blob: 6f8495e2185b36c4d4e08402a111a59a3deeb01b [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
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * 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>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800106#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800107#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108
109#include <asm/uaccess.h>
110#include <asm/cacheflush.h>
111#include <asm/tlbflush.h>
112#include <asm/page.h>
113
114/*
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
118 *
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
123 */
124
125#ifdef CONFIG_DEBUG_SLAB
126#define DEBUG 1
127#define STATS 1
128#define FORCED_DEBUG 1
129#else
130#define DEBUG 0
131#define STATS 0
132#define FORCED_DEBUG 0
133#endif
134
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135/* Shouldn't this be in a header file somewhere? */
136#define BYTES_PER_WORD sizeof(void *)
137
138#ifndef cache_line_size
139#define cache_line_size() L1_CACHE_BYTES
140#endif
141
142#ifndef ARCH_KMALLOC_MINALIGN
143/*
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
150 */
151#define ARCH_KMALLOC_MINALIGN 0
152#endif
153
154#ifndef ARCH_SLAB_MINALIGN
155/*
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
161 */
162#define ARCH_SLAB_MINALIGN 0
163#endif
164
165#ifndef ARCH_KMALLOC_FLAGS
166#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
167#endif
168
169/* Legal flag mask for kmem_cache_create(). */
170#if DEBUG
171# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
176 SLAB_DESTROY_BY_RCU)
177#else
178# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU)
182#endif
183
184/*
185 * kmem_bufctl_t:
186 *
187 * Bufctl's are used for linking objs within a slab
188 * linked offsets.
189 *
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 */
202
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700203typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
207
208/* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
210 */
211static unsigned long offslab_limit;
212
213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
247 kmem_cache_t *cachep;
248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
269 void *entry[0]; /*
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
273 * [0] is for gcc 2.95. It should really be [].
274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
277/* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
279 */
280#define BOOT_CPUCACHE_ENTRIES 1
281struct arraycache_init {
282 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800283 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
286/*
Christoph Lametere498be72005-09-09 13:03:32 -0700287 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 */
289struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned long next_reap;
295 int free_touched;
296 unsigned int free_limit;
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300};
301
Christoph Lametere498be72005-09-09 13:03:32 -0700302/*
303 * Need this for bootstrapping a per node allocator.
304 */
305#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307#define CACHE_CACHE 0
308#define SIZE_AC 1
309#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Christoph Lametere498be72005-09-09 13:03:32 -0700311/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700312 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700313 * a constant is passed to it. Mostly the same as
314 * what is in linux/slab.h except it returns an
315 * index.
316 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700317static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700318{
319 if (__builtin_constant_p(size)) {
320 int i = 0;
321
322#define CACHE(x) \
323 if (size <=x) \
324 return i; \
325 else \
326 i++;
327#include "linux/kmalloc_sizes.h"
328#undef CACHE
329 {
330 extern void __bad_size(void);
331 __bad_size();
332 }
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700333 } else
334 BUG();
Christoph Lametere498be72005-09-09 13:03:32 -0700335 return 0;
336}
337
338#define INDEX_AC index_of(sizeof(struct arraycache_init))
339#define INDEX_L3 index_of(sizeof(struct kmem_list3))
340
341static inline void kmem_list3_init(struct kmem_list3 *parent)
342{
343 INIT_LIST_HEAD(&parent->slabs_full);
344 INIT_LIST_HEAD(&parent->slabs_partial);
345 INIT_LIST_HEAD(&parent->slabs_free);
346 parent->shared = NULL;
347 parent->alien = NULL;
348 spin_lock_init(&parent->list_lock);
349 parent->free_objects = 0;
350 parent->free_touched = 0;
351}
352
353#define MAKE_LIST(cachep, listp, slab, nodeid) \
354 do { \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
357 } while (0)
358
359#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
360 do { \
361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
364 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
366/*
367 * kmem_cache_t
368 *
369 * manages a cache.
370 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800371
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800372struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800374 struct array_cache *array[NR_CPUS];
375 unsigned int batchcount;
376 unsigned int limit;
377 unsigned int shared;
378 unsigned int objsize;
Christoph Lametere498be72005-09-09 13:03:32 -0700379/* 2) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800380 struct kmem_list3 *nodelists[MAX_NUMNODES];
381 unsigned int flags; /* constant flags */
382 unsigned int num; /* # of objs per slab */
383 spinlock_t spinlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
385/* 3) cache_grow/shrink */
386 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800387 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
389 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800390 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800392 size_t colour; /* cache colouring range */
393 unsigned int colour_off; /* colour offset */
394 unsigned int colour_next; /* cache colouring */
395 kmem_cache_t *slabp_cache;
396 unsigned int slab_size;
397 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /* constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800400 void (*ctor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /* de-constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800403 void (*dtor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405/* 4) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 const char *name;
407 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409/* 5) statistics */
410#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800411 unsigned long num_active;
412 unsigned long num_allocations;
413 unsigned long high_mark;
414 unsigned long grown;
415 unsigned long reaped;
416 unsigned long errors;
417 unsigned long max_freeable;
418 unsigned long node_allocs;
419 unsigned long node_frees;
420 atomic_t allochit;
421 atomic_t allocmiss;
422 atomic_t freehit;
423 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424#endif
425#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800426 int dbghead;
427 int reallen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428#endif
429};
430
431#define CFLGS_OFF_SLAB (0x80000000UL)
432#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
433
434#define BATCHREFILL_LIMIT 16
435/* Optimization question: fewer reaps means less
436 * probability for unnessary cpucache drain/refill cycles.
437 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100438 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 * which could lock up otherwise freeable slabs.
440 */
441#define REAPTIMEOUT_CPUC (2*HZ)
442#define REAPTIMEOUT_LIST3 (4*HZ)
443
444#if STATS
445#define STATS_INC_ACTIVE(x) ((x)->num_active++)
446#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
447#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
448#define STATS_INC_GROWN(x) ((x)->grown++)
449#define STATS_INC_REAPED(x) ((x)->reaped++)
450#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
451 (x)->high_mark = (x)->num_active; \
452 } while (0)
453#define STATS_INC_ERR(x) ((x)->errors++)
454#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700455#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456#define STATS_SET_FREEABLE(x, i) \
457 do { if ((x)->max_freeable < i) \
458 (x)->max_freeable = i; \
459 } while (0)
460
461#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
462#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
463#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
464#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
465#else
466#define STATS_INC_ACTIVE(x) do { } while (0)
467#define STATS_DEC_ACTIVE(x) do { } while (0)
468#define STATS_INC_ALLOCED(x) do { } while (0)
469#define STATS_INC_GROWN(x) do { } while (0)
470#define STATS_INC_REAPED(x) do { } while (0)
471#define STATS_SET_HIGH(x) do { } while (0)
472#define STATS_INC_ERR(x) do { } while (0)
473#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700474#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475#define STATS_SET_FREEABLE(x, i) \
476 do { } while (0)
477
478#define STATS_INC_ALLOCHIT(x) do { } while (0)
479#define STATS_INC_ALLOCMISS(x) do { } while (0)
480#define STATS_INC_FREEHIT(x) do { } while (0)
481#define STATS_INC_FREEMISS(x) do { } while (0)
482#endif
483
484#if DEBUG
485/* Magic nums for obj red zoning.
486 * Placed in the first word before and the first word after an obj.
487 */
488#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
489#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
490
491/* ...and for poisoning */
492#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
493#define POISON_FREE 0x6b /* for use-after-free poisoning */
494#define POISON_END 0xa5 /* end-byte of poisoning */
495
496/* memory layout of objects:
497 * 0 : objp
498 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
499 * the end of an object is aligned with the end of the real
500 * allocation. Catches writes behind the end of the allocation.
501 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
502 * redzone word.
503 * cachep->dbghead: The real object.
504 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
505 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
506 */
507static int obj_dbghead(kmem_cache_t *cachep)
508{
509 return cachep->dbghead;
510}
511
512static int obj_reallen(kmem_cache_t *cachep)
513{
514 return cachep->reallen;
515}
516
517static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
518{
519 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
520 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
521}
522
523static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
524{
525 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
526 if (cachep->flags & SLAB_STORE_USER)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800527 return (unsigned long *)(objp + cachep->objsize -
528 2 * BYTES_PER_WORD);
529 return (unsigned long *)(objp + cachep->objsize - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
532static void **dbg_userword(kmem_cache_t *cachep, void *objp)
533{
534 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800535 return (void **)(objp + cachep->objsize - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
538#else
539
540#define obj_dbghead(x) 0
541#define obj_reallen(cachep) (cachep->objsize)
542#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
543#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
544#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
545
546#endif
547
548/*
549 * Maximum size of an obj (in 2^order pages)
550 * and absolute limit for the gfp order.
551 */
552#if defined(CONFIG_LARGE_ALLOCS)
553#define MAX_OBJ_ORDER 13 /* up to 32Mb */
554#define MAX_GFP_ORDER 13 /* up to 32Mb */
555#elif defined(CONFIG_MMU)
556#define MAX_OBJ_ORDER 5 /* 32 pages */
557#define MAX_GFP_ORDER 5 /* 32 pages */
558#else
559#define MAX_OBJ_ORDER 8 /* up to 1Mb */
560#define MAX_GFP_ORDER 8 /* up to 1Mb */
561#endif
562
563/*
564 * Do not go above this order unless 0 objects fit into the slab.
565 */
566#define BREAK_GFP_ORDER_HI 1
567#define BREAK_GFP_ORDER_LO 0
568static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
569
Pekka Enberg065d41c2005-11-13 16:06:46 -0800570/* Functions for storing/retrieving the cachep and or slab from the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 * global 'mem_map'. These are used to find the slab an obj belongs to.
572 * With kfree(), these are used to find the cache which an obj belongs to.
573 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800574static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
575{
576 page->lru.next = (struct list_head *)cache;
577}
578
579static inline struct kmem_cache *page_get_cache(struct page *page)
580{
581 return (struct kmem_cache *)page->lru.next;
582}
583
584static inline void page_set_slab(struct page *page, struct slab *slab)
585{
586 page->lru.prev = (struct list_head *)slab;
587}
588
589static inline struct slab *page_get_slab(struct page *page)
590{
591 return (struct slab *)page->lru.prev;
592}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594/* These are the default caches for kmalloc. Custom caches can have other sizes. */
595struct cache_sizes malloc_sizes[] = {
596#define CACHE(x) { .cs_size = (x) },
597#include <linux/kmalloc_sizes.h>
598 CACHE(ULONG_MAX)
599#undef CACHE
600};
601EXPORT_SYMBOL(malloc_sizes);
602
603/* Must match cache_sizes above. Out of line to keep cache footprint low. */
604struct cache_names {
605 char *name;
606 char *name_dma;
607};
608
609static struct cache_names __initdata cache_names[] = {
610#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
611#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800612 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613#undef CACHE
614};
615
616static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800617 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800619 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621/* internal cache of cache description objs */
622static kmem_cache_t cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800623 .batchcount = 1,
624 .limit = BOOT_CPUCACHE_ENTRIES,
625 .shared = 1,
626 .objsize = sizeof(kmem_cache_t),
627 .flags = SLAB_NO_REAP,
628 .spinlock = SPIN_LOCK_UNLOCKED,
629 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800631 .reallen = sizeof(kmem_cache_t),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632#endif
633};
634
635/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800636static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637static struct list_head cache_chain;
638
639/*
640 * vm_enough_memory() looks at this to determine how many
641 * slab-allocated pages are possibly freeable under pressure
642 *
643 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
644 */
645atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
647/*
648 * chicken and egg problem: delay the per-cpu array allocation
649 * until the general caches are up.
