blob: 71370256a7eb11e8b6eeb0c88cbd0e23955ab163 [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.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * 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;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 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{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800319 extern void __bad_size(void);
320
Christoph Lametere498be72005-09-09 13:03:32 -0700321 if (__builtin_constant_p(size)) {
322 int i = 0;
323
324#define CACHE(x) \
325 if (size <=x) \
326 return i; \
327 else \
328 i++;
329#include "linux/kmalloc_sizes.h"
330#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800331 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700332 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800333 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700334 return 0;
335}
336
337#define INDEX_AC index_of(sizeof(struct arraycache_init))
338#define INDEX_L3 index_of(sizeof(struct kmem_list3))
339
Pekka Enberg5295a742006-02-01 03:05:48 -0800340static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700341{
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 spin_lock_init(&parent->list_lock);
348 parent->free_objects = 0;
349 parent->free_touched = 0;
350}
351
352#define MAKE_LIST(cachep, listp, slab, nodeid) \
353 do { \
354 INIT_LIST_HEAD(listp); \
355 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
356 } while (0)
357
358#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
359 do { \
360 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
363 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
365/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800366 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 *
368 * manages a cache.
369 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800370
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800371struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
375 unsigned int limit;
376 unsigned int shared;
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800377 unsigned int buffer_size;
Christoph Lametere498be72005-09-09 13:03:32 -0700378/* 2) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800379 struct kmem_list3 *nodelists[MAX_NUMNODES];
380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
382 spinlock_t spinlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
384/* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800386 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800389 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 size_t colour; /* cache colouring range */
392 unsigned int colour_off; /* colour offset */
393 unsigned int colour_next; /* cache colouring */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800394 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800395 unsigned int slab_size;
396 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800399 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800402 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
404/* 4) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 const char *name;
406 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408/* 5) statistics */
409#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800410 unsigned long num_active;
411 unsigned long num_allocations;
412 unsigned long high_mark;
413 unsigned long grown;
414 unsigned long reaped;
415 unsigned long errors;
416 unsigned long max_freeable;
417 unsigned long node_allocs;
418 unsigned long node_frees;
419 atomic_t allochit;
420 atomic_t allocmiss;
421 atomic_t freehit;
422 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423#endif
424#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800425 /*
426 * If debugging is enabled, then the allocator can add additional
427 * fields and/or padding to every object. buffer_size contains the total
428 * object size including these internal fields, the following two
429 * variables contain the offset to the user object and its size.
430 */
431 int obj_offset;
432 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434};
435
436#define CFLGS_OFF_SLAB (0x80000000UL)
437#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
438
439#define BATCHREFILL_LIMIT 16
440/* Optimization question: fewer reaps means less
441 * probability for unnessary cpucache drain/refill cycles.
442 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100443 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 * which could lock up otherwise freeable slabs.
445 */
446#define REAPTIMEOUT_CPUC (2*HZ)
447#define REAPTIMEOUT_LIST3 (4*HZ)
448
449#if STATS
450#define STATS_INC_ACTIVE(x) ((x)->num_active++)
451#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
452#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
453#define STATS_INC_GROWN(x) ((x)->grown++)
454#define STATS_INC_REAPED(x) ((x)->reaped++)
455#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
456 (x)->high_mark = (x)->num_active; \
457 } while (0)
458#define STATS_INC_ERR(x) ((x)->errors++)
459#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700460#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461#define STATS_SET_FREEABLE(x, i) \
462 do { if ((x)->max_freeable < i) \
463 (x)->max_freeable = i; \
464 } while (0)
465
466#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
467#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
468#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
469#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
470#else
471#define STATS_INC_ACTIVE(x) do { } while (0)
472#define STATS_DEC_ACTIVE(x) do { } while (0)
473#define STATS_INC_ALLOCED(x) do { } while (0)
474#define STATS_INC_GROWN(x) do { } while (0)
475#define STATS_INC_REAPED(x) do { } while (0)
476#define STATS_SET_HIGH(x) do { } while (0)
477#define STATS_INC_ERR(x) do { } while (0)
478#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700479#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480#define STATS_SET_FREEABLE(x, i) \
481 do { } while (0)
482
483#define STATS_INC_ALLOCHIT(x) do { } while (0)
484#define STATS_INC_ALLOCMISS(x) do { } while (0)
485#define STATS_INC_FREEHIT(x) do { } while (0)
486#define STATS_INC_FREEMISS(x) do { } while (0)
487#endif
488
489#if DEBUG
490/* Magic nums for obj red zoning.
491 * Placed in the first word before and the first word after an obj.
492 */
493#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
494#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
495
496/* ...and for poisoning */
497#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
498#define POISON_FREE 0x6b /* for use-after-free poisoning */
499#define POISON_END 0xa5 /* end-byte of poisoning */
500
501/* memory layout of objects:
502 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800503 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 * the end of an object is aligned with the end of the real
505 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800506 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800508 * cachep->obj_offset: The real object.
509 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
510 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800512static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800514 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515}
516
Pekka Enberg343e0d72006-02-01 03:05:50 -0800517static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800519 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520}
521
Pekka Enberg343e0d72006-02-01 03:05:50 -0800522static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523{
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800532 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800533 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800534 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
Pekka Enberg343e0d72006-02-01 03:05:50 -0800537static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
543#else
544
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800545#define obj_offset(x) 0
546#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
548#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
550
551#endif
552
553/*
554 * Maximum size of an obj (in 2^order pages)
555 * and absolute limit for the gfp order.
556 */
557#if defined(CONFIG_LARGE_ALLOCS)
558#define MAX_OBJ_ORDER 13 /* up to 32Mb */
559#define MAX_GFP_ORDER 13 /* up to 32Mb */
560#elif defined(CONFIG_MMU)
561#define MAX_OBJ_ORDER 5 /* 32 pages */
562#define MAX_GFP_ORDER 5 /* 32 pages */
563#else
564#define MAX_OBJ_ORDER 8 /* up to 1Mb */
565#define MAX_GFP_ORDER 8 /* up to 1Mb */
566#endif
567
568/*
569 * Do not go above this order unless 0 objects fit into the slab.
570 */
571#define BREAK_GFP_ORDER_HI 1
572#define BREAK_GFP_ORDER_LO 0
573static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
574
Pekka Enberg065d41c2005-11-13 16:06:46 -0800575/* Functions for storing/retrieving the cachep and or slab from the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 * global 'mem_map'. These are used to find the slab an obj belongs to.
577 * With kfree(), these are used to find the cache which an obj belongs to.
578 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800579static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
580{
581 page->lru.next = (struct list_head *)cache;
582}
583
584static inline struct kmem_cache *page_get_cache(struct page *page)
585{
586 return (struct kmem_cache *)page->lru.next;
587}
588
589static inline void page_set_slab(struct page *page, struct slab *slab)
590{
591 page->lru.prev = (struct list_head *)slab;
592}
593
594static inline struct slab *page_get_slab(struct page *page)
595{
596 return (struct slab *)page->lru.prev;
597}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800599static inline struct kmem_cache *virt_to_cache(const void *obj)
600{
601 struct page *page = virt_to_page(obj);
602 return page_get_cache(page);
603}
604
605static inline struct slab *virt_to_slab(const void *obj)
606{
607 struct page *page = virt_to_page(obj);
608 return page_get_slab(page);
609}
610
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611/* These are the default caches for kmalloc. Custom caches can have other sizes. */
612struct cache_sizes malloc_sizes[] = {
613#define CACHE(x) { .cs_size = (x) },
614#include <linux/kmalloc_sizes.h>
615 CACHE(ULONG_MAX)
616#undef CACHE
617};
618EXPORT_SYMBOL(malloc_sizes);
619
620/* Must match cache_sizes above. Out of line to keep cache footprint low. */
621struct cache_names {
622 char *name;
623 char *name_dma;
624};
625
626static struct cache_names __initdata cache_names[] = {
627#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
628#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800629 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630#undef CACHE
631};
632
633static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800634 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800636 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
638/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800639static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800640 .batchcount = 1,
641 .limit = BOOT_CPUCACHE_ENTRIES,
642 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800643 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800644 .flags = SLAB_NO_REAP,
645 .spinlock = SPIN_LOCK_UNLOCKED,
646 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800648 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649#endif
650};
651
652/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800653static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654static struct list_head cache_chain;
655
656/*
657 * vm_enough_memory() looks at this to determine how many
658 * slab-allocated pages are possibly freeable under pressure
659 *
660 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
661 */
662atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
664/*
665 * chicken and egg problem: delay the per-cpu array allocation
666 * until the general caches are up.
667 */
668static enum {
669 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700670 PARTIAL_AC,
671 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 FULL
673} g_cpucache_up;
674
675static DEFINE_PER_CPU(struct work_struct, reap_work);
676
Pekka Enberg343e0d72006-02-01 03:05:50 -0800677static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node);
678static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800679static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800680static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
Pekka Enberg343e0d72006-02-01 03:05:50 -0800682static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683{
684 return cachep->array[smp_processor_id()];
685}
686
Pekka Enberg343e0d72006-02-01 03:05:50 -0800687static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
689 struct cache_sizes *csizep = malloc_sizes;
690
691#if DEBUG
692 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800693 * kmem_cache_create(), or __kmalloc(), before
694 * the generic caches are initialized.
695 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700696 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697#endif
698 while (size > csizep->cs_size)
699 csizep++;
700
701 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700702 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 * has cs_{dma,}cachep==NULL. Thus no special case
704 * for large kmalloc calls required.
705 */
706 if (unlikely(gfpflags & GFP_DMA))
707 return csizep->cs_dmacachep;
708 return csizep->cs_cachep;
709}
710
Pekka Enberg343e0d72006-02-01 03:05:50 -0800711struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700712{
713 return __find_general_cachep(size, gfpflags);
714}
715EXPORT_SYMBOL(kmem_find_general_cachep);
716
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800717static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800719 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
720}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800722/* Calculate the number of objects and left-over bytes for a given
723 buffer size. */
724static void cache_estimate(unsigned long gfporder, size_t buffer_size,
725 size_t align, int flags, size_t *left_over,
726 unsigned int *num)
727{
728 int nr_objs;
729 size_t mgmt_size;
730 size_t slab_size = PAGE_SIZE << gfporder;
731
732 /*
733 * The slab management structure can be either off the slab or
734 * on it. For the latter case, the memory allocated for a
735 * slab is used for:
736 *
737 * - The struct slab
738 * - One kmem_bufctl_t for each object
739 * - Padding to respect alignment of @align
740 * - @buffer_size bytes for each object
741 *
742 * If the slab management structure is off the slab, then the
743 * alignment will already be calculated into the size. Because
744 * the slabs are all pages aligned, the objects will be at the
745 * correct alignment when allocated.
746 */
747 if (flags & CFLGS_OFF_SLAB) {
748 mgmt_size = 0;
749 nr_objs = slab_size / buffer_size;
750
751 if (nr_objs > SLAB_LIMIT)
752 nr_objs = SLAB_LIMIT;
753 } else {
754 /*
755 * Ignore padding for the initial guess. The padding
756 * is at most @align-1 bytes, and @buffer_size is at
757 * least @align. In the worst case, this result will
758 * be one greater than the number of objects that fit
759 * into the memory allocation when taking the padding
760 * into account.
