cachep->name = name;
1319 /* Don't let CPUs to come and go */
1322 if (g_cpucache_up == FULL) {
1323 enable_cpucache(cachep);
1325 if (g_cpucache_up == NONE) {
1326 /* Note: the first kmem_cache_create must create
1327 * the cache that's used by kmalloc(24), otherwise
1328 * the creation of further caches will BUG().
1330 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1331 g_cpucache_up = PARTIAL;
1333 cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL);
1335 BUG_ON(!ac_data(cachep));
1336 ac_data(cachep)->avail = 0;
1337 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1338 ac_data(cachep)->batchcount = 1;
1339 ac_data(cachep)->touched = 0;
1340 cachep->batchcount = 1;
1341 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1342 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1346 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1347 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1349 /* Need the semaphore to access the chain. */
1350 down(&cache_chain_sem);
1352 struct list_head *p;
1353 mm_segment_t old_fs;
1357 list_for_each(p, &cache_chain) {
1358 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1360 /* This happens when the module gets unloaded and doesn't
1361 destroy its slab cache and noone else reuses the vmalloc
1362 area of the module. Print a warning. */
1363 if (__get_user(tmp,pc->name)) {
1364 printk("SLAB: cache with size %d has lost its name\n",
1368 if (!strcmp(pc->name,name)) {
1369 printk("kmem_cache_create: duplicate cache %s\n",name);
1370 up(&cache_chain_sem);
1371 unlock_cpu_hotplug();
1378 /* cache setup completed, link it into the list */
1379 list_add(&cachep->next, &cache_chain);
1380 up(&cache_chain_sem);
1381 unlock_cpu_hotplug();
1386 EXPORT_SYMBOL(kmem_cache_create);
1388 static inline void check_irq_off(void)
1391 BUG_ON(!irqs_disabled());
1395 static inline void check_irq_on(void)
1398 BUG_ON(irqs_disabled());
1402 static inline void check_spinlock_acquired(kmem_cache_t *cachep)
1406 BUG_ON(spin_trylock(&cachep->spinlock));
1411 * Waits for all CPUs to execute func().
1413 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1418 local_irq_disable();
1422 if (smp_call_function(func, arg, 1, 1))
1428 static void drain_array_locked(kmem_cache_t* cachep,
1429 struct array_cache *ac, int force);
1431 static void do_drain(void *arg)
1433 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1434 struct array_cache *ac;
1437 ac = ac_data(cachep);
1438 spin_lock(&cachep->spinlock);
1439 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1440 spin_unlock(&cachep->spinlock);
1444 static void drain_cpu_caches(kmem_cache_t *cachep)
1446 smp_call_function_all_cpus(do_drain, cachep);
1448 spin_lock_irq(&cachep->spinlock);
1449 if (cachep->lists.shared)
1450 drain_array_locked(cachep, cachep->lists.shared, 1);
1451 spin_unlock_irq(&cachep->spinlock);
1455 /* NUMA shrink all list3s */
1456 static int __cache_shrink(kmem_cache_t *cachep)
1461 drain_cpu_caches(cachep);
1464 spin_lock_irq(&cachep->spinlock);
1467 struct list_head *p;
1469 p = cachep->lists.slabs_free.prev;
1470 if (p == &cachep->lists.slabs_free)
1473 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1478 list_del(&slabp->list);
1480 cachep->lists.free_objects -= cachep->num;
1481 spin_unlock_irq(&cachep->spinlock);
1482 slab_destroy(cachep, slabp);
1483 spin_lock_irq(&cachep->spinlock);
1485 ret = !list_empty(&cachep->lists.slabs_full) ||
1486 !list_empty(&cachep->lists.slabs_partial);
1487 spin_unlock_irq(&cachep->spinlock);
1492 * kmem_cache_shrink - Shrink a cache.
1493 * @cachep: The cache to shrink.
1495 * Releases as many slabs as possible for a cache.
1496 * To help debugging, a zero exit status indicates all slabs were released.
1498 int kmem_cache_shrink(kmem_cache_t *cachep)
1500 if (!cachep || in_interrupt())
1503 return __cache_shrink(cachep);
1506 EXPORT_SYMBOL(kmem_cache_shrink);
1509 * kmem_cache_destroy - delete a cache
1510 * @cachep: the cache to destroy
1512 * Remove a kmem_cache_t object from the slab cache.
1513 * Returns 0 on success.
1515 * It is expected this function will be called by a module when it is
1516 * unloaded. This will remove the cache completely, and avoid a duplicate
1517 * cache being allocated each time a module is loaded and unloaded, if the
1518 * module doesn't have persistent in-kernel storage across loads and unloads.
1520 * The cache must be empty before calling this function.
1522 * The caller must guarantee that noone will allocate memory from the cache
1523 * during the kmem_cache_destroy().