650 */
651static enum {
652 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700653 PARTIAL_AC,
654 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 FULL
656} g_cpucache_up;
657
658static DEFINE_PER_CPU(struct work_struct, reap_work);
659
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800660static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
661static void enable_cpucache(kmem_cache_t *cachep);
662static void cache_reap(void *unused);
Christoph Lametere498be72005-09-09 13:03:32 -0700663static int __node_shrink(kmem_cache_t *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665static inline struct array_cache *ac_data(kmem_cache_t *cachep)
666{
667 return cachep->array[smp_processor_id()];
668}
669
Al Virodd0fc662005-10-07 07:46:04 +0100670static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671{
672 struct cache_sizes *csizep = malloc_sizes;
673
674#if DEBUG
675 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800676 * kmem_cache_create(), or __kmalloc(), before
677 * the generic caches are initialized.
678 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700679 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680#endif
681 while (size > csizep->cs_size)
682 csizep++;
683
684 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700685 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 * has cs_{dma,}cachep==NULL. Thus no special case
687 * for large kmalloc calls required.
688 */
689 if (unlikely(gfpflags & GFP_DMA))
690 return csizep->cs_dmacachep;
691 return csizep->cs_cachep;
692}
693
Al Virodd0fc662005-10-07 07:46:04 +0100694kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700695{
696 return __find_general_cachep(size, gfpflags);
697}
698EXPORT_SYMBOL(kmem_find_general_cachep);
699
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700/* Cal the num objs, wastage, and bytes left over for a given slab size. */
701static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800702 int flags, size_t *left_over, unsigned int *num)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703{
704 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800705 size_t wastage = PAGE_SIZE << gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 size_t extra = 0;
707 size_t base = 0;
708
709 if (!(flags & CFLGS_OFF_SLAB)) {
710 base = sizeof(struct slab);
711 extra = sizeof(kmem_bufctl_t);
712 }
713 i = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800714 while (i * size + ALIGN(base + i * extra, align) <= wastage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 i++;
716 if (i > 0)
717 i--;
718
719 if (i > SLAB_LIMIT)
720 i = SLAB_LIMIT;
721
722 *num = i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800723 wastage -= i * size;
724 wastage -= ALIGN(base + i * extra, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 *left_over = wastage;
726}
727
728#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
729
730static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
731{
732 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800733 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 dump_stack();
735}
736
737/*
738 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
739 * via the workqueue/eventd.
740 * Add the CPU number into the expiration time to minimize the possibility of
741 * the CPUs getting into lockstep and contending for the global cache chain
742 * lock.
743 */
744static void __devinit start_cpu_timer(int cpu)
745{
746 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
747
748 /*
749 * When this gets called from do_initcalls via cpucache_init(),
750 * init_workqueues() has already run, so keventd will be setup
751 * at that time.
752 */
753 if (keventd_up() && reap_work->func == NULL) {
754 INIT_WORK(reap_work, cache_reap, NULL);
755 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
756 }
757}
758
Christoph Lametere498be72005-09-09 13:03:32 -0700759static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800760 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800762 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 struct array_cache *nc = NULL;
764
Christoph Lametere498be72005-09-09 13:03:32 -0700765 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766 if (nc) {
767 nc->avail = 0;
768 nc->limit = entries;
769 nc->batchcount = batchcount;
770 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700771 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 }
773 return nc;
774}
775
Christoph Lametere498be72005-09-09 13:03:32 -0700776#ifdef CONFIG_NUMA
Christoph Lameterdc85da12006-01-18 17:42:36 -0800777static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int);
778
Christoph Lametere498be72005-09-09 13:03:32 -0700779static inline struct array_cache **alloc_alien_cache(int node, int limit)
780{
781 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800782 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700783 int i;
784
785 if (limit > 1)
786 limit = 12;
787 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
788 if (ac_ptr) {
789 for_each_node(i) {
790 if (i == node || !node_online(i)) {
791 ac_ptr[i] = NULL;
792 continue;
793 }
794 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
795 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800796 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700797 kfree(ac_ptr[i]);
798 kfree(ac_ptr);
799 return NULL;
800 }
801 }
802 }
803 return ac_ptr;
804}
805
806static inline void free_alien_cache(struct array_cache **ac_ptr)
807{
808 int i;
809
810 if (!ac_ptr)
811 return;
812
813 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800814 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700815
816 kfree(ac_ptr);
817}
818
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800819static inline void __drain_alien_cache(kmem_cache_t *cachep,
820 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700821{
822 struct kmem_list3 *rl3 = cachep->nodelists[node];
823
824 if (ac->avail) {
825 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700826 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700827 ac->avail = 0;
828 spin_unlock(&rl3->list_lock);
829 }
830}
831
832static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
833{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800834 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700835 struct array_cache *ac;
836 unsigned long flags;
837
838 for_each_online_node(i) {
839 ac = l3->alien[i];
840 if (ac) {
841 spin_lock_irqsave(&ac->lock, flags);
842 __drain_alien_cache(cachep, ac, i);
843 spin_unlock_irqrestore(&ac->lock, flags);
844 }
845 }
846}
847#else
848#define alloc_alien_cache(node, limit) do { } while (0)
849#define free_alien_cache(ac_ptr) do { } while (0)
850#define drain_alien_cache(cachep, l3) do { } while (0)
851#endif
852
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800854 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855{
856 long cpu = (long)hcpu;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800857 kmem_cache_t *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700858 struct kmem_list3 *l3 = NULL;
859 int node = cpu_to_node(cpu);
860 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 switch (action) {
863 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800864 mutex_lock(&cache_chain_mutex);
Christoph Lametere498be72005-09-09 13:03:32 -0700865 /* we need to do this right in the beginning since
866 * alloc_arraycache's are going to use this list.
867 * kmalloc_node allows us to add the slab to the right
868 * kmem_list3 and not this cpu's kmem_list3
869 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Christoph Lametere498be72005-09-09 13:03:32 -0700871 list_for_each_entry(cachep, &cache_chain, next) {
872 /* setup the size64 kmemlist for cpu before we can
873 * begin anything. Make sure some other cpu on this
874 * node has not already allocated this
875 */
876 if (!cachep->nodelists[node]) {
877 if (!(l3 = kmalloc_node(memsize,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800878 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -0700879 goto bad;
880 kmem_list3_init(l3);
881 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800882 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -0700883
884 cachep->nodelists[node] = l3;
885 }
886
887 spin_lock_irq(&cachep->nodelists[node]->list_lock);
888 cachep->nodelists[node]->free_limit =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800889 (1 + nr_cpus_node(node)) *
890 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -0700891 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
892 }
893
894 /* Now we can go ahead with allocating the shared array's
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800895 & array cache's */
Christoph Lametere498be72005-09-09 13:03:32 -0700896 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -0800897 struct array_cache *nc;
898
Christoph Lametere498be72005-09-09 13:03:32 -0700899 nc = alloc_arraycache(node, cachep->limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800900 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 if (!nc)
902 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
Christoph Lametere498be72005-09-09 13:03:32 -0700905 l3 = cachep->nodelists[node];
906 BUG_ON(!l3);
907 if (!l3->shared) {
908 if (!(nc = alloc_arraycache(node,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800909 cachep->shared *
910 cachep->batchcount,
911 0xbaadf00d)))
912 goto bad;
Christoph Lametere498be72005-09-09 13:03:32 -0700913
914 /* we are serialised from CPU_DEAD or
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800915 CPU_UP_CANCELLED by the cpucontrol lock */
Christoph Lametere498be72005-09-09 13:03:32 -0700916 l3->shared = nc;
917 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800919 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 break;
921 case CPU_ONLINE:
922 start_cpu_timer(cpu);
923 break;
924#ifdef CONFIG_HOTPLUG_CPU
925 case CPU_DEAD:
926 /* fall thru */
927 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800928 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
930 list_for_each_entry(cachep, &cache_chain, next) {
931 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700932 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
Christoph Lametere498be72005-09-09 13:03:32 -0700934 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 spin_lock_irq(&cachep->spinlock);
936 /* cpu is dead; no one can alloc from it. */
937 nc = cachep->array[cpu];
938 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700939 l3 = cachep->nodelists[node];
940
941 if (!l3)
942 goto unlock_cache;
943
944 spin_lock(&l3->list_lock);
945
946 /* Free limit for this kmem_list3 */
947 l3->free_limit -= cachep->batchcount;
948 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -0700949 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700950
951 if (!cpus_empty(mask)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 spin_unlock(&l3->list_lock);
953 goto unlock_cache;
954 }
Christoph Lametere498be72005-09-09 13:03:32 -0700955
956 if (l3->shared) {
957 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800958 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700959 kfree(l3->shared);
960 l3->shared = NULL;
961 }
962 if (l3->alien) {
963 drain_alien_cache(cachep, l3);
964 free_alien_cache(l3->alien);
965 l3->alien = NULL;
966 }
967
968 /* free slabs belonging to this node */
969 if (__node_shrink(cachep, node)) {
970 cachep->nodelists[node] = NULL;
971 spin_unlock(&l3->list_lock);
972 kfree(l3);
973 } else {
974 spin_unlock(&l3->list_lock);
975 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800976 unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 spin_unlock_irq(&cachep->spinlock);
978 kfree(nc);
979 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800980 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981 break;
982#endif
983 }
984 return NOTIFY_OK;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800985 bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800986 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 return NOTIFY_BAD;
988}
989
990static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
991
Christoph Lametere498be72005-09-09 13:03:32 -0700992/*
993 * swap the static kmem_list3 with kmalloced memory
994 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800995static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -0700996{
997 struct kmem_list3 *ptr;
998
999 BUG_ON(cachep->nodelists[nodeid] != list);
1000 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1001 BUG_ON(!ptr);
1002
1003 local_irq_disable();
1004 memcpy(ptr, list, sizeof(struct kmem_list3));
1005 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1006 cachep->nodelists[nodeid] = ptr;
1007 local_irq_enable();
1008}
1009
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010/* Initialisation.
1011 * Called after the gfp() functions have been enabled, and before smp_init().
1012 */
1013void __init kmem_cache_init(void)
1014{
1015 size_t left_over;
1016 struct cache_sizes *sizes;
1017 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001018 int i;
1019
1020 for (i = 0; i < NUM_INIT_LISTS; i++) {
1021 kmem_list3_init(&initkmem_list3[i]);
1022 if (i < MAX_NUMNODES)
1023 cache_cache.nodelists[i] = NULL;
1024 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025
1026 /*
1027 * Fragmentation resistance on low memory - only use bigger
1028 * page orders on machines with more than 32MB of memory.
1029 */
1030 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1031 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1032
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033 /* Bootstrap is tricky, because several objects are allocated
1034 * from caches that do not exist yet:
1035 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1036 * structures of all caches, except cache_cache itself: cache_cache
1037 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001038 * Initially an __init data area is used for the head array and the
1039 * kmem_list3 structures, it's replaced with a kmalloc allocated
1040 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001042 * The kmem_cache_t for the new cache is allocated normally.
1043 * An __init data area is used for the head array.
1044 * 3) Create the remaining kmalloc caches, with minimally sized
1045 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001046 * 4) Replace the __init data head arrays for cache_cache and the first
1047 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001048 * 5) Replace the __init data for kmem_list3 for cache_cache and
1049 * the other cache's with kmalloc allocated memory.
1050 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 */
1052
1053 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 INIT_LIST_HEAD(&cache_chain);
1055 list_add(&cache_cache.next, &cache_chain);
1056 cache_cache.colour_off = cache_line_size();
1057 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001058 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
1060 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1061
1062 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001063 &left_over, &cache_cache.num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (!cache_cache.num)
1065 BUG();
1066
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001067 cache_cache.colour = left_over / cache_cache.colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 cache_cache.colour_next = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001069 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1070 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071
1072 /* 2+3) create the kmalloc caches */
1073 sizes = malloc_sizes;
1074 names = cache_names;
1075
Christoph Lametere498be72005-09-09 13:03:32 -07001076 /* Initialize the caches that provide memory for the array cache
1077 * and the kmem_list3 structures first.