761 */
762 nr_objs = (slab_size - sizeof(struct slab)) /
763 (buffer_size + sizeof(kmem_bufctl_t));
764
765 /*
766 * This calculated number will be either the right
767 * amount, or one greater than what we want.
768 */
769 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
770 > slab_size)
771 nr_objs--;
772
773 if (nr_objs > SLAB_LIMIT)
774 nr_objs = SLAB_LIMIT;
775
776 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800778 *num = nr_objs;
779 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780}
781
782#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
783
Pekka Enberg343e0d72006-02-01 03:05:50 -0800784static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
786 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800787 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 dump_stack();
789}
790
791/*
792 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
793 * via the workqueue/eventd.
794 * Add the CPU number into the expiration time to minimize the possibility of
795 * the CPUs getting into lockstep and contending for the global cache chain
796 * lock.
797 */
798static void __devinit start_cpu_timer(int cpu)
799{
800 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
801
802 /*
803 * When this gets called from do_initcalls via cpucache_init(),
804 * init_workqueues() has already run, so keventd will be setup
805 * at that time.
806 */
807 if (keventd_up() && reap_work->func == NULL) {
808 INIT_WORK(reap_work, cache_reap, NULL);
809 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
810 }
811}
812
Christoph Lametere498be72005-09-09 13:03:32 -0700813static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800814 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800816 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 struct array_cache *nc = NULL;
818
Christoph Lametere498be72005-09-09 13:03:32 -0700819 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 if (nc) {
821 nc->avail = 0;
822 nc->limit = entries;
823 nc->batchcount = batchcount;
824 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700825 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 }
827 return nc;
828}
829
Christoph Lametere498be72005-09-09 13:03:32 -0700830#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800831static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800832
Pekka Enberg5295a742006-02-01 03:05:48 -0800833static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700834{
835 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800836 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700837 int i;
838
839 if (limit > 1)
840 limit = 12;
841 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
842 if (ac_ptr) {
843 for_each_node(i) {
844 if (i == node || !node_online(i)) {
845 ac_ptr[i] = NULL;
846 continue;
847 }
848 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
849 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800850 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700851 kfree(ac_ptr[i]);
852 kfree(ac_ptr);
853 return NULL;
854 }
855 }
856 }
857 return ac_ptr;
858}
859
Pekka Enberg5295a742006-02-01 03:05:48 -0800860static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700861{
862 int i;
863
864 if (!ac_ptr)
865 return;
866
867 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800868 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700869
870 kfree(ac_ptr);
871}
872
Pekka Enberg343e0d72006-02-01 03:05:50 -0800873static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800874 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700875{
876 struct kmem_list3 *rl3 = cachep->nodelists[node];
877
878 if (ac->avail) {
879 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700880 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700881 ac->avail = 0;
882 spin_unlock(&rl3->list_lock);
883 }
884}
885
Pekka Enberg343e0d72006-02-01 03:05:50 -0800886static void drain_alien_cache(struct kmem_cache *cachep, struct kmem_list3 *l3)
Christoph Lametere498be72005-09-09 13:03:32 -0700887{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800888 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700889 struct array_cache *ac;
890 unsigned long flags;
891
892 for_each_online_node(i) {
893 ac = l3->alien[i];
894 if (ac) {
895 spin_lock_irqsave(&ac->lock, flags);
896 __drain_alien_cache(cachep, ac, i);
897 spin_unlock_irqrestore(&ac->lock, flags);
898 }
899 }
900}
901#else
902#define alloc_alien_cache(node, limit) do { } while (0)
903#define free_alien_cache(ac_ptr) do { } while (0)
904#define drain_alien_cache(cachep, l3) do { } while (0)
905#endif
906
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800908 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909{
910 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800911 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700912 struct kmem_list3 *l3 = NULL;
913 int node = cpu_to_node(cpu);
914 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
916 switch (action) {
917 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800918 mutex_lock(&cache_chain_mutex);
Christoph Lametere498be72005-09-09 13:03:32 -0700919 /* we need to do this right in the beginning since
920 * alloc_arraycache's are going to use this list.
921 * kmalloc_node allows us to add the slab to the right
922 * kmem_list3 and not this cpu's kmem_list3
923 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924
Christoph Lametere498be72005-09-09 13:03:32 -0700925 list_for_each_entry(cachep, &cache_chain, next) {
926 /* setup the size64 kmemlist for cpu before we can
927 * begin anything. Make sure some other cpu on this
928 * node has not already allocated this
929 */
930 if (!cachep->nodelists[node]) {
931 if (!(l3 = kmalloc_node(memsize,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800932 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -0700933 goto bad;
934 kmem_list3_init(l3);
935 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800936 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -0700937
938 cachep->nodelists[node] = l3;
939 }
940
941 spin_lock_irq(&cachep->nodelists[node]->list_lock);
942 cachep->nodelists[node]->free_limit =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800943 (1 + nr_cpus_node(node)) *
944 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -0700945 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
946 }
947
948 /* Now we can go ahead with allocating the shared array's
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800949 & array cache's */
Christoph Lametere498be72005-09-09 13:03:32 -0700950 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -0800951 struct array_cache *nc;
952
Christoph Lametere498be72005-09-09 13:03:32 -0700953 nc = alloc_arraycache(node, cachep->limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800954 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 if (!nc)
956 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
Christoph Lametere498be72005-09-09 13:03:32 -0700959 l3 = cachep->nodelists[node];
960 BUG_ON(!l3);
961 if (!l3->shared) {
962 if (!(nc = alloc_arraycache(node,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800963 cachep->shared *
964 cachep->batchcount,
965 0xbaadf00d)))
966 goto bad;
Christoph Lametere498be72005-09-09 13:03:32 -0700967
968 /* we are serialised from CPU_DEAD or
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800969 CPU_UP_CANCELLED by the cpucontrol lock */
Christoph Lametere498be72005-09-09 13:03:32 -0700970 l3->shared = nc;
971 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800973 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 break;
975 case CPU_ONLINE:
976 start_cpu_timer(cpu);
977 break;
978#ifdef CONFIG_HOTPLUG_CPU
979 case CPU_DEAD:
980 /* fall thru */
981 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800982 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983
984 list_for_each_entry(cachep, &cache_chain, next) {
985 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700986 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
Christoph Lametere498be72005-09-09 13:03:32 -0700988 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 spin_lock_irq(&cachep->spinlock);
990 /* cpu is dead; no one can alloc from it. */
991 nc = cachep->array[cpu];
992 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700993 l3 = cachep->nodelists[node];
994
995 if (!l3)
996 goto unlock_cache;
997
998 spin_lock(&l3->list_lock);
999
1000 /* Free limit for this kmem_list3 */
1001 l3->free_limit -= cachep->batchcount;
1002 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001003 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001004
1005 if (!cpus_empty(mask)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001006 spin_unlock(&l3->list_lock);
1007 goto unlock_cache;
1008 }
Christoph Lametere498be72005-09-09 13:03:32 -07001009
1010 if (l3->shared) {
1011 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001012 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001013 kfree(l3->shared);
1014 l3->shared = NULL;
1015 }
1016 if (l3->alien) {
1017 drain_alien_cache(cachep, l3);
1018 free_alien_cache(l3->alien);
1019 l3->alien = NULL;
1020 }
1021
1022 /* free slabs belonging to this node */
1023 if (__node_shrink(cachep, node)) {
1024 cachep->nodelists[node] = NULL;
1025 spin_unlock(&l3->list_lock);
1026 kfree(l3);
1027 } else {
1028 spin_unlock(&l3->list_lock);
1029 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001030 unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 spin_unlock_irq(&cachep->spinlock);
1032 kfree(nc);
1033 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001034 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 break;
1036#endif
1037 }
1038 return NOTIFY_OK;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001039 bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001040 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 return NOTIFY_BAD;
1042}
1043
1044static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1045
Christoph Lametere498be72005-09-09 13:03:32 -07001046/*
1047 * swap the static kmem_list3 with kmalloced memory
1048 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001049static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001050{
1051 struct kmem_list3 *ptr;
1052
1053 BUG_ON(cachep->nodelists[nodeid] != list);
1054 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1055 BUG_ON(!ptr);
1056
1057 local_irq_disable();
1058 memcpy(ptr, list, sizeof(struct kmem_list3));
1059 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1060 cachep->nodelists[nodeid] = ptr;
1061 local_irq_enable();
1062}
1063
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064/* Initialisation.
1065 * Called after the gfp() functions have been enabled, and before smp_init().
1066 */
1067void __init kmem_cache_init(void)
1068{
1069 size_t left_over;
1070 struct cache_sizes *sizes;
1071 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001072 int i;
1073
1074 for (i = 0; i < NUM_INIT_LISTS; i++) {
1075 kmem_list3_init(&initkmem_list3[i]);
1076 if (i < MAX_NUMNODES)
1077 cache_cache.nodelists[i] = NULL;
1078 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001079
1080 /*
1081 * Fragmentation resistance on low memory - only use bigger
1082 * page orders on machines with more than 32MB of memory.
1083 */
1084 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1085 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1086
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 /* Bootstrap is tricky, because several objects are allocated
1088 * from caches that do not exist yet:
Pekka Enberg343e0d72006-02-01 03:05:50 -08001089 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 * structures of all caches, except cache_cache itself: cache_cache
1091 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001092 * Initially an __init data area is used for the head array and the
1093 * kmem_list3 structures, it's replaced with a kmalloc allocated
1094 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001096 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001097 * An __init data area is used for the head array.
1098 * 3) Create the remaining kmalloc caches, with minimally sized
1099 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 * 4) Replace the __init data head arrays for cache_cache and the first
1101 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001102 * 5) Replace the __init data for kmem_list3 for cache_cache and
1103 * the other cache's with kmalloc allocated memory.
1104 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 */
1106
1107 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 INIT_LIST_HEAD(&cache_chain);
1109 list_add(&cache_cache.next, &cache_chain);
1110 cache_cache.colour_off = cache_line_size();
1111 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001112 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001114 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001116 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001117 &left_over, &cache_cache.num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 if (!cache_cache.num)
1119 BUG();
1120
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001121 cache_cache.colour = left_over / cache_cache.colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 cache_cache.colour_next = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001123 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1124 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125
1126 /* 2+3) create the kmalloc caches */
1127 sizes = malloc_sizes;
1128 names = cache_names;
1129
Christoph Lametere498be72005-09-09 13:03:32 -07001130 /* Initialize the caches that provide memory for the array cache
1131 * and the kmem_list3 structures first.
1132 * Without this, further allocations will bug
1133 */
1134
1135 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001136 sizes[INDEX_AC].cs_size,
1137 ARCH_KMALLOC_MINALIGN,
1138 (ARCH_KMALLOC_FLAGS |
1139 SLAB_PANIC), NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001140
1141 if (INDEX_AC != INDEX_L3)
1142 sizes[INDEX_L3].cs_cachep =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001143 kmem_cache_create(names[INDEX_L3].name,
1144 sizes[INDEX_L3].cs_size,
1145 ARCH_KMALLOC_MINALIGN,
1146 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1147 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001148
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001150 /*
1151 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 * This should be particularly beneficial on SMP boxes, as it
1153 * eliminates "false sharing".
