ia64/linux-2.6.18-xen.hg

view mm/slab.c @ 912:dd42cdb0ab89

[IA64] Build blktap2 driver by default in x86 builds.

add CONFIG_XEN_BLKDEV_TAP2=y to buildconfigs/linux-defconfig_xen_ia64.

Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
author Isaku Yamahata <yamahata@valinux.co.jp>
date Mon Jun 29 12:09:16 2009 +0900 (2009-06-29)
parents 3e8752eb6d9c
children
line source
1 /*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
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
21 *
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.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
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.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
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.
46 *
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.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
89 #include <linux/config.h>
90 #include <linux/slab.h>
91 #include <linux/mm.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.h>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
117 /*
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
128 #ifdef CONFIG_DEBUG_SLAB
129 #define DEBUG 1
130 #define STATS 1
131 #define FORCED_DEBUG 1
132 #else
133 #define DEBUG 0
134 #define STATS 0
135 #define FORCED_DEBUG 0
136 #endif
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
143 #endif
145 #ifndef ARCH_KMALLOC_MINALIGN
146 /*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
153 */
154 #define ARCH_KMALLOC_MINALIGN 0
155 #endif
157 #ifndef ARCH_SLAB_MINALIGN
158 /*
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
164 */
165 #define ARCH_SLAB_MINALIGN 0
166 #endif
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170 #endif
172 /* Legal flag mask for kmem_cache_create(). */
173 #if DEBUG
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
176 SLAB_CACHE_DMA | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
180 #else
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
185 #endif
187 /*
188 * kmem_bufctl_t:
189 *
190 * Bufctl's are used for linking objs within a slab
191 * linked offsets.
192 *
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 */
206 typedef unsigned int kmem_bufctl_t;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
212 /*
213 * struct slab
214 *
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 */
219 struct slab {
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
224 kmem_bufctl_t free;
225 unsigned short nodeid;
226 };
228 /*
229 * struct slab_rcu
230 *
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
238 *
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
241 *
242 * We assume struct slab_rcu can overlay struct slab when destroying.
243 */
244 struct slab_rcu {
245 struct rcu_head head;
246 struct kmem_cache *cachep;
247 void *addr;
248 };
250 /*
251 * struct array_cache
252 *
253 * Purpose:
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
257 *
258 * The limit is stored in the per-cpu structure to reduce the data cache
259 * footprint.
260 *
261 */
262 struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
267 spinlock_t lock;
268 void *entry[0]; /*
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
271 * the entries.
272 * [0] is for gcc 2.95. It should really be [].
273 */
274 };
276 /*
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
279 */
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
284 };
286 /*
287 * The slab lists for all objects.
288 */
289 struct kmem_list3 {
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned int free_limit;
295 unsigned int colour_next; /* Per-node cache coloring */
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
301 };
303 /*
304 * Need this for bootstrapping a per node allocator.
305 */
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
309 #define SIZE_AC 1
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 int node);
316 static void enable_cpucache(struct kmem_cache *cachep);
317 static void cache_reap(void *unused);
319 /*
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
322 */
323 static __always_inline int index_of(const size_t size)
324 {
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size)) {
328 int i = 0;
330 #define CACHE(x) \
331 if (size <=x) \
332 return i; \
333 else \
334 i++;
335 #include "linux/kmalloc_sizes.h"
336 #undef CACHE
337 __bad_size();
338 } else
339 __bad_size();
340 return 0;
341 }
343 static int slab_early_init = 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3 *parent)
349 {
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
355 parent->colour_next = 0;
356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
359 }
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
362 do { \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
365 } while (0)
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 do { \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 } while (0)
374 /*
375 * struct kmem_cache
376 *
377 * manages a cache.
378 */
380 struct kmem_cache {
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache *array[NR_CPUS];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount;
385 unsigned int limit;
386 unsigned int shared;
388 unsigned int buffer_size;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3 *nodelists[MAX_NUMNODES];
392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder;
399 /* force GFP flags, e.g. GFP_DMA */
400 gfp_t gfpflags;
402 size_t colour; /* cache colouring range */
403 unsigned int colour_off; /* colour offset */
404 struct kmem_cache *slabp_cache;
405 unsigned int slab_size;
406 unsigned int dflags; /* dynamic flags */
408 /* constructor func */
409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
411 /* de-constructor func */
412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
414 /* 5) cache creation/removal */
415 const char *name;
416 struct list_head next;
418 /* 6) statistics */
419 #if STATS
420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
423 unsigned long grown;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
429 unsigned long node_overflow;
430 atomic_t allochit;
431 atomic_t allocmiss;
432 atomic_t freehit;
433 atomic_t freemiss;
434 #endif
435 #if DEBUG
436 /*
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
441 */
442 int obj_offset;
443 int obj_size;
444 #endif
445 };
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
451 /*
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
454 *
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
457 */
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
461 #if STATS
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
468 do { \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
471 } while (0)
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
477 do { \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
480 } while (0)
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
485 #else
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
501 #endif
503 #if DEBUG
505 /*
506 * memory layout of objects:
507 * 0 : objp
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
512 * redzone word.
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
517 */
518 static int obj_offset(struct kmem_cache *cachep)
519 {
520 return cachep->obj_offset;
521 }
523 static int obj_size(struct kmem_cache *cachep)
524 {
525 return cachep->obj_size;
526 }
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
529 {
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
532 }
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
535 {
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
539 2 * BYTES_PER_WORD);
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
541 }
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
544 {
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
547 }
549 #else
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
557 #endif
559 /*
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
561 * order.
562 */
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
569 #else
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
572 #endif
574 /*
575 * Do not go above this order unless 0 objects fit into the slab.
576 */
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
581 /*
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
585 */
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587 {
588 page->lru.next = (struct list_head *)cache;
589 }
591 static inline struct kmem_cache *page_get_cache(struct page *page)
592 {
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 BUG_ON(!PageSlab(page));
596 return (struct kmem_cache *)page->lru.next;
597 }
599 static inline void page_set_slab(struct page *page, struct slab *slab)
600 {
601 page->lru.prev = (struct list_head *)slab;
602 }
604 static inline struct slab *page_get_slab(struct page *page)
605 {
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 BUG_ON(!PageSlab(page));
609 return (struct slab *)page->lru.prev;
610 }
612 static inline struct kmem_cache *virt_to_cache(const void *obj)
613 {
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
616 }
618 static inline struct slab *virt_to_slab(const void *obj)
619 {
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
622 }
624 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 unsigned int idx)
626 {
627 return slab->s_mem + cache->buffer_size * idx;
628 }
630 static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
632 {
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
634 }
636 /*
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 */
639 struct cache_sizes malloc_sizes[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
642 CACHE(ULONG_MAX)
643 #undef CACHE
644 };
645 EXPORT_SYMBOL(malloc_sizes);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
648 struct cache_names {
649 char *name;
650 char *name_dma;
651 };
653 static struct cache_names __initdata cache_names[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
656 {NULL,}
657 #undef CACHE
658 };
660 static struct arraycache_init initarray_cache __initdata =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
662 static struct arraycache_init initarray_generic =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache = {
667 .batchcount = 1,
668 .limit = BOOT_CPUCACHE_ENTRIES,
669 .shared = 1,
670 .buffer_size = sizeof(struct kmem_cache),
671 .name = "kmem_cache",
672 #if DEBUG
673 .obj_size = sizeof(struct kmem_cache),
674 #endif
675 };
677 #ifdef CONFIG_LOCKDEP
679 /*
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
685 */
686 static struct lock_class_key on_slab_key;
688 static inline void init_lock_keys(struct cache_sizes *s)
689 {
690 int q;
692 for (q = 0; q < MAX_NUMNODES; q++) {
693 if (!s->cs_cachep->nodelists[q] || OFF_SLAB(s->cs_cachep))
694 continue;
695 lockdep_set_class(&s->cs_cachep->nodelists[q]->list_lock,
696 &on_slab_key);
697 }
698 }
700 #else
701 static inline void init_lock_keys(struct cache_sizes *s)
702 {
703 }
704 #endif
708 /* Guard access to the cache-chain. */
709 static DEFINE_MUTEX(cache_chain_mutex);
710 static struct list_head cache_chain;
712 /*
713 * vm_enough_memory() looks at this to determine how many slab-allocated pages
714 * are possibly freeable under pressure
715 *
716 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
717 */
718 atomic_t slab_reclaim_pages;
720 /*
721 * chicken and egg problem: delay the per-cpu array allocation
722 * until the general caches are up.
723 */
724 static enum {
725 NONE,
726 PARTIAL_AC,
727 PARTIAL_L3,
728 FULL
729 } g_cpucache_up;
731 /*
732 * used by boot code to determine if it can use slab based allocator
733 */
734 int slab_is_available(void)
735 {
736 return g_cpucache_up == FULL;
737 }
739 static DEFINE_PER_CPU(struct work_struct, reap_work);
741 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
742 {
743 return cachep->array[smp_processor_id()];
744 }
746 static inline struct kmem_cache *__find_general_cachep(size_t size,
747 gfp_t gfpflags)
748 {
749 struct cache_sizes *csizep = malloc_sizes;
751 #if DEBUG
752 /* This happens if someone tries to call
753 * kmem_cache_create(), or __kmalloc(), before
754 * the generic caches are initialized.
755 */
756 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
757 #endif
758 while (size > csizep->cs_size)
759 csizep++;
761 /*
762 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
763 * has cs_{dma,}cachep==NULL. Thus no special case
764 * for large kmalloc calls required.
765 */
766 if (unlikely(gfpflags & GFP_DMA))
767 return csizep->cs_dmacachep;
768 return csizep->cs_cachep;
769 }
771 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
772 {
773 return __find_general_cachep(size, gfpflags);
774 }
775 EXPORT_SYMBOL(kmem_find_general_cachep);
777 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
778 {
779 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
780 }
782 /*
783 * Calculate the number of objects and left-over bytes for a given buffer size.
784 */
785 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
786 size_t align, int flags, size_t *left_over,
787 unsigned int *num)
788 {
789 int nr_objs;
790 size_t mgmt_size;
791 size_t slab_size = PAGE_SIZE << gfporder;
793 /*
794 * The slab management structure can be either off the slab or
795 * on it. For the latter case, the memory allocated for a
796 * slab is used for:
797 *
798 * - The struct slab
799 * - One kmem_bufctl_t for each object
800 * - Padding to respect alignment of @align
801 * - @buffer_size bytes for each object
802 *
803 * If the slab management structure is off the slab, then the
804 * alignment will already be calculated into the size. Because
805 * the slabs are all pages aligned, the objects will be at the
806 * correct alignment when allocated.
807 */
808 if (flags & CFLGS_OFF_SLAB) {
809 mgmt_size = 0;
810 nr_objs = slab_size / buffer_size;
812 if (nr_objs > SLAB_LIMIT)
813 nr_objs = SLAB_LIMIT;
814 } else {
815 /*
816 * Ignore padding for the initial guess. The padding
817 * is at most @align-1 bytes, and @buffer_size is at
818 * least @align. In the worst case, this result will
819 * be one greater than the number of objects that fit
820 * into the memory allocation when taking the padding
821 * into account.
822 */
823 nr_objs = (slab_size - sizeof(struct slab)) /
824 (buffer_size + sizeof(kmem_bufctl_t));
826 /*
827 * This calculated number will be either the right
828 * amount, or one greater than what we want.
829 */
830 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
831 > slab_size)
832 nr_objs--;
834 if (nr_objs > SLAB_LIMIT)
835 nr_objs = SLAB_LIMIT;
837 mgmt_size = slab_mgmt_size(nr_objs, align);
838 }
839 *num = nr_objs;
840 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
841 }
843 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
845 static void __slab_error(const char *function, struct kmem_cache *cachep,
846 char *msg)
847 {
848 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
849 function, cachep->name, msg);
850 dump_stack();
851 }
853 #ifdef CONFIG_NUMA
854 /*
855 * Special reaping functions for NUMA systems called from cache_reap().
856 * These take care of doing round robin flushing of alien caches (containing
857 * objects freed on different nodes from which they were allocated) and the
858 * flushing of remote pcps by calling drain_node_pages.
859 */
860 static DEFINE_PER_CPU(unsigned long, reap_node);
862 static void init_reap_node(int cpu)
863 {
864 int node;
866 node = next_node(cpu_to_node(cpu), node_online_map);
867 if (node == MAX_NUMNODES)
868 node = first_node(node_online_map);
870 per_cpu(reap_node, cpu) = node;
871 }
873 static void next_reap_node(void)
874 {
875 int node = __get_cpu_var(reap_node);
877 /*
878 * Also drain per cpu pages on remote zones
879 */
880 if (node != numa_node_id())
881 drain_node_pages(node);
883 node = next_node(node, node_online_map);
884 if (unlikely(node >= MAX_NUMNODES))
885 node = first_node(node_online_map);
886 __get_cpu_var(reap_node) = node;
887 }
889 #else
890 #define init_reap_node(cpu) do { } while (0)
891 #define next_reap_node(void) do { } while (0)
892 #endif
894 /*
895 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
896 * via the workqueue/eventd.
897 * Add the CPU number into the expiration time to minimize the possibility of
898 * the CPUs getting into lockstep and contending for the global cache chain
899 * lock.
