blob: 1e523ed47c619d385465a7080f44fabe68d115cb [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,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 FULL
599} g_cpucache_up;
600
Mike Kravetz39d24e62006-05-15 09:44:13 -0700601/*
602 * used by boot code to determine if it can use slab based allocator
603 */
604int slab_is_available(void)
605{
Pekka Enberg8429db52009-06-12 15:58:59 +0300606 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700607}
608
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200609#ifdef CONFIG_LOCKDEP
610
611/*
612 * Slab sometimes uses the kmalloc slabs to store the slab headers
613 * for other slabs "off slab".
614 * The locking for this is tricky in that it nests within the locks
615 * of all other slabs in a few places; to deal with this special
616 * locking we put on-slab caches into a separate lock-class.
617 *
618 * We set lock class for alien array caches which are up during init.
619 * The lock annotation will be lost if all cpus of a node goes down and
620 * then comes back up during hotplug
621 */
622static struct lock_class_key on_slab_l3_key;
623static struct lock_class_key on_slab_alc_key;
624
625static void init_node_lock_keys(int q)
626{
627 struct cache_sizes *s = malloc_sizes;
628
629 if (g_cpucache_up != FULL)
630 return;
631
632 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
633 struct array_cache **alc;
634 struct kmem_list3 *l3;
635 int r;
636
637 l3 = s->cs_cachep->nodelists[q];
638 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200639 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200640 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
641 alc = l3->alien;
642 /*
643 * FIXME: This check for BAD_ALIEN_MAGIC
644 * should go away when common slab code is taught to
645 * work even without alien caches.
646 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
647 * for alloc_alien_cache,
648 */
649 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
Pekka Enberg00afa752009-12-27 14:33:14 +0200650 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200651 for_each_node(r) {
652 if (alc[r])
653 lockdep_set_class(&alc[r]->lock,
654 &on_slab_alc_key);
655 }
656 }
657}
658
659static inline void init_lock_keys(void)
660{
661 int node;
662
663 for_each_node(node)
664 init_node_lock_keys(node);
665}
666#else
667static void init_node_lock_keys(int q)
668{
669}
670
671static inline void init_lock_keys(void)
672{
673}
674#endif
675
676/*
677 * Guard access to the cache-chain.
678 */
679static DEFINE_MUTEX(cache_chain_mutex);
680static struct list_head cache_chain;
681
Tejun Heo1871e522009-10-29 22:34:13 +0900682static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
Pekka Enberg343e0d72006-02-01 03:05:50 -0800684static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685{
686 return cachep->array[smp_processor_id()];
687}
688
Andrew Mortona737b3e2006-03-22 00:08:11 -0800689static inline struct kmem_cache *__find_general_cachep(size_t size,
690 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 struct cache_sizes *csizep = malloc_sizes;
693
694#if DEBUG
695 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800696 * kmem_cache_create(), or __kmalloc(), before
697 * the generic caches are initialized.
698 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700699 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700701 if (!size)
702 return ZERO_SIZE_PTR;
703
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 while (size > csizep->cs_size)
705 csizep++;
706
707 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700708 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 * has cs_{dma,}cachep==NULL. Thus no special case
710 * for large kmalloc calls required.
711 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800712#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 if (unlikely(gfpflags & GFP_DMA))
714 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800715#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 return csizep->cs_cachep;
717}
718
Adrian Bunkb2213852006-09-25 23:31:02 -0700719static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700720{
721 return __find_general_cachep(size, gfpflags);
722}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700723
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800724static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800726 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
727}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
Andrew Mortona737b3e2006-03-22 00:08:11 -0800729/*
730 * Calculate the number of objects and left-over bytes for a given buffer size.
731 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800732static void cache_estimate(unsigned long gfporder, size_t buffer_size,
733 size_t align, int flags, size_t *left_over,
734 unsigned int *num)
735{
736 int nr_objs;
737 size_t mgmt_size;
738 size_t slab_size = PAGE_SIZE << gfporder;
739
740 /*
741 * The slab management structure can be either off the slab or
742 * on it. For the latter case, the memory allocated for a
743 * slab is used for:
744 *
745 * - The struct slab
746 * - One kmem_bufctl_t for each object
747 * - Padding to respect alignment of @align
748 * - @buffer_size bytes for each object
749 *
750 * If the slab management structure is off the slab, then the
751 * alignment will already be calculated into the size. Because
752 * the slabs are all pages aligned, the objects will be at the
753 * correct alignment when allocated.
754 */
755 if (flags & CFLGS_OFF_SLAB) {
756 mgmt_size = 0;
757 nr_objs = slab_size / buffer_size;
758
759 if (nr_objs > SLAB_LIMIT)
760 nr_objs = SLAB_LIMIT;
761 } else {
762 /*
763 * Ignore padding for the initial guess. The padding
764 * is at most @align-1 bytes, and @buffer_size is at
765 * least @align. In the worst case, this result will
766 * be one greater than the number of objects that fit
767 * into the memory allocation when taking the padding
768 * into account.
769 */
770 nr_objs = (slab_size - sizeof(struct slab)) /
771 (buffer_size + sizeof(kmem_bufctl_t));
772
773 /*
774 * This calculated number will be either the right
775 * amount, or one greater than what we want.
776 */
777 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
778 > slab_size)
779 nr_objs--;
780
781 if (nr_objs > SLAB_LIMIT)
782 nr_objs = SLAB_LIMIT;
783
784 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800786 *num = nr_objs;
787 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788}
789
Harvey Harrisond40cee22008-04-30 00:55:07 -0700790#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
Andrew Mortona737b3e2006-03-22 00:08:11 -0800792static void __slab_error(const char *function, struct kmem_cache *cachep,
793 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794{
795 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800796 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 dump_stack();
798}
799
Paul Menage3395ee02006-12-06 20:32:16 -0800800/*
801 * By default on NUMA we use alien caches to stage the freeing of
802 * objects allocated from other nodes. This causes massive memory
803 * inefficiencies when using fake NUMA setup to split memory into a
804 * large number of small nodes, so it can be disabled on the command
805 * line
806 */
807
808static int use_alien_caches __read_mostly = 1;
809static int __init noaliencache_setup(char *s)
810{
811 use_alien_caches = 0;
812 return 1;
813}
814__setup("noaliencache", noaliencache_setup);
815
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800816#ifdef CONFIG_NUMA
817/*
818 * Special reaping functions for NUMA systems called from cache_reap().
819 * These take care of doing round robin flushing of alien caches (containing
820 * objects freed on different nodes from which they were allocated) and the
821 * flushing of remote pcps by calling drain_node_pages.
822 */
Tejun Heo1871e522009-10-29 22:34:13 +0900823static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800824
825static void init_reap_node(int cpu)
826{
827 int node;
828
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700829 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800830 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800831 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800832
Tejun Heo1871e522009-10-29 22:34:13 +0900833 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800834}
835
836static void next_reap_node(void)
837{
Christoph Lameter909ea962010-12-08 16:22:55 +0100838 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800839
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800840 node = next_node(node, node_online_map);
841 if (unlikely(node >= MAX_NUMNODES))
842 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100843 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800844}
845
846#else
847#define init_reap_node(cpu) do { } while (0)
848#define next_reap_node(void) do { } while (0)
849#endif
850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851/*
852 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
853 * via the workqueue/eventd.
854 * Add the CPU number into the expiration time to minimize the possibility of
855 * the CPUs getting into lockstep and contending for the global cache chain
856 * lock.
857 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700858static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859{
Tejun Heo1871e522009-10-29 22:34:13 +0900860 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 /*
863 * When this gets called from do_initcalls via cpucache_init(),
864 * init_workqueues() has already run, so keventd will be setup
865 * at that time.
866 */
David Howells52bad642006-11-22 14:54:01 +0000867 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800868 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700869 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800870 schedule_delayed_work_on(cpu, reap_work,
871 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 }
873}
874
Christoph Lametere498be72005-09-09 13:03:32 -0700875static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300876 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800878 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 struct array_cache *nc = NULL;
880
Pekka Enberg83b519e2009-06-10 19:40:04 +0300881 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100882 /*
883 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300884 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100885 * cache the pointers are not cleared and they could be counted as
886 * valid references during a kmemleak scan. Therefore, kmemleak must
887 * not scan such objects.
888 */
889 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 if (nc) {
891 nc->avail = 0;
892 nc->limit = entries;
893 nc->batchcount = batchcount;
894 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700895 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 }
897 return nc;
898}
899
Christoph Lameter3ded1752006-03-25 03:06:44 -0800900/*
901 * Transfer objects in one arraycache to another.
902 * Locking must be handled by the caller.
903 *
904 * Return the number of entries transferred.
905 */
906static int transfer_objects(struct array_cache *to,
907 struct array_cache *from, unsigned int max)
908{
909 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700910 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800911
912 if (!nr)
913 return 0;
914
915 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
916 sizeof(void *) *nr);
917
918 from->avail -= nr;
919 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800920 return nr;
921}
922
Christoph Lameter765c4502006-09-27 01:50:08 -0700923#ifndef CONFIG_NUMA
924
925#define drain_alien_cache(cachep, alien) do { } while (0)
926#define reap_alien(cachep, l3) do { } while (0)
927
Pekka Enberg83b519e2009-06-10 19:40:04 +0300928static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700929{
930 return (struct array_cache **)BAD_ALIEN_MAGIC;
931}
932
933static inline void free_alien_cache(struct array_cache **ac_ptr)
934{
935}
936
937static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
938{
939 return 0;
940}
941
942static inline void *alternate_node_alloc(struct kmem_cache *cachep,
943 gfp_t flags)
944{
945 return NULL;
946}
947
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800948static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700949 gfp_t flags, int nodeid)
950{
951 return NULL;
952}
953
954#else /* CONFIG_NUMA */
955
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800956static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800957static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800958
Pekka Enberg83b519e2009-06-10 19:40:04 +0300959static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700960{
961 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800962 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700963 int i;
964
965 if (limit > 1)
966 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800967 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700968 if (ac_ptr) {
969 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800970 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700971 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300972 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700973 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800974 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700975 kfree(ac_ptr[i]);
976 kfree(ac_ptr);
977 return NULL;
978 }
979 }
980 }
981 return ac_ptr;
982}
983
Pekka Enberg5295a742006-02-01 03:05:48 -0800984static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700985{
986 int i;
987
988 if (!ac_ptr)
989 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700990 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800991 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700992 kfree(ac_ptr);
993}
994
Pekka Enberg343e0d72006-02-01 03:05:50 -0800995static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800996 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700997{
998 struct kmem_list3 *rl3 = cachep->nodelists[node];
999
1000 if (ac->avail) {
1001 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001002 /*
1003 * Stuff objects into the remote nodes shared array first.
1004 * That way we could avoid the overhead of putting the objects
1005 * into the free lists and getting them back later.
