blob: 1c8f5ee230d5dfeaf59fdcf5a1b6334cfa5f3691 [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 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * 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 | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800173 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
176 SLAB_DESTROY_BY_RCU)
177#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 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;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800269 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
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
307#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
310#define SIZE_AC 1
311#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametere498be72005-09-09 13:03:32 -0700313/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800314 * This function must be completely optimized away if a constant is passed to
315 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700316 */
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;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800347 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700348 spin_lock_init(&parent->list_lock);
349 parent->free_objects = 0;
350 parent->free_touched = 0;
351}
352
Andrew Mortona737b3e2006-03-22 00:08:11 -0800353#define MAKE_LIST(cachep, listp, slab, nodeid) \
354 do { \
355 INIT_LIST_HEAD(listp); \
356 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700357 } while (0)
358
Andrew Mortona737b3e2006-03-22 00:08:11 -0800359#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
360 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700361 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
364 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
366/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800367 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 *
369 * manages a cache.
370 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800371
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800372struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800374 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800375/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800376 unsigned int batchcount;
377 unsigned int limit;
378 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800379
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800380 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800381/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383
Andrew Mortona737b3e2006-03-22 00:08:11 -0800384 unsigned int flags; /* constant flags */
385 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800387/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800389 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
391 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800392 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800395 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800396 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800398 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800401 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800404 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800406/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 const char *name;
408 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800410/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800412 unsigned long num_active;
413 unsigned long num_allocations;
414 unsigned long high_mark;
415 unsigned long grown;
416 unsigned long reaped;
417 unsigned long errors;
418 unsigned long max_freeable;
419 unsigned long node_allocs;
420 unsigned long node_frees;
421 atomic_t allochit;
422 atomic_t allocmiss;
423 atomic_t freehit;
424 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425#endif
426#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800427 /*
428 * If debugging is enabled, then the allocator can add additional
429 * fields and/or padding to every object. buffer_size contains the total
430 * object size including these internal fields, the following two
431 * variables contain the offset to the user object and its size.
432 */
433 int obj_offset;
434 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435#endif
436};
437
438#define CFLGS_OFF_SLAB (0x80000000UL)
439#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
440
441#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800442/*
443 * Optimization question: fewer reaps means less probability for unnessary
444 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100446 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 * which could lock up otherwise freeable slabs.
448 */
449#define REAPTIMEOUT_CPUC (2*HZ)
450#define REAPTIMEOUT_LIST3 (4*HZ)
451
452#if STATS
453#define STATS_INC_ACTIVE(x) ((x)->num_active++)
454#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
455#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
456#define STATS_INC_GROWN(x) ((x)->grown++)
457#define STATS_INC_REAPED(x) ((x)->reaped++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800458#define STATS_SET_HIGH(x) \
459 do { \
460 if ((x)->num_active > (x)->high_mark) \
461 (x)->high_mark = (x)->num_active; \
462 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463#define STATS_INC_ERR(x) ((x)->errors++)
464#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700465#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800466#define STATS_SET_FREEABLE(x, i) \
467 do { \
468 if ((x)->max_freeable < i) \
469 (x)->max_freeable = i; \
470 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
472#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
473#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
474#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
475#else
476#define STATS_INC_ACTIVE(x) do { } while (0)
477#define STATS_DEC_ACTIVE(x) do { } while (0)
478#define STATS_INC_ALLOCED(x) do { } while (0)
479#define STATS_INC_GROWN(x) do { } while (0)
480#define STATS_INC_REAPED(x) do { } while (0)
481#define STATS_SET_HIGH(x) do { } while (0)
482#define STATS_INC_ERR(x) do { } while (0)
483#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700484#define STATS_INC_NODEFREES(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800485#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486#define STATS_INC_ALLOCHIT(x) do { } while (0)
487#define STATS_INC_ALLOCMISS(x) do { } while (0)
488#define STATS_INC_FREEHIT(x) do { } while (0)
489#define STATS_INC_FREEMISS(x) do { } while (0)
490#endif
491
492#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -0800493/*
494 * Magic nums for obj red zoning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700495 * Placed in the first word before and the first word after an obj.
496 */
497#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
498#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
499
500/* ...and for poisoning */
501#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
502#define POISON_FREE 0x6b /* for use-after-free poisoning */
503#define POISON_END 0xa5 /* end-byte of poisoning */
504
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505/*
506 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800518static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532}
533
Pekka Enberg343e0d72006-02-01 03:05:50 -0800534static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800538 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800539 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
Pekka Enberg343e0d72006-02-01 03:05:50 -0800543static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544{
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547}
548
549#else
550
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800551#define obj_offset(x) 0
552#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
556
557#endif
558
559/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
561 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 */
563#if defined(CONFIG_LARGE_ALLOCS)
564#define MAX_OBJ_ORDER 13 /* up to 32Mb */
565#define MAX_GFP_ORDER 13 /* up to 32Mb */
566#elif defined(CONFIG_MMU)
567#define MAX_OBJ_ORDER 5 /* 32 pages */
568#define MAX_GFP_ORDER 5 /* 32 pages */
569#else
570#define MAX_OBJ_ORDER 8 /* up to 1Mb */
571#define MAX_GFP_ORDER 8 /* up to 1Mb */
572#endif
573
574/*
575 * Do not go above this order unless 0 objects fit into the slab.
576 */
577#define BREAK_GFP_ORDER_HI 1
578#define BREAK_GFP_ORDER_LO 0
579static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
580
Andrew Mortona737b3e2006-03-22 00:08:11 -0800581/*
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800586static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587{
588 page->lru.next = (struct list_head *)cache;
589}
590
591static inline struct kmem_cache *page_get_cache(struct page *page)
592{
Nick Piggin84097512006-03-22 00:08:34 -0800593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800595 return (struct kmem_cache *)page->lru.next;
596}
597
598static inline void page_set_slab(struct page *page, struct slab *slab)
599{
600 page->lru.prev = (struct list_head *)slab;
601}
602
603static inline struct slab *page_get_slab(struct page *page)
604{
Nick Piggin84097512006-03-22 00:08:34 -0800605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800607 return (struct slab *)page->lru.prev;
608}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800610static inline struct kmem_cache *virt_to_cache(const void *obj)
611{
612 struct page *page = virt_to_page(obj);
613 return page_get_cache(page);
614}
615
616static inline struct slab *virt_to_slab(const void *obj)
617{
618 struct page *page = virt_to_page(obj);
619 return page_get_slab(page);
620}
621
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800622static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
623 unsigned int idx)
624{
625 return slab->s_mem + cache->buffer_size * idx;
626}
627
628static inline unsigned int obj_to_index(struct kmem_cache *cache,
629 struct slab *slab, void *obj)
630{
631 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
632}
633
Andrew Mortona737b3e2006-03-22 00:08:11 -0800634/*
635 * These are the default caches for kmalloc. Custom caches can have other sizes.
636 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637struct cache_sizes malloc_sizes[] = {
638#define CACHE(x) { .cs_size = (x) },
639#include <linux/kmalloc_sizes.h>
640 CACHE(ULONG_MAX)
641#undef CACHE
642};
643EXPORT_SYMBOL(malloc_sizes);
644
645/* Must match cache_sizes above. Out of line to keep cache footprint low. */
646struct cache_names {
647 char *name;
648 char *name_dma;
649};
650
651static struct cache_names __initdata cache_names[] = {
652#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
653#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800654 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655#undef CACHE
656};
657
658static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800659 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800664static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800665 .batchcount = 1,
666 .limit = BOOT_CPUCACHE_ENTRIES,
667 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800668 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800669 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800671 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672#endif
673};
674
675/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800676static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677static struct list_head cache_chain;
678
679/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800680 * vm_enough_memory() looks at this to determine how many slab-allocated pages
681 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 *
683 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
684 */
685atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
687/*
688 * chicken and egg problem: delay the per-cpu array allocation
689 * until the general caches are up.
690 */
691static enum {
692 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700693 PARTIAL_AC,
694 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 FULL
696} g_cpucache_up;
697
698static DEFINE_PER_CPU(struct work_struct, reap_work);
699
Andrew Mortona737b3e2006-03-22 00:08:11 -0800700static void free_block(struct kmem_cache *cachep, void **objpp, int len,
701 int node);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800702static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800703static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800704static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
Pekka Enberg343e0d72006-02-01 03:05:50 -0800706static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707{
708 return cachep->array[smp_processor_id()];
709}
710
Andrew Mortona737b3e2006-03-22 00:08:11 -0800711static inline struct kmem_cache *__find_general_cachep(size_t size,
712 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
714 struct cache_sizes *csizep = malloc_sizes;
715
716#if DEBUG
717 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800718 * kmem_cache_create(), or __kmalloc(), before
719 * the generic caches are initialized.
720 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700721 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722#endif
723 while (size > csizep->cs_size)
724 csizep++;
725
726 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700727 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 * has cs_{dma,}cachep==NULL. Thus no special case
729 * for large kmalloc calls required.
730 */
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
733 return csizep->cs_cachep;
734}
735
Pekka Enberg343e0d72006-02-01 03:05:50 -0800736struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700737{
738 return __find_general_cachep(size, gfpflags);
739}
740EXPORT_SYMBOL(kmem_find_general_cachep);
741
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800742static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
745}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746
Andrew Mortona737b3e2006-03-22 00:08:11 -0800747/*
748 * Calculate the number of objects and left-over bytes for a given buffer size.
749 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800750static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
752 unsigned int *num)
753{
754 int nr_objs;
755 size_t mgmt_size;
756 size_t slab_size = PAGE_SIZE << gfporder;
757
758 /*
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
761 * slab is used for:
762 *
763 * - The struct slab
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
767 *
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
772 */
773 if (flags & CFLGS_OFF_SLAB) {
774 mgmt_size = 0;
775 nr_objs = slab_size / buffer_size;
776
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
779 } else {
780 /*
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
786 * into account.
787 */
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
790
791 /*
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
794 */
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
796 > slab_size)
797 nr_objs--;
798
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
801
802 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800804 *num = nr_objs;
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806}
807
808#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
809
Andrew Mortona737b3e2006-03-22 00:08:11 -0800810static void __slab_error(const char *function, struct kmem_cache *cachep,
811 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812{
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800814 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 dump_stack();
816}
817
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800818#ifdef CONFIG_NUMA
819/*
820 * Special reaping functions for NUMA systems called from cache_reap().
821 * These take care of doing round robin flushing of alien caches (containing
822 * objects freed on different nodes from which they were allocated) and the
823 * flushing of remote pcps by calling drain_node_pages.
