blob: cf19ff2ab5e2bca587a645c241b24b65a7531901 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 kmem_cache_t and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
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 semaphore 'cache_chain_sem'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
89#include <linux/config.h>
90#include <linux/slab.h>
91#include <linux/mm.h>
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
97#include <linux/seq_file.h>
98#include <linux/notifier.h>
99#include <linux/kallsyms.h>
100#include <linux/cpu.h>
101#include <linux/sysctl.h>
102#include <linux/module.h>
103#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700104#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700105#include <linux/nodemask.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
107#include <asm/uaccess.h>
108#include <asm/cacheflush.h>
109#include <asm/tlbflush.h>
110#include <asm/page.h>
111
112/*
113 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
114 * SLAB_RED_ZONE & SLAB_POISON.
115 * 0 for faster, smaller code (especially in the critical paths).
116 *
117 * STATS - 1 to collect stats for /proc/slabinfo.
118 * 0 for faster, smaller code (especially in the critical paths).
119 *
120 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
121 */
122
123#ifdef CONFIG_DEBUG_SLAB
124#define DEBUG 1
125#define STATS 1
126#define FORCED_DEBUG 1
127#else
128#define DEBUG 0
129#define STATS 0
130#define FORCED_DEBUG 0
131#endif
132
133
134/* Shouldn't this be in a header file somewhere? */
135#define BYTES_PER_WORD sizeof(void *)
136
137#ifndef cache_line_size
138#define cache_line_size() L1_CACHE_BYTES
139#endif
140
141#ifndef ARCH_KMALLOC_MINALIGN
142/*
143 * Enforce a minimum alignment for the kmalloc caches.
144 * Usually, the kmalloc caches are cache_line_size() aligned, except when
145 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
146 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
147 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
148 * Note that this flag disables some debug features.
149 */
150#define ARCH_KMALLOC_MINALIGN 0
151#endif
152
153#ifndef ARCH_SLAB_MINALIGN
154/*
155 * Enforce a minimum alignment for all caches.
156 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
157 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
158 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
159 * some debug features.
160 */
161#define ARCH_SLAB_MINALIGN 0
162#endif
163
164#ifndef ARCH_KMALLOC_FLAGS
165#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
166#endif
167
168/* Legal flag mask for kmem_cache_create(). */
169#if DEBUG
170# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
172 SLAB_NO_REAP | SLAB_CACHE_DMA | \
173 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
174 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
175 SLAB_DESTROY_BY_RCU)
176#else
177# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
178 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU)
181#endif
182
183/*
184 * kmem_bufctl_t:
185 *
186 * Bufctl's are used for linking objs within a slab
187 * linked offsets.
188 *
189 * This implementation relies on "struct page" for locating the cache &
190 * slab an object belongs to.
191 * This allows the bufctl structure to be small (one int), but limits
192 * the number of objects a slab (not a cache) can contain when off-slab
193 * bufctls are used. The limit is the size of the largest general cache
194 * that does not use off-slab slabs.
195 * For 32bit archs with 4 kB pages, is this 56.
196 * This is not serious, as it is only for large objects, when it is unwise
197 * to have too many per slab.
198 * Note: This limit can be raised by introducing a general cache whose size
199 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
200 */
201
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700202typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
204#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
205#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
206
207/* Max number of objs-per-slab for caches which use off-slab slabs.
208 * Needed to avoid a possible looping condition in cache_grow().
209 */
210static unsigned long offslab_limit;
211
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 */
219struct 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;
Christoph Lametere498be72005-09-09 13:03:32 -0700225 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
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 */
244struct slab_rcu {
245 struct rcu_head head;
246 kmem_cache_t *cachep;
247 void *addr;
248};
249
250/*
251 * struct array_cache
252 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 * 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 */
262struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 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 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
276/* bootstrap: The caches do not work without cpuarrays anymore,
277 * but the cpuarrays are allocated from the generic caches...
278 */
279#define BOOT_CPUCACHE_ENTRIES 1
280struct arraycache_init {
281 struct array_cache cache;
282 void * entries[BOOT_CPUCACHE_ENTRIES];
283};
284
285/*
Christoph Lametere498be72005-09-09 13:03:32 -0700286 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 */
288struct kmem_list3 {
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 unsigned long next_reap;
Christoph Lametere498be72005-09-09 13:03:32 -0700294 int free_touched;
295 unsigned int free_limit;
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299};
300
Christoph Lametere498be72005-09-09 13:03:32 -0700301/*
302 * Need this for bootstrapping a per node allocator.
303 */
304#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
306#define CACHE_CACHE 0
307#define SIZE_AC 1
308#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Christoph Lametere498be72005-09-09 13:03:32 -0700310/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700311 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700312 * a constant is passed to it. Mostly the same as
313 * what is in linux/slab.h except it returns an
314 * index.
315 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700316static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700317{
318 if (__builtin_constant_p(size)) {
319 int i = 0;
320
321#define CACHE(x) \
322 if (size <=x) \
323 return i; \
324 else \
325 i++;
326#include "linux/kmalloc_sizes.h"
327#undef CACHE
328 {
329 extern void __bad_size(void);
330 __bad_size();
331 }
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700332 } else
333 BUG();
Christoph Lametere498be72005-09-09 13:03:32 -0700334 return 0;
335}
336
337#define INDEX_AC index_of(sizeof(struct arraycache_init))
338#define INDEX_L3 index_of(sizeof(struct kmem_list3))
339
340static inline void kmem_list3_init(struct kmem_list3 *parent)
341{
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 spin_lock_init(&parent->list_lock);
348 parent->free_objects = 0;
349 parent->free_touched = 0;
350}
351
352#define MAKE_LIST(cachep, listp, slab, nodeid) \
353 do { \
354 INIT_LIST_HEAD(listp); \
355 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
356 } while (0)
357
358#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
359 do { \
360 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
363 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
365/*
366 * kmem_cache_t
367 *
368 * manages a cache.
369 */
370
371struct kmem_cache_s {
372/* 1) per-cpu data, touched during every alloc/free */
373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
375 unsigned int limit;
Christoph Lametere498be72005-09-09 13:03:32 -0700376 unsigned int shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 unsigned int objsize;
Christoph Lametere498be72005-09-09 13:03:32 -0700378/* 2) touched by every alloc & free from the backend */
379 struct kmem_list3 *nodelists[MAX_NUMNODES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 spinlock_t spinlock;
383
384/* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
386 unsigned int gfporder;
387
388 /* force GFP flags, e.g. GFP_DMA */
389 unsigned int gfpflags;
390
391 size_t colour; /* cache colouring range */
392 unsigned int colour_off; /* colour offset */
393 unsigned int colour_next; /* cache colouring */
394 kmem_cache_t *slabp_cache;
395 unsigned int slab_size;
396 unsigned int dflags; /* dynamic flags */
397
398 /* constructor func */
399 void (*ctor)(void *, kmem_cache_t *, unsigned long);
400
401 /* de-constructor func */
402 void (*dtor)(void *, kmem_cache_t *, unsigned long);
403
404/* 4) cache creation/removal */
405 const char *name;
406 struct list_head next;
407
408/* 5) statistics */
409#if STATS
410 unsigned long num_active;
411 unsigned long num_allocations;
412 unsigned long high_mark;
413 unsigned long grown;
414 unsigned long reaped;
415 unsigned long errors;
416 unsigned long max_freeable;
417 unsigned long node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -0700418 unsigned long node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 atomic_t allochit;
420 atomic_t allocmiss;
421 atomic_t freehit;
422 atomic_t freemiss;
423#endif
424#if DEBUG
425 int dbghead;
426 int reallen;
427#endif
428};
429
430#define CFLGS_OFF_SLAB (0x80000000UL)
431#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
432
433#define BATCHREFILL_LIMIT 16
434/* Optimization question: fewer reaps means less
435 * probability for unnessary cpucache drain/refill cycles.
436 *
437 * OTHO the cpuarrays can contain lots of objects,
438 * which could lock up otherwise freeable slabs.
439 */
440#define REAPTIMEOUT_CPUC (2*HZ)
441#define REAPTIMEOUT_LIST3 (4*HZ)
442
443#if STATS
444#define STATS_INC_ACTIVE(x) ((x)->num_active++)
445#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
446#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
447#define STATS_INC_GROWN(x) ((x)->grown++)
448#define STATS_INC_REAPED(x) ((x)->reaped++)
449#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
450 (x)->high_mark = (x)->num_active; \
451 } while (0)
452#define STATS_INC_ERR(x) ((x)->errors++)
453#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700454#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455#define STATS_SET_FREEABLE(x, i) \
456 do { if ((x)->max_freeable < i) \
457 (x)->max_freeable = i; \
458 } while (0)
459
460#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
461#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
462#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
463#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
464#else
465#define STATS_INC_ACTIVE(x) do { } while (0)
466#define STATS_DEC_ACTIVE(x) do { } while (0)
467#define STATS_INC_ALLOCED(x) do { } while (0)
468#define STATS_INC_GROWN(x) do { } while (0)
469#define STATS_INC_REAPED(x) do { } while (0)
470#define STATS_SET_HIGH(x) do { } while (0)
471#define STATS_INC_ERR(x) do { } while (0)
472#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700473#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474#define STATS_SET_FREEABLE(x, i) \
475 do { } while (0)
476
477#define STATS_INC_ALLOCHIT(x) do { } while (0)
478#define STATS_INC_ALLOCMISS(x) do { } while (0)
479#define STATS_INC_FREEHIT(x) do { } while (0)
480#define STATS_INC_FREEMISS(x) do { } while (0)
481#endif
482
483#if DEBUG
484/* Magic nums for obj red zoning.
485 * Placed in the first word before and the first word after an obj.
486 */
487#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
488#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
489
490/* ...and for poisoning */
491#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
492#define POISON_FREE 0x6b /* for use-after-free poisoning */
493#define POISON_END 0xa5 /* end-byte of poisoning */
494
495/* memory layout of objects:
496 * 0 : objp
497 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
498 * the end of an object is aligned with the end of the real
499 * allocation. Catches writes behind the end of the allocation.
500 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
501 * redzone word.
502 * cachep->dbghead: The real object.
503 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
504 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
505 */
506static int obj_dbghead(kmem_cache_t *cachep)
507{
508 return cachep->dbghead;
509}
510
511static int obj_reallen(kmem_cache_t *cachep)
512{
513 return cachep->reallen;
514}
515
516static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
517{
518 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
519 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
520}
521
522static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
523{
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
525 if (cachep->flags & SLAB_STORE_USER)
526 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
527 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
528}
529
530static void **dbg_userword(kmem_cache_t *cachep, void *objp)
531{
532 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
533 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
534}
535
536#else
537
538#define obj_dbghead(x) 0
539#define obj_reallen(cachep) (cachep->objsize)
540#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
541#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
542#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
543
544#endif
545
546/*
547 * Maximum size of an obj (in 2^order pages)
548 * and absolute limit for the gfp order.
549 */
550#if defined(CONFIG_LARGE_ALLOCS)
551#define MAX_OBJ_ORDER 13 /* up to 32Mb */
552#define MAX_GFP_ORDER 13 /* up to 32Mb */
553#elif defined(CONFIG_MMU)
554#define MAX_OBJ_ORDER 5 /* 32 pages */
555#define MAX_GFP_ORDER 5 /* 32 pages */
556#else
557#define MAX_OBJ_ORDER 8 /* up to 1Mb */
558#define MAX_GFP_ORDER 8 /* up to 1Mb */
559#endif
560
561/*
562 * Do not go above this order unless 0 objects fit into the slab.
563 */
564#define BREAK_GFP_ORDER_HI 1
565#define BREAK_GFP_ORDER_LO 0
566static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
567
568/* Macros for storing/retrieving the cachep and or slab from the
569 * global 'mem_map'. These are used to find the slab an obj belongs to.
570 * With kfree(), these are used to find the cache which an obj belongs to.
571 */
572#define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
573#define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
574#define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
575#define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
576
577/* These are the default caches for kmalloc. Custom caches can have other sizes. */
578struct cache_sizes malloc_sizes[] = {
579#define CACHE(x) { .cs_size = (x) },
580#include <linux/kmalloc_sizes.h>
581 CACHE(ULONG_MAX)
582#undef CACHE
583};
584EXPORT_SYMBOL(malloc_sizes);
585
586/* Must match cache_sizes above. Out of line to keep cache footprint low. */
587struct cache_names {
588 char *name;
589 char *name_dma;
590};
591
592static struct cache_names __initdata cache_names[] = {
593#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
594#include <linux/kmalloc_sizes.h>
595 { NULL, }
596#undef CACHE
597};
598
599static struct arraycache_init initarray_cache __initdata =
600 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
601static struct arraycache_init initarray_generic =
602 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
603
604/* internal cache of cache description objs */
605static kmem_cache_t cache_cache = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 .batchcount = 1,
607 .limit = BOOT_CPUCACHE_ENTRIES,
Christoph Lametere498be72005-09-09 13:03:32 -0700608 .shared = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 .objsize = sizeof(kmem_cache_t),
610 .flags = SLAB_NO_REAP,
611 .spinlock = SPIN_LOCK_UNLOCKED,
612 .name = "kmem_cache",
613#if DEBUG
614 .reallen = sizeof(kmem_cache_t),
615#endif
616};
617
618/* Guard access to the cache-chain. */
619static struct semaphore cache_chain_sem;
620static struct list_head cache_chain;
621
622/*
623 * vm_enough_memory() looks at this to determine how many
624 * slab-allocated pages are possibly freeable under pressure
625 *
626 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
627 */
628atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630/*
631 * chicken and egg problem: delay the per-cpu array allocation
632 * until the general caches are up.