1006 */
shin, jacob693f7d32006-04-28 10:54:37 -05001007 if (rl3->shared)
1008 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001009
Christoph Lameterff694162005-09-22 21:44:02 -07001010 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001011 ac->avail = 0;
1012 spin_unlock(&rl3->list_lock);
1013 }
1014}
1015
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001016/*
1017 * Called from cache_reap() to regularly drain alien caches round robin.
1018 */
1019static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1020{
Christoph Lameter909ea962010-12-08 16:22:55 +01001021 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001022
1023 if (l3->alien) {
1024 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001025
1026 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001027 __drain_alien_cache(cachep, ac, node);
1028 spin_unlock_irq(&ac->lock);
1029 }
1030 }
1031}
1032
Andrew Mortona737b3e2006-03-22 00:08:11 -08001033static void drain_alien_cache(struct kmem_cache *cachep,
1034 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001035{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001036 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001037 struct array_cache *ac;
1038 unsigned long flags;
1039
1040 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001041 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001042 if (ac) {
1043 spin_lock_irqsave(&ac->lock, flags);
1044 __drain_alien_cache(cachep, ac, i);
1045 spin_unlock_irqrestore(&ac->lock, flags);
1046 }
1047 }
1048}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001049
Ingo Molnar873623d2006-07-13 14:44:38 +02001050static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001051{
1052 struct slab *slabp = virt_to_slab(objp);
1053 int nodeid = slabp->nodeid;
1054 struct kmem_list3 *l3;
1055 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001056 int node;
1057
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001058 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001059
1060 /*
1061 * Make sure we are not freeing a object from another node to the array
1062 * cache on this cpu.
1063 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001064 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001065 return 0;
1066
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001067 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001068 STATS_INC_NODEFREES(cachep);
1069 if (l3->alien && l3->alien[nodeid]) {
1070 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001071 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001072 if (unlikely(alien->avail == alien->limit)) {
1073 STATS_INC_ACOVERFLOW(cachep);
1074 __drain_alien_cache(cachep, alien, nodeid);
1075 }
1076 alien->entry[alien->avail++] = objp;
1077 spin_unlock(&alien->lock);
1078 } else {
1079 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1080 free_block(cachep, &objp, 1, nodeid);
1081 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1082 }
1083 return 1;
1084}
Christoph Lametere498be72005-09-09 13:03:32 -07001085#endif
1086
David Rientjes8f9f8d92010-03-27 19:40:47 -07001087/*
1088 * Allocates and initializes nodelists for a node on each slab cache, used for
1089 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1090 * will be allocated off-node since memory is not yet online for the new node.
1091 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1092 * already in use.
1093 *
1094 * Must hold cache_chain_mutex.
1095 */
1096static int init_cache_nodelists_node(int node)
1097{
1098 struct kmem_cache *cachep;
1099 struct kmem_list3 *l3;
1100 const int memsize = sizeof(struct kmem_list3);
1101
1102 list_for_each_entry(cachep, &cache_chain, next) {
1103 /*
1104 * Set up the size64 kmemlist for cpu before we can
1105 * begin anything. Make sure some other cpu on this
1106 * node has not already allocated this
1107 */
1108 if (!cachep->nodelists[node]) {
1109 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1110 if (!l3)
1111 return -ENOMEM;
1112 kmem_list3_init(l3);
1113 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1114 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1115
1116 /*
1117 * The l3s don't come and go as CPUs come and
1118 * go. cache_chain_mutex is sufficient
1119 * protection here.
1120 */
1121 cachep->nodelists[node] = l3;
1122 }
1123
1124 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1125 cachep->nodelists[node]->free_limit =
1126 (1 + nr_cpus_node(node)) *
1127 cachep->batchcount + cachep->num;
1128 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1129 }
1130 return 0;
1131}
1132
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001133static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001135 struct kmem_cache *cachep;
1136 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001137 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301138 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001139
1140 list_for_each_entry(cachep, &cache_chain, next) {
1141 struct array_cache *nc;
1142 struct array_cache *shared;
1143 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001144
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001145 /* cpu is dead; no one can alloc from it. */
1146 nc = cachep->array[cpu];
1147 cachep->array[cpu] = NULL;
1148 l3 = cachep->nodelists[node];
1149
1150 if (!l3)
1151 goto free_array_cache;
1152
1153 spin_lock_irq(&l3->list_lock);
1154
1155 /* Free limit for this kmem_list3 */
1156 l3->free_limit -= cachep->batchcount;
1157 if (nc)
1158 free_block(cachep, nc->entry, nc->avail, node);
1159
Rusty Russell58463c12009-12-17 11:43:12 -06001160 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001161 spin_unlock_irq(&l3->list_lock);
1162 goto free_array_cache;
1163 }
1164
1165 shared = l3->shared;
1166 if (shared) {
1167 free_block(cachep, shared->entry,
1168 shared->avail, node);
1169 l3->shared = NULL;
1170 }
1171
1172 alien = l3->alien;
1173 l3->alien = NULL;
1174
1175 spin_unlock_irq(&l3->list_lock);
1176
1177 kfree(shared);
1178 if (alien) {
1179 drain_alien_cache(cachep, alien);
1180 free_alien_cache(alien);
1181 }
1182free_array_cache:
1183 kfree(nc);
1184 }
1185 /*
1186 * In the previous loop, all the objects were freed to
1187 * the respective cache's slabs, now we can go ahead and
1188 * shrink each nodelist to its limit.
1189 */
1190 list_for_each_entry(cachep, &cache_chain, next) {
1191 l3 = cachep->nodelists[node];
1192 if (!l3)
1193 continue;
1194 drain_freelist(cachep, l3, l3->free_objects);
1195 }
1196}
1197
1198static int __cpuinit cpuup_prepare(long cpu)
1199{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001200 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001201 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001202 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001203 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001205 /*
1206 * We need to do this right in the beginning since
1207 * alloc_arraycache's are going to use this list.
1208 * kmalloc_node allows us to add the slab to the right
1209 * kmem_list3 and not this cpu's kmem_list3
1210 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001211 err = init_cache_nodelists_node(node);
1212 if (err < 0)
1213 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001214
1215 /*
1216 * Now we can go ahead with allocating the shared arrays and
1217 * array caches
1218 */
1219 list_for_each_entry(cachep, &cache_chain, next) {
1220 struct array_cache *nc;
1221 struct array_cache *shared = NULL;
1222 struct array_cache **alien = NULL;
1223
1224 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001225 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001226 if (!nc)
1227 goto bad;
1228 if (cachep->shared) {
1229 shared = alloc_arraycache(node,
1230 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001231 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001232 if (!shared) {
1233 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001234 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001235 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001236 }
1237 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001238 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001239 if (!alien) {
1240 kfree(shared);
1241 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001242 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001243 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001244 }
1245 cachep->array[cpu] = nc;
1246 l3 = cachep->nodelists[node];
1247 BUG_ON(!l3);
1248
1249 spin_lock_irq(&l3->list_lock);
1250 if (!l3->shared) {
1251 /*
1252 * We are serialised from CPU_DEAD or
1253 * CPU_UP_CANCELLED by the cpucontrol lock
1254 */
1255 l3->shared = shared;
1256 shared = NULL;
1257 }
1258#ifdef CONFIG_NUMA
1259 if (!l3->alien) {
1260 l3->alien = alien;
1261 alien = NULL;
1262 }
1263#endif
1264 spin_unlock_irq(&l3->list_lock);
1265 kfree(shared);
1266 free_alien_cache(alien);
1267 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001268 init_node_lock_keys(node);
1269
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001270 return 0;
1271bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001272 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001273 return -ENOMEM;
1274}
1275
1276static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1277 unsigned long action, void *hcpu)
1278{
1279 long cpu = (long)hcpu;
1280 int err = 0;
1281
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001283 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001284 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001285 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001286 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001287 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 break;
1289 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001290 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 start_cpu_timer(cpu);
1292 break;
1293#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001294 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001295 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001296 /*
1297 * Shutdown cache reaper. Note that the cache_chain_mutex is
1298 * held so that if cache_reap() is invoked it cannot do
1299 * anything expensive but will only modify reap_work
1300 * and reschedule the timer.
1301 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001302 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001303 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001304 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001305 break;
1306 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001307 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001308 start_cpu_timer(cpu);
1309 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001311 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001312 /*
1313 * Even if all the cpus of a node are down, we don't free the
1314 * kmem_list3 of any cache. This to avoid a race between
1315 * cpu_down, and a kmalloc allocation from another cpu for
1316 * memory from the node of the cpu going down. The list3
1317 * structure is usually allocated from kmem_cache_create() and
1318 * gets destroyed at kmem_cache_destroy().
1319 */
Simon Arlott183ff222007-10-20 01:27:18 +02001320 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001321#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001323 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001324 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001325 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001326 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001329 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330}
1331
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001332static struct notifier_block __cpuinitdata cpucache_notifier = {
1333 &cpuup_callback, NULL, 0
1334};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335
David Rientjes8f9f8d92010-03-27 19:40:47 -07001336#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1337/*
1338 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1339 * Returns -EBUSY if all objects cannot be drained so that the node is not
1340 * removed.
1341 *
1342 * Must hold cache_chain_mutex.
1343 */
1344static int __meminit drain_cache_nodelists_node(int node)
1345{
1346 struct kmem_cache *cachep;
1347 int ret = 0;
1348
1349 list_for_each_entry(cachep, &cache_chain, next) {
1350 struct kmem_list3 *l3;
1351
1352 l3 = cachep->nodelists[node];
1353 if (!l3)
1354 continue;
1355
1356 drain_freelist(cachep, l3, l3->free_objects);
1357
1358 if (!list_empty(&l3->slabs_full) ||
1359 !list_empty(&l3->slabs_partial)) {
1360 ret = -EBUSY;
1361 break;
1362 }
1363 }
1364 return ret;
1365}
1366
1367static int __meminit slab_memory_callback(struct notifier_block *self,
1368 unsigned long action, void *arg)
1369{
1370 struct memory_notify *mnb = arg;
1371 int ret = 0;
1372 int nid;
1373
1374 nid = mnb->status_change_nid;
1375 if (nid < 0)
1376 goto out;
1377
1378 switch (action) {
1379 case MEM_GOING_ONLINE:
1380 mutex_lock(&cache_chain_mutex);
1381 ret = init_cache_nodelists_node(nid);
1382 mutex_unlock(&cache_chain_mutex);
1383 break;
1384 case MEM_GOING_OFFLINE:
1385 mutex_lock(&cache_chain_mutex);
1386 ret = drain_cache_nodelists_node(nid);
1387 mutex_unlock(&cache_chain_mutex);
1388 break;
1389 case MEM_ONLINE:
1390 case MEM_OFFLINE:
1391 case MEM_CANCEL_ONLINE:
1392 case MEM_CANCEL_OFFLINE:
1393 break;
1394 }
1395out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001396 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001397}
1398#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1399
Christoph Lametere498be72005-09-09 13:03:32 -07001400/*
1401 * swap the static kmem_list3 with kmalloced memory
1402 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001403static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1404 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001405{
1406 struct kmem_list3 *ptr;
1407
Pekka Enberg83b519e2009-06-10 19:40:04 +03001408 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001409 BUG_ON(!ptr);
1410
Christoph Lametere498be72005-09-09 13:03:32 -07001411 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001412 /*
1413 * Do not assume that spinlocks can be initialized via memcpy:
1414 */
1415 spin_lock_init(&ptr->list_lock);
1416
Christoph Lametere498be72005-09-09 13:03:32 -07001417 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1418 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001419}
1420
Andrew Mortona737b3e2006-03-22 00:08:11 -08001421/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001422 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1423 * size of kmem_list3.
