blob: 7451bdacaf18875044ea97ed3d49da8c401d8249 [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
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * 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
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Zhaolei02af61b2009-04-10 14:26:18 +0800105#include <linux/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200117#include <linux/kmemcheck.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119#include <asm/cacheflush.h>
120#include <asm/tlbflush.h>
121#include <asm/page.h>
122
123/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 */
132
133#ifdef CONFIG_DEBUG_SLAB
134#define DEBUG 1
135#define STATS 1
136#define FORCED_DEBUG 1
137#else
138#define DEBUG 0
139#define STATS 0
140#define FORCED_DEBUG 0
141#endif
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/* Shouldn't this be in a header file somewhere? */
144#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400145#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147#ifndef ARCH_KMALLOC_MINALIGN
148/*
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700157#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158#endif
159
160#ifndef ARCH_SLAB_MINALIGN
161/*
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
167 */
168#define ARCH_SLAB_MINALIGN 0
169#endif
170
171#ifndef ARCH_KMALLOC_FLAGS
172#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173#endif
174
175/* Legal flag mask for kmem_cache_create(). */
176#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700177# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800179 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700180 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800185# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700186 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190#endif
191
192/*
193 * kmem_bufctl_t:
194 *
195 * Bufctl's are used for linking objs within a slab
196 * linked offsets.
197 *
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
209 */
210
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700211typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800214#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217/*
218 * struct slab
219 *
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 */
224struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800225 struct list_head list;
226 unsigned long colouroff;
227 void *s_mem; /* including colour offset */
228 unsigned int inuse; /* num of objs active in slab */
229 kmem_bufctl_t free;
230 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231};
232
233/*
234 * struct slab_rcu
235 *
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
243 *
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
246 *
247 * We assume struct slab_rcu can overlay struct slab when destroying.
248 */
249struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800250 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800251 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800252 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253};
254
255/*
256 * struct array_cache
257 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 * Purpose:
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
262 *
263 * The limit is stored in the per-cpu structure to reduce the data cache
264 * footprint.
265 *
266 */
267struct array_cache {
268 unsigned int avail;
269 unsigned int limit;
270 unsigned int batchcount;
271 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700272 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700273 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
276 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278};
279
Andrew Mortona737b3e2006-03-22 00:08:11 -0800280/*
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 */
284#define BOOT_CPUCACHE_ENTRIES 1
285struct arraycache_init {
286 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800287 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288};
289
290/*
Christoph Lametere498be72005-09-09 13:03:32 -0700291 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 */
293struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800294 struct list_head slabs_partial; /* partial list first, better asm code */
295 struct list_head slabs_full;
296 struct list_head slabs_free;
297 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800299 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800300 spinlock_t list_lock;
301 struct array_cache *shared; /* shared per node */
302 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800303 unsigned long next_reap; /* updated without locking */
304 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305};
306
Christoph Lametere498be72005-09-09 13:03:32 -0700307/*
308 * Need this for bootstrapping a per node allocator.
309 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200310#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700311struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
312#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200313#define SIZE_AC MAX_NUMNODES
314#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315
Christoph Lametered11d9e2006-06-30 01:55:45 -0700316static int drain_freelist(struct kmem_cache *cache,
317 struct kmem_list3 *l3, int tofree);
318static void free_block(struct kmem_cache *cachep, void **objpp, int len,
319 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300320static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000321static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700322
Christoph Lametere498be72005-09-09 13:03:32 -0700323/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700326 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700327static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700328{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800329 extern void __bad_size(void);
330
Christoph Lametere498be72005-09-09 13:03:32 -0700331 if (__builtin_constant_p(size)) {
332 int i = 0;
333
334#define CACHE(x) \
335 if (size <=x) \
336 return i; \
337 else \
338 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800339#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700340#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700342 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800343 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700344 return 0;
345}
346
Ingo Molnare0a42722006-06-23 02:03:46 -0700347static int slab_early_init = 1;
348
Christoph Lametere498be72005-09-09 13:03:32 -0700349#define INDEX_AC index_of(sizeof(struct arraycache_init))
350#define INDEX_L3 index_of(sizeof(struct kmem_list3))
351
Pekka Enberg5295a742006-02-01 03:05:48 -0800352static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700353{
354 INIT_LIST_HEAD(&parent->slabs_full);
355 INIT_LIST_HEAD(&parent->slabs_partial);
356 INIT_LIST_HEAD(&parent->slabs_free);
357 parent->shared = NULL;
358 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800359 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700360 spin_lock_init(&parent->list_lock);
361 parent->free_objects = 0;
362 parent->free_touched = 0;
363}
364
Andrew Mortona737b3e2006-03-22 00:08:11 -0800365#define MAKE_LIST(cachep, listp, slab, nodeid) \
366 do { \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700369 } while (0)
370
Andrew Mortona737b3e2006-03-22 00:08:11 -0800371#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
372 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
376 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378#define CFLGS_OFF_SLAB (0x80000000UL)
379#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
380
381#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800382/*
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100386 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 * which could lock up otherwise freeable slabs.
388 */
389#define REAPTIMEOUT_CPUC (2*HZ)
390#define REAPTIMEOUT_LIST3 (4*HZ)
391
392#if STATS
393#define STATS_INC_ACTIVE(x) ((x)->num_active++)
394#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700397#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800398#define STATS_SET_HIGH(x) \
399 do { \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
402 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403#define STATS_INC_ERR(x) ((x)->errors++)
404#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700405#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700406#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407#define STATS_SET_FREEABLE(x, i) \
408 do { \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
411 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
416#else
417#define STATS_INC_ACTIVE(x) do { } while (0)
418#define STATS_DEC_ACTIVE(x) do { } while (0)
419#define STATS_INC_ALLOCED(x) do { } while (0)
420#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700421#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422#define STATS_SET_HIGH(x) do { } while (0)
423#define STATS_INC_ERR(x) do { } while (0)
424#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700425#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700426#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800427#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428#define STATS_INC_ALLOCHIT(x) do { } while (0)
429#define STATS_INC_ALLOCMISS(x) do { } while (0)
430#define STATS_INC_FREEHIT(x) do { } while (0)
431#define STATS_INC_FREEMISS(x) do { } while (0)
432#endif
433
434#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Andrew Mortona737b3e2006-03-22 00:08:11 -0800436/*
437 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800449static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800451 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
Pekka Enberg343e0d72006-02-01 03:05:50 -0800454static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800456 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
David Woodhouseb46b8f12007-05-08 00:22:59 -0700459static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700462 return (unsigned long long*) (objp + obj_offset(cachep) -
463 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464}
465
David Woodhouseb46b8f12007-05-08 00:22:59 -0700466static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
468 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
469 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700470 return (unsigned long long *)(objp + cachep->buffer_size -
471 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400472 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700473 return (unsigned long long *) (objp + cachep->buffer_size -
474 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Pekka Enberg343e0d72006-02-01 03:05:50 -0800477static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478{
479 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800480 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481}
482
483#else
484
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800485#define obj_offset(x) 0
486#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700487#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
490
491#endif
492
Li Zefan0f24f122009-12-11 15:45:30 +0800493#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300494size_t slab_buffer_size(struct kmem_cache *cachep)
495{
496 return cachep->buffer_size;
497}
498EXPORT_SYMBOL(slab_buffer_size);
499#endif
500
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 * Do not go above this order unless 0 objects fit into the slab.
503 */
504#define BREAK_GFP_ORDER_HI 1
505#define BREAK_GFP_ORDER_LO 0
506static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
507
Andrew Mortona737b3e2006-03-22 00:08:11 -0800508/*
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800513static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
514{
515 page->lru.next = (struct list_head *)cache;
516}
517
518static inline struct kmem_cache *page_get_cache(struct page *page)
519{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700520 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700521 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800522 return (struct kmem_cache *)page->lru.next;
523}
524
525static inline void page_set_slab(struct page *page, struct slab *slab)
526{
527 page->lru.prev = (struct list_head *)slab;
528}
529
530static inline struct slab *page_get_slab(struct page *page)
531{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700532 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800533 return (struct slab *)page->lru.prev;
534}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800536static inline struct kmem_cache *virt_to_cache(const void *obj)
537{
Christoph Lameterb49af682007-05-06 14:49:41 -0700538 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800539 return page_get_cache(page);
540}
541
542static inline struct slab *virt_to_slab(const void *obj)
543{
Christoph Lameterb49af682007-05-06 14:49:41 -0700544 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800545 return page_get_slab(page);
546}
547
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800548static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
549 unsigned int idx)
550{
551 return slab->s_mem + cache->buffer_size * idx;
552}
553
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800554/*
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
559 */
560static inline unsigned int obj_to_index(const struct kmem_cache *cache,
561 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800562{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800563 u32 offset = (obj - slab->s_mem);
564 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800565}
566
Andrew Mortona737b3e2006-03-22 00:08:11 -0800567/*
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
569 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570struct cache_sizes malloc_sizes[] = {
571#define CACHE(x) { .cs_size = (x) },
572#include <linux/kmalloc_sizes.h>
573 CACHE(ULONG_MAX)
574#undef CACHE
575};
576EXPORT_SYMBOL(malloc_sizes);
577
578/* Must match cache_sizes above. Out of line to keep cache footprint low. */
579struct cache_names {
580 char *name;
581 char *name_dma;
582};
583
584static struct cache_names __initdata cache_names[] = {
585#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800587 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588#undef CACHE
589};
590
591static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800592 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800594 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800597static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800598 .batchcount = 1,
599 .limit = BOOT_CPUCACHE_ENTRIES,
600 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800601 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800602 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603};
604
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700605#define BAD_ALIEN_MAGIC 0x01020304ul
606
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 * chicken and egg problem: delay the per-cpu array allocation
609 * until the general caches are up.
610 */
611static enum {
612 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700613 PARTIAL_AC,
614 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300615 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 FULL
617} g_cpucache_up;
618
Mike Kravetz39d24e62006-05-15 09:44:13 -0700619/*
620 * used by boot code to determine if it can use slab based allocator
621 */
622int slab_is_available(void)
623{
Pekka Enberg8429db52009-06-12 15:58:59 +0300624 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700625}
626
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200627#ifdef CONFIG_LOCKDEP
628
629/*
630 * Slab sometimes uses the kmalloc slabs to store the slab headers
631 * for other slabs "off slab".
632 * The locking for this is tricky in that it nests within the locks
633 * of all other slabs in a few places; to deal with this special
634 * locking we put on-slab caches into a separate lock-class.
635 *
636 * We set lock class for alien array caches which are up during init.
637 * The lock annotation will be lost if all cpus of a node goes down and
638 * then comes back up during hotplug
639 */
640static struct lock_class_key on_slab_l3_key;
641static struct lock_class_key on_slab_alc_key;
642
643static void init_node_lock_keys(int q)
644{
645 struct cache_sizes *s = malloc_sizes;
646
647 if (g_cpucache_up != FULL)
648 return;
649
650 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
651 struct array_cache **alc;
652 struct kmem_list3 *l3;
653 int r;
654
655 l3 = s->cs_cachep->nodelists[q];
656 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200657 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200658 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
659 alc = l3->alien;
660 /*
661 * FIXME: This check for BAD_ALIEN_MAGIC
662 * should go away when common slab code is taught to
663 * work even without alien caches.
664 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
665 * for alloc_alien_cache,
666 */
667 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
Pekka Enberg00afa752009-12-27 14:33:14 +0200668 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200669 for_each_node(r) {
670 if (alc[r])
671 lockdep_set_class(&alc[r]->lock,
672 &on_slab_alc_key);
673 }
674 }
675}
676
677static inline void init_lock_keys(void)
678{
679 int node;
680
681 for_each_node(node)
682 init_node_lock_keys(node);
683}
684#else
685static void init_node_lock_keys(int q)
686{
687}
688
689static inline void init_lock_keys(void)
690{
691}
692#endif
693
694/*
695 * Guard access to the cache-chain.
696 */
697static DEFINE_MUTEX(cache_chain_mutex);
698static struct list_head cache_chain;
699
Tejun Heo1871e522009-10-29 22:34:13 +0900700static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
Pekka Enberg343e0d72006-02-01 03:05:50 -0800702static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703{
704 return cachep->array[smp_processor_id()];
705}
706
Andrew Mortona737b3e2006-03-22 00:08:11 -0800707static inline struct kmem_cache *__find_general_cachep(size_t size,
708 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
710 struct cache_sizes *csizep = malloc_sizes;
711
712#if DEBUG
713 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
716 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700717 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700719 if (!size)
720 return ZERO_SIZE_PTR;
721
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 while (size > csizep->cs_size)
723 csizep++;
724
725 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
729 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800730#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800733#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 return csizep->cs_cachep;
735}
736
Adrian Bunkb2213852006-09-25 23:31:02 -0700737static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700738{
739 return __find_general_cachep(size, gfpflags);
740}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700741
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
Harvey Harrisond40cee22008-04-30 00:55:07 -0700808#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809
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
Paul Menage3395ee02006-12-06 20:32:16 -0800818/*
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
823 * line
824 */
825
826static int use_alien_caches __read_mostly = 1;
827static int __init noaliencache_setup(char *s)
828{
829 use_alien_caches = 0;
830 return 1;
831}
832__setup("noaliencache", noaliencache_setup);
833
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800834#ifdef CONFIG_NUMA
835/*
836 * Special reaping functions for NUMA systems called from cache_reap().