900 */
901 static void __devinit start_cpu_timer(int cpu)
902 {
903 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
905 /*
906 * When this gets called from do_initcalls via cpucache_init(),
907 * init_workqueues() has already run, so keventd will be setup
908 * at that time.
909 */
910 if (keventd_up() && reap_work->func == NULL) {
911 init_reap_node(cpu);
912 INIT_WORK(reap_work, cache_reap, NULL);
913 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
914 }
915 }
917 static struct array_cache *alloc_arraycache(int node, int entries,
918 int batchcount)
919 {
920 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
921 struct array_cache *nc = NULL;
923 nc = kmalloc_node(memsize, GFP_KERNEL, node);
924 if (nc) {
925 nc->avail = 0;
926 nc->limit = entries;
927 nc->batchcount = batchcount;
928 nc->touched = 0;
929 spin_lock_init(&nc->lock);
930 }
931 return nc;
932 }
934 /*
935 * Transfer objects in one arraycache to another.
936 * Locking must be handled by the caller.
937 *
938 * Return the number of entries transferred.
939 */
940 static int transfer_objects(struct array_cache *to,
941 struct array_cache *from, unsigned int max)
942 {
943 /* Figure out how many entries to transfer */
944 int nr = min(min(from->avail, max), to->limit - to->avail);
946 if (!nr)
947 return 0;
949 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
950 sizeof(void *) *nr);
952 from->avail -= nr;
953 to->avail += nr;
954 to->touched = 1;
955 return nr;
956 }
958 #ifdef CONFIG_NUMA
959 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
960 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
962 static struct array_cache **alloc_alien_cache(int node, int limit)
963 {
964 struct array_cache **ac_ptr;
965 int memsize = sizeof(void *) * MAX_NUMNODES;
966 int i;
968 if (limit > 1)
969 limit = 12;
970 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
971 if (ac_ptr) {
972 for_each_node(i) {
973 if (i == node || !node_online(i)) {
974 ac_ptr[i] = NULL;
975 continue;
976 }
977 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
978 if (!ac_ptr[i]) {
979 for (i--; i <= 0; i--)
980 kfree(ac_ptr[i]);
981 kfree(ac_ptr);
982 return NULL;
983 }
984 }
985 }
986 return ac_ptr;
987 }
989 static void free_alien_cache(struct array_cache **ac_ptr)
990 {
991 int i;
993 if (!ac_ptr)
994 return;
995 for_each_node(i)
996 kfree(ac_ptr[i]);
997 kfree(ac_ptr);
998 }
1000 static void __drain_alien_cache(struct kmem_cache *cachep,
1001 struct array_cache *ac, int node)
1003 struct kmem_list3 *rl3 = cachep->nodelists[node];
1005 if (ac->avail) {
1006 spin_lock(&rl3->list_lock);
1007 /*
1008 * Stuff objects into the remote nodes shared array first.
1009 * That way we could avoid the overhead of putting the objects
1010 * into the free lists and getting them back later.
1011 */
1012 if (rl3->shared)
1013 transfer_objects(rl3->shared, ac, ac->limit);
1015 free_block(cachep, ac->entry, ac->avail, node);
1016 ac->avail = 0;
1017 spin_unlock(&rl3->list_lock);
1021 /*
1022 * Called from cache_reap() to regularly drain alien caches round robin.
1023 */
1024 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1026 int node = __get_cpu_var(reap_node);
1028 if (l3->alien) {
1029 struct array_cache *ac = l3->alien[node];
1031 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1032 __drain_alien_cache(cachep, ac, node);
1033 spin_unlock_irq(&ac->lock);
1038 static void drain_alien_cache(struct kmem_cache *cachep,
1039 struct array_cache **alien)
1041 int i = 0;
1042 struct array_cache *ac;
1043 unsigned long flags;
1045 for_each_online_node(i) {
1046 ac = alien[i];
1047 if (ac) {
1048 spin_lock_irqsave(&ac->lock, flags);
1049 __drain_alien_cache(cachep, ac, i);
1050 spin_unlock_irqrestore(&ac->lock, flags);
1055 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1057 struct slab *slabp = virt_to_slab(objp);
1058 int nodeid = slabp->nodeid;
1059 struct kmem_list3 *l3;
1060 struct array_cache *alien = NULL;
1062 /*
1063 * Make sure we are not freeing a object from another node to the array
1064 * cache on this cpu.
1065 */
1066 if (likely(slabp->nodeid == numa_node_id()))
1067 return 0;
1069 l3 = cachep->nodelists[numa_node_id()];
1070 STATS_INC_NODEFREES(cachep);
1071 if (l3->alien && l3->alien[nodeid]) {
1072 alien = l3->alien[nodeid];
1073 spin_lock(&alien->lock);
1074 if (unlikely(alien->avail == alien->limit)) {
1075 STATS_INC_ACOVERFLOW(cachep);
1076 __drain_alien_cache(cachep, alien, nodeid);
1078 alien->entry[alien->avail++] = objp;
1079 spin_unlock(&alien->lock);
1080 } else {
1081 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1082 free_block(cachep, &objp, 1, nodeid);
1083 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1085 return 1;
1088 #else
1090 #define drain_alien_cache(cachep, alien) do { } while (0)
1091 #define reap_alien(cachep, l3) do { } while (0)
1093 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1095 return (struct array_cache **) 0x01020304ul;
1098 static inline void free_alien_cache(struct array_cache **ac_ptr)
1102 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1104 return 0;
1107 #endif
1109 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1110 unsigned long action, void *hcpu)
1112 long cpu = (long)hcpu;
1113 struct kmem_cache *cachep;
1114 struct kmem_list3 *l3 = NULL;
1115 int node = cpu_to_node(cpu);
1116 int memsize = sizeof(struct kmem_list3);
1118 switch (action) {
1119 case CPU_UP_PREPARE:
1120 mutex_lock(&cache_chain_mutex);
1121 /*
1122 * We need to do this right in the beginning since
1123 * alloc_arraycache's are going to use this list.
1124 * kmalloc_node allows us to add the slab to the right
1125 * kmem_list3 and not this cpu's kmem_list3
1126 */
1128 list_for_each_entry(cachep, &cache_chain, next) {
1129 /*
1130 * Set up the size64 kmemlist for cpu before we can
1131 * begin anything. Make sure some other cpu on this
1132 * node has not already allocated this
1133 */
1134 if (!cachep->nodelists[node]) {
1135 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1136 if (!l3)
1137 goto bad;
1138 kmem_list3_init(l3);
1139 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1140 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1142 /*
1143 * The l3s don't come and go as CPUs come and
1144 * go. cache_chain_mutex is sufficient
1145 * protection here.
1146 */
1147 cachep->nodelists[node] = l3;
1150 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1151 cachep->nodelists[node]->free_limit =
1152 (1 + nr_cpus_node(node)) *
1153 cachep->batchcount + cachep->num;
1154 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1157 /*
1158 * Now we can go ahead with allocating the shared arrays and
1159 * array caches
1160 */
1161 list_for_each_entry(cachep, &cache_chain, next) {
1162 struct array_cache *nc;
1163 struct array_cache *shared;
1164 struct array_cache **alien;
1166 nc = alloc_arraycache(node, cachep->limit,
1167 cachep->batchcount);
1168 if (!nc)
1169 goto bad;
1170 shared = alloc_arraycache(node,
1171 cachep->shared * cachep->batchcount,
1172 0xbaadf00d);
1173 if (!shared)
1174 goto bad;
1176 alien = alloc_alien_cache(node, cachep->limit);
1177 if (!alien)
1178 goto bad;
1179 cachep->array[cpu] = nc;
1180 l3 = cachep->nodelists[node];
1181 BUG_ON(!l3);
1183 spin_lock_irq(&l3->list_lock);
1184 if (!l3->shared) {
1185 /*
1186 * We are serialised from CPU_DEAD or
1187 * CPU_UP_CANCELLED by the cpucontrol lock
1188 */
1189 l3->shared = shared;
1190 shared = NULL;
1192 #ifdef CONFIG_NUMA
1193 if (!l3->alien) {
1194 l3->alien = alien;
1195 alien = NULL;
1197 #endif
1198 spin_unlock_irq(&l3->list_lock);
1199 kfree(shared);
1200 free_alien_cache(alien);
1202 mutex_unlock(&cache_chain_mutex);
1203 break;
1204 case CPU_ONLINE:
1205 start_cpu_timer(cpu);
1206 break;
1207 #ifdef CONFIG_HOTPLUG_CPU
1208 case CPU_DEAD:
1209 /*
1210 * Even if all the cpus of a node are down, we don't free the
1211 * kmem_list3 of any cache. This to avoid a race between
1212 * cpu_down, and a kmalloc allocation from another cpu for
1213 * memory from the node of the cpu going down. The list3
1214 * structure is usually allocated from kmem_cache_create() and
1215 * gets destroyed at kmem_cache_destroy().
1216 */
1217 /* fall thru */
1218 case CPU_UP_CANCELED:
1219 mutex_lock(&cache_chain_mutex);
1220 list_for_each_entry(cachep, &cache_chain, next) {
1221 struct array_cache *nc;
1222 struct array_cache *shared;
1223 struct array_cache **alien;
1224 cpumask_t mask;
1226 mask = node_to_cpumask(node);
1227 /* cpu is dead; no one can alloc from it. */
1228 nc = cachep->array[cpu];
1229 cachep->array[cpu] = NULL;
1230 l3 = cachep->nodelists[node];
1232 if (!l3)
1233 goto free_array_cache;
1235 spin_lock_irq(&l3->list_lock);
1237 /* Free limit for this kmem_list3 */
1238 l3->free_limit -= cachep->batchcount;
1239 if (nc)
1240 free_block(cachep, nc->entry, nc->avail, node);
1242 if (!cpus_empty(mask)) {
1243 spin_unlock_irq(&l3->list_lock);
1244 goto free_array_cache;
1247 shared = l3->shared;
1248 if (shared) {
1249 free_block(cachep, l3->shared->entry,
1250 l3->shared->avail, node);
1251 l3->shared = NULL;
1254 alien = l3->alien;
1255 l3->alien = NULL;
1257 spin_unlock_irq(&l3->list_lock);
1259 kfree(shared);
1260 if (alien) {
1261 drain_alien_cache(cachep, alien);
1262 free_alien_cache(alien);
1264 free_array_cache:
1265 kfree(nc);
1267 /*
1268 * In the previous loop, all the objects were freed to
1269 * the respective cache's slabs, now we can go ahead and
1270 * shrink each nodelist to its limit.
1271 */
1272 list_for_each_entry(cachep, &cache_chain, next) {
1273 l3 = cachep->nodelists[node];
1274 if (!l3)
1275 continue;
1276 drain_freelist(cachep, l3, l3->free_objects);
1278 mutex_unlock(&cache_chain_mutex);
1279 break;
1280 #endif
1282 return NOTIFY_OK;
1283 bad:
1284 mutex_unlock(&cache_chain_mutex);
1285 return NOTIFY_BAD;
1288 static struct notifier_block __cpuinitdata cpucache_notifier = {
1289 &cpuup_callback, NULL, 0
1290 };
1292 /*
1293 * swap the static kmem_list3 with kmalloced memory
1294 */
1295 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1296 int nodeid)
1298 struct kmem_list3 *ptr;
1300 BUG_ON(cachep->nodelists[nodeid] != list);
1301 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1302 BUG_ON(!ptr);
1304 local_irq_disable();
1305 memcpy(ptr, list, sizeof(struct kmem_list3));
1306 /*
1307 * Do not assume that spinlocks can be initialized via memcpy:
1308 */
1309 spin_lock_init(&ptr->list_lock);
1311 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1312 cachep->nodelists[nodeid] = ptr;
1313 local_irq_enable();
1316 /*
1317 * Initialisation. Called after the page allocator have been initialised and
1318 * before smp_init().
1319 */
1320 void __init kmem_cache_init(void)
1322 size_t left_over;
1323 struct cache_sizes *sizes;
1324 struct cache_names *names;
1325 int i;
1326 int order;
1328 for (i = 0; i < NUM_INIT_LISTS; i++) {
1329 kmem_list3_init(&initkmem_list3[i]);
1330 if (i < MAX_NUMNODES)
1331 cache_cache.nodelists[i] = NULL;
1334 /*
1335 * Fragmentation resistance on low memory - only use bigger
1336 * page orders on machines with more than 32MB of memory.
1337 */
1338 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1339 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1341 /* Bootstrap is tricky, because several objects are allocated
1342 * from caches that do not exist yet:
1343 * 1) initialize the cache_cache cache: it contains the struct
1344 * kmem_cache structures of all caches, except cache_cache itself:
1345 * cache_cache is statically allocated.
1346 * Initially an __init data area is used for the head array and the
1347 * kmem_list3 structures, it's replaced with a kmalloc allocated
1348 * array at the end of the bootstrap.
1349 * 2) Create the first kmalloc cache.
1350 * The struct kmem_cache for the new cache is allocated normally.
1351 * An __init data area is used for the head array.
1352 * 3) Create the remaining kmalloc caches, with minimally sized
1353 * head arrays.
1354 * 4) Replace the __init data head arrays for cache_cache and the first
1355 * kmalloc cache with kmalloc allocated arrays.
1356 * 5) Replace the __init data for kmem_list3 for cache_cache and
1357 * the other cache's with kmalloc allocated memory.
