blob: 83311c9aaf9de0ad8494f04703c248f155cfc2c1 [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 Dobriyana0ec95a2008-10-06 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>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120#include <asm/cacheflush.h>
121#include <asm/tlbflush.h>
122#include <asm/page.h>
123
124/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
130 *
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 */
133
134#ifdef CONFIG_DEBUG_SLAB
135#define DEBUG 1
136#define STATS 1
137#define FORCED_DEBUG 1
138#else
139#define DEBUG 0
140#define STATS 0
141#define FORCED_DEBUG 0
142#endif
143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144/* Shouldn't this be in a header file somewhere? */
145#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400146#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148#ifndef ARCH_KMALLOC_FLAGS
149#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150#endif
151
152/* Legal flag mask for kmem_cache_create(). */
153#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700154# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800156 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700157 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700159 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200160 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800162# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700163 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700165 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200166 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167#endif
168
169/*
170 * kmem_bufctl_t:
171 *
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
174 *
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
186 */
187
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700188typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800191#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 * struct slab_rcu
196 *
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
204 *
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207 */
208struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800209 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800210 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800211 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212};
213
214/*
Lai Jiangshan5bfe53a2011-03-10 15:22:24 +0800215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
222 union {
223 struct {
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
230 };
231 struct slab_rcu __slab_cover_slab_rcu;
232 };
233};
234
235/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 * struct array_cache
237 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 * Purpose:
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
242 *
243 * The limit is stored in the per-cpu structure to reduce the data cache
244 * footprint.
245 *
246 */
247struct array_cache {
248 unsigned int avail;
249 unsigned int limit;
250 unsigned int batchcount;
251 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700252 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700253 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800254 * Must have this definition in here for the proper
255 * alignment of array_cache. Also simplifies accessing
256 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800257 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258};
259
Andrew Mortona737b3e2006-03-22 00:08:11 -0800260/*
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 */
264#define BOOT_CPUCACHE_ENTRIES 1
265struct arraycache_init {
266 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800267 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268};
269
270/*
Christoph Lametere498be72005-09-09 13:03:32 -0700271 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 */
273struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800278 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800279 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800280 spinlock_t list_lock;
281 struct array_cache *shared; /* shared per node */
282 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800283 unsigned long next_reap; /* updated without locking */
284 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
Christoph Lametere498be72005-09-09 13:03:32 -0700287/*
288 * Need this for bootstrapping a per node allocator.
289 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200290#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
H Hartley Sweeten68a1b192011-01-11 17:49:32 -0600291static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700292#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200293#define SIZE_AC MAX_NUMNODES
294#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295
Christoph Lametered11d9e2006-06-30 01:55:45 -0700296static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_list3 *l3, int tofree);
298static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300300static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000301static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700302
Christoph Lametere498be72005-09-09 13:03:32 -0700303/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700306 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700307static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700308{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800309 extern void __bad_size(void);
310
Christoph Lametere498be72005-09-09 13:03:32 -0700311 if (__builtin_constant_p(size)) {
312 int i = 0;
313
314#define CACHE(x) \
315 if (size <=x) \
316 return i; \
317 else \
318 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800319#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700320#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800321 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700322 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800323 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700324 return 0;
325}
326
Ingo Molnare0a42722006-06-23 02:03:46 -0700327static int slab_early_init = 1;
328
Christoph Lametere498be72005-09-09 13:03:32 -0700329#define INDEX_AC index_of(sizeof(struct arraycache_init))
330#define INDEX_L3 index_of(sizeof(struct kmem_list3))
331
Pekka Enberg5295a742006-02-01 03:05:48 -0800332static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700333{
334 INIT_LIST_HEAD(&parent->slabs_full);
335 INIT_LIST_HEAD(&parent->slabs_partial);
336 INIT_LIST_HEAD(&parent->slabs_free);
337 parent->shared = NULL;
338 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800339 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700340 spin_lock_init(&parent->list_lock);
341 parent->free_objects = 0;
342 parent->free_touched = 0;
343}
344
Andrew Mortona737b3e2006-03-22 00:08:11 -0800345#define MAKE_LIST(cachep, listp, slab, nodeid) \
346 do { \
347 INIT_LIST_HEAD(listp); \
348 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700349 } while (0)
350
Andrew Mortona737b3e2006-03-22 00:08:11 -0800351#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700353 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
356 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358#define CFLGS_OFF_SLAB (0x80000000UL)
359#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
360
361#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800362/*
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100366 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 * which could lock up otherwise freeable slabs.
368 */
369#define REAPTIMEOUT_CPUC (2*HZ)
370#define REAPTIMEOUT_LIST3 (4*HZ)
371
372#if STATS
373#define STATS_INC_ACTIVE(x) ((x)->num_active++)
374#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700377#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800378#define STATS_SET_HIGH(x) \
379 do { \
380 if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
382 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383#define STATS_INC_ERR(x) ((x)->errors++)
384#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700385#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700386#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800387#define STATS_SET_FREEABLE(x, i) \
388 do { \
389 if ((x)->max_freeable < i) \
390 (x)->max_freeable = i; \
391 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396#else
397#define STATS_INC_ACTIVE(x) do { } while (0)
398#define STATS_DEC_ACTIVE(x) do { } while (0)
399#define STATS_INC_ALLOCED(x) do { } while (0)
400#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700401#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402#define STATS_SET_HIGH(x) do { } while (0)
403#define STATS_INC_ERR(x) do { } while (0)
404#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700405#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700406#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408#define STATS_INC_ALLOCHIT(x) do { } while (0)
409#define STATS_INC_ALLOCMISS(x) do { } while (0)
410#define STATS_INC_FREEHIT(x) do { } while (0)
411#define STATS_INC_FREEMISS(x) do { } while (0)
412#endif
413
414#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Andrew Mortona737b3e2006-03-22 00:08:11 -0800416/*
417 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800429static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800431 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432}
433
Pekka Enberg343e0d72006-02-01 03:05:50 -0800434static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800436 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
David Woodhouseb46b8f12007-05-08 00:22:59 -0700439static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440{
441 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700442 return (unsigned long long*) (objp + obj_offset(cachep) -
443 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444}
445
David Woodhouseb46b8f12007-05-08 00:22:59 -0700446static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447{
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700450 return (unsigned long long *)(objp + cachep->buffer_size -
451 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400452 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700453 return (unsigned long long *) (objp + cachep->buffer_size -
454 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455}
456
Pekka Enberg343e0d72006-02-01 03:05:50 -0800457static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458{
459 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800460 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461}
462
463#else
464
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800465#define obj_offset(x) 0
466#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700467#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470
471#endif
472
Li Zefan0f24f122009-12-11 15:45:30 +0800473#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300474size_t slab_buffer_size(struct kmem_cache *cachep)
475{
476 return cachep->buffer_size;
477}
478EXPORT_SYMBOL(slab_buffer_size);
479#endif
480
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 * Do not go above this order unless 0 objects fit into the slab.
483 */
484#define BREAK_GFP_ORDER_HI 1
485#define BREAK_GFP_ORDER_LO 0
486static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
487
Andrew Mortona737b3e2006-03-22 00:08:11 -0800488/*
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800493static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
494{
495 page->lru.next = (struct list_head *)cache;
496}
497
498static inline struct kmem_cache *page_get_cache(struct page *page)
499{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700500 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700501 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800502 return (struct kmem_cache *)page->lru.next;
503}
504
505static inline void page_set_slab(struct page *page, struct slab *slab)
506{
507 page->lru.prev = (struct list_head *)slab;
508}
509
510static inline struct slab *page_get_slab(struct page *page)
511{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700512 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800513 return (struct slab *)page->lru.prev;
514}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800516static inline struct kmem_cache *virt_to_cache(const void *obj)
517{
Christoph Lameterb49af682007-05-06 14:49:41 -0700518 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800519 return page_get_cache(page);
520}
521
522static inline struct slab *virt_to_slab(const void *obj)
523{
Christoph Lameterb49af682007-05-06 14:49:41 -0700524 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800525 return page_get_slab(page);
526}
527
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800528static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529 unsigned int idx)
530{
531 return slab->s_mem + cache->buffer_size * idx;
532}
533
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800534/*
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 */
540static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800542{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800545}
546
Andrew Mortona737b3e2006-03-22 00:08:11 -0800547/*
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550struct cache_sizes malloc_sizes[] = {
551#define CACHE(x) { .cs_size = (x) },
552#include <linux/kmalloc_sizes.h>
553 CACHE(ULONG_MAX)
554#undef CACHE
555};
556EXPORT_SYMBOL(malloc_sizes);
557
558/* Must match cache_sizes above. Out of line to keep cache footprint low. */
559struct cache_names {
560 char *name;
561 char *name_dma;
562};
563
564static struct cache_names __initdata cache_names[] = {
565#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800567 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568#undef CACHE
569};
570
571static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
576/* internal cache of cache description objs */
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200577static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
Pekka Enberg343e0d72006-02-01 03:05:50 -0800578static struct kmem_cache cache_cache = {
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200579 .nodelists = cache_cache_nodelists,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800580 .batchcount = 1,
581 .limit = BOOT_CPUCACHE_ENTRIES,
582 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800583 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800584 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585};
586
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700587#define BAD_ALIEN_MAGIC 0x01020304ul
588
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 * chicken and egg problem: delay the per-cpu array allocation
591 * until the general caches are up.
592 */
593static enum {
594 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700595 PARTIAL_AC,
596 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300597 EARLY,
Peter Zijlstra52cef182011-11-28 21:12:40 +0100598 LATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 FULL
600} g_cpucache_up;
601
Mike Kravetz39d24e62006-05-15 09:44:13 -0700602/*
603 * used by boot code to determine if it can use slab based allocator
604 */
605int slab_is_available(void)
606{
Pekka Enberg8429db52009-06-12 15:58:59 +0300607 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700608}
609
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200610#ifdef CONFIG_LOCKDEP
611
612/*
613 * Slab sometimes uses the kmalloc slabs to store the slab headers
614 * for other slabs "off slab".
615 * The locking for this is tricky in that it nests within the locks
616 * of all other slabs in a few places; to deal with this special
617 * locking we put on-slab caches into a separate lock-class.
618 *
619 * We set lock class for alien array caches which are up during init.
620 * The lock annotation will be lost if all cpus of a node goes down and
621 * then comes back up during hotplug
622 */
623static struct lock_class_key on_slab_l3_key;
624static struct lock_class_key on_slab_alc_key;
625
Peter Zijlstra83835b32011-07-22 15:26:05 +0200626static struct lock_class_key debugobj_l3_key;
627static struct lock_class_key debugobj_alc_key;
628
629static void slab_set_lock_classes(struct kmem_cache *cachep,
630 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
631 int q)
632{
633 struct array_cache **alc;
634 struct kmem_list3 *l3;
635 int r;
636
637 l3 = cachep->nodelists[q];
638 if (!l3)
639 return;
640
641 lockdep_set_class(&l3->list_lock, l3_key);
642 alc = l3->alien;
643 /*
644 * FIXME: This check for BAD_ALIEN_MAGIC
645 * should go away when common slab code is taught to
646 * work even without alien caches.
647 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
648 * for alloc_alien_cache,
649 */
650 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
651 return;
652 for_each_node(r) {
653 if (alc[r])
654 lockdep_set_class(&alc[r]->lock, alc_key);
655 }
656}
657
658static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
659{
660 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
661}
662
663static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
664{
665 int node;
666
667 for_each_online_node(node)
668 slab_set_debugobj_lock_classes_node(cachep, node);
669}
670
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200671static void init_node_lock_keys(int q)
672{
673 struct cache_sizes *s = malloc_sizes;
674
Peter Zijlstra52cef182011-11-28 21:12:40 +0100675 if (g_cpucache_up < LATE)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200676 return;
677
678 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200679 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200680
681 l3 = s->cs_cachep->nodelists[q];
682 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200683 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200684
685 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
686 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200687 }
688}
689
690static inline void init_lock_keys(void)
691{
692 int node;
693
694 for_each_node(node)
695 init_node_lock_keys(node);
696}
697#else
698static void init_node_lock_keys(int q)
699{
700}
701
702static inline void init_lock_keys(void)
703{
704}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200705
706static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
707{
708}
709
710static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
711{
712}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200713#endif
714
715/*
716 * Guard access to the cache-chain.
717 */
718static DEFINE_MUTEX(cache_chain_mutex);
719static struct list_head cache_chain;
720
Tejun Heo1871e522009-10-29 22:34:13 +0900721static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
Pekka Enberg343e0d72006-02-01 03:05:50 -0800723static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724{
725 return cachep->array[smp_processor_id()];
726}
727
Andrew Mortona737b3e2006-03-22 00:08:11 -0800728static inline struct kmem_cache *__find_general_cachep(size_t size,
729 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730{
731 struct cache_sizes *csizep = malloc_sizes;
732
733#if DEBUG
734 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800735 * kmem_cache_create(), or __kmalloc(), before
736 * the generic caches are initialized.
737 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700738 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700740 if (!size)
741 return ZERO_SIZE_PTR;
742
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 while (size > csizep->cs_size)
744 csizep++;
745
746 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700747 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 * has cs_{dma,}cachep==NULL. Thus no special case
749 * for large kmalloc calls required.
750 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800751#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 if (unlikely(gfpflags & GFP_DMA))
753 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800754#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 return csizep->cs_cachep;
756}
757
Adrian Bunkb2213852006-09-25 23:31:02 -0700758static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700759{
760 return __find_general_cachep(size, gfpflags);
761}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700762
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800763static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800765 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
766}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767
Andrew Mortona737b3e2006-03-22 00:08:11 -0800768/*
769 * Calculate the number of objects and left-over bytes for a given buffer size.