1525 int kmem_cache_destroy (kmem_cache_t * cachep)
1529 if (!cachep || in_interrupt())
1532 /* Don't let CPUs to come and go */
1535 /* Find the cache in the chain of caches. */
1536 down(&cache_chain_sem);
1538 * the chain is never empty, cache_cache is never destroyed
1540 list_del(&cachep->next);
1541 up(&cache_chain_sem);
1543 if (__cache_shrink(cachep)) {
1544 slab_error(cachep, "Can't free all objects");
1545 down(&cache_chain_sem);
1546 list_add(&cachep->next,&cache_chain);
1547 up(&cache_chain_sem);
1548 unlock_cpu_hotplug();
1552 /* no cpu_online check required here since we clear the percpu
1553 * array on cpu offline and set this to NULL.
1555 for (i = 0; i < NR_CPUS; i++)
1556 kfree(cachep->array[i]);
1558 /* NUMA: free the list3 structures */
1559 kfree(cachep->lists.shared);
1560 cachep->lists.shared = NULL;
1561 kmem_cache_free(&cache_cache, cachep);
1563 unlock_cpu_hotplug();
1568 EXPORT_SYMBOL(kmem_cache_destroy);
1570 /* Get the memory for a slab management obj. */
1571 static inline struct slab* alloc_slabmgmt (kmem_cache_t *cachep,
1572 void *objp, int colour_off, int local_flags)
1576 if (OFF_SLAB(cachep)) {
1577 /* Slab management obj is off-slab. */
1578 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1582 slabp = objp+colour_off;
1583 colour_off += cachep->slab_size;
1586 slabp->colouroff = colour_off;
1587 slabp->s_mem = objp+colour_off;
1592 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1594 return (kmem_bufctl_t *)(slabp+1);
1597 static void cache_init_objs (kmem_cache_t * cachep,
1598 struct slab * slabp, unsigned long ctor_flags)
1602 for (i = 0; i < cachep->num; i++) {
1603 void* objp = slabp->s_mem+cachep->objsize*i;
1605 /* need to poison the objs? */
1606 if (cachep->flags & SLAB_POISON)
1607 poison_obj(cachep, objp, POISON_FREE);
1608 if (cachep->flags & SLAB_STORE_USER)
1609 *dbg_userword(cachep, objp) = NULL;
1611 if (cachep->flags & SLAB_RED_ZONE) {
1612 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1613 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1616 * Constructors are not allowed to allocate memory from
1617 * the same cache which they are a constructor for.
1618 * Otherwise, deadlock. They must also be threaded.
1620 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1621 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1623 if (cachep->flags & SLAB_RED_ZONE) {
1624 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1625 slab_error(cachep, "constructor overwrote the"
1626 " end of an object");
1627 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1628 slab_error(cachep, "constructor overwrote the"
1629 " start of an object");
1631 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1632 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1635 cachep->ctor(objp, cachep, ctor_flags);
1637 slab_bufctl(slabp)[i] = i+1;
1639 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1643 static void kmem_flagcheck(kmem_cache_t *cachep, int flags)
1645 if (flags & SLAB_DMA) {
1646 if (!(cachep->gfpflags & GFP_DMA))
1649 if (cachep->gfpflags & GFP_DMA)
1655 * Grow (by 1) the number of slabs within a cache. This is called by
1656 * kmem_cache_alloc() when there are no active objs left in a cache.
1658 static int cache_grow (kmem_cache_t * cachep, int flags)
1664 unsigned int i, local_flags;
1665 unsigned long ctor_flags;
1667 /* Be lazy and only check for valid flags here,
1668 * keeping it out of the critical path in kmem_cache_alloc().
1670 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1672 if (flags & SLAB_NO_GROW)
1675 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1676 local_flags = (flags & SLAB_LEVEL_MASK);
1677 if (!(local_flags & __GFP_WAIT))
1679 * Not allowed to sleep. Need to tell a constructor about
1680 * this - it might need to know...
1682 ctor_flags |= SLAB_CTOR_ATOMIC;
1684 /* About to mess with non-constant members - lock. */
1686 spin_lock(&cachep->spinlock);
1688 /* Get colour for the slab, and cal the next value. */
1689 offset = cachep->colour_next;
1690 cachep->colour_next++;
1691 if (cachep->colour_next >= cachep->colour)
1692 cachep->colour_next = 0;
1693 offset *= cachep->colour_off;
1695 spin_unlock(&cachep->spinlock);
1697 if (local_flags & __GFP_WAIT)
1701 * The test for missing atomic flag is performed here, rather than
1702 * the more obvious place, simply to reduce the critical path length
1703 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1704 * will eventually be caught here (where it matters).
1706 kmem_flagcheck(cachep, flags);
1709 /* Get mem for the objs. */
1710 if (!(objp = kmem_getpages(cachep, flags)))
1713 /* Get slab management. */
1714 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1717 /* Nasty!!!!!! I hope this is OK. */
1718 i = 1 << cachep->gfporder;
1719 page = virt_to_page(objp);
1721 SET_PAGE_CACHE(page, cachep);
1722 SET_PAGE_SLAB(page, slabp);
1726 cache_init_objs(cachep, slabp, ctor_flags);
1728 if (local_flags & __GFP_WAIT)
1729 local_irq_disable();
1731 spin_lock(&cachep->spinlock);
1733 /* Make slab active. */
1734 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1735 STATS_INC_GROWN(cachep);
1736 list3_data(cachep)->free_objects += cachep->num;
1737 spin_unlock(&cachep->spinlock);
1740 kmem_freepages(cachep, objp);
1742 if (local_flags & __GFP_WAIT)
1743 local_irq_disable();
1748 * Perform extra freeing checks:
1749 * - detect bad pointers.