837 * These take care of doing round robin flushing of alien caches (containing
838 * objects freed on different nodes from which they were allocated) and the
839 * flushing of remote pcps by calling drain_node_pages.
840 */
Tejun Heo1871e522009-10-29 22:34:13 +0900841static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800842
843static void init_reap_node(int cpu)
844{
845 int node;
846
847 node = next_node(cpu_to_node(cpu), node_online_map);
848 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800849 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800850
Tejun Heo1871e522009-10-29 22:34:13 +0900851 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800852}
853
854static void next_reap_node(void)
855{
Tejun Heo1871e522009-10-29 22:34:13 +0900856 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800857
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800858 node = next_node(node, node_online_map);
859 if (unlikely(node >= MAX_NUMNODES))
860 node = first_node(node_online_map);
Tejun Heo1871e522009-10-29 22:34:13 +0900861 __get_cpu_var(slab_reap_node) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800862}
863
864#else
865#define init_reap_node(cpu) do { } while (0)
866#define next_reap_node(void) do { } while (0)
867#endif
868
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869/*
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
874 * lock.
875 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700876static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877{
Tejun Heo1871e522009-10-29 22:34:13 +0900878 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879
880 /*
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
883 * at that time.
884 */
David Howells52bad642006-11-22 14:54:01 +0000885 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800886 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000887 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800888 schedule_delayed_work_on(cpu, reap_work,
889 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 }
891}
892
Christoph Lametere498be72005-09-09 13:03:32 -0700893static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300894 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800896 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 struct array_cache *nc = NULL;
898
Pekka Enberg83b519e2009-06-10 19:40:04 +0300899 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100900 /*
901 * The array_cache structures contain pointers to free object.
902 * However, when such objects are allocated or transfered to another
903 * cache the pointers are not cleared and they could be counted as
904 * valid references during a kmemleak scan. Therefore, kmemleak must
905 * not scan such objects.
906 */
907 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 if (nc) {
909 nc->avail = 0;
910 nc->limit = entries;
911 nc->batchcount = batchcount;
912 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700913 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 }
915 return nc;
916}
917
Christoph Lameter3ded1752006-03-25 03:06:44 -0800918/*
919 * Transfer objects in one arraycache to another.
920 * Locking must be handled by the caller.
921 *
922 * Return the number of entries transferred.
923 */
924static int transfer_objects(struct array_cache *to,
925 struct array_cache *from, unsigned int max)
926{
927 /* Figure out how many entries to transfer */
928 int nr = min(min(from->avail, max), to->limit - to->avail);
929
930 if (!nr)
931 return 0;
932
933 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
934 sizeof(void *) *nr);
935
936 from->avail -= nr;
937 to->avail += nr;
938 to->touched = 1;
939 return nr;
940}
941
Christoph Lameter765c4502006-09-27 01:50:08 -0700942#ifndef CONFIG_NUMA
943
944#define drain_alien_cache(cachep, alien) do { } while (0)
945#define reap_alien(cachep, l3) do { } while (0)
946
Pekka Enberg83b519e2009-06-10 19:40:04 +0300947static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700948{
949 return (struct array_cache **)BAD_ALIEN_MAGIC;
950}
951
952static inline void free_alien_cache(struct array_cache **ac_ptr)
953{
954}
955
956static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
957{
958 return 0;
959}
960
961static inline void *alternate_node_alloc(struct kmem_cache *cachep,
962 gfp_t flags)
963{
964 return NULL;
965}
966
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800967static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700968 gfp_t flags, int nodeid)
969{
970 return NULL;
971}
972
973#else /* CONFIG_NUMA */
974
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800975static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800976static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800977
Pekka Enberg83b519e2009-06-10 19:40:04 +0300978static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700979{
980 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800981 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700982 int i;
983
984 if (limit > 1)
985 limit = 12;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300986 ac_ptr = kmalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700987 if (ac_ptr) {
988 for_each_node(i) {
989 if (i == node || !node_online(i)) {
990 ac_ptr[i] = NULL;
991 continue;
992 }
Pekka Enberg83b519e2009-06-10 19:40:04 +0300993 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700994 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800995 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700996 kfree(ac_ptr[i]);
997 kfree(ac_ptr);
998 return NULL;
999 }
1000 }
1001 }
1002 return ac_ptr;
1003}
1004
Pekka Enberg5295a742006-02-01 03:05:48 -08001005static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001006{
1007 int i;
1008
1009 if (!ac_ptr)
1010 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001011 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001012 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001013 kfree(ac_ptr);
1014}
1015
Pekka Enberg343e0d72006-02-01 03:05:50 -08001016static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001017 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001018{
1019 struct kmem_list3 *rl3 = cachep->nodelists[node];
1020
1021 if (ac->avail) {
1022 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001023 /*
1024 * Stuff objects into the remote nodes shared array first.
1025 * That way we could avoid the overhead of putting the objects
1026 * into the free lists and getting them back later.
1027 */
shin, jacob693f7d32006-04-28 10:54:37 -05001028 if (rl3->shared)
1029 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001030
Christoph Lameterff694162005-09-22 21:44:02 -07001031 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001032 ac->avail = 0;
1033 spin_unlock(&rl3->list_lock);
1034 }
1035}
1036
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001037/*
1038 * Called from cache_reap() to regularly drain alien caches round robin.
1039 */
1040static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1041{
Tejun Heo1871e522009-10-29 22:34:13 +09001042 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001043
1044 if (l3->alien) {
1045 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001046
1047 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001048 __drain_alien_cache(cachep, ac, node);
1049 spin_unlock_irq(&ac->lock);
1050 }
1051 }
1052}
1053
Andrew Mortona737b3e2006-03-22 00:08:11 -08001054static void drain_alien_cache(struct kmem_cache *cachep,
1055 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001056{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001057 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001058 struct array_cache *ac;
1059 unsigned long flags;
1060
1061 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001062 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001063 if (ac) {
1064 spin_lock_irqsave(&ac->lock, flags);
1065 __drain_alien_cache(cachep, ac, i);
1066 spin_unlock_irqrestore(&ac->lock, flags);
1067 }
1068 }
1069}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001070
Ingo Molnar873623d2006-07-13 14:44:38 +02001071static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001072{
1073 struct slab *slabp = virt_to_slab(objp);
1074 int nodeid = slabp->nodeid;
1075 struct kmem_list3 *l3;
1076 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001077 int node;
1078
1079 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001080
1081 /*
1082 * Make sure we are not freeing a object from another node to the array
1083 * cache on this cpu.
1084 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001085 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001086 return 0;
1087
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001088 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001089 STATS_INC_NODEFREES(cachep);
1090 if (l3->alien && l3->alien[nodeid]) {
1091 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001092 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001093 if (unlikely(alien->avail == alien->limit)) {
1094 STATS_INC_ACOVERFLOW(cachep);
1095 __drain_alien_cache(cachep, alien, nodeid);
1096 }
1097 alien->entry[alien->avail++] = objp;
1098 spin_unlock(&alien->lock);
1099 } else {
1100 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1101 free_block(cachep, &objp, 1, nodeid);
1102 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1103 }
1104 return 1;
1105}
Christoph Lametere498be72005-09-09 13:03:32 -07001106#endif
1107
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001108static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001110 struct kmem_cache *cachep;
1111 struct kmem_list3 *l3 = NULL;
1112 int node = cpu_to_node(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301113 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001114
1115 list_for_each_entry(cachep, &cache_chain, next) {
1116 struct array_cache *nc;
1117 struct array_cache *shared;
1118 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001119
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001120 /* cpu is dead; no one can alloc from it. */
1121 nc = cachep->array[cpu];
1122 cachep->array[cpu] = NULL;
1123 l3 = cachep->nodelists[node];
1124
1125 if (!l3)
1126 goto free_array_cache;
1127
1128 spin_lock_irq(&l3->list_lock);
1129
1130 /* Free limit for this kmem_list3 */
1131 l3->free_limit -= cachep->batchcount;
1132 if (nc)
1133 free_block(cachep, nc->entry, nc->avail, node);
1134
Rusty Russell58463c12009-12-17 11:43:12 -06001135 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001136 spin_unlock_irq(&l3->list_lock);
1137 goto free_array_cache;
1138 }
1139
1140 shared = l3->shared;
1141 if (shared) {
1142 free_block(cachep, shared->entry,
1143 shared->avail, node);
1144 l3->shared = NULL;
1145 }
1146
1147 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);
1156 }
1157free_array_cache:
1158 kfree(nc);
1159 }
1160 /*
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 drain_freelist(cachep, l3, l3->free_objects);
1170 }
1171}
1172
1173static int __cpuinit cpuup_prepare(long cpu)
1174{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001175 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001176 struct kmem_list3 *l3 = NULL;
1177 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001178 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001180 /*
1181 * We need to do this right in the beginning since
1182 * alloc_arraycache's are going to use this list.
1183 * kmalloc_node allows us to add the slab to the right
1184 * kmem_list3 and not this cpu's kmem_list3
1185 */
1186
1187 list_for_each_entry(cachep, &cache_chain, next) {
1188 /*
1189 * Set up the size64 kmemlist for cpu before we can
1190 * begin anything. Make sure some other cpu on this
1191 * node has not already allocated this
1192 */
1193 if (!cachep->nodelists[node]) {
1194 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1195 if (!l3)
1196 goto bad;
1197 kmem_list3_init(l3);
1198 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1199 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1200
1201 /*
1202 * The l3s don't come and go as CPUs come and
1203 * go. cache_chain_mutex is sufficient
1204 * protection here.
1205 */
1206 cachep->nodelists[node] = l3;
1207 }
1208
1209 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1210 cachep->nodelists[node]->free_limit =
1211 (1 + nr_cpus_node(node)) *
1212 cachep->batchcount + cachep->num;
1213 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1214 }
1215
1216 /*
1217 * Now we can go ahead with allocating the shared arrays and
1218 * array caches
1219 */
1220 list_for_each_entry(cachep, &cache_chain, next) {
1221 struct array_cache *nc;
1222 struct array_cache *shared = NULL;
1223 struct array_cache **alien = NULL;
1224
1225 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001226 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001227 if (!nc)
1228 goto bad;
1229 if (cachep->shared) {
1230 shared = alloc_arraycache(node,
1231 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001232 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001233 if (!shared) {
1234 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001235 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001236 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001237 }
1238 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001239 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001240 if (!alien) {
1241 kfree(shared);
1242 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001243 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001244 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001245 }
1246 cachep->array[cpu] = nc;
1247 l3 = cachep->nodelists[node];
1248 BUG_ON(!l3);
1249
1250 spin_lock_irq(&l3->list_lock);
1251 if (!l3->shared) {
1252 /*
1253 * We are serialised from CPU_DEAD or
1254 * CPU_UP_CANCELLED by the cpucontrol lock
1255 */
1256 l3->shared = shared;
1257 shared = NULL;
1258 }
1259#ifdef CONFIG_NUMA
1260 if (!l3->alien) {
1261 l3->alien = alien;
1262 alien = NULL;
1263 }
1264#endif
1265 spin_unlock_irq(&l3->list_lock);
1266 kfree(shared);
1267 free_alien_cache(alien);
1268 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001269 init_node_lock_keys(node);
1270
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001271 return 0;
1272bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001273 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001274 return -ENOMEM;
1275}
1276
1277static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1278 unsigned long action, void *hcpu)
1279{
1280 long cpu = (long)hcpu;
1281 int err = 0;
1282
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001284 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001285 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001286 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001287 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001288 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 break;
1290 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001291 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 start_cpu_timer(cpu);
1293 break;
1294#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001295 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001296 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001297 /*
1298 * Shutdown cache reaper. Note that the cache_chain_mutex is
1299 * held so that if cache_reap() is invoked it cannot do
1300 * anything expensive but will only modify reap_work
1301 * and reschedule the timer.
1302 */
Tejun Heo1871e522009-10-29 22:34:13 +09001303 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001304 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001305 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001306 break;
1307 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001308 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001309 start_cpu_timer(cpu);
1310 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001312 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001313 /*
1314 * Even if all the cpus of a node are down, we don't free the
1315 * kmem_list3 of any cache. This to avoid a race between
1316 * cpu_down, and a kmalloc allocation from another cpu for
1317 * memory from the node of the cpu going down. The list3
1318 * structure is usually allocated from kmem_cache_create() and
1319 * gets destroyed at kmem_cache_destroy().