1358 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1359 */
1361 /* 1) create the cache_cache */
1362 INIT_LIST_HEAD(&cache_chain);
1363 list_add(&cache_cache.next, &cache_chain);
1364 cache_cache.colour_off = cache_line_size();
1365 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1366 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1368 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1369 cache_line_size());
1371 for (order = 0; order < MAX_ORDER; order++) {
1372 cache_estimate(order, cache_cache.buffer_size,
1373 cache_line_size(), 0, &left_over, &cache_cache.num);
1374 if (cache_cache.num)
1375 break;
1377 BUG_ON(!cache_cache.num);
1378 cache_cache.gfporder = order;
1379 cache_cache.colour = left_over / cache_cache.colour_off;
1380 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1381 sizeof(struct slab), cache_line_size());
1383 /* 2+3) create the kmalloc caches */
1384 sizes = malloc_sizes;
1385 names = cache_names;
1387 /*
1388 * Initialize the caches that provide memory for the array cache and the
1389 * kmem_list3 structures first. Without this, further allocations will
1390 * bug.
1391 */
1393 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1394 sizes[INDEX_AC].cs_size,
1395 ARCH_KMALLOC_MINALIGN,
1396 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1397 NULL, NULL);
1399 if (INDEX_AC != INDEX_L3) {
1400 sizes[INDEX_L3].cs_cachep =
1401 kmem_cache_create(names[INDEX_L3].name,
1402 sizes[INDEX_L3].cs_size,
1403 ARCH_KMALLOC_MINALIGN,
1404 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1405 NULL, NULL);
1408 slab_early_init = 0;
1410 while (sizes->cs_size != ULONG_MAX) {
1411 /*
1412 * For performance, all the general caches are L1 aligned.
1413 * This should be particularly beneficial on SMP boxes, as it
1414 * eliminates "false sharing".
1415 * Note for systems short on memory removing the alignment will
1416 * allow tighter packing of the smaller caches.
1417 */
1418 if (!sizes->cs_cachep) {
1419 sizes->cs_cachep = kmem_cache_create(names->name,
1420 sizes->cs_size,
1421 ARCH_KMALLOC_MINALIGN,
1422 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1423 NULL, NULL);
1425 init_lock_keys(sizes);
1427 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1428 sizes->cs_size,
1429 ARCH_KMALLOC_MINALIGN,
1430 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1431 SLAB_PANIC,
1432 NULL, NULL);
1433 sizes++;
1434 names++;
1436 /* 4) Replace the bootstrap head arrays */
1438 struct array_cache *ptr;
1440 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1442 local_irq_disable();
1443 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1444 memcpy(ptr, cpu_cache_get(&cache_cache),
1445 sizeof(struct arraycache_init));
1446 /*
1447 * Do not assume that spinlocks can be initialized via memcpy:
1448 */
1449 spin_lock_init(&ptr->lock);
1451 cache_cache.array[smp_processor_id()] = ptr;
1452 local_irq_enable();
1454 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1456 local_irq_disable();
1457 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1458 != &initarray_generic.cache);
1459 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1460 sizeof(struct arraycache_init));
1461 /*
1462 * Do not assume that spinlocks can be initialized via memcpy:
1463 */
1464 spin_lock_init(&ptr->lock);
1466 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1467 ptr;
1468 local_irq_enable();
1470 /* 5) Replace the bootstrap kmem_list3's */
1472 int node;
1473 /* Replace the static kmem_list3 structures for the boot cpu */
1474 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1475 numa_node_id());
1477 for_each_online_node(node) {
1478 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1479 &initkmem_list3[SIZE_AC + node], node);
1481 if (INDEX_AC != INDEX_L3) {
1482 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1483 &initkmem_list3[SIZE_L3 + node],
1484 node);
1489 /* 6) resize the head arrays to their final sizes */
1491 struct kmem_cache *cachep;
1492 mutex_lock(&cache_chain_mutex);
1493 list_for_each_entry(cachep, &cache_chain, next)
1494 enable_cpucache(cachep);
1495 mutex_unlock(&cache_chain_mutex);
1498 /* Done! */
1499 g_cpucache_up = FULL;
1501 /*
1502 * Register a cpu startup notifier callback that initializes
1503 * cpu_cache_get for all new cpus
1504 */
1505 register_cpu_notifier(&cpucache_notifier);
1507 /*
1508 * The reap timers are started later, with a module init call: That part
1509 * of the kernel is not yet operational.
1510 */
1513 static int __init cpucache_init(void)
1515 int cpu;
1517 /*
1518 * Register the timers that return unneeded pages to the page allocator
1519 */
1520 for_each_online_cpu(cpu)
1521 start_cpu_timer(cpu);
1522 return 0;
1524 __initcall(cpucache_init);
1526 /*
1527 * Interface to system's page allocator. No need to hold the cache-lock.
1529 * If we requested dmaable memory, we will get it. Even if we
1530 * did not request dmaable memory, we might get it, but that
1531 * would be relatively rare and ignorable.
1532 */
1533 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1535 struct page *page;
1536 int nr_pages;
1537 int i;
1539 #ifndef CONFIG_MMU
1540 /*
1541 * Nommu uses slab's for process anonymous memory allocations, and thus
1542 * requires __GFP_COMP to properly refcount higher order allocations
1543 */
1544 flags |= __GFP_COMP;
1545 #endif
1546 flags |= cachep->gfpflags;
1548 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1549 if (!page)
1550 return NULL;
1552 nr_pages = (1 << cachep->gfporder);
1553 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1554 atomic_add(nr_pages, &slab_reclaim_pages);
1555 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
1556 for (i = 0; i < nr_pages; i++)
1557 __SetPageSlab(page + i);
1558 return page_address(page);
1561 /*
1562 * Interface to system's page release.
1563 */
1564 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1566 unsigned long i = (1 << cachep->gfporder);
1567 struct page *page = virt_to_page(addr);
1568 const unsigned long nr_freed = i;
1570 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
1571 while (i--) {
1572 BUG_ON(!PageSlab(page));
1573 __ClearPageSlab(page);
1574 page++;
1576 if (current->reclaim_state)
1577 current->reclaim_state->reclaimed_slab += nr_freed;
1578 free_pages((unsigned long)addr, cachep->gfporder);
1579 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1580 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1583 static void kmem_rcu_free(struct rcu_head *head)
1585 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1586 struct kmem_cache *cachep = slab_rcu->cachep;
1588 kmem_freepages(cachep, slab_rcu->addr);
1589 if (OFF_SLAB(cachep))
1590 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1593 #if DEBUG
1595 #ifdef CONFIG_DEBUG_PAGEALLOC
1596 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1597 unsigned long caller)
1599 int size = obj_size(cachep);
1601 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1603 if (size < 5 * sizeof(unsigned long))
1604 return;
1606 *addr++ = 0x12345678;
1607 *addr++ = caller;
1608 *addr++ = smp_processor_id();
1609 size -= 3 * sizeof(unsigned long);
1611 unsigned long *sptr = &caller;
1612 unsigned long svalue;
1614 while (!kstack_end(sptr)) {
1615 svalue = *sptr++;
1616 if (kernel_text_address(svalue)) {
1617 *addr++ = svalue;
1618 size -= sizeof(unsigned long);
1619 if (size <= sizeof(unsigned long))
1620 break;
1625 *addr++ = 0x87654321;
1627 #endif
1629 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1631 int size = obj_size(cachep);
1632 addr = &((char *)addr)[obj_offset(cachep)];
1634 memset(addr, val, size);
1635 *(unsigned char *)(addr + size - 1) = POISON_END;
1638 static void dump_line(char *data, int offset, int limit)
1640 int i;
1641 printk(KERN_ERR "%03x:", offset);
1642 for (i = 0; i < limit; i++)
1643 printk(" %02x", (unsigned char)data[offset + i]);
1644 printk("\n");
1646 #endif
1648 #if DEBUG
1650 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1652 int i, size;
1653 char *realobj;
1655 if (cachep->flags & SLAB_RED_ZONE) {
1656 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1657 *dbg_redzone1(cachep, objp),
1658 *dbg_redzone2(cachep, objp));
1661 if (cachep->flags & SLAB_STORE_USER) {
1662 printk(KERN_ERR "Last user: [<%p>]",
1663 *dbg_userword(cachep, objp));
1664 print_symbol("(%s)",
1665 (unsigned long)*dbg_userword(cachep, objp));
1666 printk("\n");
1668 realobj = (char *)objp + obj_offset(cachep);
1669 size = obj_size(cachep);
1670 for (i = 0; i < size && lines; i += 16, lines--) {
1671 int limit;
1672 limit = 16;
1673 if (i + limit > size)
1674 limit = size - i;
1675 dump_line(realobj, i, limit);
1679 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1681 char *realobj;
1682 int size, i;
1683 int lines = 0;
1685 realobj = (char *)objp + obj_offset(cachep);
1686 size = obj_size(cachep);
1688 for (i = 0; i < size; i++) {
1689 char exp = POISON_FREE;
1690 if (i == size - 1)
1691 exp = POISON_END;
1692 if (realobj[i] != exp) {
1693 int limit;
1694 /* Mismatch ! */
1695 /* Print header */
1696 if (lines == 0) {
1697 printk(KERN_ERR
1698 "Slab corruption: start=%p, len=%d\n",
1699 realobj, size);
1700 print_objinfo(cachep, objp, 0);
1702 /* Hexdump the affected line */
1703 i = (i / 16) * 16;
1704 limit = 16;
1705 if (i + limit > size)
1706 limit = size - i;
1707 dump_line(realobj, i, limit);
1708 i += 16;
1709 lines++;
1710 /* Limit to 5 lines */
1711 if (lines > 5)
1712 break;
1715 if (lines != 0) {
1716 /* Print some data about the neighboring objects, if they
1717 * exist:
1718 */
1719 struct slab *slabp = virt_to_slab(objp);
1720 unsigned int objnr;
1722 objnr = obj_to_index(cachep, slabp, objp);
1723 if (objnr) {
1724 objp = index_to_obj(cachep, slabp, objnr - 1);
1725 realobj = (char *)objp + obj_offset(cachep);
1726 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1727 realobj, size);
1728 print_objinfo(cachep, objp, 2);
1730 if (objnr + 1 < cachep->num) {
1731 objp = index_to_obj(cachep, slabp, objnr + 1);
1732 realobj = (char *)objp + obj_offset(cachep);
1733 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1734 realobj, size);
1735 print_objinfo(cachep, objp, 2);
1739 #endif
1741 #if DEBUG
1742 /**
1743 * slab_destroy_objs - destroy a slab and its objects
1744 * @cachep: cache pointer being destroyed
1745 * @slabp: slab pointer being destroyed
1747 * Call the registered destructor for each object in a slab that is being
1748 * destroyed.
1749 */
1750 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1752 int i;
1753 for (i = 0; i < cachep->num; i++) {
1754 void *objp = index_to_obj(cachep, slabp, i);
1756 if (cachep->flags & SLAB_POISON) {
1757 #ifdef CONFIG_DEBUG_PAGEALLOC
1758 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1759 OFF_SLAB(cachep))
1760 kernel_map_pages(virt_to_page(objp),
1761 cachep->buffer_size / PAGE_SIZE, 1);
1762 else
1763 check_poison_obj(cachep, objp);
1764 #else
1765 check_poison_obj(cachep, objp);
1766 #endif
1768 if (cachep->flags & SLAB_RED_ZONE) {
1769 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1770 slab_error(cachep, "start of a freed object "
1771 "was overwritten");
1772 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1773 slab_error(cachep, "end of a freed object "
1774 "was overwritten");
1776 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1777 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1780 #else
1781 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1783 if (cachep->dtor) {
1784 int i;
1785 for (i = 0; i < cachep->num; i++) {
1786 void *objp = index_to_obj(cachep, slabp, i);
1787 (cachep->dtor) (objp, cachep, 0);
1791 #endif
1793 /**
1794 * slab_destroy - destroy and release all objects in a slab
1795 * @cachep: cache pointer being destroyed
1796 * @slabp: slab pointer being destroyed
1798 * Destroy all the objs in a slab, and release the mem back to the system.
1799 * Before calling the slab must have been unlinked from the cache. The
1800 * cache-lock is not held/needed.
1801 */
1802 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1804 void *addr = slabp->s_mem - slabp->colouroff;
1806 slab_destroy_objs(cachep, slabp);
1807 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1808 struct slab_rcu *slab_rcu;
1810 slab_rcu = (struct slab_rcu *)slabp;
1811 slab_rcu->cachep = cachep;
1812 slab_rcu->addr = addr;
1813 call_rcu(&slab_rcu->head, kmem_rcu_free);
1814 } else {
1815 kmem_freepages(cachep, addr);
1816 if (OFF_SLAB(cachep))
1817 kmem_cache_free(cachep->slabp_cache, slabp);
1821 /*
1822 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1823 * size of kmem_list3.
1824 */
1825 static void set_up_list3s(struct kmem_cache *cachep, int index)
1827 int node;
1829 for_each_online_node(node) {
1830 cachep->nodelists[node] = &initkmem_list3[index + node];
1831 cachep->nodelists[node]->next_reap = jiffies +
1832 REAPTIMEOUT_LIST3 +
1833 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1837 /**
1838 * calculate_slab_order - calculate size (page order) of slabs
1839 * @cachep: pointer to the cache that is being created
1840 * @size: size of objects to be created in this cache.
1841 * @align: required alignment for the objects.
1842 * @flags: slab allocation flags
1844 * Also calculates the number of objects per slab.