824 */
825static DEFINE_PER_CPU(unsigned long, reap_node);
826
827static void init_reap_node(int cpu)
828{
829 int node;
830
831 node = next_node(cpu_to_node(cpu), node_online_map);
832 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800833 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800834
835 __get_cpu_var(reap_node) = node;
836}
837
838static void next_reap_node(void)
839{
840 int node = __get_cpu_var(reap_node);
841
842 /*
843 * Also drain per cpu pages on remote zones
844 */
845 if (node != numa_node_id())
846 drain_node_pages(node);
847
848 node = next_node(node, node_online_map);
849 if (unlikely(node >= MAX_NUMNODES))
850 node = first_node(node_online_map);
851 __get_cpu_var(reap_node) = node;
852}
853
854#else
855#define init_reap_node(cpu) do { } while (0)
856#define next_reap_node(void) do { } while (0)
857#endif
858
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859/*
860 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
861 * via the workqueue/eventd.
862 * Add the CPU number into the expiration time to minimize the possibility of
863 * the CPUs getting into lockstep and contending for the global cache chain
864 * lock.
865 */
866static void __devinit start_cpu_timer(int cpu)
867{
868 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
869
870 /*
871 * When this gets called from do_initcalls via cpucache_init(),
872 * init_workqueues() has already run, so keventd will be setup
873 * at that time.
874 */
875 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800876 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 INIT_WORK(reap_work, cache_reap, NULL);
878 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
879 }
880}
881
Christoph Lametere498be72005-09-09 13:03:32 -0700882static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800883 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800885 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 struct array_cache *nc = NULL;
887
Christoph Lametere498be72005-09-09 13:03:32 -0700888 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 if (nc) {
890 nc->avail = 0;
891 nc->limit = entries;
892 nc->batchcount = batchcount;
893 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700894 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 }
896 return nc;
897}
898
Christoph Lametere498be72005-09-09 13:03:32 -0700899#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800900static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800901
Pekka Enberg5295a742006-02-01 03:05:48 -0800902static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700903{
904 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800905 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700906 int i;
907
908 if (limit > 1)
909 limit = 12;
910 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
911 if (ac_ptr) {
912 for_each_node(i) {
913 if (i == node || !node_online(i)) {
914 ac_ptr[i] = NULL;
915 continue;
916 }
917 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
918 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800919 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700920 kfree(ac_ptr[i]);
921 kfree(ac_ptr);
922 return NULL;
923 }
924 }
925 }
926 return ac_ptr;
927}
928
Pekka Enberg5295a742006-02-01 03:05:48 -0800929static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700930{
931 int i;
932
933 if (!ac_ptr)
934 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700935 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800936 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700937 kfree(ac_ptr);
938}
939
Pekka Enberg343e0d72006-02-01 03:05:50 -0800940static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800941 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700942{
943 struct kmem_list3 *rl3 = cachep->nodelists[node];
944
945 if (ac->avail) {
946 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700947 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700948 ac->avail = 0;
949 spin_unlock(&rl3->list_lock);
950 }
951}
952
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800953/*
954 * Called from cache_reap() to regularly drain alien caches round robin.
955 */
956static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
957{
958 int node = __get_cpu_var(reap_node);
959
960 if (l3->alien) {
961 struct array_cache *ac = l3->alien[node];
962 if (ac && ac->avail) {
963 spin_lock_irq(&ac->lock);
964 __drain_alien_cache(cachep, ac, node);
965 spin_unlock_irq(&ac->lock);
966 }
967 }
968}
969
Andrew Mortona737b3e2006-03-22 00:08:11 -0800970static void drain_alien_cache(struct kmem_cache *cachep,
971 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700972{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800973 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700974 struct array_cache *ac;
975 unsigned long flags;
976
977 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800978 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -0700979 if (ac) {
980 spin_lock_irqsave(&ac->lock, flags);
981 __drain_alien_cache(cachep, ac, i);
982 spin_unlock_irqrestore(&ac->lock, flags);
983 }
984 }
985}
986#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800987
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800988#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800989#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800990
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800991static inline struct array_cache **alloc_alien_cache(int node, int limit)
992{
993 return (struct array_cache **) 0x01020304ul;
994}
995
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800996static inline void free_alien_cache(struct array_cache **ac_ptr)
997{
998}
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800999
Christoph Lametere498be72005-09-09 13:03:32 -07001000#endif
1001
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001003 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004{
1005 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001006 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001007 struct kmem_list3 *l3 = NULL;
1008 int node = cpu_to_node(cpu);
1009 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
1011 switch (action) {
1012 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001013 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001014 /*
1015 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001016 * alloc_arraycache's are going to use this list.
1017 * kmalloc_node allows us to add the slab to the right
1018 * kmem_list3 and not this cpu's kmem_list3
1019 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020
Christoph Lametere498be72005-09-09 13:03:32 -07001021 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001022 /*
1023 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001024 * begin anything. Make sure some other cpu on this
1025 * node has not already allocated this
1026 */
1027 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001028 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1029 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001030 goto bad;
1031 kmem_list3_init(l3);
1032 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001033 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001034
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001035 /*
1036 * The l3s don't come and go as CPUs come and
1037 * go. cache_chain_mutex is sufficient
1038 * protection here.
1039 */
Christoph Lametere498be72005-09-09 13:03:32 -07001040 cachep->nodelists[node] = l3;
1041 }
1042
1043 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1044 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001045 (1 + nr_cpus_node(node)) *
1046 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001047 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1048 }
1049
Andrew Mortona737b3e2006-03-22 00:08:11 -08001050 /*
1051 * Now we can go ahead with allocating the shared arrays and
1052 * array caches
1053 */
Christoph Lametere498be72005-09-09 13:03:32 -07001054 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001055 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001056 struct array_cache *shared;
1057 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001058
Christoph Lametere498be72005-09-09 13:03:32 -07001059 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001060 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 if (!nc)
1062 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001063 shared = alloc_arraycache(node,
1064 cachep->shared * cachep->batchcount,
1065 0xbaadf00d);
1066 if (!shared)
1067 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001068
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001069 alien = alloc_alien_cache(node, cachep->limit);
1070 if (!alien)
1071 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001073 l3 = cachep->nodelists[node];
1074 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001075
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001076 spin_lock_irq(&l3->list_lock);
1077 if (!l3->shared) {
1078 /*
1079 * We are serialised from CPU_DEAD or
1080 * CPU_UP_CANCELLED by the cpucontrol lock
1081 */
1082 l3->shared = shared;
1083 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001084 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001085#ifdef CONFIG_NUMA
1086 if (!l3->alien) {
1087 l3->alien = alien;
1088 alien = NULL;
1089 }
1090#endif
1091 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001092 kfree(shared);
1093 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001095 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 break;
1097 case CPU_ONLINE:
1098 start_cpu_timer(cpu);
1099 break;
1100#ifdef CONFIG_HOTPLUG_CPU
1101 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001102 /*
1103 * Even if all the cpus of a node are down, we don't free the
1104 * kmem_list3 of any cache. This to avoid a race between
1105 * cpu_down, and a kmalloc allocation from another cpu for
1106 * memory from the node of the cpu going down. The list3
1107 * structure is usually allocated from kmem_cache_create() and
1108 * gets destroyed at kmem_cache_destroy().
1109 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 /* fall thru */
1111 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001112 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 list_for_each_entry(cachep, &cache_chain, next) {
1114 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001115 struct array_cache *shared;
1116 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001117 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
Christoph Lametere498be72005-09-09 13:03:32 -07001119 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 /* cpu is dead; no one can alloc from it. */
1121 nc = cachep->array[cpu];
1122 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001123 l3 = cachep->nodelists[node];
1124
1125 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001126 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001127
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001128 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001129
1130 /* Free limit for this kmem_list3 */
1131 l3->free_limit -= cachep->batchcount;
1132 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001133 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001134
1135 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001136 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001137 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001138 }
Christoph Lametere498be72005-09-09 13:03:32 -07001139
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001140 shared = l3->shared;
1141 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001142 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001143 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001144 l3->shared = NULL;
1145 }
Christoph Lametere498be72005-09-09 13:03:32 -07001146
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001147 alien = l3->alien;
1148 l3->alien = NULL;
1149
1150 spin_unlock_irq(&l3->list_lock);
1151
1152 kfree(shared);
1153 if (alien) {
1154 drain_alien_cache(cachep, alien);
1155 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001156 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001157free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 kfree(nc);
1159 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001160 /*
1161 * In the previous loop, all the objects were freed to
1162 * the respective cache's slabs, now we can go ahead and
1163 * shrink each nodelist to its limit.
1164 */
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 l3 = cachep->nodelists[node];
1167 if (!l3)
1168 continue;
1169 spin_lock_irq(&l3->list_lock);
1170 /* free slabs belonging to this node */
1171 __node_shrink(cachep, node);
1172 spin_unlock_irq(&l3->list_lock);
1173 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001174 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 break;
1176#endif
1177 }
1178 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001179bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001180 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 return NOTIFY_BAD;
1182}
1183
1184static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1185
Christoph Lametere498be72005-09-09 13:03:32 -07001186/*
1187 * swap the static kmem_list3 with kmalloced memory
1188 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001189static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1190 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001191{
1192 struct kmem_list3 *ptr;
1193
1194 BUG_ON(cachep->nodelists[nodeid] != list);
1195 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1196 BUG_ON(!ptr);
1197
1198 local_irq_disable();
1199 memcpy(ptr, list, sizeof(struct kmem_list3));
1200 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1201 cachep->nodelists[nodeid] = ptr;
1202 local_irq_enable();
1203}
1204
Andrew Mortona737b3e2006-03-22 00:08:11 -08001205/*
1206 * Initialisation. Called after the page allocator have been initialised and
1207 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 */
1209void __init kmem_cache_init(void)
1210{
1211 size_t left_over;
1212 struct cache_sizes *sizes;
1213 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001214 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001215 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001216
1217 for (i = 0; i < NUM_INIT_LISTS; i++) {
1218 kmem_list3_init(&initkmem_list3[i]);
1219 if (i < MAX_NUMNODES)
1220 cache_cache.nodelists[i] = NULL;
1221 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
1223 /*
1224 * Fragmentation resistance on low memory - only use bigger
1225 * page orders on machines with more than 32MB of memory.
1226 */
1227 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1228 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1229
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 /* Bootstrap is tricky, because several objects are allocated
1231 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001232 * 1) initialize the cache_cache cache: it contains the struct
1233 * kmem_cache structures of all caches, except cache_cache itself:
1234 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001235 * Initially an __init data area is used for the head array and the
1236 * kmem_list3 structures, it's replaced with a kmalloc allocated
1237 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001239 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001240 * An __init data area is used for the head array.
1241 * 3) Create the remaining kmalloc caches, with minimally sized
1242 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 * 4) Replace the __init data head arrays for cache_cache and the first
1244 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001245 * 5) Replace the __init data for kmem_list3 for cache_cache and
1246 * the other cache's with kmalloc allocated memory.