1320 */
Simon Arlott183ff222007-10-20 01:27:18 +02001321 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001322#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001324 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001325 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001326 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001327 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001330 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331}
1332
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001333static struct notifier_block __cpuinitdata cpucache_notifier = {
1334 &cpuup_callback, NULL, 0
1335};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336
Christoph Lametere498be72005-09-09 13:03:32 -07001337/*
1338 * swap the static kmem_list3 with kmalloced memory
1339 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001340static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1341 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001342{
1343 struct kmem_list3 *ptr;
1344
Pekka Enberg83b519e2009-06-10 19:40:04 +03001345 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001346 BUG_ON(!ptr);
1347
Christoph Lametere498be72005-09-09 13:03:32 -07001348 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001349 /*
1350 * Do not assume that spinlocks can be initialized via memcpy:
1351 */
1352 spin_lock_init(&ptr->list_lock);
1353
Christoph Lametere498be72005-09-09 13:03:32 -07001354 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1355 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001356}
1357
Andrew Mortona737b3e2006-03-22 00:08:11 -08001358/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001359 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1360 * size of kmem_list3.
1361 */
1362static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1363{
1364 int node;
1365
1366 for_each_online_node(node) {
1367 cachep->nodelists[node] = &initkmem_list3[index + node];
1368 cachep->nodelists[node]->next_reap = jiffies +
1369 REAPTIMEOUT_LIST3 +
1370 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1371 }
1372}
1373
1374/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001375 * Initialisation. Called after the page allocator have been initialised and
1376 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 */
1378void __init kmem_cache_init(void)
1379{
1380 size_t left_over;
1381 struct cache_sizes *sizes;
1382 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001383 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001384 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001385 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001386
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001387 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001388 use_alien_caches = 0;
1389
Christoph Lametere498be72005-09-09 13:03:32 -07001390 for (i = 0; i < NUM_INIT_LISTS; i++) {
1391 kmem_list3_init(&initkmem_list3[i]);
1392 if (i < MAX_NUMNODES)
1393 cache_cache.nodelists[i] = NULL;
1394 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001395 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
1397 /*
1398 * Fragmentation resistance on low memory - only use bigger
1399 * page orders on machines with more than 32MB of memory.
1400 */
Jan Beulich44813742009-09-21 17:03:05 -07001401 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 /* Bootstrap is tricky, because several objects are allocated
1405 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001406 * 1) initialize the cache_cache cache: it contains the struct
1407 * kmem_cache structures of all caches, except cache_cache itself:
1408 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001409 * Initially an __init data area is used for the head array and the
1410 * kmem_list3 structures, it's replaced with a kmalloc allocated
1411 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001413 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001414 * An __init data area is used for the head array.
1415 * 3) Create the remaining kmalloc caches, with minimally sized
1416 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 * 4) Replace the __init data head arrays for cache_cache and the first
1418 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001419 * 5) Replace the __init data for kmem_list3 for cache_cache and
1420 * the other cache's with kmalloc allocated memory.
1421 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 */
1423
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001424 node = numa_node_id();
1425
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 INIT_LIST_HEAD(&cache_chain);
1428 list_add(&cache_cache.next, &cache_chain);
1429 cache_cache.colour_off = cache_line_size();
1430 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001431 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432
Eric Dumazet8da34302007-05-06 14:49:29 -07001433 /*
1434 * struct kmem_cache size depends on nr_node_ids, which
1435 * can be less than MAX_NUMNODES.
1436 */
1437 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1438 nr_node_ids * sizeof(struct kmem_list3 *);
1439#if DEBUG
1440 cache_cache.obj_size = cache_cache.buffer_size;
1441#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001442 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1443 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001444 cache_cache.reciprocal_buffer_size =
1445 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
Jack Steiner07ed76b2006-03-07 21:55:46 -08001447 for (order = 0; order < MAX_ORDER; order++) {
1448 cache_estimate(order, cache_cache.buffer_size,
1449 cache_line_size(), 0, &left_over, &cache_cache.num);
1450 if (cache_cache.num)
1451 break;
1452 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001453 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001454 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001455 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001456 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1457 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
1459 /* 2+3) create the kmalloc caches */
1460 sizes = malloc_sizes;
1461 names = cache_names;
1462
Andrew Mortona737b3e2006-03-22 00:08:11 -08001463 /*
1464 * Initialize the caches that provide memory for the array cache and the
1465 * kmem_list3 structures first. Without this, further allocations will
1466 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001467 */
1468
1469 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001470 sizes[INDEX_AC].cs_size,
1471 ARCH_KMALLOC_MINALIGN,
1472 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001473 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001474
Andrew Mortona737b3e2006-03-22 00:08:11 -08001475 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001476 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001477 kmem_cache_create(names[INDEX_L3].name,
1478 sizes[INDEX_L3].cs_size,
1479 ARCH_KMALLOC_MINALIGN,
1480 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001481 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001482 }
Christoph Lametere498be72005-09-09 13:03:32 -07001483
Ingo Molnare0a42722006-06-23 02:03:46 -07001484 slab_early_init = 0;
1485
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001487 /*
1488 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 * This should be particularly beneficial on SMP boxes, as it
1490 * eliminates "false sharing".
1491 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001492 * allow tighter packing of the smaller caches.
1493 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001494 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001495 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001496 sizes->cs_size,
1497 ARCH_KMALLOC_MINALIGN,
1498 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001499 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001500 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001501#ifdef CONFIG_ZONE_DMA
1502 sizes->cs_dmacachep = kmem_cache_create(
1503 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001504 sizes->cs_size,
1505 ARCH_KMALLOC_MINALIGN,
1506 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1507 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001508 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001509#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 sizes++;
1511 names++;
1512 }
1513 /* 4) Replace the bootstrap head arrays */
1514 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001515 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001516
Pekka Enberg83b519e2009-06-10 19:40:04 +03001517 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001518
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001519 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1520 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001521 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001522 /*
1523 * Do not assume that spinlocks can be initialized via memcpy:
1524 */
1525 spin_lock_init(&ptr->lock);
1526
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001528
Pekka Enberg83b519e2009-06-10 19:40:04 +03001529 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001530
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001531 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001532 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001533 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001534 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001535 /*
1536 * Do not assume that spinlocks can be initialized via memcpy:
1537 */
1538 spin_lock_init(&ptr->lock);
1539
Christoph Lametere498be72005-09-09 13:03:32 -07001540 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001541 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 }
Christoph Lametere498be72005-09-09 13:03:32 -07001543 /* 5) Replace the bootstrap kmem_list3's */
1544 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001545 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546
Mel Gorman9c09a952008-01-24 05:49:54 -08001547 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001548 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001549
Christoph Lametere498be72005-09-09 13:03:32 -07001550 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001551 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001552
1553 if (INDEX_AC != INDEX_L3) {
1554 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001555 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001556 }
1557 }
1558 }
1559
Pekka Enberg8429db52009-06-12 15:58:59 +03001560 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001561}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001562
Pekka Enberg8429db52009-06-12 15:58:59 +03001563void __init kmem_cache_init_late(void)
1564{
1565 struct kmem_cache *cachep;
1566
Pekka Enberg8429db52009-06-12 15:58:59 +03001567 /* 6) resize the head arrays to their final sizes */
1568 mutex_lock(&cache_chain_mutex);
1569 list_for_each_entry(cachep, &cache_chain, next)
1570 if (enable_cpucache(cachep, GFP_NOWAIT))
1571 BUG();
1572 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001573
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 /* Done! */
1575 g_cpucache_up = FULL;
1576
Pekka Enbergec5a36f2009-06-29 09:57:10 +03001577 /* Annotate slab for lockdep -- annotate the malloc caches */
1578 init_lock_keys();
1579
Andrew Mortona737b3e2006-03-22 00:08:11 -08001580 /*
1581 * Register a cpu startup notifier callback that initializes
1582 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 */
1584 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585
Andrew Mortona737b3e2006-03-22 00:08:11 -08001586 /*
1587 * The reap timers are started later, with a module init call: That part
1588 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 */
1590}
1591
1592static int __init cpucache_init(void)
1593{
1594 int cpu;
1595
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 /*
1597 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 */
Christoph Lametere498be72005-09-09 13:03:32 -07001599 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001600 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 return 0;
1602}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603__initcall(cpucache_init);
1604
1605/*
1606 * Interface to system's page allocator. No need to hold the cache-lock.
1607 *
1608 * If we requested dmaable memory, we will get it. Even if we
1609 * did not request dmaable memory, we might get it, but that
1610 * would be relatively rare and ignorable.
1611 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001612static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613{
1614 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001615 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 int i;
1617
Luke Yangd6fef9d2006-04-10 22:52:56 -07001618#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001619 /*
1620 * Nommu uses slab's for process anonymous memory allocations, and thus
1621 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001622 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001623 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001624#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001625
Christoph Lameter3c517a62006-12-06 20:33:29 -08001626 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001627 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1628 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001629
Linus Torvalds517d0862009-06-16 19:50:13 -07001630 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 if (!page)
1632 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001634 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001636 add_zone_page_state(page_zone(page),
1637 NR_SLAB_RECLAIMABLE, nr_pages);
1638 else
1639 add_zone_page_state(page_zone(page),
1640 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001641 for (i = 0; i < nr_pages; i++)
1642 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001643
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001644 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1645 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1646
1647 if (cachep->ctor)
1648 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1649 else
1650 kmemcheck_mark_unallocated_pages(page, nr_pages);
1651 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001652
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001653 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654}
1655
1656/*
1657 * Interface to system's page release.
1658 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001659static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001661 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 struct page *page = virt_to_page(addr);
1663 const unsigned long nr_freed = i;
1664
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001665 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001666
Christoph Lameter972d1a72006-09-25 23:31:51 -07001667 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1668 sub_zone_page_state(page_zone(page),
1669 NR_SLAB_RECLAIMABLE, nr_freed);
1670 else
1671 sub_zone_page_state(page_zone(page),
1672 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001674 BUG_ON(!PageSlab(page));
1675 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 page++;
1677 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 if (current->reclaim_state)
1679 current->reclaim_state->reclaimed_slab += nr_freed;
1680 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681}
1682
1683static void kmem_rcu_free(struct rcu_head *head)
1684{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001685 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001686 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
1688 kmem_freepages(cachep, slab_rcu->addr);
1689 if (OFF_SLAB(cachep))
1690 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1691}
1692
1693#if DEBUG
1694
1695#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001696static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001697 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001699 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001701 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001703 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 return;
1705
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001706 *addr++ = 0x12345678;
1707 *addr++ = caller;
1708 *addr++ = smp_processor_id();
1709 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 {
1711 unsigned long *sptr = &caller;
1712 unsigned long svalue;
1713
1714 while (!kstack_end(sptr)) {
1715 svalue = *sptr++;
1716 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001717 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 size -= sizeof(unsigned long);
1719 if (size <= sizeof(unsigned long))
1720 break;
1721 }
1722 }
1723
1724 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001725 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726}
1727#endif
1728
Pekka Enberg343e0d72006-02-01 03:05:50 -08001729static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001731 int size = obj_size(cachep);
1732 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733
1734 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001735 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736}
1737
1738static void dump_line(char *data, int offset, int limit)
1739{
1740 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001741 unsigned char error = 0;
1742 int bad_count = 0;
1743
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001745 for (i = 0; i < limit; i++) {
1746 if (data[offset + i] != POISON_FREE) {
1747 error = data[offset + i];
1748 bad_count++;
1749 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001750 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001751 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001753
1754 if (bad_count == 1) {
1755 error ^= POISON_FREE;
1756 if (!(error & (error - 1))) {
1757 printk(KERN_ERR "Single bit error detected. Probably "
1758 "bad RAM.\n");
1759#ifdef CONFIG_X86
1760 printk(KERN_ERR "Run memtest86+ or a similar memory "
1761 "test tool.\n");
1762#else
1763 printk(KERN_ERR "Run a memory test tool.\n");
1764#endif
1765 }
1766 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767}
1768#endif
1769
1770#if DEBUG
1771
Pekka Enberg343e0d72006-02-01 03:05:50 -08001772static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773{
1774 int i, size;
1775 char *realobj;
1776
1777 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001778 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001779 *dbg_redzone1(cachep, objp),
1780 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 }
1782
1783 if (cachep->flags & SLAB_STORE_USER) {
1784 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001785 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001787 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 printk("\n");
1789 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001790 realobj = (char *)objp + obj_offset(cachep);
1791 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001792 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 int limit;
1794 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 if (i + limit > size)
1796 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 dump_line(realobj, i, limit);
1798 }
1799}
1800
Pekka Enberg343e0d72006-02-01 03:05:50 -08001801static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802{
1803 char *realobj;
1804 int size, i;
1805 int lines = 0;
1806
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001807 realobj = (char *)objp + obj_offset(cachep);
1808 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001810 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001812 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 exp = POISON_END;
1814 if (realobj[i] != exp) {
1815 int limit;
1816 /* Mismatch ! */
1817 /* Print header */
1818 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001819 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001820 "Slab corruption: %s start=%p, len=%d\n",
1821 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 print_objinfo(cachep, objp, 0);
1823 }
1824 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001825 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 if (i + limit > size)
1828 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 dump_line(realobj, i, limit);
1830 i += 16;
1831 lines++;
1832 /* Limit to 5 lines */
1833 if (lines > 5)
1834 break;
1835 }
1836 }
1837 if (lines != 0) {
1838 /* Print some data about the neighboring objects, if they
1839 * exist:
1840 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001841 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001842 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001844 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001846 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001847 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001849 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 print_objinfo(cachep, objp, 2);
1851 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001852 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001853 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001854 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001856 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 print_objinfo(cachep, objp, 2);
1858 }
1859 }
1860}
1861#endif
1862
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301864static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001865{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 int i;
1867 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001868 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869
1870 if (cachep->flags & SLAB_POISON) {
1871#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001872 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1873 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001874 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001875 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 else
1877 check_poison_obj(cachep, objp);
1878#else
1879 check_poison_obj(cachep, objp);
1880#endif
1881 }
1882 if (cachep->flags & SLAB_RED_ZONE) {
1883 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1884 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001885 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1887 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001888 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001891}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301893static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001894{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001895}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896#endif
1897
Randy Dunlap911851e2006-03-22 00:08:14 -08001898/**
1899 * slab_destroy - destroy and release all objects in a slab
1900 * @cachep: cache pointer being destroyed
1901 * @slabp: slab pointer being destroyed
1902 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001903 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001904 * Before calling the slab must have been unlinked from the cache. The
1905 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001907static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001908{
1909 void *addr = slabp->s_mem - slabp->colouroff;
1910
Rabin Vincente79aec22008-07-04 00:40:32 +05301911 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1913 struct slab_rcu *slab_rcu;
1914
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 slab_rcu->cachep = cachep;
1917 slab_rcu->addr = addr;
1918 call_rcu(&slab_rcu->head, kmem_rcu_free);
1919 } else {
1920 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001921 if (OFF_SLAB(cachep))
1922 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 }
1924}
1925
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001926static void __kmem_cache_destroy(struct kmem_cache *cachep)
1927{
1928 int i;
1929 struct kmem_list3 *l3;
1930
1931 for_each_online_cpu(i)
1932 kfree(cachep->array[i]);
1933
1934 /* NUMA: free the list3 structures */
1935 for_each_online_node(i) {
1936 l3 = cachep->nodelists[i];
1937 if (l3) {
1938 kfree(l3->shared);
1939 free_alien_cache(l3->alien);
1940 kfree(l3);
1941 }
1942 }
1943 kmem_cache_free(&cache_cache, cachep);
1944}
1945
1946
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001948 * calculate_slab_order - calculate size (page order) of slabs
1949 * @cachep: pointer to the cache that is being created
1950 * @size: size of objects to be created in this cache.