1846 * This could be made much more intelligent. For now, try to avoid using
1847 * high order pages for slabs. When the gfp() functions are more friendly
1848 * towards high-order requests, this should be changed.
1849 */
1850 static size_t calculate_slab_order(struct kmem_cache *cachep,
1851 size_t size, size_t align, unsigned long flags)
1853 unsigned long offslab_limit;
1854 size_t left_over = 0;
1855 int gfporder;
1857 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1858 unsigned int num;
1859 size_t remainder;
1861 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1862 if (!num)
1863 continue;
1865 if (flags & CFLGS_OFF_SLAB) {
1866 /*
1867 * Max number of objs-per-slab for caches which
1868 * use off-slab slabs. Needed to avoid a possible
1869 * looping condition in cache_grow().
1870 */
1871 offslab_limit = size - sizeof(struct slab);
1872 offslab_limit /= sizeof(kmem_bufctl_t);
1874 if (num > offslab_limit)
1875 break;
1878 /* Found something acceptable - save it away */
1879 cachep->num = num;
1880 cachep->gfporder = gfporder;
1881 left_over = remainder;
1883 /*
1884 * A VFS-reclaimable slab tends to have most allocations
1885 * as GFP_NOFS and we really don't want to have to be allocating
1886 * higher-order pages when we are unable to shrink dcache.
1887 */
1888 if (flags & SLAB_RECLAIM_ACCOUNT)
1889 break;
1891 /*
1892 * Large number of objects is good, but very large slabs are
1893 * currently bad for the gfp()s.
1894 */
1895 if (gfporder >= slab_break_gfp_order)
1896 break;
1898 /*
1899 * Acceptable internal fragmentation?
1900 */
1901 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1902 break;
1904 return left_over;
1907 static void setup_cpu_cache(struct kmem_cache *cachep)
1909 if (g_cpucache_up == FULL) {
1910 enable_cpucache(cachep);
1911 return;
1913 if (g_cpucache_up == NONE) {
1914 /*
1915 * Note: the first kmem_cache_create must create the cache
1916 * that's used by kmalloc(24), otherwise the creation of
1917 * further caches will BUG().
1918 */
1919 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1921 /*
1922 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1923 * the first cache, then we need to set up all its list3s,
1924 * otherwise the creation of further caches will BUG().
1925 */
1926 set_up_list3s(cachep, SIZE_AC);
1927 if (INDEX_AC == INDEX_L3)
1928 g_cpucache_up = PARTIAL_L3;
1929 else
1930 g_cpucache_up = PARTIAL_AC;
1931 } else {
1932 cachep->array[smp_processor_id()] =
1933 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1935 if (g_cpucache_up == PARTIAL_AC) {
1936 set_up_list3s(cachep, SIZE_L3);
1937 g_cpucache_up = PARTIAL_L3;
1938 } else {
1939 int node;
1940 for_each_online_node(node) {
1941 cachep->nodelists[node] =
1942 kmalloc_node(sizeof(struct kmem_list3),
1943 GFP_KERNEL, node);
1944 BUG_ON(!cachep->nodelists[node]);
1945 kmem_list3_init(cachep->nodelists[node]);
1949 cachep->nodelists[numa_node_id()]->next_reap =
1950 jiffies + REAPTIMEOUT_LIST3 +
1951 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1953 cpu_cache_get(cachep)->avail = 0;
1954 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1955 cpu_cache_get(cachep)->batchcount = 1;
1956 cpu_cache_get(cachep)->touched = 0;
1957 cachep->batchcount = 1;
1958 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1961 /**
1962 * kmem_cache_create - Create a cache.
1963 * @name: A string which is used in /proc/slabinfo to identify this cache.
1964 * @size: The size of objects to be created in this cache.
1965 * @align: The required alignment for the objects.
1966 * @flags: SLAB flags
1967 * @ctor: A constructor for the objects.
1968 * @dtor: A destructor for the objects.
1970 * Returns a ptr to the cache on success, NULL on failure.
1971 * Cannot be called within a int, but can be interrupted.
1972 * The @ctor is run when new pages are allocated by the cache
1973 * and the @dtor is run before the pages are handed back.
1975 * @name must be valid until the cache is destroyed. This implies that
1976 * the module calling this has to destroy the cache before getting unloaded.
1978 * The flags are
1980 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1981 * to catch references to uninitialised memory.
1983 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1984 * for buffer overruns.
1986 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1987 * cacheline. This can be beneficial if you're counting cycles as closely
1988 * as davem.
1989 */
1990 struct kmem_cache *
1991 kmem_cache_create (const char *name, size_t size, size_t align,
1992 unsigned long flags,
1993 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1994 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1996 size_t left_over, slab_size, ralign;
1997 struct kmem_cache *cachep = NULL, *pc;
1999 /*
2000 * Sanity checks... these are all serious usage bugs.
2001 */
2002 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2003 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2004 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2005 name);
2006 BUG();
2009 /*
2010 * Prevent CPUs from coming and going.
2011 * lock_cpu_hotplug() nests outside cache_chain_mutex
2012 */
2013 lock_cpu_hotplug();
2015 mutex_lock(&cache_chain_mutex);
2017 list_for_each_entry(pc, &cache_chain, next) {
2018 mm_segment_t old_fs = get_fs();
2019 char tmp;
2020 int res;
2022 /*
2023 * This happens when the module gets unloaded and doesn't
2024 * destroy its slab cache and no-one else reuses the vmalloc
2025 * area of the module. Print a warning.
2026 */
2027 set_fs(KERNEL_DS);
2028 res = __get_user(tmp, pc->name);
2029 set_fs(old_fs);
2030 if (res) {
2031 printk("SLAB: cache with size %d has lost its name\n",
2032 pc->buffer_size);
2033 continue;
2036 if (!strcmp(pc->name, name)) {
2037 printk("kmem_cache_create: duplicate cache %s\n", name);
2038 dump_stack();
2039 goto oops;
2043 #if DEBUG
2044 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2045 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2046 /* No constructor, but inital state check requested */
2047 printk(KERN_ERR "%s: No con, but init state check "
2048 "requested - %s\n", __FUNCTION__, name);
2049 flags &= ~SLAB_DEBUG_INITIAL;
2051 #if FORCED_DEBUG
2052 /*
2053 * Enable redzoning and last user accounting, except for caches with
2054 * large objects, if the increased size would increase the object size
2055 * above the next power of two: caches with object sizes just above a
2056 * power of two have a significant amount of internal fragmentation.
2057 */
2058 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2059 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2060 if (!(flags & SLAB_DESTROY_BY_RCU))
2061 flags |= SLAB_POISON;
2062 #endif
2063 if (flags & SLAB_DESTROY_BY_RCU)
2064 BUG_ON(flags & SLAB_POISON);
2065 #endif
2066 if (flags & SLAB_DESTROY_BY_RCU)
2067 BUG_ON(dtor);
2069 /*
2070 * Always checks flags, a caller might be expecting debug support which
2071 * isn't available.
2072 */
2073 BUG_ON(flags & ~CREATE_MASK);
2075 /*
2076 * Check that size is in terms of words. This is needed to avoid
2077 * unaligned accesses for some archs when redzoning is used, and makes
2078 * sure any on-slab bufctl's are also correctly aligned.
2079 */
2080 if (size & (BYTES_PER_WORD - 1)) {
2081 size += (BYTES_PER_WORD - 1);
2082 size &= ~(BYTES_PER_WORD - 1);
2085 /* calculate the final buffer alignment: */
2087 /* 1) arch recommendation: can be overridden for debug */
2088 if (flags & SLAB_HWCACHE_ALIGN) {
2089 /*
2090 * Default alignment: as specified by the arch code. Except if
2091 * an object is really small, then squeeze multiple objects into
2092 * one cacheline.
2093 */
2094 ralign = cache_line_size();
2095 while (size <= ralign / 2)
2096 ralign /= 2;
2097 } else {
2098 ralign = BYTES_PER_WORD;
2100 /* 2) arch mandated alignment: disables debug if necessary */
2101 if (ralign < ARCH_SLAB_MINALIGN) {
2102 ralign = ARCH_SLAB_MINALIGN;
2103 if (ralign > BYTES_PER_WORD)
2104 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2106 /* 3) caller mandated alignment: disables debug if necessary */
2107 if (ralign < align) {
2108 ralign = align;
2109 if (ralign > BYTES_PER_WORD)
2110 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2112 /*
2113 * 4) Store it. Note that the debug code below can reduce
2114 * the alignment to BYTES_PER_WORD.
2115 */
2116 align = ralign;
2118 /* Get cache's description obj. */
2119 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2120 if (!cachep)
2121 goto oops;
2123 #if DEBUG
2124 cachep->obj_size = size;
2126 if (flags & SLAB_RED_ZONE) {
2127 /* redzoning only works with word aligned caches */
2128 align = BYTES_PER_WORD;
2130 /* add space for red zone words */
2131 cachep->obj_offset += BYTES_PER_WORD;
2132 size += 2 * BYTES_PER_WORD;
2134 if (flags & SLAB_STORE_USER) {
2135 /* user store requires word alignment and
2136 * one word storage behind the end of the real
2137 * object.
2138 */
2139 align = BYTES_PER_WORD;
2140 size += BYTES_PER_WORD;
2142 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2143 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2144 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2145 cachep->obj_offset += PAGE_SIZE - size;
2146 size = PAGE_SIZE;
2148 #endif
2149 #endif
2151 /*
2152 * Determine if the slab management is 'on' or 'off' slab.
2153 * (bootstrapping cannot cope with offslab caches so don't do
2154 * it too early on.)
2155 */
2156 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2157 /*
2158 * Size is large, assume best to place the slab management obj
2159 * off-slab (should allow better packing of objs).
2160 */
2161 flags |= CFLGS_OFF_SLAB;
2163 size = ALIGN(size, align);
2165 left_over = calculate_slab_order(cachep, size, align, flags);
2167 if (!cachep->num) {
2168 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2169 kmem_cache_free(&cache_cache, cachep);
2170 cachep = NULL;
2171 goto oops;
2173 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2174 + sizeof(struct slab), align);
2176 /*
2177 * If the slab has been placed off-slab, and we have enough space then
2178 * move it on-slab. This is at the expense of any extra colouring.
2179 */
2180 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2181 flags &= ~CFLGS_OFF_SLAB;
2182 left_over -= slab_size;
2185 if (flags & CFLGS_OFF_SLAB) {
2186 /* really off slab. No need for manual alignment */
2187 slab_size =
2188 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2191 cachep->colour_off = cache_line_size();
2192 /* Offset must be a multiple of the alignment. */
2193 if (cachep->colour_off < align)
2194 cachep->colour_off = align;
2195 cachep->colour = left_over / cachep->colour_off;
2196 cachep->slab_size = slab_size;
2197 cachep->flags = flags;
2198 cachep->gfpflags = 0;
2199 if (flags & SLAB_CACHE_DMA)
2200 cachep->gfpflags |= GFP_DMA;
2201 cachep->buffer_size = size;
2203 if (flags & CFLGS_OFF_SLAB)
2204 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2205 cachep->ctor = ctor;
2206 cachep->dtor = dtor;
2207 cachep->name = name;
2210 setup_cpu_cache(cachep);
2212 /* cache setup completed, link it into the list */
2213 list_add(&cachep->next, &cache_chain);
2214 oops:
2215 if (!cachep && (flags & SLAB_PANIC))
2216 panic("kmem_cache_create(): failed to create slab `%s'\n",
2217 name);
2218 mutex_unlock(&cache_chain_mutex);
2219 unlock_cpu_hotplug();
2220 return cachep;
2222 EXPORT_SYMBOL(kmem_cache_create);
2224 #if DEBUG
2225 static void check_irq_off(void)
2227 BUG_ON(!irqs_disabled());
2230 static void check_irq_on(void)
2232 BUG_ON(irqs_disabled());
2235 static void check_spinlock_acquired(struct kmem_cache *cachep)
2237 #ifdef CONFIG_SMP
2238 check_irq_off();
2239 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2240 #endif
2243 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2245 #ifdef CONFIG_SMP
2246 check_irq_off();
2247 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2248 #endif
2251 #else
2252 #define check_irq_off() do { } while(0)
2253 #define check_irq_on() do { } while(0)
2254 #define check_spinlock_acquired(x) do { } while(0)
2255 #define check_spinlock_acquired_node(x, y) do { } while(0)
2256 #endif
2258 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2259 struct array_cache *ac,
2260 int force, int node);
2262 static void do_drain(void *arg)
2264 struct kmem_cache *cachep = arg;
2265 struct array_cache *ac;
2266 int node = numa_node_id();
2268 check_irq_off();
2269 ac = cpu_cache_get(cachep);
2270 spin_lock(&cachep->nodelists[node]->list_lock);
2271 free_block(cachep, ac->entry, ac->avail, node);
2272 spin_unlock(&cachep->nodelists[node]->list_lock);
2273 ac->avail = 0;
2276 static void drain_cpu_caches(struct kmem_cache *cachep)
2278 struct kmem_list3 *l3;
2279 int node;
2281 on_each_cpu(do_drain, cachep, 1, 1);
2282 check_irq_on();
2283 for_each_online_node(node) {
2284 l3 = cachep->nodelists[node];
2285 if (l3 && l3->alien)
2286 drain_alien_cache(cachep, l3->alien);
2289 for_each_online_node(node) {
2290 l3 = cachep->nodelists[node];
2291 if (l3)
2292 drain_array(cachep, l3, l3->shared, 1, node);
2296 /*
2297 * Remove slabs from the list of free slabs.