1247 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 */
1249
1250 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 INIT_LIST_HEAD(&cache_chain);
1252 list_add(&cache_cache.next, &cache_chain);
1253 cache_cache.colour_off = cache_line_size();
1254 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001255 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Andrew Mortona737b3e2006-03-22 00:08:11 -08001257 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1258 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Jack Steiner07ed76b2006-03-07 21:55:46 -08001260 for (order = 0; order < MAX_ORDER; order++) {
1261 cache_estimate(order, cache_cache.buffer_size,
1262 cache_line_size(), 0, &left_over, &cache_cache.num);
1263 if (cache_cache.num)
1264 break;
1265 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 if (!cache_cache.num)
1267 BUG();
Jack Steiner07ed76b2006-03-07 21:55:46 -08001268 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001269 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001270 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1271 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272
1273 /* 2+3) create the kmalloc caches */
1274 sizes = malloc_sizes;
1275 names = cache_names;
1276
Andrew Mortona737b3e2006-03-22 00:08:11 -08001277 /*
1278 * Initialize the caches that provide memory for the array cache and the
1279 * kmem_list3 structures first. Without this, further allocations will
1280 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001281 */
1282
1283 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001284 sizes[INDEX_AC].cs_size,
1285 ARCH_KMALLOC_MINALIGN,
1286 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1287 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001288
Andrew Mortona737b3e2006-03-22 00:08:11 -08001289 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001290 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001291 kmem_cache_create(names[INDEX_L3].name,
1292 sizes[INDEX_L3].cs_size,
1293 ARCH_KMALLOC_MINALIGN,
1294 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1295 NULL, NULL);
1296 }
Christoph Lametere498be72005-09-09 13:03:32 -07001297
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001299 /*
1300 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 * This should be particularly beneficial on SMP boxes, as it
1302 * eliminates "false sharing".
1303 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001304 * allow tighter packing of the smaller caches.
1305 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001306 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001307 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001308 sizes->cs_size,
1309 ARCH_KMALLOC_MINALIGN,
1310 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1311 NULL, NULL);
1312 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313
1314 /* Inc off-slab bufctl limit until the ceiling is hit. */
1315 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001316 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 offslab_limit /= sizeof(kmem_bufctl_t);
1318 }
1319
1320 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001321 sizes->cs_size,
1322 ARCH_KMALLOC_MINALIGN,
1323 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1324 SLAB_PANIC,
1325 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 sizes++;
1327 names++;
1328 }
1329 /* 4) Replace the bootstrap head arrays */
1330 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001331 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001332
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001334
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001336 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1337 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001338 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 cache_cache.array[smp_processor_id()] = ptr;
1340 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001341
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001343
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001345 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001346 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001347 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001348 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001349 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001350 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 local_irq_enable();
1352 }
Christoph Lametere498be72005-09-09 13:03:32 -07001353 /* 5) Replace the bootstrap kmem_list3's */
1354 {
1355 int node;
1356 /* Replace the static kmem_list3 structures for the boot cpu */
1357 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001358 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
Christoph Lametere498be72005-09-09 13:03:32 -07001360 for_each_online_node(node) {
1361 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001362 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001363
1364 if (INDEX_AC != INDEX_L3) {
1365 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001366 &initkmem_list3[SIZE_L3 + node],
1367 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001368 }
1369 }
1370 }
1371
1372 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001374 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001375 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001377 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001378 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 }
1380
1381 /* Done! */
1382 g_cpucache_up = FULL;
1383
Andrew Mortona737b3e2006-03-22 00:08:11 -08001384 /*
1385 * Register a cpu startup notifier callback that initializes
1386 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 */
1388 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389
Andrew Mortona737b3e2006-03-22 00:08:11 -08001390 /*
1391 * The reap timers are started later, with a module init call: That part
1392 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 */
1394}
1395
1396static int __init cpucache_init(void)
1397{
1398 int cpu;
1399
Andrew Mortona737b3e2006-03-22 00:08:11 -08001400 /*
1401 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 */
Christoph Lametere498be72005-09-09 13:03:32 -07001403 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001404 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 return 0;
1406}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407__initcall(cpucache_init);
1408
1409/*
1410 * Interface to system's page allocator. No need to hold the cache-lock.
1411 *
1412 * If we requested dmaable memory, we will get it. Even if we
1413 * did not request dmaable memory, we might get it, but that
1414 * would be relatively rare and ignorable.
1415 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001416static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417{
1418 struct page *page;
1419 void *addr;
1420 int i;
1421
1422 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001423 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 if (!page)
1425 return NULL;
1426 addr = page_address(page);
1427
1428 i = (1 << cachep->gfporder);
1429 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1430 atomic_add(i, &slab_reclaim_pages);
1431 add_page_state(nr_slab, i);
1432 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001433 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 page++;
1435 }
1436 return addr;
1437}
1438
1439/*
1440 * Interface to system's page release.
1441 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001442static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001444 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 struct page *page = virt_to_page(addr);
1446 const unsigned long nr_freed = i;
1447
1448 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001449 BUG_ON(!PageSlab(page));
1450 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 page++;
1452 }
1453 sub_page_state(nr_slab, nr_freed);
1454 if (current->reclaim_state)
1455 current->reclaim_state->reclaimed_slab += nr_freed;
1456 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001457 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1458 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459}
1460
1461static void kmem_rcu_free(struct rcu_head *head)
1462{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001463 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001464 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
1466 kmem_freepages(cachep, slab_rcu->addr);
1467 if (OFF_SLAB(cachep))
1468 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1469}
1470
1471#if DEBUG
1472
1473#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001474static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001475 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001477 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001479 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001481 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 return;
1483
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001484 *addr++ = 0x12345678;
1485 *addr++ = caller;
1486 *addr++ = smp_processor_id();
1487 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 {
1489 unsigned long *sptr = &caller;
1490 unsigned long svalue;
1491
1492 while (!kstack_end(sptr)) {
1493 svalue = *sptr++;
1494 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001495 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 size -= sizeof(unsigned long);
1497 if (size <= sizeof(unsigned long))
1498 break;
1499 }
1500 }
1501
1502 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001503 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504}
1505#endif
1506
Pekka Enberg343e0d72006-02-01 03:05:50 -08001507static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001509 int size = obj_size(cachep);
1510 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
1512 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001513 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514}
1515
1516static void dump_line(char *data, int offset, int limit)
1517{
1518 int i;
1519 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001520 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001521 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 printk("\n");
1523}
1524#endif
1525
1526#if DEBUG
1527
Pekka Enberg343e0d72006-02-01 03:05:50 -08001528static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529{
1530 int i, size;
1531 char *realobj;
1532
1533 if (cachep->flags & SLAB_RED_ZONE) {
1534 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001535 *dbg_redzone1(cachep, objp),
1536 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 }
1538
1539 if (cachep->flags & SLAB_STORE_USER) {
1540 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001541 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001543 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 printk("\n");
1545 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001546 realobj = (char *)objp + obj_offset(cachep);
1547 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001548 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549 int limit;
1550 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001551 if (i + limit > size)
1552 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 dump_line(realobj, i, limit);
1554 }
1555}
1556
Pekka Enberg343e0d72006-02-01 03:05:50 -08001557static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558{
1559 char *realobj;
1560 int size, i;
1561 int lines = 0;
1562
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001563 realobj = (char *)objp + obj_offset(cachep);
1564 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001566 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001568 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 exp = POISON_END;
1570 if (realobj[i] != exp) {
1571 int limit;
1572 /* Mismatch ! */
1573 /* Print header */
1574 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001575 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001576 "Slab corruption: start=%p, len=%d\n",
1577 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 print_objinfo(cachep, objp, 0);
1579 }
1580 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001581 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001583 if (i + limit > size)
1584 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 dump_line(realobj, i, limit);
1586 i += 16;
1587 lines++;
1588 /* Limit to 5 lines */
1589 if (lines > 5)
1590 break;
1591 }
1592 }
1593 if (lines != 0) {
1594 /* Print some data about the neighboring objects, if they
1595 * exist:
1596 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001597 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001598 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001600 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001602 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001603 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001605 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 print_objinfo(cachep, objp, 2);
1607 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001608 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001609 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001610 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001612 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 print_objinfo(cachep, objp, 2);
1614 }
1615 }
1616}
1617#endif
1618
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001620/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001621 * slab_destroy_objs - destroy a slab and its objects
1622 * @cachep: cache pointer being destroyed
1623 * @slabp: slab pointer being destroyed
1624 *
1625 * Call the registered destructor for each object in a slab that is being
1626 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001627 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001628static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001629{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 int i;
1631 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001632 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633
1634 if (cachep->flags & SLAB_POISON) {
1635#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001636 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1637 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001638 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001639 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 else
1641 check_poison_obj(cachep, objp);
1642#else
1643 check_poison_obj(cachep, objp);
1644#endif
1645 }
1646 if (cachep->flags & SLAB_RED_ZONE) {
1647 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1648 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001649 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1651 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001652 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 }
1654 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001655 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001657}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001659static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001660{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 if (cachep->dtor) {
1662 int i;
1663 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001664 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001665 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 }
1667 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001668}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669#endif
1670
Randy Dunlap911851e2006-03-22 00:08:14 -08001671/**
1672 * slab_destroy - destroy and release all objects in a slab
1673 * @cachep: cache pointer being destroyed
1674 * @slabp: slab pointer being destroyed
1675 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001676 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001677 * Before calling the slab must have been unlinked from the cache. The
1678 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001679 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001680static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001681{
1682 void *addr = slabp->s_mem - slabp->colouroff;
1683
1684 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1686 struct slab_rcu *slab_rcu;
1687
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001688 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 slab_rcu->cachep = cachep;
1690 slab_rcu->addr = addr;
1691 call_rcu(&slab_rcu->head, kmem_rcu_free);
1692 } else {
1693 kmem_freepages(cachep, addr);
1694 if (OFF_SLAB(cachep))
1695 kmem_cache_free(cachep->slabp_cache, slabp);
1696 }
1697}
1698
Andrew Mortona737b3e2006-03-22 00:08:11 -08001699/*
1700 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1701 * size of kmem_list3.
1702 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001703static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001704{
1705 int node;
1706
1707 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001708 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001709 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001710 REAPTIMEOUT_LIST3 +
1711 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001712 }
1713}
1714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001716 * calculate_slab_order - calculate size (page order) of slabs
1717 * @cachep: pointer to the cache that is being created
1718 * @size: size of objects to be created in this cache.
1719 * @align: required alignment for the objects.
1720 * @flags: slab allocation flags
1721 *
1722 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001723 *
1724 * This could be made much more intelligent. For now, try to avoid using
1725 * high order pages for slabs. When the gfp() functions are more friendly
1726 * towards high-order requests, this should be changed.