1424 */
1425static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1426{
1427 int node;
1428
1429 for_each_online_node(node) {
1430 cachep->nodelists[node] = &initkmem_list3[index + node];
1431 cachep->nodelists[node]->next_reap = jiffies +
1432 REAPTIMEOUT_LIST3 +
1433 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1434 }
1435}
1436
1437/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001438 * Initialisation. Called after the page allocator have been initialised and
1439 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 */
1441void __init kmem_cache_init(void)
1442{
1443 size_t left_over;
1444 struct cache_sizes *sizes;
1445 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001446 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001447 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001448 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001449
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001450 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001451 use_alien_caches = 0;
1452
Christoph Lametere498be72005-09-09 13:03:32 -07001453 for (i = 0; i < NUM_INIT_LISTS; i++) {
1454 kmem_list3_init(&initkmem_list3[i]);
1455 if (i < MAX_NUMNODES)
1456 cache_cache.nodelists[i] = NULL;
1457 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001458 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459
1460 /*
1461 * Fragmentation resistance on low memory - only use bigger
1462 * page orders on machines with more than 32MB of memory.
1463 */
Jan Beulich44813742009-09-21 17:03:05 -07001464 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 /* Bootstrap is tricky, because several objects are allocated
1468 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001469 * 1) initialize the cache_cache cache: it contains the struct
1470 * kmem_cache structures of all caches, except cache_cache itself:
1471 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001472 * Initially an __init data area is used for the head array and the
1473 * kmem_list3 structures, it's replaced with a kmalloc allocated
1474 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001476 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001477 * An __init data area is used for the head array.
1478 * 3) Create the remaining kmalloc caches, with minimally sized
1479 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 * 4) Replace the __init data head arrays for cache_cache and the first
1481 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001482 * 5) Replace the __init data for kmem_list3 for cache_cache and
1483 * the other cache's with kmalloc allocated memory.
1484 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 */
1486
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001487 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001488
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 INIT_LIST_HEAD(&cache_chain);
1491 list_add(&cache_cache.next, &cache_chain);
1492 cache_cache.colour_off = cache_line_size();
1493 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001494 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
Eric Dumazet8da34302007-05-06 14:49:29 -07001496 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001497 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001498 */
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001499 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1500 nr_node_ids * sizeof(struct kmem_list3 *);
Eric Dumazet8da34302007-05-06 14:49:29 -07001501#if DEBUG
1502 cache_cache.obj_size = cache_cache.buffer_size;
1503#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001504 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1505 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001506 cache_cache.reciprocal_buffer_size =
1507 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Jack Steiner07ed76b2006-03-07 21:55:46 -08001509 for (order = 0; order < MAX_ORDER; order++) {
1510 cache_estimate(order, cache_cache.buffer_size,
1511 cache_line_size(), 0, &left_over, &cache_cache.num);
1512 if (cache_cache.num)
1513 break;
1514 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001515 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001516 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001517 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001518 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1519 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520
1521 /* 2+3) create the kmalloc caches */
1522 sizes = malloc_sizes;
1523 names = cache_names;
1524
Andrew Mortona737b3e2006-03-22 00:08:11 -08001525 /*
1526 * Initialize the caches that provide memory for the array cache and the
1527 * kmem_list3 structures first. Without this, further allocations will
1528 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001529 */
1530
1531 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001532 sizes[INDEX_AC].cs_size,
1533 ARCH_KMALLOC_MINALIGN,
1534 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001535 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001536
Andrew Mortona737b3e2006-03-22 00:08:11 -08001537 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001538 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001539 kmem_cache_create(names[INDEX_L3].name,
1540 sizes[INDEX_L3].cs_size,
1541 ARCH_KMALLOC_MINALIGN,
1542 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001543 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001544 }
Christoph Lametere498be72005-09-09 13:03:32 -07001545
Ingo Molnare0a42722006-06-23 02:03:46 -07001546 slab_early_init = 0;
1547
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001549 /*
1550 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 * This should be particularly beneficial on SMP boxes, as it
1552 * eliminates "false sharing".
1553 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001554 * allow tighter packing of the smaller caches.
1555 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001556 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001557 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001558 sizes->cs_size,
1559 ARCH_KMALLOC_MINALIGN,
1560 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001561 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001562 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001563#ifdef CONFIG_ZONE_DMA
1564 sizes->cs_dmacachep = kmem_cache_create(
1565 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001566 sizes->cs_size,
1567 ARCH_KMALLOC_MINALIGN,
1568 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1569 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001570 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001571#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 sizes++;
1573 names++;
1574 }
1575 /* 4) Replace the bootstrap head arrays */
1576 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001577 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001578
Pekka Enberg83b519e2009-06-10 19:40:04 +03001579 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001580
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001581 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1582 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001583 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001584 /*
1585 * Do not assume that spinlocks can be initialized via memcpy:
1586 */
1587 spin_lock_init(&ptr->lock);
1588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001590
Pekka Enberg83b519e2009-06-10 19:40:04 +03001591 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001592
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001593 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001594 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001595 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001597 /*
1598 * Do not assume that spinlocks can be initialized via memcpy:
1599 */
1600 spin_lock_init(&ptr->lock);
1601
Christoph Lametere498be72005-09-09 13:03:32 -07001602 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001603 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 }
Christoph Lametere498be72005-09-09 13:03:32 -07001605 /* 5) Replace the bootstrap kmem_list3's */
1606 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001607 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608
Mel Gorman9c09a952008-01-24 05:49:54 -08001609 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001610 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001611
Christoph Lametere498be72005-09-09 13:03:32 -07001612 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001613 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001614
1615 if (INDEX_AC != INDEX_L3) {
1616 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001617 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001618 }
1619 }
1620 }
1621
Pekka Enberg8429db52009-06-12 15:58:59 +03001622 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001623}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001624
Pekka Enberg8429db52009-06-12 15:58:59 +03001625void __init kmem_cache_init_late(void)
1626{
1627 struct kmem_cache *cachep;
1628
Pekka Enberg8429db52009-06-12 15:58:59 +03001629 /* 6) resize the head arrays to their final sizes */
1630 mutex_lock(&cache_chain_mutex);
1631 list_for_each_entry(cachep, &cache_chain, next)
1632 if (enable_cpucache(cachep, GFP_NOWAIT))
1633 BUG();
1634 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001635
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 /* Done! */
1637 g_cpucache_up = FULL;
1638
Pekka Enbergec5a36f2009-06-29 09:57:10 +03001639 /* Annotate slab for lockdep -- annotate the malloc caches */
1640 init_lock_keys();
1641
Andrew Mortona737b3e2006-03-22 00:08:11 -08001642 /*
1643 * Register a cpu startup notifier callback that initializes
1644 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 */
1646 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
David Rientjes8f9f8d92010-03-27 19:40:47 -07001648#ifdef CONFIG_NUMA
1649 /*
1650 * Register a memory hotplug callback that initializes and frees
1651 * nodelists.
1652 */
1653 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1654#endif
1655
Andrew Mortona737b3e2006-03-22 00:08:11 -08001656 /*
1657 * The reap timers are started later, with a module init call: That part
1658 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 */
1660}
1661
1662static int __init cpucache_init(void)
1663{
1664 int cpu;
1665
Andrew Mortona737b3e2006-03-22 00:08:11 -08001666 /*
1667 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 */
Christoph Lametere498be72005-09-09 13:03:32 -07001669 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001670 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 return 0;
1672}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673__initcall(cpucache_init);
1674
1675/*
1676 * Interface to system's page allocator. No need to hold the cache-lock.
1677 *
1678 * If we requested dmaable memory, we will get it. Even if we
1679 * did not request dmaable memory, we might get it, but that
1680 * would be relatively rare and ignorable.
1681 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001682static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
1684 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001685 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 int i;
1687
Luke Yangd6fef9d2006-04-10 22:52:56 -07001688#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001689 /*
1690 * Nommu uses slab's for process anonymous memory allocations, and thus
1691 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001692 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001693 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001694#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001695
Christoph Lameter3c517a62006-12-06 20:33:29 -08001696 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001697 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1698 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001699
Linus Torvalds517d0862009-06-16 19:50:13 -07001700 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 if (!page)
1702 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001704 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001706 add_zone_page_state(page_zone(page),
1707 NR_SLAB_RECLAIMABLE, nr_pages);
1708 else
1709 add_zone_page_state(page_zone(page),
1710 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001711 for (i = 0; i < nr_pages; i++)
1712 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001713
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001714 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1715 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1716
1717 if (cachep->ctor)
1718 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1719 else
1720 kmemcheck_mark_unallocated_pages(page, nr_pages);
1721 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001722
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001723 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724}
1725
1726/*
1727 * Interface to system's page release.