1951 * @align: required alignment for the objects.
1952 * @flags: slab allocation flags
1953 *
1954 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001955 *
1956 * This could be made much more intelligent. For now, try to avoid using
1957 * high order pages for slabs. When the gfp() functions are more friendly
1958 * towards high-order requests, this should be changed.
1959 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001960static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001961 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001962{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001963 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001964 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001965 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001966
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001967 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001968 unsigned int num;
1969 size_t remainder;
1970
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001971 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001972 if (!num)
1973 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001974
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001975 if (flags & CFLGS_OFF_SLAB) {
1976 /*
1977 * Max number of objs-per-slab for caches which
1978 * use off-slab slabs. Needed to avoid a possible
1979 * looping condition in cache_grow().
1980 */
1981 offslab_limit = size - sizeof(struct slab);
1982 offslab_limit /= sizeof(kmem_bufctl_t);
1983
1984 if (num > offslab_limit)
1985 break;
1986 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001988 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001989 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001990 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001991 left_over = remainder;
1992
1993 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001994 * A VFS-reclaimable slab tends to have most allocations
1995 * as GFP_NOFS and we really don't want to have to be allocating
1996 * higher-order pages when we are unable to shrink dcache.
1997 */
1998 if (flags & SLAB_RECLAIM_ACCOUNT)
1999 break;
2000
2001 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002002 * Large number of objects is good, but very large slabs are
2003 * currently bad for the gfp()s.
2004 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002005 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002006 break;
2007
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002008 /*
2009 * Acceptable internal fragmentation?
2010 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002011 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012 break;
2013 }
2014 return left_over;
2015}
2016
Pekka Enberg83b519e2009-06-10 19:40:04 +03002017static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002018{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002019 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002020 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002021
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002022 if (g_cpucache_up == NONE) {
2023 /*
2024 * Note: the first kmem_cache_create must create the cache
2025 * that's used by kmalloc(24), otherwise the creation of
2026 * further caches will BUG().
2027 */
2028 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2029
2030 /*
2031 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2032 * the first cache, then we need to set up all its list3s,
2033 * otherwise the creation of further caches will BUG().
2034 */
2035 set_up_list3s(cachep, SIZE_AC);
2036 if (INDEX_AC == INDEX_L3)
2037 g_cpucache_up = PARTIAL_L3;
2038 else
2039 g_cpucache_up = PARTIAL_AC;
2040 } else {
2041 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002042 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002043
2044 if (g_cpucache_up == PARTIAL_AC) {
2045 set_up_list3s(cachep, SIZE_L3);
2046 g_cpucache_up = PARTIAL_L3;
2047 } else {
2048 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002049 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002050 cachep->nodelists[node] =
2051 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002052 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002053 BUG_ON(!cachep->nodelists[node]);
2054 kmem_list3_init(cachep->nodelists[node]);
2055 }
2056 }
2057 }
2058 cachep->nodelists[numa_node_id()]->next_reap =
2059 jiffies + REAPTIMEOUT_LIST3 +
2060 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2061
2062 cpu_cache_get(cachep)->avail = 0;
2063 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2064 cpu_cache_get(cachep)->batchcount = 1;
2065 cpu_cache_get(cachep)->touched = 0;
2066 cachep->batchcount = 1;
2067 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002068 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002069}
2070
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002071/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 * kmem_cache_create - Create a cache.
2073 * @name: A string which is used in /proc/slabinfo to identify this cache.
2074 * @size: The size of objects to be created in this cache.
2075 * @align: The required alignment for the objects.
2076 * @flags: SLAB flags
2077 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 *
2079 * Returns a ptr to the cache on success, NULL on failure.
2080 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002081 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 *
2083 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002084 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002085 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2086 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002087 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 * The flags are
2089 *
2090 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2091 * to catch references to uninitialised memory.
2092 *
2093 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2094 * for buffer overruns.
2095 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2097 * cacheline. This can be beneficial if you're counting cycles as closely
2098 * as davem.
2099 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002100struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002102 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103{
2104 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002105 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002106 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107
2108 /*
2109 * Sanity checks... these are all serious usage bugs.
2110 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002111 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002112 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002113 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002114 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002115 BUG();
2116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002118 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002119 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302120 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002121 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002122 if (slab_is_available()) {
2123 get_online_cpus();
2124 mutex_lock(&cache_chain_mutex);
2125 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002126
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002127 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002128 char tmp;
2129 int res;
2130
2131 /*
2132 * This happens when the module gets unloaded and doesn't
2133 * destroy its slab cache and no-one else reuses the vmalloc
2134 * area of the module. Print a warning.
2135 */
Andrew Morton138ae662006-12-06 20:36:41 -08002136 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002137 if (res) {
matzeb4169522007-05-06 14:49:52 -07002138 printk(KERN_ERR
2139 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002140 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002141 continue;
2142 }
2143
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002144 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002145 printk(KERN_ERR
2146 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002147 dump_stack();
2148 goto oops;
2149 }
2150 }
2151
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152#if DEBUG
2153 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154#if FORCED_DEBUG
2155 /*
2156 * Enable redzoning and last user accounting, except for caches with
2157 * large objects, if the increased size would increase the object size
2158 * above the next power of two: caches with object sizes just above a
2159 * power of two have a significant amount of internal fragmentation.
2160 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002161 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2162 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002163 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 if (!(flags & SLAB_DESTROY_BY_RCU))
2165 flags |= SLAB_POISON;
2166#endif
2167 if (flags & SLAB_DESTROY_BY_RCU)
2168 BUG_ON(flags & SLAB_POISON);
2169#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002171 * Always checks flags, a caller might be expecting debug support which
2172 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002174 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 /*
2177 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 * unaligned accesses for some archs when redzoning is used, and makes
2179 * sure any on-slab bufctl's are also correctly aligned.
2180 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002181 if (size & (BYTES_PER_WORD - 1)) {
2182 size += (BYTES_PER_WORD - 1);
2183 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 }
2185
Andrew Mortona737b3e2006-03-22 00:08:11 -08002186 /* calculate the final buffer alignment: */
2187
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 /* 1) arch recommendation: can be overridden for debug */
2189 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002190 /*
2191 * Default alignment: as specified by the arch code. Except if
2192 * an object is really small, then squeeze multiple objects into
2193 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 */
2195 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002196 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 ralign /= 2;
2198 } else {
2199 ralign = BYTES_PER_WORD;
2200 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002201
2202 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002203 * Redzoning and user store require word alignment or possibly larger.
2204 * Note this will be overridden by architecture or caller mandated
2205 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002206 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002207 if (flags & SLAB_STORE_USER)
2208 ralign = BYTES_PER_WORD;
2209
2210 if (flags & SLAB_RED_ZONE) {
2211 ralign = REDZONE_ALIGN;
2212 /* If redzoning, ensure that the second redzone is suitably
2213 * aligned, by adjusting the object size accordingly. */
2214 size += REDZONE_ALIGN - 1;
2215 size &= ~(REDZONE_ALIGN - 1);
2216 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002217
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002218 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 if (ralign < ARCH_SLAB_MINALIGN) {
2220 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002222 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 if (ralign < align) {
2224 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002226 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002227 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002228 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002229 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002230 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 */
2232 align = ralign;
2233
Pekka Enberg83b519e2009-06-10 19:40:04 +03002234 if (slab_is_available())
2235 gfp = GFP_KERNEL;
2236 else
2237 gfp = GFP_NOWAIT;
2238
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002240 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002242 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243
2244#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002245 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246
Pekka Enbergca5f9702006-09-25 23:31:25 -07002247 /*
2248 * Both debugging options require word-alignment which is calculated
2249 * into align above.
2250 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002253 cachep->obj_offset += sizeof(unsigned long long);
2254 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 }
2256 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002257 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002258 * the real object. But if the second red zone needs to be
2259 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002261 if (flags & SLAB_RED_ZONE)
2262 size += REDZONE_ALIGN;
2263 else
2264 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 }
2266#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002267 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002268 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2269 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 size = PAGE_SIZE;
2271 }
2272#endif
2273#endif
2274
Ingo Molnare0a42722006-06-23 02:03:46 -07002275 /*
2276 * Determine if the slab management is 'on' or 'off' slab.
2277 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002278 * it too early on. Always use on-slab management when
2279 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002280 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002281 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2282 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 /*
2284 * Size is large, assume best to place the slab management obj
2285 * off-slab (should allow better packing of objs).
2286 */
2287 flags |= CFLGS_OFF_SLAB;
2288
2289 size = ALIGN(size, align);
2290
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002291 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292
2293 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002294 printk(KERN_ERR
2295 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 kmem_cache_free(&cache_cache, cachep);
2297 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002298 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002300 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2301 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302
2303 /*
2304 * If the slab has been placed off-slab, and we have enough space then
2305 * move it on-slab. This is at the expense of any extra colouring.
2306 */
2307 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2308 flags &= ~CFLGS_OFF_SLAB;
2309 left_over -= slab_size;
2310 }
2311
2312 if (flags & CFLGS_OFF_SLAB) {
2313 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002314 slab_size =
2315 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302316
2317#ifdef CONFIG_PAGE_POISONING
2318 /* If we're going to use the generic kernel_map_pages()
2319 * poisoning, then it's going to smash the contents of
2320 * the redzone and userword anyhow, so switch them off.
2321 */
2322 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2323 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2324#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 }
2326
2327 cachep->colour_off = cache_line_size();
2328 /* Offset must be a multiple of the alignment. */
2329 if (cachep->colour_off < align)
2330 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002331 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 cachep->slab_size = slab_size;
2333 cachep->flags = flags;
2334 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002335 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002337 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002338 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002340 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002341 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002342 /*
2343 * This is a possibility for one of the malloc_sizes caches.
2344 * But since we go off slab only for object size greater than
2345 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2346 * this should not happen at all.
2347 * But leave a BUG_ON for some lucky dude.