1727 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001728static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001729 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001730{
1731 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001732 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001733
Andrew Mortona737b3e2006-03-22 00:08:11 -08001734 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001735 unsigned int num;
1736 size_t remainder;
1737
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001738 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001739 if (!num)
1740 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001741
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001742 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001743 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001744 break;
1745
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001746 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001747 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001748 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001749 left_over = remainder;
1750
1751 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001752 * A VFS-reclaimable slab tends to have most allocations
1753 * as GFP_NOFS and we really don't want to have to be allocating
1754 * higher-order pages when we are unable to shrink dcache.
1755 */
1756 if (flags & SLAB_RECLAIM_ACCOUNT)
1757 break;
1758
1759 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001760 * Large number of objects is good, but very large slabs are
1761 * currently bad for the gfp()s.
1762 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001763 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001764 break;
1765
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001766 /*
1767 * Acceptable internal fragmentation?
1768 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001769 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001770 break;
1771 }
1772 return left_over;
1773}
1774
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001775static void setup_cpu_cache(struct kmem_cache *cachep)
1776{
1777 if (g_cpucache_up == FULL) {
1778 enable_cpucache(cachep);
1779 return;
1780 }
1781 if (g_cpucache_up == NONE) {
1782 /*
1783 * Note: the first kmem_cache_create must create the cache
1784 * that's used by kmalloc(24), otherwise the creation of
1785 * further caches will BUG().
1786 */
1787 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1788
1789 /*
1790 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1791 * the first cache, then we need to set up all its list3s,
1792 * otherwise the creation of further caches will BUG().
1793 */
1794 set_up_list3s(cachep, SIZE_AC);
1795 if (INDEX_AC == INDEX_L3)
1796 g_cpucache_up = PARTIAL_L3;
1797 else
1798 g_cpucache_up = PARTIAL_AC;
1799 } else {
1800 cachep->array[smp_processor_id()] =
1801 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1802
1803 if (g_cpucache_up == PARTIAL_AC) {
1804 set_up_list3s(cachep, SIZE_L3);
1805 g_cpucache_up = PARTIAL_L3;
1806 } else {
1807 int node;
1808 for_each_online_node(node) {
1809 cachep->nodelists[node] =
1810 kmalloc_node(sizeof(struct kmem_list3),
1811 GFP_KERNEL, node);
1812 BUG_ON(!cachep->nodelists[node]);
1813 kmem_list3_init(cachep->nodelists[node]);
1814 }
1815 }
1816 }
1817 cachep->nodelists[numa_node_id()]->next_reap =
1818 jiffies + REAPTIMEOUT_LIST3 +
1819 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1820
1821 cpu_cache_get(cachep)->avail = 0;
1822 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1823 cpu_cache_get(cachep)->batchcount = 1;
1824 cpu_cache_get(cachep)->touched = 0;
1825 cachep->batchcount = 1;
1826 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1827}
1828
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001829/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 * kmem_cache_create - Create a cache.
1831 * @name: A string which is used in /proc/slabinfo to identify this cache.
1832 * @size: The size of objects to be created in this cache.
1833 * @align: The required alignment for the objects.
1834 * @flags: SLAB flags
1835 * @ctor: A constructor for the objects.
1836 * @dtor: A destructor for the objects.
1837 *
1838 * Returns a ptr to the cache on success, NULL on failure.
1839 * Cannot be called within a int, but can be interrupted.
1840 * The @ctor is run when new pages are allocated by the cache
1841 * and the @dtor is run before the pages are handed back.
1842 *
1843 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001844 * the module calling this has to destroy the cache before getting unloaded.
1845 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 * The flags are
1847 *
1848 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1849 * to catch references to uninitialised memory.
1850 *
1851 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1852 * for buffer overruns.
1853 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1855 * cacheline. This can be beneficial if you're counting cycles as closely
1856 * as davem.
1857 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001858struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001860 unsigned long flags,
1861 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001862 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863{
1864 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001865 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001866 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867
1868 /*
1869 * Sanity checks... these are all serious usage bugs.
1870 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001871 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001872 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001873 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1874 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 BUG();
1876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001878 /*
1879 * Prevent CPUs from coming and going.
1880 * lock_cpu_hotplug() nests outside cache_chain_mutex
1881 */
1882 lock_cpu_hotplug();
1883
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001884 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001885
1886 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001887 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001888 mm_segment_t old_fs = get_fs();
1889 char tmp;
1890 int res;
1891
1892 /*
1893 * This happens when the module gets unloaded and doesn't
1894 * destroy its slab cache and no-one else reuses the vmalloc
1895 * area of the module. Print a warning.
1896 */
1897 set_fs(KERNEL_DS);
1898 res = __get_user(tmp, pc->name);
1899 set_fs(old_fs);
1900 if (res) {
1901 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001902 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001903 continue;
1904 }
1905
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001906 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001907 printk("kmem_cache_create: duplicate cache %s\n", name);
1908 dump_stack();
1909 goto oops;
1910 }
1911 }
1912
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913#if DEBUG
1914 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1915 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1916 /* No constructor, but inital state check requested */
1917 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001918 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 flags &= ~SLAB_DEBUG_INITIAL;
1920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921#if FORCED_DEBUG
1922 /*
1923 * Enable redzoning and last user accounting, except for caches with
1924 * large objects, if the increased size would increase the object size
1925 * above the next power of two: caches with object sizes just above a
1926 * power of two have a significant amount of internal fragmentation.
1927 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001928 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001929 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 if (!(flags & SLAB_DESTROY_BY_RCU))
1931 flags |= SLAB_POISON;
1932#endif
1933 if (flags & SLAB_DESTROY_BY_RCU)
1934 BUG_ON(flags & SLAB_POISON);
1935#endif
1936 if (flags & SLAB_DESTROY_BY_RCU)
1937 BUG_ON(dtor);
1938
1939 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001940 * Always checks flags, a caller might be expecting debug support which
1941 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 */
1943 if (flags & ~CREATE_MASK)
1944 BUG();
1945
Andrew Mortona737b3e2006-03-22 00:08:11 -08001946 /*
1947 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948 * unaligned accesses for some archs when redzoning is used, and makes
1949 * sure any on-slab bufctl's are also correctly aligned.
1950 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001951 if (size & (BYTES_PER_WORD - 1)) {
1952 size += (BYTES_PER_WORD - 1);
1953 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
1955
Andrew Mortona737b3e2006-03-22 00:08:11 -08001956 /* calculate the final buffer alignment: */
1957
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 /* 1) arch recommendation: can be overridden for debug */
1959 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001960 /*
1961 * Default alignment: as specified by the arch code. Except if
1962 * an object is really small, then squeeze multiple objects into
1963 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 */
1965 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001966 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 ralign /= 2;
1968 } else {
1969 ralign = BYTES_PER_WORD;
1970 }
1971 /* 2) arch mandated alignment: disables debug if necessary */
1972 if (ralign < ARCH_SLAB_MINALIGN) {
1973 ralign = ARCH_SLAB_MINALIGN;
1974 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001975 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 }
1977 /* 3) caller mandated alignment: disables debug if necessary */
1978 if (ralign < align) {
1979 ralign = align;
1980 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001981 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08001983 /*
1984 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 * the alignment to BYTES_PER_WORD.
1986 */
1987 align = ralign;
1988
1989 /* Get cache's description obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001990 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001992 goto oops;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001993 memset(cachep, 0, sizeof(struct kmem_cache));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994
1995#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001996 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
1998 if (flags & SLAB_RED_ZONE) {
1999 /* redzoning only works with word aligned caches */
2000 align = BYTES_PER_WORD;
2001
2002 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002003 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002004 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 }
2006 if (flags & SLAB_STORE_USER) {
2007 /* user store requires word alignment and
2008 * one word storage behind the end of the real
2009 * object.
2010 */
2011 align = BYTES_PER_WORD;
2012 size += BYTES_PER_WORD;
2013 }
2014#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002015 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002016 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2017 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 size = PAGE_SIZE;
2019 }
2020#endif
2021#endif
2022
2023 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002024 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 /*
2026 * Size is large, assume best to place the slab management obj
2027 * off-slab (should allow better packing of objs).
2028 */
2029 flags |= CFLGS_OFF_SLAB;
2030
2031 size = ALIGN(size, align);
2032
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002033 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034
2035 if (!cachep->num) {
2036 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2037 kmem_cache_free(&cache_cache, cachep);
2038 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002039 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002041 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2042 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
2044 /*
2045 * If the slab has been placed off-slab, and we have enough space then
2046 * move it on-slab. This is at the expense of any extra colouring.
2047 */
2048 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2049 flags &= ~CFLGS_OFF_SLAB;
2050 left_over -= slab_size;
2051 }
2052
2053 if (flags & CFLGS_OFF_SLAB) {
2054 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002055 slab_size =
2056 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 }
2058
2059 cachep->colour_off = cache_line_size();
2060 /* Offset must be a multiple of the alignment. */
2061 if (cachep->colour_off < align)
2062 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002063 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 cachep->slab_size = slab_size;
2065 cachep->flags = flags;
2066 cachep->gfpflags = 0;
2067 if (flags & SLAB_CACHE_DMA)
2068 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002069 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070
2071 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002072 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 cachep->ctor = ctor;
2074 cachep->dtor = dtor;
2075 cachep->name = name;
2076
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002078 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 /* cache setup completed, link it into the list */
2081 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002082oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 if (!cachep && (flags & SLAB_PANIC))
2084 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002085 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002086 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002087 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 return cachep;
2089}
2090EXPORT_SYMBOL(kmem_cache_create);
2091
2092#if DEBUG
2093static void check_irq_off(void)
2094{
2095 BUG_ON(!irqs_disabled());
2096}
2097
2098static void check_irq_on(void)
2099{
2100 BUG_ON(irqs_disabled());
2101}
2102
Pekka Enberg343e0d72006-02-01 03:05:50 -08002103static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
2105#ifdef CONFIG_SMP
2106 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002107 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108#endif
2109}
Christoph Lametere498be72005-09-09 13:03:32 -07002110
Pekka Enberg343e0d72006-02-01 03:05:50 -08002111static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002112{
2113#ifdef CONFIG_SMP
2114 check_irq_off();
2115 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2116#endif
2117}
2118
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119#else
2120#define check_irq_off() do { } while(0)
2121#define check_irq_on() do { } while(0)
2122#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002123#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124#endif
2125
Christoph Lameteraab22072006-03-22 00:09:06 -08002126static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2127 struct array_cache *ac,
2128 int force, int node);
2129
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130static void do_drain(void *arg)
2131{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002132 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002134 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135
2136 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002137 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002138 spin_lock(&cachep->nodelists[node]->list_lock);
2139 free_block(cachep, ac->entry, ac->avail, node);
2140 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 ac->avail = 0;
2142}
2143
Pekka Enberg343e0d72006-02-01 03:05:50 -08002144static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145{
Christoph Lametere498be72005-09-09 13:03:32 -07002146 struct kmem_list3 *l3;
2147 int node;
2148
Andrew Mortona07fa392006-03-22 00:08:17 -08002149 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002151 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002152 l3 = cachep->nodelists[node];
2153 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002154 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002155 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002156 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002157 }
2158 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159}
2160
Pekka Enberg343e0d72006-02-01 03:05:50 -08002161static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
2163 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002164 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 int ret;
2166
Christoph Lametere498be72005-09-09 13:03:32 -07002167 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 struct list_head *p;
2169
Christoph Lametere498be72005-09-09 13:03:32 -07002170 p = l3->slabs_free.prev;
2171 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 break;
2173
Christoph Lametere498be72005-09-09 13:03:32 -07002174 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175#if DEBUG
2176 if (slabp->inuse)
2177 BUG();
2178#endif
2179 list_del(&slabp->list);
2180
Christoph Lametere498be72005-09-09 13:03:32 -07002181 l3->free_objects -= cachep->num;
2182 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002184 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002186 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 return ret;
2188}
2189
Pekka Enberg343e0d72006-02-01 03:05:50 -08002190static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002191{
2192 int ret = 0, i = 0;
2193 struct kmem_list3 *l3;
2194
2195 drain_cpu_caches(cachep);
2196
2197 check_irq_on();
2198 for_each_online_node(i) {
2199 l3 = cachep->nodelists[i];
2200 if (l3) {
2201 spin_lock_irq(&l3->list_lock);
2202 ret += __node_shrink(cachep, i);
2203 spin_unlock_irq(&l3->list_lock);
2204 }
2205 }
2206 return (ret ? 1 : 0);
2207}
2208
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209/**
2210 * kmem_cache_shrink - Shrink a cache.