1078 * Without this, further allocations will bug
1079 */
1080
1081 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001082 sizes[INDEX_AC].cs_size,
1083 ARCH_KMALLOC_MINALIGN,
1084 (ARCH_KMALLOC_FLAGS |
1085 SLAB_PANIC), NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001086
1087 if (INDEX_AC != INDEX_L3)
1088 sizes[INDEX_L3].cs_cachep =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001089 kmem_cache_create(names[INDEX_L3].name,
1090 sizes[INDEX_L3].cs_size,
1091 ARCH_KMALLOC_MINALIGN,
1092 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1093 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001094
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001096 /*
1097 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 * This should be particularly beneficial on SMP boxes, as it
1099 * eliminates "false sharing".
1100 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001101 * allow tighter packing of the smaller caches.
1102 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001103 if (!sizes->cs_cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07001104 sizes->cs_cachep = kmem_cache_create(names->name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001105 sizes->cs_size,
1106 ARCH_KMALLOC_MINALIGN,
1107 (ARCH_KMALLOC_FLAGS
1108 | SLAB_PANIC),
1109 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110
1111 /* Inc off-slab bufctl limit until the ceiling is hit. */
1112 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001113 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 offslab_limit /= sizeof(kmem_bufctl_t);
1115 }
1116
1117 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001118 sizes->cs_size,
1119 ARCH_KMALLOC_MINALIGN,
1120 (ARCH_KMALLOC_FLAGS |
1121 SLAB_CACHE_DMA |
1122 SLAB_PANIC), NULL,
1123 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
1125 sizes++;
1126 names++;
1127 }
1128 /* 4) Replace the bootstrap head arrays */
1129 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001130 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001131
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 local_irq_disable();
1135 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001136 memcpy(ptr, ac_data(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001137 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 cache_cache.array[smp_processor_id()] = ptr;
1139 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001140
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001142
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001144 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001145 != &initarray_generic.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001146 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001147 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001148 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001149 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 local_irq_enable();
1151 }
Christoph Lametere498be72005-09-09 13:03:32 -07001152 /* 5) Replace the bootstrap kmem_list3's */
1153 {
1154 int node;
1155 /* Replace the static kmem_list3 structures for the boot cpu */
1156 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001157 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
Christoph Lametere498be72005-09-09 13:03:32 -07001159 for_each_online_node(node) {
1160 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001161 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001162
1163 if (INDEX_AC != INDEX_L3) {
1164 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001165 &initkmem_list3[SIZE_L3 + node],
1166 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001167 }
1168 }
1169 }
1170
1171 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 {
1173 kmem_cache_t *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001174 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001176 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001177 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 }
1179
1180 /* Done! */
1181 g_cpucache_up = FULL;
1182
1183 /* Register a cpu startup notifier callback
1184 * that initializes ac_data for all new cpus
1185 */
1186 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187
1188 /* The reap timers are started later, with a module init call:
1189 * That part of the kernel is not yet operational.
1190 */
1191}
1192
1193static int __init cpucache_init(void)
1194{
1195 int cpu;
1196
1197 /*
1198 * Register the timers that return unneeded
1199 * pages to gfp.
1200 */
Christoph Lametere498be72005-09-09 13:03:32 -07001201 for_each_online_cpu(cpu)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001202 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203
1204 return 0;
1205}
1206
1207__initcall(cpucache_init);
1208
1209/*
1210 * Interface to system's page allocator. No need to hold the cache-lock.
1211 *
1212 * If we requested dmaable memory, we will get it. Even if we
1213 * did not request dmaable memory, we might get it, but that
1214 * would be relatively rare and ignorable.
1215 */
Al Virodd0fc662005-10-07 07:46:04 +01001216static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217{
1218 struct page *page;
1219 void *addr;
1220 int i;
1221
1222 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001223 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 if (!page)
1225 return NULL;
1226 addr = page_address(page);
1227
1228 i = (1 << cachep->gfporder);
1229 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1230 atomic_add(i, &slab_reclaim_pages);
1231 add_page_state(nr_slab, i);
1232 while (i--) {
1233 SetPageSlab(page);
1234 page++;
1235 }
1236 return addr;
1237}
1238
1239/*
1240 * Interface to system's page release.
1241 */
1242static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1243{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001244 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 struct page *page = virt_to_page(addr);
1246 const unsigned long nr_freed = i;
1247
1248 while (i--) {
1249 if (!TestClearPageSlab(page))
1250 BUG();
1251 page++;
1252 }
1253 sub_page_state(nr_slab, nr_freed);
1254 if (current->reclaim_state)
1255 current->reclaim_state->reclaimed_slab += nr_freed;
1256 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001257 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1258 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259}
1260
1261static void kmem_rcu_free(struct rcu_head *head)
1262{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001263 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 kmem_cache_t *cachep = slab_rcu->cachep;
1265
1266 kmem_freepages(cachep, slab_rcu->addr);
1267 if (OFF_SLAB(cachep))
1268 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1269}
1270
1271#if DEBUG
1272
1273#ifdef CONFIG_DEBUG_PAGEALLOC
1274static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001275 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276{
1277 int size = obj_reallen(cachep);
1278
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001279 addr = (unsigned long *)&((char *)addr)[obj_dbghead(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001281 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 return;
1283
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001284 *addr++ = 0x12345678;
1285 *addr++ = caller;
1286 *addr++ = smp_processor_id();
1287 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 {
1289 unsigned long *sptr = &caller;
1290 unsigned long svalue;
1291
1292 while (!kstack_end(sptr)) {
1293 svalue = *sptr++;
1294 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001295 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 size -= sizeof(unsigned long);
1297 if (size <= sizeof(unsigned long))
1298 break;
1299 }
1300 }
1301
1302 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001303 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304}
1305#endif
1306
1307static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1308{
1309 int size = obj_reallen(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001310 addr = &((char *)addr)[obj_dbghead(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311
1312 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001313 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314}
1315
1316static void dump_line(char *data, int offset, int limit)
1317{
1318 int i;
1319 printk(KERN_ERR "%03x:", offset);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001320 for (i = 0; i < limit; i++) {
1321 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 }
1323 printk("\n");
1324}
1325#endif
1326
1327#if DEBUG
1328
1329static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1330{
1331 int i, size;
1332 char *realobj;
1333
1334 if (cachep->flags & SLAB_RED_ZONE) {
1335 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001336 *dbg_redzone1(cachep, objp),
1337 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 }
1339
1340 if (cachep->flags & SLAB_STORE_USER) {
1341 printk(KERN_ERR "Last user: [<%p>]",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001342 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 print_symbol("(%s)",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001344 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 printk("\n");
1346 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001347 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 size = obj_reallen(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001349 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 int limit;
1351 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001352 if (i + limit > size)
1353 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 dump_line(realobj, i, limit);
1355 }
1356}
1357
1358static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1359{
1360 char *realobj;
1361 int size, i;
1362 int lines = 0;
1363
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001364 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 size = obj_reallen(cachep);
1366
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001367 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001369 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 exp = POISON_END;
1371 if (realobj[i] != exp) {
1372 int limit;
1373 /* Mismatch ! */
1374 /* Print header */
1375 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001376 printk(KERN_ERR
1377 "Slab corruption: start=%p, len=%d\n",
1378 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 print_objinfo(cachep, objp, 0);
1380 }
1381 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001382 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001384 if (i + limit > size)
1385 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 dump_line(realobj, i, limit);
1387 i += 16;
1388 lines++;
1389 /* Limit to 5 lines */
1390 if (lines > 5)
1391 break;
1392 }
1393 }
1394 if (lines != 0) {
1395 /* Print some data about the neighboring objects, if they
1396 * exist:
1397 */
Pekka Enberg065d41c2005-11-13 16:06:46 -08001398 struct slab *slabp = page_get_slab(virt_to_page(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 int objnr;
1400
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001401 objnr = (objp - slabp->s_mem) / cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 if (objnr) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001403 objp = slabp->s_mem + (objnr - 1) * cachep->objsize;
1404 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001406 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 print_objinfo(cachep, objp, 2);
1408 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001409 if (objnr + 1 < cachep->num) {
1410 objp = slabp->s_mem + (objnr + 1) * cachep->objsize;
1411 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001413 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 print_objinfo(cachep, objp, 2);
1415 }
1416 }
1417}
1418#endif
1419
1420/* Destroy all the objs in a slab, and release the mem back to the system.
1421 * Before calling the slab must have been unlinked from the cache.
1422 * The cache-lock is not held/needed.
1423 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001424static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425{
1426 void *addr = slabp->s_mem - slabp->colouroff;
1427
1428#if DEBUG
1429 int i;
1430 for (i = 0; i < cachep->num; i++) {
1431 void *objp = slabp->s_mem + cachep->objsize * i;
1432
1433 if (cachep->flags & SLAB_POISON) {
1434#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001435 if ((cachep->objsize % PAGE_SIZE) == 0
1436 && OFF_SLAB(cachep))
1437 kernel_map_pages(virt_to_page(objp),
1438 cachep->objsize / PAGE_SIZE,
1439 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 else
1441 check_poison_obj(cachep, objp);
1442#else
1443 check_poison_obj(cachep, objp);
1444#endif
1445 }
1446 if (cachep->flags & SLAB_RED_ZONE) {
1447 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1448 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001449 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1451 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001452 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 }
1454 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001455 (cachep->dtor) (objp + obj_dbghead(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 }
1457#else
1458 if (cachep->dtor) {
1459 int i;
1460 for (i = 0; i < cachep->num; i++) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001461 void *objp = slabp->s_mem + cachep->objsize * i;
1462 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 }
1464 }
1465#endif
1466
1467 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1468 struct slab_rcu *slab_rcu;
1469
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001470 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471 slab_rcu->cachep = cachep;
1472 slab_rcu->addr = addr;
1473 call_rcu(&slab_rcu->head, kmem_rcu_free);
1474 } else {
1475 kmem_freepages(cachep, addr);
1476 if (OFF_SLAB(cachep))
1477 kmem_cache_free(cachep->slabp_cache, slabp);
1478 }
1479}
1480
Christoph Lametere498be72005-09-09 13:03:32 -07001481/* For setting up all the kmem_list3s for cache whose objsize is same
1482 as size of kmem_list3. */
1483static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1484{
1485 int node;
1486
1487 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001488 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001489 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001490 REAPTIMEOUT_LIST3 +
1491 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001492 }
1493}
1494
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495/**
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001496 * calculate_slab_order - calculate size (page order) of slabs and the number
1497 * of objects per slab.
1498 *
1499 * This could be made much more intelligent. For now, try to avoid using
1500 * high order pages for slabs. When the gfp() functions are more friendly
1501 * towards high-order requests, this should be changed.
1502 */
1503static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1504 size_t align, gfp_t flags)
1505{
1506 size_t left_over = 0;
1507
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 for (;; cachep->gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001509 unsigned int num;
1510 size_t remainder;
1511
1512 if (cachep->gfporder > MAX_GFP_ORDER) {
1513 cachep->num = 0;
1514 break;
1515 }
1516
1517 cache_estimate(cachep->gfporder, size, align, flags,
1518 &remainder, &num);
1519 if (!num)
1520 continue;
1521 /* More than offslab_limit objects will cause problems */
1522 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1523 break;
1524
1525 cachep->num = num;
1526 left_over = remainder;
1527
1528 /*
1529 * Large number of objects is good, but very large slabs are
1530 * currently bad for the gfp()s.
1531 */
1532 if (cachep->gfporder >= slab_break_gfp_order)
1533 break;
1534
1535 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1536 /* Acceptable internal fragmentation */
1537 break;
1538 }
1539 return left_over;
1540}
1541
1542/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 * kmem_cache_create - Create a cache.
1544 * @name: A string which is used in /proc/slabinfo to identify this cache.
1545 * @size: The size of objects to be created in this cache.
1546 * @align: The required alignment for the objects.
1547 * @flags: SLAB flags
1548 * @ctor: A constructor for the objects.
1549 * @dtor: A destructor for the objects.
1550 *
1551 * Returns a ptr to the cache on success, NULL on failure.
1552 * Cannot be called within a int, but can be interrupted.