770 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800771static void cache_estimate(unsigned long gfporder, size_t buffer_size,
772 size_t align, int flags, size_t *left_over,
773 unsigned int *num)
774{
775 int nr_objs;
776 size_t mgmt_size;
777 size_t slab_size = PAGE_SIZE << gfporder;
778
779 /*
780 * The slab management structure can be either off the slab or
781 * on it. For the latter case, the memory allocated for a
782 * slab is used for:
783 *
784 * - The struct slab
785 * - One kmem_bufctl_t for each object
786 * - Padding to respect alignment of @align
787 * - @buffer_size bytes for each object
788 *
789 * If the slab management structure is off the slab, then the
790 * alignment will already be calculated into the size. Because
791 * the slabs are all pages aligned, the objects will be at the
792 * correct alignment when allocated.
793 */
794 if (flags & CFLGS_OFF_SLAB) {
795 mgmt_size = 0;
796 nr_objs = slab_size / buffer_size;
797
798 if (nr_objs > SLAB_LIMIT)
799 nr_objs = SLAB_LIMIT;
800 } else {
801 /*
802 * Ignore padding for the initial guess. The padding
803 * is at most @align-1 bytes, and @buffer_size is at
804 * least @align. In the worst case, this result will
805 * be one greater than the number of objects that fit
806 * into the memory allocation when taking the padding
807 * into account.
808 */
809 nr_objs = (slab_size - sizeof(struct slab)) /
810 (buffer_size + sizeof(kmem_bufctl_t));
811
812 /*
813 * This calculated number will be either the right
814 * amount, or one greater than what we want.
815 */
816 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
817 > slab_size)
818 nr_objs--;
819
820 if (nr_objs > SLAB_LIMIT)
821 nr_objs = SLAB_LIMIT;
822
823 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800825 *num = nr_objs;
826 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827}
828
Harvey Harrisond40cee22008-04-30 00:55:07 -0700829#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
Andrew Mortona737b3e2006-03-22 00:08:11 -0800831static void __slab_error(const char *function, struct kmem_cache *cachep,
832 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833{
834 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800835 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 dump_stack();
837}
838
Paul Menage3395ee02006-12-06 20:32:16 -0800839/*
840 * By default on NUMA we use alien caches to stage the freeing of
841 * objects allocated from other nodes. This causes massive memory
842 * inefficiencies when using fake NUMA setup to split memory into a
843 * large number of small nodes, so it can be disabled on the command
844 * line
845 */
846
847static int use_alien_caches __read_mostly = 1;
848static int __init noaliencache_setup(char *s)
849{
850 use_alien_caches = 0;
851 return 1;
852}
853__setup("noaliencache", noaliencache_setup);
854
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800855#ifdef CONFIG_NUMA
856/*
857 * Special reaping functions for NUMA systems called from cache_reap().
858 * These take care of doing round robin flushing of alien caches (containing
859 * objects freed on different nodes from which they were allocated) and the
860 * flushing of remote pcps by calling drain_node_pages.
861 */
Tejun Heo1871e522009-10-29 22:34:13 +0900862static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800863
864static void init_reap_node(int cpu)
865{
866 int node;
867
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700868 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800869 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800870 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800871
Tejun Heo1871e522009-10-29 22:34:13 +0900872 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800873}
874
875static void next_reap_node(void)
876{
Christoph Lameter909ea962010-12-08 16:22:55 +0100877 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800879 node = next_node(node, node_online_map);
880 if (unlikely(node >= MAX_NUMNODES))
881 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100882 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800883}
884
885#else
886#define init_reap_node(cpu) do { } while (0)
887#define next_reap_node(void) do { } while (0)
888#endif
889
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890/*
891 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
892 * via the workqueue/eventd.
893 * Add the CPU number into the expiration time to minimize the possibility of
894 * the CPUs getting into lockstep and contending for the global cache chain
895 * lock.
896 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700897static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898{
Tejun Heo1871e522009-10-29 22:34:13 +0900899 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
901 /*
902 * When this gets called from do_initcalls via cpucache_init(),
903 * init_workqueues() has already run, so keventd will be setup
904 * at that time.
905 */
David Howells52bad642006-11-22 14:54:01 +0000906 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800907 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700908 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800909 schedule_delayed_work_on(cpu, reap_work,
910 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 }
912}
913
Christoph Lametere498be72005-09-09 13:03:32 -0700914static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300915 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800917 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 struct array_cache *nc = NULL;
919
Pekka Enberg83b519e2009-06-10 19:40:04 +0300920 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100921 /*
922 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300923 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100924 * cache the pointers are not cleared and they could be counted as
925 * valid references during a kmemleak scan. Therefore, kmemleak must
926 * not scan such objects.
927 */
928 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 if (nc) {
930 nc->avail = 0;
931 nc->limit = entries;
932 nc->batchcount = batchcount;
933 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700934 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 }
936 return nc;
937}
938
Christoph Lameter3ded1752006-03-25 03:06:44 -0800939/*
940 * Transfer objects in one arraycache to another.
941 * Locking must be handled by the caller.
942 *
943 * Return the number of entries transferred.
944 */
945static int transfer_objects(struct array_cache *to,
946 struct array_cache *from, unsigned int max)
947{
948 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700949 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800950
951 if (!nr)
952 return 0;
953
954 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
955 sizeof(void *) *nr);
956
957 from->avail -= nr;
958 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800959 return nr;
960}
961
Christoph Lameter765c4502006-09-27 01:50:08 -0700962#ifndef CONFIG_NUMA
963
964#define drain_alien_cache(cachep, alien) do { } while (0)
965#define reap_alien(cachep, l3) do { } while (0)
966
Pekka Enberg83b519e2009-06-10 19:40:04 +0300967static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700968{
969 return (struct array_cache **)BAD_ALIEN_MAGIC;
970}
971
972static inline void free_alien_cache(struct array_cache **ac_ptr)
973{
974}
975
976static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
977{
978 return 0;
979}
980
981static inline void *alternate_node_alloc(struct kmem_cache *cachep,
982 gfp_t flags)
983{
984 return NULL;
985}
986
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800987static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700988 gfp_t flags, int nodeid)
989{
990 return NULL;
991}
992
993#else /* CONFIG_NUMA */
994
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800995static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800996static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800997
Pekka Enberg83b519e2009-06-10 19:40:04 +0300998static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700999{
1000 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001001 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001002 int i;
1003
1004 if (limit > 1)
1005 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001006 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001007 if (ac_ptr) {
1008 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001009 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001010 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001011 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001012 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001013 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001014 kfree(ac_ptr[i]);
1015 kfree(ac_ptr);
1016 return NULL;
1017 }
1018 }
1019 }
1020 return ac_ptr;
1021}
1022
Pekka Enberg5295a742006-02-01 03:05:48 -08001023static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001024{
1025 int i;
1026
1027 if (!ac_ptr)
1028 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001029 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001030 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001031 kfree(ac_ptr);
1032}
1033
Pekka Enberg343e0d72006-02-01 03:05:50 -08001034static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001035 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001036{
1037 struct kmem_list3 *rl3 = cachep->nodelists[node];
1038
1039 if (ac->avail) {
1040 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001041 /*
1042 * Stuff objects into the remote nodes shared array first.
1043 * That way we could avoid the overhead of putting the objects
1044 * into the free lists and getting them back later.
1045 */
shin, jacob693f7d32006-04-28 10:54:37 -05001046 if (rl3->shared)
1047 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001048
Christoph Lameterff694162005-09-22 21:44:02 -07001049 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001050 ac->avail = 0;
1051 spin_unlock(&rl3->list_lock);
1052 }
1053}
1054
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001055/*
1056 * Called from cache_reap() to regularly drain alien caches round robin.
1057 */
1058static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1059{
Christoph Lameter909ea962010-12-08 16:22:55 +01001060 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001061
1062 if (l3->alien) {
1063 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001064
1065 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001066 __drain_alien_cache(cachep, ac, node);
1067 spin_unlock_irq(&ac->lock);
1068 }
1069 }
1070}
1071
Andrew Mortona737b3e2006-03-22 00:08:11 -08001072static void drain_alien_cache(struct kmem_cache *cachep,
1073 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001074{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001075 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001076 struct array_cache *ac;
1077 unsigned long flags;
1078
1079 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001080 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001081 if (ac) {
1082 spin_lock_irqsave(&ac->lock, flags);
1083 __drain_alien_cache(cachep, ac, i);
1084 spin_unlock_irqrestore(&ac->lock, flags);
1085 }
1086 }
1087}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001088
Ingo Molnar873623d2006-07-13 14:44:38 +02001089static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001090{
1091 struct slab *slabp = virt_to_slab(objp);
1092 int nodeid = slabp->nodeid;
1093 struct kmem_list3 *l3;
1094 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001095 int node;
1096
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001097 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001098
1099 /*
1100 * Make sure we are not freeing a object from another node to the array
1101 * cache on this cpu.
1102 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001103 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001104 return 0;
1105
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001106 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001107 STATS_INC_NODEFREES(cachep);
1108 if (l3->alien && l3->alien[nodeid]) {
1109 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001110 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001111 if (unlikely(alien->avail == alien->limit)) {
1112 STATS_INC_ACOVERFLOW(cachep);
1113 __drain_alien_cache(cachep, alien, nodeid);
1114 }
1115 alien->entry[alien->avail++] = objp;
1116 spin_unlock(&alien->lock);
1117 } else {
1118 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1119 free_block(cachep, &objp, 1, nodeid);
1120 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1121 }
1122 return 1;
1123}
Christoph Lametere498be72005-09-09 13:03:32 -07001124#endif
1125
David Rientjes8f9f8d92010-03-27 19:40:47 -07001126/*
1127 * Allocates and initializes nodelists for a node on each slab cache, used for
1128 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1129 * will be allocated off-node since memory is not yet online for the new node.
1130 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1131 * already in use.
1132 *
1133 * Must hold cache_chain_mutex.
1134 */
1135static int init_cache_nodelists_node(int node)
1136{
1137 struct kmem_cache *cachep;
1138 struct kmem_list3 *l3;
1139 const int memsize = sizeof(struct kmem_list3);
1140
1141 list_for_each_entry(cachep, &cache_chain, next) {
1142 /*
1143 * Set up the size64 kmemlist for cpu before we can
1144 * begin anything. Make sure some other cpu on this
1145 * node has not already allocated this
1146 */
1147 if (!cachep->nodelists[node]) {
1148 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1149 if (!l3)
1150 return -ENOMEM;
1151 kmem_list3_init(l3);
1152 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1153 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1154
1155 /*
1156 * The l3s don't come and go as CPUs come and
1157 * go. cache_chain_mutex is sufficient
1158 * protection here.
1159 */
1160 cachep->nodelists[node] = l3;
1161 }
1162
1163 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1164 cachep->nodelists[node]->free_limit =
1165 (1 + nr_cpus_node(node)) *
1166 cachep->batchcount + cachep->num;
1167 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1168 }
1169 return 0;
1170}
1171
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001172static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001174 struct kmem_cache *cachep;
1175 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001176 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301177 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001178
1179 list_for_each_entry(cachep, &cache_chain, next) {
1180 struct array_cache *nc;
1181 struct array_cache *shared;
1182 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001183
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001184 /* cpu is dead; no one can alloc from it. */
1185 nc = cachep->array[cpu];
1186 cachep->array[cpu] = NULL;
1187 l3 = cachep->nodelists[node];
1188
1189 if (!l3)
1190 goto free_array_cache;
1191
1192 spin_lock_irq(&l3->list_lock);
1193
1194 /* Free limit for this kmem_list3 */
1195 l3->free_limit -= cachep->batchcount;
1196 if (nc)
1197 free_block(cachep, nc->entry, nc->avail, node);
1198
Rusty Russell58463c12009-12-17 11:43:12 -06001199 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001200 spin_unlock_irq(&l3->list_lock);
1201 goto free_array_cache;
1202 }
1203
1204 shared = l3->shared;
1205 if (shared) {
1206 free_block(cachep, shared->entry,
1207 shared->avail, node);
1208 l3->shared = NULL;
1209 }
1210
1211 alien = l3->alien;
1212 l3->alien = NULL;
1213
1214 spin_unlock_irq(&l3->list_lock);
1215
1216 kfree(shared);
1217 if (alien) {
1218 drain_alien_cache(cachep, alien);
1219 free_alien_cache(alien);
1220 }
1221free_array_cache:
1222 kfree(nc);
1223 }
1224 /*
1225 * In the previous loop, all the objects were freed to
1226 * the respective cache's slabs, now we can go ahead and
1227 * shrink each nodelist to its limit.
1228 */
1229 list_for_each_entry(cachep, &cache_chain, next) {
1230 l3 = cachep->nodelists[node];
1231 if (!l3)
1232 continue;
1233 drain_freelist(cachep, l3, l3->free_objects);
1234 }
1235}
1236
1237static int __cpuinit cpuup_prepare(long cpu)
1238{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001239 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001240 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001241 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001242 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001244 /*
1245 * We need to do this right in the beginning since
1246 * alloc_arraycache's are going to use this list.