2211 * @cachep: The cache to shrink.
2212 *
2213 * Releases as many slabs as possible for a cache.
2214 * To help debugging, a zero exit status indicates all slabs were released.
2215 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002216int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217{
2218 if (!cachep || in_interrupt())
2219 BUG();
2220
2221 return __cache_shrink(cachep);
2222}
2223EXPORT_SYMBOL(kmem_cache_shrink);
2224
2225/**
2226 * kmem_cache_destroy - delete a cache
2227 * @cachep: the cache to destroy
2228 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002229 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 * Returns 0 on success.
2231 *
2232 * It is expected this function will be called by a module when it is
2233 * unloaded. This will remove the cache completely, and avoid a duplicate
2234 * cache being allocated each time a module is loaded and unloaded, if the
2235 * module doesn't have persistent in-kernel storage across loads and unloads.
2236 *
2237 * The cache must be empty before calling this function.
2238 *
2239 * The caller must guarantee that noone will allocate memory from the cache
2240 * during the kmem_cache_destroy().
2241 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002242int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243{
2244 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002245 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246
2247 if (!cachep || in_interrupt())
2248 BUG();
2249
2250 /* Don't let CPUs to come and go */
2251 lock_cpu_hotplug();
2252
2253 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002254 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 /*
2256 * the chain is never empty, cache_cache is never destroyed
2257 */
2258 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002259 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
2261 if (__cache_shrink(cachep)) {
2262 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002263 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002264 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002265 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 unlock_cpu_hotplug();
2267 return 1;
2268 }
2269
2270 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002271 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
Christoph Lametere498be72005-09-09 13:03:32 -07002273 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002274 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
2276 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002277 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002278 l3 = cachep->nodelists[i];
2279 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002280 kfree(l3->shared);
2281 free_alien_cache(l3->alien);
2282 kfree(l3);
2283 }
2284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 return 0;
2288}
2289EXPORT_SYMBOL(kmem_cache_destroy);
2290
2291/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002292static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002293 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294{
2295 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002296
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 if (OFF_SLAB(cachep)) {
2298 /* Slab management obj is off-slab. */
2299 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2300 if (!slabp)
2301 return NULL;
2302 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002303 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 colour_off += cachep->slab_size;
2305 }
2306 slabp->inuse = 0;
2307 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002308 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 return slabp;
2310}
2311
2312static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2313{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002314 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315}
2316
Pekka Enberg343e0d72006-02-01 03:05:50 -08002317static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002318 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319{
2320 int i;
2321
2322 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002323 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324#if DEBUG
2325 /* need to poison the objs? */
2326 if (cachep->flags & SLAB_POISON)
2327 poison_obj(cachep, objp, POISON_FREE);
2328 if (cachep->flags & SLAB_STORE_USER)
2329 *dbg_userword(cachep, objp) = NULL;
2330
2331 if (cachep->flags & SLAB_RED_ZONE) {
2332 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2333 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2334 }
2335 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002336 * Constructors are not allowed to allocate memory from the same
2337 * cache which they are a constructor for. Otherwise, deadlock.
2338 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 */
2340 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002341 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002342 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343
2344 if (cachep->flags & SLAB_RED_ZONE) {
2345 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2346 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002347 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2349 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002350 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002352 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2353 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002354 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002355 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356#else
2357 if (cachep->ctor)
2358 cachep->ctor(objp, cachep, ctor_flags);
2359#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002360 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002362 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 slabp->free = 0;
2364}
2365
Pekka Enberg343e0d72006-02-01 03:05:50 -08002366static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002368 if (flags & SLAB_DMA)
2369 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2370 else
2371 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372}
2373
Andrew Mortona737b3e2006-03-22 00:08:11 -08002374static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2375 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002376{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002377 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002378 kmem_bufctl_t next;
2379
2380 slabp->inuse++;
2381 next = slab_bufctl(slabp)[slabp->free];
2382#if DEBUG
2383 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2384 WARN_ON(slabp->nodeid != nodeid);
2385#endif
2386 slabp->free = next;
2387
2388 return objp;
2389}
2390
Andrew Mortona737b3e2006-03-22 00:08:11 -08002391static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2392 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002393{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002394 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002395
2396#if DEBUG
2397 /* Verify that the slab belongs to the intended node */
2398 WARN_ON(slabp->nodeid != nodeid);
2399
2400 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2401 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002402 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002403 BUG();
2404 }
2405#endif
2406 slab_bufctl(slabp)[objnr] = slabp->free;
2407 slabp->free = objnr;
2408 slabp->inuse--;
2409}
2410
Andrew Mortona737b3e2006-03-22 00:08:11 -08002411static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2412 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413{
2414 int i;
2415 struct page *page;
2416
2417 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002419
2420 i = 1;
2421 if (likely(!PageCompound(page)))
2422 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002424 page_set_cache(page, cachep);
2425 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 page++;
2427 } while (--i);
2428}
2429
2430/*
2431 * Grow (by 1) the number of slabs within a cache. This is called by
2432 * kmem_cache_alloc() when there are no active objs left in a cache.
2433 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002434static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002436 struct slab *slabp;
2437 void *objp;
2438 size_t offset;
2439 gfp_t local_flags;
2440 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002441 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442
Andrew Mortona737b3e2006-03-22 00:08:11 -08002443 /*
2444 * Be lazy and only check for valid flags here, keeping it out of the
2445 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002447 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 BUG();
2449 if (flags & SLAB_NO_GROW)
2450 return 0;
2451
2452 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2453 local_flags = (flags & SLAB_LEVEL_MASK);
2454 if (!(local_flags & __GFP_WAIT))
2455 /*
2456 * Not allowed to sleep. Need to tell a constructor about
2457 * this - it might need to know...
2458 */
2459 ctor_flags |= SLAB_CTOR_ATOMIC;
2460
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002461 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002463 l3 = cachep->nodelists[nodeid];
2464 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
2466 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002467 offset = l3->colour_next;
2468 l3->colour_next++;
2469 if (l3->colour_next >= cachep->colour)
2470 l3->colour_next = 0;
2471 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002473 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474
2475 if (local_flags & __GFP_WAIT)
2476 local_irq_enable();
2477
2478 /*
2479 * The test for missing atomic flag is performed here, rather than
2480 * the more obvious place, simply to reduce the critical path length
2481 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2482 * will eventually be caught here (where it matters).
2483 */
2484 kmem_flagcheck(cachep, flags);
2485
Andrew Mortona737b3e2006-03-22 00:08:11 -08002486 /*
2487 * Get mem for the objs. Attempt to allocate a physical page from
2488 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002489 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002490 objp = kmem_getpages(cachep, flags, nodeid);
2491 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 goto failed;
2493
2494 /* Get slab management. */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002495 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags);
2496 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 goto opps1;
2498
Christoph Lametere498be72005-09-09 13:03:32 -07002499 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 set_slab_attr(cachep, slabp, objp);
2501
2502 cache_init_objs(cachep, slabp, ctor_flags);
2503
2504 if (local_flags & __GFP_WAIT)
2505 local_irq_disable();
2506 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002507 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
2509 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002510 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002512 l3->free_objects += cachep->num;
2513 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002515opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002517failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 if (local_flags & __GFP_WAIT)
2519 local_irq_disable();
2520 return 0;
2521}
2522
2523#if DEBUG
2524
2525/*
2526 * Perform extra freeing checks:
2527 * - detect bad pointers.