1750 * - POISON/RED_ZONE checking
1751 * - destructor calls, for caches with POISON+dtor
1753 static inline void kfree_debugcheck(const void *objp)
1758 if (!virt_addr_valid(objp)) {
1759 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1760 (unsigned long)objp);
1763 page = virt_to_page(objp);
1764 if (!PageSlab(page)) {
1765 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1771 static inline void *cache_free_debugcheck (kmem_cache_t * cachep, void * objp, void *caller)
1778 objp -= obj_dbghead(cachep);
1779 kfree_debugcheck(objp);
1780 page = virt_to_page(objp);
1782 if (GET_PAGE_CACHE(page) != cachep) {
1783 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1784 GET_PAGE_CACHE(page),cachep);
1785 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1786 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1789 slabp = GET_PAGE_SLAB(page);
1791 if (cachep->flags & SLAB_RED_ZONE) {
1792 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1793 slab_error(cachep, "double free, or memory outside"
1794 " object was overwritten");
1795 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1796 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1798 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1799 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1801 if (cachep->flags & SLAB_STORE_USER)
1802 *dbg_userword(cachep, objp) = caller;
1804 objnr = (objp-slabp->s_mem)/cachep->objsize;
1806 BUG_ON(objnr >= cachep->num);
1807 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1809 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1810 /* Need to call the slab's constructor so the
1811 * caller can perform a verify of its state (debugging).
1812 * Called without the cache-lock held.
1814 cachep->ctor(objp+obj_dbghead(cachep),
1815 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1817 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1818 /* we want to cache poison the object,
1819 * call the destruction callback
1821 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1823 if (cachep->flags & SLAB_POISON) {
1824 #ifdef CONFIG_DEBUG_PAGEALLOC
1825 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1826 store_stackinfo(cachep, objp, (unsigned long)caller);
1827 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1829 poison_obj(cachep, objp, POISON_FREE);
1832 poison_obj(cachep, objp, POISON_FREE);
1839 static inline void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
1845 check_spinlock_acquired(cachep);
1846 /* Check slab's freelist to see if this obj is there. */
1847 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
1849 if (entries > cachep->num || i < 0 || i >= cachep->num)
1852 if (entries != cachep->num - slabp->inuse) {
1855 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
1856 cachep->name, cachep->num, slabp, slabp->inuse);
1857 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
1859 printk("\n%03x:", i);
1860 printk(" %02x", ((unsigned char*)slabp)[i]);
1868 static void* cache_alloc_refill(kmem_cache_t* cachep, int flags)
1871 struct kmem_list3 *l3;
1872 struct array_cache *ac;
1875 ac = ac_data(cachep);
1877 batchcount = ac->batchcount;
1878 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
1879 /* if there was little recent activity on this
1880 * cache, then perform only a partial refill.
1881 * Otherwise we could generate refill bouncing.
1883 batchcount = BATCHREFILL_LIMIT;
1885 l3 = list3_data(cachep);
1887 BUG_ON(ac->avail > 0);
1888 spin_lock(&cachep->spinlock);
1890 struct array_cache *shared_array = l3->shared;
1891 if (shared_array->avail) {
1892 if (batchcount > shared_array->avail)
1893 batchcount = shared_array->avail;
1894 shared_array->avail -= batchcount;
1895 ac->avail = batchcount;
1896 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
1897 sizeof(void*)*batchcount);
1898 shared_array->touched = 1;
1902 while (batchcount > 0) {
1903 struct list_head *entry;
1905 /* Get slab alloc is to come from. */
1906 entry = l3->slabs_partial.next;
1907 if (entry == &l3->slabs_partial) {
1908 l3->free_touched = 1;
1909 entry = l3->slabs_free.next;
1910 if (entry == &l3->slabs_free)
1914 slabp = list_entry(entry, struct slab, list);
1915 check_slabp(cachep, slabp);
1916 check_spinlock_acquired(cachep);
1917 while (slabp->inuse < cachep->num && batchcount--) {
1919 STATS_INC_ALLOCED(cachep);
1920 STATS_INC_ACTIVE(cachep);
1921 STATS_SET_HIGH(cachep);
1923 /* get obj pointer */
1924 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
1927 next = slab_bufctl(slabp)[slabp->free];
1929 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
1933 check_slabp(cachep, slabp);
1935 /* move slabp to correct slabp list: */
1936 list_del(&slabp->list);
1937 if (slabp->free == BUFCTL_END)
1938 list_add(&slabp->list, &l3->slabs_full);
1940 list_add(&slabp->list, &l3->slabs_partial);
1944 l3->free_objects -= ac->avail;
1946 spin_unlock(&cachep->spinlock);
1948 if (unlikely(!ac->avail)) {
1950 x = cache_grow(cachep, flags);
1952 // cache_grow can reenable interrupts, then ac could change.