1247 * kmalloc_node allows us to add the slab to the right
1248 * kmem_list3 and not this cpu's kmem_list3
1249 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001250 err = init_cache_nodelists_node(node);
1251 if (err < 0)
1252 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001253
1254 /*
1255 * Now we can go ahead with allocating the shared arrays and
1256 * array caches
1257 */
1258 list_for_each_entry(cachep, &cache_chain, next) {
1259 struct array_cache *nc;
1260 struct array_cache *shared = NULL;
1261 struct array_cache **alien = NULL;
1262
1263 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001264 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001265 if (!nc)
1266 goto bad;
1267 if (cachep->shared) {
1268 shared = alloc_arraycache(node,
1269 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001270 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001271 if (!shared) {
1272 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001273 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001274 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001275 }
1276 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001277 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001278 if (!alien) {
1279 kfree(shared);
1280 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001281 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001282 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001283 }
1284 cachep->array[cpu] = nc;
1285 l3 = cachep->nodelists[node];
1286 BUG_ON(!l3);
1287
1288 spin_lock_irq(&l3->list_lock);
1289 if (!l3->shared) {
1290 /*
1291 * We are serialised from CPU_DEAD or
1292 * CPU_UP_CANCELLED by the cpucontrol lock
1293 */
1294 l3->shared = shared;
1295 shared = NULL;
1296 }
1297#ifdef CONFIG_NUMA
1298 if (!l3->alien) {
1299 l3->alien = alien;
1300 alien = NULL;
1301 }
1302#endif
1303 spin_unlock_irq(&l3->list_lock);
1304 kfree(shared);
1305 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001306 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1307 slab_set_debugobj_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001308 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001309 init_node_lock_keys(node);
1310
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001311 return 0;
1312bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001313 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001314 return -ENOMEM;
1315}
1316
1317static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1318 unsigned long action, void *hcpu)
1319{
1320 long cpu = (long)hcpu;
1321 int err = 0;
1322
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001324 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001325 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001326 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001327 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001328 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 break;
1330 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001331 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 start_cpu_timer(cpu);
1333 break;
1334#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001335 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001336 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001337 /*
1338 * Shutdown cache reaper. Note that the cache_chain_mutex is
1339 * held so that if cache_reap() is invoked it cannot do
1340 * anything expensive but will only modify reap_work
1341 * and reschedule the timer.
1342 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001343 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001344 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001345 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001346 break;
1347 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001348 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001349 start_cpu_timer(cpu);
1350 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001352 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001353 /*
1354 * Even if all the cpus of a node are down, we don't free the
1355 * kmem_list3 of any cache. This to avoid a race between
1356 * cpu_down, and a kmalloc allocation from another cpu for
1357 * memory from the node of the cpu going down. The list3
1358 * structure is usually allocated from kmem_cache_create() and
1359 * gets destroyed at kmem_cache_destroy().
1360 */
Simon Arlott183ff222007-10-20 01:27:18 +02001361 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001362#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001364 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001365 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001366 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001367 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001370 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371}
1372
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001373static struct notifier_block __cpuinitdata cpucache_notifier = {
1374 &cpuup_callback, NULL, 0
1375};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
David Rientjes8f9f8d92010-03-27 19:40:47 -07001377#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1378/*
1379 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1380 * Returns -EBUSY if all objects cannot be drained so that the node is not
1381 * removed.
1382 *
1383 * Must hold cache_chain_mutex.
1384 */
1385static int __meminit drain_cache_nodelists_node(int node)
1386{
1387 struct kmem_cache *cachep;
1388 int ret = 0;
1389
1390 list_for_each_entry(cachep, &cache_chain, next) {
1391 struct kmem_list3 *l3;
1392
1393 l3 = cachep->nodelists[node];
1394 if (!l3)
1395 continue;
1396
1397 drain_freelist(cachep, l3, l3->free_objects);
1398
1399 if (!list_empty(&l3->slabs_full) ||
1400 !list_empty(&l3->slabs_partial)) {
1401 ret = -EBUSY;
1402 break;
1403 }
1404 }
1405 return ret;
1406}
1407
1408static int __meminit slab_memory_callback(struct notifier_block *self,
1409 unsigned long action, void *arg)
1410{
1411 struct memory_notify *mnb = arg;
1412 int ret = 0;
1413 int nid;
1414
1415 nid = mnb->status_change_nid;
1416 if (nid < 0)
1417 goto out;
1418
1419 switch (action) {
1420 case MEM_GOING_ONLINE:
1421 mutex_lock(&cache_chain_mutex);
1422 ret = init_cache_nodelists_node(nid);
1423 mutex_unlock(&cache_chain_mutex);
1424 break;
1425 case MEM_GOING_OFFLINE:
1426 mutex_lock(&cache_chain_mutex);
1427 ret = drain_cache_nodelists_node(nid);
1428 mutex_unlock(&cache_chain_mutex);
1429 break;
1430 case MEM_ONLINE:
1431 case MEM_OFFLINE:
1432 case MEM_CANCEL_ONLINE:
1433 case MEM_CANCEL_OFFLINE:
1434 break;
1435 }
1436out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001437 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001438}
1439#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1440
Christoph Lametere498be72005-09-09 13:03:32 -07001441/*
1442 * swap the static kmem_list3 with kmalloced memory
1443 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001444static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1445 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001446{
1447 struct kmem_list3 *ptr;
1448
Pekka Enberg83b519e2009-06-10 19:40:04 +03001449 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001450 BUG_ON(!ptr);
1451
Christoph Lametere498be72005-09-09 13:03:32 -07001452 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001453 /*
1454 * Do not assume that spinlocks can be initialized via memcpy:
1455 */
1456 spin_lock_init(&ptr->list_lock);
1457
Christoph Lametere498be72005-09-09 13:03:32 -07001458 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1459 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001460}
1461
Andrew Mortona737b3e2006-03-22 00:08:11 -08001462/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001463 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1464 * size of kmem_list3.
1465 */
1466static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1467{
1468 int node;
1469
1470 for_each_online_node(node) {
1471 cachep->nodelists[node] = &initkmem_list3[index + node];
1472 cachep->nodelists[node]->next_reap = jiffies +
1473 REAPTIMEOUT_LIST3 +
1474 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1475 }
1476}
1477
1478/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001479 * Initialisation. Called after the page allocator have been initialised and
1480 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 */
1482void __init kmem_cache_init(void)
1483{
1484 size_t left_over;
1485 struct cache_sizes *sizes;
1486 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001487 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001488 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001489 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001490
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001491 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001492 use_alien_caches = 0;
1493
Christoph Lametere498be72005-09-09 13:03:32 -07001494 for (i = 0; i < NUM_INIT_LISTS; i++) {
1495 kmem_list3_init(&initkmem_list3[i]);
1496 if (i < MAX_NUMNODES)
1497 cache_cache.nodelists[i] = NULL;
1498 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001499 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
1501 /*
1502 * Fragmentation resistance on low memory - only use bigger
1503 * page orders on machines with more than 32MB of memory.
1504 */
Jan Beulich44813742009-09-21 17:03:05 -07001505 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1507
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 /* Bootstrap is tricky, because several objects are allocated
1509 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001510 * 1) initialize the cache_cache cache: it contains the struct
1511 * kmem_cache structures of all caches, except cache_cache itself:
1512 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001513 * Initially an __init data area is used for the head array and the
1514 * kmem_list3 structures, it's replaced with a kmalloc allocated
1515 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001517 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001518 * An __init data area is used for the head array.
1519 * 3) Create the remaining kmalloc caches, with minimally sized
1520 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 * 4) Replace the __init data head arrays for cache_cache and the first
1522 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001523 * 5) Replace the __init data for kmem_list3 for cache_cache and
1524 * the other cache's with kmalloc allocated memory.
1525 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 */
1527
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001528 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001529
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 INIT_LIST_HEAD(&cache_chain);
1532 list_add(&cache_cache.next, &cache_chain);
1533 cache_cache.colour_off = cache_line_size();
1534 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001535 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536
Eric Dumazet8da34302007-05-06 14:49:29 -07001537 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001538 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001539 */
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001540 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1541 nr_node_ids * sizeof(struct kmem_list3 *);
Eric Dumazet8da34302007-05-06 14:49:29 -07001542#if DEBUG
1543 cache_cache.obj_size = cache_cache.buffer_size;
1544#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001545 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1546 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001547 cache_cache.reciprocal_buffer_size =
1548 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549
Jack Steiner07ed76b2006-03-07 21:55:46 -08001550 for (order = 0; order < MAX_ORDER; order++) {
1551 cache_estimate(order, cache_cache.buffer_size,
1552 cache_line_size(), 0, &left_over, &cache_cache.num);
1553 if (cache_cache.num)
1554 break;
1555 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001556 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001557 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001558 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001559 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1560 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561
1562 /* 2+3) create the kmalloc caches */
1563 sizes = malloc_sizes;
1564 names = cache_names;
1565
Andrew Mortona737b3e2006-03-22 00:08:11 -08001566 /*
1567 * Initialize the caches that provide memory for the array cache and the
1568 * kmem_list3 structures first. Without this, further allocations will
1569 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001570 */
1571
1572 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001573 sizes[INDEX_AC].cs_size,
1574 ARCH_KMALLOC_MINALIGN,
1575 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001576 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001577
Andrew Mortona737b3e2006-03-22 00:08:11 -08001578 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001579 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001580 kmem_cache_create(names[INDEX_L3].name,
1581 sizes[INDEX_L3].cs_size,
1582 ARCH_KMALLOC_MINALIGN,
1583 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001584 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001585 }
Christoph Lametere498be72005-09-09 13:03:32 -07001586
Ingo Molnare0a42722006-06-23 02:03:46 -07001587 slab_early_init = 0;
1588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001590 /*
1591 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 * This should be particularly beneficial on SMP boxes, as it
1593 * eliminates "false sharing".
1594 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001595 * allow tighter packing of the smaller caches.
1596 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001597 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001598 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001599 sizes->cs_size,
1600 ARCH_KMALLOC_MINALIGN,
1601 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001602 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001603 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001604#ifdef CONFIG_ZONE_DMA
1605 sizes->cs_dmacachep = kmem_cache_create(
1606 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001607 sizes->cs_size,
1608 ARCH_KMALLOC_MINALIGN,
1609 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1610 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001611 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001612#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 sizes++;
1614 names++;
1615 }
1616 /* 4) Replace the bootstrap head arrays */
1617 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001618 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001619
Pekka Enberg83b519e2009-06-10 19:40:04 +03001620 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001621
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001622 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1623 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001624 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001625 /*
1626 * Do not assume that spinlocks can be initialized via memcpy:
1627 */
1628 spin_lock_init(&ptr->lock);
1629
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001631
Pekka Enberg83b519e2009-06-10 19:40:04 +03001632 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001633
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001634 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001635 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001636 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001637 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001638 /*
1639 * Do not assume that spinlocks can be initialized via memcpy:
1640 */
1641 spin_lock_init(&ptr->lock);
1642
Christoph Lametere498be72005-09-09 13:03:32 -07001643 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001644 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 }
Christoph Lametere498be72005-09-09 13:03:32 -07001646 /* 5) Replace the bootstrap kmem_list3's */
1647 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001648 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Mel Gorman9c09a952008-01-24 05:49:54 -08001650 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001651 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001652
Christoph Lametere498be72005-09-09 13:03:32 -07001653 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001654 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001655
1656 if (INDEX_AC != INDEX_L3) {
1657 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001658 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001659 }
1660 }
1661 }
1662
Pekka Enberg8429db52009-06-12 15:58:59 +03001663 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001664}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001665
Pekka Enberg8429db52009-06-12 15:58:59 +03001666void __init kmem_cache_init_late(void)
1667{
1668 struct kmem_cache *cachep;
1669
Peter Zijlstra52cef182011-11-28 21:12:40 +01001670 g_cpucache_up = LATE;
1671
Peter Zijlstra30765b92011-07-28 23:22:56 +02001672 /* Annotate slab for lockdep -- annotate the malloc caches */
1673 init_lock_keys();
1674
Pekka Enberg8429db52009-06-12 15:58:59 +03001675 /* 6) resize the head arrays to their final sizes */
1676 mutex_lock(&cache_chain_mutex);
1677 list_for_each_entry(cachep, &cache_chain, next)
1678 if (enable_cpucache(cachep, GFP_NOWAIT))
1679 BUG();
1680 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001681
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 /* Done! */
1683 g_cpucache_up = FULL;
1684
Andrew Mortona737b3e2006-03-22 00:08:11 -08001685 /*
1686 * Register a cpu startup notifier callback that initializes
1687 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 */
1689 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
David Rientjes8f9f8d92010-03-27 19:40:47 -07001691#ifdef CONFIG_NUMA
1692 /*
1693 * Register a memory hotplug callback that initializes and frees
1694 * nodelists.
1695 */
1696 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1697#endif
1698
Andrew Mortona737b3e2006-03-22 00:08:11 -08001699 /*
1700 * The reap timers are started later, with a module init call: That part
1701 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 */
1703}
1704
1705static int __init cpucache_init(void)
1706{
1707 int cpu;
1708
Andrew Mortona737b3e2006-03-22 00:08:11 -08001709 /*
1710 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 */
Christoph Lametere498be72005-09-09 13:03:32 -07001712 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001713 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 return 0;
1715}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716__initcall(cpucache_init);
1717
1718/*
1719 * Interface to system's page allocator. No need to hold the cache-lock.
1720 *
1721 * If we requested dmaable memory, we will get it. Even if we
1722 * did not request dmaable memory, we might get it, but that
1723 * would be relatively rare and ignorable.
1724 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001725static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726{
1727 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001728 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 int i;
1730
Luke Yangd6fef9d2006-04-10 22:52:56 -07001731#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001732 /*
1733 * Nommu uses slab's for process anonymous memory allocations, and thus
1734 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001735 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001736 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001737#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001738
Christoph Lameter3c517a62006-12-06 20:33:29 -08001739 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001740 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1741 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001742
Linus Torvalds517d0862009-06-16 19:50:13 -07001743 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 if (!page)
1745 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001747 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001749 add_zone_page_state(page_zone(page),
1750 NR_SLAB_RECLAIMABLE, nr_pages);
1751 else
1752 add_zone_page_state(page_zone(page),
1753 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001754 for (i = 0; i < nr_pages; i++)
1755 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001756
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001757 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1758 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1759
1760 if (cachep->ctor)
1761 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1762 else
1763 kmemcheck_mark_unallocated_pages(page, nr_pages);
1764 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001765
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001766 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767}
1768
1769/*
1770 * Interface to system's page release.