1553 * The @ctor is run when new pages are allocated by the cache
1554 * and the @dtor is run before the pages are handed back.
1555 *
1556 * @name must be valid until the cache is destroyed. This implies that
1557 * the module calling this has to destroy the cache before getting
1558 * unloaded.
1559 *
1560 * The flags are
1561 *
1562 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1563 * to catch references to uninitialised memory.
1564 *
1565 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1566 * for buffer overruns.
1567 *
1568 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1569 * memory pressure.
1570 *
1571 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1572 * cacheline. This can be beneficial if you're counting cycles as closely
1573 * as davem.
1574 */
1575kmem_cache_t *
1576kmem_cache_create (const char *name, size_t size, size_t align,
1577 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1578 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1579{
1580 size_t left_over, slab_size, ralign;
1581 kmem_cache_t *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001582 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583
1584 /*
1585 * Sanity checks... these are all serious usage bugs.
1586 */
1587 if ((!name) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001588 in_interrupt() ||
1589 (size < BYTES_PER_WORD) ||
1590 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1591 printk(KERN_ERR "%s: Early error in slab %s\n",
1592 __FUNCTION__, name);
1593 BUG();
1594 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001596 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001597
1598 list_for_each(p, &cache_chain) {
1599 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1600 mm_segment_t old_fs = get_fs();
1601 char tmp;
1602 int res;
1603
1604 /*
1605 * This happens when the module gets unloaded and doesn't
1606 * destroy its slab cache and no-one else reuses the vmalloc
1607 * area of the module. Print a warning.
1608 */
1609 set_fs(KERNEL_DS);
1610 res = __get_user(tmp, pc->name);
1611 set_fs(old_fs);
1612 if (res) {
1613 printk("SLAB: cache with size %d has lost its name\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001614 pc->objsize);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001615 continue;
1616 }
1617
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001618 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001619 printk("kmem_cache_create: duplicate cache %s\n", name);
1620 dump_stack();
1621 goto oops;
1622 }
1623 }
1624
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625#if DEBUG
1626 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1627 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1628 /* No constructor, but inital state check requested */
1629 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001630 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 flags &= ~SLAB_DEBUG_INITIAL;
1632 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633#if FORCED_DEBUG
1634 /*
1635 * Enable redzoning and last user accounting, except for caches with
1636 * large objects, if the increased size would increase the object size
1637 * above the next power of two: caches with object sizes just above a
1638 * power of two have a significant amount of internal fragmentation.
1639 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001640 if ((size < 4096
1641 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1642 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 if (!(flags & SLAB_DESTROY_BY_RCU))
1644 flags |= SLAB_POISON;
1645#endif
1646 if (flags & SLAB_DESTROY_BY_RCU)
1647 BUG_ON(flags & SLAB_POISON);
1648#endif
1649 if (flags & SLAB_DESTROY_BY_RCU)
1650 BUG_ON(dtor);
1651
1652 /*
1653 * Always checks flags, a caller might be expecting debug
1654 * support which isn't available.
1655 */
1656 if (flags & ~CREATE_MASK)
1657 BUG();
1658
1659 /* Check that size is in terms of words. This is needed to avoid
1660 * unaligned accesses for some archs when redzoning is used, and makes
1661 * sure any on-slab bufctl's are also correctly aligned.
1662 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001663 if (size & (BYTES_PER_WORD - 1)) {
1664 size += (BYTES_PER_WORD - 1);
1665 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 }
1667
1668 /* calculate out the final buffer alignment: */
1669 /* 1) arch recommendation: can be overridden for debug */
1670 if (flags & SLAB_HWCACHE_ALIGN) {
1671 /* Default alignment: as specified by the arch code.
1672 * Except if an object is really small, then squeeze multiple
1673 * objects into one cacheline.
1674 */
1675 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001676 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 ralign /= 2;
1678 } else {
1679 ralign = BYTES_PER_WORD;
1680 }
1681 /* 2) arch mandated alignment: disables debug if necessary */
1682 if (ralign < ARCH_SLAB_MINALIGN) {
1683 ralign = ARCH_SLAB_MINALIGN;
1684 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001685 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 }
1687 /* 3) caller mandated alignment: disables debug if necessary */
1688 if (ralign < align) {
1689 ralign = align;
1690 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001691 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 }
1693 /* 4) Store it. Note that the debug code below can reduce
1694 * the alignment to BYTES_PER_WORD.
1695 */
1696 align = ralign;
1697
1698 /* Get cache's description obj. */
1699 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1700 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001701 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 memset(cachep, 0, sizeof(kmem_cache_t));
1703
1704#if DEBUG
1705 cachep->reallen = size;
1706
1707 if (flags & SLAB_RED_ZONE) {
1708 /* redzoning only works with word aligned caches */
1709 align = BYTES_PER_WORD;
1710
1711 /* add space for red zone words */
1712 cachep->dbghead += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001713 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 }
1715 if (flags & SLAB_STORE_USER) {
1716 /* user store requires word alignment and
1717 * one word storage behind the end of the real
1718 * object.
1719 */
1720 align = BYTES_PER_WORD;
1721 size += BYTES_PER_WORD;
1722 }
1723#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001724 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1725 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 cachep->dbghead += PAGE_SIZE - size;
1727 size = PAGE_SIZE;
1728 }
1729#endif
1730#endif
1731
1732 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001733 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 /*
1735 * Size is large, assume best to place the slab management obj
1736 * off-slab (should allow better packing of objs).
1737 */
1738 flags |= CFLGS_OFF_SLAB;
1739
1740 size = ALIGN(size, align);
1741
1742 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1743 /*
1744 * A VFS-reclaimable slab tends to have most allocations
1745 * as GFP_NOFS and we really don't want to have to be allocating
1746 * higher-order pages when we are unable to shrink dcache.
1747 */
1748 cachep->gfporder = 0;
1749 cache_estimate(cachep->gfporder, size, align, flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001750 &left_over, &cachep->num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001751 } else
1752 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753
1754 if (!cachep->num) {
1755 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1756 kmem_cache_free(&cache_cache, cachep);
1757 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001758 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001760 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1761 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762
1763 /*
1764 * If the slab has been placed off-slab, and we have enough space then
1765 * move it on-slab. This is at the expense of any extra colouring.
1766 */
1767 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1768 flags &= ~CFLGS_OFF_SLAB;
1769 left_over -= slab_size;
1770 }
1771
1772 if (flags & CFLGS_OFF_SLAB) {
1773 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001774 slab_size =
1775 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 }
1777
1778 cachep->colour_off = cache_line_size();
1779 /* Offset must be a multiple of the alignment. */
1780 if (cachep->colour_off < align)
1781 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001782 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 cachep->slab_size = slab_size;
1784 cachep->flags = flags;
1785 cachep->gfpflags = 0;
1786 if (flags & SLAB_CACHE_DMA)
1787 cachep->gfpflags |= GFP_DMA;
1788 spin_lock_init(&cachep->spinlock);
1789 cachep->objsize = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790
1791 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001792 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 cachep->ctor = ctor;
1794 cachep->dtor = dtor;
1795 cachep->name = name;
1796
1797 /* Don't let CPUs to come and go */
1798 lock_cpu_hotplug();
1799
1800 if (g_cpucache_up == FULL) {
1801 enable_cpucache(cachep);
1802 } else {
1803 if (g_cpucache_up == NONE) {
1804 /* Note: the first kmem_cache_create must create
1805 * the cache that's used by kmalloc(24), otherwise
1806 * the creation of further caches will BUG().
1807 */
Christoph Lametere498be72005-09-09 13:03:32 -07001808 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 &initarray_generic.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001810
1811 /* If the cache that's used by
1812 * kmalloc(sizeof(kmem_list3)) is the first cache,
1813 * then we need to set up all its list3s, otherwise
1814 * the creation of further caches will BUG().
1815 */
1816 set_up_list3s(cachep, SIZE_AC);
1817 if (INDEX_AC == INDEX_L3)
1818 g_cpucache_up = PARTIAL_L3;
1819 else
1820 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001822 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001823 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001824
1825 if (g_cpucache_up == PARTIAL_AC) {
1826 set_up_list3s(cachep, SIZE_L3);
1827 g_cpucache_up = PARTIAL_L3;
1828 } else {
1829 int node;
1830 for_each_online_node(node) {
1831
1832 cachep->nodelists[node] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 kmalloc_node(sizeof
1834 (struct kmem_list3),
1835 GFP_KERNEL, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001836 BUG_ON(!cachep->nodelists[node]);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001837 kmem_list3_init(cachep->
1838 nodelists[node]);
Christoph Lametere498be72005-09-09 13:03:32 -07001839 }
1840 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 }
Christoph Lametere498be72005-09-09 13:03:32 -07001842 cachep->nodelists[numa_node_id()]->next_reap =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001843 jiffies + REAPTIMEOUT_LIST3 +
1844 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001845
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 BUG_ON(!ac_data(cachep));
1847 ac_data(cachep)->avail = 0;
1848 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1849 ac_data(cachep)->batchcount = 1;
1850 ac_data(cachep)->touched = 0;
1851 cachep->batchcount = 1;
1852 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001853 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 /* cache setup completed, link it into the list */
1856 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 unlock_cpu_hotplug();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 if (!cachep && (flags & SLAB_PANIC))
1860 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001862 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 return cachep;
1864}
1865EXPORT_SYMBOL(kmem_cache_create);
1866
1867#if DEBUG
1868static void check_irq_off(void)
1869{
1870 BUG_ON(!irqs_disabled());
1871}
1872
1873static void check_irq_on(void)
1874{
1875 BUG_ON(irqs_disabled());
1876}
1877
1878static void check_spinlock_acquired(kmem_cache_t *cachep)
1879{
1880#ifdef CONFIG_SMP
1881 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001882 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883#endif
1884}
Christoph Lametere498be72005-09-09 13:03:32 -07001885
1886static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1887{
1888#ifdef CONFIG_SMP
1889 check_irq_off();
1890 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1891#endif
1892}
1893
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894#else
1895#define check_irq_off() do { } while(0)
1896#define check_irq_on() do { } while(0)
1897#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001898#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899#endif
1900
1901/*
1902 * Waits for all CPUs to execute func().
1903 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001904static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905{
1906 check_irq_on();
1907 preempt_disable();
1908
1909 local_irq_disable();
1910 func(arg);
1911 local_irq_enable();
1912
1913 if (smp_call_function(func, arg, 1, 1))
1914 BUG();
1915
1916 preempt_enable();
1917}
1918
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001919static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1920 int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921
1922static void do_drain(void *arg)
1923{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 kmem_cache_t *cachep = (kmem_cache_t *) arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07001926 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
1928 check_irq_off();
1929 ac = ac_data(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07001930 spin_lock(&cachep->nodelists[node]->list_lock);
1931 free_block(cachep, ac->entry, ac->avail, node);
1932 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 ac->avail = 0;
1934}
1935
1936static void drain_cpu_caches(kmem_cache_t *cachep)
1937{
Christoph Lametere498be72005-09-09 13:03:32 -07001938 struct kmem_list3 *l3;
1939 int node;
1940
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 smp_call_function_all_cpus(do_drain, cachep);
1942 check_irq_on();
1943 spin_lock_irq(&cachep->spinlock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001944 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07001945 l3 = cachep->nodelists[node];
1946 if (l3) {
1947 spin_lock(&l3->list_lock);
1948 drain_array_locked(cachep, l3->shared, 1, node);
1949 spin_unlock(&l3->list_lock);
1950 if (l3->alien)
1951 drain_alien_cache(cachep, l3);
1952 }
1953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 spin_unlock_irq(&cachep->spinlock);
1955}
1956
Christoph Lametere498be72005-09-09 13:03:32 -07001957static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958{
1959 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07001960 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 int ret;
1962
Christoph Lametere498be72005-09-09 13:03:32 -07001963 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 struct list_head *p;
1965
Christoph Lametere498be72005-09-09 13:03:32 -07001966 p = l3->slabs_free.prev;
1967 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 break;
1969
Christoph Lametere498be72005-09-09 13:03:32 -07001970 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971#if DEBUG
1972 if (slabp->inuse)
1973 BUG();
1974#endif
1975 list_del(&slabp->list);
1976
Christoph Lametere498be72005-09-09 13:03:32 -07001977 l3->free_objects -= cachep->num;
1978 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07001980 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001982 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 return ret;
1984}
1985
Christoph Lametere498be72005-09-09 13:03:32 -07001986static int __cache_shrink(kmem_cache_t *cachep)
1987{
1988 int ret = 0, i = 0;
1989 struct kmem_list3 *l3;
1990
1991 drain_cpu_caches(cachep);
1992
1993 check_irq_on();
1994 for_each_online_node(i) {
1995 l3 = cachep->nodelists[i];
1996 if (l3) {
1997 spin_lock_irq(&l3->list_lock);
1998 ret += __node_shrink(cachep, i);
1999 spin_unlock_irq(&l3->list_lock);
2000 }
2001 }
2002 return (ret ? 1 : 0);
2003}
2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005/**
2006 * kmem_cache_shrink - Shrink a cache.