1953 ac = ac_data(cachep);
1954 if (!x && ac->avail == 0) // no objects in sight? abort
1957 if (!ac->avail) // objects refilled by interrupt?
1961 return ac_entry(ac)[--ac->avail];
1965 cache_alloc_debugcheck_before(kmem_cache_t *cachep, int flags)
1967 might_sleep_if(flags & __GFP_WAIT);
1969 kmem_flagcheck(cachep, flags);
1973 static inline void *
1974 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
1975 unsigned long flags, void *objp, void *caller)
1980 if (cachep->flags & SLAB_POISON) {
1981 #ifdef CONFIG_DEBUG_PAGEALLOC
1982 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
1983 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
1985 check_poison_obj(cachep, objp);
1987 check_poison_obj(cachep, objp);
1989 poison_obj(cachep, objp, POISON_INUSE);
1991 if (cachep->flags & SLAB_STORE_USER)
1992 *dbg_userword(cachep, objp) = caller;
1994 if (cachep->flags & SLAB_RED_ZONE) {
1995 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
1996 slab_error(cachep, "double free, or memory outside"
1997 " object was overwritten");
1998 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1999 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2001 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2002 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2004 objp += obj_dbghead(cachep);
2005 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2006 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2008 if (!(flags & __GFP_WAIT))
2009 ctor_flags |= SLAB_CTOR_ATOMIC;
2011 cachep->ctor(objp, cachep, ctor_flags);
2018 static inline void * __cache_alloc (kmem_cache_t *cachep, int flags)
2020 unsigned long save_flags;
2022 struct array_cache *ac;
2024 cache_alloc_debugcheck_before(cachep, flags);
2026 local_irq_save(save_flags);
2027 ac = ac_data(cachep);
2028 if (likely(ac->avail)) {
2029 STATS_INC_ALLOCHIT(cachep);
2031 objp = ac_entry(ac)[--ac->avail];
2033 STATS_INC_ALLOCMISS(cachep);
2034 objp = cache_alloc_refill(cachep, flags);
2036 local_irq_restore(save_flags);
2037 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2042 * NUMA: different approach needed if the spinlock is moved into
2046 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2050 check_spinlock_acquired(cachep);
2052 /* NUMA: move add into loop */
2053 cachep->lists.free_objects += nr_objects;
2055 for (i = 0; i < nr_objects; i++) {
2056 void *objp = objpp[i];
2060 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2061 list_del(&slabp->list);
2062 objnr = (objp - slabp->s_mem) / cachep->objsize;
2063 check_slabp(cachep, slabp);
2065 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2066 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2067 cachep->name, objp);
2071 slab_bufctl(slabp)[objnr] = slabp->free;
2072 slabp->free = objnr;
2073 STATS_DEC_ACTIVE(cachep);
2075 check_slabp(cachep, slabp);
2077 /* fixup slab chains */
2078 if (slabp->inuse == 0) {
2079 if (cachep->lists.free_objects > cachep->free_limit) {
2080 cachep->lists.free_objects -= cachep->num;
2081 slab_destroy(cachep, slabp);
2083 list_add(&slabp->list,
2084 &list3_data_ptr(cachep, objp)->slabs_free);
2087 /* Unconditionally move a slab to the end of the
2088 * partial list on free - maximum time for the
2089 * other objects to be freed, too.
2091 list_add_tail(&slabp->list,
2092 &list3_data_ptr(cachep, objp)->slabs_partial);
2097 static void cache_flusharray (kmem_cache_t* cachep, struct array_cache *ac)
2101 batchcount = ac->batchcount;
2103 BUG_ON(!batchcount || batchcount > ac->avail);
2106 spin_lock(&cachep->spinlock);
2107 if (cachep->lists.shared) {
2108 struct array_cache *shared_array = cachep->lists.shared;
2109 int max = shared_array->limit-shared_array->avail;
2111 if (batchcount > max)
2113 memcpy(&ac_entry(shared_array)[shared_array->avail],
2115 sizeof(void*)*batchcount);
2116 shared_array->avail += batchcount;
2121 free_block(cachep, &ac_entry(ac)[0], batchcount);
2126 struct list_head *p;
2128 p = list3_data(cachep)->slabs_free.next;
2129 while (p != &(list3_data(cachep)->slabs_free)) {
2132 slabp = list_entry(p, struct slab, list);
2133 BUG_ON(slabp->inuse);
2138 STATS_SET_FREEABLE(cachep, i);
2141 spin_unlock(&cachep->spinlock);
2142 ac->avail -= batchcount;
2143 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2144 sizeof(void*)*ac->avail);
2149 * Release an obj back to its cache. If the obj has a constructed
2150 * state, it must be in this state _before_ it is released.
2152 * Called with disabled ints.