633 */
634static enum {
635 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700636 PARTIAL_AC,
637 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 FULL
639} g_cpucache_up;
640
641static DEFINE_PER_CPU(struct work_struct, reap_work);
642
643static void free_block(kmem_cache_t* cachep, void** objpp, int len);
644static void enable_cpucache (kmem_cache_t *cachep);
645static void cache_reap (void *unused);
Christoph Lametere498be72005-09-09 13:03:32 -0700646static int __node_shrink(kmem_cache_t *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
648static inline struct array_cache *ac_data(kmem_cache_t *cachep)
649{
650 return cachep->array[smp_processor_id()];
651}
652
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700653static inline kmem_cache_t *__find_general_cachep(size_t size,
654 unsigned int __nocast gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655{
656 struct cache_sizes *csizep = malloc_sizes;
657
658#if DEBUG
659 /* This happens if someone tries to call
660 * kmem_cache_create(), or __kmalloc(), before
661 * the generic caches are initialized.
662 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700663 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664#endif
665 while (size > csizep->cs_size)
666 csizep++;
667
668 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700669 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 * has cs_{dma,}cachep==NULL. Thus no special case
671 * for large kmalloc calls required.
672 */
673 if (unlikely(gfpflags & GFP_DMA))
674 return csizep->cs_dmacachep;
675 return csizep->cs_cachep;
676}
677
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700678kmem_cache_t *kmem_find_general_cachep(size_t size,
679 unsigned int __nocast gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700680{
681 return __find_general_cachep(size, gfpflags);
682}
683EXPORT_SYMBOL(kmem_find_general_cachep);
684
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685/* Cal the num objs, wastage, and bytes left over for a given slab size. */
686static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
687 int flags, size_t *left_over, unsigned int *num)
688{
689 int i;
690 size_t wastage = PAGE_SIZE<<gfporder;
691 size_t extra = 0;
692 size_t base = 0;
693
694 if (!(flags & CFLGS_OFF_SLAB)) {
695 base = sizeof(struct slab);
696 extra = sizeof(kmem_bufctl_t);
697 }
698 i = 0;
699 while (i*size + ALIGN(base+i*extra, align) <= wastage)
700 i++;
701 if (i > 0)
702 i--;
703
704 if (i > SLAB_LIMIT)
705 i = SLAB_LIMIT;
706
707 *num = i;
708 wastage -= i*size;
709 wastage -= ALIGN(base+i*extra, align);
710 *left_over = wastage;
711}
712
713#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
714
715static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
716{
717 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
718 function, cachep->name, msg);
719 dump_stack();
720}
721
722/*
723 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
724 * via the workqueue/eventd.
725 * Add the CPU number into the expiration time to minimize the possibility of
726 * the CPUs getting into lockstep and contending for the global cache chain
727 * lock.
728 */
729static void __devinit start_cpu_timer(int cpu)
730{
731 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
732
733 /*
734 * When this gets called from do_initcalls via cpucache_init(),
735 * init_workqueues() has already run, so keventd will be setup
736 * at that time.
737 */
738 if (keventd_up() && reap_work->func == NULL) {
739 INIT_WORK(reap_work, cache_reap, NULL);
740 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
741 }
742}
743
Christoph Lametere498be72005-09-09 13:03:32 -0700744static struct array_cache *alloc_arraycache(int node, int entries,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 int batchcount)
746{
747 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
748 struct array_cache *nc = NULL;
749
Christoph Lametere498be72005-09-09 13:03:32 -0700750 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 if (nc) {
752 nc->avail = 0;
753 nc->limit = entries;
754 nc->batchcount = batchcount;
755 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700756 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 }
758 return nc;
759}
760
Christoph Lametere498be72005-09-09 13:03:32 -0700761#ifdef CONFIG_NUMA
762static inline struct array_cache **alloc_alien_cache(int node, int limit)
763{
764 struct array_cache **ac_ptr;
765 int memsize = sizeof(void*)*MAX_NUMNODES;
766 int i;
767
768 if (limit > 1)
769 limit = 12;
770 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
771 if (ac_ptr) {
772 for_each_node(i) {
773 if (i == node || !node_online(i)) {
774 ac_ptr[i] = NULL;
775 continue;
776 }
777 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
778 if (!ac_ptr[i]) {
779 for (i--; i <=0; i--)
780 kfree(ac_ptr[i]);
781 kfree(ac_ptr);
782 return NULL;
783 }
784 }
785 }
786 return ac_ptr;
787}
788
789static inline void free_alien_cache(struct array_cache **ac_ptr)
790{
791 int i;
792
793 if (!ac_ptr)
794 return;
795
796 for_each_node(i)
797 kfree(ac_ptr[i]);
798
799 kfree(ac_ptr);
800}
801
802static inline void __drain_alien_cache(kmem_cache_t *cachep, struct array_cache *ac, int node)
803{
804 struct kmem_list3 *rl3 = cachep->nodelists[node];
805
806 if (ac->avail) {
807 spin_lock(&rl3->list_lock);
808 free_block(cachep, ac->entry, ac->avail);
809 ac->avail = 0;
810 spin_unlock(&rl3->list_lock);
811 }
812}
813
814static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
815{
816 int i=0;
817 struct array_cache *ac;
818 unsigned long flags;
819
820 for_each_online_node(i) {
821 ac = l3->alien[i];
822 if (ac) {
823 spin_lock_irqsave(&ac->lock, flags);
824 __drain_alien_cache(cachep, ac, i);
825 spin_unlock_irqrestore(&ac->lock, flags);
826 }
827 }
828}
829#else
830#define alloc_alien_cache(node, limit) do { } while (0)
831#define free_alien_cache(ac_ptr) do { } while (0)
832#define drain_alien_cache(cachep, l3) do { } while (0)
833#endif
834
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835static int __devinit cpuup_callback(struct notifier_block *nfb,
836 unsigned long action, void *hcpu)
837{
838 long cpu = (long)hcpu;
839 kmem_cache_t* cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700840 struct kmem_list3 *l3 = NULL;
841 int node = cpu_to_node(cpu);
842 int memsize = sizeof(struct kmem_list3);
843 struct array_cache *nc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845 switch (action) {
846 case CPU_UP_PREPARE:
847 down(&cache_chain_sem);
Christoph Lametere498be72005-09-09 13:03:32 -0700848 /* we need to do this right in the beginning since
849 * alloc_arraycache's are going to use this list.
850 * kmalloc_node allows us to add the slab to the right
851 * kmem_list3 and not this cpu's kmem_list3
852 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853
Christoph Lametere498be72005-09-09 13:03:32 -0700854 list_for_each_entry(cachep, &cache_chain, next) {
855 /* setup the size64 kmemlist for cpu before we can
856 * begin anything. Make sure some other cpu on this
857 * node has not already allocated this
858 */
859 if (!cachep->nodelists[node]) {
860 if (!(l3 = kmalloc_node(memsize,
861 GFP_KERNEL, node)))
862 goto bad;
863 kmem_list3_init(l3);
864 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
865 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
866
867 cachep->nodelists[node] = l3;
868 }
869
870 spin_lock_irq(&cachep->nodelists[node]->list_lock);
871 cachep->nodelists[node]->free_limit =
872 (1 + nr_cpus_node(node)) *
873 cachep->batchcount + cachep->num;
874 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
875 }
876
877 /* Now we can go ahead with allocating the shared array's
878 & array cache's */
879 list_for_each_entry(cachep, &cache_chain, next) {
880 nc = alloc_arraycache(node, cachep->limit,
881 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 if (!nc)
883 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885
Christoph Lametere498be72005-09-09 13:03:32 -0700886 l3 = cachep->nodelists[node];
887 BUG_ON(!l3);
888 if (!l3->shared) {
889 if (!(nc = alloc_arraycache(node,
890 cachep->shared*cachep->batchcount,
891 0xbaadf00d)))
892 goto bad;
893
894 /* we are serialised from CPU_DEAD or
895 CPU_UP_CANCELLED by the cpucontrol lock */
896 l3->shared = nc;
897 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 }
899 up(&cache_chain_sem);
900 break;
901 case CPU_ONLINE:
902 start_cpu_timer(cpu);
903 break;
904#ifdef CONFIG_HOTPLUG_CPU
905 case CPU_DEAD:
906 /* fall thru */
907 case CPU_UP_CANCELED:
908 down(&cache_chain_sem);
909
910 list_for_each_entry(cachep, &cache_chain, next) {
911 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700912 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913
Christoph Lametere498be72005-09-09 13:03:32 -0700914 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 spin_lock_irq(&cachep->spinlock);
916 /* cpu is dead; no one can alloc from it. */
917 nc = cachep->array[cpu];
918 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700919 l3 = cachep->nodelists[node];
920
921 if (!l3)
922 goto unlock_cache;
923
924 spin_lock(&l3->list_lock);
925
926 /* Free limit for this kmem_list3 */
927 l3->free_limit -= cachep->batchcount;
928 if (nc)
929 free_block(cachep, nc->entry, nc->avail);
930
931 if (!cpus_empty(mask)) {
932 spin_unlock(&l3->list_lock);
933 goto unlock_cache;
934 }
935
936 if (l3->shared) {
937 free_block(cachep, l3->shared->entry,
938 l3->shared->avail);
939 kfree(l3->shared);
940 l3->shared = NULL;
941 }
942 if (l3->alien) {
943 drain_alien_cache(cachep, l3);
944 free_alien_cache(l3->alien);
945 l3->alien = NULL;
946 }
947
948 /* free slabs belonging to this node */
949 if (__node_shrink(cachep, node)) {
950 cachep->nodelists[node] = NULL;
951 spin_unlock(&l3->list_lock);
952 kfree(l3);
953 } else {
954 spin_unlock(&l3->list_lock);
955 }
956unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 spin_unlock_irq(&cachep->spinlock);
958 kfree(nc);
959 }
960 up(&cache_chain_sem);
961 break;
962#endif
963 }
964 return NOTIFY_OK;
965bad:
966 up(&cache_chain_sem);
967 return NOTIFY_BAD;
968}
969
970static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
971
Christoph Lametere498be72005-09-09 13:03:32 -0700972/*
973 * swap the static kmem_list3 with kmalloced memory
974 */
975static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list,
976 int nodeid)
977{
978 struct kmem_list3 *ptr;
979
980 BUG_ON(cachep->nodelists[nodeid] != list);
981 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
982 BUG_ON(!ptr);
983
984 local_irq_disable();
985 memcpy(ptr, list, sizeof(struct kmem_list3));
986 MAKE_ALL_LISTS(cachep, ptr, nodeid);
987 cachep->nodelists[nodeid] = ptr;
988 local_irq_enable();
989}
990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991/* Initialisation.
992 * Called after the gfp() functions have been enabled, and before smp_init().
993 */
994void __init kmem_cache_init(void)
995{
996 size_t left_over;
997 struct cache_sizes *sizes;
998 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -0700999 int i;
1000
1001 for (i = 0; i < NUM_INIT_LISTS; i++) {
1002 kmem_list3_init(&initkmem_list3[i]);
1003 if (i < MAX_NUMNODES)
1004 cache_cache.nodelists[i] = NULL;
1005 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
1007 /*
1008 * Fragmentation resistance on low memory - only use bigger
1009 * page orders on machines with more than 32MB of memory.
1010 */
1011 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1012 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1013
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 /* Bootstrap is tricky, because several objects are allocated
1015 * from caches that do not exist yet:
1016 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1017 * structures of all caches, except cache_cache itself: cache_cache
1018 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001019 * Initially an __init data area is used for the head array and the
1020 * kmem_list3 structures, it's replaced with a kmalloc allocated
1021 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001023 * The kmem_cache_t for the new cache is allocated normally.
1024 * An __init data area is used for the head array.
1025 * 3) Create the remaining kmalloc caches, with minimally sized
1026 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 * 4) Replace the __init data head arrays for cache_cache and the first
1028 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001029 * 5) Replace the __init data for kmem_list3 for cache_cache and
1030 * the other cache's with kmalloc allocated memory.
1031 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032 */
1033
1034 /* 1) create the cache_cache */
1035 init_MUTEX(&cache_chain_sem);
1036 INIT_LIST_HEAD(&cache_chain);
1037 list_add(&cache_cache.next, &cache_chain);
1038 cache_cache.colour_off = cache_line_size();
1039 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001040 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041
1042 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1043
1044 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
1045 &left_over, &cache_cache.num);
1046 if (!cache_cache.num)
1047 BUG();
1048
1049 cache_cache.colour = left_over/cache_cache.colour_off;
1050 cache_cache.colour_next = 0;
1051 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
1052 sizeof(struct slab), cache_line_size());
1053
1054 /* 2+3) create the kmalloc caches */
1055 sizes = malloc_sizes;
1056 names = cache_names;
1057
Christoph Lametere498be72005-09-09 13:03:32 -07001058 /* Initialize the caches that provide memory for the array cache
1059 * and the kmem_list3 structures first.