2298 * Specify the number of slabs to drain in tofree.
2300 * Returns the actual number of slabs released.
2301 */
2302 static int drain_freelist(struct kmem_cache *cache,
2303 struct kmem_list3 *l3, int tofree)
2305 struct list_head *p;
2306 int nr_freed;
2307 struct slab *slabp;
2309 nr_freed = 0;
2310 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2312 spin_lock_irq(&l3->list_lock);
2313 p = l3->slabs_free.prev;
2314 if (p == &l3->slabs_free) {
2315 spin_unlock_irq(&l3->list_lock);
2316 goto out;
2319 slabp = list_entry(p, struct slab, list);
2320 #if DEBUG
2321 BUG_ON(slabp->inuse);
2322 #endif
2323 list_del(&slabp->list);
2324 /*
2325 * Safe to drop the lock. The slab is no longer linked
2326 * to the cache.
2327 */
2328 l3->free_objects -= cache->num;
2329 spin_unlock_irq(&l3->list_lock);
2330 slab_destroy(cache, slabp);
2331 nr_freed++;
2333 out:
2334 return nr_freed;
2337 static int __cache_shrink(struct kmem_cache *cachep)
2339 int ret = 0, i = 0;
2340 struct kmem_list3 *l3;
2342 drain_cpu_caches(cachep);
2344 check_irq_on();
2345 for_each_online_node(i) {
2346 l3 = cachep->nodelists[i];
2347 if (!l3)
2348 continue;
2350 drain_freelist(cachep, l3, l3->free_objects);
2352 ret += !list_empty(&l3->slabs_full) ||
2353 !list_empty(&l3->slabs_partial);
2355 return (ret ? 1 : 0);
2358 /**
2359 * kmem_cache_shrink - Shrink a cache.
2360 * @cachep: The cache to shrink.
2362 * Releases as many slabs as possible for a cache.
2363 * To help debugging, a zero exit status indicates all slabs were released.
2364 */
2365 int kmem_cache_shrink(struct kmem_cache *cachep)
2367 BUG_ON(!cachep || in_interrupt());
2369 return __cache_shrink(cachep);
2371 EXPORT_SYMBOL(kmem_cache_shrink);
2373 /**
2374 * kmem_cache_destroy - delete a cache
2375 * @cachep: the cache to destroy
2377 * Remove a struct kmem_cache object from the slab cache.
2378 * Returns 0 on success.
2380 * It is expected this function will be called by a module when it is
2381 * unloaded. This will remove the cache completely, and avoid a duplicate
2382 * cache being allocated each time a module is loaded and unloaded, if the
2383 * module doesn't have persistent in-kernel storage across loads and unloads.
2385 * The cache must be empty before calling this function.
2387 * The caller must guarantee that noone will allocate memory from the cache
2388 * during the kmem_cache_destroy().
2389 */
2390 int kmem_cache_destroy(struct kmem_cache *cachep)
2392 int i;
2393 struct kmem_list3 *l3;
2395 BUG_ON(!cachep || in_interrupt());
2397 /* Don't let CPUs to come and go */
2398 lock_cpu_hotplug();
2400 /* Find the cache in the chain of caches. */
2401 mutex_lock(&cache_chain_mutex);
2402 /*
2403 * the chain is never empty, cache_cache is never destroyed
2404 */
2405 list_del(&cachep->next);
2406 mutex_unlock(&cache_chain_mutex);
2408 if (__cache_shrink(cachep)) {
2409 slab_error(cachep, "Can't free all objects");
2410 mutex_lock(&cache_chain_mutex);
2411 list_add(&cachep->next, &cache_chain);
2412 mutex_unlock(&cache_chain_mutex);
2413 unlock_cpu_hotplug();
2414 return 1;
2417 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2418 synchronize_rcu();
2420 for_each_online_cpu(i)
2421 kfree(cachep->array[i]);
2423 /* NUMA: free the list3 structures */
2424 for_each_online_node(i) {
2425 l3 = cachep->nodelists[i];
2426 if (l3) {
2427 kfree(l3->shared);
2428 free_alien_cache(l3->alien);
2429 kfree(l3);
2432 kmem_cache_free(&cache_cache, cachep);
2433 unlock_cpu_hotplug();
2434 return 0;
2436 EXPORT_SYMBOL(kmem_cache_destroy);
2438 /* Get the memory for a slab management obj. */
2439 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2440 int colour_off, gfp_t local_flags,
2441 int nodeid)
2443 struct slab *slabp;
2445 if (OFF_SLAB(cachep)) {
2446 /* Slab management obj is off-slab. */
2447 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2448 local_flags, nodeid);
2449 if (!slabp)
2450 return NULL;
2451 } else {
2452 slabp = objp + colour_off;
2453 colour_off += cachep->slab_size;
2455 slabp->inuse = 0;
2456 slabp->colouroff = colour_off;
2457 slabp->s_mem = objp + colour_off;
2458 slabp->nodeid = nodeid;
2459 return slabp;
2462 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2464 return (kmem_bufctl_t *) (slabp + 1);
2467 static void cache_init_objs(struct kmem_cache *cachep,
2468 struct slab *slabp, unsigned long ctor_flags)
2470 int i;
2472 for (i = 0; i < cachep->num; i++) {
2473 void *objp = index_to_obj(cachep, slabp, i);
2474 #if DEBUG
2475 /* need to poison the objs? */
2476 if (cachep->flags & SLAB_POISON)
2477 poison_obj(cachep, objp, POISON_FREE);
2478 if (cachep->flags & SLAB_STORE_USER)
2479 *dbg_userword(cachep, objp) = NULL;
2481 if (cachep->flags & SLAB_RED_ZONE) {
2482 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2483 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2485 /*
2486 * Constructors are not allowed to allocate memory from the same
2487 * cache which they are a constructor for. Otherwise, deadlock.
2488 * They must also be threaded.
2489 */
2490 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2491 cachep->ctor(objp + obj_offset(cachep), cachep,
2492 ctor_flags);
2494 if (cachep->flags & SLAB_RED_ZONE) {
2495 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2496 slab_error(cachep, "constructor overwrote the"
2497 " end of an object");
2498 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2499 slab_error(cachep, "constructor overwrote the"
2500 " start of an object");
2502 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2503 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2504 kernel_map_pages(virt_to_page(objp),
2505 cachep->buffer_size / PAGE_SIZE, 0);
2506 #else
2507 if (cachep->ctor)
2508 cachep->ctor(objp, cachep, ctor_flags);
2509 #endif
2510 slab_bufctl(slabp)[i] = i + 1;
2512 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2513 slabp->free = 0;
2516 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2518 if (flags & SLAB_DMA)
2519 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2520 else
2521 BUG_ON(cachep->gfpflags & GFP_DMA);
2524 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2525 int nodeid)
2527 void *objp = index_to_obj(cachep, slabp, slabp->free);
2528 kmem_bufctl_t next;
2530 slabp->inuse++;
2531 next = slab_bufctl(slabp)[slabp->free];
2532 #if DEBUG
2533 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2534 WARN_ON(slabp->nodeid != nodeid);
2535 #endif
2536 slabp->free = next;
2538 return objp;
2541 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2542 void *objp, int nodeid)
2544 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2546 #if DEBUG
2547 /* Verify that the slab belongs to the intended node */
2548 WARN_ON(slabp->nodeid != nodeid);
2550 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2551 printk(KERN_ERR "slab: double free detected in cache "
2552 "'%s', objp %p\n", cachep->name, objp);
2553 BUG();
2555 #endif
2556 slab_bufctl(slabp)[objnr] = slabp->free;
2557 slabp->free = objnr;
2558 slabp->inuse--;
2561 /*
2562 * Map pages beginning at addr to the given cache and slab. This is required
2563 * for the slab allocator to be able to lookup the cache and slab of a
2564 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2565 */
2566 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2567 void *addr)
2569 int nr_pages;
2570 struct page *page;
2572 page = virt_to_page(addr);
2574 nr_pages = 1;
2575 if (likely(!PageCompound(page)))
2576 nr_pages <<= cache->gfporder;
2578 do {
2579 page_set_cache(page, cache);
2580 page_set_slab(page, slab);
2581 page++;
2582 } while (--nr_pages);
2585 /*
2586 * Grow (by 1) the number of slabs within a cache. This is called by
2587 * kmem_cache_alloc() when there are no active objs left in a cache.
2588 */
2589 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2591 struct slab *slabp;
2592 void *objp;
2593 size_t offset;
2594 gfp_t local_flags;
2595 unsigned long ctor_flags;
2596 struct kmem_list3 *l3;
2598 /*
2599 * Be lazy and only check for valid flags here, keeping it out of the
2600 * critical path in kmem_cache_alloc().
2601 */
2602 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2603 if (flags & SLAB_NO_GROW)
2604 return 0;
2606 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2607 local_flags = (flags & SLAB_LEVEL_MASK);
2608 if (!(local_flags & __GFP_WAIT))
2609 /*
2610 * Not allowed to sleep. Need to tell a constructor about
2611 * this - it might need to know...
2612 */
2613 ctor_flags |= SLAB_CTOR_ATOMIC;
2615 /* Take the l3 list lock to change the colour_next on this node */
2616 check_irq_off();
2617 l3 = cachep->nodelists[nodeid];
2618 spin_lock(&l3->list_lock);
2620 /* Get colour for the slab, and cal the next value. */
2621 offset = l3->colour_next;
2622 l3->colour_next++;
2623 if (l3->colour_next >= cachep->colour)
2624 l3->colour_next = 0;
2625 spin_unlock(&l3->list_lock);
2627 offset *= cachep->colour_off;
2629 if (local_flags & __GFP_WAIT)
2630 local_irq_enable();
2632 /*
2633 * The test for missing atomic flag is performed here, rather than
2634 * the more obvious place, simply to reduce the critical path length
2635 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2636 * will eventually be caught here (where it matters).
2637 */
2638 kmem_flagcheck(cachep, flags);
2640 /*
2641 * Get mem for the objs. Attempt to allocate a physical page from
2642 * 'nodeid'.
2643 */
2644 objp = kmem_getpages(cachep, flags, nodeid);
2645 if (!objp)
2646 goto failed;
2648 /* Get slab management. */
2649 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2650 if (!slabp)
2651 goto opps1;
2653 slabp->nodeid = nodeid;
2654 slab_map_pages(cachep, slabp, objp);
2656 cache_init_objs(cachep, slabp, ctor_flags);
2658 if (local_flags & __GFP_WAIT)
2659 local_irq_disable();
2660 check_irq_off();
2661 spin_lock(&l3->list_lock);
2663 /* Make slab active. */
2664 list_add_tail(&slabp->list, &(l3->slabs_free));
2665 STATS_INC_GROWN(cachep);
2666 l3->free_objects += cachep->num;
2667 spin_unlock(&l3->list_lock);
2668 return 1;
2669 opps1:
2670 kmem_freepages(cachep, objp);
2671 failed:
2672 if (local_flags & __GFP_WAIT)
2673 local_irq_disable();
2674 return 0;
2677 #if DEBUG
2679 /*
2680 * Perform extra freeing checks:
2681 * - detect bad pointers.
2682 * - POISON/RED_ZONE checking
2683 * - destructor calls, for caches with POISON+dtor
2684 */
2685 static void kfree_debugcheck(const void *objp)
2687 struct page *page;
2689 if (!virt_addr_valid(objp)) {
2690 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2691 (unsigned long)objp);
2692 BUG();
2694 page = virt_to_page(objp);
2695 if (!PageSlab(page)) {
2696 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2697 (unsigned long)objp);
2698 BUG();
2702 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2704 unsigned long redzone1, redzone2;
2706 redzone1 = *dbg_redzone1(cache, obj);
2707 redzone2 = *dbg_redzone2(cache, obj);
2709 /*
2710 * Redzone is ok.
2711 */
2712 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2713 return;
2715 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2716 slab_error(cache, "double free detected");
2717 else
2718 slab_error(cache, "memory outside object was overwritten");
2720 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2721 obj, redzone1, redzone2);
2724 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2725 void *caller)
2727 struct page *page;
2728 unsigned int objnr;
2729 struct slab *slabp;
2731 objp -= obj_offset(cachep);
2732 kfree_debugcheck(objp);
2733 page = virt_to_page(objp);
2735 slabp = page_get_slab(page);
2737 if (cachep->flags & SLAB_RED_ZONE) {
2738 verify_redzone_free(cachep, objp);
2739 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2740 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2742 if (cachep->flags & SLAB_STORE_USER)
2743 *dbg_userword(cachep, objp) = caller;
2745 objnr = obj_to_index(cachep, slabp, objp);
2747 BUG_ON(objnr >= cachep->num);
2748 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2750 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2751 /*
2752 * Need to call the slab's constructor so the caller can
2753 * perform a verify of its state (debugging). Called without
2754 * the cache-lock held.