1728 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001729static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001731 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 struct page *page = virt_to_page(addr);
1733 const unsigned long nr_freed = i;
1734
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001735 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001736
Christoph Lameter972d1a72006-09-25 23:31:51 -07001737 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1738 sub_zone_page_state(page_zone(page),
1739 NR_SLAB_RECLAIMABLE, nr_freed);
1740 else
1741 sub_zone_page_state(page_zone(page),
1742 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001744 BUG_ON(!PageSlab(page));
1745 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 page++;
1747 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 if (current->reclaim_state)
1749 current->reclaim_state->reclaimed_slab += nr_freed;
1750 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751}
1752
1753static void kmem_rcu_free(struct rcu_head *head)
1754{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001755 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001756 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757
1758 kmem_freepages(cachep, slab_rcu->addr);
1759 if (OFF_SLAB(cachep))
1760 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1761}
1762
1763#if DEBUG
1764
1765#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001766static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001769 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001771 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 return;
1775
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001776 *addr++ = 0x12345678;
1777 *addr++ = caller;
1778 *addr++ = smp_processor_id();
1779 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 {
1781 unsigned long *sptr = &caller;
1782 unsigned long svalue;
1783
1784 while (!kstack_end(sptr)) {
1785 svalue = *sptr++;
1786 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001787 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 size -= sizeof(unsigned long);
1789 if (size <= sizeof(unsigned long))
1790 break;
1791 }
1792 }
1793
1794 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796}
1797#endif
1798
Pekka Enberg343e0d72006-02-01 03:05:50 -08001799static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001801 int size = obj_size(cachep);
1802 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803
1804 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001805 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806}
1807
1808static void dump_line(char *data, int offset, int limit)
1809{
1810 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001811 unsigned char error = 0;
1812 int bad_count = 0;
1813
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001815 for (i = 0; i < limit; i++) {
1816 if (data[offset + i] != POISON_FREE) {
1817 error = data[offset + i];
1818 bad_count++;
1819 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001820 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001821 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001823
1824 if (bad_count == 1) {
1825 error ^= POISON_FREE;
1826 if (!(error & (error - 1))) {
1827 printk(KERN_ERR "Single bit error detected. Probably "
1828 "bad RAM.\n");
1829#ifdef CONFIG_X86
1830 printk(KERN_ERR "Run memtest86+ or a similar memory "
1831 "test tool.\n");
1832#else
1833 printk(KERN_ERR "Run a memory test tool.\n");
1834#endif
1835 }
1836 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837}
1838#endif
1839
1840#if DEBUG
1841
Pekka Enberg343e0d72006-02-01 03:05:50 -08001842static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
1844 int i, size;
1845 char *realobj;
1846
1847 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001848 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001849 *dbg_redzone1(cachep, objp),
1850 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 }
1852
1853 if (cachep->flags & SLAB_STORE_USER) {
1854 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001855 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001857 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 printk("\n");
1859 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001860 realobj = (char *)objp + obj_offset(cachep);
1861 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001862 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 int limit;
1864 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001865 if (i + limit > size)
1866 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 dump_line(realobj, i, limit);
1868 }
1869}
1870
Pekka Enberg343e0d72006-02-01 03:05:50 -08001871static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872{
1873 char *realobj;
1874 int size, i;
1875 int lines = 0;
1876
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001877 realobj = (char *)objp + obj_offset(cachep);
1878 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001882 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 exp = POISON_END;
1884 if (realobj[i] != exp) {
1885 int limit;
1886 /* Mismatch ! */
1887 /* Print header */
1888 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001889 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001890 "Slab corruption: %s start=%p, len=%d\n",
1891 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 print_objinfo(cachep, objp, 0);
1893 }
1894 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001897 if (i + limit > size)
1898 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 dump_line(realobj, i, limit);
1900 i += 16;
1901 lines++;
1902 /* Limit to 5 lines */
1903 if (lines > 5)
1904 break;
1905 }
1906 }
1907 if (lines != 0) {
1908 /* Print some data about the neighboring objects, if they
1909 * exist:
1910 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001911 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001912 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001914 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001916 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001917 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001919 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 print_objinfo(cachep, objp, 2);
1921 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001923 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001924 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001926 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 print_objinfo(cachep, objp, 2);
1928 }
1929 }
1930}
1931#endif
1932
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301934static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001935{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 int i;
1937 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001938 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
1940 if (cachep->flags & SLAB_POISON) {
1941#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001942 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1943 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001944 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001945 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 else
1947 check_poison_obj(cachep, objp);
1948#else
1949 check_poison_obj(cachep, objp);
1950#endif
1951 }
1952 if (cachep->flags & SLAB_RED_ZONE) {
1953 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1954 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001955 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1957 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001958 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001961}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301963static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001964{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001965}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966#endif
1967
Randy Dunlap911851e2006-03-22 00:08:14 -08001968/**
1969 * slab_destroy - destroy and release all objects in a slab
1970 * @cachep: cache pointer being destroyed
1971 * @slabp: slab pointer being destroyed
1972 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001973 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001974 * Before calling the slab must have been unlinked from the cache. The
1975 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001976 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001977static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001978{
1979 void *addr = slabp->s_mem - slabp->colouroff;
1980
Rabin Vincente79aec22008-07-04 00:40:32 +05301981 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1983 struct slab_rcu *slab_rcu;
1984
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001985 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 slab_rcu->cachep = cachep;
1987 slab_rcu->addr = addr;
1988 call_rcu(&slab_rcu->head, kmem_rcu_free);
1989 } else {
1990 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001991 if (OFF_SLAB(cachep))
1992 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 }
1994}
1995
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001996static void __kmem_cache_destroy(struct kmem_cache *cachep)
1997{
1998 int i;
1999 struct kmem_list3 *l3;
2000
2001 for_each_online_cpu(i)
2002 kfree(cachep->array[i]);
2003
2004 /* NUMA: free the list3 structures */
2005 for_each_online_node(i) {
2006 l3 = cachep->nodelists[i];
2007 if (l3) {
2008 kfree(l3->shared);
2009 free_alien_cache(l3->alien);
2010 kfree(l3);
2011 }
2012 }
2013 kmem_cache_free(&cache_cache, cachep);
2014}
2015
2016
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002018 * calculate_slab_order - calculate size (page order) of slabs
2019 * @cachep: pointer to the cache that is being created
2020 * @size: size of objects to be created in this cache.
2021 * @align: required alignment for the objects.
2022 * @flags: slab allocation flags
2023 *
2024 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025 *
2026 * This could be made much more intelligent. For now, try to avoid using
2027 * high order pages for slabs. When the gfp() functions are more friendly
2028 * towards high-order requests, this should be changed.
2029 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002030static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002031 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002033 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002034 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002035 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002036
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002037 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002038 unsigned int num;
2039 size_t remainder;
2040
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002041 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002042 if (!num)
2043 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002044
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002045 if (flags & CFLGS_OFF_SLAB) {
2046 /*
2047 * Max number of objs-per-slab for caches which
2048 * use off-slab slabs. Needed to avoid a possible
2049 * looping condition in cache_grow().
2050 */
2051 offslab_limit = size - sizeof(struct slab);
2052 offslab_limit /= sizeof(kmem_bufctl_t);
2053
2054 if (num > offslab_limit)
2055 break;
2056 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002059 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002060 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002061 left_over = remainder;
2062
2063 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002064 * A VFS-reclaimable slab tends to have most allocations
2065 * as GFP_NOFS and we really don't want to have to be allocating
2066 * higher-order pages when we are unable to shrink dcache.
2067 */
2068 if (flags & SLAB_RECLAIM_ACCOUNT)
2069 break;
2070
2071 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002072 * Large number of objects is good, but very large slabs are
2073 * currently bad for the gfp()s.
2074 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002075 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002076 break;
2077
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002078 /*
2079 * Acceptable internal fragmentation?
2080 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002081 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002082 break;
2083 }
2084 return left_over;
2085}
2086
Pekka Enberg83b519e2009-06-10 19:40:04 +03002087static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002088{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002089 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002090 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002091
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002092 if (g_cpucache_up == NONE) {
2093 /*
2094 * Note: the first kmem_cache_create must create the cache
2095 * that's used by kmalloc(24), otherwise the creation of
2096 * further caches will BUG().
2097 */
2098 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2099
2100 /*
2101 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2102 * the first cache, then we need to set up all its list3s,
2103 * otherwise the creation of further caches will BUG().
2104 */
2105 set_up_list3s(cachep, SIZE_AC);
2106 if (INDEX_AC == INDEX_L3)
2107 g_cpucache_up = PARTIAL_L3;
2108 else
2109 g_cpucache_up = PARTIAL_AC;
2110 } else {
2111 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002112 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002113
2114 if (g_cpucache_up == PARTIAL_AC) {
2115 set_up_list3s(cachep, SIZE_L3);
2116 g_cpucache_up = PARTIAL_L3;
2117 } else {
2118 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002119 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002120 cachep->nodelists[node] =
2121 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002122 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002123 BUG_ON(!cachep->nodelists[node]);
2124 kmem_list3_init(cachep->nodelists[node]);
2125 }
2126 }
2127 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002128 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002129 jiffies + REAPTIMEOUT_LIST3 +
2130 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2131
2132 cpu_cache_get(cachep)->avail = 0;
2133 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2134 cpu_cache_get(cachep)->batchcount = 1;
2135 cpu_cache_get(cachep)->touched = 0;
2136 cachep->batchcount = 1;
2137 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002138 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002139}
2140
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002141/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 * kmem_cache_create - Create a cache.
2143 * @name: A string which is used in /proc/slabinfo to identify this cache.
2144 * @size: The size of objects to be created in this cache.
2145 * @align: The required alignment for the objects.
2146 * @flags: SLAB flags
2147 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 *
2149 * Returns a ptr to the cache on success, NULL on failure.
2150 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002151 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 *
2153 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002154 * the module calling this has to destroy the cache before getting unloaded.
2155 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 * The flags are
2157 *
2158 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2159 * to catch references to uninitialised memory.
2160 *
2161 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2162 * for buffer overruns.
2163 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2165 * cacheline. This can be beneficial if you're counting cycles as closely
2166 * as davem.
2167 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002168struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002170 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171{
2172 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002173 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002174 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175
2176 /*
2177 * Sanity checks... these are all serious usage bugs.
2178 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002179 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002180 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002181 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002182 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002183 BUG();
2184 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002186 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002187 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302188 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002189 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002190 if (slab_is_available()) {
2191 get_online_cpus();
2192 mutex_lock(&cache_chain_mutex);
2193 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002194
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002195 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002196 char tmp;
2197 int res;
2198
2199 /*
2200 * This happens when the module gets unloaded and doesn't
2201 * destroy its slab cache and no-one else reuses the vmalloc
2202 * area of the module. Print a warning.
2203 */
Andrew Morton138ae662006-12-06 20:36:41 -08002204 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002205 if (res) {
matzeb4169522007-05-06 14:49:52 -07002206 printk(KERN_ERR
2207 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002208 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002209 continue;
2210 }
2211
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002212 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002213 printk(KERN_ERR
2214 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002215 dump_stack();
2216 goto oops;
2217 }
2218 }
2219
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220#if DEBUG
2221 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222#if FORCED_DEBUG
2223 /*
2224 * Enable redzoning and last user accounting, except for caches with
2225 * large objects, if the increased size would increase the object size
2226 * above the next power of two: caches with object sizes just above a
2227 * power of two have a significant amount of internal fragmentation.
2228 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002229 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2230 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002231 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 if (!(flags & SLAB_DESTROY_BY_RCU))
2233 flags |= SLAB_POISON;
2234#endif
2235 if (flags & SLAB_DESTROY_BY_RCU)
2236 BUG_ON(flags & SLAB_POISON);
2237#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002239 * Always checks flags, a caller might be expecting debug support which
2240 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002242 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243
Andrew Mortona737b3e2006-03-22 00:08:11 -08002244 /*
2245 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 * unaligned accesses for some archs when redzoning is used, and makes
2247 * sure any on-slab bufctl's are also correctly aligned.
2248 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002249 if (size & (BYTES_PER_WORD - 1)) {
2250 size += (BYTES_PER_WORD - 1);
2251 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 }
2253
Andrew Mortona737b3e2006-03-22 00:08:11 -08002254 /* calculate the final buffer alignment: */
2255
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 /* 1) arch recommendation: can be overridden for debug */
2257 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002258 /*
2259 * Default alignment: as specified by the arch code. Except if
2260 * an object is really small, then squeeze multiple objects into
2261 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 */
2263 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002264 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 ralign /= 2;
2266 } else {
2267 ralign = BYTES_PER_WORD;
2268 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002269
2270 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002271 * Redzoning and user store require word alignment or possibly larger.