1154 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001155 * allow tighter packing of the smaller caches.
1156 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001157 if (!sizes->cs_cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07001158 sizes->cs_cachep = kmem_cache_create(names->name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001159 sizes->cs_size,
1160 ARCH_KMALLOC_MINALIGN,
1161 (ARCH_KMALLOC_FLAGS
1162 | SLAB_PANIC),
1163 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
1165 /* Inc off-slab bufctl limit until the ceiling is hit. */
1166 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001167 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 offslab_limit /= sizeof(kmem_bufctl_t);
1169 }
1170
1171 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001172 sizes->cs_size,
1173 ARCH_KMALLOC_MINALIGN,
1174 (ARCH_KMALLOC_FLAGS |
1175 SLAB_CACHE_DMA |
1176 SLAB_PANIC), NULL,
1177 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
1179 sizes++;
1180 names++;
1181 }
1182 /* 4) Replace the bootstrap head arrays */
1183 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001184 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001185
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001187
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001189 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1190 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001191 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 cache_cache.array[smp_processor_id()] = ptr;
1193 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001194
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001196
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001198 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001199 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001200 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001201 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001202 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001203 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 local_irq_enable();
1205 }
Christoph Lametere498be72005-09-09 13:03:32 -07001206 /* 5) Replace the bootstrap kmem_list3's */
1207 {
1208 int node;
1209 /* Replace the static kmem_list3 structures for the boot cpu */
1210 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001211 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212
Christoph Lametere498be72005-09-09 13:03:32 -07001213 for_each_online_node(node) {
1214 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001215 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001216
1217 if (INDEX_AC != INDEX_L3) {
1218 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001219 &initkmem_list3[SIZE_L3 + node],
1220 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001221 }
1222 }
1223 }
1224
1225 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001227 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001228 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001230 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001231 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 }
1233
1234 /* Done! */
1235 g_cpucache_up = FULL;
1236
1237 /* Register a cpu startup notifier callback
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001238 * that initializes cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 */
1240 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241
1242 /* The reap timers are started later, with a module init call:
1243 * That part of the kernel is not yet operational.
1244 */
1245}
1246
1247static int __init cpucache_init(void)
1248{
1249 int cpu;
1250
1251 /*
1252 * Register the timers that return unneeded
1253 * pages to gfp.
1254 */
Christoph Lametere498be72005-09-09 13:03:32 -07001255 for_each_online_cpu(cpu)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001256 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257
1258 return 0;
1259}
1260
1261__initcall(cpucache_init);
1262
1263/*
1264 * Interface to system's page allocator. No need to hold the cache-lock.
1265 *
1266 * If we requested dmaable memory, we will get it. Even if we
1267 * did not request dmaable memory, we might get it, but that
1268 * would be relatively rare and ignorable.
1269 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001270static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271{
1272 struct page *page;
1273 void *addr;
1274 int i;
1275
1276 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001277 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 if (!page)
1279 return NULL;
1280 addr = page_address(page);
1281
1282 i = (1 << cachep->gfporder);
1283 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1284 atomic_add(i, &slab_reclaim_pages);
1285 add_page_state(nr_slab, i);
1286 while (i--) {
1287 SetPageSlab(page);
1288 page++;
1289 }
1290 return addr;
1291}
1292
1293/*
1294 * Interface to system's page release.
1295 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001296static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001298 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 struct page *page = virt_to_page(addr);
1300 const unsigned long nr_freed = i;
1301
1302 while (i--) {
1303 if (!TestClearPageSlab(page))
1304 BUG();
1305 page++;
1306 }
1307 sub_page_state(nr_slab, nr_freed);
1308 if (current->reclaim_state)
1309 current->reclaim_state->reclaimed_slab += nr_freed;
1310 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001311 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1312 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313}
1314
1315static void kmem_rcu_free(struct rcu_head *head)
1316{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001317 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001318 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
1320 kmem_freepages(cachep, slab_rcu->addr);
1321 if (OFF_SLAB(cachep))
1322 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1323}
1324
1325#if DEBUG
1326
1327#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001328static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001329 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001331 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001333 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001335 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 return;
1337
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001338 *addr++ = 0x12345678;
1339 *addr++ = caller;
1340 *addr++ = smp_processor_id();
1341 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 {
1343 unsigned long *sptr = &caller;
1344 unsigned long svalue;
1345
1346 while (!kstack_end(sptr)) {
1347 svalue = *sptr++;
1348 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001349 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 size -= sizeof(unsigned long);
1351 if (size <= sizeof(unsigned long))
1352 break;
1353 }
1354 }
1355
1356 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001357 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358}
1359#endif
1360
Pekka Enberg343e0d72006-02-01 03:05:50 -08001361static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001363 int size = obj_size(cachep);
1364 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
1366 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001367 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368}
1369
1370static void dump_line(char *data, int offset, int limit)
1371{
1372 int i;
1373 printk(KERN_ERR "%03x:", offset);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001374 for (i = 0; i < limit; i++) {
1375 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 }
1377 printk("\n");
1378}
1379#endif
1380
1381#if DEBUG
1382
Pekka Enberg343e0d72006-02-01 03:05:50 -08001383static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
1385 int i, size;
1386 char *realobj;
1387
1388 if (cachep->flags & SLAB_RED_ZONE) {
1389 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001390 *dbg_redzone1(cachep, objp),
1391 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 }
1393
1394 if (cachep->flags & SLAB_STORE_USER) {
1395 printk(KERN_ERR "Last user: [<%p>]",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001396 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 print_symbol("(%s)",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001398 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 printk("\n");
1400 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001401 realobj = (char *)objp + obj_offset(cachep);
1402 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001403 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 int limit;
1405 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001406 if (i + limit > size)
1407 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 dump_line(realobj, i, limit);
1409 }
1410}
1411
Pekka Enberg343e0d72006-02-01 03:05:50 -08001412static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413{
1414 char *realobj;
1415 int size, i;
1416 int lines = 0;
1417
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001418 realobj = (char *)objp + obj_offset(cachep);
1419 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001421 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001423 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 exp = POISON_END;
1425 if (realobj[i] != exp) {
1426 int limit;
1427 /* Mismatch ! */
1428 /* Print header */
1429 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001430 printk(KERN_ERR
1431 "Slab corruption: start=%p, len=%d\n",
1432 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 print_objinfo(cachep, objp, 0);
1434 }
1435 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001436 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001438 if (i + limit > size)
1439 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 dump_line(realobj, i, limit);
1441 i += 16;
1442 lines++;
1443 /* Limit to 5 lines */
1444 if (lines > 5)
1445 break;
1446 }
1447 }
1448 if (lines != 0) {
1449 /* Print some data about the neighboring objects, if they
1450 * exist:
1451 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001452 struct slab *slabp = virt_to_slab(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 int objnr;
1454
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001455 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 if (objnr) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001457 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1458 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001460 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 print_objinfo(cachep, objp, 2);
1462 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001463 if (objnr + 1 < cachep->num) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001464 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1465 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001467 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 print_objinfo(cachep, objp, 2);
1469 }
1470 }
1471}
1472#endif
1473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001475/**
1476 * slab_destroy_objs - call the registered destructor for each object in
1477 * a slab that is to be destroyed.
1478 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001479static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001480{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 int i;
1482 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001483 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
1485 if (cachep->flags & SLAB_POISON) {
1486#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001487 if ((cachep->buffer_size % PAGE_SIZE) == 0
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001488 && OFF_SLAB(cachep))
1489 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001490 cachep->buffer_size / PAGE_SIZE,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001491 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 else
1493 check_poison_obj(cachep, objp);
1494#else
1495 check_poison_obj(cachep, objp);
1496#endif
1497 }
1498 if (cachep->flags & SLAB_RED_ZONE) {
1499 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1500 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001501 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1503 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001504 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 }
1506 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001507 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001509}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001511static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001512{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 if (cachep->dtor) {
1514 int i;
1515 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001516 void *objp = slabp->s_mem + cachep->buffer_size * i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001517 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 }
1519 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001520}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521#endif
1522
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001523/**
1524 * Destroy all the objs in a slab, and release the mem back to the system.
1525 * Before calling the slab must have been unlinked from the cache.
1526 * The cache-lock is not held/needed.
1527 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001528static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001529{
1530 void *addr = slabp->s_mem - slabp->colouroff;
1531
1532 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1534 struct slab_rcu *slab_rcu;
1535
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 slab_rcu->cachep = cachep;
1538 slab_rcu->addr = addr;
1539 call_rcu(&slab_rcu->head, kmem_rcu_free);
1540 } else {
1541 kmem_freepages(cachep, addr);
1542 if (OFF_SLAB(cachep))
1543 kmem_cache_free(cachep->slabp_cache, slabp);
1544 }
1545}
1546
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001547/* For setting up all the kmem_list3s for cache whose buffer_size is same
Christoph Lametere498be72005-09-09 13:03:32 -07001548 as size of kmem_list3. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001549static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001550{
1551 int node;
1552
1553 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001554 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001555 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001556 REAPTIMEOUT_LIST3 +
1557 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001558 }
1559}
1560
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001562 * calculate_slab_order - calculate size (page order) of slabs
1563 * @cachep: pointer to the cache that is being created
1564 * @size: size of objects to be created in this cache.
1565 * @align: required alignment for the objects.
1566 * @flags: slab allocation flags
1567 *
1568 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001569 *
1570 * This could be made much more intelligent. For now, try to avoid using
1571 * high order pages for slabs. When the gfp() functions are more friendly
1572 * towards high-order requests, this should be changed.
1573 */
Randy Dunlapee13d782006-02-01 03:05:53 -08001574static inline size_t calculate_slab_order(struct kmem_cache *cachep,
1575 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001576{
1577 size_t left_over = 0;
1578
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001579 for (;; cachep->gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001580 unsigned int num;
1581 size_t remainder;
1582
1583 if (cachep->gfporder > MAX_GFP_ORDER) {
1584 cachep->num = 0;
1585 break;
1586 }
1587
1588 cache_estimate(cachep->gfporder, size, align, flags,
1589 &remainder, &num);
1590 if (!num)
1591 continue;
1592 /* More than offslab_limit objects will cause problems */
1593 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1594 break;
1595
1596 cachep->num = num;
1597 left_over = remainder;
1598
1599 /*
1600 * Large number of objects is good, but very large slabs are
1601 * currently bad for the gfp()s.
1602 */
1603 if (cachep->gfporder >= slab_break_gfp_order)
1604 break;
1605
1606 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1607 /* Acceptable internal fragmentation */
1608 break;
1609 }
1610 return left_over;
1611}
1612
1613/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 * kmem_cache_create - Create a cache.
1615 * @name: A string which is used in /proc/slabinfo to identify this cache.
1616 * @size: The size of objects to be created in this cache.
1617 * @align: The required alignment for the objects.
1618 * @flags: SLAB flags
1619 * @ctor: A constructor for the objects.
1620 * @dtor: A destructor for the objects.
1621 *
1622 * Returns a ptr to the cache on success, NULL on failure.