2007 * @cachep: The cache to shrink.
2008 *
2009 * Releases as many slabs as possible for a cache.
2010 * To help debugging, a zero exit status indicates all slabs were released.
2011 */
2012int kmem_cache_shrink(kmem_cache_t *cachep)
2013{
2014 if (!cachep || in_interrupt())
2015 BUG();
2016
2017 return __cache_shrink(cachep);
2018}
2019EXPORT_SYMBOL(kmem_cache_shrink);
2020
2021/**
2022 * kmem_cache_destroy - delete a cache
2023 * @cachep: the cache to destroy
2024 *
2025 * Remove a kmem_cache_t object from the slab cache.
2026 * Returns 0 on success.
2027 *
2028 * It is expected this function will be called by a module when it is
2029 * unloaded. This will remove the cache completely, and avoid a duplicate
2030 * cache being allocated each time a module is loaded and unloaded, if the
2031 * module doesn't have persistent in-kernel storage across loads and unloads.
2032 *
2033 * The cache must be empty before calling this function.
2034 *
2035 * The caller must guarantee that noone will allocate memory from the cache
2036 * during the kmem_cache_destroy().
2037 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002038int kmem_cache_destroy(kmem_cache_t *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039{
2040 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002041 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042
2043 if (!cachep || in_interrupt())
2044 BUG();
2045
2046 /* Don't let CPUs to come and go */
2047 lock_cpu_hotplug();
2048
2049 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002050 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 /*
2052 * the chain is never empty, cache_cache is never destroyed
2053 */
2054 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002055 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056
2057 if (__cache_shrink(cachep)) {
2058 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002059 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002060 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002061 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 unlock_cpu_hotplug();
2063 return 1;
2064 }
2065
2066 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002067 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
Christoph Lametere498be72005-09-09 13:03:32 -07002069 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002070 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071
2072 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002073 for_each_online_node(i) {
2074 if ((l3 = cachep->nodelists[i])) {
2075 kfree(l3->shared);
2076 free_alien_cache(l3->alien);
2077 kfree(l3);
2078 }
2079 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 kmem_cache_free(&cache_cache, cachep);
2081
2082 unlock_cpu_hotplug();
2083
2084 return 0;
2085}
2086EXPORT_SYMBOL(kmem_cache_destroy);
2087
2088/* Get the memory for a slab management obj. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002089static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2090 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091{
2092 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 if (OFF_SLAB(cachep)) {
2095 /* Slab management obj is off-slab. */
2096 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2097 if (!slabp)
2098 return NULL;
2099 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002100 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 colour_off += cachep->slab_size;
2102 }
2103 slabp->inuse = 0;
2104 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002105 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106
2107 return slabp;
2108}
2109
2110static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2111{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002112 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113}
2114
2115static void cache_init_objs(kmem_cache_t *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002116 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117{
2118 int i;
2119
2120 for (i = 0; i < cachep->num; i++) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002121 void *objp = slabp->s_mem + cachep->objsize * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122#if DEBUG
2123 /* need to poison the objs? */
2124 if (cachep->flags & SLAB_POISON)
2125 poison_obj(cachep, objp, POISON_FREE);
2126 if (cachep->flags & SLAB_STORE_USER)
2127 *dbg_userword(cachep, objp) = NULL;
2128
2129 if (cachep->flags & SLAB_RED_ZONE) {
2130 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2131 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2132 }
2133 /*
2134 * Constructors are not allowed to allocate memory from
2135 * the same cache which they are a constructor for.
2136 * Otherwise, deadlock. They must also be threaded.
2137 */
2138 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002139 cachep->ctor(objp + obj_dbghead(cachep), cachep,
2140 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
2142 if (cachep->flags & SLAB_RED_ZONE) {
2143 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2144 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002145 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2147 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002148 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002150 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2151 && cachep->flags & SLAB_POISON)
2152 kernel_map_pages(virt_to_page(objp),
2153 cachep->objsize / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154#else
2155 if (cachep->ctor)
2156 cachep->ctor(objp, cachep, ctor_flags);
2157#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002158 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002160 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 slabp->free = 0;
2162}
2163
Al Viro6daa0e22005-10-21 03:18:50 -04002164static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165{
2166 if (flags & SLAB_DMA) {
2167 if (!(cachep->gfpflags & GFP_DMA))
2168 BUG();
2169 } else {
2170 if (cachep->gfpflags & GFP_DMA)
2171 BUG();
2172 }
2173}
2174
2175static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2176{
2177 int i;
2178 struct page *page;
2179
2180 /* Nasty!!!!!! I hope this is OK. */
2181 i = 1 << cachep->gfporder;
2182 page = virt_to_page(objp);
2183 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002184 page_set_cache(page, cachep);
2185 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 page++;
2187 } while (--i);
2188}
2189
2190/*
2191 * Grow (by 1) the number of slabs within a cache. This is called by
2192 * kmem_cache_alloc() when there are no active objs left in a cache.
2193 */
Al Virodd0fc662005-10-07 07:46:04 +01002194static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002196 struct slab *slabp;
2197 void *objp;
2198 size_t offset;
2199 gfp_t local_flags;
2200 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002201 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202
2203 /* Be lazy and only check for valid flags here,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002204 * keeping it out of the critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002206 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 BUG();
2208 if (flags & SLAB_NO_GROW)
2209 return 0;
2210
2211 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2212 local_flags = (flags & SLAB_LEVEL_MASK);
2213 if (!(local_flags & __GFP_WAIT))
2214 /*
2215 * Not allowed to sleep. Need to tell a constructor about
2216 * this - it might need to know...
2217 */
2218 ctor_flags |= SLAB_CTOR_ATOMIC;
2219
2220 /* About to mess with non-constant members - lock. */
2221 check_irq_off();
2222 spin_lock(&cachep->spinlock);
2223
2224 /* Get colour for the slab, and cal the next value. */
2225 offset = cachep->colour_next;
2226 cachep->colour_next++;
2227 if (cachep->colour_next >= cachep->colour)
2228 cachep->colour_next = 0;
2229 offset *= cachep->colour_off;
2230
2231 spin_unlock(&cachep->spinlock);
2232
Christoph Lametere498be72005-09-09 13:03:32 -07002233 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 if (local_flags & __GFP_WAIT)
2235 local_irq_enable();
2236
2237 /*
2238 * The test for missing atomic flag is performed here, rather than
2239 * the more obvious place, simply to reduce the critical path length
2240 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2241 * will eventually be caught here (where it matters).
2242 */
2243 kmem_flagcheck(cachep, flags);
2244
Christoph Lametere498be72005-09-09 13:03:32 -07002245 /* Get mem for the objs.
2246 * Attempt to allocate a physical page from 'nodeid',
2247 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2249 goto failed;
2250
2251 /* Get slab management. */
2252 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2253 goto opps1;
2254
Christoph Lametere498be72005-09-09 13:03:32 -07002255 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 set_slab_attr(cachep, slabp, objp);
2257
2258 cache_init_objs(cachep, slabp, ctor_flags);
2259
2260 if (local_flags & __GFP_WAIT)
2261 local_irq_disable();
2262 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002263 l3 = cachep->nodelists[nodeid];
2264 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
2266 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002267 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002269 l3->free_objects += cachep->num;
2270 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002272 opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 kmem_freepages(cachep, objp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002274 failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 if (local_flags & __GFP_WAIT)
2276 local_irq_disable();
2277 return 0;
2278}
2279
2280#if DEBUG
2281
2282/*
2283 * Perform extra freeing checks:
2284 * - detect bad pointers.
2285 * - POISON/RED_ZONE checking
2286 * - destructor calls, for caches with POISON+dtor
2287 */
2288static void kfree_debugcheck(const void *objp)
2289{
2290 struct page *page;
2291
2292 if (!virt_addr_valid(objp)) {
2293 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002294 (unsigned long)objp);
2295 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 }
2297 page = virt_to_page(objp);
2298 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002299 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2300 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 BUG();
2302 }
2303}
2304
2305static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002306 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307{
2308 struct page *page;
2309 unsigned int objnr;
2310 struct slab *slabp;
2311
2312 objp -= obj_dbghead(cachep);
2313 kfree_debugcheck(objp);
2314 page = virt_to_page(objp);
2315
Pekka Enberg065d41c2005-11-13 16:06:46 -08002316 if (page_get_cache(page) != cachep) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002317 printk(KERN_ERR
2318 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2319 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002321 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2322 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 WARN_ON(1);
2324 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002325 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326
2327 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002328 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2329 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2330 slab_error(cachep,
2331 "double free, or memory outside"
2332 " object was overwritten");
2333 printk(KERN_ERR
2334 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2335 objp, *dbg_redzone1(cachep, objp),
2336 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 }
2338 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2339 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2340 }
2341 if (cachep->flags & SLAB_STORE_USER)
2342 *dbg_userword(cachep, objp) = caller;
2343
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002344 objnr = (objp - slabp->s_mem) / cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345
2346 BUG_ON(objnr >= cachep->num);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002347 BUG_ON(objp != slabp->s_mem + objnr * cachep->objsize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348
2349 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2350 /* Need to call the slab's constructor so the
2351 * caller can perform a verify of its state (debugging).
2352 * Called without the cache-lock held.
2353 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002354 cachep->ctor(objp + obj_dbghead(cachep),
2355 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 }
2357 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2358 /* we want to cache poison the object,
2359 * call the destruction callback
2360 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002361 cachep->dtor(objp + obj_dbghead(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 }
2363 if (cachep->flags & SLAB_POISON) {
2364#ifdef CONFIG_DEBUG_PAGEALLOC
2365 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2366 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002367 kernel_map_pages(virt_to_page(objp),
2368 cachep->objsize / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 } else {
2370 poison_obj(cachep, objp, POISON_FREE);
2371 }
2372#else
2373 poison_obj(cachep, objp, POISON_FREE);
2374#endif
2375 }
2376 return objp;
2377}
2378
2379static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2380{
2381 kmem_bufctl_t i;
2382 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002383
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 /* Check slab's freelist to see if this obj is there. */
2385 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2386 entries++;
2387 if (entries > cachep->num || i >= cachep->num)
2388 goto bad;
2389 }
2390 if (entries != cachep->num - slabp->inuse) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002391 bad:
2392 printk(KERN_ERR
2393 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2394 cachep->name, cachep->num, slabp, slabp->inuse);
2395 for (i = 0;
2396 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2397 i++) {
2398 if ((i % 16) == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002400 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 }
2402 printk("\n");
2403 BUG();
2404 }
2405}
2406#else
2407#define kfree_debugcheck(x) do { } while(0)
2408#define cache_free_debugcheck(x,objp,z) (objp)
2409#define check_slabp(x,y) do { } while(0)
2410#endif
2411
Al Virodd0fc662005-10-07 07:46:04 +01002412static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413{
2414 int batchcount;
2415 struct kmem_list3 *l3;
2416 struct array_cache *ac;
2417
2418 check_irq_off();
2419 ac = ac_data(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002420 retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 batchcount = ac->batchcount;
2422 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2423 /* if there was little recent activity on this
2424 * cache, then perform only a partial refill.