2348 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002349 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002350 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 cachep->name = name;
2353
Pekka Enberg83b519e2009-06-10 19:40:04 +03002354 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002355 __kmem_cache_destroy(cachep);
2356 cachep = NULL;
2357 goto oops;
2358 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 /* cache setup completed, link it into the list */
2361 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002362oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 if (!cachep && (flags & SLAB_PANIC))
2364 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002365 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002366 if (slab_is_available()) {
2367 mutex_unlock(&cache_chain_mutex);
2368 put_online_cpus();
2369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 return cachep;
2371}
2372EXPORT_SYMBOL(kmem_cache_create);
2373
2374#if DEBUG
2375static void check_irq_off(void)
2376{
2377 BUG_ON(!irqs_disabled());
2378}
2379
2380static void check_irq_on(void)
2381{
2382 BUG_ON(irqs_disabled());
2383}
2384
Pekka Enberg343e0d72006-02-01 03:05:50 -08002385static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386{
2387#ifdef CONFIG_SMP
2388 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002389 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390#endif
2391}
Christoph Lametere498be72005-09-09 13:03:32 -07002392
Pekka Enberg343e0d72006-02-01 03:05:50 -08002393static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002394{
2395#ifdef CONFIG_SMP
2396 check_irq_off();
2397 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2398#endif
2399}
2400
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401#else
2402#define check_irq_off() do { } while(0)
2403#define check_irq_on() do { } while(0)
2404#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002405#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406#endif
2407
Christoph Lameteraab22072006-03-22 00:09:06 -08002408static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2409 struct array_cache *ac,
2410 int force, int node);
2411
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412static void do_drain(void *arg)
2413{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002414 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002416 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417
2418 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002419 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002420 spin_lock(&cachep->nodelists[node]->list_lock);
2421 free_block(cachep, ac->entry, ac->avail, node);
2422 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 ac->avail = 0;
2424}
2425
Pekka Enberg343e0d72006-02-01 03:05:50 -08002426static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427{
Christoph Lametere498be72005-09-09 13:03:32 -07002428 struct kmem_list3 *l3;
2429 int node;
2430
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002431 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002433 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002434 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002435 if (l3 && l3->alien)
2436 drain_alien_cache(cachep, l3->alien);
2437 }
2438
2439 for_each_online_node(node) {
2440 l3 = cachep->nodelists[node];
2441 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002442 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002443 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444}
2445
Christoph Lametered11d9e2006-06-30 01:55:45 -07002446/*
2447 * Remove slabs from the list of free slabs.
2448 * Specify the number of slabs to drain in tofree.
2449 *
2450 * Returns the actual number of slabs released.
2451 */
2452static int drain_freelist(struct kmem_cache *cache,
2453 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002455 struct list_head *p;
2456 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458
Christoph Lametered11d9e2006-06-30 01:55:45 -07002459 nr_freed = 0;
2460 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461
Christoph Lametered11d9e2006-06-30 01:55:45 -07002462 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002463 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002464 if (p == &l3->slabs_free) {
2465 spin_unlock_irq(&l3->list_lock);
2466 goto out;
2467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Christoph Lametered11d9e2006-06-30 01:55:45 -07002469 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002471 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472#endif
2473 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002474 /*
2475 * Safe to drop the lock. The slab is no longer linked
2476 * to the cache.
2477 */
2478 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002479 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002480 slab_destroy(cache, slabp);
2481 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002483out:
2484 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485}
2486
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002487/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002488static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002489{
2490 int ret = 0, i = 0;
2491 struct kmem_list3 *l3;
2492
2493 drain_cpu_caches(cachep);
2494
2495 check_irq_on();
2496 for_each_online_node(i) {
2497 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002498 if (!l3)
2499 continue;
2500
2501 drain_freelist(cachep, l3, l3->free_objects);
2502
2503 ret += !list_empty(&l3->slabs_full) ||
2504 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002505 }
2506 return (ret ? 1 : 0);
2507}
2508
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509/**
2510 * kmem_cache_shrink - Shrink a cache.
2511 * @cachep: The cache to shrink.
2512 *
2513 * Releases as many slabs as possible for a cache.
2514 * To help debugging, a zero exit status indicates all slabs were released.
2515 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002516int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002518 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002519 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002521 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002522 mutex_lock(&cache_chain_mutex);
2523 ret = __cache_shrink(cachep);
2524 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002525 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002526 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527}
2528EXPORT_SYMBOL(kmem_cache_shrink);
2529
2530/**
2531 * kmem_cache_destroy - delete a cache
2532 * @cachep: the cache to destroy
2533 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002534 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 *
2536 * It is expected this function will be called by a module when it is
2537 * unloaded. This will remove the cache completely, and avoid a duplicate
2538 * cache being allocated each time a module is loaded and unloaded, if the
2539 * module doesn't have persistent in-kernel storage across loads and unloads.
2540 *
2541 * The cache must be empty before calling this function.
2542 *
2543 * The caller must guarantee that noone will allocate memory from the cache
2544 * during the kmem_cache_destroy().
2545 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002546void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002548 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002551 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002552 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 /*
2554 * the chain is never empty, cache_cache is never destroyed
2555 */
2556 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 if (__cache_shrink(cachep)) {
2558 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002559 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002560 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002561 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002562 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 }
2564
2565 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002566 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002568 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002569 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002570 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571}
2572EXPORT_SYMBOL(kmem_cache_destroy);
2573
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002574/*
2575 * Get the memory for a slab management obj.
2576 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2577 * always come from malloc_sizes caches. The slab descriptor cannot
2578 * come from the same cache which is getting created because,
2579 * when we are searching for an appropriate cache for these
2580 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2581 * If we are creating a malloc_sizes cache here it would not be visible to
2582 * kmem_find_general_cachep till the initialization is complete.
2583 * Hence we cannot have slabp_cache same as the original cache.
2584 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002585static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002586 int colour_off, gfp_t local_flags,
2587 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588{
2589 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 if (OFF_SLAB(cachep)) {
2592 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002593 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002594 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002595 /*
2596 * If the first object in the slab is leaked (it's allocated
2597 * but no one has a reference to it), we want to make sure
2598 * kmemleak does not treat the ->s_mem pointer as a reference
2599 * to the object. Otherwise we will not report the leak.
2600 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002601 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2602 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 if (!slabp)
2604 return NULL;
2605 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002606 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 colour_off += cachep->slab_size;
2608 }
2609 slabp->inuse = 0;
2610 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002611 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002612 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002613 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 return slabp;
2615}
2616
2617static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2618{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002619 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620}
2621
Pekka Enberg343e0d72006-02-01 03:05:50 -08002622static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002623 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624{
2625 int i;
2626
2627 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002628 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629#if DEBUG
2630 /* need to poison the objs? */
2631 if (cachep->flags & SLAB_POISON)
2632 poison_obj(cachep, objp, POISON_FREE);
2633 if (cachep->flags & SLAB_STORE_USER)
2634 *dbg_userword(cachep, objp) = NULL;
2635
2636 if (cachep->flags & SLAB_RED_ZONE) {
2637 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2638 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2639 }
2640 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002641 * Constructors are not allowed to allocate memory from the same
2642 * cache which they are a constructor for. Otherwise, deadlock.
2643 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 */
2645 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002646 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647
2648 if (cachep->flags & SLAB_RED_ZONE) {
2649 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2650 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002651 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2653 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002654 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002656 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2657 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002658 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002659 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660#else
2661 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002662 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002664 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002666 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667}
2668
Pekka Enberg343e0d72006-02-01 03:05:50 -08002669static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002671 if (CONFIG_ZONE_DMA_FLAG) {
2672 if (flags & GFP_DMA)
2673 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2674 else
2675 BUG_ON(cachep->gfpflags & GFP_DMA);
2676 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677}
2678
Andrew Mortona737b3e2006-03-22 00:08:11 -08002679static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2680 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002681{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002682 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002683 kmem_bufctl_t next;
2684
2685 slabp->inuse++;
2686 next = slab_bufctl(slabp)[slabp->free];
2687#if DEBUG
2688 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2689 WARN_ON(slabp->nodeid != nodeid);
2690#endif
2691 slabp->free = next;
2692
2693 return objp;
2694}
2695
Andrew Mortona737b3e2006-03-22 00:08:11 -08002696static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2697 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002698{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002699 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002700
2701#if DEBUG
2702 /* Verify that the slab belongs to the intended node */
2703 WARN_ON(slabp->nodeid != nodeid);
2704
Al Viro871751e2006-03-25 03:06:39 -08002705 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002706 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002707 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002708 BUG();
2709 }
2710#endif
2711 slab_bufctl(slabp)[objnr] = slabp->free;
2712 slabp->free = objnr;
2713 slabp->inuse--;
2714}
2715
Pekka Enberg47768742006-06-23 02:03:07 -07002716/*
2717 * Map pages beginning at addr to the given cache and slab. This is required
2718 * for the slab allocator to be able to lookup the cache and slab of a
2719 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2720 */
2721static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2722 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723{
Pekka Enberg47768742006-06-23 02:03:07 -07002724 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 struct page *page;
2726
Pekka Enberg47768742006-06-23 02:03:07 -07002727 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002728
Pekka Enberg47768742006-06-23 02:03:07 -07002729 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002730 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002731 nr_pages <<= cache->gfporder;
2732
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002734 page_set_cache(page, cache);
2735 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002737 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738}
2739
2740/*
2741 * Grow (by 1) the number of slabs within a cache. This is called by
2742 * kmem_cache_alloc() when there are no active objs left in a cache.
2743 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002744static int cache_grow(struct kmem_cache *cachep,
2745 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002747 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002748 size_t offset;
2749 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002750 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751
Andrew Mortona737b3e2006-03-22 00:08:11 -08002752 /*
2753 * Be lazy and only check for valid flags here, keeping it out of the
2754 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002756 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2757 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002759 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002761 l3 = cachep->nodelists[nodeid];
2762 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
2764 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002765 offset = l3->colour_next;
2766 l3->colour_next++;
2767 if (l3->colour_next >= cachep->colour)
2768 l3->colour_next = 0;
2769 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002771 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772
2773 if (local_flags & __GFP_WAIT)
2774 local_irq_enable();
2775
2776 /*
2777 * The test for missing atomic flag is performed here, rather than
2778 * the more obvious place, simply to reduce the critical path length
2779 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2780 * will eventually be caught here (where it matters).
2781 */
2782 kmem_flagcheck(cachep, flags);
2783
Andrew Mortona737b3e2006-03-22 00:08:11 -08002784 /*
2785 * Get mem for the objs. Attempt to allocate a physical page from
2786 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002787 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002788 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002789 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002790 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 goto failed;
2792
2793 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002794 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002795 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002796 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 goto opps1;
2798
Pekka Enberg47768742006-06-23 02:03:07 -07002799 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800
Christoph Lametera35afb82007-05-16 22:10:57 -07002801 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802
2803 if (local_flags & __GFP_WAIT)
2804 local_irq_disable();
2805 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002806 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
2808 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002809 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002811 l3->free_objects += cachep->num;
2812 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002814opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002816failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 if (local_flags & __GFP_WAIT)
2818 local_irq_disable();
2819 return 0;
2820}
2821
2822#if DEBUG
2823
2824/*
2825 * Perform extra freeing checks:
2826 * - detect bad pointers.
2827 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 */
2829static void kfree_debugcheck(const void *objp)
2830{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 if (!virt_addr_valid(objp)) {
2832 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002833 (unsigned long)objp);
2834 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836}
2837
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002838static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2839{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002840 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002841
2842 redzone1 = *dbg_redzone1(cache, obj);
2843 redzone2 = *dbg_redzone2(cache, obj);
2844
2845 /*
2846 * Redzone is ok.
2847 */
2848 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2849 return;
2850
2851 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2852 slab_error(cache, "double free detected");
2853 else
2854 slab_error(cache, "memory outside object was overwritten");
2855
David Woodhouseb46b8f12007-05-08 00:22:59 -07002856 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002857 obj, redzone1, redzone2);
2858}
2859
Pekka Enberg343e0d72006-02-01 03:05:50 -08002860static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002861 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862{
2863 struct page *page;
2864 unsigned int objnr;
2865 struct slab *slabp;
2866
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002867 BUG_ON(virt_to_cache(objp) != cachep);
2868
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002869 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002871 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872
Pekka Enberg065d41c2005-11-13 16:06:46 -08002873 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874
2875 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002876 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2879 }
2880 if (cachep->flags & SLAB_STORE_USER)
2881 *dbg_userword(cachep, objp) = caller;
2882
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002883 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884
2885 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002886 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887
Al Viro871751e2006-03-25 03:06:39 -08002888#ifdef CONFIG_DEBUG_SLAB_LEAK
2889 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2890#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 if (cachep->flags & SLAB_POISON) {
2892#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002893 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002895 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002896 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 } else {
2898 poison_obj(cachep, objp, POISON_FREE);
2899 }
2900#else
2901 poison_obj(cachep, objp, POISON_FREE);
2902#endif
2903 }
2904 return objp;
2905}
2906
Pekka Enberg343e0d72006-02-01 03:05:50 -08002907static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908{
2909 kmem_bufctl_t i;
2910 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002911
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 /* Check slab's freelist to see if this obj is there. */
2913 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2914 entries++;
2915 if (entries > cachep->num || i >= cachep->num)
2916 goto bad;
2917 }
2918 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002919bad:
2920 printk(KERN_ERR "slab: Internal list corruption detected in "
2921 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2922 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002923 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002924 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002925 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002926 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002928 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 }
2930 printk("\n");
2931 BUG();
2932 }
2933}
2934#else
2935#define kfree_debugcheck(x) do { } while(0)
2936#define cache_free_debugcheck(x,objp,z) (objp)
2937#define check_slabp(x,y) do { } while(0)
2938#endif
2939
Pekka Enberg343e0d72006-02-01 03:05:50 -08002940static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941{
2942 int batchcount;
2943 struct kmem_list3 *l3;
2944 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002945 int node;
2946
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002948 check_irq_off();
2949 node = numa_node_id();
2950 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 batchcount = ac->batchcount;
2952 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002953 /*
2954 * If there was little recent activity on this cache, then
2955 * perform only a partial refill. Otherwise we could generate
2956 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 */
2958 batchcount = BATCHREFILL_LIMIT;
2959 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002960 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961
Christoph Lametere498be72005-09-09 13:03:32 -07002962 BUG_ON(ac->avail > 0 || !l3);
2963 spin_lock(&l3->list_lock);
2964
Christoph Lameter3ded1752006-03-25 03:06:44 -08002965 /* See if we can refill from the shared array */
2966 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2967 goto alloc_done;
2968
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 while (batchcount > 0) {
2970 struct list_head *entry;
2971 struct slab *slabp;
2972 /* Get slab alloc is to come from. */
2973 entry = l3->slabs_partial.next;
2974 if (entry == &l3->slabs_partial) {
2975 l3->free_touched = 1;
2976 entry = l3->slabs_free.next;
2977 if (entry == &l3->slabs_free)
2978 goto must_grow;
2979 }
2980
2981 slabp = list_entry(entry, struct slab, list);
2982 check_slabp(cachep, slabp);
2983 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07002984
2985 /*
2986 * The slab was either on partial or free list so
2987 * there must be at least one object available for
2988 * allocation.