1771 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001772static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001774 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 struct page *page = virt_to_page(addr);
1776 const unsigned long nr_freed = i;
1777
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001778 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001779
Christoph Lameter972d1a72006-09-25 23:31:51 -07001780 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1781 sub_zone_page_state(page_zone(page),
1782 NR_SLAB_RECLAIMABLE, nr_freed);
1783 else
1784 sub_zone_page_state(page_zone(page),
1785 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001787 BUG_ON(!PageSlab(page));
1788 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 page++;
1790 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 if (current->reclaim_state)
1792 current->reclaim_state->reclaimed_slab += nr_freed;
1793 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794}
1795
1796static void kmem_rcu_free(struct rcu_head *head)
1797{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001798 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001799 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800
1801 kmem_freepages(cachep, slab_rcu->addr);
1802 if (OFF_SLAB(cachep))
1803 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1804}
1805
1806#if DEBUG
1807
1808#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001809static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001810 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001812 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001814 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001816 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 return;
1818
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001819 *addr++ = 0x12345678;
1820 *addr++ = caller;
1821 *addr++ = smp_processor_id();
1822 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 {
1824 unsigned long *sptr = &caller;
1825 unsigned long svalue;
1826
1827 while (!kstack_end(sptr)) {
1828 svalue = *sptr++;
1829 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001830 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 size -= sizeof(unsigned long);
1832 if (size <= sizeof(unsigned long))
1833 break;
1834 }
1835 }
1836
1837 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001838 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839}
1840#endif
1841
Pekka Enberg343e0d72006-02-01 03:05:50 -08001842static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001844 int size = obj_size(cachep);
1845 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846
1847 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001848 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849}
1850
1851static void dump_line(char *data, int offset, int limit)
1852{
1853 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001854 unsigned char error = 0;
1855 int bad_count = 0;
1856
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001857 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001858 for (i = 0; i < limit; i++) {
1859 if (data[offset + i] != POISON_FREE) {
1860 error = data[offset + i];
1861 bad_count++;
1862 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001863 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001864 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1865 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001866
1867 if (bad_count == 1) {
1868 error ^= POISON_FREE;
1869 if (!(error & (error - 1))) {
1870 printk(KERN_ERR "Single bit error detected. Probably "
1871 "bad RAM.\n");
1872#ifdef CONFIG_X86
1873 printk(KERN_ERR "Run memtest86+ or a similar memory "
1874 "test tool.\n");
1875#else
1876 printk(KERN_ERR "Run a memory test tool.\n");
1877#endif
1878 }
1879 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880}
1881#endif
1882
1883#if DEBUG
1884
Pekka Enberg343e0d72006-02-01 03:05:50 -08001885static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886{
1887 int i, size;
1888 char *realobj;
1889
1890 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001891 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001892 *dbg_redzone1(cachep, objp),
1893 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 }
1895
1896 if (cachep->flags & SLAB_STORE_USER) {
1897 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001898 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001900 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 printk("\n");
1902 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001903 realobj = (char *)objp + obj_offset(cachep);
1904 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001905 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 int limit;
1907 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001908 if (i + limit > size)
1909 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 dump_line(realobj, i, limit);
1911 }
1912}
1913
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915{
1916 char *realobj;
1917 int size, i;
1918 int lines = 0;
1919
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001920 realobj = (char *)objp + obj_offset(cachep);
1921 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001923 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001925 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 exp = POISON_END;
1927 if (realobj[i] != exp) {
1928 int limit;
1929 /* Mismatch ! */
1930 /* Print header */
1931 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001933 "Slab corruption: %s start=%p, len=%d\n",
1934 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 print_objinfo(cachep, objp, 0);
1936 }
1937 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001938 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001940 if (i + limit > size)
1941 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 dump_line(realobj, i, limit);
1943 i += 16;
1944 lines++;
1945 /* Limit to 5 lines */
1946 if (lines > 5)
1947 break;
1948 }
1949 }
1950 if (lines != 0) {
1951 /* Print some data about the neighboring objects, if they
1952 * exist:
1953 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001954 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001955 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001957 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001959 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001960 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001962 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 print_objinfo(cachep, objp, 2);
1964 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001966 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001967 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001969 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 print_objinfo(cachep, objp, 2);
1971 }
1972 }
1973}
1974#endif
1975
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301977static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001978{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 int i;
1980 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001981 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982
1983 if (cachep->flags & SLAB_POISON) {
1984#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001985 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1986 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001987 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001988 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 else
1990 check_poison_obj(cachep, objp);
1991#else
1992 check_poison_obj(cachep, objp);
1993#endif
1994 }
1995 if (cachep->flags & SLAB_RED_ZONE) {
1996 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1997 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001998 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2000 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002001 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002004}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302006static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002007{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002008}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009#endif
2010
Randy Dunlap911851e2006-03-22 00:08:14 -08002011/**
2012 * slab_destroy - destroy and release all objects in a slab
2013 * @cachep: cache pointer being destroyed
2014 * @slabp: slab pointer being destroyed
2015 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002016 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002017 * Before calling the slab must have been unlinked from the cache. The
2018 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002019 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002020static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002021{
2022 void *addr = slabp->s_mem - slabp->colouroff;
2023
Rabin Vincente79aec22008-07-04 00:40:32 +05302024 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2026 struct slab_rcu *slab_rcu;
2027
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002028 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 slab_rcu->cachep = cachep;
2030 slab_rcu->addr = addr;
2031 call_rcu(&slab_rcu->head, kmem_rcu_free);
2032 } else {
2033 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002034 if (OFF_SLAB(cachep))
2035 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 }
2037}
2038
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002039static void __kmem_cache_destroy(struct kmem_cache *cachep)
2040{
2041 int i;
2042 struct kmem_list3 *l3;
2043
2044 for_each_online_cpu(i)
2045 kfree(cachep->array[i]);
2046
2047 /* NUMA: free the list3 structures */
2048 for_each_online_node(i) {
2049 l3 = cachep->nodelists[i];
2050 if (l3) {
2051 kfree(l3->shared);
2052 free_alien_cache(l3->alien);
2053 kfree(l3);
2054 }
2055 }
2056 kmem_cache_free(&cache_cache, cachep);
2057}
2058
2059
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002061 * calculate_slab_order - calculate size (page order) of slabs
2062 * @cachep: pointer to the cache that is being created
2063 * @size: size of objects to be created in this cache.
2064 * @align: required alignment for the objects.
2065 * @flags: slab allocation flags
2066 *
2067 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002068 *
2069 * This could be made much more intelligent. For now, try to avoid using
2070 * high order pages for slabs. When the gfp() functions are more friendly
2071 * towards high-order requests, this should be changed.
2072 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002073static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002074 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002075{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002076 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002077 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002078 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002079
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002080 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002081 unsigned int num;
2082 size_t remainder;
2083
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002084 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002085 if (!num)
2086 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002087
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002088 if (flags & CFLGS_OFF_SLAB) {
2089 /*
2090 * Max number of objs-per-slab for caches which
2091 * use off-slab slabs. Needed to avoid a possible
2092 * looping condition in cache_grow().
2093 */
2094 offslab_limit = size - sizeof(struct slab);
2095 offslab_limit /= sizeof(kmem_bufctl_t);
2096
2097 if (num > offslab_limit)
2098 break;
2099 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002100
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002101 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002102 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002103 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002104 left_over = remainder;
2105
2106 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002107 * A VFS-reclaimable slab tends to have most allocations
2108 * as GFP_NOFS and we really don't want to have to be allocating
2109 * higher-order pages when we are unable to shrink dcache.
2110 */
2111 if (flags & SLAB_RECLAIM_ACCOUNT)
2112 break;
2113
2114 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002115 * Large number of objects is good, but very large slabs are
2116 * currently bad for the gfp()s.
2117 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002118 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002119 break;
2120
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002121 /*
2122 * Acceptable internal fragmentation?
2123 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002124 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002125 break;
2126 }
2127 return left_over;
2128}
2129
Pekka Enberg83b519e2009-06-10 19:40:04 +03002130static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002131{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002132 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002133 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002134
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002135 if (g_cpucache_up == NONE) {
2136 /*
2137 * Note: the first kmem_cache_create must create the cache
2138 * that's used by kmalloc(24), otherwise the creation of
2139 * further caches will BUG().
2140 */
2141 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2142
2143 /*
2144 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2145 * the first cache, then we need to set up all its list3s,
2146 * otherwise the creation of further caches will BUG().
2147 */
2148 set_up_list3s(cachep, SIZE_AC);
2149 if (INDEX_AC == INDEX_L3)
2150 g_cpucache_up = PARTIAL_L3;
2151 else
2152 g_cpucache_up = PARTIAL_AC;
2153 } else {
2154 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002155 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002156
2157 if (g_cpucache_up == PARTIAL_AC) {
2158 set_up_list3s(cachep, SIZE_L3);
2159 g_cpucache_up = PARTIAL_L3;
2160 } else {
2161 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002162 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002163 cachep->nodelists[node] =
2164 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002165 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002166 BUG_ON(!cachep->nodelists[node]);
2167 kmem_list3_init(cachep->nodelists[node]);
2168 }
2169 }
2170 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002171 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002172 jiffies + REAPTIMEOUT_LIST3 +
2173 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2174
2175 cpu_cache_get(cachep)->avail = 0;
2176 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2177 cpu_cache_get(cachep)->batchcount = 1;
2178 cpu_cache_get(cachep)->touched = 0;
2179 cachep->batchcount = 1;
2180 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002181 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002182}
2183
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002184/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 * kmem_cache_create - Create a cache.
2186 * @name: A string which is used in /proc/slabinfo to identify this cache.
2187 * @size: The size of objects to be created in this cache.
2188 * @align: The required alignment for the objects.
2189 * @flags: SLAB flags
2190 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 *
2192 * Returns a ptr to the cache on success, NULL on failure.
2193 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002194 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 *
2196 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002197 * the module calling this has to destroy the cache before getting unloaded.
2198 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 * The flags are
2200 *
2201 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2202 * to catch references to uninitialised memory.
2203 *
2204 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2205 * for buffer overruns.
2206 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2208 * cacheline. This can be beneficial if you're counting cycles as closely
2209 * as davem.
2210 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002211struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002213 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214{
2215 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002216 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002217 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218
2219 /*
2220 * Sanity checks... these are all serious usage bugs.
2221 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002222 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002223 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002224 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002225 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002226 BUG();
2227 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002229 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002230 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302231 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002232 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002233 if (slab_is_available()) {
2234 get_online_cpus();
2235 mutex_lock(&cache_chain_mutex);
2236 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002237
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002238 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002239 char tmp;
2240 int res;
2241
2242 /*
2243 * This happens when the module gets unloaded and doesn't
2244 * destroy its slab cache and no-one else reuses the vmalloc
2245 * area of the module. Print a warning.
2246 */
Andrew Morton138ae662006-12-06 20:36:41 -08002247 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002248 if (res) {
matzeb4169522007-05-06 14:49:52 -07002249 printk(KERN_ERR
2250 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002251 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002252 continue;
2253 }
2254
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002255 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002256 printk(KERN_ERR
2257 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002258 dump_stack();
2259 goto oops;
2260 }
2261 }
2262
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263#if DEBUG
2264 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265#if FORCED_DEBUG
2266 /*
2267 * Enable redzoning and last user accounting, except for caches with
2268 * large objects, if the increased size would increase the object size
2269 * above the next power of two: caches with object sizes just above a
2270 * power of two have a significant amount of internal fragmentation.
2271 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002272 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2273 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002274 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 if (!(flags & SLAB_DESTROY_BY_RCU))
2276 flags |= SLAB_POISON;
2277#endif
2278 if (flags & SLAB_DESTROY_BY_RCU)
2279 BUG_ON(flags & SLAB_POISON);
2280#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002282 * Always checks flags, a caller might be expecting debug support which
2283 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002285 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Andrew Mortona737b3e2006-03-22 00:08:11 -08002287 /*
2288 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 * unaligned accesses for some archs when redzoning is used, and makes
2290 * sure any on-slab bufctl's are also correctly aligned.
2291 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002292 if (size & (BYTES_PER_WORD - 1)) {
2293 size += (BYTES_PER_WORD - 1);
2294 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 }
2296
Andrew Mortona737b3e2006-03-22 00:08:11 -08002297 /* calculate the final buffer alignment: */
2298
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 /* 1) arch recommendation: can be overridden for debug */
2300 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002301 /*
2302 * Default alignment: as specified by the arch code. Except if
2303 * an object is really small, then squeeze multiple objects into
2304 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 */
2306 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002307 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 ralign /= 2;
2309 } else {
2310 ralign = BYTES_PER_WORD;
2311 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002312
2313 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002314 * Redzoning and user store require word alignment or possibly larger.
2315 * Note this will be overridden by architecture or caller mandated
2316 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002317 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002318 if (flags & SLAB_STORE_USER)
2319 ralign = BYTES_PER_WORD;
2320
2321 if (flags & SLAB_RED_ZONE) {
2322 ralign = REDZONE_ALIGN;
2323 /* If redzoning, ensure that the second redzone is suitably
2324 * aligned, by adjusting the object size accordingly. */
2325 size += REDZONE_ALIGN - 1;
2326 size &= ~(REDZONE_ALIGN - 1);
2327 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002328
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002329 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 if (ralign < ARCH_SLAB_MINALIGN) {
2331 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002333 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 if (ralign < align) {
2335 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002337 /* disable debug if necessary */
2338 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002339 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002340 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002341 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 */
2343 align = ralign;
2344
Pekka Enberg83b519e2009-06-10 19:40:04 +03002345 if (slab_is_available())
2346 gfp = GFP_KERNEL;
2347 else
2348 gfp = GFP_NOWAIT;
2349
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002351 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002353 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002355 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002357 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
Pekka Enbergca5f9702006-09-25 23:31:25 -07002359 /*
2360 * Both debugging options require word-alignment which is calculated
2361 * into align above.