2425 * Otherwise we could generate refill bouncing.
2426 */
2427 batchcount = BATCHREFILL_LIMIT;
2428 }
Christoph Lametere498be72005-09-09 13:03:32 -07002429 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430
Christoph Lametere498be72005-09-09 13:03:32 -07002431 BUG_ON(ac->avail > 0 || !l3);
2432 spin_lock(&l3->list_lock);
2433
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 if (l3->shared) {
2435 struct array_cache *shared_array = l3->shared;
2436 if (shared_array->avail) {
2437 if (batchcount > shared_array->avail)
2438 batchcount = shared_array->avail;
2439 shared_array->avail -= batchcount;
2440 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002441 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002442 &(shared_array->entry[shared_array->avail]),
2443 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 shared_array->touched = 1;
2445 goto alloc_done;
2446 }
2447 }
2448 while (batchcount > 0) {
2449 struct list_head *entry;
2450 struct slab *slabp;
2451 /* Get slab alloc is to come from. */
2452 entry = l3->slabs_partial.next;
2453 if (entry == &l3->slabs_partial) {
2454 l3->free_touched = 1;
2455 entry = l3->slabs_free.next;
2456 if (entry == &l3->slabs_free)
2457 goto must_grow;
2458 }
2459
2460 slabp = list_entry(entry, struct slab, list);
2461 check_slabp(cachep, slabp);
2462 check_spinlock_acquired(cachep);
2463 while (slabp->inuse < cachep->num && batchcount--) {
2464 kmem_bufctl_t next;
2465 STATS_INC_ALLOCED(cachep);
2466 STATS_INC_ACTIVE(cachep);
2467 STATS_SET_HIGH(cachep);
2468
2469 /* get obj pointer */
Christoph Lametere498be72005-09-09 13:03:32 -07002470 ac->entry[ac->avail++] = slabp->s_mem +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002471 slabp->free * cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
2473 slabp->inuse++;
2474 next = slab_bufctl(slabp)[slabp->free];
2475#if DEBUG
2476 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002477 WARN_ON(numa_node_id() != slabp->nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002479 slabp->free = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 }
2481 check_slabp(cachep, slabp);
2482
2483 /* move slabp to correct slabp list: */
2484 list_del(&slabp->list);
2485 if (slabp->free == BUFCTL_END)
2486 list_add(&slabp->list, &l3->slabs_full);
2487 else
2488 list_add(&slabp->list, &l3->slabs_partial);
2489 }
2490
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002491 must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 l3->free_objects -= ac->avail;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002493 alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002494 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
2496 if (unlikely(!ac->avail)) {
2497 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002498 x = cache_grow(cachep, flags, numa_node_id());
2499
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 // cache_grow can reenable interrupts, then ac could change.
2501 ac = ac_data(cachep);
2502 if (!x && ac->avail == 0) // no objects in sight? abort
2503 return NULL;
2504
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002505 if (!ac->avail) // objects refilled by interrupt?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 goto retry;
2507 }
2508 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002509 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510}
2511
2512static inline void
Al Virodd0fc662005-10-07 07:46:04 +01002513cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514{
2515 might_sleep_if(flags & __GFP_WAIT);
2516#if DEBUG
2517 kmem_flagcheck(cachep, flags);
2518#endif
2519}
2520
2521#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002522static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2523 void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002525 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002527 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528#ifdef CONFIG_DEBUG_PAGEALLOC
2529 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002530 kernel_map_pages(virt_to_page(objp),
2531 cachep->objsize / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 else
2533 check_poison_obj(cachep, objp);
2534#else
2535 check_poison_obj(cachep, objp);
2536#endif
2537 poison_obj(cachep, objp, POISON_INUSE);
2538 }
2539 if (cachep->flags & SLAB_STORE_USER)
2540 *dbg_userword(cachep, objp) = caller;
2541
2542 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002543 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2544 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2545 slab_error(cachep,
2546 "double free, or memory outside"
2547 " object was overwritten");
2548 printk(KERN_ERR
2549 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2550 objp, *dbg_redzone1(cachep, objp),
2551 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 }
2553 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2554 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2555 }
2556 objp += obj_dbghead(cachep);
2557 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002558 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559
2560 if (!(flags & __GFP_WAIT))
2561 ctor_flags |= SLAB_CTOR_ATOMIC;
2562
2563 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002564 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 return objp;
2566}
2567#else
2568#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2569#endif
2570
Al Virodd0fc662005-10-07 07:46:04 +01002571static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002573 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574 struct array_cache *ac;
2575
Christoph Lameterdc85da12006-01-18 17:42:36 -08002576#ifdef CONFIG_NUMA
Christoph Lameter86c562a2006-01-18 17:42:37 -08002577 if (unlikely(current->mempolicy && !in_interrupt())) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08002578 int nid = slab_node(current->mempolicy);
2579
2580 if (nid != numa_node_id())
2581 return __cache_alloc_node(cachep, flags, nid);
2582 }
2583#endif
2584
Alok N Kataria5c382302005-09-27 21:45:46 -07002585 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 ac = ac_data(cachep);
2587 if (likely(ac->avail)) {
2588 STATS_INC_ALLOCHIT(cachep);
2589 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002590 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 } else {
2592 STATS_INC_ALLOCMISS(cachep);
2593 objp = cache_alloc_refill(cachep, flags);
2594 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002595 return objp;
2596}
2597
Al Virodd0fc662005-10-07 07:46:04 +01002598static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Alok N Kataria5c382302005-09-27 21:45:46 -07002599{
2600 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002601 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002602
2603 cache_alloc_debugcheck_before(cachep, flags);
2604
2605 local_irq_save(save_flags);
2606 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002608 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002609 __builtin_return_address(0));
Eric Dumazet34342e82005-09-03 15:55:06 -07002610 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 return objp;
2612}
2613
Christoph Lametere498be72005-09-09 13:03:32 -07002614#ifdef CONFIG_NUMA
2615/*
2616 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 */
Al Viro6daa0e22005-10-21 03:18:50 -04002618static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002619{
2620 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002621 struct slab *slabp;
2622 struct kmem_list3 *l3;
2623 void *obj;
2624 kmem_bufctl_t next;
2625 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002627 l3 = cachep->nodelists[nodeid];
2628 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002629
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002630 retry:
2631 spin_lock(&l3->list_lock);
2632 entry = l3->slabs_partial.next;
2633 if (entry == &l3->slabs_partial) {
2634 l3->free_touched = 1;
2635 entry = l3->slabs_free.next;
2636 if (entry == &l3->slabs_free)
2637 goto must_grow;
2638 }
Christoph Lametere498be72005-09-09 13:03:32 -07002639
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 slabp = list_entry(entry, struct slab, list);
2641 check_spinlock_acquired_node(cachep, nodeid);
2642 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002643
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002644 STATS_INC_NODEALLOCS(cachep);
2645 STATS_INC_ACTIVE(cachep);
2646 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002647
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002648 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002649
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002650 /* get obj pointer */
2651 obj = slabp->s_mem + slabp->free * cachep->objsize;
2652 slabp->inuse++;
2653 next = slab_bufctl(slabp)[slabp->free];
Christoph Lametere498be72005-09-09 13:03:32 -07002654#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002655 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
Christoph Lametere498be72005-09-09 13:03:32 -07002656#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002657 slabp->free = next;
2658 check_slabp(cachep, slabp);
2659 l3->free_objects--;
2660 /* move slabp to correct slabp list: */
2661 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002662
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002663 if (slabp->free == BUFCTL_END) {
2664 list_add(&slabp->list, &l3->slabs_full);
2665 } else {
2666 list_add(&slabp->list, &l3->slabs_partial);
2667 }
Christoph Lametere498be72005-09-09 13:03:32 -07002668
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002669 spin_unlock(&l3->list_lock);
2670 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002671
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002672 must_grow:
2673 spin_unlock(&l3->list_lock);
2674 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002675
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 if (!x)
2677 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002678
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002679 goto retry;
2680 done:
2681 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002682}
2683#endif
2684
2685/*
2686 * Caller needs to acquire correct kmem_list's list_lock
2687 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002688static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2689 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690{
2691 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002692 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693
2694 for (i = 0; i < nr_objects; i++) {
2695 void *objp = objpp[i];
2696 struct slab *slabp;
2697 unsigned int objnr;
2698
Pekka Enberg065d41c2005-11-13 16:06:46 -08002699 slabp = page_get_slab(virt_to_page(objp));
Christoph Lameterff694162005-09-22 21:44:02 -07002700 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 list_del(&slabp->list);
2702 objnr = (objp - slabp->s_mem) / cachep->objsize;
Christoph Lameterff694162005-09-22 21:44:02 -07002703 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002705
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706#if DEBUG
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002707 /* Verify that the slab belongs to the intended node */
2708 WARN_ON(slabp->nodeid != node);
2709
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
Christoph Lametere498be72005-09-09 13:03:32 -07002711 printk(KERN_ERR "slab: double free detected in cache "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002712 "'%s', objp %p\n", cachep->name, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 BUG();
2714 }
2715#endif
2716 slab_bufctl(slabp)[objnr] = slabp->free;
2717 slabp->free = objnr;
2718 STATS_DEC_ACTIVE(cachep);
2719 slabp->inuse--;
Christoph Lametere498be72005-09-09 13:03:32 -07002720 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 check_slabp(cachep, slabp);
2722
2723 /* fixup slab chains */
2724 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002725 if (l3->free_objects > l3->free_limit) {
2726 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 slab_destroy(cachep, slabp);
2728 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002729 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 }
2731 } else {
2732 /* Unconditionally move a slab to the end of the
2733 * partial list on free - maximum time for the
2734 * other objects to be freed, too.
2735 */
Christoph Lametere498be72005-09-09 13:03:32 -07002736 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 }
2738 }
2739}
2740
2741static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2742{
2743 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002744 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002745 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
2747 batchcount = ac->batchcount;
2748#if DEBUG
2749 BUG_ON(!batchcount || batchcount > ac->avail);
2750#endif
2751 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002752 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002753 spin_lock(&l3->list_lock);
2754 if (l3->shared) {
2755 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002756 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 if (max) {
2758 if (batchcount > max)
2759 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002760 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002761 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 shared_array->avail += batchcount;
2763 goto free_done;
2764 }
2765 }
2766
Christoph Lameterff694162005-09-22 21:44:02 -07002767 free_block(cachep, ac->entry, batchcount, node);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002768 free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769#if STATS
2770 {
2771 int i = 0;
2772 struct list_head *p;
2773
Christoph Lametere498be72005-09-09 13:03:32 -07002774 p = l3->slabs_free.next;
2775 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 struct slab *slabp;
2777
2778 slabp = list_entry(p, struct slab, list);
2779 BUG_ON(slabp->inuse);
2780
2781 i++;
2782 p = p->next;
2783 }
2784 STATS_SET_FREEABLE(cachep, i);
2785 }
2786#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002787 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002789 memmove(ac->entry, &(ac->entry[batchcount]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002790 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791}
2792
2793/*
2794 * __cache_free
2795 * Release an obj back to its cache. If the obj has a constructed
2796 * state, it must be in this state _before_ it is released.
2797 *
2798 * Called with disabled ints.
2799 */
2800static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2801{
2802 struct array_cache *ac = ac_data(cachep);
2803
2804 check_irq_off();
2805 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2806
Christoph Lametere498be72005-09-09 13:03:32 -07002807 /* Make sure we are not freeing a object from another
2808 * node to the array cache on this cpu.