2528 * - POISON/RED_ZONE checking
2529 * - destructor calls, for caches with POISON+dtor
2530 */
2531static void kfree_debugcheck(const void *objp)
2532{
2533 struct page *page;
2534
2535 if (!virt_addr_valid(objp)) {
2536 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002537 (unsigned long)objp);
2538 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
2540 page = virt_to_page(objp);
2541 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002542 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2543 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 BUG();
2545 }
2546}
2547
Pekka Enberg343e0d72006-02-01 03:05:50 -08002548static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002549 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550{
2551 struct page *page;
2552 unsigned int objnr;
2553 struct slab *slabp;
2554
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002555 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 kfree_debugcheck(objp);
2557 page = virt_to_page(objp);
2558
Pekka Enberg065d41c2005-11-13 16:06:46 -08002559 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002560 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2561 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002562 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002564 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2565 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 WARN_ON(1);
2567 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002568 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569
2570 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002571 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2572 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2573 slab_error(cachep, "double free, or memory outside"
2574 " object was overwritten");
2575 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2576 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002577 objp, *dbg_redzone1(cachep, objp),
2578 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 }
2580 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2581 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2582 }
2583 if (cachep->flags & SLAB_STORE_USER)
2584 *dbg_userword(cachep, objp) = caller;
2585
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002586 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
2588 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002589 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590
2591 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002592 /*
2593 * Need to call the slab's constructor so the caller can
2594 * perform a verify of its state (debugging). Called without
2595 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002597 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002598 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 }
2600 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2601 /* we want to cache poison the object,
2602 * call the destruction callback
2603 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002604 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 }
2606 if (cachep->flags & SLAB_POISON) {
2607#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002608 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002611 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 } else {
2613 poison_obj(cachep, objp, POISON_FREE);
2614 }
2615#else
2616 poison_obj(cachep, objp, POISON_FREE);
2617#endif
2618 }
2619 return objp;
2620}
2621
Pekka Enberg343e0d72006-02-01 03:05:50 -08002622static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623{
2624 kmem_bufctl_t i;
2625 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 /* Check slab's freelist to see if this obj is there. */
2628 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2629 entries++;
2630 if (entries > cachep->num || i >= cachep->num)
2631 goto bad;
2632 }
2633 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002634bad:
2635 printk(KERN_ERR "slab: Internal list corruption detected in "
2636 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2637 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002639 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002641 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002643 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
2645 printk("\n");
2646 BUG();
2647 }
2648}
2649#else
2650#define kfree_debugcheck(x) do { } while(0)
2651#define cache_free_debugcheck(x,objp,z) (objp)
2652#define check_slabp(x,y) do { } while(0)
2653#endif
2654
Pekka Enberg343e0d72006-02-01 03:05:50 -08002655static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656{
2657 int batchcount;
2658 struct kmem_list3 *l3;
2659 struct array_cache *ac;
2660
2661 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002662 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002663retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 batchcount = ac->batchcount;
2665 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002666 /*
2667 * If there was little recent activity on this cache, then
2668 * perform only a partial refill. Otherwise we could generate
2669 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 */
2671 batchcount = BATCHREFILL_LIMIT;
2672 }
Christoph Lametere498be72005-09-09 13:03:32 -07002673 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674
Christoph Lametere498be72005-09-09 13:03:32 -07002675 BUG_ON(ac->avail > 0 || !l3);
2676 spin_lock(&l3->list_lock);
2677
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 if (l3->shared) {
2679 struct array_cache *shared_array = l3->shared;
2680 if (shared_array->avail) {
2681 if (batchcount > shared_array->avail)
2682 batchcount = shared_array->avail;
2683 shared_array->avail -= batchcount;
2684 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002685 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002686 &(shared_array->entry[shared_array->avail]),
2687 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 shared_array->touched = 1;
2689 goto alloc_done;
2690 }
2691 }
2692 while (batchcount > 0) {
2693 struct list_head *entry;
2694 struct slab *slabp;
2695 /* Get slab alloc is to come from. */
2696 entry = l3->slabs_partial.next;
2697 if (entry == &l3->slabs_partial) {
2698 l3->free_touched = 1;
2699 entry = l3->slabs_free.next;
2700 if (entry == &l3->slabs_free)
2701 goto must_grow;
2702 }
2703
2704 slabp = list_entry(entry, struct slab, list);
2705 check_slabp(cachep, slabp);
2706 check_spinlock_acquired(cachep);
2707 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 STATS_INC_ALLOCED(cachep);
2709 STATS_INC_ACTIVE(cachep);
2710 STATS_SET_HIGH(cachep);
2711
Matthew Dobson78d382d2006-02-01 03:05:47 -08002712 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2713 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 }
2715 check_slabp(cachep, slabp);
2716
2717 /* move slabp to correct slabp list: */
2718 list_del(&slabp->list);
2719 if (slabp->free == BUFCTL_END)
2720 list_add(&slabp->list, &l3->slabs_full);
2721 else
2722 list_add(&slabp->list, &l3->slabs_partial);
2723 }
2724
Andrew Mortona737b3e2006-03-22 00:08:11 -08002725must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002727alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002728 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729
2730 if (unlikely(!ac->avail)) {
2731 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002732 x = cache_grow(cachep, flags, numa_node_id());
2733
Andrew Mortona737b3e2006-03-22 00:08:11 -08002734 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002735 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002736 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 return NULL;
2738
Andrew Mortona737b3e2006-03-22 00:08:11 -08002739 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 goto retry;
2741 }
2742 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002743 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744}
2745
Andrew Mortona737b3e2006-03-22 00:08:11 -08002746static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2747 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748{
2749 might_sleep_if(flags & __GFP_WAIT);
2750#if DEBUG
2751 kmem_flagcheck(cachep, flags);
2752#endif
2753}
2754
2755#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002756static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2757 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002759 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002761 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002763 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002764 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002765 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 else
2767 check_poison_obj(cachep, objp);
2768#else
2769 check_poison_obj(cachep, objp);
2770#endif
2771 poison_obj(cachep, objp, POISON_INUSE);
2772 }
2773 if (cachep->flags & SLAB_STORE_USER)
2774 *dbg_userword(cachep, objp) = caller;
2775
2776 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002777 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2778 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2779 slab_error(cachep, "double free, or memory outside"
2780 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002781 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002782 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2783 objp, *dbg_redzone1(cachep, objp),
2784 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785 }
2786 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2787 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2788 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002789 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002791 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792
2793 if (!(flags & __GFP_WAIT))
2794 ctor_flags |= SLAB_CTOR_ATOMIC;
2795
2796 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002797 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 return objp;
2799}
2800#else
2801#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2802#endif
2803
Pekka Enberg343e0d72006-02-01 03:05:50 -08002804static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002806 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 struct array_cache *ac;
2808
Christoph Lameterdc85da12006-01-18 17:42:36 -08002809#ifdef CONFIG_NUMA
Christoph Lameter86c562a2006-01-18 17:42:37 -08002810 if (unlikely(current->mempolicy && !in_interrupt())) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08002811 int nid = slab_node(current->mempolicy);
2812
2813 if (nid != numa_node_id())
2814 return __cache_alloc_node(cachep, flags, nid);
2815 }
2816#endif
2817
Alok N Kataria5c382302005-09-27 21:45:46 -07002818 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002819 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 if (likely(ac->avail)) {
2821 STATS_INC_ALLOCHIT(cachep);
2822 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002823 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 } else {
2825 STATS_INC_ALLOCMISS(cachep);
2826 objp = cache_alloc_refill(cachep, flags);
2827 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002828 return objp;
2829}
2830
Andrew Mortona737b3e2006-03-22 00:08:11 -08002831static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2832 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002833{
2834 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002835 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002836
2837 cache_alloc_debugcheck_before(cachep, flags);
2838
2839 local_irq_save(save_flags);
2840 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002842 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002843 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002844 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 return objp;
2846}
2847
Christoph Lametere498be72005-09-09 13:03:32 -07002848#ifdef CONFIG_NUMA
2849/*
2850 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002852static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2853 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002854{
2855 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002856 struct slab *slabp;
2857 struct kmem_list3 *l3;
2858 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002859 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002861 l3 = cachep->nodelists[nodeid];
2862 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002863
Andrew Mortona737b3e2006-03-22 00:08:11 -08002864retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002865 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002866 spin_lock(&l3->list_lock);
2867 entry = l3->slabs_partial.next;
2868 if (entry == &l3->slabs_partial) {
2869 l3->free_touched = 1;
2870 entry = l3->slabs_free.next;
2871 if (entry == &l3->slabs_free)
2872 goto must_grow;
2873 }
Christoph Lametere498be72005-09-09 13:03:32 -07002874
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002875 slabp = list_entry(entry, struct slab, list);
2876 check_spinlock_acquired_node(cachep, nodeid);
2877 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002878
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002879 STATS_INC_NODEALLOCS(cachep);
2880 STATS_INC_ACTIVE(cachep);
2881 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002882
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002884
Matthew Dobson78d382d2006-02-01 03:05:47 -08002885 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002886 check_slabp(cachep, slabp);
2887 l3->free_objects--;
2888 /* move slabp to correct slabp list: */
2889 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002890
Andrew Mortona737b3e2006-03-22 00:08:11 -08002891 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002892 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002893 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002894 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002895
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002896 spin_unlock(&l3->list_lock);
2897 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002898
Andrew Mortona737b3e2006-03-22 00:08:11 -08002899must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002900 spin_unlock(&l3->list_lock);
2901 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002902
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002903 if (!x)
2904 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002905
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002906 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002907done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002908 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002909}
2910#endif
2911
2912/*
2913 * Caller needs to acquire correct kmem_list's list_lock
2914 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002915static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
2918 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002919 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920
2921 for (i = 0; i < nr_objects; i++) {
2922 void *objp = objpp[i];
2923 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002925 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002926 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002928 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002930 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002932 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 check_slabp(cachep, slabp);
2934
2935 /* fixup slab chains */
2936 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002937 if (l3->free_objects > l3->free_limit) {
2938 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 slab_destroy(cachep, slabp);
2940 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002941 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 }
2943 } else {
2944 /* Unconditionally move a slab to the end of the
2945 * partial list on free - maximum time for the
2946 * other objects to be freed, too.
2947 */
Christoph Lametere498be72005-09-09 13:03:32 -07002948 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 }
2950 }
2951}
2952
Pekka Enberg343e0d72006-02-01 03:05:50 -08002953static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954{
2955 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002956 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002957 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958
2959 batchcount = ac->batchcount;
2960#if DEBUG
2961 BUG_ON(!batchcount || batchcount > ac->avail);
2962#endif
2963 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002964 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002965 spin_lock(&l3->list_lock);
2966 if (l3->shared) {
2967 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002968 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 if (max) {
2970 if (batchcount > max)
2971 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002972 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002973 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 shared_array->avail += batchcount;
2975 goto free_done;
2976 }
2977 }
2978
Christoph Lameterff694162005-09-22 21:44:02 -07002979 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002980free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981#if STATS
2982 {
2983 int i = 0;
2984 struct list_head *p;
2985
Christoph Lametere498be72005-09-09 13:03:32 -07002986 p = l3->slabs_free.next;
2987 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 struct slab *slabp;
2989
2990 slabp = list_entry(p, struct slab, list);
2991 BUG_ON(slabp->inuse);
2992
2993 i++;
2994 p = p->next;
2995 }
2996 STATS_SET_FREEABLE(cachep, i);
2997 }
2998#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002999 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003001 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002}
3003
3004/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003005 * Release an obj back to its cache. If the obj has a constructed state, it must
3006 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003008static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003010 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011
3012 check_irq_off();
3013 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3014
Christoph Lametere498be72005-09-09 13:03:32 -07003015 /* Make sure we are not freeing a object from another
3016 * node to the array cache on this cpu.