2362 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002365 cachep->obj_offset += sizeof(unsigned long long);
2366 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 }
2368 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002369 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002370 * the real object. But if the second red zone needs to be
2371 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002373 if (flags & SLAB_RED_ZONE)
2374 size += REDZONE_ALIGN;
2375 else
2376 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 }
2378#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002379 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002380 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2381 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 size = PAGE_SIZE;
2383 }
2384#endif
2385#endif
2386
Ingo Molnare0a42722006-06-23 02:03:46 -07002387 /*
2388 * Determine if the slab management is 'on' or 'off' slab.
2389 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002390 * it too early on. Always use on-slab management when
2391 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002392 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002393 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2394 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 /*
2396 * Size is large, assume best to place the slab management obj
2397 * off-slab (should allow better packing of objs).
2398 */
2399 flags |= CFLGS_OFF_SLAB;
2400
2401 size = ALIGN(size, align);
2402
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002403 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404
2405 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002406 printk(KERN_ERR
2407 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 kmem_cache_free(&cache_cache, cachep);
2409 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002410 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002412 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2413 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414
2415 /*
2416 * If the slab has been placed off-slab, and we have enough space then
2417 * move it on-slab. This is at the expense of any extra colouring.
2418 */
2419 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2420 flags &= ~CFLGS_OFF_SLAB;
2421 left_over -= slab_size;
2422 }
2423
2424 if (flags & CFLGS_OFF_SLAB) {
2425 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002426 slab_size =
2427 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302428
2429#ifdef CONFIG_PAGE_POISONING
2430 /* If we're going to use the generic kernel_map_pages()
2431 * poisoning, then it's going to smash the contents of
2432 * the redzone and userword anyhow, so switch them off.
2433 */
2434 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2435 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2436#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 }
2438
2439 cachep->colour_off = cache_line_size();
2440 /* Offset must be a multiple of the alignment. */
2441 if (cachep->colour_off < align)
2442 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002443 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 cachep->slab_size = slab_size;
2445 cachep->flags = flags;
2446 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002447 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002449 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002450 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002452 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002453 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002454 /*
2455 * This is a possibility for one of the malloc_sizes caches.
2456 * But since we go off slab only for object size greater than
2457 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2458 * this should not happen at all.
2459 * But leave a BUG_ON for some lucky dude.
2460 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002461 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 cachep->name = name;
2465
Pekka Enberg83b519e2009-06-10 19:40:04 +03002466 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002467 __kmem_cache_destroy(cachep);
2468 cachep = NULL;
2469 goto oops;
2470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471
Peter Zijlstra83835b32011-07-22 15:26:05 +02002472 if (flags & SLAB_DEBUG_OBJECTS) {
2473 /*
2474 * Would deadlock through slab_destroy()->call_rcu()->
2475 * debug_object_activate()->kmem_cache_alloc().
2476 */
2477 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2478
2479 slab_set_debugobj_lock_classes(cachep);
2480 }
2481
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 /* cache setup completed, link it into the list */
2483 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002484oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 if (!cachep && (flags & SLAB_PANIC))
2486 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002487 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002488 if (slab_is_available()) {
2489 mutex_unlock(&cache_chain_mutex);
2490 put_online_cpus();
2491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 return cachep;
2493}
2494EXPORT_SYMBOL(kmem_cache_create);
2495
2496#if DEBUG
2497static void check_irq_off(void)
2498{
2499 BUG_ON(!irqs_disabled());
2500}
2501
2502static void check_irq_on(void)
2503{
2504 BUG_ON(irqs_disabled());
2505}
2506
Pekka Enberg343e0d72006-02-01 03:05:50 -08002507static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508{
2509#ifdef CONFIG_SMP
2510 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002511 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512#endif
2513}
Christoph Lametere498be72005-09-09 13:03:32 -07002514
Pekka Enberg343e0d72006-02-01 03:05:50 -08002515static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002516{
2517#ifdef CONFIG_SMP
2518 check_irq_off();
2519 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2520#endif
2521}
2522
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523#else
2524#define check_irq_off() do { } while(0)
2525#define check_irq_on() do { } while(0)
2526#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002527#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528#endif
2529
Christoph Lameteraab22072006-03-22 00:09:06 -08002530static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2531 struct array_cache *ac,
2532 int force, int node);
2533
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534static void do_drain(void *arg)
2535{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002536 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002538 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539
2540 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002541 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002542 spin_lock(&cachep->nodelists[node]->list_lock);
2543 free_block(cachep, ac->entry, ac->avail, node);
2544 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 ac->avail = 0;
2546}
2547
Pekka Enberg343e0d72006-02-01 03:05:50 -08002548static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549{
Christoph Lametere498be72005-09-09 13:03:32 -07002550 struct kmem_list3 *l3;
2551 int node;
2552
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002553 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002555 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002556 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002557 if (l3 && l3->alien)
2558 drain_alien_cache(cachep, l3->alien);
2559 }
2560
2561 for_each_online_node(node) {
2562 l3 = cachep->nodelists[node];
2563 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002564 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002565 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566}
2567
Christoph Lametered11d9e2006-06-30 01:55:45 -07002568/*
2569 * Remove slabs from the list of free slabs.
2570 * Specify the number of slabs to drain in tofree.
2571 *
2572 * Returns the actual number of slabs released.
2573 */
2574static int drain_freelist(struct kmem_cache *cache,
2575 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002577 struct list_head *p;
2578 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
Christoph Lametered11d9e2006-06-30 01:55:45 -07002581 nr_freed = 0;
2582 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583
Christoph Lametered11d9e2006-06-30 01:55:45 -07002584 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002585 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002586 if (p == &l3->slabs_free) {
2587 spin_unlock_irq(&l3->list_lock);
2588 goto out;
2589 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590
Christoph Lametered11d9e2006-06-30 01:55:45 -07002591 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002593 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594#endif
2595 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002596 /*
2597 * Safe to drop the lock. The slab is no longer linked
2598 * to the cache.
2599 */
2600 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002601 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002602 slab_destroy(cache, slabp);
2603 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002605out:
2606 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607}
2608
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002609/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002610static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002611{
2612 int ret = 0, i = 0;
2613 struct kmem_list3 *l3;
2614
2615 drain_cpu_caches(cachep);
2616
2617 check_irq_on();
2618 for_each_online_node(i) {
2619 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002620 if (!l3)
2621 continue;
2622
2623 drain_freelist(cachep, l3, l3->free_objects);
2624
2625 ret += !list_empty(&l3->slabs_full) ||
2626 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002627 }
2628 return (ret ? 1 : 0);
2629}
2630
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631/**
2632 * kmem_cache_shrink - Shrink a cache.
2633 * @cachep: The cache to shrink.
2634 *
2635 * Releases as many slabs as possible for a cache.
2636 * To help debugging, a zero exit status indicates all slabs were released.
2637 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002638int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002640 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002641 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002643 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002644 mutex_lock(&cache_chain_mutex);
2645 ret = __cache_shrink(cachep);
2646 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002647 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002648 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649}
2650EXPORT_SYMBOL(kmem_cache_shrink);
2651
2652/**
2653 * kmem_cache_destroy - delete a cache
2654 * @cachep: the cache to destroy
2655 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002656 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 *
2658 * It is expected this function will be called by a module when it is
2659 * unloaded. This will remove the cache completely, and avoid a duplicate
2660 * cache being allocated each time a module is loaded and unloaded, if the
2661 * module doesn't have persistent in-kernel storage across loads and unloads.
2662 *
2663 * The cache must be empty before calling this function.
2664 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002665 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 * during the kmem_cache_destroy().
2667 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002668void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002670 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002673 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002674 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 /*
2676 * the chain is never empty, cache_cache is never destroyed
2677 */
2678 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 if (__cache_shrink(cachep)) {
2680 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002681 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002682 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002683 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002684 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 }
2686
2687 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002688 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002690 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002691 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002692 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693}
2694EXPORT_SYMBOL(kmem_cache_destroy);
2695
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002696/*
2697 * Get the memory for a slab management obj.
2698 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2699 * always come from malloc_sizes caches. The slab descriptor cannot
2700 * come from the same cache which is getting created because,
2701 * when we are searching for an appropriate cache for these
2702 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2703 * If we are creating a malloc_sizes cache here it would not be visible to
2704 * kmem_find_general_cachep till the initialization is complete.
2705 * Hence we cannot have slabp_cache same as the original cache.
2706 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002707static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002708 int colour_off, gfp_t local_flags,
2709 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710{
2711 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 if (OFF_SLAB(cachep)) {
2714 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002715 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002716 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002717 /*
2718 * If the first object in the slab is leaked (it's allocated
2719 * but no one has a reference to it), we want to make sure
2720 * kmemleak does not treat the ->s_mem pointer as a reference
2721 * to the object. Otherwise we will not report the leak.
2722 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002723 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2724 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 if (!slabp)
2726 return NULL;
2727 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002728 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 colour_off += cachep->slab_size;
2730 }
2731 slabp->inuse = 0;
2732 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002733 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002734 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002735 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 return slabp;
2737}
2738
2739static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2740{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742}
2743
Pekka Enberg343e0d72006-02-01 03:05:50 -08002744static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002745 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746{
2747 int i;
2748
2749 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002750 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751#if DEBUG
2752 /* need to poison the objs? */
2753 if (cachep->flags & SLAB_POISON)
2754 poison_obj(cachep, objp, POISON_FREE);
2755 if (cachep->flags & SLAB_STORE_USER)
2756 *dbg_userword(cachep, objp) = NULL;
2757
2758 if (cachep->flags & SLAB_RED_ZONE) {
2759 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2760 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2761 }
2762 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002763 * Constructors are not allowed to allocate memory from the same
2764 * cache which they are a constructor for. Otherwise, deadlock.
2765 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 */
2767 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002768 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769
2770 if (cachep->flags & SLAB_RED_ZONE) {
2771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002773 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2775 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002776 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002778 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2779 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002780 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002781 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782#else
2783 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002784 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002786 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002788 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789}
2790
Pekka Enberg343e0d72006-02-01 03:05:50 -08002791static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002793 if (CONFIG_ZONE_DMA_FLAG) {
2794 if (flags & GFP_DMA)
2795 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2796 else
2797 BUG_ON(cachep->gfpflags & GFP_DMA);
2798 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799}
2800
Andrew Mortona737b3e2006-03-22 00:08:11 -08002801static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2802 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002803{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002804 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002805 kmem_bufctl_t next;
2806
2807 slabp->inuse++;
2808 next = slab_bufctl(slabp)[slabp->free];
2809#if DEBUG
2810 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2811 WARN_ON(slabp->nodeid != nodeid);
2812#endif
2813 slabp->free = next;
2814
2815 return objp;
2816}
2817
Andrew Mortona737b3e2006-03-22 00:08:11 -08002818static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2819 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002820{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002821 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002822
2823#if DEBUG
2824 /* Verify that the slab belongs to the intended node */
2825 WARN_ON(slabp->nodeid != nodeid);
2826
Al Viro871751e2006-03-25 03:06:39 -08002827 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002828 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002829 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002830 BUG();
2831 }
2832#endif
2833 slab_bufctl(slabp)[objnr] = slabp->free;
2834 slabp->free = objnr;
2835 slabp->inuse--;
2836}
2837
Pekka Enberg47768742006-06-23 02:03:07 -07002838/*
2839 * Map pages beginning at addr to the given cache and slab. This is required
2840 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002841 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002842 */
2843static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2844 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845{
Pekka Enberg47768742006-06-23 02:03:07 -07002846 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 struct page *page;
2848
Pekka Enberg47768742006-06-23 02:03:07 -07002849 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002850
Pekka Enberg47768742006-06-23 02:03:07 -07002851 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002852 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002853 nr_pages <<= cache->gfporder;
2854
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002856 page_set_cache(page, cache);
2857 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002859 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860}
2861
2862/*
2863 * Grow (by 1) the number of slabs within a cache. This is called by
2864 * kmem_cache_alloc() when there are no active objs left in a cache.
2865 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002866static int cache_grow(struct kmem_cache *cachep,
2867 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002869 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002870 size_t offset;
2871 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002872 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
Andrew Mortona737b3e2006-03-22 00:08:11 -08002874 /*
2875 * Be lazy and only check for valid flags here, keeping it out of the
2876 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002878 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2879 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002881 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002883 l3 = cachep->nodelists[nodeid];
2884 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885
2886 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002887 offset = l3->colour_next;
2888 l3->colour_next++;
2889 if (l3->colour_next >= cachep->colour)
2890 l3->colour_next = 0;
2891 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002893 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894
2895 if (local_flags & __GFP_WAIT)
2896 local_irq_enable();
2897
2898 /*
2899 * The test for missing atomic flag is performed here, rather than
2900 * the more obvious place, simply to reduce the critical path length
2901 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2902 * will eventually be caught here (where it matters).