1623 * Cannot be called within a int, but can be interrupted.
1624 * The @ctor is run when new pages are allocated by the cache
1625 * and the @dtor is run before the pages are handed back.
1626 *
1627 * @name must be valid until the cache is destroyed. This implies that
1628 * the module calling this has to destroy the cache before getting
1629 * unloaded.
1630 *
1631 * The flags are
1632 *
1633 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1634 * to catch references to uninitialised memory.
1635 *
1636 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1637 * for buffer overruns.
1638 *
1639 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1640 * memory pressure.
1641 *
1642 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1643 * cacheline. This can be beneficial if you're counting cycles as closely
1644 * as davem.
1645 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001646struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647kmem_cache_create (const char *name, size_t size, size_t align,
Pekka Enberg343e0d72006-02-01 03:05:50 -08001648 unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long),
1649 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650{
1651 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001652 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001653 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654
1655 /*
1656 * Sanity checks... these are all serious usage bugs.
1657 */
1658 if ((!name) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001659 in_interrupt() ||
1660 (size < BYTES_PER_WORD) ||
1661 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1662 printk(KERN_ERR "%s: Early error in slab %s\n",
1663 __FUNCTION__, name);
1664 BUG();
1665 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001667 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001668
1669 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001670 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001671 mm_segment_t old_fs = get_fs();
1672 char tmp;
1673 int res;
1674
1675 /*
1676 * This happens when the module gets unloaded and doesn't
1677 * destroy its slab cache and no-one else reuses the vmalloc
1678 * area of the module. Print a warning.
1679 */
1680 set_fs(KERNEL_DS);
1681 res = __get_user(tmp, pc->name);
1682 set_fs(old_fs);
1683 if (res) {
1684 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001685 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001686 continue;
1687 }
1688
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001689 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001690 printk("kmem_cache_create: duplicate cache %s\n", name);
1691 dump_stack();
1692 goto oops;
1693 }
1694 }
1695
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696#if DEBUG
1697 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1698 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1699 /* No constructor, but inital state check requested */
1700 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001701 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 flags &= ~SLAB_DEBUG_INITIAL;
1703 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704#if FORCED_DEBUG
1705 /*
1706 * Enable redzoning and last user accounting, except for caches with
1707 * large objects, if the increased size would increase the object size
1708 * above the next power of two: caches with object sizes just above a
1709 * power of two have a significant amount of internal fragmentation.
1710 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001711 if ((size < 4096
1712 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1713 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 if (!(flags & SLAB_DESTROY_BY_RCU))
1715 flags |= SLAB_POISON;
1716#endif
1717 if (flags & SLAB_DESTROY_BY_RCU)
1718 BUG_ON(flags & SLAB_POISON);
1719#endif
1720 if (flags & SLAB_DESTROY_BY_RCU)
1721 BUG_ON(dtor);
1722
1723 /*
1724 * Always checks flags, a caller might be expecting debug
1725 * support which isn't available.
1726 */
1727 if (flags & ~CREATE_MASK)
1728 BUG();
1729
1730 /* Check that size is in terms of words. This is needed to avoid
1731 * unaligned accesses for some archs when redzoning is used, and makes
1732 * sure any on-slab bufctl's are also correctly aligned.
1733 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001734 if (size & (BYTES_PER_WORD - 1)) {
1735 size += (BYTES_PER_WORD - 1);
1736 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 }
1738
1739 /* calculate out the final buffer alignment: */
1740 /* 1) arch recommendation: can be overridden for debug */
1741 if (flags & SLAB_HWCACHE_ALIGN) {
1742 /* Default alignment: as specified by the arch code.
1743 * Except if an object is really small, then squeeze multiple
1744 * objects into one cacheline.
1745 */
1746 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 ralign /= 2;
1749 } else {
1750 ralign = BYTES_PER_WORD;
1751 }
1752 /* 2) arch mandated alignment: disables debug if necessary */
1753 if (ralign < ARCH_SLAB_MINALIGN) {
1754 ralign = ARCH_SLAB_MINALIGN;
1755 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001756 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 }
1758 /* 3) caller mandated alignment: disables debug if necessary */
1759 if (ralign < align) {
1760 ralign = align;
1761 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001762 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 }
1764 /* 4) Store it. Note that the debug code below can reduce
1765 * the alignment to BYTES_PER_WORD.
1766 */
1767 align = ralign;
1768
1769 /* Get cache's description obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001770 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001772 goto oops;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001773 memset(cachep, 0, sizeof(struct kmem_cache));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
1775#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001776 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777
1778 if (flags & SLAB_RED_ZONE) {
1779 /* redzoning only works with word aligned caches */
1780 align = BYTES_PER_WORD;
1781
1782 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001783 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001784 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 }
1786 if (flags & SLAB_STORE_USER) {
1787 /* user store requires word alignment and
1788 * one word storage behind the end of the real
1789 * object.
1790 */
1791 align = BYTES_PER_WORD;
1792 size += BYTES_PER_WORD;
1793 }
1794#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001796 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1797 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 size = PAGE_SIZE;
1799 }
1800#endif
1801#endif
1802
1803 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001804 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 /*
1806 * Size is large, assume best to place the slab management obj
1807 * off-slab (should allow better packing of objs).
1808 */
1809 flags |= CFLGS_OFF_SLAB;
1810
1811 size = ALIGN(size, align);
1812
1813 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1814 /*
1815 * A VFS-reclaimable slab tends to have most allocations
1816 * as GFP_NOFS and we really don't want to have to be allocating
1817 * higher-order pages when we are unable to shrink dcache.
1818 */
1819 cachep->gfporder = 0;
1820 cache_estimate(cachep->gfporder, size, align, flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001821 &left_over, &cachep->num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001822 } else
1823 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824
1825 if (!cachep->num) {
1826 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1827 kmem_cache_free(&cache_cache, cachep);
1828 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001829 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001831 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1832 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
1834 /*
1835 * If the slab has been placed off-slab, and we have enough space then
1836 * move it on-slab. This is at the expense of any extra colouring.
1837 */
1838 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1839 flags &= ~CFLGS_OFF_SLAB;
1840 left_over -= slab_size;
1841 }
1842
1843 if (flags & CFLGS_OFF_SLAB) {
1844 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001845 slab_size =
1846 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 }
1848
1849 cachep->colour_off = cache_line_size();
1850 /* Offset must be a multiple of the alignment. */
1851 if (cachep->colour_off < align)
1852 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001853 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 cachep->slab_size = slab_size;
1855 cachep->flags = flags;
1856 cachep->gfpflags = 0;
1857 if (flags & SLAB_CACHE_DMA)
1858 cachep->gfpflags |= GFP_DMA;
1859 spin_lock_init(&cachep->spinlock);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001860 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861
1862 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001863 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 cachep->ctor = ctor;
1865 cachep->dtor = dtor;
1866 cachep->name = name;
1867
1868 /* Don't let CPUs to come and go */
1869 lock_cpu_hotplug();
1870
1871 if (g_cpucache_up == FULL) {
1872 enable_cpucache(cachep);
1873 } else {
1874 if (g_cpucache_up == NONE) {
1875 /* Note: the first kmem_cache_create must create
1876 * the cache that's used by kmalloc(24), otherwise
1877 * the creation of further caches will BUG().
1878 */
Christoph Lametere498be72005-09-09 13:03:32 -07001879 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 &initarray_generic.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001881
1882 /* If the cache that's used by
1883 * kmalloc(sizeof(kmem_list3)) is the first cache,
1884 * then we need to set up all its list3s, otherwise
1885 * the creation of further caches will BUG().
1886 */
1887 set_up_list3s(cachep, SIZE_AC);
1888 if (INDEX_AC == INDEX_L3)
1889 g_cpucache_up = PARTIAL_L3;
1890 else
1891 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001893 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001894 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001895
1896 if (g_cpucache_up == PARTIAL_AC) {
1897 set_up_list3s(cachep, SIZE_L3);
1898 g_cpucache_up = PARTIAL_L3;
1899 } else {
1900 int node;
1901 for_each_online_node(node) {
1902
1903 cachep->nodelists[node] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001904 kmalloc_node(sizeof
1905 (struct kmem_list3),
1906 GFP_KERNEL, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001907 BUG_ON(!cachep->nodelists[node]);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001908 kmem_list3_init(cachep->
1909 nodelists[node]);
Christoph Lametere498be72005-09-09 13:03:32 -07001910 }
1911 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 }
Christoph Lametere498be72005-09-09 13:03:32 -07001913 cachep->nodelists[numa_node_id()]->next_reap =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001914 jiffies + REAPTIMEOUT_LIST3 +
1915 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001916
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001917 BUG_ON(!cpu_cache_get(cachep));
1918 cpu_cache_get(cachep)->avail = 0;
1919 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1920 cpu_cache_get(cachep)->batchcount = 1;
1921 cpu_cache_get(cachep)->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 cachep->batchcount = 1;
1923 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 /* cache setup completed, link it into the list */
1927 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 unlock_cpu_hotplug();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001929 oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 if (!cachep && (flags & SLAB_PANIC))
1931 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001933 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 return cachep;
1935}
1936EXPORT_SYMBOL(kmem_cache_create);
1937
1938#if DEBUG
1939static void check_irq_off(void)
1940{
1941 BUG_ON(!irqs_disabled());
1942}
1943
1944static void check_irq_on(void)
1945{
1946 BUG_ON(irqs_disabled());
1947}
1948
Pekka Enberg343e0d72006-02-01 03:05:50 -08001949static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950{
1951#ifdef CONFIG_SMP
1952 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001953 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954#endif
1955}
Christoph Lametere498be72005-09-09 13:03:32 -07001956
Pekka Enberg343e0d72006-02-01 03:05:50 -08001957static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001958{
1959#ifdef CONFIG_SMP
1960 check_irq_off();
1961 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1962#endif
1963}
1964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965#else
1966#define check_irq_off() do { } while(0)
1967#define check_irq_on() do { } while(0)
1968#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001969#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970#endif
1971
1972/*
1973 * Waits for all CPUs to execute func().