2989 */
roel kluin249b9f32008-10-29 17:18:07 -04002990 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07002991
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 STATS_INC_ALLOCED(cachep);
2994 STATS_INC_ACTIVE(cachep);
2995 STATS_SET_HIGH(cachep);
2996
Matthew Dobson78d382d2006-02-01 03:05:47 -08002997 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002998 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 }
3000 check_slabp(cachep, slabp);
3001
3002 /* move slabp to correct slabp list: */
3003 list_del(&slabp->list);
3004 if (slabp->free == BUFCTL_END)
3005 list_add(&slabp->list, &l3->slabs_full);
3006 else
3007 list_add(&slabp->list, &l3->slabs_partial);
3008 }
3009
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003013 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014
3015 if (unlikely(!ac->avail)) {
3016 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003017 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003018
Andrew Mortona737b3e2006-03-22 00:08:11 -08003019 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003020 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 return NULL;
3023
Andrew Mortona737b3e2006-03-22 00:08:11 -08003024 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 goto retry;
3026 }
3027 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003028 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029}
3030
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3032 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033{
3034 might_sleep_if(flags & __GFP_WAIT);
3035#if DEBUG
3036 kmem_flagcheck(cachep, flags);
3037#endif
3038}
3039
3040#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003041static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3042 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003044 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003046 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003048 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003049 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003050 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 else
3052 check_poison_obj(cachep, objp);
3053#else
3054 check_poison_obj(cachep, objp);
3055#endif
3056 poison_obj(cachep, objp, POISON_INUSE);
3057 }
3058 if (cachep->flags & SLAB_STORE_USER)
3059 *dbg_userword(cachep, objp) = caller;
3060
3061 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3063 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3064 slab_error(cachep, "double free, or memory outside"
3065 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003066 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003067 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003068 objp, *dbg_redzone1(cachep, objp),
3069 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 }
3071 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3072 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3073 }
Al Viro871751e2006-03-25 03:06:39 -08003074#ifdef CONFIG_DEBUG_SLAB_LEAK
3075 {
3076 struct slab *slabp;
3077 unsigned objnr;
3078
Christoph Lameterb49af682007-05-06 14:49:41 -07003079 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003080 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3081 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3082 }
3083#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003084 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003085 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003086 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003087#if ARCH_SLAB_MINALIGN
3088 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3089 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3090 objp, ARCH_SLAB_MINALIGN);
3091 }
3092#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 return objp;
3094}
3095#else
3096#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3097#endif
3098
Akinobu Mita773ff602008-12-23 19:37:01 +09003099static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003100{
3101 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003102 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003103
Akinobu Mita773ff602008-12-23 19:37:01 +09003104 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003105}
3106
Pekka Enberg343e0d72006-02-01 03:05:50 -08003107static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003109 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 struct array_cache *ac;
3111
Alok N Kataria5c382302005-09-27 21:45:46 -07003112 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003113
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003114 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 if (likely(ac->avail)) {
3116 STATS_INC_ALLOCHIT(cachep);
3117 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003118 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 } else {
3120 STATS_INC_ALLOCMISS(cachep);
3121 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003122 /*
3123 * the 'ac' may be updated by cache_alloc_refill(),
3124 * and kmemleak_erase() requires its correct value.
3125 */
3126 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003128 /*
3129 * To avoid a false negative, if an object that is in one of the
3130 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3131 * treat the array pointers as a reference to the object.
3132 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003133 if (objp)
3134 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003135 return objp;
3136}
3137
Christoph Lametere498be72005-09-09 13:03:32 -07003138#ifdef CONFIG_NUMA
3139/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003140 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003141 *
3142 * If we are in_interrupt, then process context, including cpusets and
3143 * mempolicy, may not apply and should not be used for allocation policy.
3144 */
3145static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3146{
3147 int nid_alloc, nid_here;
3148
Christoph Lameter765c4502006-09-27 01:50:08 -07003149 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003150 return NULL;
3151 nid_alloc = nid_here = numa_node_id();
3152 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3153 nid_alloc = cpuset_mem_spread_node();
3154 else if (current->mempolicy)
3155 nid_alloc = slab_node(current->mempolicy);
3156 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003157 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003158 return NULL;
3159}
3160
3161/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003162 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003163 * certain node and fall back is permitted. First we scan all the
3164 * available nodelists for available objects. If that fails then we
3165 * perform an allocation without specifying a node. This allows the page
3166 * allocator to do its reclaim / fallback magic. We then insert the
3167 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003168 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003169static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003170{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003171 struct zonelist *zonelist;
3172 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003173 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003174 struct zone *zone;
3175 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003176 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003177 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003178
3179 if (flags & __GFP_THISNODE)
3180 return NULL;
3181
Mel Gorman0e884602008-04-28 02:12:14 -07003182 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003183 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003184
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185retry:
3186 /*
3187 * Look through allowed nodes for objects available
3188 * from existing per node queues.
3189 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003190 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3191 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003192
Mel Gorman54a6eb52008-04-28 02:12:16 -07003193 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003194 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003195 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003196 obj = ____cache_alloc_node(cache,
3197 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003198 if (obj)
3199 break;
3200 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003201 }
3202
Christoph Lametercfce6602007-05-06 14:50:17 -07003203 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003204 /*
3205 * This allocation will be performed within the constraints
3206 * of the current cpuset / memory policy requirements.
3207 * We may trigger various forms of reclaim on the allowed
3208 * set and go into memory reserves if necessary.
3209 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003210 if (local_flags & __GFP_WAIT)
3211 local_irq_enable();
3212 kmem_flagcheck(cache, flags);
Mel Gorman6484eb32009-06-16 15:31:54 -07003213 obj = kmem_getpages(cache, local_flags, numa_node_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003214 if (local_flags & __GFP_WAIT)
3215 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003216 if (obj) {
3217 /*
3218 * Insert into the appropriate per node queues
3219 */
3220 nid = page_to_nid(virt_to_page(obj));
3221 if (cache_grow(cache, flags, nid, obj)) {
3222 obj = ____cache_alloc_node(cache,
3223 flags | GFP_THISNODE, nid);
3224 if (!obj)
3225 /*
3226 * Another processor may allocate the
3227 * objects in the slab since we are
3228 * not holding any locks.
3229 */
3230 goto retry;
3231 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003232 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003233 obj = NULL;
3234 }
3235 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003236 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003237 return obj;
3238}
3239
3240/*
Christoph Lametere498be72005-09-09 13:03:32 -07003241 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003243static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003244 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003245{
3246 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003247 struct slab *slabp;
3248 struct kmem_list3 *l3;
3249 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003250 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003252 l3 = cachep->nodelists[nodeid];
3253 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003254
Andrew Mortona737b3e2006-03-22 00:08:11 -08003255retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003256 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003257 spin_lock(&l3->list_lock);
3258 entry = l3->slabs_partial.next;
3259 if (entry == &l3->slabs_partial) {
3260 l3->free_touched = 1;
3261 entry = l3->slabs_free.next;
3262 if (entry == &l3->slabs_free)
3263 goto must_grow;
3264 }
Christoph Lametere498be72005-09-09 13:03:32 -07003265
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003266 slabp = list_entry(entry, struct slab, list);
3267 check_spinlock_acquired_node(cachep, nodeid);
3268 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003269
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003270 STATS_INC_NODEALLOCS(cachep);
3271 STATS_INC_ACTIVE(cachep);
3272 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003274 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003275
Matthew Dobson78d382d2006-02-01 03:05:47 -08003276 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003277 check_slabp(cachep, slabp);
3278 l3->free_objects--;
3279 /* move slabp to correct slabp list: */
3280 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Andrew Mortona737b3e2006-03-22 00:08:11 -08003282 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003283 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003284 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003285 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003286
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003287 spin_unlock(&l3->list_lock);
3288 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003289
Andrew Mortona737b3e2006-03-22 00:08:11 -08003290must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003292 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003293 if (x)
3294 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003295
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003296 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003297
Andrew Mortona737b3e2006-03-22 00:08:11 -08003298done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003299 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003300}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003301
3302/**
3303 * kmem_cache_alloc_node - Allocate an object on the specified node
3304 * @cachep: The cache to allocate from.
3305 * @flags: See kmalloc().
3306 * @nodeid: node number of the target node.
3307 * @caller: return address of caller, used for debug information
3308 *
3309 * Identical to kmem_cache_alloc but it will allocate memory on the given
3310 * node, which can improve the performance for cpu bound structures.
3311 *
3312 * Fallback to other node is possible if __GFP_THISNODE is not set.
3313 */
3314static __always_inline void *
3315__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3316 void *caller)
3317{
3318 unsigned long save_flags;
3319 void *ptr;
3320
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003321 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003322
Nick Piggincf40bd12009-01-21 08:12:39 +01003323 lockdep_trace_alloc(flags);
3324
Akinobu Mita773ff602008-12-23 19:37:01 +09003325 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003326 return NULL;
3327
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003328 cache_alloc_debugcheck_before(cachep, flags);
3329 local_irq_save(save_flags);
3330
Tim Blechmann8e15b792009-11-30 18:59:34 +01003331 if (nodeid == -1)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003332 nodeid = numa_node_id();
3333
3334 if (unlikely(!cachep->nodelists[nodeid])) {
3335 /* Node not bootstrapped yet */
3336 ptr = fallback_alloc(cachep, flags);
3337 goto out;
3338 }
3339
3340 if (nodeid == numa_node_id()) {
3341 /*
3342 * Use the locally cached objects if possible.
3343 * However ____cache_alloc does not allow fallback
3344 * to other nodes. It may fail while we still have
3345 * objects on other nodes available.
3346 */
3347 ptr = ____cache_alloc(cachep, flags);
3348 if (ptr)
3349 goto out;
3350 }
3351 /* ___cache_alloc_node can fall back to other nodes */
3352 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3353 out:
3354 local_irq_restore(save_flags);
3355 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003356 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3357 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003358
Pekka Enbergc175eea2008-05-09 20:35:53 +02003359 if (likely(ptr))
3360 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3361
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003362 if (unlikely((flags & __GFP_ZERO) && ptr))
3363 memset(ptr, 0, obj_size(cachep));
3364
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003365 return ptr;
3366}
3367
3368static __always_inline void *
3369__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3370{
3371 void *objp;
3372
3373 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3374 objp = alternate_node_alloc(cache, flags);
3375 if (objp)
3376 goto out;
3377 }
3378 objp = ____cache_alloc(cache, flags);
3379
3380 /*
3381 * We may just have run out of memory on the local node.