2154 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
2156 struct array_cache *ac = ac_data(cachep);
2159 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2161 if (likely(ac->avail < ac->limit)) {
2162 STATS_INC_FREEHIT(cachep);
2163 ac_entry(ac)[ac->avail++] = objp;
2166 STATS_INC_FREEMISS(cachep);
2167 cache_flusharray(cachep, ac);
2168 ac_entry(ac)[ac->avail++] = objp;
2173 * kmem_cache_alloc - Allocate an object
2174 * @cachep: The cache to allocate from.
2175 * @flags: See kmalloc().
2177 * Allocate an object from this cache. The flags are only relevant
2178 * if the cache has no available objects.
2180 void * kmem_cache_alloc (kmem_cache_t *cachep, int flags)
2182 return __cache_alloc(cachep, flags);
2185 EXPORT_SYMBOL(kmem_cache_alloc);
2188 * kmem_ptr_validate - check if an untrusted pointer might
2190 * @cachep: the cache we're checking against
2191 * @ptr: pointer to validate
2193 * This verifies that the untrusted pointer looks sane:
2194 * it is _not_ a guarantee that the pointer is actually
2195 * part of the slab cache in question, but it at least
2196 * validates that the pointer can be dereferenced and
2197 * looks half-way sane.
2199 * Currently only used for dentry validation.
2201 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2203 unsigned long addr = (unsigned long) ptr;
2204 unsigned long min_addr = PAGE_OFFSET;
2205 unsigned long align_mask = BYTES_PER_WORD-1;
2206 unsigned long size = cachep->objsize;
2209 if (unlikely(addr < min_addr))
2211 if (unlikely(addr > (unsigned long)high_memory - size))
2213 if (unlikely(addr & align_mask))
2215 if (unlikely(!kern_addr_valid(addr)))
2217 if (unlikely(!kern_addr_valid(addr + size - 1)))
2219 page = virt_to_page(ptr);
2220 if (unlikely(!PageSlab(page)))
2222 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2230 * kmalloc - allocate memory
2231 * @size: how many bytes of memory are required.
2232 * @flags: the type of memory to allocate.
2234 * kmalloc is the normal method of allocating memory
2237 * The @flags argument may be one of:
2239 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2241 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2243 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2245 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2246 * must be suitable for DMA. This can mean different things on different
2247 * platforms. For example, on i386, it means that the memory must come
2248 * from the first 16MB.
2250 void * __kmalloc (size_t size, int flags)
2252 struct cache_sizes *csizep = malloc_sizes;
2254 for (; csizep->cs_size; csizep++) {
2255 if (size > csizep->cs_size)
2258 /* This happens if someone tries to call
2259 * kmem_cache_create(), or kmalloc(), before
2260 * the generic caches are initialized.
2262 BUG_ON(csizep->cs_cachep == NULL);
2264 return __cache_alloc(flags & GFP_DMA ?
2265 csizep->cs_dmacachep : csizep->cs_cachep, flags);
2270 EXPORT_SYMBOL(__kmalloc);
2274 * __alloc_percpu - allocate one copy of the object for every present
2275 * cpu in the system, zeroing them.
2276 * Objects should be dereferenced using per_cpu_ptr/get_cpu_ptr
2279 * @size: how many bytes of memory are required.
2280 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2282 void *__alloc_percpu(size_t size, size_t align)
2285 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2290 for (i = 0; i < NR_CPUS; i++) {
2291 if (!cpu_possible(i))
2293 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
2294 if (!pdata->ptrs[i])
2296 memset(pdata->ptrs[i], 0, size);
2299 /* Catch derefs w/o wrappers */
2300 return (void *) (~(unsigned long) pdata);
2304 if (!cpu_possible(i))
2306 kfree(pdata->ptrs[i]);
2312 EXPORT_SYMBOL(__alloc_percpu);
2316 * kmem_cache_free - Deallocate an object
2317 * @cachep: The cache the allocation was from.
2318 * @objp: The previously allocated object.
2320 * Free an object which was previously allocated from this
2323 void kmem_cache_free (kmem_cache_t *cachep, void *objp)
2325 unsigned long flags;
2327 local_irq_save(flags);
2328 __cache_free(cachep, objp);
2329 local_irq_restore(flags);
2332 EXPORT_SYMBOL(kmem_cache_free);
2335 * kfree - free previously allocated memory
2336 * @objp: pointer returned by kmalloc.
2338 * Don't free memory not originally allocated by kmalloc()
2339 * or you will run into trouble.
2341 void kfree (const void *objp)
2344 unsigned long flags;
2348 local_irq_save(flags);
2349 kfree_debugcheck(objp);
2350 c = GET_PAGE_CACHE(virt_to_page(objp));
2351 __cache_free(c, (void*)objp);
2352 local_irq_restore(flags);
2355 EXPORT_SYMBOL(kfree);
2359 * free_percpu - free previously allocated percpu memory
2360 * @objp: pointer returned by alloc_percpu.
2362 * Don't free memory not originally allocated by alloc_percpu()
2363 * The complemented objp is to check for that.