2755 */
2756 cachep->ctor(objp + obj_offset(cachep),
2757 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2759 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2760 /* we want to cache poison the object,
2761 * call the destruction callback
2762 */
2763 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2765 #ifdef CONFIG_DEBUG_SLAB_LEAK
2766 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2767 #endif
2768 if (cachep->flags & SLAB_POISON) {
2769 #ifdef CONFIG_DEBUG_PAGEALLOC
2770 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2771 store_stackinfo(cachep, objp, (unsigned long)caller);
2772 kernel_map_pages(virt_to_page(objp),
2773 cachep->buffer_size / PAGE_SIZE, 0);
2774 } else {
2775 poison_obj(cachep, objp, POISON_FREE);
2777 #else
2778 poison_obj(cachep, objp, POISON_FREE);
2779 #endif
2781 return objp;
2784 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2786 kmem_bufctl_t i;
2787 int entries = 0;
2789 /* Check slab's freelist to see if this obj is there. */
2790 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2791 entries++;
2792 if (entries > cachep->num || i >= cachep->num)
2793 goto bad;
2795 if (entries != cachep->num - slabp->inuse) {
2796 bad:
2797 printk(KERN_ERR "slab: Internal list corruption detected in "
2798 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2799 cachep->name, cachep->num, slabp, slabp->inuse);
2800 for (i = 0;
2801 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2802 i++) {
2803 if (i % 16 == 0)
2804 printk("\n%03x:", i);
2805 printk(" %02x", ((unsigned char *)slabp)[i]);
2807 printk("\n");
2808 BUG();
2811 #else
2812 #define kfree_debugcheck(x) do { } while(0)
2813 #define cache_free_debugcheck(x,objp,z) (objp)
2814 #define check_slabp(x,y) do { } while(0)
2815 #endif
2817 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2819 int batchcount;
2820 struct kmem_list3 *l3;
2821 struct array_cache *ac;
2823 check_irq_off();
2824 ac = cpu_cache_get(cachep);
2825 retry:
2826 batchcount = ac->batchcount;
2827 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2828 /*
2829 * If there was little recent activity on this cache, then
2830 * perform only a partial refill. Otherwise we could generate
2831 * refill bouncing.
2832 */
2833 batchcount = BATCHREFILL_LIMIT;
2835 l3 = cachep->nodelists[numa_node_id()];
2837 BUG_ON(ac->avail > 0 || !l3);
2838 spin_lock(&l3->list_lock);
2840 /* See if we can refill from the shared array */
2841 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2842 goto alloc_done;
2844 while (batchcount > 0) {
2845 struct list_head *entry;
2846 struct slab *slabp;
2847 /* Get slab alloc is to come from. */
2848 entry = l3->slabs_partial.next;
2849 if (entry == &l3->slabs_partial) {
2850 l3->free_touched = 1;
2851 entry = l3->slabs_free.next;
2852 if (entry == &l3->slabs_free)
2853 goto must_grow;
2856 slabp = list_entry(entry, struct slab, list);
2857 check_slabp(cachep, slabp);
2858 check_spinlock_acquired(cachep);
2859 while (slabp->inuse < cachep->num && batchcount--) {
2860 STATS_INC_ALLOCED(cachep);
2861 STATS_INC_ACTIVE(cachep);
2862 STATS_SET_HIGH(cachep);
2864 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2865 numa_node_id());
2867 check_slabp(cachep, slabp);
2869 /* move slabp to correct slabp list: */
2870 list_del(&slabp->list);
2871 if (slabp->free == BUFCTL_END)
2872 list_add(&slabp->list, &l3->slabs_full);
2873 else
2874 list_add(&slabp->list, &l3->slabs_partial);
2877 must_grow:
2878 l3->free_objects -= ac->avail;
2879 alloc_done:
2880 spin_unlock(&l3->list_lock);
2882 if (unlikely(!ac->avail)) {
2883 int x;
2884 x = cache_grow(cachep, flags, numa_node_id());
2886 /* cache_grow can reenable interrupts, then ac could change. */
2887 ac = cpu_cache_get(cachep);
2888 if (!x && ac->avail == 0) /* no objects in sight? abort */
2889 return NULL;
2891 if (!ac->avail) /* objects refilled by interrupt? */
2892 goto retry;
2894 ac->touched = 1;
2895 return ac->entry[--ac->avail];
2898 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2899 gfp_t flags)
2901 might_sleep_if(flags & __GFP_WAIT);
2902 #if DEBUG
2903 kmem_flagcheck(cachep, flags);
2904 #endif
2907 #if DEBUG
2908 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2909 gfp_t flags, void *objp, void *caller)
2911 if (!objp)
2912 return objp;
2913 if (cachep->flags & SLAB_POISON) {
2914 #ifdef CONFIG_DEBUG_PAGEALLOC
2915 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2916 kernel_map_pages(virt_to_page(objp),
2917 cachep->buffer_size / PAGE_SIZE, 1);
2918 else
2919 check_poison_obj(cachep, objp);
2920 #else
2921 check_poison_obj(cachep, objp);
2922 #endif
2923 poison_obj(cachep, objp, POISON_INUSE);
2925 if (cachep->flags & SLAB_STORE_USER)
2926 *dbg_userword(cachep, objp) = caller;
2928 if (cachep->flags & SLAB_RED_ZONE) {
2929 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2930 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2931 slab_error(cachep, "double free, or memory outside"
2932 " object was overwritten");
2933 printk(KERN_ERR
2934 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2935 objp, *dbg_redzone1(cachep, objp),
2936 *dbg_redzone2(cachep, objp));
2938 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2939 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2941 #ifdef CONFIG_DEBUG_SLAB_LEAK
2943 struct slab *slabp;
2944 unsigned objnr;
2946 slabp = page_get_slab(virt_to_page(objp));
2947 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2948 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2950 #endif
2951 objp += obj_offset(cachep);
2952 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2953 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2955 if (!(flags & __GFP_WAIT))
2956 ctor_flags |= SLAB_CTOR_ATOMIC;
2958 cachep->ctor(objp, cachep, ctor_flags);
2960 return objp;
2962 #else
2963 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2964 #endif
2966 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2968 void *objp;
2969 struct array_cache *ac;
2971 #ifdef CONFIG_NUMA
2972 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2973 objp = alternate_node_alloc(cachep, flags);
2974 if (objp != NULL)
2975 return objp;
2977 #endif
2979 check_irq_off();
2980 ac = cpu_cache_get(cachep);
2981 if (likely(ac->avail)) {
2982 STATS_INC_ALLOCHIT(cachep);
2983 ac->touched = 1;
2984 objp = ac->entry[--ac->avail];
2985 } else {
2986 STATS_INC_ALLOCMISS(cachep);
2987 objp = cache_alloc_refill(cachep, flags);
2989 return objp;
2992 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2993 gfp_t flags, void *caller)
2995 unsigned long save_flags;
2996 void *objp;
2998 cache_alloc_debugcheck_before(cachep, flags);
3000 local_irq_save(save_flags);
3001 objp = ____cache_alloc(cachep, flags);
3002 local_irq_restore(save_flags);
3003 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3004 caller);
3005 prefetchw(objp);
3006 return objp;
3009 #ifdef CONFIG_NUMA
3010 /*
3011 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3013 * If we are in_interrupt, then process context, including cpusets and
3014 * mempolicy, may not apply and should not be used for allocation policy.
3015 */
3016 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3018 int nid_alloc, nid_here;
3020 if (in_interrupt())
3021 return NULL;
3022 nid_alloc = nid_here = numa_node_id();
3023 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3024 nid_alloc = cpuset_mem_spread_node();
3025 else if (current->mempolicy)
3026 nid_alloc = slab_node(current->mempolicy);
3027 if (nid_alloc != nid_here)
3028 return __cache_alloc_node(cachep, flags, nid_alloc);
3029 return NULL;
3032 /*
3033 * A interface to enable slab creation on nodeid
3034 */
3035 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3036 int nodeid)
3038 struct list_head *entry;
3039 struct slab *slabp;
3040 struct kmem_list3 *l3;
3041 void *obj;
3042 int x;
3044 l3 = cachep->nodelists[nodeid];
3045 BUG_ON(!l3);
3047 retry:
3048 check_irq_off();
3049 spin_lock(&l3->list_lock);
3050 entry = l3->slabs_partial.next;
3051 if (entry == &l3->slabs_partial) {
3052 l3->free_touched = 1;
3053 entry = l3->slabs_free.next;
3054 if (entry == &l3->slabs_free)
3055 goto must_grow;
3058 slabp = list_entry(entry, struct slab, list);
3059 check_spinlock_acquired_node(cachep, nodeid);
3060 check_slabp(cachep, slabp);
3062 STATS_INC_NODEALLOCS(cachep);
3063 STATS_INC_ACTIVE(cachep);
3064 STATS_SET_HIGH(cachep);
3066 BUG_ON(slabp->inuse == cachep->num);
3068 obj = slab_get_obj(cachep, slabp, nodeid);
3069 check_slabp(cachep, slabp);
3070 l3->free_objects--;
3071 /* move slabp to correct slabp list: */
3072 list_del(&slabp->list);
3074 if (slabp->free == BUFCTL_END)
3075 list_add(&slabp->list, &l3->slabs_full);
3076 else
3077 list_add(&slabp->list, &l3->slabs_partial);
3079 spin_unlock(&l3->list_lock);
3080 goto done;
3082 must_grow:
3083 spin_unlock(&l3->list_lock);
3084 x = cache_grow(cachep, flags, nodeid);
3086 if (!x)
3087 return NULL;
3089 goto retry;
3090 done:
3091 return obj;
3093 #endif
3095 /*
3096 * Caller needs to acquire correct kmem_list's list_lock
3097 */
3098 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3099 int node)
3101 int i;
3102 struct kmem_list3 *l3;
3104 for (i = 0; i < nr_objects; i++) {
3105 void *objp = objpp[i];
3106 struct slab *slabp;
3108 slabp = virt_to_slab(objp);
3109 l3 = cachep->nodelists[node];
3110 list_del(&slabp->list);
3111 check_spinlock_acquired_node(cachep, node);
3112 check_slabp(cachep, slabp);
3113 slab_put_obj(cachep, slabp, objp, node);
3114 STATS_DEC_ACTIVE(cachep);
3115 l3->free_objects++;
3116 check_slabp(cachep, slabp);
3118 /* fixup slab chains */
3119 if (slabp->inuse == 0) {
3120 if (l3->free_objects > l3->free_limit) {
3121 l3->free_objects -= cachep->num;
3122 slab_destroy(cachep, slabp);
3123 } else {
3124 list_add(&slabp->list, &l3->slabs_free);
3126 } else {
3127 /* Unconditionally move a slab to the end of the
3128 * partial list on free - maximum time for the
3129 * other objects to be freed, too.
3130 */
3131 list_add_tail(&slabp->list, &l3->slabs_partial);
3136 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3138 int batchcount;
3139 struct kmem_list3 *l3;
3140 int node = numa_node_id();
3142 batchcount = ac->batchcount;
3143 #if DEBUG
3144 BUG_ON(!batchcount || batchcount > ac->avail);
3145 #endif
3146 check_irq_off();
3147 l3 = cachep->nodelists[node];
3148 spin_lock(&l3->list_lock);
3149 if (l3->shared) {
3150 struct array_cache *shared_array = l3->shared;
3151 int max = shared_array->limit - shared_array->avail;
3152 if (max) {
3153 if (batchcount > max)
3154 batchcount = max;
3155 memcpy(&(shared_array->entry[shared_array->avail]),
3156 ac->entry, sizeof(void *) * batchcount);
3157 shared_array->avail += batchcount;
3158 goto free_done;
3162 free_block(cachep, ac->entry, batchcount, node);
3163 free_done:
3164 #if STATS
3166 int i = 0;
3167 struct list_head *p;
3169 p = l3->slabs_free.next;
3170 while (p != &(l3->slabs_free)) {
3171 struct slab *slabp;
3173 slabp = list_entry(p, struct slab, list);
3174 BUG_ON(slabp->inuse);
3176 i++;
3177 p = p->next;
3179 STATS_SET_FREEABLE(cachep, i);
3181 #endif
3182 spin_unlock(&l3->list_lock);
3183 ac->avail -= batchcount;
3184 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3187 /*
3188 * Release an obj back to its cache. If the obj has a constructed state, it must
3189 * be in this state _before_ it is released. Called with disabled ints.
3190 */
3191 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3193 struct array_cache *ac = cpu_cache_get(cachep);
3195 check_irq_off();
3196 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3198 if (cache_free_alien(cachep, objp))
3199 return;
3201 if (likely(ac->avail < ac->limit)) {
3202 STATS_INC_FREEHIT(cachep);
3203 ac->entry[ac->avail++] = objp;
3204 return;
3205 } else {
3206 STATS_INC_FREEMISS(cachep);
3207 cache_flusharray(cachep, ac);
3208 ac->entry[ac->avail++] = objp;
3212 /**
3213 * kmem_cache_alloc - Allocate an object
3214 * @cachep: The cache to allocate from.
3215 * @flags: See kmalloc().
3217 * Allocate an object from this cache. The flags are only relevant
3218 * if the cache has no available objects.
3219 */
3220 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3222 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3224 EXPORT_SYMBOL(kmem_cache_alloc);
3226 /**
3227 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3228 * @cache: The cache to allocate from.
3229 * @flags: See kmalloc().
3231 * Allocate an object from this cache and set the allocated memory to zero.
3232 * The flags are only relevant if the cache has no available objects.
3233 */
3234 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3236 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3237 if (ret)
3238 memset(ret, 0, obj_size(cache));
3239 return ret;
3241 EXPORT_SYMBOL(kmem_cache_zalloc);
3243 /**
3244 * kmem_ptr_validate - check if an untrusted pointer might
3245 * be a slab entry.