1974 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001975static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976{
1977 check_irq_on();
1978 preempt_disable();
1979
1980 local_irq_disable();
1981 func(arg);
1982 local_irq_enable();
1983
1984 if (smp_call_function(func, arg, 1, 1))
1985 BUG();
1986
1987 preempt_enable();
1988}
1989
Pekka Enberg343e0d72006-02-01 03:05:50 -08001990static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001991 int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992
1993static void do_drain(void *arg)
1994{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001995 struct kmem_cache *cachep = (struct kmem_cache *) arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07001997 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998
1999 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002000 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002001 spin_lock(&cachep->nodelists[node]->list_lock);
2002 free_block(cachep, ac->entry, ac->avail, node);
2003 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 ac->avail = 0;
2005}
2006
Pekka Enberg343e0d72006-02-01 03:05:50 -08002007static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008{
Christoph Lametere498be72005-09-09 13:03:32 -07002009 struct kmem_list3 *l3;
2010 int node;
2011
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 smp_call_function_all_cpus(do_drain, cachep);
2013 check_irq_on();
2014 spin_lock_irq(&cachep->spinlock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002015 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002016 l3 = cachep->nodelists[node];
2017 if (l3) {
2018 spin_lock(&l3->list_lock);
2019 drain_array_locked(cachep, l3->shared, 1, node);
2020 spin_unlock(&l3->list_lock);
2021 if (l3->alien)
2022 drain_alien_cache(cachep, l3);
2023 }
2024 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 spin_unlock_irq(&cachep->spinlock);
2026}
2027
Pekka Enberg343e0d72006-02-01 03:05:50 -08002028static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029{
2030 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002031 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 int ret;
2033
Christoph Lametere498be72005-09-09 13:03:32 -07002034 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 struct list_head *p;
2036
Christoph Lametere498be72005-09-09 13:03:32 -07002037 p = l3->slabs_free.prev;
2038 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 break;
2040
Christoph Lametere498be72005-09-09 13:03:32 -07002041 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042#if DEBUG
2043 if (slabp->inuse)
2044 BUG();
2045#endif
2046 list_del(&slabp->list);
2047
Christoph Lametere498be72005-09-09 13:03:32 -07002048 l3->free_objects -= cachep->num;
2049 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002051 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002053 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054 return ret;
2055}
2056
Pekka Enberg343e0d72006-02-01 03:05:50 -08002057static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002058{
2059 int ret = 0, i = 0;
2060 struct kmem_list3 *l3;
2061
2062 drain_cpu_caches(cachep);
2063
2064 check_irq_on();
2065 for_each_online_node(i) {
2066 l3 = cachep->nodelists[i];
2067 if (l3) {
2068 spin_lock_irq(&l3->list_lock);
2069 ret += __node_shrink(cachep, i);
2070 spin_unlock_irq(&l3->list_lock);
2071 }
2072 }
2073 return (ret ? 1 : 0);
2074}
2075
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076/**
2077 * kmem_cache_shrink - Shrink a cache.
2078 * @cachep: The cache to shrink.
2079 *
2080 * Releases as many slabs as possible for a cache.
2081 * To help debugging, a zero exit status indicates all slabs were released.
2082 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002083int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084{
2085 if (!cachep || in_interrupt())
2086 BUG();
2087
2088 return __cache_shrink(cachep);
2089}
2090EXPORT_SYMBOL(kmem_cache_shrink);
2091
2092/**
2093 * kmem_cache_destroy - delete a cache
2094 * @cachep: the cache to destroy
2095 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002096 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 * Returns 0 on success.
2098 *
2099 * It is expected this function will be called by a module when it is
2100 * unloaded. This will remove the cache completely, and avoid a duplicate
2101 * cache being allocated each time a module is loaded and unloaded, if the
2102 * module doesn't have persistent in-kernel storage across loads and unloads.
2103 *
2104 * The cache must be empty before calling this function.
2105 *
2106 * The caller must guarantee that noone will allocate memory from the cache
2107 * during the kmem_cache_destroy().
2108 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002109int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110{
2111 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002112 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
2114 if (!cachep || in_interrupt())
2115 BUG();
2116
2117 /* Don't let CPUs to come and go */
2118 lock_cpu_hotplug();
2119
2120 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002121 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 /*
2123 * the chain is never empty, cache_cache is never destroyed
2124 */
2125 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002126 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127
2128 if (__cache_shrink(cachep)) {
2129 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002130 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002131 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002132 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 unlock_cpu_hotplug();
2134 return 1;
2135 }
2136
2137 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002138 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139
Christoph Lametere498be72005-09-09 13:03:32 -07002140 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002141 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142
2143 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002144 for_each_online_node(i) {
2145 if ((l3 = cachep->nodelists[i])) {
2146 kfree(l3->shared);
2147 free_alien_cache(l3->alien);
2148 kfree(l3);
2149 }
2150 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 kmem_cache_free(&cache_cache, cachep);
2152
2153 unlock_cpu_hotplug();
2154
2155 return 0;
2156}
2157EXPORT_SYMBOL(kmem_cache_destroy);
2158
2159/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002160static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002161 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
2163 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002164
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 if (OFF_SLAB(cachep)) {
2166 /* Slab management obj is off-slab. */
2167 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2168 if (!slabp)
2169 return NULL;
2170 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002171 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 colour_off += cachep->slab_size;
2173 }
2174 slabp->inuse = 0;
2175 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002176 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177
2178 return slabp;
2179}
2180
2181static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2182{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002183 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184}
2185
Pekka Enberg343e0d72006-02-01 03:05:50 -08002186static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002187 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188{
2189 int i;
2190
2191 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002192 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193#if DEBUG
2194 /* need to poison the objs? */
2195 if (cachep->flags & SLAB_POISON)
2196 poison_obj(cachep, objp, POISON_FREE);
2197 if (cachep->flags & SLAB_STORE_USER)
2198 *dbg_userword(cachep, objp) = NULL;
2199
2200 if (cachep->flags & SLAB_RED_ZONE) {
2201 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2202 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2203 }
2204 /*
2205 * Constructors are not allowed to allocate memory from
2206 * the same cache which they are a constructor for.
2207 * Otherwise, deadlock. They must also be threaded.
2208 */
2209 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002210 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002211 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212
2213 if (cachep->flags & SLAB_RED_ZONE) {
2214 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2215 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2218 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002219 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002221 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002222 && cachep->flags & SLAB_POISON)
2223 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002224 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225#else
2226 if (cachep->ctor)
2227 cachep->ctor(objp, cachep, ctor_flags);
2228#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002229 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002231 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 slabp->free = 0;
2233}
2234
Pekka Enberg343e0d72006-02-01 03:05:50 -08002235static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236{
2237 if (flags & SLAB_DMA) {
2238 if (!(cachep->gfpflags & GFP_DMA))
2239 BUG();
2240 } else {
2241 if (cachep->gfpflags & GFP_DMA)
2242 BUG();
2243 }
2244}
2245
Pekka Enberg343e0d72006-02-01 03:05:50 -08002246static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002247{
2248 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2249 kmem_bufctl_t next;
2250
2251 slabp->inuse++;
2252 next = slab_bufctl(slabp)[slabp->free];
2253#if DEBUG
2254 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2255 WARN_ON(slabp->nodeid != nodeid);
2256#endif
2257 slabp->free = next;
2258
2259 return objp;
2260}
2261
Pekka Enberg343e0d72006-02-01 03:05:50 -08002262static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
Matthew Dobson78d382d2006-02-01 03:05:47 -08002263 int nodeid)
2264{
2265 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2266
2267#if DEBUG
2268 /* Verify that the slab belongs to the intended node */
2269 WARN_ON(slabp->nodeid != nodeid);
2270
2271 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2272 printk(KERN_ERR "slab: double free detected in cache "
2273 "'%s', objp %p\n", cachep->name, objp);
2274 BUG();
2275 }
2276#endif
2277 slab_bufctl(slabp)[objnr] = slabp->free;
2278 slabp->free = objnr;
2279 slabp->inuse--;
2280}
2281
Pekka Enberg343e0d72006-02-01 03:05:50 -08002282static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283{
2284 int i;
2285 struct page *page;
2286
2287 /* Nasty!!!!!! I hope this is OK. */
2288 i = 1 << cachep->gfporder;
2289 page = virt_to_page(objp);
2290 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002291 page_set_cache(page, cachep);
2292 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 page++;
2294 } while (--i);
2295}
2296
2297/*
2298 * Grow (by 1) the number of slabs within a cache. This is called by
2299 * kmem_cache_alloc() when there are no active objs left in a cache.
2300 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002301static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002303 struct slab *slabp;
2304 void *objp;
2305 size_t offset;
2306 gfp_t local_flags;
2307 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002308 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309
2310 /* Be lazy and only check for valid flags here,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002311 * keeping it out of the critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002313 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 BUG();
2315 if (flags & SLAB_NO_GROW)
2316 return 0;
2317
2318 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2319 local_flags = (flags & SLAB_LEVEL_MASK);
2320 if (!(local_flags & __GFP_WAIT))
2321 /*
2322 * Not allowed to sleep. Need to tell a constructor about
2323 * this - it might need to know...
2324 */
2325 ctor_flags |= SLAB_CTOR_ATOMIC;
2326
2327 /* About to mess with non-constant members - lock. */
2328 check_irq_off();
2329 spin_lock(&cachep->spinlock);
2330
2331 /* Get colour for the slab, and cal the next value. */
2332 offset = cachep->colour_next;
2333 cachep->colour_next++;
2334 if (cachep->colour_next >= cachep->colour)
2335 cachep->colour_next = 0;
2336 offset *= cachep->colour_off;
2337
2338 spin_unlock(&cachep->spinlock);
2339
Christoph Lametere498be72005-09-09 13:03:32 -07002340 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 if (local_flags & __GFP_WAIT)
2342 local_irq_enable();
2343
2344 /*
2345 * The test for missing atomic flag is performed here, rather than
2346 * the more obvious place, simply to reduce the critical path length
2347 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2348 * will eventually be caught here (where it matters).
2349 */
2350 kmem_flagcheck(cachep, flags);
2351
Christoph Lametere498be72005-09-09 13:03:32 -07002352 /* Get mem for the objs.
2353 * Attempt to allocate a physical page from 'nodeid',
2354 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2356 goto failed;
2357
2358 /* Get slab management. */
2359 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2360 goto opps1;
2361
Christoph Lametere498be72005-09-09 13:03:32 -07002362 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 set_slab_attr(cachep, slabp, objp);
2364
2365 cache_init_objs(cachep, slabp, ctor_flags);
2366
2367 if (local_flags & __GFP_WAIT)
2368 local_irq_disable();
2369 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002370 l3 = cachep->nodelists[nodeid];
2371 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372
2373 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002374 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002376 l3->free_objects += cachep->num;
2377 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002379 opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 kmem_freepages(cachep, objp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002381 failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 if (local_flags & __GFP_WAIT)
2383 local_irq_disable();
2384 return 0;
2385}
2386
2387#if DEBUG
2388
2389/*
2390 * Perform extra freeing checks:
2391 * - detect bad pointers.
2392 * - POISON/RED_ZONE checking
2393 * - destructor calls, for caches with POISON+dtor
2394 */
2395static void kfree_debugcheck(const void *objp)
2396{
2397 struct page *page;
2398
2399 if (!virt_addr_valid(objp)) {
2400 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002401 (unsigned long)objp);
2402 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 }
2404 page = virt_to_page(objp);
2405 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002406 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2407 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 BUG();
2409 }
2410}
2411
Pekka Enberg343e0d72006-02-01 03:05:50 -08002412static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002413 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
2415 struct page *page;
2416 unsigned int objnr;
2417 struct slab *slabp;
2418
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002419 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 kfree_debugcheck(objp);
2421 page = virt_to_page(objp);
2422
Pekka Enberg065d41c2005-11-13 16:06:46 -08002423 if (page_get_cache(page) != cachep) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002424 printk(KERN_ERR
2425 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2426 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002428 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2429 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 WARN_ON(1);
2431 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002432 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433
2434 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002435 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2436 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2437 slab_error(cachep,
2438 "double free, or memory outside"
2439 " object was overwritten");
2440 printk(KERN_ERR
2441 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2442 objp, *dbg_redzone1(cachep, objp),
2443 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 }
2445 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2446 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2447 }
2448 if (cachep->flags & SLAB_STORE_USER)
2449 *dbg_userword(cachep, objp) = caller;
2450
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002451 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452
2453 BUG_ON(objnr >= cachep->num);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002454 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
2456 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2457 /* Need to call the slab's constructor so the
2458 * caller can perform a verify of its state (debugging).