2272 * Note this will be overridden by architecture or caller mandated
2273 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002274 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002275 if (flags & SLAB_STORE_USER)
2276 ralign = BYTES_PER_WORD;
2277
2278 if (flags & SLAB_RED_ZONE) {
2279 ralign = REDZONE_ALIGN;
2280 /* If redzoning, ensure that the second redzone is suitably
2281 * aligned, by adjusting the object size accordingly. */
2282 size += REDZONE_ALIGN - 1;
2283 size &= ~(REDZONE_ALIGN - 1);
2284 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002285
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002286 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 if (ralign < ARCH_SLAB_MINALIGN) {
2288 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002290 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 if (ralign < align) {
2292 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002294 /* disable debug if necessary */
2295 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002296 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002297 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002298 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 */
2300 align = ralign;
2301
Pekka Enberg83b519e2009-06-10 19:40:04 +03002302 if (slab_is_available())
2303 gfp = GFP_KERNEL;
2304 else
2305 gfp = GFP_NOWAIT;
2306
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002308 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002310 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002312 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002314 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315
Pekka Enbergca5f9702006-09-25 23:31:25 -07002316 /*
2317 * Both debugging options require word-alignment which is calculated
2318 * into align above.
2319 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002322 cachep->obj_offset += sizeof(unsigned long long);
2323 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 }
2325 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002326 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002327 * the real object. But if the second red zone needs to be
2328 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002330 if (flags & SLAB_RED_ZONE)
2331 size += REDZONE_ALIGN;
2332 else
2333 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 }
2335#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002336 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002337 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2338 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 size = PAGE_SIZE;
2340 }
2341#endif
2342#endif
2343
Ingo Molnare0a42722006-06-23 02:03:46 -07002344 /*
2345 * Determine if the slab management is 'on' or 'off' slab.
2346 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002347 * it too early on. Always use on-slab management when
2348 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002349 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002350 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2351 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 /*
2353 * Size is large, assume best to place the slab management obj
2354 * off-slab (should allow better packing of objs).
2355 */
2356 flags |= CFLGS_OFF_SLAB;
2357
2358 size = ALIGN(size, align);
2359
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002360 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
2362 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002363 printk(KERN_ERR
2364 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 kmem_cache_free(&cache_cache, cachep);
2366 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002367 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002369 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2370 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371
2372 /*
2373 * If the slab has been placed off-slab, and we have enough space then
2374 * move it on-slab. This is at the expense of any extra colouring.
2375 */
2376 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2377 flags &= ~CFLGS_OFF_SLAB;
2378 left_over -= slab_size;
2379 }
2380
2381 if (flags & CFLGS_OFF_SLAB) {
2382 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002383 slab_size =
2384 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302385
2386#ifdef CONFIG_PAGE_POISONING
2387 /* If we're going to use the generic kernel_map_pages()
2388 * poisoning, then it's going to smash the contents of
2389 * the redzone and userword anyhow, so switch them off.
2390 */
2391 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2392 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2393#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
2395
2396 cachep->colour_off = cache_line_size();
2397 /* Offset must be a multiple of the alignment. */
2398 if (cachep->colour_off < align)
2399 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002400 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 cachep->slab_size = slab_size;
2402 cachep->flags = flags;
2403 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002404 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002406 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002407 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002409 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002410 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002411 /*
2412 * This is a possibility for one of the malloc_sizes caches.
2413 * But since we go off slab only for object size greater than
2414 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2415 * this should not happen at all.
2416 * But leave a BUG_ON for some lucky dude.
2417 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002418 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002419 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 cachep->name = name;
2422
Pekka Enberg83b519e2009-06-10 19:40:04 +03002423 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002424 __kmem_cache_destroy(cachep);
2425 cachep = NULL;
2426 goto oops;
2427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 /* cache setup completed, link it into the list */
2430 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002431oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 if (!cachep && (flags & SLAB_PANIC))
2433 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002434 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002435 if (slab_is_available()) {
2436 mutex_unlock(&cache_chain_mutex);
2437 put_online_cpus();
2438 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 return cachep;
2440}
2441EXPORT_SYMBOL(kmem_cache_create);
2442
2443#if DEBUG
2444static void check_irq_off(void)
2445{
2446 BUG_ON(!irqs_disabled());
2447}
2448
2449static void check_irq_on(void)
2450{
2451 BUG_ON(irqs_disabled());
2452}
2453
Pekka Enberg343e0d72006-02-01 03:05:50 -08002454static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455{
2456#ifdef CONFIG_SMP
2457 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002458 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459#endif
2460}
Christoph Lametere498be72005-09-09 13:03:32 -07002461
Pekka Enberg343e0d72006-02-01 03:05:50 -08002462static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002463{
2464#ifdef CONFIG_SMP
2465 check_irq_off();
2466 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2467#endif
2468}
2469
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470#else
2471#define check_irq_off() do { } while(0)
2472#define check_irq_on() do { } while(0)
2473#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002474#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475#endif
2476
Christoph Lameteraab22072006-03-22 00:09:06 -08002477static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2478 struct array_cache *ac,
2479 int force, int node);
2480
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481static void do_drain(void *arg)
2482{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002483 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002485 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486
2487 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002488 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002489 spin_lock(&cachep->nodelists[node]->list_lock);
2490 free_block(cachep, ac->entry, ac->avail, node);
2491 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 ac->avail = 0;
2493}
2494
Pekka Enberg343e0d72006-02-01 03:05:50 -08002495static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496{
Christoph Lametere498be72005-09-09 13:03:32 -07002497 struct kmem_list3 *l3;
2498 int node;
2499
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002500 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002502 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002503 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002504 if (l3 && l3->alien)
2505 drain_alien_cache(cachep, l3->alien);
2506 }
2507
2508 for_each_online_node(node) {
2509 l3 = cachep->nodelists[node];
2510 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002511 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002512 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513}
2514
Christoph Lametered11d9e2006-06-30 01:55:45 -07002515/*
2516 * Remove slabs from the list of free slabs.
2517 * Specify the number of slabs to drain in tofree.
2518 *
2519 * Returns the actual number of slabs released.
2520 */
2521static int drain_freelist(struct kmem_cache *cache,
2522 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002524 struct list_head *p;
2525 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
Christoph Lametered11d9e2006-06-30 01:55:45 -07002528 nr_freed = 0;
2529 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530
Christoph Lametered11d9e2006-06-30 01:55:45 -07002531 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002532 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002533 if (p == &l3->slabs_free) {
2534 spin_unlock_irq(&l3->list_lock);
2535 goto out;
2536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537
Christoph Lametered11d9e2006-06-30 01:55:45 -07002538 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002540 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541#endif
2542 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002543 /*
2544 * Safe to drop the lock. The slab is no longer linked
2545 * to the cache.
2546 */
2547 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002548 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002549 slab_destroy(cache, slabp);
2550 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002552out:
2553 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554}
2555
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002556/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002557static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002558{
2559 int ret = 0, i = 0;
2560 struct kmem_list3 *l3;
2561
2562 drain_cpu_caches(cachep);
2563
2564 check_irq_on();
2565 for_each_online_node(i) {
2566 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002567 if (!l3)
2568 continue;
2569
2570 drain_freelist(cachep, l3, l3->free_objects);
2571
2572 ret += !list_empty(&l3->slabs_full) ||
2573 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002574 }
2575 return (ret ? 1 : 0);
2576}
2577
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578/**
2579 * kmem_cache_shrink - Shrink a cache.
2580 * @cachep: The cache to shrink.
2581 *
2582 * Releases as many slabs as possible for a cache.
2583 * To help debugging, a zero exit status indicates all slabs were released.
2584 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002585int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002587 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002588 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002590 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002591 mutex_lock(&cache_chain_mutex);
2592 ret = __cache_shrink(cachep);
2593 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002594 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002595 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596}
2597EXPORT_SYMBOL(kmem_cache_shrink);
2598
2599/**
2600 * kmem_cache_destroy - delete a cache
2601 * @cachep: the cache to destroy
2602 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002603 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 *
2605 * It is expected this function will be called by a module when it is
2606 * unloaded. This will remove the cache completely, and avoid a duplicate
2607 * cache being allocated each time a module is loaded and unloaded, if the
2608 * module doesn't have persistent in-kernel storage across loads and unloads.
2609 *
2610 * The cache must be empty before calling this function.
2611 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002612 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 * during the kmem_cache_destroy().
2614 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002615void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002617 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002620 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002621 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 /*
2623 * the chain is never empty, cache_cache is never destroyed
2624 */
2625 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 if (__cache_shrink(cachep)) {
2627 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002628 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002629 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002630 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002631 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
2633
2634 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002635 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002637 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002638 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002639 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640}
2641EXPORT_SYMBOL(kmem_cache_destroy);
2642
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002643/*
2644 * Get the memory for a slab management obj.
2645 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2646 * always come from malloc_sizes caches. The slab descriptor cannot
2647 * come from the same cache which is getting created because,
2648 * when we are searching for an appropriate cache for these
2649 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2650 * If we are creating a malloc_sizes cache here it would not be visible to
2651 * kmem_find_general_cachep till the initialization is complete.
2652 * Hence we cannot have slabp_cache same as the original cache.
2653 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002654static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002655 int colour_off, gfp_t local_flags,
2656 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657{
2658 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002659
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 if (OFF_SLAB(cachep)) {
2661 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002662 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002663 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002664 /*
2665 * If the first object in the slab is leaked (it's allocated
2666 * but no one has a reference to it), we want to make sure
2667 * kmemleak does not treat the ->s_mem pointer as a reference
2668 * to the object. Otherwise we will not report the leak.
2669 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002670 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2671 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 if (!slabp)
2673 return NULL;
2674 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002675 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 colour_off += cachep->slab_size;
2677 }
2678 slabp->inuse = 0;
2679 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002680 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002681 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002682 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 return slabp;
2684}
2685
2686static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2687{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002688 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689}
2690
Pekka Enberg343e0d72006-02-01 03:05:50 -08002691static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002692 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693{
2694 int i;
2695
2696 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002697 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698#if DEBUG
2699 /* need to poison the objs? */
2700 if (cachep->flags & SLAB_POISON)
2701 poison_obj(cachep, objp, POISON_FREE);
2702 if (cachep->flags & SLAB_STORE_USER)
2703 *dbg_userword(cachep, objp) = NULL;
2704
2705 if (cachep->flags & SLAB_RED_ZONE) {
2706 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2707 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2708 }
2709 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002710 * Constructors are not allowed to allocate memory from the same
2711 * cache which they are a constructor for. Otherwise, deadlock.