3017 */
3018#ifdef CONFIG_NUMA
3019 {
3020 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003021 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003022 if (unlikely(slabp->nodeid != numa_node_id())) {
3023 struct array_cache *alien = NULL;
3024 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003025 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003026
Andrew Mortona737b3e2006-03-22 00:08:11 -08003027 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003028 STATS_INC_NODEFREES(cachep);
3029 if (l3->alien && l3->alien[nodeid]) {
3030 alien = l3->alien[nodeid];
3031 spin_lock(&alien->lock);
3032 if (unlikely(alien->avail == alien->limit))
3033 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003034 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003035 alien->entry[alien->avail++] = objp;
3036 spin_unlock(&alien->lock);
3037 } else {
3038 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003039 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003040 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003041 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003042 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003043 }
3044 return;
3045 }
3046 }
3047#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 if (likely(ac->avail < ac->limit)) {
3049 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003050 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 return;
3052 } else {
3053 STATS_INC_FREEMISS(cachep);
3054 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003055 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 }
3057}
3058
3059/**
3060 * kmem_cache_alloc - Allocate an object
3061 * @cachep: The cache to allocate from.
3062 * @flags: See kmalloc().
3063 *
3064 * Allocate an object from this cache. The flags are only relevant
3065 * if the cache has no available objects.
3066 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003067void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003069 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070}
3071EXPORT_SYMBOL(kmem_cache_alloc);
3072
3073/**
3074 * kmem_ptr_validate - check if an untrusted pointer might
3075 * be a slab entry.
3076 * @cachep: the cache we're checking against
3077 * @ptr: pointer to validate
3078 *
3079 * This verifies that the untrusted pointer looks sane:
3080 * it is _not_ a guarantee that the pointer is actually
3081 * part of the slab cache in question, but it at least
3082 * validates that the pointer can be dereferenced and
3083 * looks half-way sane.
3084 *
3085 * Currently only used for dentry validation.
3086 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003087int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003089 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003092 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 struct page *page;
3094
3095 if (unlikely(addr < min_addr))
3096 goto out;
3097 if (unlikely(addr > (unsigned long)high_memory - size))
3098 goto out;
3099 if (unlikely(addr & align_mask))
3100 goto out;
3101 if (unlikely(!kern_addr_valid(addr)))
3102 goto out;
3103 if (unlikely(!kern_addr_valid(addr + size - 1)))
3104 goto out;
3105 page = virt_to_page(ptr);
3106 if (unlikely(!PageSlab(page)))
3107 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003108 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 goto out;
3110 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003111out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 return 0;
3113}
3114
3115#ifdef CONFIG_NUMA
3116/**
3117 * kmem_cache_alloc_node - Allocate an object on the specified node
3118 * @cachep: The cache to allocate from.
3119 * @flags: See kmalloc().
3120 * @nodeid: node number of the target node.
3121 *
3122 * Identical to kmem_cache_alloc, except that this function is slow
3123 * and can sleep. And it will allocate memory on the given node, which
3124 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003125 * New and improved: it will now make sure that the object gets
3126 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003128void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129{
Christoph Lametere498be72005-09-09 13:03:32 -07003130 unsigned long save_flags;
3131 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132
Christoph Lametere498be72005-09-09 13:03:32 -07003133 cache_alloc_debugcheck_before(cachep, flags);
3134 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003135
3136 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003137 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003138 ptr = ____cache_alloc(cachep, flags);
3139 else
3140 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003141 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003142
3143 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3144 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145
Christoph Lametere498be72005-09-09 13:03:32 -07003146 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147}
3148EXPORT_SYMBOL(kmem_cache_alloc_node);
3149
Al Virodd0fc662005-10-07 07:46:04 +01003150void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003151{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003152 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003153
3154 cachep = kmem_find_general_cachep(size, flags);
3155 if (unlikely(cachep == NULL))
3156 return NULL;
3157 return kmem_cache_alloc_node(cachep, flags, node);
3158}
3159EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160#endif
3161
3162/**
3163 * kmalloc - allocate memory
3164 * @size: how many bytes of memory are required.
3165 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003166 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 *
3168 * kmalloc is the normal method of allocating memory
3169 * in the kernel.
3170 *
3171 * The @flags argument may be one of:
3172 *
3173 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3174 *
3175 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3176 *
3177 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3178 *
3179 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3180 * must be suitable for DMA. This can mean different things on different
3181 * platforms. For example, on i386, it means that the memory must come
3182 * from the first 16MB.
3183 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003184static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3185 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003187 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003189 /* If you want to save a few bytes .text space: replace
3190 * __ with kmem_.
3191 * Then kmalloc uses the uninlined functions instead of the inline
3192 * functions.
3193 */
3194 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003195 if (unlikely(cachep == NULL))
3196 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003197 return __cache_alloc(cachep, flags, caller);
3198}
3199
3200#ifndef CONFIG_DEBUG_SLAB
3201
3202void *__kmalloc(size_t size, gfp_t flags)
3203{
3204 return __do_kmalloc(size, flags, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205}
3206EXPORT_SYMBOL(__kmalloc);
3207
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003208#else
3209
3210void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3211{
3212 return __do_kmalloc(size, flags, caller);
3213}
3214EXPORT_SYMBOL(__kmalloc_track_caller);
3215
3216#endif
3217
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218#ifdef CONFIG_SMP
3219/**
3220 * __alloc_percpu - allocate one copy of the object for every present
3221 * cpu in the system, zeroing them.
3222 * Objects should be dereferenced using the per_cpu_ptr macro only.
3223 *
3224 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003226void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227{
3228 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003229 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230
3231 if (!pdata)
3232 return NULL;
3233
Christoph Lametere498be72005-09-09 13:03:32 -07003234 /*
3235 * Cannot use for_each_online_cpu since a cpu may come online
3236 * and we have no way of figuring out how to fix the array
3237 * that we have allocated then....
3238 */
3239 for_each_cpu(i) {
3240 int node = cpu_to_node(i);
3241
3242 if (node_online(node))
3243 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3244 else
3245 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246
3247 if (!pdata->ptrs[i])
3248 goto unwind_oom;
3249 memset(pdata->ptrs[i], 0, size);
3250 }
3251
3252 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003253 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254
Andrew Mortona737b3e2006-03-22 00:08:11 -08003255unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 while (--i >= 0) {
3257 if (!cpu_possible(i))
3258 continue;
3259 kfree(pdata->ptrs[i]);
3260 }
3261 kfree(pdata);
3262 return NULL;
3263}
3264EXPORT_SYMBOL(__alloc_percpu);
3265#endif
3266
3267/**
3268 * kmem_cache_free - Deallocate an object
3269 * @cachep: The cache the allocation was from.
3270 * @objp: The previously allocated object.
3271 *
3272 * Free an object which was previously allocated from this
3273 * cache.
3274 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003275void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276{
3277 unsigned long flags;
3278
3279 local_irq_save(flags);
3280 __cache_free(cachep, objp);
3281 local_irq_restore(flags);
3282}
3283EXPORT_SYMBOL(kmem_cache_free);
3284
3285/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 * kfree - free previously allocated memory
3287 * @objp: pointer returned by kmalloc.
3288 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003289 * If @objp is NULL, no operation is performed.
3290 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 * Don't free memory not originally allocated by kmalloc()
3292 * or you will run into trouble.
3293 */
3294void kfree(const void *objp)
3295{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003296 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 unsigned long flags;
3298
3299 if (unlikely(!objp))
3300 return;
3301 local_irq_save(flags);
3302 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003303 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003304 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003305 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 local_irq_restore(flags);
3307}
3308EXPORT_SYMBOL(kfree);
3309
3310#ifdef CONFIG_SMP
3311/**
3312 * free_percpu - free previously allocated percpu memory
3313 * @objp: pointer returned by alloc_percpu.
3314 *
3315 * Don't free memory not originally allocated by alloc_percpu()
3316 * The complemented objp is to check for that.
3317 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319{
3320 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Christoph Lametere498be72005-09-09 13:03:32 -07003323 /*
3324 * We allocate for all cpus so we cannot use for online cpu here.
3325 */
3326 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003327 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 kfree(p);
3329}
3330EXPORT_SYMBOL(free_percpu);
3331#endif
3332
Pekka Enberg343e0d72006-02-01 03:05:50 -08003333unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003335 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336}
3337EXPORT_SYMBOL(kmem_cache_size);
3338
Pekka Enberg343e0d72006-02-01 03:05:50 -08003339const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003340{
3341 return cachep->name;
3342}
3343EXPORT_SYMBOL_GPL(kmem_cache_name);
3344
Christoph Lametere498be72005-09-09 13:03:32 -07003345/*
3346 * This initializes kmem_list3 for all nodes.
3347 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003348static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003349{
3350 int node;
3351 struct kmem_list3 *l3;
3352 int err = 0;
3353
3354 for_each_online_node(node) {
3355 struct array_cache *nc = NULL, *new;
3356 struct array_cache **new_alien = NULL;
3357#ifdef CONFIG_NUMA
Andrew Mortona737b3e2006-03-22 00:08:11 -08003358 new_alien = alloc_alien_cache(node, cachep->limit);
3359 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003360 goto fail;
3361#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08003362 new = alloc_arraycache(node, cachep->shared*cachep->batchcount,
3363 0xbaadf00d);
3364 if (!new)
Christoph Lametere498be72005-09-09 13:03:32 -07003365 goto fail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003366 l3 = cachep->nodelists[node];
3367 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07003368 spin_lock_irq(&l3->list_lock);
3369
Andrew Mortona737b3e2006-03-22 00:08:11 -08003370 nc = cachep->nodelists[node]->shared;
3371 if (nc)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003372 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003373
3374 l3->shared = new;
3375 if (!cachep->nodelists[node]->alien) {
3376 l3->alien = new_alien;
3377 new_alien = NULL;
3378 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003379 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003380 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003381 spin_unlock_irq(&l3->list_lock);
3382 kfree(nc);
3383 free_alien_cache(new_alien);
3384 continue;
3385 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003386 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3387 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07003388 goto fail;
3389
3390 kmem_list3_init(l3);
3391 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003392 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003393 l3->shared = new;
3394 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003395 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003396 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003397 cachep->nodelists[node] = l3;
3398 }
3399 return err;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003400fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003401 err = -ENOMEM;
3402 return err;
3403}
3404
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003406 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 struct array_cache *new[NR_CPUS];
3408};
3409
3410static void do_ccupdate_local(void *info)
3411{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003412 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413 struct array_cache *old;
3414
3415 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003416 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003417
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3419 new->new[smp_processor_id()] = old;
3420}
3421
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003422/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003423static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3424 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425{
3426 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003427 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003429 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003430 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003431 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3432 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003433 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003434 for (i--; i >= 0; i--)
3435 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003436 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437 }
3438 }
3439 new.cachep = cachep;
3440
Andrew Mortona07fa392006-03-22 00:08:17 -08003441 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003442
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 cachep->batchcount = batchcount;
3445 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003446 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447
Christoph Lametere498be72005-09-09 13:03:32 -07003448 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 struct array_cache *ccold = new.new[i];
3450 if (!ccold)
3451 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003452 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003453 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003454 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455 kfree(ccold);
3456 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
Christoph Lametere498be72005-09-09 13:03:32 -07003458 err = alloc_kmemlist(cachep);
3459 if (err) {
3460 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003462 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 return 0;
3465}
3466
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003467/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003468static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469{
3470 int err;
3471 int limit, shared;
3472
Andrew Mortona737b3e2006-03-22 00:08:11 -08003473 /*
3474 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 * - create a LIFO ordering, i.e. return objects that are cache-warm
3476 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003477 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 * bufctl chains: array operations are cheaper.