2903 */
2904 kmem_flagcheck(cachep, flags);
2905
Andrew Mortona737b3e2006-03-22 00:08:11 -08002906 /*
2907 * Get mem for the objs. Attempt to allocate a physical page from
2908 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002909 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002910 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002911 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002912 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 goto failed;
2914
2915 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002916 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002917 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002918 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 goto opps1;
2920
Pekka Enberg47768742006-06-23 02:03:07 -07002921 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922
Christoph Lametera35afb82007-05-16 22:10:57 -07002923 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
2925 if (local_flags & __GFP_WAIT)
2926 local_irq_disable();
2927 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002928 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929
2930 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002931 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002933 l3->free_objects += cachep->num;
2934 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002936opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002938failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 if (local_flags & __GFP_WAIT)
2940 local_irq_disable();
2941 return 0;
2942}
2943
2944#if DEBUG
2945
2946/*
2947 * Perform extra freeing checks:
2948 * - detect bad pointers.
2949 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 */
2951static void kfree_debugcheck(const void *objp)
2952{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 if (!virt_addr_valid(objp)) {
2954 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002955 (unsigned long)objp);
2956 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958}
2959
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002960static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2961{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002962 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002963
2964 redzone1 = *dbg_redzone1(cache, obj);
2965 redzone2 = *dbg_redzone2(cache, obj);
2966
2967 /*
2968 * Redzone is ok.
2969 */
2970 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2971 return;
2972
2973 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2974 slab_error(cache, "double free detected");
2975 else
2976 slab_error(cache, "memory outside object was overwritten");
2977
David Woodhouseb46b8f12007-05-08 00:22:59 -07002978 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002979 obj, redzone1, redzone2);
2980}
2981
Pekka Enberg343e0d72006-02-01 03:05:50 -08002982static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002983 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984{
2985 struct page *page;
2986 unsigned int objnr;
2987 struct slab *slabp;
2988
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002989 BUG_ON(virt_to_cache(objp) != cachep);
2990
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002991 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002993 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994
Pekka Enberg065d41c2005-11-13 16:06:46 -08002995 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996
2997 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002998 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3000 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3001 }
3002 if (cachep->flags & SLAB_STORE_USER)
3003 *dbg_userword(cachep, objp) = caller;
3004
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003005 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006
3007 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003008 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009
Al Viro871751e2006-03-25 03:06:39 -08003010#ifdef CONFIG_DEBUG_SLAB_LEAK
3011 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3012#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 if (cachep->flags & SLAB_POISON) {
3014#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08003015 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003017 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003018 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 } else {
3020 poison_obj(cachep, objp, POISON_FREE);
3021 }
3022#else
3023 poison_obj(cachep, objp, POISON_FREE);
3024#endif
3025 }
3026 return objp;
3027}
3028
Pekka Enberg343e0d72006-02-01 03:05:50 -08003029static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030{
3031 kmem_bufctl_t i;
3032 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003033
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 /* Check slab's freelist to see if this obj is there. */
3035 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3036 entries++;
3037 if (entries > cachep->num || i >= cachep->num)
3038 goto bad;
3039 }
3040 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003041bad:
3042 printk(KERN_ERR "slab: Internal list corruption detected in "
3043 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3044 cachep->name, cachep->num, slabp, slabp->inuse);
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003045 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3046 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3047 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 BUG();
3049 }
3050}
3051#else
3052#define kfree_debugcheck(x) do { } while(0)
3053#define cache_free_debugcheck(x,objp,z) (objp)
3054#define check_slabp(x,y) do { } while(0)
3055#endif
3056
Pekka Enberg343e0d72006-02-01 03:05:50 -08003057static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058{
3059 int batchcount;
3060 struct kmem_list3 *l3;
3061 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003062 int node;
3063
Andrew Mortona737b3e2006-03-22 00:08:11 -08003064retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003065 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003066 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003067 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 batchcount = ac->batchcount;
3069 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003070 /*
3071 * If there was little recent activity on this cache, then
3072 * perform only a partial refill. Otherwise we could generate
3073 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074 */
3075 batchcount = BATCHREFILL_LIMIT;
3076 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003077 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078
Christoph Lametere498be72005-09-09 13:03:32 -07003079 BUG_ON(ac->avail > 0 || !l3);
3080 spin_lock(&l3->list_lock);
3081
Christoph Lameter3ded1752006-03-25 03:06:44 -08003082 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003083 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3084 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003085 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003086 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003087
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 while (batchcount > 0) {
3089 struct list_head *entry;
3090 struct slab *slabp;
3091 /* Get slab alloc is to come from. */
3092 entry = l3->slabs_partial.next;
3093 if (entry == &l3->slabs_partial) {
3094 l3->free_touched = 1;
3095 entry = l3->slabs_free.next;
3096 if (entry == &l3->slabs_free)
3097 goto must_grow;
3098 }
3099
3100 slabp = list_entry(entry, struct slab, list);
3101 check_slabp(cachep, slabp);
3102 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003103
3104 /*
3105 * The slab was either on partial or free list so
3106 * there must be at least one object available for
3107 * allocation.
3108 */
roel kluin249b9f32008-10-29 17:18:07 -04003109 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003110
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 STATS_INC_ALLOCED(cachep);
3113 STATS_INC_ACTIVE(cachep);
3114 STATS_SET_HIGH(cachep);
3115
Matthew Dobson78d382d2006-02-01 03:05:47 -08003116 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003117 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 }
3119 check_slabp(cachep, slabp);
3120
3121 /* move slabp to correct slabp list: */
3122 list_del(&slabp->list);
3123 if (slabp->free == BUFCTL_END)
3124 list_add(&slabp->list, &l3->slabs_full);
3125 else
3126 list_add(&slabp->list, &l3->slabs_partial);
3127 }
3128
Andrew Mortona737b3e2006-03-22 00:08:11 -08003129must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003131alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003132 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133
3134 if (unlikely(!ac->avail)) {
3135 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003136 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003137
Andrew Mortona737b3e2006-03-22 00:08:11 -08003138 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003139 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003140 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 return NULL;
3142
Andrew Mortona737b3e2006-03-22 00:08:11 -08003143 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 goto retry;
3145 }
3146 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003147 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148}
3149
Andrew Mortona737b3e2006-03-22 00:08:11 -08003150static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3151 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152{
3153 might_sleep_if(flags & __GFP_WAIT);
3154#if DEBUG
3155 kmem_flagcheck(cachep, flags);
3156#endif
3157}
3158
3159#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003160static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3161 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003163 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003165 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003167 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003168 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003169 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 else
3171 check_poison_obj(cachep, objp);
3172#else
3173 check_poison_obj(cachep, objp);
3174#endif
3175 poison_obj(cachep, objp, POISON_INUSE);
3176 }
3177 if (cachep->flags & SLAB_STORE_USER)
3178 *dbg_userword(cachep, objp) = caller;
3179
3180 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003181 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3182 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3183 slab_error(cachep, "double free, or memory outside"
3184 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003185 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003186 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003187 objp, *dbg_redzone1(cachep, objp),
3188 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 }
3190 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3191 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3192 }
Al Viro871751e2006-03-25 03:06:39 -08003193#ifdef CONFIG_DEBUG_SLAB_LEAK
3194 {
3195 struct slab *slabp;
3196 unsigned objnr;
3197
Christoph Lameterb49af682007-05-06 14:49:41 -07003198 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003199 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3200 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3201 }
3202#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003203 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003204 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003205 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003206 if (ARCH_SLAB_MINALIGN &&
3207 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003208 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003209 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 return objp;
3212}
3213#else
3214#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3215#endif
3216
Akinobu Mita773ff602008-12-23 19:37:01 +09003217static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003218{
3219 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003220 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003221
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003222 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003223}
3224
Pekka Enberg343e0d72006-02-01 03:05:50 -08003225static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003227 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228 struct array_cache *ac;
3229
Alok N Kataria5c382302005-09-27 21:45:46 -07003230 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003231
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003232 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233 if (likely(ac->avail)) {
3234 STATS_INC_ALLOCHIT(cachep);
3235 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003236 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 } else {
3238 STATS_INC_ALLOCMISS(cachep);
3239 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003240 /*
3241 * the 'ac' may be updated by cache_alloc_refill(),
3242 * and kmemleak_erase() requires its correct value.
3243 */
3244 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003245 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003246 /*
3247 * To avoid a false negative, if an object that is in one of the
3248 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3249 * treat the array pointers as a reference to the object.
3250 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003251 if (objp)
3252 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003253 return objp;
3254}
3255
Christoph Lametere498be72005-09-09 13:03:32 -07003256#ifdef CONFIG_NUMA
3257/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003258 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003259 *
3260 * If we are in_interrupt, then process context, including cpusets and
3261 * mempolicy, may not apply and should not be used for allocation policy.
3262 */
3263static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3264{
3265 int nid_alloc, nid_here;
3266
Christoph Lameter765c4502006-09-27 01:50:08 -07003267 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003268 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003269 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003270 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003271 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003272 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003273 else if (current->mempolicy)
3274 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003275 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003276 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003277 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003278 return NULL;
3279}
3280
3281/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003282 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003283 * certain node and fall back is permitted. First we scan all the
3284 * available nodelists for available objects. If that fails then we
3285 * perform an allocation without specifying a node. This allows the page
3286 * allocator to do its reclaim / fallback magic. We then insert the
3287 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003288 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003289static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003290{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003291 struct zonelist *zonelist;
3292 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003293 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003294 struct zone *zone;
3295 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003296 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003297 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003298
3299 if (flags & __GFP_THISNODE)
3300 return NULL;
3301
Miao Xiec0ff7452010-05-24 14:32:08 -07003302 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003303 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003304 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003305
Christoph Lameter3c517a62006-12-06 20:33:29 -08003306retry:
3307 /*
3308 * Look through allowed nodes for objects available
3309 * from existing per node queues.
3310 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003311 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3312 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003313
Mel Gorman54a6eb52008-04-28 02:12:16 -07003314 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003315 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003316 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003317 obj = ____cache_alloc_node(cache,
3318 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003319 if (obj)
3320 break;
3321 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003322 }
3323
Christoph Lametercfce6602007-05-06 14:50:17 -07003324 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003325 /*
3326 * This allocation will be performed within the constraints
3327 * of the current cpuset / memory policy requirements.
3328 * We may trigger various forms of reclaim on the allowed
3329 * set and go into memory reserves if necessary.
3330 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003331 if (local_flags & __GFP_WAIT)
3332 local_irq_enable();
3333 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003334 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003335 if (local_flags & __GFP_WAIT)
3336 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003337 if (obj) {
3338 /*
3339 * Insert into the appropriate per node queues
3340 */
3341 nid = page_to_nid(virt_to_page(obj));
3342 if (cache_grow(cache, flags, nid, obj)) {
3343 obj = ____cache_alloc_node(cache,
3344 flags | GFP_THISNODE, nid);
3345 if (!obj)
3346 /*
3347 * Another processor may allocate the
3348 * objects in the slab since we are
3349 * not holding any locks.
3350 */
3351 goto retry;
3352 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003353 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003354 obj = NULL;
3355 }
3356 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003357 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003358 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003359 return obj;
3360}
3361
3362/*
Christoph Lametere498be72005-09-09 13:03:32 -07003363 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003365static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003366 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003367{
3368 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003369 struct slab *slabp;
3370 struct kmem_list3 *l3;
3371 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003372 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003374 l3 = cachep->nodelists[nodeid];
3375 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003376
Andrew Mortona737b3e2006-03-22 00:08:11 -08003377retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003378 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003379 spin_lock(&l3->list_lock);
3380 entry = l3->slabs_partial.next;
3381 if (entry == &l3->slabs_partial) {
3382 l3->free_touched = 1;
3383 entry = l3->slabs_free.next;
3384 if (entry == &l3->slabs_free)
3385 goto must_grow;
3386 }
Christoph Lametere498be72005-09-09 13:03:32 -07003387
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003388 slabp = list_entry(entry, struct slab, list);
3389 check_spinlock_acquired_node(cachep, nodeid);
3390 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003391
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003392 STATS_INC_NODEALLOCS(cachep);
3393 STATS_INC_ACTIVE(cachep);
3394 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003395
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003396 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003397
Matthew Dobson78d382d2006-02-01 03:05:47 -08003398 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003399 check_slabp(cachep, slabp);
3400 l3->free_objects--;
3401 /* move slabp to correct slabp list: */
3402 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003403
Andrew Mortona737b3e2006-03-22 00:08:11 -08003404 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003405 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003406 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003407 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003408
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003409 spin_unlock(&l3->list_lock);
3410 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003411
Andrew Mortona737b3e2006-03-22 00:08:11 -08003412must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003413 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003414 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003415 if (x)
3416 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003417
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003418 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003419
Andrew Mortona737b3e2006-03-22 00:08:11 -08003420done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003421 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003422}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003423
3424/**
3425 * kmem_cache_alloc_node - Allocate an object on the specified node
3426 * @cachep: The cache to allocate from.
3427 * @flags: See kmalloc().
3428 * @nodeid: node number of the target node.
3429 * @caller: return address of caller, used for debug information
3430 *
3431 * Identical to kmem_cache_alloc but it will allocate memory on the given
3432 * node, which can improve the performance for cpu bound structures.
3433 *
3434 * Fallback to other node is possible if __GFP_THISNODE is not set.
3435 */
3436static __always_inline void *
3437__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3438 void *caller)
3439{
3440 unsigned long save_flags;
3441 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003442 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003443
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003444 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003445
Nick Piggincf40bd12009-01-21 08:12:39 +01003446 lockdep_trace_alloc(flags);
3447
Akinobu Mita773ff602008-12-23 19:37:01 +09003448 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003449 return NULL;
3450
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003451 cache_alloc_debugcheck_before(cachep, flags);
3452 local_irq_save(save_flags);
3453
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003454 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003455 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003456
3457 if (unlikely(!cachep->nodelists[nodeid])) {
3458 /* Node not bootstrapped yet */
3459 ptr = fallback_alloc(cachep, flags);
3460 goto out;
3461 }
3462
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003463 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003464 /*
3465 * Use the locally cached objects if possible.