2459 * Called without the cache-lock held.
2460 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002461 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002462 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 }
2464 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2465 /* we want to cache poison the object,
2466 * call the destruction callback
2467 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002468 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 }
2470 if (cachep->flags & SLAB_POISON) {
2471#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002472 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002474 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002475 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 } else {
2477 poison_obj(cachep, objp, POISON_FREE);
2478 }
2479#else
2480 poison_obj(cachep, objp, POISON_FREE);
2481#endif
2482 }
2483 return objp;
2484}
2485
Pekka Enberg343e0d72006-02-01 03:05:50 -08002486static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487{
2488 kmem_bufctl_t i;
2489 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002490
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 /* Check slab's freelist to see if this obj is there. */
2492 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2493 entries++;
2494 if (entries > cachep->num || i >= cachep->num)
2495 goto bad;
2496 }
2497 if (entries != cachep->num - slabp->inuse) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002498 bad:
2499 printk(KERN_ERR
2500 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2501 cachep->name, cachep->num, slabp, slabp->inuse);
2502 for (i = 0;
2503 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2504 i++) {
2505 if ((i % 16) == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002507 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 }
2509 printk("\n");
2510 BUG();
2511 }
2512}
2513#else
2514#define kfree_debugcheck(x) do { } while(0)
2515#define cache_free_debugcheck(x,objp,z) (objp)
2516#define check_slabp(x,y) do { } while(0)
2517#endif
2518
Pekka Enberg343e0d72006-02-01 03:05:50 -08002519static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520{
2521 int batchcount;
2522 struct kmem_list3 *l3;
2523 struct array_cache *ac;
2524
2525 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002526 ac = cpu_cache_get(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002527 retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 batchcount = ac->batchcount;
2529 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2530 /* if there was little recent activity on this
2531 * cache, then perform only a partial refill.
2532 * Otherwise we could generate refill bouncing.
2533 */
2534 batchcount = BATCHREFILL_LIMIT;
2535 }
Christoph Lametere498be72005-09-09 13:03:32 -07002536 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537
Christoph Lametere498be72005-09-09 13:03:32 -07002538 BUG_ON(ac->avail > 0 || !l3);
2539 spin_lock(&l3->list_lock);
2540
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 if (l3->shared) {
2542 struct array_cache *shared_array = l3->shared;
2543 if (shared_array->avail) {
2544 if (batchcount > shared_array->avail)
2545 batchcount = shared_array->avail;
2546 shared_array->avail -= batchcount;
2547 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002548 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002549 &(shared_array->entry[shared_array->avail]),
2550 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 shared_array->touched = 1;
2552 goto alloc_done;
2553 }
2554 }
2555 while (batchcount > 0) {
2556 struct list_head *entry;
2557 struct slab *slabp;
2558 /* Get slab alloc is to come from. */
2559 entry = l3->slabs_partial.next;
2560 if (entry == &l3->slabs_partial) {
2561 l3->free_touched = 1;
2562 entry = l3->slabs_free.next;
2563 if (entry == &l3->slabs_free)
2564 goto must_grow;
2565 }
2566
2567 slabp = list_entry(entry, struct slab, list);
2568 check_slabp(cachep, slabp);
2569 check_spinlock_acquired(cachep);
2570 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 STATS_INC_ALLOCED(cachep);
2572 STATS_INC_ACTIVE(cachep);
2573 STATS_SET_HIGH(cachep);
2574
Matthew Dobson78d382d2006-02-01 03:05:47 -08002575 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2576 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 }
2578 check_slabp(cachep, slabp);
2579
2580 /* move slabp to correct slabp list: */
2581 list_del(&slabp->list);
2582 if (slabp->free == BUFCTL_END)
2583 list_add(&slabp->list, &l3->slabs_full);
2584 else
2585 list_add(&slabp->list, &l3->slabs_partial);
2586 }
2587
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002588 must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 l3->free_objects -= ac->avail;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590 alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002591 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592
2593 if (unlikely(!ac->avail)) {
2594 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002595 x = cache_grow(cachep, flags, numa_node_id());
2596
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 // cache_grow can reenable interrupts, then ac could change.
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002598 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 if (!x && ac->avail == 0) // no objects in sight? abort
2600 return NULL;
2601
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002602 if (!ac->avail) // objects refilled by interrupt?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 goto retry;
2604 }
2605 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002606 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607}
2608
2609static inline void
Pekka Enberg343e0d72006-02-01 03:05:50 -08002610cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611{
2612 might_sleep_if(flags & __GFP_WAIT);
2613#if DEBUG
2614 kmem_flagcheck(cachep, flags);
2615#endif
2616}
2617
2618#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -08002619static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002620 void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002622 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002624 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002626 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002627 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002628 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 else
2630 check_poison_obj(cachep, objp);
2631#else
2632 check_poison_obj(cachep, objp);
2633#endif
2634 poison_obj(cachep, objp, POISON_INUSE);
2635 }
2636 if (cachep->flags & SLAB_STORE_USER)
2637 *dbg_userword(cachep, objp) = caller;
2638
2639 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2641 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2642 slab_error(cachep,
2643 "double free, or memory outside"
2644 " object was overwritten");
2645 printk(KERN_ERR
2646 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2647 objp, *dbg_redzone1(cachep, objp),
2648 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 }
2650 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2651 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2652 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002653 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002655 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656
2657 if (!(flags & __GFP_WAIT))
2658 ctor_flags |= SLAB_CTOR_ATOMIC;
2659
2660 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 return objp;
2663}
2664#else
2665#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2666#endif
2667
Pekka Enberg343e0d72006-02-01 03:05:50 -08002668static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002670 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 struct array_cache *ac;
2672
Christoph Lameterdc85da12006-01-18 17:42:36 -08002673#ifdef CONFIG_NUMA
Christoph Lameter86c562a2006-01-18 17:42:37 -08002674 if (unlikely(current->mempolicy && !in_interrupt())) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08002675 int nid = slab_node(current->mempolicy);
2676
2677 if (nid != numa_node_id())
2678 return __cache_alloc_node(cachep, flags, nid);
2679 }
2680#endif
2681
Alok N Kataria5c382302005-09-27 21:45:46 -07002682 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002683 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 if (likely(ac->avail)) {
2685 STATS_INC_ALLOCHIT(cachep);
2686 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002687 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 } else {
2689 STATS_INC_ALLOCMISS(cachep);
2690 objp = cache_alloc_refill(cachep, flags);
2691 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002692 return objp;
2693}
2694
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002695static __always_inline void *
2696__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002697{
2698 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002699 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002700
2701 cache_alloc_debugcheck_before(cachep, flags);
2702
2703 local_irq_save(save_flags);
2704 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002706 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002707 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002708 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 return objp;
2710}
2711
Christoph Lametere498be72005-09-09 13:03:32 -07002712#ifdef CONFIG_NUMA
2713/*
2714 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002716static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002717{
2718 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002719 struct slab *slabp;
2720 struct kmem_list3 *l3;
2721 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002722 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002724 l3 = cachep->nodelists[nodeid];
2725 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002726
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002727 retry:
2728 spin_lock(&l3->list_lock);
2729 entry = l3->slabs_partial.next;
2730 if (entry == &l3->slabs_partial) {
2731 l3->free_touched = 1;
2732 entry = l3->slabs_free.next;
2733 if (entry == &l3->slabs_free)
2734 goto must_grow;
2735 }
Christoph Lametere498be72005-09-09 13:03:32 -07002736
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002737 slabp = list_entry(entry, struct slab, list);
2738 check_spinlock_acquired_node(cachep, nodeid);
2739 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002740
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 STATS_INC_NODEALLOCS(cachep);
2742 STATS_INC_ACTIVE(cachep);
2743 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002744
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002745 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002746
Matthew Dobson78d382d2006-02-01 03:05:47 -08002747 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002748 check_slabp(cachep, slabp);
2749 l3->free_objects--;
2750 /* move slabp to correct slabp list: */
2751 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002752
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002753 if (slabp->free == BUFCTL_END) {
2754 list_add(&slabp->list, &l3->slabs_full);
2755 } else {
2756 list_add(&slabp->list, &l3->slabs_partial);
2757 }
Christoph Lametere498be72005-09-09 13:03:32 -07002758
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002759 spin_unlock(&l3->list_lock);
2760 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002761
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002762 must_grow:
2763 spin_unlock(&l3->list_lock);
2764 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002765
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002766 if (!x)
2767 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002768
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002769 goto retry;
2770 done:
2771 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002772}
2773#endif
2774
2775/*
2776 * Caller needs to acquire correct kmem_list's list_lock
2777 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002778static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002779 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780{
2781 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002782 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
2784 for (i = 0; i < nr_objects; i++) {
2785 void *objp = objpp[i];
2786 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08002788 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002789 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002791 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002793 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002795 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 check_slabp(cachep, slabp);
2797
2798 /* fixup slab chains */
2799 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002800 if (l3->free_objects > l3->free_limit) {
2801 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 slab_destroy(cachep, slabp);
2803 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002804 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 }
2806 } else {
2807 /* Unconditionally move a slab to the end of the
2808 * partial list on free - maximum time for the
2809 * other objects to be freed, too.
2810 */
Christoph Lametere498be72005-09-09 13:03:32 -07002811 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 }
2813 }
2814}
2815
Pekka Enberg343e0d72006-02-01 03:05:50 -08002816static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817{
2818 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002819 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002820 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
2822 batchcount = ac->batchcount;
2823#if DEBUG
2824 BUG_ON(!batchcount || batchcount > ac->avail);
2825#endif
2826 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002827 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002828 spin_lock(&l3->list_lock);
2829 if (l3->shared) {
2830 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002831 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 if (max) {
2833 if (batchcount > max)
2834 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002835 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002836 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 shared_array->avail += batchcount;
2838 goto free_done;
2839 }
2840 }
2841
Christoph Lameterff694162005-09-22 21:44:02 -07002842 free_block(cachep, ac->entry, batchcount, node);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002843 free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844#if STATS
2845 {
2846 int i = 0;
2847 struct list_head *p;
2848
Christoph Lametere498be72005-09-09 13:03:32 -07002849 p = l3->slabs_free.next;
2850 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 struct slab *slabp;
2852
2853 slabp = list_entry(p, struct slab, list);
2854 BUG_ON(slabp->inuse);
2855
2856 i++;
2857 p = p->next;
2858 }
2859 STATS_SET_FREEABLE(cachep, i);
2860 }
2861#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002862 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002864 memmove(ac->entry, &(ac->entry[batchcount]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002865 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866}
2867
2868/*
2869 * __cache_free
2870 * Release an obj back to its cache. If the obj has a constructed
2871 * state, it must be in this state _before_ it is released.
2872 *
2873 * Called with disabled ints.