1060 * Without this, further allocations will bug
1061 */
1062
1063 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1064 sizes[INDEX_AC].cs_size, ARCH_KMALLOC_MINALIGN,
1065 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1066
1067 if (INDEX_AC != INDEX_L3)
1068 sizes[INDEX_L3].cs_cachep =
1069 kmem_cache_create(names[INDEX_L3].name,
1070 sizes[INDEX_L3].cs_size, ARCH_KMALLOC_MINALIGN,
1071 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1072
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001074 /*
1075 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 * This should be particularly beneficial on SMP boxes, as it
1077 * eliminates "false sharing".
1078 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001079 * allow tighter packing of the smaller caches.
1080 */
1081 if(!sizes->cs_cachep)
1082 sizes->cs_cachep = kmem_cache_create(names->name,
1083 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1084 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085
1086 /* Inc off-slab bufctl limit until the ceiling is hit. */
1087 if (!(OFF_SLAB(sizes->cs_cachep))) {
1088 offslab_limit = sizes->cs_size-sizeof(struct slab);
1089 offslab_limit /= sizeof(kmem_bufctl_t);
1090 }
1091
1092 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1093 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1094 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
1095 NULL, NULL);
1096
1097 sizes++;
1098 names++;
1099 }
1100 /* 4) Replace the bootstrap head arrays */
1101 {
1102 void * ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001103
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001105
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 local_irq_disable();
1107 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001108 memcpy(ptr, ac_data(&cache_cache),
1109 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 cache_cache.array[smp_processor_id()] = ptr;
1111 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001112
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001114
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001116 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1117 != &initarray_generic.cache);
1118 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001120 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1121 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 local_irq_enable();
1123 }
Christoph Lametere498be72005-09-09 13:03:32 -07001124 /* 5) Replace the bootstrap kmem_list3's */
1125 {
1126 int node;
1127 /* Replace the static kmem_list3 structures for the boot cpu */
1128 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1129 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
Christoph Lametere498be72005-09-09 13:03:32 -07001131 for_each_online_node(node) {
1132 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1133 &initkmem_list3[SIZE_AC+node], node);
1134
1135 if (INDEX_AC != INDEX_L3) {
1136 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1137 &initkmem_list3[SIZE_L3+node],
1138 node);
1139 }
1140 }
1141 }
1142
1143 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 {
1145 kmem_cache_t *cachep;
1146 down(&cache_chain_sem);
1147 list_for_each_entry(cachep, &cache_chain, next)
1148 enable_cpucache(cachep);
1149 up(&cache_chain_sem);
1150 }
1151
1152 /* Done! */
1153 g_cpucache_up = FULL;
1154
1155 /* Register a cpu startup notifier callback
1156 * that initializes ac_data for all new cpus
1157 */
1158 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
1160 /* The reap timers are started later, with a module init call:
1161 * That part of the kernel is not yet operational.
1162 */
1163}
1164
1165static int __init cpucache_init(void)
1166{
1167 int cpu;
1168
1169 /*
1170 * Register the timers that return unneeded
1171 * pages to gfp.
1172 */
Christoph Lametere498be72005-09-09 13:03:32 -07001173 for_each_online_cpu(cpu)
1174 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
1176 return 0;
1177}
1178
1179__initcall(cpucache_init);
1180
1181/*
1182 * Interface to system's page allocator. No need to hold the cache-lock.
1183 *
1184 * If we requested dmaable memory, we will get it. Even if we
1185 * did not request dmaable memory, we might get it, but that
1186 * would be relatively rare and ignorable.
1187 */
1188static void *kmem_getpages(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
1189{
1190 struct page *page;
1191 void *addr;
1192 int i;
1193
1194 flags |= cachep->gfpflags;
1195 if (likely(nodeid == -1)) {
1196 page = alloc_pages(flags, cachep->gfporder);
1197 } else {
1198 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1199 }
1200 if (!page)
1201 return NULL;
1202 addr = page_address(page);
1203
1204 i = (1 << cachep->gfporder);
1205 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1206 atomic_add(i, &slab_reclaim_pages);
1207 add_page_state(nr_slab, i);
1208 while (i--) {
1209 SetPageSlab(page);
1210 page++;
1211 }
1212 return addr;
1213}
1214
1215/*
1216 * Interface to system's page release.
1217 */
1218static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1219{
1220 unsigned long i = (1<<cachep->gfporder);
1221 struct page *page = virt_to_page(addr);
1222 const unsigned long nr_freed = i;
1223
1224 while (i--) {
1225 if (!TestClearPageSlab(page))
1226 BUG();
1227 page++;
1228 }
1229 sub_page_state(nr_slab, nr_freed);
1230 if (current->reclaim_state)
1231 current->reclaim_state->reclaimed_slab += nr_freed;
1232 free_pages((unsigned long)addr, cachep->gfporder);
1233 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1234 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
1235}
1236
1237static void kmem_rcu_free(struct rcu_head *head)
1238{
1239 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
1240 kmem_cache_t *cachep = slab_rcu->cachep;
1241
1242 kmem_freepages(cachep, slab_rcu->addr);
1243 if (OFF_SLAB(cachep))
1244 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1245}
1246
1247#if DEBUG
1248
1249#ifdef CONFIG_DEBUG_PAGEALLOC
1250static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1251 unsigned long caller)
1252{
1253 int size = obj_reallen(cachep);
1254
1255 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
1256
1257 if (size < 5*sizeof(unsigned long))
1258 return;
1259
1260 *addr++=0x12345678;
1261 *addr++=caller;
1262 *addr++=smp_processor_id();
1263 size -= 3*sizeof(unsigned long);
1264 {
1265 unsigned long *sptr = &caller;
1266 unsigned long svalue;
1267
1268 while (!kstack_end(sptr)) {
1269 svalue = *sptr++;
1270 if (kernel_text_address(svalue)) {
1271 *addr++=svalue;
1272 size -= sizeof(unsigned long);
1273 if (size <= sizeof(unsigned long))
1274 break;
1275 }
1276 }
1277
1278 }
1279 *addr++=0x87654321;
1280}
1281#endif
1282
1283static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1284{
1285 int size = obj_reallen(cachep);
1286 addr = &((char*)addr)[obj_dbghead(cachep)];
1287
1288 memset(addr, val, size);
1289 *(unsigned char *)(addr+size-1) = POISON_END;
1290}
1291
1292static void dump_line(char *data, int offset, int limit)
1293{
1294 int i;
1295 printk(KERN_ERR "%03x:", offset);
1296 for (i=0;i<limit;i++) {
1297 printk(" %02x", (unsigned char)data[offset+i]);
1298 }
1299 printk("\n");
1300}
1301#endif
1302
1303#if DEBUG
1304
1305static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1306{
1307 int i, size;
1308 char *realobj;
1309
1310 if (cachep->flags & SLAB_RED_ZONE) {
1311 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1312 *dbg_redzone1(cachep, objp),
1313 *dbg_redzone2(cachep, objp));
1314 }
1315
1316 if (cachep->flags & SLAB_STORE_USER) {
1317 printk(KERN_ERR "Last user: [<%p>]",
1318 *dbg_userword(cachep, objp));
1319 print_symbol("(%s)",
1320 (unsigned long)*dbg_userword(cachep, objp));
1321 printk("\n");
1322 }
1323 realobj = (char*)objp+obj_dbghead(cachep);
1324 size = obj_reallen(cachep);
1325 for (i=0; i<size && lines;i+=16, lines--) {
1326 int limit;
1327 limit = 16;
1328 if (i+limit > size)
1329 limit = size-i;
1330 dump_line(realobj, i, limit);
1331 }
1332}
1333
1334static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1335{
1336 char *realobj;
1337 int size, i;
1338 int lines = 0;
1339
1340 realobj = (char*)objp+obj_dbghead(cachep);
1341 size = obj_reallen(cachep);
1342
1343 for (i=0;i<size;i++) {
1344 char exp = POISON_FREE;
1345 if (i == size-1)
1346 exp = POISON_END;
1347 if (realobj[i] != exp) {
1348 int limit;
1349 /* Mismatch ! */
1350 /* Print header */
1351 if (lines == 0) {
1352 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1353 realobj, size);
1354 print_objinfo(cachep, objp, 0);
1355 }
1356 /* Hexdump the affected line */
1357 i = (i/16)*16;
1358 limit = 16;
1359 if (i+limit > size)
1360 limit = size-i;
1361 dump_line(realobj, i, limit);
1362 i += 16;
1363 lines++;
1364 /* Limit to 5 lines */
1365 if (lines > 5)
1366 break;
1367 }
1368 }
1369 if (lines != 0) {
1370 /* Print some data about the neighboring objects, if they
1371 * exist:
1372 */
1373 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1374 int objnr;
1375
1376 objnr = (objp-slabp->s_mem)/cachep->objsize;
1377 if (objnr) {
1378 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1379 realobj = (char*)objp+obj_dbghead(cachep);
1380 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1381 realobj, size);
1382 print_objinfo(cachep, objp, 2);
1383 }
1384 if (objnr+1 < cachep->num) {
1385 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1386 realobj = (char*)objp+obj_dbghead(cachep);
1387 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1388 realobj, size);
1389 print_objinfo(cachep, objp, 2);
1390 }
1391 }
1392}
1393#endif
1394
1395/* Destroy all the objs in a slab, and release the mem back to the system.
1396 * Before calling the slab must have been unlinked from the cache.
1397 * The cache-lock is not held/needed.
1398 */
1399static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1400{
1401 void *addr = slabp->s_mem - slabp->colouroff;
1402
1403#if DEBUG
1404 int i;
1405 for (i = 0; i < cachep->num; i++) {
1406 void *objp = slabp->s_mem + cachep->objsize * i;
1407
1408 if (cachep->flags & SLAB_POISON) {
1409#ifdef CONFIG_DEBUG_PAGEALLOC
1410 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1411 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1412 else
1413 check_poison_obj(cachep, objp);
1414#else
1415 check_poison_obj(cachep, objp);
1416#endif
1417 }
1418 if (cachep->flags & SLAB_RED_ZONE) {
1419 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1420 slab_error(cachep, "start of a freed object "
1421 "was overwritten");
1422 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1423 slab_error(cachep, "end of a freed object "
1424 "was overwritten");
1425 }
1426 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1427 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1428 }
1429#else
1430 if (cachep->dtor) {
1431 int i;
1432 for (i = 0; i < cachep->num; i++) {
1433 void* objp = slabp->s_mem+cachep->objsize*i;
1434 (cachep->dtor)(objp, cachep, 0);
1435 }
1436 }
1437#endif
1438
1439 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1440 struct slab_rcu *slab_rcu;
1441
1442 slab_rcu = (struct slab_rcu *) slabp;
1443 slab_rcu->cachep = cachep;
1444 slab_rcu->addr = addr;
1445 call_rcu(&slab_rcu->head, kmem_rcu_free);
1446 } else {
1447 kmem_freepages(cachep, addr);
1448 if (OFF_SLAB(cachep))
1449 kmem_cache_free(cachep->slabp_cache, slabp);
1450 }
1451}
1452
Christoph Lametere498be72005-09-09 13:03:32 -07001453/* For setting up all the kmem_list3s for cache whose objsize is same
1454 as size of kmem_list3. */
1455static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1456{
1457 int node;
1458
1459 for_each_online_node(node) {
1460 cachep->nodelists[node] = &initkmem_list3[index+node];
1461 cachep->nodelists[node]->next_reap = jiffies +
1462 REAPTIMEOUT_LIST3 +
1463 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1464 }
1465}
1466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467/**
1468 * kmem_cache_create - Create a cache.
1469 * @name: A string which is used in /proc/slabinfo to identify this cache.
1470 * @size: The size of objects to be created in this cache.
1471 * @align: The required alignment for the objects.
1472 * @flags: SLAB flags
1473 * @ctor: A constructor for the objects.
1474 * @dtor: A destructor for the objects.
1475 *
1476 * Returns a ptr to the cache on success, NULL on failure.
1477 * Cannot be called within a int, but can be interrupted.
1478 * The @ctor is run when new pages are allocated by the cache
1479 * and the @dtor is run before the pages are handed back.
1480 *
1481 * @name must be valid until the cache is destroyed. This implies that
1482 * the module calling this has to destroy the cache before getting
1483 * unloaded.
1484 *
1485 * The flags are
1486 *
1487 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1488 * to catch references to uninitialised memory.
1489 *
1490 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1491 * for buffer overruns.
1492 *
1493 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1494 * memory pressure.
1495 *
1496 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1497 * cacheline. This can be beneficial if you're counting cycles as closely
1498 * as davem.
1499 */
1500kmem_cache_t *
1501kmem_cache_create (const char *name, size_t size, size_t align,
1502 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1503 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1504{
1505 size_t left_over, slab_size, ralign;
1506 kmem_cache_t *cachep = NULL;
1507
1508 /*
1509 * Sanity checks... these are all serious usage bugs.