3479 * The numbers are guessed, we should auto-tune as described by
3480 * Bonwick.
3481 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003482 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003484 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003486 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003488 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 limit = 54;
3490 else
3491 limit = 120;
3492
Andrew Mortona737b3e2006-03-22 00:08:11 -08003493 /*
3494 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 * allocation behaviour: Most allocs on one cpu, most free operations
3496 * on another cpu. For these cases, an efficient object passing between
3497 * cpus is necessary. This is provided by a shared array. The array
3498 * replaces Bonwick's magazine layer.
3499 * On uniprocessor, it's functionally equivalent (but less efficient)
3500 * to a larger limit. Thus disabled by default.
3501 */
3502 shared = 0;
3503#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003504 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 shared = 8;
3506#endif
3507
3508#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003509 /*
3510 * With debugging enabled, large batchcount lead to excessively long
3511 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 */
3513 if (limit > 32)
3514 limit = 32;
3515#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003516 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 if (err)
3518 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003519 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520}
3521
Christoph Lameter1b552532006-03-22 00:09:07 -08003522/*
3523 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003524 * necessary. Note that the l3 listlock also protects the array_cache
3525 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003526 */
3527void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3528 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
3530 int tofree;
3531
Christoph Lameter1b552532006-03-22 00:09:07 -08003532 if (!ac || !ac->avail)
3533 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 if (ac->touched && !force) {
3535 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003536 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003537 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003538 if (ac->avail) {
3539 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3540 if (tofree > ac->avail)
3541 tofree = (ac->avail + 1) / 2;
3542 free_block(cachep, ac->entry, tofree, node);
3543 ac->avail -= tofree;
3544 memmove(ac->entry, &(ac->entry[tofree]),
3545 sizeof(void *) * ac->avail);
3546 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003547 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 }
3549}
3550
3551/**
3552 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003553 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 *
3555 * Called from workqueue/eventd every few seconds.
3556 * Purpose:
3557 * - clear the per-cpu caches for this CPU.
3558 * - return freeable pages to the main free memory pool.
3559 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003560 * If we cannot acquire the cache chain mutex then just give up - we'll try
3561 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 */
3563static void cache_reap(void *unused)
3564{
3565 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003566 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003567 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003569 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003571 schedule_delayed_work(&__get_cpu_var(reap_work),
3572 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 return;
3574 }
3575
3576 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003577 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003578 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 int tofree;
3580 struct slab *slabp;
3581
Pekka Enberg343e0d72006-02-01 03:05:50 -08003582 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 check_irq_on();
3584
Christoph Lameter35386e32006-03-22 00:09:05 -08003585 /*
3586 * We only take the l3 lock if absolutely necessary and we
3587 * have established with reasonable certainty that
3588 * we can do some work if the lock was obtained.
3589 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003590 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003591
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003592 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593
Christoph Lameteraab22072006-03-22 00:09:06 -08003594 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595
Christoph Lameter35386e32006-03-22 00:09:05 -08003596 /*
3597 * These are racy checks but it does not matter
3598 * if we skip one check or scan twice.
3599 */
Christoph Lametere498be72005-09-09 13:03:32 -07003600 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003601 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602
Christoph Lametere498be72005-09-09 13:03:32 -07003603 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604
Christoph Lameteraab22072006-03-22 00:09:06 -08003605 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606
Christoph Lametere498be72005-09-09 13:03:32 -07003607 if (l3->free_touched) {
3608 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003609 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610 }
3611
Andrew Mortona737b3e2006-03-22 00:08:11 -08003612 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3613 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003615 /*
3616 * Do not lock if there are no free blocks.
3617 */
3618 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 break;
3620
Christoph Lameter35386e32006-03-22 00:09:05 -08003621 spin_lock_irq(&l3->list_lock);
3622 p = l3->slabs_free.next;
3623 if (p == &(l3->slabs_free)) {
3624 spin_unlock_irq(&l3->list_lock);
3625 break;
3626 }
3627
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 slabp = list_entry(p, struct slab, list);
3629 BUG_ON(slabp->inuse);
3630 list_del(&slabp->list);
3631 STATS_INC_REAPED(searchp);
3632
Andrew Mortona737b3e2006-03-22 00:08:11 -08003633 /*
3634 * Safe to drop the lock. The slab is no longer linked
3635 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 * cache_chain_lock
3637 */
Christoph Lametere498be72005-09-09 13:03:32 -07003638 l3->free_objects -= searchp->num;
3639 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003641 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003642next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643 cond_resched();
3644 }
3645 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003646 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003647 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003648 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003649 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650}
3651
3652#ifdef CONFIG_PROC_FS
3653
Pekka Enberg85289f92006-01-08 01:00:36 -08003654static void print_slabinfo_header(struct seq_file *m)
3655{
3656 /*
3657 * Output format version, so at least we can change it
3658 * without _too_ many complaints.
3659 */
3660#if STATS
3661 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3662#else
3663 seq_puts(m, "slabinfo - version: 2.1\n");
3664#endif
3665 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3666 "<objperslab> <pagesperslab>");
3667 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3668 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3669#if STATS
3670 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3671 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3672 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3673#endif
3674 seq_putc(m, '\n');
3675}
3676
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677static void *s_start(struct seq_file *m, loff_t *pos)
3678{
3679 loff_t n = *pos;
3680 struct list_head *p;
3681
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003682 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003683 if (!n)
3684 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685 p = cache_chain.next;
3686 while (n--) {
3687 p = p->next;
3688 if (p == &cache_chain)
3689 return NULL;
3690 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003691 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692}
3693
3694static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3695{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003696 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003698 return cachep->next.next == &cache_chain ?
3699 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700}
3701
3702static void s_stop(struct seq_file *m, void *p)
3703{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003704 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705}
3706
3707static int s_show(struct seq_file *m, void *p)
3708{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003709 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003711 struct slab *slabp;
3712 unsigned long active_objs;
3713 unsigned long num_objs;
3714 unsigned long active_slabs = 0;
3715 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003716 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003718 int node;
3719 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 active_objs = 0;
3722 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003723 for_each_online_node(node) {
3724 l3 = cachep->nodelists[node];
3725 if (!l3)
3726 continue;
3727
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003728 check_irq_on();
3729 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003730
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003731 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003732 slabp = list_entry(q, struct slab, list);
3733 if (slabp->inuse != cachep->num && !error)
3734 error = "slabs_full accounting error";
3735 active_objs += cachep->num;
3736 active_slabs++;
3737 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003738 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003739 slabp = list_entry(q, struct slab, list);
3740 if (slabp->inuse == cachep->num && !error)
3741 error = "slabs_partial inuse accounting error";
3742 if (!slabp->inuse && !error)
3743 error = "slabs_partial/inuse accounting error";
3744 active_objs += slabp->inuse;
3745 active_slabs++;
3746 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003747 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003748 slabp = list_entry(q, struct slab, list);
3749 if (slabp->inuse && !error)
3750 error = "slabs_free/inuse accounting error";
3751 num_slabs++;
3752 }
3753 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003754 if (l3->shared)
3755 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003756
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003757 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003759 num_slabs += active_slabs;
3760 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003761 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 error = "free_objects accounting error";
3763
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003764 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765 if (error)
3766 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3767
3768 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003769 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003770 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003772 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003773 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003774 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003776 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 unsigned long high = cachep->high_mark;
3778 unsigned long allocs = cachep->num_allocations;
3779 unsigned long grown = cachep->grown;
3780 unsigned long reaped = cachep->reaped;
3781 unsigned long errors = cachep->errors;
3782 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003784 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785
Christoph Lametere498be72005-09-09 13:03:32 -07003786 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Andrew Mortona737b3e2006-03-22 00:08:11 -08003787 %4lu %4lu %4lu %4lu", allocs, high, grown,
3788 reaped, errors, max_freeable, node_allocs,
3789 node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 }
3791 /* cpu stats */
3792 {
3793 unsigned long allochit = atomic_read(&cachep->allochit);
3794 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3795 unsigned long freehit = atomic_read(&cachep->freehit);
3796 unsigned long freemiss = atomic_read(&cachep->freemiss);
3797
3798 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003799 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 }
3801#endif
3802 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 return 0;
3804}
3805
3806/*
3807 * slabinfo_op - iterator that generates /proc/slabinfo
3808 *
3809 * Output layout:
3810 * cache-name
3811 * num-active-objs
3812 * total-objs
3813 * object size
3814 * num-active-slabs
3815 * total-slabs
3816 * num-pages-per-slab
3817 * + further values on SMP and with statistics enabled
3818 */
3819
3820struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003821 .start = s_start,
3822 .next = s_next,
3823 .stop = s_stop,
3824 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825};
3826
3827#define MAX_SLABINFO_WRITE 128
3828/**
3829 * slabinfo_write - Tuning for the slab allocator
3830 * @file: unused
3831 * @buffer: user buffer
3832 * @count: data length
3833 * @ppos: unused
3834 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003835ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3836 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003838 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839 int limit, batchcount, shared, res;
3840 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003841
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 if (count > MAX_SLABINFO_WRITE)
3843 return -EINVAL;
3844 if (copy_from_user(&kbuf, buffer, count))
3845 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003846 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847
3848 tmp = strchr(kbuf, ' ');
3849 if (!tmp)
3850 return -EINVAL;
3851 *tmp = '\0';
3852 tmp++;
3853 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3854 return -EINVAL;
3855
3856 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003857 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003859 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003860 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
Andrew Mortona737b3e2006-03-22 00:08:11 -08003862 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003864 if (limit < 1 || batchcount < 1 ||
3865 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003866 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003868 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003869 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 }
3871 break;
3872 }
3873 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003874 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 if (res >= 0)
3876 res = count;
3877 return res;
3878}
3879#endif
3880
Manfred Spraul00e145b2005-09-03 15:55:07 -07003881/**
3882 * ksize - get the actual amount of memory allocated for a given object
3883 * @objp: Pointer to the object
3884 *
3885 * kmalloc may internally round up allocations and return more memory
3886 * than requested. ksize() can be used to determine the actual amount of
3887 * memory allocated. The caller may use this additional memory, even though
3888 * a smaller amount of memory was initially specified with the kmalloc call.
3889 * The caller must guarantee that objp points to a valid object previously
3890 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3891 * must not be freed during the duration of the call.
3892 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893unsigned int ksize(const void *objp)
3894{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003895 if (unlikely(objp == NULL))
3896 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003898 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899}