2874 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002875static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002877 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878
2879 check_irq_off();
2880 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2881
Christoph Lametere498be72005-09-09 13:03:32 -07002882 /* Make sure we are not freeing a object from another
2883 * node to the array cache on this cpu.
2884 */
2885#ifdef CONFIG_NUMA
2886 {
2887 struct slab *slabp;
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08002888 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07002889 if (unlikely(slabp->nodeid != numa_node_id())) {
2890 struct array_cache *alien = NULL;
2891 int nodeid = slabp->nodeid;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002892 struct kmem_list3 *l3 =
2893 cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07002894
2895 STATS_INC_NODEFREES(cachep);
2896 if (l3->alien && l3->alien[nodeid]) {
2897 alien = l3->alien[nodeid];
2898 spin_lock(&alien->lock);
2899 if (unlikely(alien->avail == alien->limit))
2900 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002901 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002902 alien->entry[alien->avail++] = objp;
2903 spin_unlock(&alien->lock);
2904 } else {
2905 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002906 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002907 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002908 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002909 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002910 }
2911 return;
2912 }
2913 }
2914#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 if (likely(ac->avail < ac->limit)) {
2916 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002917 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 return;
2919 } else {
2920 STATS_INC_FREEMISS(cachep);
2921 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002922 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 }
2924}
2925
2926/**
2927 * kmem_cache_alloc - Allocate an object
2928 * @cachep: The cache to allocate from.
2929 * @flags: See kmalloc().
2930 *
2931 * Allocate an object from this cache. The flags are only relevant
2932 * if the cache has no available objects.
2933 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002934void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002936 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937}
2938EXPORT_SYMBOL(kmem_cache_alloc);
2939
2940/**
2941 * kmem_ptr_validate - check if an untrusted pointer might
2942 * be a slab entry.
2943 * @cachep: the cache we're checking against
2944 * @ptr: pointer to validate
2945 *
2946 * This verifies that the untrusted pointer looks sane:
2947 * it is _not_ a guarantee that the pointer is actually
2948 * part of the slab cache in question, but it at least
2949 * validates that the pointer can be dereferenced and
2950 * looks half-way sane.
2951 *
2952 * Currently only used for dentry validation.
2953 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002954int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002956 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002958 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002959 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 struct page *page;
2961
2962 if (unlikely(addr < min_addr))
2963 goto out;
2964 if (unlikely(addr > (unsigned long)high_memory - size))
2965 goto out;
2966 if (unlikely(addr & align_mask))
2967 goto out;
2968 if (unlikely(!kern_addr_valid(addr)))
2969 goto out;
2970 if (unlikely(!kern_addr_valid(addr + size - 1)))
2971 goto out;
2972 page = virt_to_page(ptr);
2973 if (unlikely(!PageSlab(page)))
2974 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002975 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 goto out;
2977 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002978 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 return 0;
2980}
2981
2982#ifdef CONFIG_NUMA
2983/**
2984 * kmem_cache_alloc_node - Allocate an object on the specified node
2985 * @cachep: The cache to allocate from.
2986 * @flags: See kmalloc().
2987 * @nodeid: node number of the target node.
2988 *
2989 * Identical to kmem_cache_alloc, except that this function is slow
2990 * and can sleep. And it will allocate memory on the given node, which
2991 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002992 * New and improved: it will now make sure that the object gets
2993 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002995void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996{
Christoph Lametere498be72005-09-09 13:03:32 -07002997 unsigned long save_flags;
2998 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999
Christoph Lametere498be72005-09-09 13:03:32 -07003000 cache_alloc_debugcheck_before(cachep, flags);
3001 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003002
3003 if (nodeid == -1 || nodeid == numa_node_id() ||
3004 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003005 ptr = ____cache_alloc(cachep, flags);
3006 else
3007 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003008 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003009
3010 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3011 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
Christoph Lametere498be72005-09-09 13:03:32 -07003013 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014}
3015EXPORT_SYMBOL(kmem_cache_alloc_node);
3016
Al Virodd0fc662005-10-07 07:46:04 +01003017void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003018{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003019 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003020
3021 cachep = kmem_find_general_cachep(size, flags);
3022 if (unlikely(cachep == NULL))
3023 return NULL;
3024 return kmem_cache_alloc_node(cachep, flags, node);
3025}
3026EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027#endif
3028
3029/**
3030 * kmalloc - allocate memory
3031 * @size: how many bytes of memory are required.
3032 * @flags: the type of memory to allocate.
3033 *
3034 * kmalloc is the normal method of allocating memory
3035 * in the kernel.
3036 *
3037 * The @flags argument may be one of:
3038 *
3039 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3040 *
3041 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3042 *
3043 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3044 *
3045 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3046 * must be suitable for DMA. This can mean different things on different
3047 * platforms. For example, on i386, it means that the memory must come
3048 * from the first 16MB.
3049 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003050static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3051 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003053 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003055 /* If you want to save a few bytes .text space: replace
3056 * __ with kmem_.
3057 * Then kmalloc uses the uninlined functions instead of the inline
3058 * functions.
3059 */
3060 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003061 if (unlikely(cachep == NULL))
3062 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003063 return __cache_alloc(cachep, flags, caller);
3064}
3065
3066#ifndef CONFIG_DEBUG_SLAB
3067
3068void *__kmalloc(size_t size, gfp_t flags)
3069{
3070 return __do_kmalloc(size, flags, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071}
3072EXPORT_SYMBOL(__kmalloc);
3073
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003074#else
3075
3076void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3077{
3078 return __do_kmalloc(size, flags, caller);
3079}
3080EXPORT_SYMBOL(__kmalloc_track_caller);
3081
3082#endif
3083
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084#ifdef CONFIG_SMP
3085/**
3086 * __alloc_percpu - allocate one copy of the object for every present
3087 * cpu in the system, zeroing them.
3088 * Objects should be dereferenced using the per_cpu_ptr macro only.
3089 *
3090 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003092void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093{
3094 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003095 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096
3097 if (!pdata)
3098 return NULL;
3099
Christoph Lametere498be72005-09-09 13:03:32 -07003100 /*
3101 * Cannot use for_each_online_cpu since a cpu may come online
3102 * and we have no way of figuring out how to fix the array
3103 * that we have allocated then....
3104 */
3105 for_each_cpu(i) {
3106 int node = cpu_to_node(i);
3107
3108 if (node_online(node))
3109 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3110 else
3111 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112
3113 if (!pdata->ptrs[i])
3114 goto unwind_oom;
3115 memset(pdata->ptrs[i], 0, size);
3116 }
3117
3118 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003119 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003121 unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 while (--i >= 0) {
3123 if (!cpu_possible(i))
3124 continue;
3125 kfree(pdata->ptrs[i]);
3126 }
3127 kfree(pdata);
3128 return NULL;
3129}
3130EXPORT_SYMBOL(__alloc_percpu);
3131#endif
3132
3133/**
3134 * kmem_cache_free - Deallocate an object
3135 * @cachep: The cache the allocation was from.
3136 * @objp: The previously allocated object.
3137 *
3138 * Free an object which was previously allocated from this
3139 * cache.
3140 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003141void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142{
3143 unsigned long flags;
3144
3145 local_irq_save(flags);
3146 __cache_free(cachep, objp);
3147 local_irq_restore(flags);
3148}
3149EXPORT_SYMBOL(kmem_cache_free);
3150
3151/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 * kfree - free previously allocated memory
3153 * @objp: pointer returned by kmalloc.
3154 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003155 * If @objp is NULL, no operation is performed.
3156 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 * Don't free memory not originally allocated by kmalloc()
3158 * or you will run into trouble.
3159 */
3160void kfree(const void *objp)
3161{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003162 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 unsigned long flags;
3164
3165 if (unlikely(!objp))
3166 return;
3167 local_irq_save(flags);
3168 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003169 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003170 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003171 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 local_irq_restore(flags);
3173}
3174EXPORT_SYMBOL(kfree);
3175
3176#ifdef CONFIG_SMP
3177/**
3178 * free_percpu - free previously allocated percpu memory
3179 * @objp: pointer returned by alloc_percpu.
3180 *
3181 * Don't free memory not originally allocated by alloc_percpu()
3182 * The complemented objp is to check for that.
3183 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003184void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185{
3186 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003187 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188
Christoph Lametere498be72005-09-09 13:03:32 -07003189 /*
3190 * We allocate for all cpus so we cannot use for online cpu here.
3191 */
3192 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003193 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 kfree(p);
3195}
3196EXPORT_SYMBOL(free_percpu);
3197#endif
3198
Pekka Enberg343e0d72006-02-01 03:05:50 -08003199unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003201 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202}
3203EXPORT_SYMBOL(kmem_cache_size);
3204
Pekka Enberg343e0d72006-02-01 03:05:50 -08003205const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003206{
3207 return cachep->name;
3208}
3209EXPORT_SYMBOL_GPL(kmem_cache_name);
3210
Christoph Lametere498be72005-09-09 13:03:32 -07003211/*
3212 * This initializes kmem_list3 for all nodes.
3213 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003214static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003215{
3216 int node;
3217 struct kmem_list3 *l3;
3218 int err = 0;
3219
3220 for_each_online_node(node) {
3221 struct array_cache *nc = NULL, *new;
3222 struct array_cache **new_alien = NULL;
3223#ifdef CONFIG_NUMA
3224 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3225 goto fail;
3226#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003227 if (!(new = alloc_arraycache(node, (cachep->shared *
3228 cachep->batchcount),
3229 0xbaadf00d)))
Christoph Lametere498be72005-09-09 13:03:32 -07003230 goto fail;
3231 if ((l3 = cachep->nodelists[node])) {
3232
3233 spin_lock_irq(&l3->list_lock);
3234
3235 if ((nc = cachep->nodelists[node]->shared))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003236 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003237
3238 l3->shared = new;
3239 if (!cachep->nodelists[node]->alien) {
3240 l3->alien = new_alien;
3241 new_alien = NULL;
3242 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003243 l3->free_limit = (1 + nr_cpus_node(node)) *
3244 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003245 spin_unlock_irq(&l3->list_lock);
3246 kfree(nc);
3247 free_alien_cache(new_alien);
3248 continue;
3249 }
3250 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003251 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -07003252 goto fail;
3253
3254 kmem_list3_init(l3);
3255 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003256 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003257 l3->shared = new;
3258 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003259 l3->free_limit = (1 + nr_cpus_node(node)) *
3260 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003261 cachep->nodelists[node] = l3;
3262 }
3263 return err;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003264 fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003265 err = -ENOMEM;
3266 return err;
3267}
3268
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003270 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 struct array_cache *new[NR_CPUS];
3272};
3273
3274static void do_ccupdate_local(void *info)
3275{
3276 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3277 struct array_cache *old;
3278
3279 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003280 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3283 new->new[smp_processor_id()] = old;
3284}
3285
Pekka Enberg343e0d72006-02-01 03:05:50 -08003286static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003287 int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288{
3289 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003290 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003292 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003293 for_each_online_cpu(i) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003294 new.new[i] =
3295 alloc_arraycache(cpu_to_node(i), limit, batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003296 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003297 for (i--; i >= 0; i--)
3298 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003299 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 }
3301 }
3302 new.cachep = cachep;
3303
3304 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003305
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 check_irq_on();
3307 spin_lock_irq(&cachep->spinlock);
3308 cachep->batchcount = batchcount;
3309 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003310 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 spin_unlock_irq(&cachep->spinlock);
3312
Christoph Lametere498be72005-09-09 13:03:32 -07003313 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 struct array_cache *ccold = new.new[i];
3315 if (!ccold)
3316 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003317 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003318 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003319 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320 kfree(ccold);
3321 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Christoph Lametere498be72005-09-09 13:03:32 -07003323 err = alloc_kmemlist(cachep);
3324 if (err) {
3325 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003326 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003327 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 return 0;
3330}
3331
Pekka Enberg343e0d72006-02-01 03:05:50 -08003332static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333{
3334 int err;
3335 int limit, shared;
3336
3337 /* The head array serves three purposes:
3338 * - create a LIFO ordering, i.e. return objects that are cache-warm
3339 * - reduce the number of spinlock operations.