3382 * ____cache_alloc_node() knows how to locate memory on other nodes
3383 */
3384 if (!objp)
3385 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3386
3387 out:
3388 return objp;
3389}
3390#else
3391
3392static __always_inline void *
3393__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3394{
3395 return ____cache_alloc(cachep, flags);
3396}
3397
3398#endif /* CONFIG_NUMA */
3399
3400static __always_inline void *
3401__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3402{
3403 unsigned long save_flags;
3404 void *objp;
3405
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003406 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003407
Nick Piggincf40bd12009-01-21 08:12:39 +01003408 lockdep_trace_alloc(flags);
3409
Akinobu Mita773ff602008-12-23 19:37:01 +09003410 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003411 return NULL;
3412
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003413 cache_alloc_debugcheck_before(cachep, flags);
3414 local_irq_save(save_flags);
3415 objp = __do_cache_alloc(cachep, flags);
3416 local_irq_restore(save_flags);
3417 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003418 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3419 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003420 prefetchw(objp);
3421
Pekka Enbergc175eea2008-05-09 20:35:53 +02003422 if (likely(objp))
3423 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3424
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003425 if (unlikely((flags & __GFP_ZERO) && objp))
3426 memset(objp, 0, obj_size(cachep));
3427
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003428 return objp;
3429}
Christoph Lametere498be72005-09-09 13:03:32 -07003430
3431/*
3432 * Caller needs to acquire correct kmem_list's list_lock
3433 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003434static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003435 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436{
3437 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003438 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439
3440 for (i = 0; i < nr_objects; i++) {
3441 void *objp = objpp[i];
3442 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003444 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003445 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003447 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003449 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003451 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 check_slabp(cachep, slabp);
3453
3454 /* fixup slab chains */
3455 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003456 if (l3->free_objects > l3->free_limit) {
3457 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003458 /* No need to drop any previously held
3459 * lock here, even if we have a off-slab slab
3460 * descriptor it is guaranteed to come from
3461 * a different cache, refer to comments before
3462 * alloc_slabmgmt.
3463 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 slab_destroy(cachep, slabp);
3465 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003466 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 }
3468 } else {
3469 /* Unconditionally move a slab to the end of the
3470 * partial list on free - maximum time for the
3471 * other objects to be freed, too.
3472 */
Christoph Lametere498be72005-09-09 13:03:32 -07003473 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 }
3475 }
3476}
3477
Pekka Enberg343e0d72006-02-01 03:05:50 -08003478static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479{
3480 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003481 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003482 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
3484 batchcount = ac->batchcount;
3485#if DEBUG
3486 BUG_ON(!batchcount || batchcount > ac->avail);
3487#endif
3488 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003489 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003490 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003491 if (l3->shared) {
3492 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003493 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 if (max) {
3495 if (batchcount > max)
3496 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003497 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003498 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 shared_array->avail += batchcount;
3500 goto free_done;
3501 }
3502 }
3503
Christoph Lameterff694162005-09-22 21:44:02 -07003504 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003505free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506#if STATS
3507 {
3508 int i = 0;
3509 struct list_head *p;
3510
Christoph Lametere498be72005-09-09 13:03:32 -07003511 p = l3->slabs_free.next;
3512 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 struct slab *slabp;
3514
3515 slabp = list_entry(p, struct slab, list);
3516 BUG_ON(slabp->inuse);
3517
3518 i++;
3519 p = p->next;
3520 }
3521 STATS_SET_FREEABLE(cachep, i);
3522 }
3523#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003524 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003526 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527}
3528
3529/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003530 * Release an obj back to its cache. If the obj has a constructed state, it must
3531 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003533static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003535 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
3537 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003538 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3540
Pekka Enbergc175eea2008-05-09 20:35:53 +02003541 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3542
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003543 /*
3544 * Skip calling cache_free_alien() when the platform is not numa.
3545 * This will avoid cache misses that happen while accessing slabp (which
3546 * is per page memory reference) to get nodeid. Instead use a global
3547 * variable to skip the call, which is mostly likely to be present in
3548 * the cache.
3549 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003550 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003551 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003552
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 if (likely(ac->avail < ac->limit)) {
3554 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003555 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 return;
3557 } else {
3558 STATS_INC_FREEMISS(cachep);
3559 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003560 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 }
3562}
3563
3564/**
3565 * kmem_cache_alloc - Allocate an object
3566 * @cachep: The cache to allocate from.
3567 * @flags: See kmalloc().
3568 *
3569 * Allocate an object from this cache. The flags are only relevant
3570 * if the cache has no available objects.
3571 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003572void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003574 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3575
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003576 trace_kmem_cache_alloc(_RET_IP_, ret,
3577 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003578
3579 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580}
3581EXPORT_SYMBOL(kmem_cache_alloc);
3582
Li Zefan0f24f122009-12-11 15:45:30 +08003583#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003584void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3585{
3586 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3587}
3588EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3589#endif
3590
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591/**
Randy Dunlap76824862008-03-19 17:00:40 -07003592 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 * @cachep: the cache we're checking against
3594 * @ptr: pointer to validate
3595 *
Randy Dunlap76824862008-03-19 17:00:40 -07003596 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 * it is _not_ a guarantee that the pointer is actually
3598 * part of the slab cache in question, but it at least
3599 * validates that the pointer can be dereferenced and
3600 * looks half-way sane.
3601 *
3602 * Currently only used for dentry validation.
3603 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003604int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003606 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003608 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003609 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610 struct page *page;
3611
3612 if (unlikely(addr < min_addr))
3613 goto out;
3614 if (unlikely(addr > (unsigned long)high_memory - size))
3615 goto out;
3616 if (unlikely(addr & align_mask))
3617 goto out;
3618 if (unlikely(!kern_addr_valid(addr)))
3619 goto out;
3620 if (unlikely(!kern_addr_valid(addr + size - 1)))
3621 goto out;
3622 page = virt_to_page(ptr);
3623 if (unlikely(!PageSlab(page)))
3624 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003625 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 goto out;
3627 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003628out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 return 0;
3630}
3631
3632#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003633void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3634{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003635 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3636 __builtin_return_address(0));
3637
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003638 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3639 obj_size(cachep), cachep->buffer_size,
3640 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003641
3642 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003643}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644EXPORT_SYMBOL(kmem_cache_alloc_node);
3645
Li Zefan0f24f122009-12-11 15:45:30 +08003646#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003647void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3648 gfp_t flags,
3649 int nodeid)
3650{
3651 return __cache_alloc_node(cachep, flags, nodeid,
3652 __builtin_return_address(0));
3653}
3654EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3655#endif
3656
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003657static __always_inline void *
3658__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003659{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003660 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003661 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003662
3663 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003664 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3665 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003666 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3667
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003668 trace_kmalloc_node((unsigned long) caller, ret,
3669 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003670
3671 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003672}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003673
Li Zefan0bb38a52009-12-11 15:45:50 +08003674#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003675void *__kmalloc_node(size_t size, gfp_t flags, int node)
3676{
3677 return __do_kmalloc_node(size, flags, node,
3678 __builtin_return_address(0));
3679}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003680EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003681
3682void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003683 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003684{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003685 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003686}
3687EXPORT_SYMBOL(__kmalloc_node_track_caller);
3688#else
3689void *__kmalloc_node(size_t size, gfp_t flags, int node)
3690{
3691 return __do_kmalloc_node(size, flags, node, NULL);
3692}
3693EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003694#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003695#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696
3697/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003698 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003700 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003701 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003703static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3704 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003706 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003707 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003709 /* If you want to save a few bytes .text space: replace
3710 * __ with kmem_.
3711 * Then kmalloc uses the uninlined functions instead of the inline
3712 * functions.
3713 */
3714 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003715 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3716 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003717 ret = __cache_alloc(cachep, flags, caller);
3718
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003719 trace_kmalloc((unsigned long) caller, ret,
3720 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003721
3722 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003723}
3724
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003725
Li Zefan0bb38a52009-12-11 15:45:50 +08003726#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003727void *__kmalloc(size_t size, gfp_t flags)
3728{
Al Viro871751e2006-03-25 03:06:39 -08003729 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730}
3731EXPORT_SYMBOL(__kmalloc);
3732
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003733void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003734{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003735 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003736}
3737EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003738
3739#else
3740void *__kmalloc(size_t size, gfp_t flags)
3741{
3742 return __do_kmalloc(size, flags, NULL);
3743}
3744EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003745#endif
3746
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747/**
3748 * kmem_cache_free - Deallocate an object
3749 * @cachep: The cache the allocation was from.
3750 * @objp: The previously allocated object.
3751 *
3752 * Free an object which was previously allocated from this
3753 * cache.
3754 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003755void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756{
3757 unsigned long flags;
3758
3759 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003760 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003761 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3762 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003763 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003765
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003766 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767}
3768EXPORT_SYMBOL(kmem_cache_free);
3769
3770/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 * kfree - free previously allocated memory
3772 * @objp: pointer returned by kmalloc.
3773 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003774 * If @objp is NULL, no operation is performed.
3775 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776 * Don't free memory not originally allocated by kmalloc()
3777 * or you will run into trouble.
3778 */
3779void kfree(const void *objp)
3780{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003781 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782 unsigned long flags;
3783
Pekka Enberg2121db72009-03-25 11:05:57 +02003784 trace_kfree(_RET_IP_, objp);
3785
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003786 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 return;
3788 local_irq_save(flags);
3789 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003790 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003791 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003792 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003793 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 local_irq_restore(flags);
3795}
3796EXPORT_SYMBOL(kfree);
3797
Pekka Enberg343e0d72006-02-01 03:05:50 -08003798unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003800 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801}
3802EXPORT_SYMBOL(kmem_cache_size);
3803
Pekka Enberg343e0d72006-02-01 03:05:50 -08003804const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003805{
3806 return cachep->name;
3807}
3808EXPORT_SYMBOL_GPL(kmem_cache_name);
3809
Christoph Lametere498be72005-09-09 13:03:32 -07003810/*
Simon Arlott183ff222007-10-20 01:27:18 +02003811 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003812 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003813static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003814{
3815 int node;
3816 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003817 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003818 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003819
Mel Gorman9c09a952008-01-24 05:49:54 -08003820 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003821
Paul Menage3395ee02006-12-06 20:32:16 -08003822 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003823 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003824 if (!new_alien)
3825 goto fail;
3826 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003827
Eric Dumazet63109842007-05-06 14:49:28 -07003828 new_shared = NULL;
3829 if (cachep->shared) {
3830 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003831 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003832 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003833 if (!new_shared) {
3834 free_alien_cache(new_alien);
3835 goto fail;
3836 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003837 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003838
Andrew Mortona737b3e2006-03-22 00:08:11 -08003839 l3 = cachep->nodelists[node];
3840 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 struct array_cache *shared = l3->shared;
3842
Christoph Lametere498be72005-09-09 13:03:32 -07003843 spin_lock_irq(&l3->list_lock);
3844
Christoph Lametercafeb022006-03-25 03:06:46 -08003845 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003846 free_block(cachep, shared->entry,
3847 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003848
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 l3->shared = new_shared;
3850 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003851 l3->alien = new_alien;
3852 new_alien = NULL;
3853 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003854 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003855 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003856 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003857 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003858 free_alien_cache(new_alien);
3859 continue;
3860 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003861 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003862 if (!l3) {
3863 free_alien_cache(new_alien);
3864 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003865 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003866 }
Christoph Lametere498be72005-09-09 13:03:32 -07003867
3868 kmem_list3_init(l3);
3869 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003870 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003871 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003872 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003873 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003874 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003875 cachep->nodelists[node] = l3;
3876 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003877 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003878
Andrew Mortona737b3e2006-03-22 00:08:11 -08003879fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003880 if (!cachep->next.next) {
3881 /* Cache is not active yet. Roll back what we did */
3882 node--;
3883 while (node >= 0) {
3884 if (cachep->nodelists[node]) {
3885 l3 = cachep->nodelists[node];
3886
3887 kfree(l3->shared);
3888 free_alien_cache(l3->alien);
3889 kfree(l3);
3890 cachep->nodelists[node] = NULL;
3891 }
3892 node--;
3893 }
3894 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003895 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003896}
3897
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003899 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900 struct array_cache *new[NR_CPUS];
3901};
3902
3903static void do_ccupdate_local(void *info)
3904{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003905 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 struct array_cache *old;
3907
3908 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003909 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003910
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3912 new->new[smp_processor_id()] = old;
3913}
3914
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003915/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003916static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003917 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003919 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003920 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921
Pekka Enberg83b519e2009-06-10 19:40:04 +03003922 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003923 if (!new)
3924 return -ENOMEM;
3925
Christoph Lametere498be72005-09-09 13:03:32 -07003926 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003927 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003928 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003929 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003930 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003931 kfree(new->new[i]);
3932 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003933 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 }
3935 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003936 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003938 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003939
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 cachep->batchcount = batchcount;
3942 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003943 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944
Christoph Lametere498be72005-09-09 13:03:32 -07003945 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003946 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 if (!ccold)
3948 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003949 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003950 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003951 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 kfree(ccold);
3953 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003954 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003955 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956}
3957
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003958/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003959static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960{
3961 int err;
3962 int limit, shared;
3963
Andrew Mortona737b3e2006-03-22 00:08:11 -08003964 /*
3965 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 * - create a LIFO ordering, i.e. return objects that are cache-warm
3967 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003968 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 * bufctl chains: array operations are cheaper.
3970 * The numbers are guessed, we should auto-tune as described by
3971 * Bonwick.
3972 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003973 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003975 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003977 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003979 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 limit = 54;
3981 else
3982 limit = 120;
3983
Andrew Mortona737b3e2006-03-22 00:08:11 -08003984 /*
3985 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 * allocation behaviour: Most allocs on one cpu, most free operations
3987 * on another cpu. For these cases, an efficient object passing between
3988 * cpus is necessary. This is provided by a shared array. The array
3989 * replaces Bonwick's magazine layer.