1510 */
1511 if ((!name) ||
1512 in_interrupt() ||
1513 (size < BYTES_PER_WORD) ||
1514 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1515 (dtor && !ctor)) {
1516 printk(KERN_ERR "%s: Early error in slab %s\n",
1517 __FUNCTION__, name);
1518 BUG();
1519 }
1520
1521#if DEBUG
1522 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1523 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1524 /* No constructor, but inital state check requested */
1525 printk(KERN_ERR "%s: No con, but init state check "
1526 "requested - %s\n", __FUNCTION__, name);
1527 flags &= ~SLAB_DEBUG_INITIAL;
1528 }
1529
1530#if FORCED_DEBUG
1531 /*
1532 * Enable redzoning and last user accounting, except for caches with
1533 * large objects, if the increased size would increase the object size
1534 * above the next power of two: caches with object sizes just above a
1535 * power of two have a significant amount of internal fragmentation.
1536 */
1537 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1538 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1539 if (!(flags & SLAB_DESTROY_BY_RCU))
1540 flags |= SLAB_POISON;
1541#endif
1542 if (flags & SLAB_DESTROY_BY_RCU)
1543 BUG_ON(flags & SLAB_POISON);
1544#endif
1545 if (flags & SLAB_DESTROY_BY_RCU)
1546 BUG_ON(dtor);
1547
1548 /*
1549 * Always checks flags, a caller might be expecting debug
1550 * support which isn't available.
1551 */
1552 if (flags & ~CREATE_MASK)
1553 BUG();
1554
1555 /* Check that size is in terms of words. This is needed to avoid
1556 * unaligned accesses for some archs when redzoning is used, and makes
1557 * sure any on-slab bufctl's are also correctly aligned.
1558 */
1559 if (size & (BYTES_PER_WORD-1)) {
1560 size += (BYTES_PER_WORD-1);
1561 size &= ~(BYTES_PER_WORD-1);
1562 }
1563
1564 /* calculate out the final buffer alignment: */
1565 /* 1) arch recommendation: can be overridden for debug */
1566 if (flags & SLAB_HWCACHE_ALIGN) {
1567 /* Default alignment: as specified by the arch code.
1568 * Except if an object is really small, then squeeze multiple
1569 * objects into one cacheline.
1570 */
1571 ralign = cache_line_size();
1572 while (size <= ralign/2)
1573 ralign /= 2;
1574 } else {
1575 ralign = BYTES_PER_WORD;
1576 }
1577 /* 2) arch mandated alignment: disables debug if necessary */
1578 if (ralign < ARCH_SLAB_MINALIGN) {
1579 ralign = ARCH_SLAB_MINALIGN;
1580 if (ralign > BYTES_PER_WORD)
1581 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1582 }
1583 /* 3) caller mandated alignment: disables debug if necessary */
1584 if (ralign < align) {
1585 ralign = align;
1586 if (ralign > BYTES_PER_WORD)
1587 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1588 }
1589 /* 4) Store it. Note that the debug code below can reduce
1590 * the alignment to BYTES_PER_WORD.
1591 */
1592 align = ralign;
1593
1594 /* Get cache's description obj. */
1595 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1596 if (!cachep)
1597 goto opps;
1598 memset(cachep, 0, sizeof(kmem_cache_t));
1599
1600#if DEBUG
1601 cachep->reallen = size;
1602
1603 if (flags & SLAB_RED_ZONE) {
1604 /* redzoning only works with word aligned caches */
1605 align = BYTES_PER_WORD;
1606
1607 /* add space for red zone words */
1608 cachep->dbghead += BYTES_PER_WORD;
1609 size += 2*BYTES_PER_WORD;
1610 }
1611 if (flags & SLAB_STORE_USER) {
1612 /* user store requires word alignment and
1613 * one word storage behind the end of the real
1614 * object.
1615 */
1616 align = BYTES_PER_WORD;
1617 size += BYTES_PER_WORD;
1618 }
1619#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lametere498be72005-09-09 13:03:32 -07001620 if (size >= malloc_sizes[INDEX_L3+1].cs_size && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 cachep->dbghead += PAGE_SIZE - size;
1622 size = PAGE_SIZE;
1623 }
1624#endif
1625#endif
1626
1627 /* Determine if the slab management is 'on' or 'off' slab. */
1628 if (size >= (PAGE_SIZE>>3))
1629 /*
1630 * Size is large, assume best to place the slab management obj
1631 * off-slab (should allow better packing of objs).
1632 */
1633 flags |= CFLGS_OFF_SLAB;
1634
1635 size = ALIGN(size, align);
1636
1637 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1638 /*
1639 * A VFS-reclaimable slab tends to have most allocations
1640 * as GFP_NOFS and we really don't want to have to be allocating
1641 * higher-order pages when we are unable to shrink dcache.
1642 */
1643 cachep->gfporder = 0;
1644 cache_estimate(cachep->gfporder, size, align, flags,
1645 &left_over, &cachep->num);
1646 } else {
1647 /*
1648 * Calculate size (in pages) of slabs, and the num of objs per
1649 * slab. This could be made much more intelligent. For now,
1650 * try to avoid using high page-orders for slabs. When the
1651 * gfp() funcs are more friendly towards high-order requests,
1652 * this should be changed.
1653 */
1654 do {
1655 unsigned int break_flag = 0;
1656cal_wastage:
1657 cache_estimate(cachep->gfporder, size, align, flags,
1658 &left_over, &cachep->num);
1659 if (break_flag)
1660 break;
1661 if (cachep->gfporder >= MAX_GFP_ORDER)
1662 break;
1663 if (!cachep->num)
1664 goto next;
1665 if (flags & CFLGS_OFF_SLAB &&
1666 cachep->num > offslab_limit) {
1667 /* This num of objs will cause problems. */
1668 cachep->gfporder--;
1669 break_flag++;
1670 goto cal_wastage;
1671 }
1672
1673 /*
1674 * Large num of objs is good, but v. large slabs are
1675 * currently bad for the gfp()s.
1676 */
1677 if (cachep->gfporder >= slab_break_gfp_order)
1678 break;
1679
1680 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1681 break; /* Acceptable internal fragmentation. */
1682next:
1683 cachep->gfporder++;
1684 } while (1);
1685 }
1686
1687 if (!cachep->num) {
1688 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1689 kmem_cache_free(&cache_cache, cachep);
1690 cachep = NULL;
1691 goto opps;
1692 }
1693 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1694 + sizeof(struct slab), align);
1695
1696 /*
1697 * If the slab has been placed off-slab, and we have enough space then
1698 * move it on-slab. This is at the expense of any extra colouring.
1699 */
1700 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1701 flags &= ~CFLGS_OFF_SLAB;
1702 left_over -= slab_size;
1703 }
1704
1705 if (flags & CFLGS_OFF_SLAB) {
1706 /* really off slab. No need for manual alignment */
1707 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1708 }
1709
1710 cachep->colour_off = cache_line_size();
1711 /* Offset must be a multiple of the alignment. */
1712 if (cachep->colour_off < align)
1713 cachep->colour_off = align;
1714 cachep->colour = left_over/cachep->colour_off;
1715 cachep->slab_size = slab_size;
1716 cachep->flags = flags;
1717 cachep->gfpflags = 0;
1718 if (flags & SLAB_CACHE_DMA)
1719 cachep->gfpflags |= GFP_DMA;
1720 spin_lock_init(&cachep->spinlock);
1721 cachep->objsize = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722
1723 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001724 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 cachep->ctor = ctor;
1726 cachep->dtor = dtor;
1727 cachep->name = name;
1728
1729 /* Don't let CPUs to come and go */
1730 lock_cpu_hotplug();
1731
1732 if (g_cpucache_up == FULL) {
1733 enable_cpucache(cachep);
1734 } else {
1735 if (g_cpucache_up == NONE) {
1736 /* Note: the first kmem_cache_create must create
1737 * the cache that's used by kmalloc(24), otherwise
1738 * the creation of further caches will BUG().
1739 */
Christoph Lametere498be72005-09-09 13:03:32 -07001740 cachep->array[smp_processor_id()] =
1741 &initarray_generic.cache;
1742
1743 /* If the cache that's used by
1744 * kmalloc(sizeof(kmem_list3)) is the first cache,
1745 * then we need to set up all its list3s, otherwise
1746 * the creation of further caches will BUG().
1747 */
1748 set_up_list3s(cachep, SIZE_AC);
1749 if (INDEX_AC == INDEX_L3)
1750 g_cpucache_up = PARTIAL_L3;
1751 else
1752 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001754 cachep->array[smp_processor_id()] =
1755 kmalloc(sizeof(struct arraycache_init),
1756 GFP_KERNEL);
1757
1758 if (g_cpucache_up == PARTIAL_AC) {
1759 set_up_list3s(cachep, SIZE_L3);
1760 g_cpucache_up = PARTIAL_L3;
1761 } else {
1762 int node;
1763 for_each_online_node(node) {
1764
1765 cachep->nodelists[node] =
1766 kmalloc_node(sizeof(struct kmem_list3),
1767 GFP_KERNEL, node);
1768 BUG_ON(!cachep->nodelists[node]);
1769 kmem_list3_init(cachep->nodelists[node]);
1770 }
1771 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 }
Christoph Lametere498be72005-09-09 13:03:32 -07001773 cachep->nodelists[numa_node_id()]->next_reap =
1774 jiffies + REAPTIMEOUT_LIST3 +
1775 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1776
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 BUG_ON(!ac_data(cachep));
1778 ac_data(cachep)->avail = 0;
1779 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1780 ac_data(cachep)->batchcount = 1;
1781 ac_data(cachep)->touched = 0;
1782 cachep->batchcount = 1;
1783 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 }
1785
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 /* Need the semaphore to access the chain. */
1787 down(&cache_chain_sem);
1788 {
1789 struct list_head *p;
1790 mm_segment_t old_fs;
1791
1792 old_fs = get_fs();
1793 set_fs(KERNEL_DS);
1794 list_for_each(p, &cache_chain) {
1795 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1796 char tmp;
1797 /* This happens when the module gets unloaded and doesn't
1798 destroy its slab cache and noone else reuses the vmalloc
1799 area of the module. Print a warning. */
1800 if (__get_user(tmp,pc->name)) {
1801 printk("SLAB: cache with size %d has lost its name\n",
1802 pc->objsize);
1803 continue;
1804 }
1805 if (!strcmp(pc->name,name)) {
1806 printk("kmem_cache_create: duplicate cache %s\n",name);
1807 up(&cache_chain_sem);
1808 unlock_cpu_hotplug();
1809 BUG();
1810 }
1811 }
1812 set_fs(old_fs);
1813 }
1814
1815 /* cache setup completed, link it into the list */
1816 list_add(&cachep->next, &cache_chain);
1817 up(&cache_chain_sem);
1818 unlock_cpu_hotplug();
1819opps:
1820 if (!cachep && (flags & SLAB_PANIC))
1821 panic("kmem_cache_create(): failed to create slab `%s'\n",
1822 name);
1823 return cachep;
1824}
1825EXPORT_SYMBOL(kmem_cache_create);
1826
1827#if DEBUG
1828static void check_irq_off(void)
1829{
1830 BUG_ON(!irqs_disabled());
1831}
1832
1833static void check_irq_on(void)
1834{
1835 BUG_ON(irqs_disabled());
1836}
1837
1838static void check_spinlock_acquired(kmem_cache_t *cachep)
1839{
1840#ifdef CONFIG_SMP
1841 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001842 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843#endif
1844}
Christoph Lametere498be72005-09-09 13:03:32 -07001845
1846static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1847{
1848#ifdef CONFIG_SMP
1849 check_irq_off();
1850 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1851#endif
1852}
1853
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854#else
1855#define check_irq_off() do { } while(0)
1856#define check_irq_on() do { } while(0)
1857#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001858#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859#endif
1860
1861/*
1862 * Waits for all CPUs to execute func().