3466 * However ____cache_alloc does not allow fallback
3467 * to other nodes. It may fail while we still have
3468 * objects on other nodes available.
3469 */
3470 ptr = ____cache_alloc(cachep, flags);
3471 if (ptr)
3472 goto out;
3473 }
3474 /* ___cache_alloc_node can fall back to other nodes */
3475 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3476 out:
3477 local_irq_restore(save_flags);
3478 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003479 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3480 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003481
Pekka Enbergc175eea2008-05-09 20:35:53 +02003482 if (likely(ptr))
3483 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3484
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003485 if (unlikely((flags & __GFP_ZERO) && ptr))
3486 memset(ptr, 0, obj_size(cachep));
3487
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003488 return ptr;
3489}
3490
3491static __always_inline void *
3492__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3493{
3494 void *objp;
3495
3496 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3497 objp = alternate_node_alloc(cache, flags);
3498 if (objp)
3499 goto out;
3500 }
3501 objp = ____cache_alloc(cache, flags);
3502
3503 /*
3504 * We may just have run out of memory on the local node.
3505 * ____cache_alloc_node() knows how to locate memory on other nodes
3506 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003507 if (!objp)
3508 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003509
3510 out:
3511 return objp;
3512}
3513#else
3514
3515static __always_inline void *
3516__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3517{
3518 return ____cache_alloc(cachep, flags);
3519}
3520
3521#endif /* CONFIG_NUMA */
3522
3523static __always_inline void *
3524__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3525{
3526 unsigned long save_flags;
3527 void *objp;
3528
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003529 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003530
Nick Piggincf40bd12009-01-21 08:12:39 +01003531 lockdep_trace_alloc(flags);
3532
Akinobu Mita773ff602008-12-23 19:37:01 +09003533 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003534 return NULL;
3535
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003536 cache_alloc_debugcheck_before(cachep, flags);
3537 local_irq_save(save_flags);
3538 objp = __do_cache_alloc(cachep, flags);
3539 local_irq_restore(save_flags);
3540 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003541 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3542 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003543 prefetchw(objp);
3544
Pekka Enbergc175eea2008-05-09 20:35:53 +02003545 if (likely(objp))
3546 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3547
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003548 if (unlikely((flags & __GFP_ZERO) && objp))
3549 memset(objp, 0, obj_size(cachep));
3550
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003551 return objp;
3552}
Christoph Lametere498be72005-09-09 13:03:32 -07003553
3554/*
3555 * Caller needs to acquire correct kmem_list's list_lock
3556 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003557static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003558 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559{
3560 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003561 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562
3563 for (i = 0; i < nr_objects; i++) {
3564 void *objp = objpp[i];
3565 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003567 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003568 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003570 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003572 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003574 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575 check_slabp(cachep, slabp);
3576
3577 /* fixup slab chains */
3578 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003579 if (l3->free_objects > l3->free_limit) {
3580 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003581 /* No need to drop any previously held
3582 * lock here, even if we have a off-slab slab
3583 * descriptor it is guaranteed to come from
3584 * a different cache, refer to comments before
3585 * alloc_slabmgmt.
3586 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 slab_destroy(cachep, slabp);
3588 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003589 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 }
3591 } else {
3592 /* Unconditionally move a slab to the end of the
3593 * partial list on free - maximum time for the
3594 * other objects to be freed, too.
3595 */
Christoph Lametere498be72005-09-09 13:03:32 -07003596 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 }
3598 }
3599}
3600
Pekka Enberg343e0d72006-02-01 03:05:50 -08003601static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602{
3603 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003604 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003605 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606
3607 batchcount = ac->batchcount;
3608#if DEBUG
3609 BUG_ON(!batchcount || batchcount > ac->avail);
3610#endif
3611 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003612 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003613 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003614 if (l3->shared) {
3615 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003616 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 if (max) {
3618 if (batchcount > max)
3619 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003620 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003621 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622 shared_array->avail += batchcount;
3623 goto free_done;
3624 }
3625 }
3626
Christoph Lameterff694162005-09-22 21:44:02 -07003627 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003628free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629#if STATS
3630 {
3631 int i = 0;
3632 struct list_head *p;
3633
Christoph Lametere498be72005-09-09 13:03:32 -07003634 p = l3->slabs_free.next;
3635 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 struct slab *slabp;
3637
3638 slabp = list_entry(p, struct slab, list);
3639 BUG_ON(slabp->inuse);
3640
3641 i++;
3642 p = p->next;
3643 }
3644 STATS_SET_FREEABLE(cachep, i);
3645 }
3646#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003647 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003649 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650}
3651
3652/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003653 * Release an obj back to its cache. If the obj has a constructed state, it must
3654 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003656static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3657 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003659 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660
3661 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003662 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003663 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664
Pekka Enbergc175eea2008-05-09 20:35:53 +02003665 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3666
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003667 /*
3668 * Skip calling cache_free_alien() when the platform is not numa.
3669 * This will avoid cache misses that happen while accessing slabp (which
3670 * is per page memory reference) to get nodeid. Instead use a global
3671 * variable to skip the call, which is mostly likely to be present in
3672 * the cache.
3673 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003674 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003675 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003676
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 if (likely(ac->avail < ac->limit)) {
3678 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003679 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 return;
3681 } else {
3682 STATS_INC_FREEMISS(cachep);
3683 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003684 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685 }
3686}
3687
3688/**
3689 * kmem_cache_alloc - Allocate an object
3690 * @cachep: The cache to allocate from.
3691 * @flags: See kmalloc().
3692 *
3693 * Allocate an object from this cache. The flags are only relevant
3694 * if the cache has no available objects.
3695 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003696void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003698 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3699
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003700 trace_kmem_cache_alloc(_RET_IP_, ret,
3701 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003702
3703 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704}
3705EXPORT_SYMBOL(kmem_cache_alloc);
3706
Li Zefan0f24f122009-12-11 15:45:30 +08003707#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003708void *
3709kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003710{
Steven Rostedt85beb582010-11-24 16:23:34 -05003711 void *ret;
3712
3713 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3714
3715 trace_kmalloc(_RET_IP_, ret,
3716 size, slab_buffer_size(cachep), flags);
3717 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003718}
Steven Rostedt85beb582010-11-24 16:23:34 -05003719EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003720#endif
3721
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003723void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3724{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003725 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3726 __builtin_return_address(0));
3727
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003728 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3729 obj_size(cachep), cachep->buffer_size,
3730 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003731
3732 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003733}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734EXPORT_SYMBOL(kmem_cache_alloc_node);
3735
Li Zefan0f24f122009-12-11 15:45:30 +08003736#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003737void *kmem_cache_alloc_node_trace(size_t size,
3738 struct kmem_cache *cachep,
3739 gfp_t flags,
3740 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003741{
Steven Rostedt85beb582010-11-24 16:23:34 -05003742 void *ret;
3743
3744 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003745 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003746 trace_kmalloc_node(_RET_IP_, ret,
3747 size, slab_buffer_size(cachep),
3748 flags, nodeid);
3749 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003750}
Steven Rostedt85beb582010-11-24 16:23:34 -05003751EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003752#endif
3753
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003754static __always_inline void *
3755__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003756{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003757 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003758
3759 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003760 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3761 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003762 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003763}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003764
Li Zefan0bb38a52009-12-11 15:45:50 +08003765#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003766void *__kmalloc_node(size_t size, gfp_t flags, int node)
3767{
3768 return __do_kmalloc_node(size, flags, node,
3769 __builtin_return_address(0));
3770}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003771EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003772
3773void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003774 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003775{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003776 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003777}
3778EXPORT_SYMBOL(__kmalloc_node_track_caller);
3779#else
3780void *__kmalloc_node(size_t size, gfp_t flags, int node)
3781{
3782 return __do_kmalloc_node(size, flags, node, NULL);
3783}
3784EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003785#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003786#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787
3788/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003789 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003791 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003792 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003794static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3795 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003797 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003798 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003800 /* If you want to save a few bytes .text space: replace
3801 * __ with kmem_.
3802 * Then kmalloc uses the uninlined functions instead of the inline
3803 * functions.
3804 */
3805 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003806 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3807 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003808 ret = __cache_alloc(cachep, flags, caller);
3809
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003810 trace_kmalloc((unsigned long) caller, ret,
3811 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003812
3813 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003814}
3815
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003816
Li Zefan0bb38a52009-12-11 15:45:50 +08003817#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003818void *__kmalloc(size_t size, gfp_t flags)
3819{
Al Viro871751e2006-03-25 03:06:39 -08003820 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821}
3822EXPORT_SYMBOL(__kmalloc);
3823
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003824void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003825{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003826 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003827}
3828EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003829
3830#else
3831void *__kmalloc(size_t size, gfp_t flags)
3832{
3833 return __do_kmalloc(size, flags, NULL);
3834}
3835EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003836#endif
3837
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838/**
3839 * kmem_cache_free - Deallocate an object
3840 * @cachep: The cache the allocation was from.
3841 * @objp: The previously allocated object.
3842 *
3843 * Free an object which was previously allocated from this
3844 * cache.
3845 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003846void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847{
3848 unsigned long flags;
3849
3850 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003851 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003852 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3853 debug_check_no_obj_freed(objp, obj_size(cachep));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003854 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003856
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003857 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858}
3859EXPORT_SYMBOL(kmem_cache_free);
3860
3861/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862 * kfree - free previously allocated memory
3863 * @objp: pointer returned by kmalloc.
3864 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003865 * If @objp is NULL, no operation is performed.
3866 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 * Don't free memory not originally allocated by kmalloc()
3868 * or you will run into trouble.
3869 */
3870void kfree(const void *objp)
3871{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003872 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 unsigned long flags;
3874
Pekka Enberg2121db72009-03-25 11:05:57 +02003875 trace_kfree(_RET_IP_, objp);
3876
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003877 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 return;
3879 local_irq_save(flags);
3880 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003881 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003882 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003883 debug_check_no_obj_freed(objp, obj_size(c));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003884 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 local_irq_restore(flags);
3886}
3887EXPORT_SYMBOL(kfree);
3888
Pekka Enberg343e0d72006-02-01 03:05:50 -08003889unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003891 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892}
3893EXPORT_SYMBOL(kmem_cache_size);
3894
Christoph Lametere498be72005-09-09 13:03:32 -07003895/*
Simon Arlott183ff222007-10-20 01:27:18 +02003896 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003897 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003898static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003899{
3900 int node;
3901 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003902 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003903 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003904
Mel Gorman9c09a952008-01-24 05:49:54 -08003905 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003906
Paul Menage3395ee02006-12-06 20:32:16 -08003907 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003908 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003909 if (!new_alien)
3910 goto fail;
3911 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003912
Eric Dumazet63109842007-05-06 14:49:28 -07003913 new_shared = NULL;
3914 if (cachep->shared) {
3915 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003916 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003917 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003918 if (!new_shared) {
3919 free_alien_cache(new_alien);
3920 goto fail;
3921 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003922 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003923
Andrew Mortona737b3e2006-03-22 00:08:11 -08003924 l3 = cachep->nodelists[node];
3925 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003926 struct array_cache *shared = l3->shared;
3927
Christoph Lametere498be72005-09-09 13:03:32 -07003928 spin_lock_irq(&l3->list_lock);
3929
Christoph Lametercafeb022006-03-25 03:06:46 -08003930 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003931 free_block(cachep, shared->entry,
3932 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003933
Christoph Lametercafeb022006-03-25 03:06:46 -08003934 l3->shared = new_shared;
3935 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003936 l3->alien = new_alien;
3937 new_alien = NULL;
3938 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003939 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003940 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003941 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003942 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003943 free_alien_cache(new_alien);
3944 continue;
3945 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003946 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003947 if (!l3) {
3948 free_alien_cache(new_alien);
3949 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003950 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003951 }
Christoph Lametere498be72005-09-09 13:03:32 -07003952
3953 kmem_list3_init(l3);
3954 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003955 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003956 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003957 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003958 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003959 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003960 cachep->nodelists[node] = l3;
3961 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003962 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003963
Andrew Mortona737b3e2006-03-22 00:08:11 -08003964fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003965 if (!cachep->next.next) {
3966 /* Cache is not active yet. Roll back what we did */
3967 node--;
3968 while (node >= 0) {
3969 if (cachep->nodelists[node]) {
3970 l3 = cachep->nodelists[node];
3971
3972 kfree(l3->shared);
3973 free_alien_cache(l3->alien);
3974 kfree(l3);
3975 cachep->nodelists[node] = NULL;
3976 }
3977 node--;
3978 }
3979 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003980 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003981}
3982
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003984 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003985 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986};
3987
3988static void do_ccupdate_local(void *info)
3989{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003990 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 struct array_cache *old;
3992
3993 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003994 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003995
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3997 new->new[smp_processor_id()] = old;
3998}
3999
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004000/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08004001static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004002 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004004 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004005 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004007 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4008 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004009 if (!new)
4010 return -ENOMEM;
4011
Christoph Lametere498be72005-09-09 13:03:32 -07004012 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004013 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004014 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004015 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004016 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004017 kfree(new->new[i]);
4018 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004019 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020 }
4021 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004022 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004024 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004025
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027 cachep->batchcount = batchcount;
4028 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004029 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030
Christoph Lametere498be72005-09-09 13:03:32 -07004031 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004032 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 if (!ccold)
4034 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004035 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4036 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4037 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 kfree(ccold);
4039 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004040 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004041 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042}
4043
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004044/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004045static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046{
4047 int err;
4048 int limit, shared;
4049
Andrew Mortona737b3e2006-03-22 00:08:11 -08004050 /*
4051 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052 * - create a LIFO ordering, i.e. return objects that are cache-warm
4053 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004054 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 * bufctl chains: array operations are cheaper.