3340 * - reduce the number of linked list operations on the slab and
3341 * bufctl chains: array operations are cheaper.
3342 * The numbers are guessed, we should auto-tune as described by
3343 * Bonwick.
3344 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003345 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003347 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003349 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003351 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352 limit = 54;
3353 else
3354 limit = 120;
3355
3356 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3357 * allocation behaviour: Most allocs on one cpu, most free operations
3358 * on another cpu. For these cases, an efficient object passing between
3359 * cpus is necessary. This is provided by a shared array. The array
3360 * replaces Bonwick's magazine layer.
3361 * On uniprocessor, it's functionally equivalent (but less efficient)
3362 * to a larger limit. Thus disabled by default.
3363 */
3364 shared = 0;
3365#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003366 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 shared = 8;
3368#endif
3369
3370#if DEBUG
3371 /* With debugging enabled, large batchcount lead to excessively
3372 * long periods with disabled local interrupts. Limit the
3373 * batchcount
3374 */
3375 if (limit > 32)
3376 limit = 32;
3377#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003378 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 if (err)
3380 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003381 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382}
3383
Pekka Enberg343e0d72006-02-01 03:05:50 -08003384static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003385 int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386{
3387 int tofree;
3388
Christoph Lametere498be72005-09-09 13:03:32 -07003389 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 if (ac->touched && !force) {
3391 ac->touched = 0;
3392 } else if (ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003393 tofree = force ? ac->avail : (ac->limit + 4) / 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 if (tofree > ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003395 tofree = (ac->avail + 1) / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 }
Christoph Lameterff694162005-09-22 21:44:02 -07003397 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003399 memmove(ac->entry, &(ac->entry[tofree]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003400 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 }
3402}
3403
3404/**
3405 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003406 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 *
3408 * Called from workqueue/eventd every few seconds.
3409 * Purpose:
3410 * - clear the per-cpu caches for this CPU.
3411 * - return freeable pages to the main free memory pool.
3412 *
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003413 * If we cannot acquire the cache chain mutex then just give up - we'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 * try again on the next iteration.
3415 */
3416static void cache_reap(void *unused)
3417{
3418 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003419 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003421 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003423 schedule_delayed_work(&__get_cpu_var(reap_work),
3424 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425 return;
3426 }
3427
3428 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003429 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003430 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 int tofree;
3432 struct slab *slabp;
3433
Pekka Enberg343e0d72006-02-01 03:05:50 -08003434 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435
3436 if (searchp->flags & SLAB_NO_REAP)
3437 goto next;
3438
3439 check_irq_on();
3440
Christoph Lametere498be72005-09-09 13:03:32 -07003441 l3 = searchp->nodelists[numa_node_id()];
3442 if (l3->alien)
3443 drain_alien_cache(searchp, l3);
3444 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003446 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003447 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448
Christoph Lametere498be72005-09-09 13:03:32 -07003449 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 goto next_unlock;
3451
Christoph Lametere498be72005-09-09 13:03:32 -07003452 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453
Christoph Lametere498be72005-09-09 13:03:32 -07003454 if (l3->shared)
3455 drain_array_locked(searchp, l3->shared, 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003456 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Christoph Lametere498be72005-09-09 13:03:32 -07003458 if (l3->free_touched) {
3459 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460 goto next_unlock;
3461 }
3462
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003463 tofree =
3464 (l3->free_limit + 5 * searchp->num -
3465 1) / (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003467 p = l3->slabs_free.next;
3468 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 break;
3470
3471 slabp = list_entry(p, struct slab, list);
3472 BUG_ON(slabp->inuse);
3473 list_del(&slabp->list);
3474 STATS_INC_REAPED(searchp);
3475
3476 /* Safe to drop the lock. The slab is no longer
3477 * linked to the cache.
3478 * searchp cannot disappear, we hold
3479 * cache_chain_lock
3480 */
Christoph Lametere498be72005-09-09 13:03:32 -07003481 l3->free_objects -= searchp->num;
3482 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003484 spin_lock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003485 } while (--tofree > 0);
3486 next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003487 spin_unlock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003488 next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 cond_resched();
3490 }
3491 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003492 mutex_unlock(&cache_chain_mutex);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003493 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003495 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496}
3497
3498#ifdef CONFIG_PROC_FS
3499
Pekka Enberg85289f92006-01-08 01:00:36 -08003500static void print_slabinfo_header(struct seq_file *m)
3501{
3502 /*
3503 * Output format version, so at least we can change it
3504 * without _too_ many complaints.
3505 */
3506#if STATS
3507 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3508#else
3509 seq_puts(m, "slabinfo - version: 2.1\n");
3510#endif
3511 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3512 "<objperslab> <pagesperslab>");
3513 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3514 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3515#if STATS
3516 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3517 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3518 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3519#endif
3520 seq_putc(m, '\n');
3521}
3522
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523static void *s_start(struct seq_file *m, loff_t *pos)
3524{
3525 loff_t n = *pos;
3526 struct list_head *p;
3527
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003528 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003529 if (!n)
3530 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 p = cache_chain.next;
3532 while (n--) {
3533 p = p->next;
3534 if (p == &cache_chain)
3535 return NULL;
3536 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003537 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538}
3539
3540static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3541{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003542 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543 ++*pos;
3544 return cachep->next.next == &cache_chain ? NULL
Pekka Enberg343e0d72006-02-01 03:05:50 -08003545 : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546}
3547
3548static void s_stop(struct seq_file *m, void *p)
3549{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003550 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551}
3552
3553static int s_show(struct seq_file *m, void *p)
3554{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003555 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003557 struct slab *slabp;
3558 unsigned long active_objs;
3559 unsigned long num_objs;
3560 unsigned long active_slabs = 0;
3561 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003562 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003564 int node;
3565 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566
3567 check_irq_on();
3568 spin_lock_irq(&cachep->spinlock);
3569 active_objs = 0;
3570 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003571 for_each_online_node(node) {
3572 l3 = cachep->nodelists[node];
3573 if (!l3)
3574 continue;
3575
3576 spin_lock(&l3->list_lock);
3577
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003578 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003579 slabp = list_entry(q, struct slab, list);
3580 if (slabp->inuse != cachep->num && !error)
3581 error = "slabs_full accounting error";
3582 active_objs += cachep->num;
3583 active_slabs++;
3584 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003585 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003586 slabp = list_entry(q, struct slab, list);
3587 if (slabp->inuse == cachep->num && !error)
3588 error = "slabs_partial inuse accounting error";
3589 if (!slabp->inuse && !error)
3590 error = "slabs_partial/inuse accounting error";
3591 active_objs += slabp->inuse;
3592 active_slabs++;
3593 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003594 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003595 slabp = list_entry(q, struct slab, list);
3596 if (slabp->inuse && !error)
3597 error = "slabs_free/inuse accounting error";
3598 num_slabs++;
3599 }
3600 free_objects += l3->free_objects;
3601 shared_avail += l3->shared->avail;
3602
3603 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003605 num_slabs += active_slabs;
3606 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003607 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 error = "free_objects accounting error";
3609
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003610 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 if (error)
3612 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3613
3614 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003615 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003616 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003618 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003619 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003620 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003622 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 unsigned long high = cachep->high_mark;
3624 unsigned long allocs = cachep->num_allocations;
3625 unsigned long grown = cachep->grown;
3626 unsigned long reaped = cachep->reaped;
3627 unsigned long errors = cachep->errors;
3628 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003630 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631
Christoph Lametere498be72005-09-09 13:03:32 -07003632 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003633 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 }
3635 /* cpu stats */
3636 {
3637 unsigned long allochit = atomic_read(&cachep->allochit);
3638 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3639 unsigned long freehit = atomic_read(&cachep->freehit);
3640 unsigned long freemiss = atomic_read(&cachep->freemiss);
3641
3642 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003643 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 }
3645#endif
3646 seq_putc(m, '\n');
3647 spin_unlock_irq(&cachep->spinlock);
3648 return 0;
3649}
3650
3651/*
3652 * slabinfo_op - iterator that generates /proc/slabinfo
3653 *
3654 * Output layout:
3655 * cache-name
3656 * num-active-objs
3657 * total-objs
3658 * object size
3659 * num-active-slabs
3660 * total-slabs
3661 * num-pages-per-slab
3662 * + further values on SMP and with statistics enabled
3663 */
3664
3665struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003666 .start = s_start,
3667 .next = s_next,
3668 .stop = s_stop,
3669 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670};
3671
3672#define MAX_SLABINFO_WRITE 128
3673/**
3674 * slabinfo_write - Tuning for the slab allocator
3675 * @file: unused
3676 * @buffer: user buffer
3677 * @count: data length
3678 * @ppos: unused
3679 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003680ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3681 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003683 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684 int limit, batchcount, shared, res;
3685 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003686
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687 if (count > MAX_SLABINFO_WRITE)
3688 return -EINVAL;
3689 if (copy_from_user(&kbuf, buffer, count))
3690 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003691 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692
3693 tmp = strchr(kbuf, ' ');
3694 if (!tmp)
3695 return -EINVAL;
3696 *tmp = '\0';
3697 tmp++;
3698 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3699 return -EINVAL;
3700
3701 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003702 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003704 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003705 struct kmem_cache *cachep = list_entry(p, struct kmem_cache,
3706 next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707
3708 if (!strcmp(cachep->name, kbuf)) {
3709 if (limit < 1 ||
3710 batchcount < 1 ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003711 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003712 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003714 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003715 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 }
3717 break;
3718 }
3719 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003720 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 if (res >= 0)
3722 res = count;
3723 return res;
3724}
3725#endif
3726
Manfred Spraul00e145b2005-09-03 15:55:07 -07003727/**
3728 * ksize - get the actual amount of memory allocated for a given object
3729 * @objp: Pointer to the object
3730 *
3731 * kmalloc may internally round up allocations and return more memory
3732 * than requested. ksize() can be used to determine the actual amount of
3733 * memory allocated. The caller may use this additional memory, even though
3734 * a smaller amount of memory was initially specified with the kmalloc call.
3735 * The caller must guarantee that objp points to a valid object previously
3736 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3737 * must not be freed during the duration of the call.
3738 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739unsigned int ksize(const void *objp)
3740{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003741 if (unlikely(objp == NULL))
3742 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003744 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745}