1863 */
1864static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1865{
1866 check_irq_on();
1867 preempt_disable();
1868
1869 local_irq_disable();
1870 func(arg);
1871 local_irq_enable();
1872
1873 if (smp_call_function(func, arg, 1, 1))
1874 BUG();
1875
1876 preempt_enable();
1877}
1878
1879static void drain_array_locked(kmem_cache_t* cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07001880 struct array_cache *ac, int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881
1882static void do_drain(void *arg)
1883{
1884 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1885 struct array_cache *ac;
1886
1887 check_irq_off();
1888 ac = ac_data(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07001889 spin_lock(&cachep->nodelists[numa_node_id()]->list_lock);
1890 free_block(cachep, ac->entry, ac->avail);
1891 spin_unlock(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 ac->avail = 0;
1893}
1894
1895static void drain_cpu_caches(kmem_cache_t *cachep)
1896{
Christoph Lametere498be72005-09-09 13:03:32 -07001897 struct kmem_list3 *l3;
1898 int node;
1899
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 smp_call_function_all_cpus(do_drain, cachep);
1901 check_irq_on();
1902 spin_lock_irq(&cachep->spinlock);
Christoph Lametere498be72005-09-09 13:03:32 -07001903 for_each_online_node(node) {
1904 l3 = cachep->nodelists[node];
1905 if (l3) {
1906 spin_lock(&l3->list_lock);
1907 drain_array_locked(cachep, l3->shared, 1, node);
1908 spin_unlock(&l3->list_lock);
1909 if (l3->alien)
1910 drain_alien_cache(cachep, l3);
1911 }
1912 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 spin_unlock_irq(&cachep->spinlock);
1914}
1915
Christoph Lametere498be72005-09-09 13:03:32 -07001916static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917{
1918 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07001919 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 int ret;
1921
Christoph Lametere498be72005-09-09 13:03:32 -07001922 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 struct list_head *p;
1924
Christoph Lametere498be72005-09-09 13:03:32 -07001925 p = l3->slabs_free.prev;
1926 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 break;
1928
Christoph Lametere498be72005-09-09 13:03:32 -07001929 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930#if DEBUG
1931 if (slabp->inuse)
1932 BUG();
1933#endif
1934 list_del(&slabp->list);
1935
Christoph Lametere498be72005-09-09 13:03:32 -07001936 l3->free_objects -= cachep->num;
1937 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07001939 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 }
Christoph Lametere498be72005-09-09 13:03:32 -07001941 ret = !list_empty(&l3->slabs_full) ||
1942 !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 return ret;
1944}
1945
Christoph Lametere498be72005-09-09 13:03:32 -07001946static int __cache_shrink(kmem_cache_t *cachep)
1947{
1948 int ret = 0, i = 0;
1949 struct kmem_list3 *l3;
1950
1951 drain_cpu_caches(cachep);
1952
1953 check_irq_on();
1954 for_each_online_node(i) {
1955 l3 = cachep->nodelists[i];
1956 if (l3) {
1957 spin_lock_irq(&l3->list_lock);
1958 ret += __node_shrink(cachep, i);
1959 spin_unlock_irq(&l3->list_lock);
1960 }
1961 }
1962 return (ret ? 1 : 0);
1963}
1964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965/**
1966 * kmem_cache_shrink - Shrink a cache.
1967 * @cachep: The cache to shrink.
1968 *
1969 * Releases as many slabs as possible for a cache.
1970 * To help debugging, a zero exit status indicates all slabs were released.
1971 */
1972int kmem_cache_shrink(kmem_cache_t *cachep)
1973{
1974 if (!cachep || in_interrupt())
1975 BUG();
1976
1977 return __cache_shrink(cachep);
1978}
1979EXPORT_SYMBOL(kmem_cache_shrink);
1980
1981/**
1982 * kmem_cache_destroy - delete a cache
1983 * @cachep: the cache to destroy
1984 *
1985 * Remove a kmem_cache_t object from the slab cache.
1986 * Returns 0 on success.
1987 *
1988 * It is expected this function will be called by a module when it is
1989 * unloaded. This will remove the cache completely, and avoid a duplicate
1990 * cache being allocated each time a module is loaded and unloaded, if the
1991 * module doesn't have persistent in-kernel storage across loads and unloads.
1992 *
1993 * The cache must be empty before calling this function.
1994 *
1995 * The caller must guarantee that noone will allocate memory from the cache
1996 * during the kmem_cache_destroy().
1997 */
1998int kmem_cache_destroy(kmem_cache_t * cachep)
1999{
2000 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002001 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002
2003 if (!cachep || in_interrupt())
2004 BUG();
2005
2006 /* Don't let CPUs to come and go */
2007 lock_cpu_hotplug();
2008
2009 /* Find the cache in the chain of caches. */
2010 down(&cache_chain_sem);
2011 /*
2012 * the chain is never empty, cache_cache is never destroyed
2013 */
2014 list_del(&cachep->next);
2015 up(&cache_chain_sem);
2016
2017 if (__cache_shrink(cachep)) {
2018 slab_error(cachep, "Can't free all objects");
2019 down(&cache_chain_sem);
2020 list_add(&cachep->next,&cache_chain);
2021 up(&cache_chain_sem);
2022 unlock_cpu_hotplug();
2023 return 1;
2024 }
2025
2026 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002027 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
Christoph Lametere498be72005-09-09 13:03:32 -07002029 for_each_online_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 kfree(cachep->array[i]);
2031
2032 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002033 for_each_online_node(i) {
2034 if ((l3 = cachep->nodelists[i])) {
2035 kfree(l3->shared);
2036 free_alien_cache(l3->alien);
2037 kfree(l3);
2038 }
2039 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 kmem_cache_free(&cache_cache, cachep);
2041
2042 unlock_cpu_hotplug();
2043
2044 return 0;
2045}
2046EXPORT_SYMBOL(kmem_cache_destroy);
2047
2048/* Get the memory for a slab management obj. */
Christoph Lametere498be72005-09-09 13:03:32 -07002049static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2050 int colour_off, unsigned int __nocast local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051{
2052 struct slab *slabp;
2053
2054 if (OFF_SLAB(cachep)) {
2055 /* Slab management obj is off-slab. */
2056 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2057 if (!slabp)
2058 return NULL;
2059 } else {
2060 slabp = objp+colour_off;
2061 colour_off += cachep->slab_size;
2062 }
2063 slabp->inuse = 0;
2064 slabp->colouroff = colour_off;
2065 slabp->s_mem = objp+colour_off;
2066
2067 return slabp;
2068}
2069
2070static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2071{
2072 return (kmem_bufctl_t *)(slabp+1);
2073}
2074
2075static void cache_init_objs(kmem_cache_t *cachep,
2076 struct slab *slabp, unsigned long ctor_flags)
2077{
2078 int i;
2079
2080 for (i = 0; i < cachep->num; i++) {
Christoph Lametere498be72005-09-09 13:03:32 -07002081 void *objp = slabp->s_mem+cachep->objsize*i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082#if DEBUG
2083 /* need to poison the objs? */
2084 if (cachep->flags & SLAB_POISON)
2085 poison_obj(cachep, objp, POISON_FREE);
2086 if (cachep->flags & SLAB_STORE_USER)
2087 *dbg_userword(cachep, objp) = NULL;
2088
2089 if (cachep->flags & SLAB_RED_ZONE) {
2090 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2091 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2092 }
2093 /*
2094 * Constructors are not allowed to allocate memory from
2095 * the same cache which they are a constructor for.
2096 * Otherwise, deadlock. They must also be threaded.
2097 */
2098 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2099 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
2100
2101 if (cachep->flags & SLAB_RED_ZONE) {
2102 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2103 slab_error(cachep, "constructor overwrote the"
2104 " end of an object");
2105 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2106 slab_error(cachep, "constructor overwrote the"
2107 " start of an object");
2108 }
2109 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2110 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2111#else
2112 if (cachep->ctor)
2113 cachep->ctor(objp, cachep, ctor_flags);
2114#endif
2115 slab_bufctl(slabp)[i] = i+1;
2116 }
2117 slab_bufctl(slabp)[i-1] = BUFCTL_END;
2118 slabp->free = 0;
2119}
2120
2121static void kmem_flagcheck(kmem_cache_t *cachep, unsigned int flags)
2122{
2123 if (flags & SLAB_DMA) {
2124 if (!(cachep->gfpflags & GFP_DMA))
2125 BUG();
2126 } else {
2127 if (cachep->gfpflags & GFP_DMA)
2128 BUG();
2129 }
2130}
2131
2132static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2133{
2134 int i;
2135 struct page *page;
2136
2137 /* Nasty!!!!!! I hope this is OK. */
2138 i = 1 << cachep->gfporder;
2139 page = virt_to_page(objp);
2140 do {
2141 SET_PAGE_CACHE(page, cachep);
2142 SET_PAGE_SLAB(page, slabp);
2143 page++;
2144 } while (--i);
2145}
2146
2147/*
2148 * Grow (by 1) the number of slabs within a cache. This is called by
2149 * kmem_cache_alloc() when there are no active objs left in a cache.
2150 */
2151static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
2152{
2153 struct slab *slabp;
2154 void *objp;
2155 size_t offset;
2156 unsigned int local_flags;
2157 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002158 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
2160 /* Be lazy and only check for valid flags here,
2161 * keeping it out of the critical path in kmem_cache_alloc().
2162 */
2163 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
2164 BUG();
2165 if (flags & SLAB_NO_GROW)
2166 return 0;
2167
2168 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2169 local_flags = (flags & SLAB_LEVEL_MASK);
2170 if (!(local_flags & __GFP_WAIT))
2171 /*
2172 * Not allowed to sleep. Need to tell a constructor about
2173 * this - it might need to know...
2174 */
2175 ctor_flags |= SLAB_CTOR_ATOMIC;
2176
2177 /* About to mess with non-constant members - lock. */
2178 check_irq_off();
2179 spin_lock(&cachep->spinlock);
2180
2181 /* Get colour for the slab, and cal the next value. */
2182 offset = cachep->colour_next;
2183 cachep->colour_next++;
2184 if (cachep->colour_next >= cachep->colour)
2185 cachep->colour_next = 0;
2186 offset *= cachep->colour_off;
2187
2188 spin_unlock(&cachep->spinlock);
2189
Christoph Lametere498be72005-09-09 13:03:32 -07002190 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 if (local_flags & __GFP_WAIT)
2192 local_irq_enable();
2193
2194 /*
2195 * The test for missing atomic flag is performed here, rather than
2196 * the more obvious place, simply to reduce the critical path length
2197 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2198 * will eventually be caught here (where it matters).
2199 */
2200 kmem_flagcheck(cachep, flags);
2201
Christoph Lametere498be72005-09-09 13:03:32 -07002202 /* Get mem for the objs.
2203 * Attempt to allocate a physical page from 'nodeid',
2204 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2206 goto failed;
2207
2208 /* Get slab management. */
2209 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2210 goto opps1;
2211
Christoph Lametere498be72005-09-09 13:03:32 -07002212 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 set_slab_attr(cachep, slabp, objp);
2214
2215 cache_init_objs(cachep, slabp, ctor_flags);
2216
2217 if (local_flags & __GFP_WAIT)
2218 local_irq_disable();
2219 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002220 l3 = cachep->nodelists[nodeid];
2221 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
2223 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002224 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002226 l3->free_objects += cachep->num;
2227 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 return 1;
2229opps1:
2230 kmem_freepages(cachep, objp);
2231failed:
2232 if (local_flags & __GFP_WAIT)
2233 local_irq_disable();
2234 return 0;
2235}
2236
2237#if DEBUG
2238
2239/*
2240 * Perform extra freeing checks:
2241 * - detect bad pointers.
2242 * - POISON/RED_ZONE checking
2243 * - destructor calls, for caches with POISON+dtor
2244 */
2245static void kfree_debugcheck(const void *objp)
2246{
2247 struct page *page;
2248
2249 if (!virt_addr_valid(objp)) {
2250 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2251 (unsigned long)objp);
2252 BUG();
2253 }
2254 page = virt_to_page(objp);
2255 if (!PageSlab(page)) {
2256 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
2257 BUG();
2258 }
2259}
2260
2261static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2262 void *caller)
2263{
2264 struct page *page;
2265 unsigned int objnr;
2266 struct slab *slabp;
2267
2268 objp -= obj_dbghead(cachep);
2269 kfree_debugcheck(objp);
2270 page = virt_to_page(objp);
2271
2272 if (GET_PAGE_CACHE(page) != cachep) {
2273 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2274 GET_PAGE_CACHE(page),cachep);
2275 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2276 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
2277 WARN_ON(1);
2278 }
2279 slabp = GET_PAGE_SLAB(page);
2280
2281 if (cachep->flags & SLAB_RED_ZONE) {
2282 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2283 slab_error(cachep, "double free, or memory outside"
2284 " object was overwritten");
2285 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2286 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2287 }
2288 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2289 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2290 }
2291 if (cachep->flags & SLAB_STORE_USER)
2292 *dbg_userword(cachep, objp) = caller;
2293
2294 objnr = (objp-slabp->s_mem)/cachep->objsize;
2295
2296 BUG_ON(objnr >= cachep->num);
2297 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
2298
2299 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2300 /* Need to call the slab's constructor so the
2301 * caller can perform a verify of its state (debugging).
2302 * Called without the cache-lock held.
2303 */
2304 cachep->ctor(objp+obj_dbghead(cachep),
2305 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
2306 }
2307 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2308 /* we want to cache poison the object,
2309 * call the destruction callback
2310 */
2311 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
2312 }
2313 if (cachep->flags & SLAB_POISON) {
2314#ifdef CONFIG_DEBUG_PAGEALLOC
2315 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2316 store_stackinfo(cachep, objp, (unsigned long)caller);
2317 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2318 } else {
2319 poison_obj(cachep, objp, POISON_FREE);
2320 }
2321#else
2322 poison_obj(cachep, objp, POISON_FREE);
2323#endif
2324 }
2325 return objp;
2326}
2327
2328static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2329{
2330 kmem_bufctl_t i;
2331 int entries = 0;
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 /* Check slab's freelist to see if this obj is there. */
2334 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2335 entries++;
2336 if (entries > cachep->num || i >= cachep->num)
2337 goto bad;
2338 }
2339 if (entries != cachep->num - slabp->inuse) {
2340bad:
2341 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2342 cachep->name, cachep->num, slabp, slabp->inuse);
2343 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
2344 if ((i%16)==0)
2345 printk("\n%03x:", i);
2346 printk(" %02x", ((unsigned char*)slabp)[i]);
2347 }
2348 printk("\n");
2349 BUG();
2350 }
2351}
2352#else
2353#define kfree_debugcheck(x) do { } while(0)
2354#define cache_free_debugcheck(x,objp,z) (objp)
2355#define check_slabp(x,y) do { } while(0)
2356#endif
2357
2358static void *cache_alloc_refill(kmem_cache_t *cachep, unsigned int __nocast flags)
2359{
2360 int batchcount;
2361 struct kmem_list3 *l3;
2362 struct array_cache *ac;
2363
2364 check_irq_off();
2365 ac = ac_data(cachep);
2366retry:
2367 batchcount = ac->batchcount;
2368 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2369 /* if there was little recent activity on this
2370 * cache, then perform only a partial refill.