2366 free_percpu(const void *objp)
2369 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2371 for (i = 0; i < NR_CPUS; i++) {
2372 if (!cpu_possible(i))
2378 EXPORT_SYMBOL(free_percpu);
2381 unsigned int kmem_cache_size(kmem_cache_t *cachep)
2383 return obj_reallen(cachep);
2386 EXPORT_SYMBOL(kmem_cache_size);
2388 kmem_cache_t * kmem_find_general_cachep (size_t size, int gfpflags)
2390 struct cache_sizes *csizep = malloc_sizes;
2392 /* This function could be moved to the header file, and
2393 * made inline so consumers can quickly determine what
2394 * cache pointer they require.
2396 for ( ; csizep->cs_size; csizep++) {
2397 if (size > csizep->cs_size)
2401 return (gfpflags & GFP_DMA) ? csizep->cs_dmacachep : csizep->cs_cachep;
2404 EXPORT_SYMBOL(kmem_find_general_cachep);
2406 struct ccupdate_struct {
2407 kmem_cache_t *cachep;
2408 struct array_cache *new[NR_CPUS];
2411 static void do_ccupdate_local(void *info)
2413 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2414 struct array_cache *old;
2417 old = ac_data(new->cachep);
2419 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2420 new->new[smp_processor_id()] = old;
2424 static int do_tune_cpucache (kmem_cache_t* cachep, int limit, int batchcount, int shared)
2426 struct ccupdate_struct new;
2427 struct array_cache *new_shared;
2430 memset(&new.new,0,sizeof(new.new));
2431 for (i = 0; i < NR_CPUS; i++) {
2432 struct array_cache *ccnew;
2434 ccnew = kmalloc(sizeof(void*)*limit+
2435 sizeof(struct array_cache), GFP_KERNEL);
2437 for (i--; i >= 0; i--) kfree(new.new[i]);
2441 ccnew->limit = limit;
2442 ccnew->batchcount = batchcount;
2446 new.cachep = cachep;
2448 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2451 spin_lock_irq(&cachep->spinlock);
2452 cachep->batchcount = batchcount;
2453 cachep->limit = limit;
2454 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2455 spin_unlock_irq(&cachep->spinlock);
2457 for (i = 0; i < NR_CPUS; i++) {
2458 struct array_cache *ccold = new.new[i];
2461 spin_lock_irq(&cachep->spinlock);
2462 free_block(cachep, ac_entry(ccold), ccold->avail);
2463 spin_unlock_irq(&cachep->spinlock);
2466 new_shared = kmalloc(sizeof(void*)*batchcount*shared+
2467 sizeof(struct array_cache), GFP_KERNEL);
2469 struct array_cache *old;
2470 new_shared->avail = 0;
2471 new_shared->limit = batchcount*shared;
2472 new_shared->batchcount = 0xbaadf00d;
2473 new_shared->touched = 0;
2475 spin_lock_irq(&cachep->spinlock);
2476 old = cachep->lists.shared;
2477 cachep->lists.shared = new_shared;
2479 free_block(cachep, ac_entry(old), old->avail);
2480 spin_unlock_irq(&cachep->spinlock);
2488 static void enable_cpucache (kmem_cache_t *cachep)
2493 /* The head array serves three purposes:
2494 * - create a LIFO ordering, i.e. return objects that are cache-warm
2495 * - reduce the number of spinlock operations.
2496 * - reduce the number of linked list operations on the slab and
2497 * bufctl chains: array operations are cheaper.
2498 * The numbers are guessed, we should auto-tune as described by
2501 if (cachep->objsize > 131072)
2503 else if (cachep->objsize > PAGE_SIZE)
2505 else if (cachep->objsize > 1024)
2507 else if (cachep->objsize > 256)
2512 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2513 * allocation behaviour: Most allocs on one cpu, most free operations
2514 * on another cpu. For these cases, an efficient object passing between
2515 * cpus is necessary. This is provided by a shared array. The array
2516 * replaces Bonwick's magazine layer.
2517 * On uniprocessor, it's functionally equivalent (but less efficient)
2518 * to a larger limit. Thus disabled by default.
2522 if (cachep->objsize <= PAGE_SIZE)
2527 /* With debugging enabled, large batchcount lead to excessively
2528 * long periods with disabled local interrupts. Limit the
2534 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2536 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2537 cachep->name, -err);
2540 static void drain_array(kmem_cache_t *cachep, struct array_cache *ac)
2547 } else if (ac->avail) {
2548 tofree = (ac->limit+4)/5;
2549 if (tofree > ac->avail) {
2550 tofree = (ac->avail+1)/2;
2552 spin_lock(&cachep->spinlock);
2553 free_block(cachep, ac_entry(ac), tofree);
2554 spin_unlock(&cachep->spinlock);
2555 ac->avail -= tofree;
2556 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2557 sizeof(void*)*ac->avail);
2561 static void drain_array_locked(kmem_cache_t *cachep,
2562 struct array_cache *ac, int force)
2566 check_spinlock_acquired(cachep);
2567 if (ac->touched && !force) {
2569 } else if (ac->avail) {
2570 tofree = force ? ac->avail : (ac->limit+4)/5;
2571 if (tofree > ac->avail) {
2572 tofree = (ac->avail+1)/2;
2574 free_block(cachep, ac_entry(ac), tofree);
2575 ac->avail -= tofree;
2576 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2577 sizeof(void*)*ac->avail);
2582 * cache_reap - Reclaim memory from caches.