3246 * @cachep: the cache we're checking against
3247 * @ptr: pointer to validate
3249 * This verifies that the untrusted pointer looks sane:
3250 * it is _not_ a guarantee that the pointer is actually
3251 * part of the slab cache in question, but it at least
3252 * validates that the pointer can be dereferenced and
3253 * looks half-way sane.
3255 * Currently only used for dentry validation.
3256 */
3257 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3259 unsigned long addr = (unsigned long)ptr;
3260 unsigned long min_addr = PAGE_OFFSET;
3261 unsigned long align_mask = BYTES_PER_WORD - 1;
3262 unsigned long size = cachep->buffer_size;
3263 struct page *page;
3265 if (unlikely(addr < min_addr))
3266 goto out;
3267 if (unlikely(addr > (unsigned long)high_memory - size))
3268 goto out;
3269 if (unlikely(addr & align_mask))
3270 goto out;
3271 if (unlikely(!kern_addr_valid(addr)))
3272 goto out;
3273 if (unlikely(!kern_addr_valid(addr + size - 1)))
3274 goto out;
3275 page = virt_to_page(ptr);
3276 if (unlikely(!PageSlab(page)))
3277 goto out;
3278 if (unlikely(page_get_cache(page) != cachep))
3279 goto out;
3280 return 1;
3281 out:
3282 return 0;
3285 #ifdef CONFIG_NUMA
3286 /**
3287 * kmem_cache_alloc_node - Allocate an object on the specified node
3288 * @cachep: The cache to allocate from.
3289 * @flags: See kmalloc().
3290 * @nodeid: node number of the target node.
3292 * Identical to kmem_cache_alloc, except that this function is slow
3293 * and can sleep. And it will allocate memory on the given node, which
3294 * can improve the performance for cpu bound structures.
3295 * New and improved: it will now make sure that the object gets
3296 * put on the correct node list so that there is no false sharing.
3297 */
3298 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3300 unsigned long save_flags;
3301 void *ptr;
3303 cache_alloc_debugcheck_before(cachep, flags);
3304 local_irq_save(save_flags);
3306 if (nodeid == -1 || nodeid == numa_node_id() ||
3307 !cachep->nodelists[nodeid])
3308 ptr = ____cache_alloc(cachep, flags);
3309 else
3310 ptr = __cache_alloc_node(cachep, flags, nodeid);
3311 local_irq_restore(save_flags);
3313 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3314 __builtin_return_address(0));
3316 return ptr;
3318 EXPORT_SYMBOL(kmem_cache_alloc_node);
3320 void *kmalloc_node(size_t size, gfp_t flags, int node)
3322 struct kmem_cache *cachep;
3324 cachep = kmem_find_general_cachep(size, flags);
3325 if (unlikely(cachep == NULL))
3326 return NULL;
3327 return kmem_cache_alloc_node(cachep, flags, node);
3329 EXPORT_SYMBOL(kmalloc_node);
3330 #endif
3332 /**
3333 * __do_kmalloc - allocate memory
3334 * @size: how many bytes of memory are required.
3335 * @flags: the type of memory to allocate (see kmalloc).
3336 * @caller: function caller for debug tracking of the caller
3337 */
3338 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3339 void *caller)
3341 struct kmem_cache *cachep;
3343 /* If you want to save a few bytes .text space: replace
3344 * __ with kmem_.
3345 * Then kmalloc uses the uninlined functions instead of the inline
3346 * functions.
3347 */
3348 cachep = __find_general_cachep(size, flags);
3349 if (unlikely(cachep == NULL))
3350 return NULL;
3351 return __cache_alloc(cachep, flags, caller);
3355 void *__kmalloc(size_t size, gfp_t flags)
3357 #ifndef CONFIG_DEBUG_SLAB
3358 return __do_kmalloc(size, flags, NULL);
3359 #else
3360 return __do_kmalloc(size, flags, __builtin_return_address(0));
3361 #endif
3363 EXPORT_SYMBOL(__kmalloc);
3365 #ifdef CONFIG_DEBUG_SLAB
3366 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3368 return __do_kmalloc(size, flags, caller);
3370 EXPORT_SYMBOL(__kmalloc_track_caller);
3371 #endif
3373 #ifdef CONFIG_SMP
3374 /**
3375 * __alloc_percpu - allocate one copy of the object for every present
3376 * cpu in the system, zeroing them.
3377 * Objects should be dereferenced using the per_cpu_ptr macro only.
3379 * @size: how many bytes of memory are required.
3380 */
3381 void *__alloc_percpu(size_t size)
3383 int i;
3384 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3386 if (!pdata)
3387 return NULL;
3389 /*
3390 * Cannot use for_each_online_cpu since a cpu may come online
3391 * and we have no way of figuring out how to fix the array
3392 * that we have allocated then....
3393 */
3394 for_each_possible_cpu(i) {
3395 int node = cpu_to_node(i);
3397 if (node_online(node))
3398 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3399 else
3400 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3402 if (!pdata->ptrs[i])
3403 goto unwind_oom;
3404 memset(pdata->ptrs[i], 0, size);
3407 /* Catch derefs w/o wrappers */
3408 return (void *)(~(unsigned long)pdata);
3410 unwind_oom:
3411 while (--i >= 0) {
3412 if (!cpu_possible(i))
3413 continue;
3414 kfree(pdata->ptrs[i]);
3416 kfree(pdata);
3417 return NULL;
3419 EXPORT_SYMBOL(__alloc_percpu);
3420 #endif
3422 /**
3423 * kmem_cache_free - Deallocate an object
3424 * @cachep: The cache the allocation was from.
3425 * @objp: The previously allocated object.
3427 * Free an object which was previously allocated from this
3428 * cache.
3429 */
3430 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3432 unsigned long flags;
3434 BUG_ON(virt_to_cache(objp) != cachep);
3436 local_irq_save(flags);
3437 __cache_free(cachep, objp);
3438 local_irq_restore(flags);
3440 EXPORT_SYMBOL(kmem_cache_free);
3442 /**
3443 * kfree - free previously allocated memory
3444 * @objp: pointer returned by kmalloc.
3446 * If @objp is NULL, no operation is performed.
3448 * Don't free memory not originally allocated by kmalloc()
3449 * or you will run into trouble.
3450 */
3451 void kfree(const void *objp)
3453 struct kmem_cache *c;
3454 unsigned long flags;
3456 if (unlikely(!objp))
3457 return;
3458 local_irq_save(flags);
3459 kfree_debugcheck(objp);
3460 c = virt_to_cache(objp);
3461 debug_check_no_locks_freed(objp, obj_size(c));
3462 __cache_free(c, (void *)objp);
3463 local_irq_restore(flags);
3465 EXPORT_SYMBOL(kfree);
3467 #ifdef CONFIG_SMP
3468 /**
3469 * free_percpu - free previously allocated percpu memory
3470 * @objp: pointer returned by alloc_percpu.
3472 * Don't free memory not originally allocated by alloc_percpu()
3473 * The complemented objp is to check for that.
3474 */
3475 void free_percpu(const void *objp)
3477 int i;
3478 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3480 /*
3481 * We allocate for all cpus so we cannot use for online cpu here.
3482 */
3483 for_each_possible_cpu(i)
3484 kfree(p->ptrs[i]);
3485 kfree(p);
3487 EXPORT_SYMBOL(free_percpu);
3488 #endif
3490 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3492 return obj_size(cachep);
3494 EXPORT_SYMBOL(kmem_cache_size);
3496 const char *kmem_cache_name(struct kmem_cache *cachep)
3498 return cachep->name;
3500 EXPORT_SYMBOL_GPL(kmem_cache_name);
3502 /*
3503 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3504 */
3505 static int alloc_kmemlist(struct kmem_cache *cachep)
3507 int node;
3508 struct kmem_list3 *l3;
3509 struct array_cache *new_shared;
3510 struct array_cache **new_alien;
3512 for_each_online_node(node) {
3514 new_alien = alloc_alien_cache(node, cachep->limit);
3515 if (!new_alien)
3516 goto fail;
3518 new_shared = alloc_arraycache(node,
3519 cachep->shared*cachep->batchcount,
3520 0xbaadf00d);
3521 if (!new_shared) {
3522 free_alien_cache(new_alien);
3523 goto fail;
3526 l3 = cachep->nodelists[node];
3527 if (l3) {
3528 struct array_cache *shared = l3->shared;
3530 spin_lock_irq(&l3->list_lock);
3532 if (shared)
3533 free_block(cachep, shared->entry,
3534 shared->avail, node);
3536 l3->shared = new_shared;
3537 if (!l3->alien) {
3538 l3->alien = new_alien;
3539 new_alien = NULL;
3541 l3->free_limit = (1 + nr_cpus_node(node)) *
3542 cachep->batchcount + cachep->num;
3543 spin_unlock_irq(&l3->list_lock);
3544 kfree(shared);
3545 free_alien_cache(new_alien);
3546 continue;
3548 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3549 if (!l3) {
3550 free_alien_cache(new_alien);
3551 kfree(new_shared);
3552 goto fail;
3555 kmem_list3_init(l3);
3556 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3557 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3558 l3->shared = new_shared;
3559 l3->alien = new_alien;
3560 l3->free_limit = (1 + nr_cpus_node(node)) *
3561 cachep->batchcount + cachep->num;
3562 cachep->nodelists[node] = l3;
3564 return 0;
3566 fail:
3567 if (!cachep->next.next) {
3568 /* Cache is not active yet. Roll back what we did */
3569 node--;
3570 while (node >= 0) {
3571 if (cachep->nodelists[node]) {
3572 l3 = cachep->nodelists[node];
3574 kfree(l3->shared);
3575 free_alien_cache(l3->alien);
3576 kfree(l3);
3577 cachep->nodelists[node] = NULL;
3579 node--;
3582 return -ENOMEM;
3585 struct ccupdate_struct {
3586 struct kmem_cache *cachep;
3587 struct array_cache *new[NR_CPUS];
3588 };
3590 static void do_ccupdate_local(void *info)
3592 struct ccupdate_struct *new = info;
3593 struct array_cache *old;
3595 check_irq_off();
3596 old = cpu_cache_get(new->cachep);
3598 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3599 new->new[smp_processor_id()] = old;
3602 /* Always called with the cache_chain_mutex held */
3603 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3604 int batchcount, int shared)
3606 struct ccupdate_struct new;
3607 int i, err;
3609 memset(&new.new, 0, sizeof(new.new));
3610 for_each_online_cpu(i) {
3611 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3612 batchcount);
3613 if (!new.new[i]) {
3614 for (i--; i >= 0; i--)
3615 kfree(new.new[i]);
3616 return -ENOMEM;
3619 new.cachep = cachep;
3621 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3623 check_irq_on();
3624 cachep->batchcount = batchcount;
3625 cachep->limit = limit;
3626 cachep->shared = shared;
3628 for_each_online_cpu(i) {
3629 struct array_cache *ccold = new.new[i];
3630 if (!ccold)
3631 continue;
3632 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3633 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3634 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3635 kfree(ccold);
3638 err = alloc_kmemlist(cachep);
3639 if (err) {
3640 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3641 cachep->name, -err);
3642 BUG();
3644 return 0;
3647 /* Called with cache_chain_mutex held always */
3648 static void enable_cpucache(struct kmem_cache *cachep)
3650 int err;
3651 int limit, shared;
3653 /*
3654 * The head array serves three purposes:
3655 * - create a LIFO ordering, i.e. return objects that are cache-warm
3656 * - reduce the number of spinlock operations.
3657 * - reduce the number of linked list operations on the slab and
3658 * bufctl chains: array operations are cheaper.
3659 * The numbers are guessed, we should auto-tune as described by
3660 * Bonwick.
3661 */
3662 if (cachep->buffer_size > 131072)
3663 limit = 1;
3664 else if (cachep->buffer_size > PAGE_SIZE)
3665 limit = 8;
3666 else if (cachep->buffer_size > 1024)
3667 limit = 24;
3668 else if (cachep->buffer_size > 256)
3669 limit = 54;
3670 else
3671 limit = 120;
3673 /*
3674 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3675 * allocation behaviour: Most allocs on one cpu, most free operations
3676 * on another cpu. For these cases, an efficient object passing between
3677 * cpus is necessary. This is provided by a shared array. The array
3678 * replaces Bonwick's magazine layer.
3679 * On uniprocessor, it's functionally equivalent (but less efficient)
3680 * to a larger limit. Thus disabled by default.
3681 */
3682 shared = 0;
3683 #ifdef CONFIG_SMP
3684 if (cachep->buffer_size <= PAGE_SIZE)
3685 shared = 8;
3686 #endif
3688 #if DEBUG
3689 /*
3690 * With debugging enabled, large batchcount lead to excessively long
3691 * periods with disabled local interrupts. Limit the batchcount
3692 */
3693 if (limit > 32)
3694 limit = 32;
3695 #endif
3696 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3697 if (err)
3698 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3699 cachep->name, -err);
3702 /*
3703 * Drain an array if it contains any elements taking the l3 lock only if
3704 * necessary. Note that the l3 listlock also protects the array_cache
3705 * if drain_array() is used on the shared array.