2712 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 */
2714 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002715 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
2717 if (cachep->flags & SLAB_RED_ZONE) {
2718 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2719 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002720 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2722 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002723 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002725 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2726 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002727 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002728 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729#else
2730 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002731 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002733 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002735 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736}
2737
Pekka Enberg343e0d72006-02-01 03:05:50 -08002738static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002740 if (CONFIG_ZONE_DMA_FLAG) {
2741 if (flags & GFP_DMA)
2742 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2743 else
2744 BUG_ON(cachep->gfpflags & GFP_DMA);
2745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746}
2747
Andrew Mortona737b3e2006-03-22 00:08:11 -08002748static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2749 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002750{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002751 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002752 kmem_bufctl_t next;
2753
2754 slabp->inuse++;
2755 next = slab_bufctl(slabp)[slabp->free];
2756#if DEBUG
2757 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2758 WARN_ON(slabp->nodeid != nodeid);
2759#endif
2760 slabp->free = next;
2761
2762 return objp;
2763}
2764
Andrew Mortona737b3e2006-03-22 00:08:11 -08002765static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2766 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002767{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002768 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002769
2770#if DEBUG
2771 /* Verify that the slab belongs to the intended node */
2772 WARN_ON(slabp->nodeid != nodeid);
2773
Al Viro871751e2006-03-25 03:06:39 -08002774 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002775 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002776 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002777 BUG();
2778 }
2779#endif
2780 slab_bufctl(slabp)[objnr] = slabp->free;
2781 slabp->free = objnr;
2782 slabp->inuse--;
2783}
2784
Pekka Enberg47768742006-06-23 02:03:07 -07002785/*
2786 * Map pages beginning at addr to the given cache and slab. This is required
2787 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002788 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002789 */
2790static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2791 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792{
Pekka Enberg47768742006-06-23 02:03:07 -07002793 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 struct page *page;
2795
Pekka Enberg47768742006-06-23 02:03:07 -07002796 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002797
Pekka Enberg47768742006-06-23 02:03:07 -07002798 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002799 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002800 nr_pages <<= cache->gfporder;
2801
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002803 page_set_cache(page, cache);
2804 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002806 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807}
2808
2809/*
2810 * Grow (by 1) the number of slabs within a cache. This is called by
2811 * kmem_cache_alloc() when there are no active objs left in a cache.
2812 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002813static int cache_grow(struct kmem_cache *cachep,
2814 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002816 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002817 size_t offset;
2818 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002819 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
Andrew Mortona737b3e2006-03-22 00:08:11 -08002821 /*
2822 * Be lazy and only check for valid flags here, keeping it out of the
2823 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002825 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2826 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002828 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002830 l3 = cachep->nodelists[nodeid];
2831 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832
2833 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002834 offset = l3->colour_next;
2835 l3->colour_next++;
2836 if (l3->colour_next >= cachep->colour)
2837 l3->colour_next = 0;
2838 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002840 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841
2842 if (local_flags & __GFP_WAIT)
2843 local_irq_enable();
2844
2845 /*
2846 * The test for missing atomic flag is performed here, rather than
2847 * the more obvious place, simply to reduce the critical path length
2848 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2849 * will eventually be caught here (where it matters).
2850 */
2851 kmem_flagcheck(cachep, flags);
2852
Andrew Mortona737b3e2006-03-22 00:08:11 -08002853 /*
2854 * Get mem for the objs. Attempt to allocate a physical page from
2855 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002856 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002857 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002858 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002859 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 goto failed;
2861
2862 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002863 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002864 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002865 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 goto opps1;
2867
Pekka Enberg47768742006-06-23 02:03:07 -07002868 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869
Christoph Lametera35afb82007-05-16 22:10:57 -07002870 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871
2872 if (local_flags & __GFP_WAIT)
2873 local_irq_disable();
2874 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002875 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
2877 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002878 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002880 l3->free_objects += cachep->num;
2881 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002883opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002885failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 if (local_flags & __GFP_WAIT)
2887 local_irq_disable();
2888 return 0;
2889}
2890
2891#if DEBUG
2892
2893/*
2894 * Perform extra freeing checks:
2895 * - detect bad pointers.
2896 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 */
2898static void kfree_debugcheck(const void *objp)
2899{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 if (!virt_addr_valid(objp)) {
2901 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002902 (unsigned long)objp);
2903 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905}
2906
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002907static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2908{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002909 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002910
2911 redzone1 = *dbg_redzone1(cache, obj);
2912 redzone2 = *dbg_redzone2(cache, obj);
2913
2914 /*
2915 * Redzone is ok.
2916 */
2917 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2918 return;
2919
2920 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2921 slab_error(cache, "double free detected");
2922 else
2923 slab_error(cache, "memory outside object was overwritten");
2924
David Woodhouseb46b8f12007-05-08 00:22:59 -07002925 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002926 obj, redzone1, redzone2);
2927}
2928
Pekka Enberg343e0d72006-02-01 03:05:50 -08002929static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931{
2932 struct page *page;
2933 unsigned int objnr;
2934 struct slab *slabp;
2935
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002936 BUG_ON(virt_to_cache(objp) != cachep);
2937
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002938 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002940 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
Pekka Enberg065d41c2005-11-13 16:06:46 -08002942 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943
2944 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002945 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2947 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2948 }
2949 if (cachep->flags & SLAB_STORE_USER)
2950 *dbg_userword(cachep, objp) = caller;
2951
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002952 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953
2954 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002955 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956
Al Viro871751e2006-03-25 03:06:39 -08002957#ifdef CONFIG_DEBUG_SLAB_LEAK
2958 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2959#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 if (cachep->flags & SLAB_POISON) {
2961#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002962 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002964 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002965 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 } else {
2967 poison_obj(cachep, objp, POISON_FREE);
2968 }
2969#else
2970 poison_obj(cachep, objp, POISON_FREE);
2971#endif
2972 }
2973 return objp;
2974}
2975
Pekka Enberg343e0d72006-02-01 03:05:50 -08002976static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977{
2978 kmem_bufctl_t i;
2979 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002980
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 /* Check slab's freelist to see if this obj is there. */
2982 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2983 entries++;
2984 if (entries > cachep->num || i >= cachep->num)
2985 goto bad;
2986 }
2987 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002988bad:
2989 printk(KERN_ERR "slab: Internal list corruption detected in "
2990 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2991 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002992 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002993 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002994 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002995 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002997 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998 }
2999 printk("\n");
3000 BUG();
3001 }
3002}
3003#else
3004#define kfree_debugcheck(x) do { } while(0)
3005#define cache_free_debugcheck(x,objp,z) (objp)
3006#define check_slabp(x,y) do { } while(0)
3007#endif
3008
Pekka Enberg343e0d72006-02-01 03:05:50 -08003009static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010{
3011 int batchcount;
3012 struct kmem_list3 *l3;
3013 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003014 int node;
3015
Andrew Mortona737b3e2006-03-22 00:08:11 -08003016retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003017 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003018 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003019 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 batchcount = ac->batchcount;
3021 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003022 /*
3023 * If there was little recent activity on this cache, then
3024 * perform only a partial refill. Otherwise we could generate
3025 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 */
3027 batchcount = BATCHREFILL_LIMIT;
3028 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003029 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030
Christoph Lametere498be72005-09-09 13:03:32 -07003031 BUG_ON(ac->avail > 0 || !l3);
3032 spin_lock(&l3->list_lock);
3033
Christoph Lameter3ded1752006-03-25 03:06:44 -08003034 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003035 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3036 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003037 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003038 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003039
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 while (batchcount > 0) {
3041 struct list_head *entry;
3042 struct slab *slabp;
3043 /* Get slab alloc is to come from. */
3044 entry = l3->slabs_partial.next;
3045 if (entry == &l3->slabs_partial) {
3046 l3->free_touched = 1;
3047 entry = l3->slabs_free.next;
3048 if (entry == &l3->slabs_free)
3049 goto must_grow;
3050 }
3051
3052 slabp = list_entry(entry, struct slab, list);
3053 check_slabp(cachep, slabp);
3054 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003055
3056 /*
3057 * The slab was either on partial or free list so
3058 * there must be at least one object available for
3059 * allocation.
3060 */
roel kluin249b9f32008-10-29 17:18:07 -04003061 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003062
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 STATS_INC_ALLOCED(cachep);
3065 STATS_INC_ACTIVE(cachep);
3066 STATS_SET_HIGH(cachep);
3067
Matthew Dobson78d382d2006-02-01 03:05:47 -08003068 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003069 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 }
3071 check_slabp(cachep, slabp);
3072
3073 /* move slabp to correct slabp list: */
3074 list_del(&slabp->list);
3075 if (slabp->free == BUFCTL_END)
3076 list_add(&slabp->list, &l3->slabs_full);
3077 else
3078 list_add(&slabp->list, &l3->slabs_partial);
3079 }
3080
Andrew Mortona737b3e2006-03-22 00:08:11 -08003081must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003083alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003084 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085
3086 if (unlikely(!ac->avail)) {
3087 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003088 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003089
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003091 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003092 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 return NULL;
3094
Andrew Mortona737b3e2006-03-22 00:08:11 -08003095 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 goto retry;
3097 }
3098 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003099 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100}
3101
Andrew Mortona737b3e2006-03-22 00:08:11 -08003102static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3103 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104{
3105 might_sleep_if(flags & __GFP_WAIT);
3106#if DEBUG
3107 kmem_flagcheck(cachep, flags);
3108#endif
3109}
3110
3111#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003112static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3113 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003115 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003117 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003119 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003120 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003121 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 else
3123 check_poison_obj(cachep, objp);
3124#else
3125 check_poison_obj(cachep, objp);
3126#endif
3127 poison_obj(cachep, objp, POISON_INUSE);
3128 }
3129 if (cachep->flags & SLAB_STORE_USER)
3130 *dbg_userword(cachep, objp) = caller;
3131
3132 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003133 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3134 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3135 slab_error(cachep, "double free, or memory outside"
3136 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003137 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003138 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003139 objp, *dbg_redzone1(cachep, objp),
3140 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 }
3142 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3143 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3144 }
Al Viro871751e2006-03-25 03:06:39 -08003145#ifdef CONFIG_DEBUG_SLAB_LEAK
3146 {
3147 struct slab *slabp;
3148 unsigned objnr;
3149
Christoph Lameterb49af682007-05-06 14:49:41 -07003150 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003151 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3152 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3153 }
3154#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003155 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003156 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003157 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003158 if (ARCH_SLAB_MINALIGN &&
3159 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003160 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003161 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003162 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 return objp;
3164}
3165#else
3166#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3167#endif
3168
Akinobu Mita773ff602008-12-23 19:37:01 +09003169static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003170{
3171 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003172 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003173
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003174 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003175}
3176
Pekka Enberg343e0d72006-02-01 03:05:50 -08003177static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003179 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 struct array_cache *ac;
3181
Alok N Kataria5c382302005-09-27 21:45:46 -07003182 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003183
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003184 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185 if (likely(ac->avail)) {
3186 STATS_INC_ALLOCHIT(cachep);
3187 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003188 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 } else {
3190 STATS_INC_ALLOCMISS(cachep);
3191 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003192 /*
3193 * the 'ac' may be updated by cache_alloc_refill(),
3194 * and kmemleak_erase() requires its correct value.