2584 * Called from a timer, every few seconds
2586 * - clear the per-cpu caches for this CPU.
2587 * - return freeable pages to the main free memory pool.
2589 * If we cannot acquire the cache chain semaphore then just give up - we'll
2590 * try again next timer interrupt.
2592 static inline void cache_reap (void)
2594 struct list_head *walk;
2597 BUG_ON(!in_interrupt());
2600 if (down_trylock(&cache_chain_sem))
2603 list_for_each(walk, &cache_chain) {
2604 kmem_cache_t *searchp;
2605 struct list_head* p;
2609 searchp = list_entry(walk, kmem_cache_t, next);
2611 if (searchp->flags & SLAB_NO_REAP)
2615 local_irq_disable();
2616 drain_array(searchp, ac_data(searchp));
2618 if(time_after(searchp->lists.next_reap, jiffies))
2621 spin_lock(&searchp->spinlock);
2622 if(time_after(searchp->lists.next_reap, jiffies)) {
2625 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2627 if (searchp->lists.shared)
2628 drain_array_locked(searchp, searchp->lists.shared, 0);
2630 if (searchp->lists.free_touched) {
2631 searchp->lists.free_touched = 0;
2635 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2637 p = list3_data(searchp)->slabs_free.next;
2638 if (p == &(list3_data(searchp)->slabs_free))
2641 slabp = list_entry(p, struct slab, list);
2642 BUG_ON(slabp->inuse);
2643 list_del(&slabp->list);
2644 STATS_INC_REAPED(searchp);
2646 /* Safe to drop the lock. The slab is no longer
2647 * linked to the cache.
2648 * searchp cannot disappear, we hold
2651 searchp->lists.free_objects -= searchp->num;
2652 spin_unlock_irq(&searchp->spinlock);
2653 slab_destroy(searchp, slabp);
2654 spin_lock_irq(&searchp->spinlock);
2655 } while(--tofree > 0);
2657 spin_unlock(&searchp->spinlock);
2664 up(&cache_chain_sem);
2668 * This is a timer handler. There is one per CPU. It is called periodially
2669 * to shrink this CPU's caches. Otherwise there could be memory tied up
2670 * for long periods (or for ever) due to load changes.
2672 static void reap_timer_fnc(unsigned long cpu)
2674 struct timer_list *rt = &__get_cpu_var(reap_timers);
2676 /* CPU hotplug can drag us off cpu: don't run on wrong CPU */
2677 if (!cpu_is_offline(cpu)) {
2679 mod_timer(rt, jiffies + REAPTIMEOUT_CPUC + cpu);
2683 #ifdef CONFIG_PROC_FS
2685 static void *s_start(struct seq_file *m, loff_t *pos)
2688 struct list_head *p;
2690 down(&cache_chain_sem);
2693 * Output format version, so at least we can change it
2694 * without _too_ many complaints.
2697 seq_puts(m, "slabinfo - version: 2.0 (statistics)\n");
2699 seq_puts(m, "slabinfo - version: 2.0\n");
2701 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2702 seq_puts(m, " : tunables <batchcount> <limit> <sharedfactor>");
2703 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2705 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <freelimit>");
2706 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2710 p = cache_chain.next;
2713 if (p == &cache_chain)
2716 return list_entry(p, kmem_cache_t, next);
2719 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2721 kmem_cache_t *cachep = p;
2723 return cachep->next.next == &cache_chain ? NULL
2724 : list_entry(cachep->next.next, kmem_cache_t, next);
2727 static void s_stop(struct seq_file *m, void *p)
2729 up(&cache_chain_sem);
2732 static int s_show(struct seq_file *m, void *p)
2734 kmem_cache_t *cachep = p;
2735 struct list_head *q;
2737 unsigned long active_objs;
2738 unsigned long num_objs;
2739 unsigned long active_slabs = 0;
2740 unsigned long num_slabs;
2743 mm_segment_t old_fs;
2747 spin_lock_irq(&cachep->spinlock);
2750 list_for_each(q,&cachep->lists.slabs_full) {
2751 slabp = list_entry(q, struct slab, list);
2752 if (slabp->inuse != cachep->num && !error)
2753 error = "slabs_full accounting error";
2754 active_objs += cachep->num;
2757 list_for_each(q,&cachep->lists.slabs_partial) {
2758 slabp = list_entry(q, struct slab, list);
2759 if (slabp->inuse == cachep->num && !error)
2760 error = "slabs_partial inuse accounting error";
2761 if (!slabp->inuse && !error)
2762 error = "slabs_partial/inuse accounting error";
2763 active_objs += slabp->inuse;
2766 list_for_each(q,&cachep->lists.slabs_free) {
2767 slabp = list_entry(q, struct slab, list);
2768 if (slabp->inuse && !error)
2769 error = "slabs_free/inuse accounting error";
2772 num_slabs+=active_slabs;
2773 num_objs = num_slabs*cachep->num;