2809 */
2810#ifdef CONFIG_NUMA
2811 {
2812 struct slab *slabp;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002813 slabp = page_get_slab(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002814 if (unlikely(slabp->nodeid != numa_node_id())) {
2815 struct array_cache *alien = NULL;
2816 int nodeid = slabp->nodeid;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002817 struct kmem_list3 *l3 =
2818 cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07002819
2820 STATS_INC_NODEFREES(cachep);
2821 if (l3->alien && l3->alien[nodeid]) {
2822 alien = l3->alien[nodeid];
2823 spin_lock(&alien->lock);
2824 if (unlikely(alien->avail == alien->limit))
2825 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002826 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002827 alien->entry[alien->avail++] = objp;
2828 spin_unlock(&alien->lock);
2829 } else {
2830 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002831 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002832 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002833 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002834 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002835 }
2836 return;
2837 }
2838 }
2839#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 if (likely(ac->avail < ac->limit)) {
2841 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002842 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 return;
2844 } else {
2845 STATS_INC_FREEMISS(cachep);
2846 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002847 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 }
2849}
2850
2851/**
2852 * kmem_cache_alloc - Allocate an object
2853 * @cachep: The cache to allocate from.
2854 * @flags: See kmalloc().
2855 *
2856 * Allocate an object from this cache. The flags are only relevant
2857 * if the cache has no available objects.
2858 */
Al Virodd0fc662005-10-07 07:46:04 +01002859void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860{
2861 return __cache_alloc(cachep, flags);
2862}
2863EXPORT_SYMBOL(kmem_cache_alloc);
2864
2865/**
2866 * kmem_ptr_validate - check if an untrusted pointer might
2867 * be a slab entry.
2868 * @cachep: the cache we're checking against
2869 * @ptr: pointer to validate
2870 *
2871 * This verifies that the untrusted pointer looks sane:
2872 * it is _not_ a guarantee that the pointer is actually
2873 * part of the slab cache in question, but it at least
2874 * validates that the pointer can be dereferenced and
2875 * looks half-way sane.
2876 *
2877 * Currently only used for dentry validation.
2878 */
2879int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2880{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002881 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 unsigned long align_mask = BYTES_PER_WORD - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 unsigned long size = cachep->objsize;
2885 struct page *page;
2886
2887 if (unlikely(addr < min_addr))
2888 goto out;
2889 if (unlikely(addr > (unsigned long)high_memory - size))
2890 goto out;
2891 if (unlikely(addr & align_mask))
2892 goto out;
2893 if (unlikely(!kern_addr_valid(addr)))
2894 goto out;
2895 if (unlikely(!kern_addr_valid(addr + size - 1)))
2896 goto out;
2897 page = virt_to_page(ptr);
2898 if (unlikely(!PageSlab(page)))
2899 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002900 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 goto out;
2902 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002903 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 return 0;
2905}
2906
2907#ifdef CONFIG_NUMA
2908/**
2909 * kmem_cache_alloc_node - Allocate an object on the specified node
2910 * @cachep: The cache to allocate from.
2911 * @flags: See kmalloc().
2912 * @nodeid: node number of the target node.
2913 *
2914 * Identical to kmem_cache_alloc, except that this function is slow
2915 * and can sleep. And it will allocate memory on the given node, which
2916 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002917 * New and improved: it will now make sure that the object gets
2918 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 */
Al Virodd0fc662005-10-07 07:46:04 +01002920void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921{
Christoph Lametere498be72005-09-09 13:03:32 -07002922 unsigned long save_flags;
2923 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Christoph Lameterff694162005-09-22 21:44:02 -07002925 if (nodeid == -1)
Christoph Lametere498be72005-09-09 13:03:32 -07002926 return __cache_alloc(cachep, flags);
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002927
Christoph Lametere498be72005-09-09 13:03:32 -07002928 if (unlikely(!cachep->nodelists[nodeid])) {
2929 /* Fall back to __cache_alloc if we run into trouble */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930 printk(KERN_WARNING
2931 "slab: not allocating in inactive node %d for cache %s\n",
2932 nodeid, cachep->name);
2933 return __cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935
Christoph Lametere498be72005-09-09 13:03:32 -07002936 cache_alloc_debugcheck_before(cachep, flags);
2937 local_irq_save(save_flags);
Alok N Kataria5c382302005-09-27 21:45:46 -07002938 if (nodeid == numa_node_id())
2939 ptr = ____cache_alloc(cachep, flags);
2940 else
2941 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002942 local_irq_restore(save_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002943 ptr =
2944 cache_alloc_debugcheck_after(cachep, flags, ptr,
2945 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946
Christoph Lametere498be72005-09-09 13:03:32 -07002947 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948}
2949EXPORT_SYMBOL(kmem_cache_alloc_node);
2950
Al Virodd0fc662005-10-07 07:46:04 +01002951void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002952{
2953 kmem_cache_t *cachep;
2954
2955 cachep = kmem_find_general_cachep(size, flags);
2956 if (unlikely(cachep == NULL))
2957 return NULL;
2958 return kmem_cache_alloc_node(cachep, flags, node);
2959}
2960EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961#endif
2962
2963/**
2964 * kmalloc - allocate memory
2965 * @size: how many bytes of memory are required.
2966 * @flags: the type of memory to allocate.
2967 *
2968 * kmalloc is the normal method of allocating memory
2969 * in the kernel.
2970 *
2971 * The @flags argument may be one of:
2972 *
2973 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2974 *
2975 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2976 *
2977 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2978 *
2979 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2980 * must be suitable for DMA. This can mean different things on different
2981 * platforms. For example, on i386, it means that the memory must come
2982 * from the first 16MB.
2983 */
Al Virodd0fc662005-10-07 07:46:04 +01002984void *__kmalloc(size_t size, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985{
2986 kmem_cache_t *cachep;
2987
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002988 /* If you want to save a few bytes .text space: replace
2989 * __ with kmem_.
2990 * Then kmalloc uses the uninlined functions instead of the inline
2991 * functions.
2992 */
2993 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07002994 if (unlikely(cachep == NULL))
2995 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 return __cache_alloc(cachep, flags);
2997}
2998EXPORT_SYMBOL(__kmalloc);
2999
3000#ifdef CONFIG_SMP
3001/**
3002 * __alloc_percpu - allocate one copy of the object for every present
3003 * cpu in the system, zeroing them.
3004 * Objects should be dereferenced using the per_cpu_ptr macro only.
3005 *
3006 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003008void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009{
3010 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003011 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
3013 if (!pdata)
3014 return NULL;
3015
Christoph Lametere498be72005-09-09 13:03:32 -07003016 /*
3017 * Cannot use for_each_online_cpu since a cpu may come online
3018 * and we have no way of figuring out how to fix the array
3019 * that we have allocated then....
3020 */
3021 for_each_cpu(i) {
3022 int node = cpu_to_node(i);
3023
3024 if (node_online(node))
3025 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3026 else
3027 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028
3029 if (!pdata->ptrs[i])
3030 goto unwind_oom;
3031 memset(pdata->ptrs[i], 0, size);
3032 }
3033
3034 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003035 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003037 unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 while (--i >= 0) {
3039 if (!cpu_possible(i))
3040 continue;
3041 kfree(pdata->ptrs[i]);
3042 }
3043 kfree(pdata);
3044 return NULL;
3045}
3046EXPORT_SYMBOL(__alloc_percpu);
3047#endif
3048
3049/**
3050 * kmem_cache_free - Deallocate an object
3051 * @cachep: The cache the allocation was from.
3052 * @objp: The previously allocated object.
3053 *
3054 * Free an object which was previously allocated from this
3055 * cache.
3056 */
3057void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3058{
3059 unsigned long flags;
3060
3061 local_irq_save(flags);
3062 __cache_free(cachep, objp);
3063 local_irq_restore(flags);
3064}
3065EXPORT_SYMBOL(kmem_cache_free);
3066
3067/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 * kfree - free previously allocated memory
3069 * @objp: pointer returned by kmalloc.
3070 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003071 * If @objp is NULL, no operation is performed.
3072 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 * Don't free memory not originally allocated by kmalloc()
3074 * or you will run into trouble.
3075 */
3076void kfree(const void *objp)
3077{
3078 kmem_cache_t *c;
3079 unsigned long flags;
3080
3081 if (unlikely(!objp))
3082 return;
3083 local_irq_save(flags);
3084 kfree_debugcheck(objp);
Pekka Enberg065d41c2005-11-13 16:06:46 -08003085 c = page_get_cache(virt_to_page(objp));
David Woodhousea4fc7ab2006-01-11 14:41:26 +00003086 mutex_debug_check_no_locks_freed(objp, obj_reallen(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003087 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 local_irq_restore(flags);
3089}
3090EXPORT_SYMBOL(kfree);
3091
3092#ifdef CONFIG_SMP
3093/**
3094 * free_percpu - free previously allocated percpu memory
3095 * @objp: pointer returned by alloc_percpu.
3096 *
3097 * Don't free memory not originally allocated by alloc_percpu()
3098 * The complemented objp is to check for that.
3099 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003100void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101{
3102 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003103 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104
Christoph Lametere498be72005-09-09 13:03:32 -07003105 /*
3106 * We allocate for all cpus so we cannot use for online cpu here.
3107 */
3108 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003109 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 kfree(p);
3111}
3112EXPORT_SYMBOL(free_percpu);
3113#endif
3114
3115unsigned int kmem_cache_size(kmem_cache_t *cachep)
3116{
3117 return obj_reallen(cachep);
3118}
3119EXPORT_SYMBOL(kmem_cache_size);
3120
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003121const char *kmem_cache_name(kmem_cache_t *cachep)
3122{
3123 return cachep->name;
3124}
3125EXPORT_SYMBOL_GPL(kmem_cache_name);
3126
Christoph Lametere498be72005-09-09 13:03:32 -07003127/*
3128 * This initializes kmem_list3 for all nodes.
3129 */
3130static int alloc_kmemlist(kmem_cache_t *cachep)
3131{
3132 int node;
3133 struct kmem_list3 *l3;
3134 int err = 0;
3135
3136 for_each_online_node(node) {
3137 struct array_cache *nc = NULL, *new;
3138 struct array_cache **new_alien = NULL;
3139#ifdef CONFIG_NUMA
3140 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3141 goto fail;
3142#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003143 if (!(new = alloc_arraycache(node, (cachep->shared *
3144 cachep->batchcount),
3145 0xbaadf00d)))
Christoph Lametere498be72005-09-09 13:03:32 -07003146 goto fail;
3147 if ((l3 = cachep->nodelists[node])) {
3148
3149 spin_lock_irq(&l3->list_lock);
3150
3151 if ((nc = cachep->nodelists[node]->shared))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003152 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003153
3154 l3->shared = new;
3155 if (!cachep->nodelists[node]->alien) {
3156 l3->alien = new_alien;
3157 new_alien = NULL;
3158 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003159 l3->free_limit = (1 + nr_cpus_node(node)) *
3160 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003161 spin_unlock_irq(&l3->list_lock);
3162 kfree(nc);
3163 free_alien_cache(new_alien);
3164 continue;
3165 }
3166 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003167 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -07003168 goto fail;
3169
3170 kmem_list3_init(l3);
3171 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003172 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003173 l3->shared = new;
3174 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003175 l3->free_limit = (1 + nr_cpus_node(node)) *
3176 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003177 cachep->nodelists[node] = l3;
3178 }
3179 return err;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003180 fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003181 err = -ENOMEM;
3182 return err;
3183}
3184
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185struct ccupdate_struct {
3186 kmem_cache_t *cachep;
3187 struct array_cache *new[NR_CPUS];
3188};
3189
3190static void do_ccupdate_local(void *info)
3191{
3192 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3193 struct array_cache *old;
3194
3195 check_irq_off();
3196 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003197
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3199 new->new[smp_processor_id()] = old;
3200}
3201
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003203 int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204{
3205 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003206 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003208 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003209 for_each_online_cpu(i) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003210 new.new[i] =
3211 alloc_arraycache(cpu_to_node(i), limit, batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003212 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003213 for (i--; i >= 0; i--)
3214 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003215 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 }
3217 }
3218 new.cachep = cachep;
3219
3220 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003221
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222 check_irq_on();
3223 spin_lock_irq(&cachep->spinlock);
3224 cachep->batchcount = batchcount;
3225 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003226 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 spin_unlock_irq(&cachep->spinlock);
3228
Christoph Lametere498be72005-09-09 13:03:32 -07003229 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 struct array_cache *ccold = new.new[i];
3231 if (!ccold)
3232 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003233 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003234 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003235 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 kfree(ccold);
3237 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238
Christoph Lametere498be72005-09-09 13:03:32 -07003239 err = alloc_kmemlist(cachep);
3240 if (err) {
3241 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003242 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003243 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 return 0;
3246}
3247
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248static void enable_cpucache(kmem_cache_t *cachep)
3249{
3250 int err;
3251 int limit, shared;
3252
3253 /* The head array serves three purposes:
3254 * - create a LIFO ordering, i.e. return objects that are cache-warm
3255 * - reduce the number of spinlock operations.