3195 */
3196 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003198 /*
3199 * To avoid a false negative, if an object that is in one of the
3200 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3201 * treat the array pointers as a reference to the object.
3202 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003203 if (objp)
3204 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003205 return objp;
3206}
3207
Christoph Lametere498be72005-09-09 13:03:32 -07003208#ifdef CONFIG_NUMA
3209/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003210 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003211 *
3212 * If we are in_interrupt, then process context, including cpusets and
3213 * mempolicy, may not apply and should not be used for allocation policy.
3214 */
3215static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3216{
3217 int nid_alloc, nid_here;
3218
Christoph Lameter765c4502006-09-27 01:50:08 -07003219 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003220 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003221 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003222 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003223 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003224 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003225 else if (current->mempolicy)
3226 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003227 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003228 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003229 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003230 return NULL;
3231}
3232
3233/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003234 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003235 * certain node and fall back is permitted. First we scan all the
3236 * available nodelists for available objects. If that fails then we
3237 * perform an allocation without specifying a node. This allows the page
3238 * allocator to do its reclaim / fallback magic. We then insert the
3239 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003240 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003241static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003242{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003243 struct zonelist *zonelist;
3244 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003245 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003246 struct zone *zone;
3247 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003248 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003249 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003250
3251 if (flags & __GFP_THISNODE)
3252 return NULL;
3253
Miao Xiec0ff7452010-05-24 14:32:08 -07003254 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003255 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003256 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003257
Christoph Lameter3c517a62006-12-06 20:33:29 -08003258retry:
3259 /*
3260 * Look through allowed nodes for objects available
3261 * from existing per node queues.
3262 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003263 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3264 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003265
Mel Gorman54a6eb52008-04-28 02:12:16 -07003266 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003267 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003268 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003269 obj = ____cache_alloc_node(cache,
3270 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003271 if (obj)
3272 break;
3273 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003274 }
3275
Christoph Lametercfce6602007-05-06 14:50:17 -07003276 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003277 /*
3278 * This allocation will be performed within the constraints
3279 * of the current cpuset / memory policy requirements.
3280 * We may trigger various forms of reclaim on the allowed
3281 * set and go into memory reserves if necessary.
3282 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003283 if (local_flags & __GFP_WAIT)
3284 local_irq_enable();
3285 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003286 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003287 if (local_flags & __GFP_WAIT)
3288 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003289 if (obj) {
3290 /*
3291 * Insert into the appropriate per node queues
3292 */
3293 nid = page_to_nid(virt_to_page(obj));
3294 if (cache_grow(cache, flags, nid, obj)) {
3295 obj = ____cache_alloc_node(cache,
3296 flags | GFP_THISNODE, nid);
3297 if (!obj)
3298 /*
3299 * Another processor may allocate the
3300 * objects in the slab since we are
3301 * not holding any locks.
3302 */
3303 goto retry;
3304 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003305 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003306 obj = NULL;
3307 }
3308 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003309 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003310 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003311 return obj;
3312}
3313
3314/*
Christoph Lametere498be72005-09-09 13:03:32 -07003315 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003317static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003318 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003319{
3320 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 struct slab *slabp;
3322 struct kmem_list3 *l3;
3323 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003324 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003326 l3 = cachep->nodelists[nodeid];
3327 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003328
Andrew Mortona737b3e2006-03-22 00:08:11 -08003329retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003330 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003331 spin_lock(&l3->list_lock);
3332 entry = l3->slabs_partial.next;
3333 if (entry == &l3->slabs_partial) {
3334 l3->free_touched = 1;
3335 entry = l3->slabs_free.next;
3336 if (entry == &l3->slabs_free)
3337 goto must_grow;
3338 }
Christoph Lametere498be72005-09-09 13:03:32 -07003339
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003340 slabp = list_entry(entry, struct slab, list);
3341 check_spinlock_acquired_node(cachep, nodeid);
3342 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003343
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003344 STATS_INC_NODEALLOCS(cachep);
3345 STATS_INC_ACTIVE(cachep);
3346 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003347
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003348 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003349
Matthew Dobson78d382d2006-02-01 03:05:47 -08003350 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003351 check_slabp(cachep, slabp);
3352 l3->free_objects--;
3353 /* move slabp to correct slabp list: */
3354 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003355
Andrew Mortona737b3e2006-03-22 00:08:11 -08003356 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003357 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003358 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003359 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003360
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003361 spin_unlock(&l3->list_lock);
3362 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003363
Andrew Mortona737b3e2006-03-22 00:08:11 -08003364must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003365 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003366 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003367 if (x)
3368 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003369
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003370 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003371
Andrew Mortona737b3e2006-03-22 00:08:11 -08003372done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003373 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003374}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003375
3376/**
3377 * kmem_cache_alloc_node - Allocate an object on the specified node
3378 * @cachep: The cache to allocate from.
3379 * @flags: See kmalloc().
3380 * @nodeid: node number of the target node.
3381 * @caller: return address of caller, used for debug information
3382 *
3383 * Identical to kmem_cache_alloc but it will allocate memory on the given
3384 * node, which can improve the performance for cpu bound structures.
3385 *
3386 * Fallback to other node is possible if __GFP_THISNODE is not set.
3387 */
3388static __always_inline void *
3389__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3390 void *caller)
3391{
3392 unsigned long save_flags;
3393 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003394 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003396 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003397
Nick Piggincf40bd12009-01-21 08:12:39 +01003398 lockdep_trace_alloc(flags);
3399
Akinobu Mita773ff602008-12-23 19:37:01 +09003400 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003401 return NULL;
3402
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003403 cache_alloc_debugcheck_before(cachep, flags);
3404 local_irq_save(save_flags);
3405
Tim Blechmann8e15b792009-11-30 18:59:34 +01003406 if (nodeid == -1)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003407 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003408
3409 if (unlikely(!cachep->nodelists[nodeid])) {
3410 /* Node not bootstrapped yet */
3411 ptr = fallback_alloc(cachep, flags);
3412 goto out;
3413 }
3414
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003415 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003416 /*
3417 * Use the locally cached objects if possible.
3418 * However ____cache_alloc does not allow fallback
3419 * to other nodes. It may fail while we still have
3420 * objects on other nodes available.
3421 */
3422 ptr = ____cache_alloc(cachep, flags);
3423 if (ptr)
3424 goto out;
3425 }
3426 /* ___cache_alloc_node can fall back to other nodes */
3427 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3428 out:
3429 local_irq_restore(save_flags);
3430 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003431 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3432 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003433
Pekka Enbergc175eea2008-05-09 20:35:53 +02003434 if (likely(ptr))
3435 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3436
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003437 if (unlikely((flags & __GFP_ZERO) && ptr))
3438 memset(ptr, 0, obj_size(cachep));
3439
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003440 return ptr;
3441}
3442
3443static __always_inline void *
3444__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3445{
3446 void *objp;
3447
3448 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3449 objp = alternate_node_alloc(cache, flags);
3450 if (objp)
3451 goto out;
3452 }
3453 objp = ____cache_alloc(cache, flags);
3454
3455 /*
3456 * We may just have run out of memory on the local node.
3457 * ____cache_alloc_node() knows how to locate memory on other nodes
3458 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003459 if (!objp)
3460 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003461
3462 out:
3463 return objp;
3464}
3465#else
3466
3467static __always_inline void *
3468__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3469{
3470 return ____cache_alloc(cachep, flags);
3471}
3472
3473#endif /* CONFIG_NUMA */
3474
3475static __always_inline void *
3476__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3477{
3478 unsigned long save_flags;
3479 void *objp;
3480
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003481 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003482
Nick Piggincf40bd12009-01-21 08:12:39 +01003483 lockdep_trace_alloc(flags);
3484
Akinobu Mita773ff602008-12-23 19:37:01 +09003485 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003486 return NULL;
3487
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003488 cache_alloc_debugcheck_before(cachep, flags);
3489 local_irq_save(save_flags);
3490 objp = __do_cache_alloc(cachep, flags);
3491 local_irq_restore(save_flags);
3492 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003493 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3494 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003495 prefetchw(objp);
3496
Pekka Enbergc175eea2008-05-09 20:35:53 +02003497 if (likely(objp))
3498 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3499
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003500 if (unlikely((flags & __GFP_ZERO) && objp))
3501 memset(objp, 0, obj_size(cachep));
3502
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003503 return objp;
3504}
Christoph Lametere498be72005-09-09 13:03:32 -07003505
3506/*
3507 * Caller needs to acquire correct kmem_list's list_lock
3508 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003509static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003510 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511{
3512 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003513 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514
3515 for (i = 0; i < nr_objects; i++) {
3516 void *objp = objpp[i];
3517 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003519 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003520 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003522 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003524 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003526 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 check_slabp(cachep, slabp);
3528
3529 /* fixup slab chains */
3530 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003531 if (l3->free_objects > l3->free_limit) {
3532 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003533 /* No need to drop any previously held
3534 * lock here, even if we have a off-slab slab
3535 * descriptor it is guaranteed to come from
3536 * a different cache, refer to comments before
3537 * alloc_slabmgmt.
3538 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 slab_destroy(cachep, slabp);
3540 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003541 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 }
3543 } else {
3544 /* Unconditionally move a slab to the end of the
3545 * partial list on free - maximum time for the
3546 * other objects to be freed, too.
3547 */
Christoph Lametere498be72005-09-09 13:03:32 -07003548 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 }
3550 }
3551}
3552
Pekka Enberg343e0d72006-02-01 03:05:50 -08003553static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554{
3555 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003556 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003557 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558
3559 batchcount = ac->batchcount;
3560#if DEBUG
3561 BUG_ON(!batchcount || batchcount > ac->avail);
3562#endif
3563 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003564 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003565 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003566 if (l3->shared) {
3567 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003568 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569 if (max) {
3570 if (batchcount > max)
3571 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003572 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003573 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574 shared_array->avail += batchcount;
3575 goto free_done;
3576 }
3577 }
3578
Christoph Lameterff694162005-09-22 21:44:02 -07003579 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003580free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581#if STATS
3582 {
3583 int i = 0;
3584 struct list_head *p;
3585
Christoph Lametere498be72005-09-09 13:03:32 -07003586 p = l3->slabs_free.next;
3587 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 struct slab *slabp;
3589
3590 slabp = list_entry(p, struct slab, list);
3591 BUG_ON(slabp->inuse);
3592
3593 i++;
3594 p = p->next;
3595 }
3596 STATS_SET_FREEABLE(cachep, i);
3597 }
3598#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003599 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003601 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602}
3603
3604/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003605 * Release an obj back to its cache. If the obj has a constructed state, it must
3606 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003608static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3609 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003611 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612
3613 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003614 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003615 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616
Pekka Enbergc175eea2008-05-09 20:35:53 +02003617 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3618
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003619 /*
3620 * Skip calling cache_free_alien() when the platform is not numa.