2371 * Otherwise we could generate refill bouncing.
2372 */
2373 batchcount = BATCHREFILL_LIMIT;
2374 }
Christoph Lametere498be72005-09-09 13:03:32 -07002375 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
Christoph Lametere498be72005-09-09 13:03:32 -07002377 BUG_ON(ac->avail > 0 || !l3);
2378 spin_lock(&l3->list_lock);
2379
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 if (l3->shared) {
2381 struct array_cache *shared_array = l3->shared;
2382 if (shared_array->avail) {
2383 if (batchcount > shared_array->avail)
2384 batchcount = shared_array->avail;
2385 shared_array->avail -= batchcount;
2386 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002387 memcpy(ac->entry,
2388 &(shared_array->entry[shared_array->avail]),
2389 sizeof(void*)*batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 shared_array->touched = 1;
2391 goto alloc_done;
2392 }
2393 }
2394 while (batchcount > 0) {
2395 struct list_head *entry;
2396 struct slab *slabp;
2397 /* Get slab alloc is to come from. */
2398 entry = l3->slabs_partial.next;
2399 if (entry == &l3->slabs_partial) {
2400 l3->free_touched = 1;
2401 entry = l3->slabs_free.next;
2402 if (entry == &l3->slabs_free)
2403 goto must_grow;
2404 }
2405
2406 slabp = list_entry(entry, struct slab, list);
2407 check_slabp(cachep, slabp);
2408 check_spinlock_acquired(cachep);
2409 while (slabp->inuse < cachep->num && batchcount--) {
2410 kmem_bufctl_t next;
2411 STATS_INC_ALLOCED(cachep);
2412 STATS_INC_ACTIVE(cachep);
2413 STATS_SET_HIGH(cachep);
2414
2415 /* get obj pointer */
Christoph Lametere498be72005-09-09 13:03:32 -07002416 ac->entry[ac->avail++] = slabp->s_mem +
2417 slabp->free*cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418
2419 slabp->inuse++;
2420 next = slab_bufctl(slabp)[slabp->free];
2421#if DEBUG
2422 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2423#endif
2424 slabp->free = next;
2425 }
2426 check_slabp(cachep, slabp);
2427
2428 /* move slabp to correct slabp list: */
2429 list_del(&slabp->list);
2430 if (slabp->free == BUFCTL_END)
2431 list_add(&slabp->list, &l3->slabs_full);
2432 else
2433 list_add(&slabp->list, &l3->slabs_partial);
2434 }
2435
2436must_grow:
2437 l3->free_objects -= ac->avail;
2438alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002439 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440
2441 if (unlikely(!ac->avail)) {
2442 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002443 x = cache_grow(cachep, flags, numa_node_id());
2444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 // cache_grow can reenable interrupts, then ac could change.
2446 ac = ac_data(cachep);
2447 if (!x && ac->avail == 0) // no objects in sight? abort
2448 return NULL;
2449
2450 if (!ac->avail) // objects refilled by interrupt?
2451 goto retry;
2452 }
2453 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002454 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455}
2456
2457static inline void
2458cache_alloc_debugcheck_before(kmem_cache_t *cachep, unsigned int __nocast flags)
2459{
2460 might_sleep_if(flags & __GFP_WAIT);
2461#if DEBUG
2462 kmem_flagcheck(cachep, flags);
2463#endif
2464}
2465
2466#if DEBUG
2467static void *
2468cache_alloc_debugcheck_after(kmem_cache_t *cachep,
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002469 unsigned int __nocast flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470{
2471 if (!objp)
2472 return objp;
2473 if (cachep->flags & SLAB_POISON) {
2474#ifdef CONFIG_DEBUG_PAGEALLOC
2475 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2476 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2477 else
2478 check_poison_obj(cachep, objp);
2479#else
2480 check_poison_obj(cachep, objp);
2481#endif
2482 poison_obj(cachep, objp, POISON_INUSE);
2483 }
2484 if (cachep->flags & SLAB_STORE_USER)
2485 *dbg_userword(cachep, objp) = caller;
2486
2487 if (cachep->flags & SLAB_RED_ZONE) {
2488 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2489 slab_error(cachep, "double free, or memory outside"
2490 " object was overwritten");
2491 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2492 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2493 }
2494 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2495 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2496 }
2497 objp += obj_dbghead(cachep);
2498 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2499 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2500
2501 if (!(flags & __GFP_WAIT))
2502 ctor_flags |= SLAB_CTOR_ATOMIC;
2503
2504 cachep->ctor(objp, cachep, ctor_flags);
2505 }
2506 return objp;
2507}
2508#else
2509#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2510#endif
2511
2512
2513static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2514{
2515 unsigned long save_flags;
2516 void* objp;
2517 struct array_cache *ac;
2518
2519 cache_alloc_debugcheck_before(cachep, flags);
2520
2521 local_irq_save(save_flags);
2522 ac = ac_data(cachep);
2523 if (likely(ac->avail)) {
2524 STATS_INC_ALLOCHIT(cachep);
2525 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002526 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 } else {
2528 STATS_INC_ALLOCMISS(cachep);
2529 objp = cache_alloc_refill(cachep, flags);
2530 }
2531 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002532 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2533 __builtin_return_address(0));
2534 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 return objp;
2536}
2537
Christoph Lametere498be72005-09-09 13:03:32 -07002538#ifdef CONFIG_NUMA
2539/*
2540 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 */
Christoph Lametere498be72005-09-09 13:03:32 -07002542static void *__cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid)
2543{
2544 struct list_head *entry;
2545 struct slab *slabp;
2546 struct kmem_list3 *l3;
2547 void *obj;
2548 kmem_bufctl_t next;
2549 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550
Christoph Lametere498be72005-09-09 13:03:32 -07002551 l3 = cachep->nodelists[nodeid];
2552 BUG_ON(!l3);
2553
2554retry:
2555 spin_lock(&l3->list_lock);
2556 entry = l3->slabs_partial.next;
2557 if (entry == &l3->slabs_partial) {
2558 l3->free_touched = 1;
2559 entry = l3->slabs_free.next;
2560 if (entry == &l3->slabs_free)
2561 goto must_grow;
2562 }
2563
2564 slabp = list_entry(entry, struct slab, list);
2565 check_spinlock_acquired_node(cachep, nodeid);
2566 check_slabp(cachep, slabp);
2567
2568 STATS_INC_NODEALLOCS(cachep);
2569 STATS_INC_ACTIVE(cachep);
2570 STATS_SET_HIGH(cachep);
2571
2572 BUG_ON(slabp->inuse == cachep->num);
2573
2574 /* get obj pointer */
2575 obj = slabp->s_mem + slabp->free*cachep->objsize;
2576 slabp->inuse++;
2577 next = slab_bufctl(slabp)[slabp->free];
2578#if DEBUG
2579 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2580#endif
2581 slabp->free = next;
2582 check_slabp(cachep, slabp);
2583 l3->free_objects--;
2584 /* move slabp to correct slabp list: */
2585 list_del(&slabp->list);
2586
2587 if (slabp->free == BUFCTL_END) {
2588 list_add(&slabp->list, &l3->slabs_full);
2589 } else {
2590 list_add(&slabp->list, &l3->slabs_partial);
2591 }
2592
2593 spin_unlock(&l3->list_lock);
2594 goto done;
2595
2596must_grow:
2597 spin_unlock(&l3->list_lock);
2598 x = cache_grow(cachep, flags, nodeid);
2599
2600 if (!x)
2601 return NULL;
2602
2603 goto retry;
2604done:
2605 return obj;
2606}
2607#endif
2608
2609/*
2610 * Caller needs to acquire correct kmem_list's list_lock
2611 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2613{
2614 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002615 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616
2617 for (i = 0; i < nr_objects; i++) {
2618 void *objp = objpp[i];
2619 struct slab *slabp;
2620 unsigned int objnr;
Christoph Lametere498be72005-09-09 13:03:32 -07002621 int nodeid = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622
2623 slabp = GET_PAGE_SLAB(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002624 nodeid = slabp->nodeid;
2625 l3 = cachep->nodelists[nodeid];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 list_del(&slabp->list);
2627 objnr = (objp - slabp->s_mem) / cachep->objsize;
Christoph Lametere498be72005-09-09 13:03:32 -07002628 check_spinlock_acquired_node(cachep, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002630
2631
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632#if DEBUG
2633 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
Christoph Lametere498be72005-09-09 13:03:32 -07002634 printk(KERN_ERR "slab: double free detected in cache "
2635 "'%s', objp %p\n", cachep->name, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 BUG();
2637 }
2638#endif
2639 slab_bufctl(slabp)[objnr] = slabp->free;
2640 slabp->free = objnr;
2641 STATS_DEC_ACTIVE(cachep);
2642 slabp->inuse--;
Christoph Lametere498be72005-09-09 13:03:32 -07002643 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 check_slabp(cachep, slabp);
2645
2646 /* fixup slab chains */
2647 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002648 if (l3->free_objects > l3->free_limit) {
2649 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 slab_destroy(cachep, slabp);
2651 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002652 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 }
2654 } else {
2655 /* Unconditionally move a slab to the end of the
2656 * partial list on free - maximum time for the
2657 * other objects to be freed, too.
2658 */
Christoph Lametere498be72005-09-09 13:03:32 -07002659 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 }
2661 }
2662}
2663
2664static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2665{
2666 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002667 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
2669 batchcount = ac->batchcount;
2670#if DEBUG
2671 BUG_ON(!batchcount || batchcount > ac->avail);
2672#endif
2673 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002674 l3 = cachep->nodelists[numa_node_id()];
2675 spin_lock(&l3->list_lock);
2676 if (l3->shared) {
2677 struct array_cache *shared_array = l3->shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 int max = shared_array->limit-shared_array->avail;
2679 if (max) {
2680 if (batchcount > max)
2681 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002682 memcpy(&(shared_array->entry[shared_array->avail]),
2683 ac->entry,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 sizeof(void*)*batchcount);
2685 shared_array->avail += batchcount;
2686 goto free_done;
2687 }
2688 }
2689
Christoph Lametere498be72005-09-09 13:03:32 -07002690 free_block(cachep, ac->entry, batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691free_done:
2692#if STATS
2693 {
2694 int i = 0;
2695 struct list_head *p;
2696
Christoph Lametere498be72005-09-09 13:03:32 -07002697 p = l3->slabs_free.next;
2698 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 struct slab *slabp;
2700
2701 slabp = list_entry(p, struct slab, list);
2702 BUG_ON(slabp->inuse);
2703
2704 i++;
2705 p = p->next;
2706 }
2707 STATS_SET_FREEABLE(cachep, i);
2708 }
2709#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002710 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002712 memmove(ac->entry, &(ac->entry[batchcount]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 sizeof(void*)*ac->avail);
2714}
2715
Christoph Lametere498be72005-09-09 13:03:32 -07002716
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717/*
2718 * __cache_free
2719 * Release an obj back to its cache. If the obj has a constructed
2720 * state, it must be in this state _before_ it is released.
2721 *
2722 * Called with disabled ints.
2723 */
2724static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2725{
2726 struct array_cache *ac = ac_data(cachep);
2727
2728 check_irq_off();
2729 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2730
Christoph Lametere498be72005-09-09 13:03:32 -07002731 /* Make sure we are not freeing a object from another
2732 * node to the array cache on this cpu.
2733 */
2734#ifdef CONFIG_NUMA
2735 {
2736 struct slab *slabp;
2737 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2738 if (unlikely(slabp->nodeid != numa_node_id())) {
2739 struct array_cache *alien = NULL;
2740 int nodeid = slabp->nodeid;
2741 struct kmem_list3 *l3 = cachep->nodelists[numa_node_id()];
2742
2743 STATS_INC_NODEFREES(cachep);
2744 if (l3->alien && l3->alien[nodeid]) {
2745 alien = l3->alien[nodeid];
2746 spin_lock(&alien->lock);
2747 if (unlikely(alien->avail == alien->limit))
2748 __drain_alien_cache(cachep,
2749 alien, nodeid);
2750 alien->entry[alien->avail++] = objp;
2751 spin_unlock(&alien->lock);
2752 } else {
2753 spin_lock(&(cachep->nodelists[nodeid])->
2754 list_lock);
2755 free_block(cachep, &objp, 1);
2756 spin_unlock(&(cachep->nodelists[nodeid])->
2757 list_lock);
2758 }
2759 return;
2760 }
2761 }
2762#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 if (likely(ac->avail < ac->limit)) {
2764 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002765 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 return;
2767 } else {
2768 STATS_INC_FREEMISS(cachep);
2769 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002770 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 }
2772}
2773
2774/**
2775 * kmem_cache_alloc - Allocate an object
2776 * @cachep: The cache to allocate from.