3256 * - reduce the number of linked list operations on the slab and
3257 * bufctl chains: array operations are cheaper.
3258 * The numbers are guessed, we should auto-tune as described by
3259 * Bonwick.
3260 */
3261 if (cachep->objsize > 131072)
3262 limit = 1;
3263 else if (cachep->objsize > PAGE_SIZE)
3264 limit = 8;
3265 else if (cachep->objsize > 1024)
3266 limit = 24;
3267 else if (cachep->objsize > 256)
3268 limit = 54;
3269 else
3270 limit = 120;
3271
3272 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3273 * allocation behaviour: Most allocs on one cpu, most free operations
3274 * on another cpu. For these cases, an efficient object passing between
3275 * cpus is necessary. This is provided by a shared array. The array
3276 * replaces Bonwick's magazine layer.
3277 * On uniprocessor, it's functionally equivalent (but less efficient)
3278 * to a larger limit. Thus disabled by default.
3279 */
3280 shared = 0;
3281#ifdef CONFIG_SMP
3282 if (cachep->objsize <= PAGE_SIZE)
3283 shared = 8;
3284#endif
3285
3286#if DEBUG
3287 /* With debugging enabled, large batchcount lead to excessively
3288 * long periods with disabled local interrupts. Limit the
3289 * batchcount
3290 */
3291 if (limit > 32)
3292 limit = 32;
3293#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003294 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295 if (err)
3296 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003297 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298}
3299
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003300static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3301 int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302{
3303 int tofree;
3304
Christoph Lametere498be72005-09-09 13:03:32 -07003305 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 if (ac->touched && !force) {
3307 ac->touched = 0;
3308 } else if (ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003309 tofree = force ? ac->avail : (ac->limit + 4) / 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 if (tofree > ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003311 tofree = (ac->avail + 1) / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 }
Christoph Lameterff694162005-09-22 21:44:02 -07003313 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003315 memmove(ac->entry, &(ac->entry[tofree]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003316 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 }
3318}
3319
3320/**
3321 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003322 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 *
3324 * Called from workqueue/eventd every few seconds.
3325 * Purpose:
3326 * - clear the per-cpu caches for this CPU.
3327 * - return freeable pages to the main free memory pool.
3328 *
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003329 * If we cannot acquire the cache chain mutex then just give up - we'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330 * try again on the next iteration.
3331 */
3332static void cache_reap(void *unused)
3333{
3334 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003335 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003337 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003339 schedule_delayed_work(&__get_cpu_var(reap_work),
3340 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341 return;
3342 }
3343
3344 list_for_each(walk, &cache_chain) {
3345 kmem_cache_t *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003346 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 int tofree;
3348 struct slab *slabp;
3349
3350 searchp = list_entry(walk, kmem_cache_t, next);
3351
3352 if (searchp->flags & SLAB_NO_REAP)
3353 goto next;
3354
3355 check_irq_on();
3356
Christoph Lametere498be72005-09-09 13:03:32 -07003357 l3 = searchp->nodelists[numa_node_id()];
3358 if (l3->alien)
3359 drain_alien_cache(searchp, l3);
3360 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361
Christoph Lametere498be72005-09-09 13:03:32 -07003362 drain_array_locked(searchp, ac_data(searchp), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003363 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364
Christoph Lametere498be72005-09-09 13:03:32 -07003365 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 goto next_unlock;
3367
Christoph Lametere498be72005-09-09 13:03:32 -07003368 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369
Christoph Lametere498be72005-09-09 13:03:32 -07003370 if (l3->shared)
3371 drain_array_locked(searchp, l3->shared, 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003372 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
Christoph Lametere498be72005-09-09 13:03:32 -07003374 if (l3->free_touched) {
3375 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 goto next_unlock;
3377 }
3378
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003379 tofree =
3380 (l3->free_limit + 5 * searchp->num -
3381 1) / (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003383 p = l3->slabs_free.next;
3384 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 break;
3386
3387 slabp = list_entry(p, struct slab, list);
3388 BUG_ON(slabp->inuse);
3389 list_del(&slabp->list);
3390 STATS_INC_REAPED(searchp);
3391
3392 /* Safe to drop the lock. The slab is no longer
3393 * linked to the cache.
3394 * searchp cannot disappear, we hold
3395 * cache_chain_lock
3396 */
Christoph Lametere498be72005-09-09 13:03:32 -07003397 l3->free_objects -= searchp->num;
3398 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003400 spin_lock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003401 } while (--tofree > 0);
3402 next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003403 spin_unlock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003404 next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 cond_resched();
3406 }
3407 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003408 mutex_unlock(&cache_chain_mutex);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003409 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003411 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412}
3413
3414#ifdef CONFIG_PROC_FS
3415
Pekka Enberg85289f92006-01-08 01:00:36 -08003416static void print_slabinfo_header(struct seq_file *m)
3417{
3418 /*
3419 * Output format version, so at least we can change it
3420 * without _too_ many complaints.
3421 */
3422#if STATS
3423 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3424#else
3425 seq_puts(m, "slabinfo - version: 2.1\n");
3426#endif
3427 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3428 "<objperslab> <pagesperslab>");
3429 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3430 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3431#if STATS
3432 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3433 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3434 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3435#endif
3436 seq_putc(m, '\n');
3437}
3438
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439static void *s_start(struct seq_file *m, loff_t *pos)
3440{
3441 loff_t n = *pos;
3442 struct list_head *p;
3443
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003444 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003445 if (!n)
3446 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 p = cache_chain.next;
3448 while (n--) {
3449 p = p->next;
3450 if (p == &cache_chain)
3451 return NULL;
3452 }
3453 return list_entry(p, kmem_cache_t, next);
3454}
3455
3456static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3457{
3458 kmem_cache_t *cachep = p;
3459 ++*pos;
3460 return cachep->next.next == &cache_chain ? NULL
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 : list_entry(cachep->next.next, kmem_cache_t, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462}
3463
3464static void s_stop(struct seq_file *m, void *p)
3465{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003466 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467}
3468
3469static int s_show(struct seq_file *m, void *p)
3470{
3471 kmem_cache_t *cachep = p;
3472 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003473 struct slab *slabp;
3474 unsigned long active_objs;
3475 unsigned long num_objs;
3476 unsigned long active_slabs = 0;
3477 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003478 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003480 int node;
3481 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482
3483 check_irq_on();
3484 spin_lock_irq(&cachep->spinlock);
3485 active_objs = 0;
3486 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003487 for_each_online_node(node) {
3488 l3 = cachep->nodelists[node];
3489 if (!l3)
3490 continue;
3491
3492 spin_lock(&l3->list_lock);
3493
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003494 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003495 slabp = list_entry(q, struct slab, list);
3496 if (slabp->inuse != cachep->num && !error)
3497 error = "slabs_full accounting error";
3498 active_objs += cachep->num;
3499 active_slabs++;
3500 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003501 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003502 slabp = list_entry(q, struct slab, list);
3503 if (slabp->inuse == cachep->num && !error)
3504 error = "slabs_partial inuse accounting error";
3505 if (!slabp->inuse && !error)
3506 error = "slabs_partial/inuse accounting error";
3507 active_objs += slabp->inuse;
3508 active_slabs++;
3509 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003510 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003511 slabp = list_entry(q, struct slab, list);
3512 if (slabp->inuse && !error)
3513 error = "slabs_free/inuse accounting error";
3514 num_slabs++;
3515 }
3516 free_objects += l3->free_objects;
3517 shared_avail += l3->shared->avail;
3518
3519 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003521 num_slabs += active_slabs;
3522 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003523 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 error = "free_objects accounting error";
3525
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003526 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 if (error)
3528 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3529
3530 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003531 name, active_objs, num_objs, cachep->objsize,
3532 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003534 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003535 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003536 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003538 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 unsigned long high = cachep->high_mark;
3540 unsigned long allocs = cachep->num_allocations;
3541 unsigned long grown = cachep->grown;
3542 unsigned long reaped = cachep->reaped;
3543 unsigned long errors = cachep->errors;
3544 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003546 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547
Christoph Lametere498be72005-09-09 13:03:32 -07003548 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003549 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550 }
3551 /* cpu stats */
3552 {
3553 unsigned long allochit = atomic_read(&cachep->allochit);
3554 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3555 unsigned long freehit = atomic_read(&cachep->freehit);
3556 unsigned long freemiss = atomic_read(&cachep->freemiss);
3557
3558 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003559 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 }
3561#endif
3562 seq_putc(m, '\n');
3563 spin_unlock_irq(&cachep->spinlock);
3564 return 0;
3565}
3566
3567/*
3568 * slabinfo_op - iterator that generates /proc/slabinfo
3569 *
3570 * Output layout:
3571 * cache-name
3572 * num-active-objs
3573 * total-objs
3574 * object size
3575 * num-active-slabs
3576 * total-slabs
3577 * num-pages-per-slab
3578 * + further values on SMP and with statistics enabled
3579 */
3580
3581struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003582 .start = s_start,
3583 .next = s_next,
3584 .stop = s_stop,
3585 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586};
3587
3588#define MAX_SLABINFO_WRITE 128
3589/**
3590 * slabinfo_write - Tuning for the slab allocator
3591 * @file: unused
3592 * @buffer: user buffer
3593 * @count: data length
3594 * @ppos: unused
3595 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003596ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3597 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003599 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 int limit, batchcount, shared, res;
3601 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003602
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 if (count > MAX_SLABINFO_WRITE)
3604 return -EINVAL;
3605 if (copy_from_user(&kbuf, buffer, count))
3606 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003607 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608
3609 tmp = strchr(kbuf, ' ');
3610 if (!tmp)
3611 return -EINVAL;
3612 *tmp = '\0';
3613 tmp++;
3614 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3615 return -EINVAL;
3616
3617 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003618 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003620 list_for_each(p, &cache_chain) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3622
3623 if (!strcmp(cachep->name, kbuf)) {
3624 if (limit < 1 ||
3625 batchcount < 1 ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003626 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003627 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003629 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003630 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 }
3632 break;
3633 }
3634 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003635 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 if (res >= 0)
3637 res = count;
3638 return res;
3639}
3640#endif
3641
Manfred Spraul00e145b2005-09-03 15:55:07 -07003642/**
3643 * ksize - get the actual amount of memory allocated for a given object
3644 * @objp: Pointer to the object
3645 *
3646 * kmalloc may internally round up allocations and return more memory
3647 * than requested. ksize() can be used to determine the actual amount of
3648 * memory allocated. The caller may use this additional memory, even though
3649 * a smaller amount of memory was initially specified with the kmalloc call.
3650 * The caller must guarantee that objp points to a valid object previously
3651 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3652 * must not be freed during the duration of the call.
3653 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654unsigned int ksize(const void *objp)
3655{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003656 if (unlikely(objp == NULL))
3657 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658
Pekka Enberg065d41c2005-11-13 16:06:46 -08003659 return obj_reallen(page_get_cache(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660}