3706 */
3707 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3708 struct array_cache *ac, int force, int node)
3710 int tofree;
3712 if (!ac || !ac->avail)
3713 return;
3714 if (ac->touched && !force) {
3715 ac->touched = 0;
3716 } else {
3717 spin_lock_irq(&l3->list_lock);
3718 if (ac->avail) {
3719 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3720 if (tofree > ac->avail)
3721 tofree = (ac->avail + 1) / 2;
3722 free_block(cachep, ac->entry, tofree, node);
3723 ac->avail -= tofree;
3724 memmove(ac->entry, &(ac->entry[tofree]),
3725 sizeof(void *) * ac->avail);
3727 spin_unlock_irq(&l3->list_lock);
3731 /**
3732 * cache_reap - Reclaim memory from caches.
3733 * @unused: unused parameter
3735 * Called from workqueue/eventd every few seconds.
3736 * Purpose:
3737 * - clear the per-cpu caches for this CPU.
3738 * - return freeable pages to the main free memory pool.
3740 * If we cannot acquire the cache chain mutex then just give up - we'll try
3741 * again on the next iteration.
3742 */
3743 static void cache_reap(void *unused)
3745 struct kmem_cache *searchp;
3746 struct kmem_list3 *l3;
3747 int node = numa_node_id();
3749 if (!mutex_trylock(&cache_chain_mutex)) {
3750 /* Give up. Setup the next iteration. */
3751 schedule_delayed_work(&__get_cpu_var(reap_work),
3752 REAPTIMEOUT_CPUC);
3753 return;
3756 list_for_each_entry(searchp, &cache_chain, next) {
3757 check_irq_on();
3759 /*
3760 * We only take the l3 lock if absolutely necessary and we
3761 * have established with reasonable certainty that
3762 * we can do some work if the lock was obtained.
3763 */
3764 l3 = searchp->nodelists[node];
3766 reap_alien(searchp, l3);
3768 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3770 /*
3771 * These are racy checks but it does not matter
3772 * if we skip one check or scan twice.
3773 */
3774 if (time_after(l3->next_reap, jiffies))
3775 goto next;
3777 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3779 drain_array(searchp, l3, l3->shared, 0, node);
3781 if (l3->free_touched)
3782 l3->free_touched = 0;
3783 else {
3784 int freed;
3786 freed = drain_freelist(searchp, l3, (l3->free_limit +
3787 5 * searchp->num - 1) / (5 * searchp->num));
3788 STATS_ADD_REAPED(searchp, freed);
3790 next:
3791 cond_resched();
3793 check_irq_on();
3794 mutex_unlock(&cache_chain_mutex);
3795 next_reap_node();
3796 refresh_cpu_vm_stats(smp_processor_id());
3797 /* Set up the next iteration */
3798 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3801 #ifdef CONFIG_PROC_FS
3803 static void print_slabinfo_header(struct seq_file *m)
3805 /*
3806 * Output format version, so at least we can change it
3807 * without _too_ many complaints.
3808 */
3809 #if STATS
3810 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3811 #else
3812 seq_puts(m, "slabinfo - version: 2.1\n");
3813 #endif
3814 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3815 "<objperslab> <pagesperslab>");
3816 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3817 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3818 #if STATS
3819 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3820 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3821 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3822 #endif
3823 seq_putc(m, '\n');
3826 static void *s_start(struct seq_file *m, loff_t *pos)
3828 loff_t n = *pos;
3829 struct list_head *p;
3831 mutex_lock(&cache_chain_mutex);
3832 if (!n)
3833 print_slabinfo_header(m);
3834 p = cache_chain.next;
3835 while (n--) {
3836 p = p->next;
3837 if (p == &cache_chain)
3838 return NULL;
3840 return list_entry(p, struct kmem_cache, next);
3843 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3845 struct kmem_cache *cachep = p;
3846 ++*pos;
3847 return cachep->next.next == &cache_chain ?
3848 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3851 static void s_stop(struct seq_file *m, void *p)
3853 mutex_unlock(&cache_chain_mutex);
3856 static int s_show(struct seq_file *m, void *p)
3858 struct kmem_cache *cachep = p;
3859 struct slab *slabp;
3860 unsigned long active_objs;
3861 unsigned long num_objs;
3862 unsigned long active_slabs = 0;
3863 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3864 const char *name;
3865 char *error = NULL;
3866 int node;
3867 struct kmem_list3 *l3;
3869 active_objs = 0;
3870 num_slabs = 0;
3871 for_each_online_node(node) {
3872 l3 = cachep->nodelists[node];
3873 if (!l3)
3874 continue;
3876 check_irq_on();
3877 spin_lock_irq(&l3->list_lock);
3879 list_for_each_entry(slabp, &l3->slabs_full, list) {
3880 if (slabp->inuse != cachep->num && !error)
3881 error = "slabs_full accounting error";
3882 active_objs += cachep->num;
3883 active_slabs++;
3885 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3886 if (slabp->inuse == cachep->num && !error)
3887 error = "slabs_partial inuse accounting error";
3888 if (!slabp->inuse && !error)
3889 error = "slabs_partial/inuse accounting error";
3890 active_objs += slabp->inuse;
3891 active_slabs++;
3893 list_for_each_entry(slabp, &l3->slabs_free, list) {
3894 if (slabp->inuse && !error)
3895 error = "slabs_free/inuse accounting error";
3896 num_slabs++;
3898 free_objects += l3->free_objects;
3899 if (l3->shared)
3900 shared_avail += l3->shared->avail;
3902 spin_unlock_irq(&l3->list_lock);
3904 num_slabs += active_slabs;
3905 num_objs = num_slabs * cachep->num;
3906 if (num_objs - active_objs != free_objects && !error)
3907 error = "free_objects accounting error";
3909 name = cachep->name;
3910 if (error)
3911 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3913 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3914 name, active_objs, num_objs, cachep->buffer_size,
3915 cachep->num, (1 << cachep->gfporder));
3916 seq_printf(m, " : tunables %4u %4u %4u",
3917 cachep->limit, cachep->batchcount, cachep->shared);
3918 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3919 active_slabs, num_slabs, shared_avail);
3920 #if STATS
3921 { /* list3 stats */
3922 unsigned long high = cachep->high_mark;
3923 unsigned long allocs = cachep->num_allocations;
3924 unsigned long grown = cachep->grown;
3925 unsigned long reaped = cachep->reaped;
3926 unsigned long errors = cachep->errors;
3927 unsigned long max_freeable = cachep->max_freeable;
3928 unsigned long node_allocs = cachep->node_allocs;
3929 unsigned long node_frees = cachep->node_frees;
3930 unsigned long overflows = cachep->node_overflow;
3932 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3933 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3934 reaped, errors, max_freeable, node_allocs,
3935 node_frees, overflows);
3937 /* cpu stats */
3939 unsigned long allochit = atomic_read(&cachep->allochit);
3940 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3941 unsigned long freehit = atomic_read(&cachep->freehit);
3942 unsigned long freemiss = atomic_read(&cachep->freemiss);
3944 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3945 allochit, allocmiss, freehit, freemiss);
3947 #endif
3948 seq_putc(m, '\n');
3949 return 0;
3952 /*
3953 * slabinfo_op - iterator that generates /proc/slabinfo
3955 * Output layout:
3956 * cache-name
3957 * num-active-objs
3958 * total-objs
3959 * object size
3960 * num-active-slabs
3961 * total-slabs
3962 * num-pages-per-slab
3963 * + further values on SMP and with statistics enabled
3964 */
3966 struct seq_operations slabinfo_op = {
3967 .start = s_start,
3968 .next = s_next,
3969 .stop = s_stop,
3970 .show = s_show,
3971 };
3973 #define MAX_SLABINFO_WRITE 128
3974 /**
3975 * slabinfo_write - Tuning for the slab allocator
3976 * @file: unused
3977 * @buffer: user buffer
3978 * @count: data length
3979 * @ppos: unused
3980 */
3981 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3982 size_t count, loff_t *ppos)
3984 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3985 int limit, batchcount, shared, res;
3986 struct kmem_cache *cachep;
3988 if (count > MAX_SLABINFO_WRITE)
3989 return -EINVAL;
3990 if (copy_from_user(&kbuf, buffer, count))
3991 return -EFAULT;
3992 kbuf[MAX_SLABINFO_WRITE] = '\0';
3994 tmp = strchr(kbuf, ' ');
3995 if (!tmp)
3996 return -EINVAL;
3997 *tmp = '\0';
3998 tmp++;
3999 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4000 return -EINVAL;
4002 /* Find the cache in the chain of caches. */
4003 mutex_lock(&cache_chain_mutex);
4004 res = -EINVAL;
4005 list_for_each_entry(cachep, &cache_chain, next) {
4006 if (!strcmp(cachep->name, kbuf)) {
4007 if (limit < 1 || batchcount < 1 ||
4008 batchcount > limit || shared < 0) {
4009 res = 0;
4010 } else {
4011 res = do_tune_cpucache(cachep, limit,
4012 batchcount, shared);
4014 break;
4017 mutex_unlock(&cache_chain_mutex);
4018 if (res >= 0)
4019 res = count;
4020 return res;
4023 #ifdef CONFIG_DEBUG_SLAB_LEAK
4025 static void *leaks_start(struct seq_file *m, loff_t *pos)
4027 loff_t n = *pos;
4028 struct list_head *p;
4030 mutex_lock(&cache_chain_mutex);
4031 p = cache_chain.next;
4032 while (n--) {
4033 p = p->next;
4034 if (p == &cache_chain)
4035 return NULL;
4037 return list_entry(p, struct kmem_cache, next);
4040 static inline int add_caller(unsigned long *n, unsigned long v)
4042 unsigned long *p;
4043 int l;
4044 if (!v)
4045 return 1;
4046 l = n[1];
4047 p = n + 2;
4048 while (l) {
4049 int i = l/2;
4050 unsigned long *q = p + 2 * i;
4051 if (*q == v) {
4052 q[1]++;
4053 return 1;
4055 if (*q > v) {
4056 l = i;
4057 } else {
4058 p = q + 2;
4059 l -= i + 1;
4062 if (++n[1] == n[0])
4063 return 0;
4064 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4065 p[0] = v;
4066 p[1] = 1;
4067 return 1;
4070 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4072 void *p;
4073 int i;
4074 if (n[0] == n[1])
4075 return;
4076 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4077 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4078 continue;
4079 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4080 return;
4084 static void show_symbol(struct seq_file *m, unsigned long address)
4086 #ifdef CONFIG_KALLSYMS
4087 char *modname;
4088 const char *name;
4089 unsigned long offset, size;
4090 char namebuf[KSYM_NAME_LEN+1];
4092 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4094 if (name) {
4095 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4096 if (modname)
4097 seq_printf(m, " [%s]", modname);
4098 return;
4100 #endif
4101 seq_printf(m, "%p", (void *)address);
4104 static int leaks_show(struct seq_file *m, void *p)
4106 struct kmem_cache *cachep = p;
4107 struct slab *slabp;
4108 struct kmem_list3 *l3;
4109 const char *name;
4110 unsigned long *n = m->private;
4111 int node;
4112 int i;
4114 if (!(cachep->flags & SLAB_STORE_USER))
4115 return 0;
4116 if (!(cachep->flags & SLAB_RED_ZONE))
4117 return 0;
4119 /* OK, we can do it */
4121 n[1] = 0;
4123 for_each_online_node(node) {
4124 l3 = cachep->nodelists[node];
4125 if (!l3)
4126 continue;
4128 check_irq_on();
4129 spin_lock_irq(&l3->list_lock);
4131 list_for_each_entry(slabp, &l3->slabs_full, list)
4132 handle_slab(n, cachep, slabp);
4133 list_for_each_entry(slabp, &l3->slabs_partial, list)
4134 handle_slab(n, cachep, slabp);
4135 spin_unlock_irq(&l3->list_lock);
4137 name = cachep->name;
4138 if (n[0] == n[1]) {
4139 /* Increase the buffer size */
4140 mutex_unlock(&cache_chain_mutex);
4141 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4142 if (!m->private) {
4143 /* Too bad, we are really out */
4144 m->private = n;
4145 mutex_lock(&cache_chain_mutex);
4146 return -ENOMEM;
4148 *(unsigned long *)m->private = n[0] * 2;
4149 kfree(n);
4150 mutex_lock(&cache_chain_mutex);
4151 /* Now make sure this entry will be retried */
4152 m->count = m->size;
4153 return 0;
4155 for (i = 0; i < n[1]; i++) {
4156 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4157 show_symbol(m, n[2*i+2]);
4158 seq_putc(m, '\n');
4160 return 0;
4163 struct seq_operations slabstats_op = {
4164 .start = leaks_start,
4165 .next = s_next,
4166 .stop = s_stop,
4167 .show = leaks_show,
4168 };
4169 #endif
4170 #endif
4172 /**
4173 * ksize - get the actual amount of memory allocated for a given object
4174 * @objp: Pointer to the object
4176 * kmalloc may internally round up allocations and return more memory
4177 * than requested. ksize() can be used to determine the actual amount of
4178 * memory allocated. The caller may use this additional memory, even though
4179 * a smaller amount of memory was initially specified with the kmalloc call.
4180 * The caller must guarantee that objp points to a valid object previously
4181 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4182 * must not be freed during the duration of the call.
4183 */
4184 unsigned int ksize(const void *objp)
4186 if (unlikely(objp == NULL))
4187 return 0;
4189 return obj_size(virt_to_cache(objp));