2777 * @flags: See kmalloc().
2778 *
2779 * Allocate an object from this cache. The flags are only relevant
2780 * if the cache has no available objects.
2781 */
2782void *kmem_cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2783{
2784 return __cache_alloc(cachep, flags);
2785}
2786EXPORT_SYMBOL(kmem_cache_alloc);
2787
2788/**
2789 * kmem_ptr_validate - check if an untrusted pointer might
2790 * be a slab entry.
2791 * @cachep: the cache we're checking against
2792 * @ptr: pointer to validate
2793 *
2794 * This verifies that the untrusted pointer looks sane:
2795 * it is _not_ a guarantee that the pointer is actually
2796 * part of the slab cache in question, but it at least
2797 * validates that the pointer can be dereferenced and
2798 * looks half-way sane.
2799 *
2800 * Currently only used for dentry validation.
2801 */
2802int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2803{
2804 unsigned long addr = (unsigned long) ptr;
2805 unsigned long min_addr = PAGE_OFFSET;
2806 unsigned long align_mask = BYTES_PER_WORD-1;
2807 unsigned long size = cachep->objsize;
2808 struct page *page;
2809
2810 if (unlikely(addr < min_addr))
2811 goto out;
2812 if (unlikely(addr > (unsigned long)high_memory - size))
2813 goto out;
2814 if (unlikely(addr & align_mask))
2815 goto out;
2816 if (unlikely(!kern_addr_valid(addr)))
2817 goto out;
2818 if (unlikely(!kern_addr_valid(addr + size - 1)))
2819 goto out;
2820 page = virt_to_page(ptr);
2821 if (unlikely(!PageSlab(page)))
2822 goto out;
2823 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2824 goto out;
2825 return 1;
2826out:
2827 return 0;
2828}
2829
2830#ifdef CONFIG_NUMA
2831/**
2832 * kmem_cache_alloc_node - Allocate an object on the specified node
2833 * @cachep: The cache to allocate from.
2834 * @flags: See kmalloc().
2835 * @nodeid: node number of the target node.
2836 *
2837 * Identical to kmem_cache_alloc, except that this function is slow
2838 * and can sleep. And it will allocate memory on the given node, which
2839 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002840 * New and improved: it will now make sure that the object gets
2841 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 */
Victor Fuscob2d55072005-09-10 00:26:36 -07002843void *kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844{
Christoph Lametere498be72005-09-09 13:03:32 -07002845 unsigned long save_flags;
2846 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847
Christoph Lametere498be72005-09-09 13:03:32 -07002848 if (nodeid == numa_node_id() || nodeid == -1)
2849 return __cache_alloc(cachep, flags);
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002850
Christoph Lametere498be72005-09-09 13:03:32 -07002851 if (unlikely(!cachep->nodelists[nodeid])) {
2852 /* Fall back to __cache_alloc if we run into trouble */
2853 printk(KERN_WARNING "slab: not allocating in inactive node %d for cache %s\n", nodeid, cachep->name);
2854 return __cache_alloc(cachep,flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
Christoph Lametere498be72005-09-09 13:03:32 -07002857 cache_alloc_debugcheck_before(cachep, flags);
2858 local_irq_save(save_flags);
2859 ptr = __cache_alloc_node(cachep, flags, nodeid);
2860 local_irq_restore(save_flags);
2861 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862
Christoph Lametere498be72005-09-09 13:03:32 -07002863 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864}
2865EXPORT_SYMBOL(kmem_cache_alloc_node);
2866
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002867void *kmalloc_node(size_t size, unsigned int __nocast flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002868{
2869 kmem_cache_t *cachep;
2870
2871 cachep = kmem_find_general_cachep(size, flags);
2872 if (unlikely(cachep == NULL))
2873 return NULL;
2874 return kmem_cache_alloc_node(cachep, flags, node);
2875}
2876EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877#endif
2878
2879/**
2880 * kmalloc - allocate memory
2881 * @size: how many bytes of memory are required.
2882 * @flags: the type of memory to allocate.
2883 *
2884 * kmalloc is the normal method of allocating memory
2885 * in the kernel.
2886 *
2887 * The @flags argument may be one of:
2888 *
2889 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2890 *
2891 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2892 *
2893 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2894 *
2895 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2896 * must be suitable for DMA. This can mean different things on different
2897 * platforms. For example, on i386, it means that the memory must come
2898 * from the first 16MB.
2899 */
2900void *__kmalloc(size_t size, unsigned int __nocast flags)
2901{
2902 kmem_cache_t *cachep;
2903
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002904 /* If you want to save a few bytes .text space: replace
2905 * __ with kmem_.
2906 * Then kmalloc uses the uninlined functions instead of the inline
2907 * functions.
2908 */
2909 cachep = __find_general_cachep(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 if (unlikely(cachep == NULL))
2911 return NULL;
2912 return __cache_alloc(cachep, flags);
2913}
2914EXPORT_SYMBOL(__kmalloc);
2915
2916#ifdef CONFIG_SMP
2917/**
2918 * __alloc_percpu - allocate one copy of the object for every present
2919 * cpu in the system, zeroing them.
2920 * Objects should be dereferenced using the per_cpu_ptr macro only.
2921 *
2922 * @size: how many bytes of memory are required.
2923 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2924 */
2925void *__alloc_percpu(size_t size, size_t align)
2926{
2927 int i;
2928 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2929
2930 if (!pdata)
2931 return NULL;
2932
Christoph Lametere498be72005-09-09 13:03:32 -07002933 /*
2934 * Cannot use for_each_online_cpu since a cpu may come online
2935 * and we have no way of figuring out how to fix the array
2936 * that we have allocated then....
2937 */
2938 for_each_cpu(i) {
2939 int node = cpu_to_node(i);
2940
2941 if (node_online(node))
2942 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
2943 else
2944 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
2946 if (!pdata->ptrs[i])
2947 goto unwind_oom;
2948 memset(pdata->ptrs[i], 0, size);
2949 }
2950
2951 /* Catch derefs w/o wrappers */
2952 return (void *) (~(unsigned long) pdata);
2953
2954unwind_oom:
2955 while (--i >= 0) {
2956 if (!cpu_possible(i))
2957 continue;
2958 kfree(pdata->ptrs[i]);
2959 }
2960 kfree(pdata);
2961 return NULL;
2962}
2963EXPORT_SYMBOL(__alloc_percpu);
2964#endif
2965
2966/**
2967 * kmem_cache_free - Deallocate an object
2968 * @cachep: The cache the allocation was from.
2969 * @objp: The previously allocated object.
2970 *
2971 * Free an object which was previously allocated from this
2972 * cache.
2973 */
2974void kmem_cache_free(kmem_cache_t *cachep, void *objp)
2975{
2976 unsigned long flags;
2977
2978 local_irq_save(flags);
2979 __cache_free(cachep, objp);
2980 local_irq_restore(flags);
2981}
2982EXPORT_SYMBOL(kmem_cache_free);
2983
2984/**
Pekka J Enbergdd392712005-09-06 15:18:31 -07002985 * kzalloc - allocate memory. The memory is set to zero.
2986 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 * @flags: the type of memory to allocate.
2988 */
Pekka J Enbergdd392712005-09-06 15:18:31 -07002989void *kzalloc(size_t size, unsigned int __nocast flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990{
Pekka J Enbergdd392712005-09-06 15:18:31 -07002991 void *ret = kmalloc(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 if (ret)
Pekka J Enbergdd392712005-09-06 15:18:31 -07002993 memset(ret, 0, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 return ret;
2995}
Pekka J Enbergdd392712005-09-06 15:18:31 -07002996EXPORT_SYMBOL(kzalloc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997
2998/**
2999 * kfree - free previously allocated memory
3000 * @objp: pointer returned by kmalloc.
3001 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003002 * If @objp is NULL, no operation is performed.
3003 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004 * Don't free memory not originally allocated by kmalloc()
3005 * or you will run into trouble.
3006 */
3007void kfree(const void *objp)
3008{
3009 kmem_cache_t *c;
3010 unsigned long flags;
3011
3012 if (unlikely(!objp))
3013 return;
3014 local_irq_save(flags);
3015 kfree_debugcheck(objp);
3016 c = GET_PAGE_CACHE(virt_to_page(objp));
3017 __cache_free(c, (void*)objp);
3018 local_irq_restore(flags);
3019}
3020EXPORT_SYMBOL(kfree);
3021
3022#ifdef CONFIG_SMP
3023/**
3024 * free_percpu - free previously allocated percpu memory
3025 * @objp: pointer returned by alloc_percpu.
3026 *
3027 * Don't free memory not originally allocated by alloc_percpu()
3028 * The complemented objp is to check for that.
3029 */
3030void
3031free_percpu(const void *objp)
3032{
3033 int i;
3034 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
3035
Christoph Lametere498be72005-09-09 13:03:32 -07003036 /*
3037 * We allocate for all cpus so we cannot use for online cpu here.
3038 */
3039 for_each_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 kfree(p);
3042}
3043EXPORT_SYMBOL(free_percpu);
3044#endif
3045
3046unsigned int kmem_cache_size(kmem_cache_t *cachep)
3047{
3048 return obj_reallen(cachep);
3049}
3050EXPORT_SYMBOL(kmem_cache_size);
3051
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003052const char *kmem_cache_name(kmem_cache_t *cachep)
3053{
3054 return cachep->name;
3055}
3056EXPORT_SYMBOL_GPL(kmem_cache_name);
3057
Christoph Lametere498be72005-09-09 13:03:32 -07003058/*
3059 * This initializes kmem_list3 for all nodes.
3060 */
3061static int alloc_kmemlist(kmem_cache_t *cachep)
3062{
3063 int node;
3064 struct kmem_list3 *l3;
3065 int err = 0;
3066
3067 for_each_online_node(node) {
3068 struct array_cache *nc = NULL, *new;
3069 struct array_cache **new_alien = NULL;
3070#ifdef CONFIG_NUMA
3071 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3072 goto fail;
3073#endif
3074 if (!(new = alloc_arraycache(node, (cachep->shared*
3075 cachep->batchcount), 0xbaadf00d)))
3076 goto fail;
3077 if ((l3 = cachep->nodelists[node])) {
3078
3079 spin_lock_irq(&l3->list_lock);
3080
3081 if ((nc = cachep->nodelists[node]->shared))
3082 free_block(cachep, nc->entry,
3083 nc->avail);
3084
3085 l3->shared = new;
3086 if (!cachep->nodelists[node]->alien) {
3087 l3->alien = new_alien;
3088 new_alien = NULL;
3089 }
3090 l3->free_limit = (1 + nr_cpus_node(node))*
3091 cachep->batchcount + cachep->num;
3092 spin_unlock_irq(&l3->list_lock);
3093 kfree(nc);
3094 free_alien_cache(new_alien);
3095 continue;
3096 }
3097 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3098 GFP_KERNEL, node)))
3099 goto fail;
3100
3101 kmem_list3_init(l3);
3102 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3103 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
3104 l3->shared = new;
3105 l3->alien = new_alien;
3106 l3->free_limit = (1 + nr_cpus_node(node))*
3107 cachep->batchcount + cachep->num;
3108 cachep->nodelists[node] = l3;
3109 }
3110 return err;
3111fail:
3112 err = -ENOMEM;
3113 return err;
3114}
3115
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116struct ccupdate_struct {
3117 kmem_cache_t *cachep;
3118 struct array_cache *new[NR_CPUS];
3119};
3120
3121static void do_ccupdate_local(void *info)
3122{
3123 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3124 struct array_cache *old;
3125
3126 check_irq_off();
3127 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003128
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3130 new->new[smp_processor_id()] = old;
3131}
3132
3133
3134static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3135 int shared)
3136{
3137 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003138 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139
3140 memset(&new.new,0,sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003141 for_each_online_cpu(i) {
3142 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, batchcount);
3143 if (!new.new[i]) {
3144 for (i--; i >= 0; i--) kfree(new.new[i]);
3145 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 }
3147 }
3148 new.cachep = cachep;
3149
3150 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003151
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 check_irq_on();
3153 spin_lock_irq(&cachep->spinlock);
3154 cachep->batchcount = batchcount;
3155 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003156 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 spin_unlock_irq(&cachep->spinlock);
3158
Christoph Lametere498be72005-09-09 13:03:32 -07003159 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 struct array_cache *ccold = new.new[i];
3161 if (!ccold)
3162 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003163 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3164 free_block(cachep, ccold->entry, ccold->avail);
3165 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 kfree(ccold);
3167 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168
Christoph Lametere498be72005-09-09 13:03:32 -07003169 err = alloc_kmemlist(cachep);
3170 if (err) {
3171 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3172 cachep->name, -err);
3173 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 return 0;
3176}
3177
3178
3179static void enable_cpucache(kmem_cache_t *cachep)
3180{
3181 int err;
3182 int limit, shared;
3183
3184 /* The head array serves three purposes:
3185 * - create a LIFO ordering, i.e. return objects that are cache-warm
3186 * - reduce the number of spinlock operations.