2774 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2775 error = "free_objects accounting error";
2777 name = cachep->name;
2780 * Check to see if `name' resides inside a module which has been
2781 * unloaded (someone forgot to destroy their cache)
2785 if (__get_user(tmp, name))
2790 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2792 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2793 name, active_objs, num_objs, cachep->objsize,
2794 cachep->num, (1<<cachep->gfporder));
2795 seq_printf(m, " : tunables %4u %4u %4u",
2796 cachep->limit, cachep->batchcount,
2797 cachep->lists.shared->limit/cachep->batchcount);
2798 seq_printf(m, " : slabdata %6lu %6lu %6u",
2799 active_slabs, num_slabs, cachep->lists.shared->avail);
2802 unsigned long high = cachep->high_mark;
2803 unsigned long allocs = cachep->num_allocations;
2804 unsigned long grown = cachep->grown;
2805 unsigned long reaped = cachep->reaped;
2806 unsigned long errors = cachep->errors;
2807 unsigned long max_freeable = cachep->max_freeable;
2808 unsigned long free_limit = cachep->free_limit;
2810 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu",
2811 allocs, high, grown, reaped, errors,
2812 max_freeable, free_limit);
2816 unsigned long allochit = atomic_read(&cachep->allochit);
2817 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2818 unsigned long freehit = atomic_read(&cachep->freehit);
2819 unsigned long freemiss = atomic_read(&cachep->freemiss);
2821 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2822 allochit, allocmiss, freehit, freemiss);
2826 spin_unlock_irq(&cachep->spinlock);
2831 * slabinfo_op - iterator that generates /proc/slabinfo
2840 * num-pages-per-slab
2841 * + further values on SMP and with statistics enabled
2844 struct seq_operations slabinfo_op = {
2851 #define MAX_SLABINFO_WRITE 128
2853 * slabinfo_write - Tuning for the slab allocator
2855 * @buffer: user buffer
2856 * @count: data length
2859 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
2860 size_t count, loff_t *ppos)
2862 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
2863 int limit, batchcount, shared, res;
2864 struct list_head *p;
2866 if (count > MAX_SLABINFO_WRITE)
2868 if (copy_from_user(&kbuf, buffer, count))
2870 kbuf[MAX_SLABINFO_WRITE] = '\0';
2872 tmp = strchr(kbuf, ' ');
2877 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
2880 /* Find the cache in the chain of caches. */
2881 down(&cache_chain_sem);
2883 list_for_each(p,&cache_chain) {
2884 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
2886 if (!strcmp(cachep->name, kbuf)) {
2889 batchcount > limit ||
2893 res = do_tune_cpucache(cachep, limit, batchcount, shared);
2898 up(&cache_chain_sem);
2905 unsigned int ksize(const void *objp)
2908 unsigned long flags;
2909 unsigned int size = 0;
2911 if (likely(objp != NULL)) {
2912 local_irq_save(flags);
2913 c = GET_PAGE_CACHE(virt_to_page(objp));
2914 size = kmem_cache_size(c);
2915 local_irq_restore(flags);
2921 void ptrinfo(unsigned long addr)
2925 printk("Dumping data about address %p.\n", (void*)addr);
2926 if (!virt_addr_valid((void*)addr)) {
2927 printk("virt addr invalid.\n");
2932 pgd_t *pgd = pgd_offset_k(addr);
2934 if (pgd_none(*pgd)) {
2935 printk("No pgd.\n");
2938 pmd = pmd_offset(pgd, addr);
2939 if (pmd_none(*pmd)) {
2940 printk("No pmd.\n");
2944 if (pmd_large(*pmd)) {
2945 printk("Large page.\n");
2949 printk("normal page, pte_val 0x%llx\n",
2950 (unsigned long long)pte_val(*pte_offset_kernel(pmd, addr)));
2954 page = virt_to_page((void*)addr);
2955 printk("struct page at %p, flags %08lx\n",
2956 page, (unsigned long)page->flags);
2957 if (PageSlab(page)) {
2960 unsigned long flags;
2964 c = GET_PAGE_CACHE(page);
2965 printk("belongs to cache %s.\n",c->name);
2967 spin_lock_irqsave(&c->spinlock, flags);
2968 s = GET_PAGE_SLAB(page);
2969 printk("slabp %p with %d inuse objects (from %d).\n",
2970 s, s->inuse, c->num);
2973 objnr = (addr-(unsigned long)s->s_mem)/c->objsize;
2974 objp = s->s_mem+c->objsize*objnr;
2975 printk("points into object no %d, starting at %p, len %d.\n",
2976 objnr, objp, c->objsize);
2977 if (objnr >= c->num) {
2978 printk("Bad obj number.\n");
2980 kernel_map_pages(virt_to_page(objp),
2981 c->objsize/PAGE_SIZE, 1);
2983 print_objinfo(c, objp, 2);
2985 spin_unlock_irqrestore(&c->spinlock, flags);