3990 * On uniprocessor, it's functionally equivalent (but less efficient)
3991 * to a larger limit. Thus disabled by default.
3992 */
3993 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003994 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996
3997#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003998 /*
3999 * With debugging enabled, large batchcount lead to excessively long
4000 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 */
4002 if (limit > 32)
4003 limit = 32;
4004#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004005 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 if (err)
4007 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004008 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004009 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010}
4011
Christoph Lameter1b552532006-03-22 00:09:07 -08004012/*
4013 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004014 * necessary. Note that the l3 listlock also protects the array_cache
4015 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004016 */
4017void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4018 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019{
4020 int tofree;
4021
Christoph Lameter1b552532006-03-22 00:09:07 -08004022 if (!ac || !ac->avail)
4023 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024 if (ac->touched && !force) {
4025 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004026 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004027 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004028 if (ac->avail) {
4029 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4030 if (tofree > ac->avail)
4031 tofree = (ac->avail + 1) / 2;
4032 free_block(cachep, ac->entry, tofree, node);
4033 ac->avail -= tofree;
4034 memmove(ac->entry, &(ac->entry[tofree]),
4035 sizeof(void *) * ac->avail);
4036 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004037 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 }
4039}
4040
4041/**
4042 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004043 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 *
4045 * Called from workqueue/eventd every few seconds.
4046 * Purpose:
4047 * - clear the per-cpu caches for this CPU.
4048 * - return freeable pages to the main free memory pool.
4049 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004050 * If we cannot acquire the cache chain mutex then just give up - we'll try
4051 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004053static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004055 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004056 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004057 int node = numa_node_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004058 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004060 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004062 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004064 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 check_irq_on();
4066
Christoph Lameter35386e32006-03-22 00:09:05 -08004067 /*
4068 * We only take the l3 lock if absolutely necessary and we
4069 * have established with reasonable certainty that
4070 * we can do some work if the lock was obtained.
4071 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004072 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004073
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004074 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075
Christoph Lameteraab22072006-03-22 00:09:06 -08004076 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077
Christoph Lameter35386e32006-03-22 00:09:05 -08004078 /*
4079 * These are racy checks but it does not matter
4080 * if we skip one check or scan twice.
4081 */
Christoph Lametere498be72005-09-09 13:03:32 -07004082 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004083 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084
Christoph Lametere498be72005-09-09 13:03:32 -07004085 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086
Christoph Lameteraab22072006-03-22 00:09:06 -08004087 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088
Christoph Lametered11d9e2006-06-30 01:55:45 -07004089 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004090 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004091 else {
4092 int freed;
4093
4094 freed = drain_freelist(searchp, l3, (l3->free_limit +
4095 5 * searchp->num - 1) / (5 * searchp->num));
4096 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004098next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099 cond_resched();
4100 }
4101 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004102 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004103 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004104out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004105 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004106 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107}
4108
Linus Torvalds158a9622008-01-02 13:04:48 -08004109#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110
Pekka Enberg85289f92006-01-08 01:00:36 -08004111static void print_slabinfo_header(struct seq_file *m)
4112{
4113 /*
4114 * Output format version, so at least we can change it
4115 * without _too_ many complaints.
4116 */
4117#if STATS
4118 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4119#else
4120 seq_puts(m, "slabinfo - version: 2.1\n");
4121#endif
4122 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4123 "<objperslab> <pagesperslab>");
4124 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4125 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4126#if STATS
4127 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004128 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004129 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4130#endif
4131 seq_putc(m, '\n');
4132}
4133
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134static void *s_start(struct seq_file *m, loff_t *pos)
4135{
4136 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004138 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004139 if (!n)
4140 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004141
4142 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143}
4144
4145static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4146{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004147 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148}
4149
4150static void s_stop(struct seq_file *m, void *p)
4151{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004152 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153}
4154
4155static int s_show(struct seq_file *m, void *p)
4156{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004157 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004158 struct slab *slabp;
4159 unsigned long active_objs;
4160 unsigned long num_objs;
4161 unsigned long active_slabs = 0;
4162 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004163 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004165 int node;
4166 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
Linus Torvalds1da177e2005-04-16 15:20:36 -07004168 active_objs = 0;
4169 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004170 for_each_online_node(node) {
4171 l3 = cachep->nodelists[node];
4172 if (!l3)
4173 continue;
4174
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004175 check_irq_on();
4176 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004177
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004178 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004179 if (slabp->inuse != cachep->num && !error)
4180 error = "slabs_full accounting error";
4181 active_objs += cachep->num;
4182 active_slabs++;
4183 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004184 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004185 if (slabp->inuse == cachep->num && !error)
4186 error = "slabs_partial inuse accounting error";
4187 if (!slabp->inuse && !error)
4188 error = "slabs_partial/inuse accounting error";
4189 active_objs += slabp->inuse;
4190 active_slabs++;
4191 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004192 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004193 if (slabp->inuse && !error)
4194 error = "slabs_free/inuse accounting error";
4195 num_slabs++;
4196 }
4197 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004198 if (l3->shared)
4199 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004200
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004201 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004203 num_slabs += active_slabs;
4204 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004205 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 error = "free_objects accounting error";
4207
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004208 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209 if (error)
4210 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4211
4212 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004213 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004214 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004216 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004217 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004218 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004220 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 unsigned long high = cachep->high_mark;
4222 unsigned long allocs = cachep->num_allocations;
4223 unsigned long grown = cachep->grown;
4224 unsigned long reaped = cachep->reaped;
4225 unsigned long errors = cachep->errors;
4226 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004228 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004229 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230
Christoph Lametere498be72005-09-09 13:03:32 -07004231 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004232 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004233 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004234 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004235 }
4236 /* cpu stats */
4237 {
4238 unsigned long allochit = atomic_read(&cachep->allochit);
4239 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4240 unsigned long freehit = atomic_read(&cachep->freehit);
4241 unsigned long freemiss = atomic_read(&cachep->freemiss);
4242
4243 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 }
4246#endif
4247 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 return 0;
4249}
4250
4251/*
4252 * slabinfo_op - iterator that generates /proc/slabinfo
4253 *
4254 * Output layout:
4255 * cache-name
4256 * num-active-objs
4257 * total-objs
4258 * object size
4259 * num-active-slabs
4260 * total-slabs
4261 * num-pages-per-slab
4262 * + further values on SMP and with statistics enabled
4263 */
4264
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004265static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004266 .start = s_start,
4267 .next = s_next,
4268 .stop = s_stop,
4269 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270};
4271
4272#define MAX_SLABINFO_WRITE 128
4273/**
4274 * slabinfo_write - Tuning for the slab allocator
4275 * @file: unused
4276 * @buffer: user buffer
4277 * @count: data length
4278 * @ppos: unused
4279 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004280ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4281 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004283 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004284 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004285 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004286
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287 if (count > MAX_SLABINFO_WRITE)
4288 return -EINVAL;
4289 if (copy_from_user(&kbuf, buffer, count))
4290 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004291 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292
4293 tmp = strchr(kbuf, ' ');
4294 if (!tmp)
4295 return -EINVAL;
4296 *tmp = '\0';
4297 tmp++;
4298 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4299 return -EINVAL;
4300
4301 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004302 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004304 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004305 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004306 if (limit < 1 || batchcount < 1 ||
4307 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004308 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004309 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004310 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004311 batchcount, shared,
4312 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313 }
4314 break;
4315 }
4316 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004317 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004318 if (res >= 0)
4319 res = count;
4320 return res;
4321}
Al Viro871751e2006-03-25 03:06:39 -08004322
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004323static int slabinfo_open(struct inode *inode, struct file *file)
4324{
4325 return seq_open(file, &slabinfo_op);
4326}
4327
4328static const struct file_operations proc_slabinfo_operations = {
4329 .open = slabinfo_open,
4330 .read = seq_read,
4331 .write = slabinfo_write,
4332 .llseek = seq_lseek,
4333 .release = seq_release,
4334};
4335
Al Viro871751e2006-03-25 03:06:39 -08004336#ifdef CONFIG_DEBUG_SLAB_LEAK
4337
4338static void *leaks_start(struct seq_file *m, loff_t *pos)
4339{
Al Viro871751e2006-03-25 03:06:39 -08004340 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004341 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004342}
4343
4344static inline int add_caller(unsigned long *n, unsigned long v)
4345{
4346 unsigned long *p;
4347 int l;
4348 if (!v)
4349 return 1;
4350 l = n[1];
4351 p = n + 2;
4352 while (l) {
4353 int i = l/2;
4354 unsigned long *q = p + 2 * i;
4355 if (*q == v) {
4356 q[1]++;
4357 return 1;
4358 }
4359 if (*q > v) {
4360 l = i;
4361 } else {
4362 p = q + 2;
4363 l -= i + 1;
4364 }
4365 }
4366 if (++n[1] == n[0])
4367 return 0;
4368 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4369 p[0] = v;
4370 p[1] = 1;
4371 return 1;
4372}
4373
4374static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4375{
4376 void *p;
4377 int i;
4378 if (n[0] == n[1])
4379 return;
4380 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4381 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4382 continue;
4383 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4384 return;
4385 }
4386}
4387
4388static void show_symbol(struct seq_file *m, unsigned long address)
4389{
4390#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004391 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004392 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004393
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004394 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004395 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004396 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004397 seq_printf(m, " [%s]", modname);
4398 return;
4399 }
4400#endif
4401 seq_printf(m, "%p", (void *)address);
4402}
4403
4404static int leaks_show(struct seq_file *m, void *p)
4405{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004406 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004407 struct slab *slabp;
4408 struct kmem_list3 *l3;
4409 const char *name;
4410 unsigned long *n = m->private;
4411 int node;
4412 int i;
4413
4414 if (!(cachep->flags & SLAB_STORE_USER))
4415 return 0;
4416 if (!(cachep->flags & SLAB_RED_ZONE))
4417 return 0;
4418
4419 /* OK, we can do it */
4420
4421 n[1] = 0;
4422
4423 for_each_online_node(node) {
4424 l3 = cachep->nodelists[node];
4425 if (!l3)
4426 continue;
4427
4428 check_irq_on();
4429 spin_lock_irq(&l3->list_lock);
4430
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004431 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004432 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004433 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004434 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004435 spin_unlock_irq(&l3->list_lock);
4436 }
4437 name = cachep->name;
4438 if (n[0] == n[1]) {
4439 /* Increase the buffer size */
4440 mutex_unlock(&cache_chain_mutex);
4441 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4442 if (!m->private) {
4443 /* Too bad, we are really out */
4444 m->private = n;
4445 mutex_lock(&cache_chain_mutex);
4446 return -ENOMEM;
4447 }
4448 *(unsigned long *)m->private = n[0] * 2;
4449 kfree(n);
4450 mutex_lock(&cache_chain_mutex);
4451 /* Now make sure this entry will be retried */
4452 m->count = m->size;
4453 return 0;
4454 }
4455 for (i = 0; i < n[1]; i++) {
4456 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4457 show_symbol(m, n[2*i+2]);
4458 seq_putc(m, '\n');
4459 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004460
Al Viro871751e2006-03-25 03:06:39 -08004461 return 0;
4462}
4463
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004464static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004465 .start = leaks_start,
4466 .next = s_next,
4467 .stop = s_stop,
4468 .show = leaks_show,
4469};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004470
4471static int slabstats_open(struct inode *inode, struct file *file)
4472{
4473 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4474 int ret = -ENOMEM;
4475 if (n) {
4476 ret = seq_open(file, &slabstats_op);
4477 if (!ret) {
4478 struct seq_file *m = file->private_data;
4479 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4480 m->private = n;
4481 n = NULL;
4482 }
4483 kfree(n);
4484 }
4485 return ret;
4486}
4487
4488static const struct file_operations proc_slabstats_operations = {
4489 .open = slabstats_open,
4490 .read = seq_read,
4491 .llseek = seq_lseek,
4492 .release = seq_release_private,
4493};
Al Viro871751e2006-03-25 03:06:39 -08004494#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004495
4496static int __init slab_proc_init(void)
4497{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004498 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004499#ifdef CONFIG_DEBUG_SLAB_LEAK
4500 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4501#endif
4502 return 0;
4503}
4504module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004505#endif
4506
Manfred Spraul00e145b2005-09-03 15:55:07 -07004507/**
4508 * ksize - get the actual amount of memory allocated for a given object
4509 * @objp: Pointer to the object
4510 *
4511 * kmalloc may internally round up allocations and return more memory
4512 * than requested. ksize() can be used to determine the actual amount of
4513 * memory allocated. The caller may use this additional memory, even though
4514 * a smaller amount of memory was initially specified with the kmalloc call.
4515 * The caller must guarantee that objp points to a valid object previously
4516 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4517 * must not be freed during the duration of the call.
4518 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004519size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004520{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004521 BUG_ON(!objp);
4522 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004523 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004524
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004525 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004526}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004527EXPORT_SYMBOL(ksize);