3187 * - reduce the number of linked list operations on the slab and
3188 * bufctl chains: array operations are cheaper.
3189 * The numbers are guessed, we should auto-tune as described by
3190 * Bonwick.
3191 */
3192 if (cachep->objsize > 131072)
3193 limit = 1;
3194 else if (cachep->objsize > PAGE_SIZE)
3195 limit = 8;
3196 else if (cachep->objsize > 1024)
3197 limit = 24;
3198 else if (cachep->objsize > 256)
3199 limit = 54;
3200 else
3201 limit = 120;
3202
3203 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3204 * allocation behaviour: Most allocs on one cpu, most free operations
3205 * on another cpu. For these cases, an efficient object passing between
3206 * cpus is necessary. This is provided by a shared array. The array
3207 * replaces Bonwick's magazine layer.
3208 * On uniprocessor, it's functionally equivalent (but less efficient)
3209 * to a larger limit. Thus disabled by default.
3210 */
3211 shared = 0;
3212#ifdef CONFIG_SMP
3213 if (cachep->objsize <= PAGE_SIZE)
3214 shared = 8;
3215#endif
3216
3217#if DEBUG
3218 /* With debugging enabled, large batchcount lead to excessively
3219 * long periods with disabled local interrupts. Limit the
3220 * batchcount
3221 */
3222 if (limit > 32)
3223 limit = 32;
3224#endif
3225 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
3226 if (err)
3227 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3228 cachep->name, -err);
3229}
3230
3231static void drain_array_locked(kmem_cache_t *cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07003232 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233{
3234 int tofree;
3235
Christoph Lametere498be72005-09-09 13:03:32 -07003236 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 if (ac->touched && !force) {
3238 ac->touched = 0;
3239 } else if (ac->avail) {
3240 tofree = force ? ac->avail : (ac->limit+4)/5;
3241 if (tofree > ac->avail) {
3242 tofree = (ac->avail+1)/2;
3243 }
Christoph Lametere498be72005-09-09 13:03:32 -07003244 free_block(cachep, ac->entry, tofree);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003246 memmove(ac->entry, &(ac->entry[tofree]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247 sizeof(void*)*ac->avail);
3248 }
3249}
3250
3251/**
3252 * cache_reap - Reclaim memory from caches.
3253 *
3254 * Called from workqueue/eventd every few seconds.
3255 * Purpose:
3256 * - clear the per-cpu caches for this CPU.
3257 * - return freeable pages to the main free memory pool.
3258 *
3259 * If we cannot acquire the cache chain semaphore then just give up - we'll
3260 * try again on the next iteration.
3261 */
3262static void cache_reap(void *unused)
3263{
3264 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003265 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266
3267 if (down_trylock(&cache_chain_sem)) {
3268 /* Give up. Setup the next iteration. */
3269 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
3270 return;
3271 }
3272
3273 list_for_each(walk, &cache_chain) {
3274 kmem_cache_t *searchp;
3275 struct list_head* p;
3276 int tofree;
3277 struct slab *slabp;
3278
3279 searchp = list_entry(walk, kmem_cache_t, next);
3280
3281 if (searchp->flags & SLAB_NO_REAP)
3282 goto next;
3283
3284 check_irq_on();
3285
Christoph Lametere498be72005-09-09 13:03:32 -07003286 l3 = searchp->nodelists[numa_node_id()];
3287 if (l3->alien)
3288 drain_alien_cache(searchp, l3);
3289 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Christoph Lametere498be72005-09-09 13:03:32 -07003291 drain_array_locked(searchp, ac_data(searchp), 0,
3292 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293
Christoph Lametere498be72005-09-09 13:03:32 -07003294 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 goto next_unlock;
3296
Christoph Lametere498be72005-09-09 13:03:32 -07003297 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298
Christoph Lametere498be72005-09-09 13:03:32 -07003299 if (l3->shared)
3300 drain_array_locked(searchp, l3->shared, 0,
3301 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
Christoph Lametere498be72005-09-09 13:03:32 -07003303 if (l3->free_touched) {
3304 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 goto next_unlock;
3306 }
3307
Christoph Lametere498be72005-09-09 13:03:32 -07003308 tofree = (l3->free_limit+5*searchp->num-1)/(5*searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003309 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003310 p = l3->slabs_free.next;
3311 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 break;
3313
3314 slabp = list_entry(p, struct slab, list);
3315 BUG_ON(slabp->inuse);
3316 list_del(&slabp->list);
3317 STATS_INC_REAPED(searchp);
3318
3319 /* Safe to drop the lock. The slab is no longer
3320 * linked to the cache.
3321 * searchp cannot disappear, we hold
3322 * cache_chain_lock
3323 */
Christoph Lametere498be72005-09-09 13:03:32 -07003324 l3->free_objects -= searchp->num;
3325 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003327 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 } while(--tofree > 0);
3329next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003330 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331next:
3332 cond_resched();
3333 }
3334 check_irq_on();
3335 up(&cache_chain_sem);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003336 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 /* Setup the next iteration */
3338 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
3339}
3340
3341#ifdef CONFIG_PROC_FS
3342
3343static void *s_start(struct seq_file *m, loff_t *pos)
3344{
3345 loff_t n = *pos;
3346 struct list_head *p;
3347
3348 down(&cache_chain_sem);
3349 if (!n) {
3350 /*
3351 * Output format version, so at least we can change it
3352 * without _too_ many complaints.
3353 */
3354#if STATS
3355 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3356#else
3357 seq_puts(m, "slabinfo - version: 2.1\n");
3358#endif
3359 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
3360 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3361 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3362#if STATS
3363 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
Christoph Lametere498be72005-09-09 13:03:32 -07003364 " <error> <maxfreeable> <nodeallocs> <remotefrees>");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3366#endif
3367 seq_putc(m, '\n');
3368 }
3369 p = cache_chain.next;
3370 while (n--) {
3371 p = p->next;
3372 if (p == &cache_chain)
3373 return NULL;
3374 }
3375 return list_entry(p, kmem_cache_t, next);
3376}
3377
3378static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3379{
3380 kmem_cache_t *cachep = p;
3381 ++*pos;
3382 return cachep->next.next == &cache_chain ? NULL
3383 : list_entry(cachep->next.next, kmem_cache_t, next);
3384}
3385
3386static void s_stop(struct seq_file *m, void *p)
3387{
3388 up(&cache_chain_sem);
3389}
3390
3391static int s_show(struct seq_file *m, void *p)
3392{
3393 kmem_cache_t *cachep = p;
3394 struct list_head *q;
3395 struct slab *slabp;
3396 unsigned long active_objs;
3397 unsigned long num_objs;
3398 unsigned long active_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003399 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3400 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003402 int node;
3403 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404
3405 check_irq_on();
3406 spin_lock_irq(&cachep->spinlock);
3407 active_objs = 0;
3408 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003409 for_each_online_node(node) {
3410 l3 = cachep->nodelists[node];
3411 if (!l3)
3412 continue;
3413
3414 spin_lock(&l3->list_lock);
3415
3416 list_for_each(q,&l3->slabs_full) {
3417 slabp = list_entry(q, struct slab, list);
3418 if (slabp->inuse != cachep->num && !error)
3419 error = "slabs_full accounting error";
3420 active_objs += cachep->num;
3421 active_slabs++;
3422 }
3423 list_for_each(q,&l3->slabs_partial) {
3424 slabp = list_entry(q, struct slab, list);
3425 if (slabp->inuse == cachep->num && !error)
3426 error = "slabs_partial inuse accounting error";
3427 if (!slabp->inuse && !error)
3428 error = "slabs_partial/inuse accounting error";
3429 active_objs += slabp->inuse;
3430 active_slabs++;
3431 }
3432 list_for_each(q,&l3->slabs_free) {
3433 slabp = list_entry(q, struct slab, list);
3434 if (slabp->inuse && !error)
3435 error = "slabs_free/inuse accounting error";
3436 num_slabs++;
3437 }
3438 free_objects += l3->free_objects;
3439 shared_avail += l3->shared->avail;
3440
3441 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 }
3443 num_slabs+=active_slabs;
3444 num_objs = num_slabs*cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003445 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 error = "free_objects accounting error";
3447
3448 name = cachep->name;
3449 if (error)
3450 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3451
3452 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3453 name, active_objs, num_objs, cachep->objsize,
3454 cachep->num, (1<<cachep->gfporder));
3455 seq_printf(m, " : tunables %4u %4u %4u",
3456 cachep->limit, cachep->batchcount,
Christoph Lametere498be72005-09-09 13:03:32 -07003457 cachep->shared);
3458 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3459 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460#if STATS
3461 { /* list3 stats */
3462 unsigned long high = cachep->high_mark;
3463 unsigned long allocs = cachep->num_allocations;
3464 unsigned long grown = cachep->grown;
3465 unsigned long reaped = cachep->reaped;
3466 unsigned long errors = cachep->errors;
3467 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003469 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
Christoph Lametere498be72005-09-09 13:03:32 -07003471 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3472 %4lu %4lu %4lu %4lu",
3473 allocs, high, grown, reaped, errors,
3474 max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 }
3476 /* cpu stats */
3477 {
3478 unsigned long allochit = atomic_read(&cachep->allochit);
3479 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3480 unsigned long freehit = atomic_read(&cachep->freehit);
3481 unsigned long freemiss = atomic_read(&cachep->freemiss);
3482
3483 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3484 allochit, allocmiss, freehit, freemiss);
3485 }
3486#endif
3487 seq_putc(m, '\n');
3488 spin_unlock_irq(&cachep->spinlock);
3489 return 0;
3490}
3491
3492/*
3493 * slabinfo_op - iterator that generates /proc/slabinfo
3494 *
3495 * Output layout:
3496 * cache-name
3497 * num-active-objs
3498 * total-objs
3499 * object size
3500 * num-active-slabs
3501 * total-slabs
3502 * num-pages-per-slab
3503 * + further values on SMP and with statistics enabled
3504 */
3505
3506struct seq_operations slabinfo_op = {
3507 .start = s_start,
3508 .next = s_next,
3509 .stop = s_stop,
3510 .show = s_show,
3511};
3512
3513#define MAX_SLABINFO_WRITE 128
3514/**
3515 * slabinfo_write - Tuning for the slab allocator
3516 * @file: unused
3517 * @buffer: user buffer
3518 * @count: data length
3519 * @ppos: unused
3520 */
3521ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3522 size_t count, loff_t *ppos)
3523{
3524 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3525 int limit, batchcount, shared, res;
3526 struct list_head *p;
3527
3528 if (count > MAX_SLABINFO_WRITE)
3529 return -EINVAL;
3530 if (copy_from_user(&kbuf, buffer, count))
3531 return -EFAULT;
3532 kbuf[MAX_SLABINFO_WRITE] = '\0';
3533
3534 tmp = strchr(kbuf, ' ');
3535 if (!tmp)
3536 return -EINVAL;
3537 *tmp = '\0';
3538 tmp++;
3539 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3540 return -EINVAL;
3541
3542 /* Find the cache in the chain of caches. */
3543 down(&cache_chain_sem);
3544 res = -EINVAL;
3545 list_for_each(p,&cache_chain) {
3546 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3547
3548 if (!strcmp(cachep->name, kbuf)) {
3549 if (limit < 1 ||
3550 batchcount < 1 ||
3551 batchcount > limit ||
3552 shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003553 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003555 res = do_tune_cpucache(cachep, limit,
3556 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 }
3558 break;
3559 }
3560 }
3561 up(&cache_chain_sem);
3562 if (res >= 0)
3563 res = count;
3564 return res;
3565}
3566#endif
3567
Manfred Spraul00e145b2005-09-03 15:55:07 -07003568/**
3569 * ksize - get the actual amount of memory allocated for a given object
3570 * @objp: Pointer to the object
3571 *
3572 * kmalloc may internally round up allocations and return more memory
3573 * than requested. ksize() can be used to determine the actual amount of
3574 * memory allocated. The caller may use this additional memory, even though
3575 * a smaller amount of memory was initially specified with the kmalloc call.
3576 * The caller must guarantee that objp points to a valid object previously
3577 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3578 * must not be freed during the duration of the call.
3579 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580unsigned int ksize(const void *objp)
3581{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003582 if (unlikely(objp == NULL))
3583 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584
Manfred Spraul00e145b2005-09-03 15:55:07 -07003585 return obj_reallen(GET_PAGE_CACHE(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586}
Paulo Marques543537b2005-06-23 00:09:02 -07003587
3588
3589/*
3590 * kstrdup - allocate space for and copy an existing string
3591 *
3592 * @s: the string to duplicate
3593 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3594 */
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07003595char *kstrdup(const char *s, unsigned int __nocast gfp)
Paulo Marques543537b2005-06-23 00:09:02 -07003596{
3597 size_t len;
3598 char *buf;
3599
3600 if (!s)
3601 return NULL;
3602
3603 len = strlen(s) + 1;
3604 buf = kmalloc(len, gfp);
3605 if (buf)
3606 memcpy(buf, s, len);
3607 return buf;
3608}
3609EXPORT_SYMBOL(kstrdup);