4056 * The numbers are guessed, we should auto-tune as described by
4057 * Bonwick.
4058 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004059 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004061 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004063 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004065 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066 limit = 54;
4067 else
4068 limit = 120;
4069
Andrew Mortona737b3e2006-03-22 00:08:11 -08004070 /*
4071 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072 * allocation behaviour: Most allocs on one cpu, most free operations
4073 * on another cpu. For these cases, an efficient object passing between
4074 * cpus is necessary. This is provided by a shared array. The array
4075 * replaces Bonwick's magazine layer.
4076 * On uniprocessor, it's functionally equivalent (but less efficient)
4077 * to a larger limit. Thus disabled by default.
4078 */
4079 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004080 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082
4083#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004084 /*
4085 * With debugging enabled, large batchcount lead to excessively long
4086 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 */
4088 if (limit > 32)
4089 limit = 32;
4090#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004091 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 if (err)
4093 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004094 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004095 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096}
4097
Christoph Lameter1b552532006-03-22 00:09:07 -08004098/*
4099 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004100 * necessary. Note that the l3 listlock also protects the array_cache
4101 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004102 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004103static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004104 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105{
4106 int tofree;
4107
Christoph Lameter1b552532006-03-22 00:09:07 -08004108 if (!ac || !ac->avail)
4109 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 if (ac->touched && !force) {
4111 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004112 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004113 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004114 if (ac->avail) {
4115 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4116 if (tofree > ac->avail)
4117 tofree = (ac->avail + 1) / 2;
4118 free_block(cachep, ac->entry, tofree, node);
4119 ac->avail -= tofree;
4120 memmove(ac->entry, &(ac->entry[tofree]),
4121 sizeof(void *) * ac->avail);
4122 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004123 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124 }
4125}
4126
4127/**
4128 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004129 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 *
4131 * Called from workqueue/eventd every few seconds.
4132 * Purpose:
4133 * - clear the per-cpu caches for this CPU.
4134 * - return freeable pages to the main free memory pool.
4135 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004136 * If we cannot acquire the cache chain mutex then just give up - we'll try
4137 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004139static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004141 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004142 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004143 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004144 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004146 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004148 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004150 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151 check_irq_on();
4152
Christoph Lameter35386e32006-03-22 00:09:05 -08004153 /*
4154 * We only take the l3 lock if absolutely necessary and we
4155 * have established with reasonable certainty that
4156 * we can do some work if the lock was obtained.
4157 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004158 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004159
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004160 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161
Christoph Lameteraab22072006-03-22 00:09:06 -08004162 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163
Christoph Lameter35386e32006-03-22 00:09:05 -08004164 /*
4165 * These are racy checks but it does not matter
4166 * if we skip one check or scan twice.
4167 */
Christoph Lametere498be72005-09-09 13:03:32 -07004168 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004169 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170
Christoph Lametere498be72005-09-09 13:03:32 -07004171 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172
Christoph Lameteraab22072006-03-22 00:09:06 -08004173 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174
Christoph Lametered11d9e2006-06-30 01:55:45 -07004175 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004176 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004177 else {
4178 int freed;
4179
4180 freed = drain_freelist(searchp, l3, (l3->free_limit +
4181 5 * searchp->num - 1) / (5 * searchp->num));
4182 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004184next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185 cond_resched();
4186 }
4187 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004188 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004189 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004190out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004191 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004192 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193}
4194
Linus Torvalds158a9622008-01-02 13:04:48 -08004195#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
Pekka Enberg85289f92006-01-08 01:00:36 -08004197static void print_slabinfo_header(struct seq_file *m)
4198{
4199 /*
4200 * Output format version, so at least we can change it
4201 * without _too_ many complaints.
4202 */
4203#if STATS
4204 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4205#else
4206 seq_puts(m, "slabinfo - version: 2.1\n");
4207#endif
4208 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4209 "<objperslab> <pagesperslab>");
4210 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4211 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4212#if STATS
4213 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004214 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004215 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4216#endif
4217 seq_putc(m, '\n');
4218}
4219
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220static void *s_start(struct seq_file *m, loff_t *pos)
4221{
4222 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004224 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004225 if (!n)
4226 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004227
4228 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229}
4230
4231static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4232{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004233 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234}
4235
4236static void s_stop(struct seq_file *m, void *p)
4237{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004238 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239}
4240
4241static int s_show(struct seq_file *m, void *p)
4242{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004243 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 struct slab *slabp;
4245 unsigned long active_objs;
4246 unsigned long num_objs;
4247 unsigned long active_slabs = 0;
4248 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004249 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004251 int node;
4252 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254 active_objs = 0;
4255 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004256 for_each_online_node(node) {
4257 l3 = cachep->nodelists[node];
4258 if (!l3)
4259 continue;
4260
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004261 check_irq_on();
4262 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004263
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004264 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004265 if (slabp->inuse != cachep->num && !error)
4266 error = "slabs_full accounting error";
4267 active_objs += cachep->num;
4268 active_slabs++;
4269 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004270 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004271 if (slabp->inuse == cachep->num && !error)
4272 error = "slabs_partial inuse accounting error";
4273 if (!slabp->inuse && !error)
4274 error = "slabs_partial/inuse accounting error";
4275 active_objs += slabp->inuse;
4276 active_slabs++;
4277 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004278 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004279 if (slabp->inuse && !error)
4280 error = "slabs_free/inuse accounting error";
4281 num_slabs++;
4282 }
4283 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004284 if (l3->shared)
4285 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004286
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004287 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004289 num_slabs += active_slabs;
4290 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004291 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292 error = "free_objects accounting error";
4293
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004294 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295 if (error)
4296 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4297
4298 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004299 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004300 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004302 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004303 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004304 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004305#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004306 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307 unsigned long high = cachep->high_mark;
4308 unsigned long allocs = cachep->num_allocations;
4309 unsigned long grown = cachep->grown;
4310 unsigned long reaped = cachep->reaped;
4311 unsigned long errors = cachep->errors;
4312 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004314 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004315 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316
Joe Perchese92dd4f2010-03-26 19:27:58 -07004317 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4318 "%4lu %4lu %4lu %4lu %4lu",
4319 allocs, high, grown,
4320 reaped, errors, max_freeable, node_allocs,
4321 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 }
4323 /* cpu stats */
4324 {
4325 unsigned long allochit = atomic_read(&cachep->allochit);
4326 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4327 unsigned long freehit = atomic_read(&cachep->freehit);
4328 unsigned long freemiss = atomic_read(&cachep->freemiss);
4329
4330 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004331 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332 }
4333#endif
4334 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335 return 0;
4336}
4337
4338/*
4339 * slabinfo_op - iterator that generates /proc/slabinfo
4340 *
4341 * Output layout:
4342 * cache-name
4343 * num-active-objs
4344 * total-objs
4345 * object size
4346 * num-active-slabs
4347 * total-slabs
4348 * num-pages-per-slab
4349 * + further values on SMP and with statistics enabled
4350 */
4351
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004352static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004353 .start = s_start,
4354 .next = s_next,
4355 .stop = s_stop,
4356 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357};
4358
4359#define MAX_SLABINFO_WRITE 128
4360/**
4361 * slabinfo_write - Tuning for the slab allocator
4362 * @file: unused
4363 * @buffer: user buffer
4364 * @count: data length
4365 * @ppos: unused
4366 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004367static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004368 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004369{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004370 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004371 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004372 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004373
Linus Torvalds1da177e2005-04-16 15:20:36 -07004374 if (count > MAX_SLABINFO_WRITE)
4375 return -EINVAL;
4376 if (copy_from_user(&kbuf, buffer, count))
4377 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004378 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004379
4380 tmp = strchr(kbuf, ' ');
4381 if (!tmp)
4382 return -EINVAL;
4383 *tmp = '\0';
4384 tmp++;
4385 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4386 return -EINVAL;
4387
4388 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004389 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004390 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004391 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004392 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004393 if (limit < 1 || batchcount < 1 ||
4394 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004395 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004396 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004397 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004398 batchcount, shared,
4399 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004400 }
4401 break;
4402 }
4403 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004404 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004405 if (res >= 0)
4406 res = count;
4407 return res;
4408}
Al Viro871751e2006-03-25 03:06:39 -08004409
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004410static int slabinfo_open(struct inode *inode, struct file *file)
4411{
4412 return seq_open(file, &slabinfo_op);
4413}
4414
4415static const struct file_operations proc_slabinfo_operations = {
4416 .open = slabinfo_open,
4417 .read = seq_read,
4418 .write = slabinfo_write,
4419 .llseek = seq_lseek,
4420 .release = seq_release,
4421};
4422
Al Viro871751e2006-03-25 03:06:39 -08004423#ifdef CONFIG_DEBUG_SLAB_LEAK
4424
4425static void *leaks_start(struct seq_file *m, loff_t *pos)
4426{
Al Viro871751e2006-03-25 03:06:39 -08004427 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004428 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004429}
4430
4431static inline int add_caller(unsigned long *n, unsigned long v)
4432{
4433 unsigned long *p;
4434 int l;
4435 if (!v)
4436 return 1;
4437 l = n[1];
4438 p = n + 2;
4439 while (l) {
4440 int i = l/2;
4441 unsigned long *q = p + 2 * i;
4442 if (*q == v) {
4443 q[1]++;
4444 return 1;
4445 }
4446 if (*q > v) {
4447 l = i;
4448 } else {
4449 p = q + 2;
4450 l -= i + 1;
4451 }
4452 }
4453 if (++n[1] == n[0])
4454 return 0;
4455 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4456 p[0] = v;
4457 p[1] = 1;
4458 return 1;
4459}
4460
4461static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4462{
4463 void *p;
4464 int i;
4465 if (n[0] == n[1])
4466 return;
4467 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4468 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4469 continue;
4470 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4471 return;
4472 }
4473}
4474
4475static void show_symbol(struct seq_file *m, unsigned long address)
4476{
4477#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004478 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004479 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004480
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004481 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004482 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004483 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004484 seq_printf(m, " [%s]", modname);
4485 return;
4486 }
4487#endif
4488 seq_printf(m, "%p", (void *)address);
4489}
4490
4491static int leaks_show(struct seq_file *m, void *p)
4492{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004493 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004494 struct slab *slabp;
4495 struct kmem_list3 *l3;
4496 const char *name;
4497 unsigned long *n = m->private;
4498 int node;
4499 int i;
4500
4501 if (!(cachep->flags & SLAB_STORE_USER))
4502 return 0;
4503 if (!(cachep->flags & SLAB_RED_ZONE))
4504 return 0;
4505
4506 /* OK, we can do it */
4507
4508 n[1] = 0;
4509
4510 for_each_online_node(node) {
4511 l3 = cachep->nodelists[node];
4512 if (!l3)
4513 continue;
4514
4515 check_irq_on();
4516 spin_lock_irq(&l3->list_lock);
4517
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004518 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004519 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004520 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004521 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004522 spin_unlock_irq(&l3->list_lock);
4523 }
4524 name = cachep->name;
4525 if (n[0] == n[1]) {
4526 /* Increase the buffer size */
4527 mutex_unlock(&cache_chain_mutex);
4528 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4529 if (!m->private) {
4530 /* Too bad, we are really out */
4531 m->private = n;
4532 mutex_lock(&cache_chain_mutex);
4533 return -ENOMEM;
4534 }
4535 *(unsigned long *)m->private = n[0] * 2;
4536 kfree(n);
4537 mutex_lock(&cache_chain_mutex);
4538 /* Now make sure this entry will be retried */
4539 m->count = m->size;
4540 return 0;
4541 }
4542 for (i = 0; i < n[1]; i++) {
4543 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4544 show_symbol(m, n[2*i+2]);
4545 seq_putc(m, '\n');
4546 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004547
Al Viro871751e2006-03-25 03:06:39 -08004548 return 0;
4549}
4550
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004551static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004552 .start = leaks_start,
4553 .next = s_next,
4554 .stop = s_stop,
4555 .show = leaks_show,
4556};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004557
4558static int slabstats_open(struct inode *inode, struct file *file)
4559{
4560 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4561 int ret = -ENOMEM;
4562 if (n) {
4563 ret = seq_open(file, &slabstats_op);
4564 if (!ret) {
4565 struct seq_file *m = file->private_data;
4566 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4567 m->private = n;
4568 n = NULL;
4569 }
4570 kfree(n);
4571 }
4572 return ret;
4573}
4574
4575static const struct file_operations proc_slabstats_operations = {
4576 .open = slabstats_open,
4577 .read = seq_read,
4578 .llseek = seq_lseek,
4579 .release = seq_release_private,
4580};
Al Viro871751e2006-03-25 03:06:39 -08004581#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004582
4583static int __init slab_proc_init(void)
4584{
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004585 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004586#ifdef CONFIG_DEBUG_SLAB_LEAK
4587 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4588#endif
4589 return 0;
4590}
4591module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004592#endif
4593
Manfred Spraul00e145b2005-09-03 15:55:07 -07004594/**
4595 * ksize - get the actual amount of memory allocated for a given object
4596 * @objp: Pointer to the object
4597 *
4598 * kmalloc may internally round up allocations and return more memory
4599 * than requested. ksize() can be used to determine the actual amount of
4600 * memory allocated. The caller may use this additional memory, even though
4601 * a smaller amount of memory was initially specified with the kmalloc call.
4602 * The caller must guarantee that objp points to a valid object previously
4603 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4604 * must not be freed during the duration of the call.
4605 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004606size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004607{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004608 BUG_ON(!objp);
4609 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004610 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004611
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004612 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004613}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004614EXPORT_SYMBOL(ksize);