3621 * This will avoid cache misses that happen while accessing slabp (which
3622 * is per page memory reference) to get nodeid. Instead use a global
3623 * variable to skip the call, which is mostly likely to be present in
3624 * the cache.
3625 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003626 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003627 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003628
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 if (likely(ac->avail < ac->limit)) {
3630 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003631 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 return;
3633 } else {
3634 STATS_INC_FREEMISS(cachep);
3635 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003636 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 }
3638}
3639
3640/**
3641 * kmem_cache_alloc - Allocate an object
3642 * @cachep: The cache to allocate from.
3643 * @flags: See kmalloc().
3644 *
3645 * Allocate an object from this cache. The flags are only relevant
3646 * if the cache has no available objects.
3647 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003648void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003650 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3651
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003652 trace_kmem_cache_alloc(_RET_IP_, ret,
3653 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003654
3655 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656}
3657EXPORT_SYMBOL(kmem_cache_alloc);
3658
Li Zefan0f24f122009-12-11 15:45:30 +08003659#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003660void *
3661kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003662{
Steven Rostedt85beb582010-11-24 16:23:34 -05003663 void *ret;
3664
3665 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3666
3667 trace_kmalloc(_RET_IP_, ret,
3668 size, slab_buffer_size(cachep), flags);
3669 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003670}
Steven Rostedt85beb582010-11-24 16:23:34 -05003671EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003672#endif
3673
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003675void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3676{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003677 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3678 __builtin_return_address(0));
3679
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003680 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3681 obj_size(cachep), cachep->buffer_size,
3682 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003683
3684 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003685}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686EXPORT_SYMBOL(kmem_cache_alloc_node);
3687
Li Zefan0f24f122009-12-11 15:45:30 +08003688#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003689void *kmem_cache_alloc_node_trace(size_t size,
3690 struct kmem_cache *cachep,
3691 gfp_t flags,
3692 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003693{
Steven Rostedt85beb582010-11-24 16:23:34 -05003694 void *ret;
3695
3696 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003697 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003698 trace_kmalloc_node(_RET_IP_, ret,
3699 size, slab_buffer_size(cachep),
3700 flags, nodeid);
3701 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003702}
Steven Rostedt85beb582010-11-24 16:23:34 -05003703EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003704#endif
3705
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003706static __always_inline void *
3707__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003708{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003709 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003710
3711 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003712 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3713 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003714 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003715}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003716
Li Zefan0bb38a52009-12-11 15:45:50 +08003717#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003718void *__kmalloc_node(size_t size, gfp_t flags, int node)
3719{
3720 return __do_kmalloc_node(size, flags, node,
3721 __builtin_return_address(0));
3722}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003723EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003724
3725void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003726 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003727{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003728 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003729}
3730EXPORT_SYMBOL(__kmalloc_node_track_caller);
3731#else
3732void *__kmalloc_node(size_t size, gfp_t flags, int node)
3733{
3734 return __do_kmalloc_node(size, flags, node, NULL);
3735}
3736EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003737#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003738#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739
3740/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003741 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003743 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003744 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003746static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3747 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003749 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003750 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003752 /* If you want to save a few bytes .text space: replace
3753 * __ with kmem_.
3754 * Then kmalloc uses the uninlined functions instead of the inline
3755 * functions.
3756 */
3757 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003758 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3759 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003760 ret = __cache_alloc(cachep, flags, caller);
3761
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003762 trace_kmalloc((unsigned long) caller, ret,
3763 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003764
3765 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003766}
3767
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003768
Li Zefan0bb38a52009-12-11 15:45:50 +08003769#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003770void *__kmalloc(size_t size, gfp_t flags)
3771{
Al Viro871751e2006-03-25 03:06:39 -08003772 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773}
3774EXPORT_SYMBOL(__kmalloc);
3775
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003776void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003777{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003778 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003779}
3780EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003781
3782#else
3783void *__kmalloc(size_t size, gfp_t flags)
3784{
3785 return __do_kmalloc(size, flags, NULL);
3786}
3787EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003788#endif
3789
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790/**
3791 * kmem_cache_free - Deallocate an object
3792 * @cachep: The cache the allocation was from.
3793 * @objp: The previously allocated object.
3794 *
3795 * Free an object which was previously allocated from this
3796 * cache.
3797 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003798void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799{
3800 unsigned long flags;
3801
3802 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003803 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003804 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3805 debug_check_no_obj_freed(objp, obj_size(cachep));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003806 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003808
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003809 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810}
3811EXPORT_SYMBOL(kmem_cache_free);
3812
3813/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814 * kfree - free previously allocated memory
3815 * @objp: pointer returned by kmalloc.
3816 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003817 * If @objp is NULL, no operation is performed.
3818 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 * Don't free memory not originally allocated by kmalloc()
3820 * or you will run into trouble.
3821 */
3822void kfree(const void *objp)
3823{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003824 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 unsigned long flags;
3826
Pekka Enberg2121db72009-03-25 11:05:57 +02003827 trace_kfree(_RET_IP_, objp);
3828
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003829 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830 return;
3831 local_irq_save(flags);
3832 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003833 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003834 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003835 debug_check_no_obj_freed(objp, obj_size(c));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003836 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837 local_irq_restore(flags);
3838}
3839EXPORT_SYMBOL(kfree);
3840
Pekka Enberg343e0d72006-02-01 03:05:50 -08003841unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003843 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844}
3845EXPORT_SYMBOL(kmem_cache_size);
3846
Christoph Lametere498be72005-09-09 13:03:32 -07003847/*
Simon Arlott183ff222007-10-20 01:27:18 +02003848 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003849 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003850static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003851{
3852 int node;
3853 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003854 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003855 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003856
Mel Gorman9c09a952008-01-24 05:49:54 -08003857 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003858
Paul Menage3395ee02006-12-06 20:32:16 -08003859 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003860 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003861 if (!new_alien)
3862 goto fail;
3863 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003864
Eric Dumazet63109842007-05-06 14:49:28 -07003865 new_shared = NULL;
3866 if (cachep->shared) {
3867 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003868 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003869 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003870 if (!new_shared) {
3871 free_alien_cache(new_alien);
3872 goto fail;
3873 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003874 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003875
Andrew Mortona737b3e2006-03-22 00:08:11 -08003876 l3 = cachep->nodelists[node];
3877 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003878 struct array_cache *shared = l3->shared;
3879
Christoph Lametere498be72005-09-09 13:03:32 -07003880 spin_lock_irq(&l3->list_lock);
3881
Christoph Lametercafeb022006-03-25 03:06:46 -08003882 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003883 free_block(cachep, shared->entry,
3884 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003885
Christoph Lametercafeb022006-03-25 03:06:46 -08003886 l3->shared = new_shared;
3887 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003888 l3->alien = new_alien;
3889 new_alien = NULL;
3890 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003891 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003892 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003893 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003894 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003895 free_alien_cache(new_alien);
3896 continue;
3897 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003898 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003899 if (!l3) {
3900 free_alien_cache(new_alien);
3901 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003902 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003903 }
Christoph Lametere498be72005-09-09 13:03:32 -07003904
3905 kmem_list3_init(l3);
3906 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003907 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003908 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003909 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003910 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003911 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003912 cachep->nodelists[node] = l3;
3913 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003914 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003915
Andrew Mortona737b3e2006-03-22 00:08:11 -08003916fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003917 if (!cachep->next.next) {
3918 /* Cache is not active yet. Roll back what we did */
3919 node--;
3920 while (node >= 0) {
3921 if (cachep->nodelists[node]) {
3922 l3 = cachep->nodelists[node];
3923
3924 kfree(l3->shared);
3925 free_alien_cache(l3->alien);
3926 kfree(l3);
3927 cachep->nodelists[node] = NULL;
3928 }
3929 node--;
3930 }
3931 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003932 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003933}
3934
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003936 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937 struct array_cache *new[NR_CPUS];
3938};
3939
3940static void do_ccupdate_local(void *info)
3941{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003942 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 struct array_cache *old;
3944
3945 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003946 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003947
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3949 new->new[smp_processor_id()] = old;
3950}
3951
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003952/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003953static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003954 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003956 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003957 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958
Pekka Enberg83b519e2009-06-10 19:40:04 +03003959 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003960 if (!new)
3961 return -ENOMEM;
3962
Christoph Lametere498be72005-09-09 13:03:32 -07003963 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003964 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003965 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003966 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003967 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003968 kfree(new->new[i]);
3969 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003970 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 }
3972 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003973 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003975 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003976
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 cachep->batchcount = batchcount;
3979 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003980 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981
Christoph Lametere498be72005-09-09 13:03:32 -07003982 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003983 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 if (!ccold)
3985 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003986 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3987 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3988 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 kfree(ccold);
3990 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003991 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003992 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993}
3994
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003995/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003996static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997{
3998 int err;
3999 int limit, shared;
4000
Andrew Mortona737b3e2006-03-22 00:08:11 -08004001 /*
4002 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 * - create a LIFO ordering, i.e. return objects that are cache-warm
4004 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004005 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 * bufctl chains: array operations are cheaper.
4007 * The numbers are guessed, we should auto-tune as described by
4008 * Bonwick.
4009 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004010 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004012 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004014 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004016 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 limit = 54;
4018 else
4019 limit = 120;
4020
Andrew Mortona737b3e2006-03-22 00:08:11 -08004021 /*
4022 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 * allocation behaviour: Most allocs on one cpu, most free operations
4024 * on another cpu. For these cases, an efficient object passing between
4025 * cpus is necessary. This is provided by a shared array. The array
4026 * replaces Bonwick's magazine layer.
4027 * On uniprocessor, it's functionally equivalent (but less efficient)
4028 * to a larger limit. Thus disabled by default.
4029 */
4030 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004031 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033
4034#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004035 /*
4036 * With debugging enabled, large batchcount lead to excessively long
4037 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 */
4039 if (limit > 32)
4040 limit = 32;
4041#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004042 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043 if (err)
4044 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004045 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004046 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047}
4048
Christoph Lameter1b552532006-03-22 00:09:07 -08004049/*
4050 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004051 * necessary. Note that the l3 listlock also protects the array_cache
4052 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004053 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004054static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004055 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056{
4057 int tofree;
4058
Christoph Lameter1b552532006-03-22 00:09:07 -08004059 if (!ac || !ac->avail)
4060 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 if (ac->touched && !force) {
4062 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004063 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004064 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004065 if (ac->avail) {
4066 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4067 if (tofree > ac->avail)
4068 tofree = (ac->avail + 1) / 2;
4069 free_block(cachep, ac->entry, tofree, node);
4070 ac->avail -= tofree;
4071 memmove(ac->entry, &(ac->entry[tofree]),
4072 sizeof(void *) * ac->avail);
4073 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004074 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 }
4076}
4077
4078/**
4079 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004080 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 *
4082 * Called from workqueue/eventd every few seconds.
4083 * Purpose:
4084 * - clear the per-cpu caches for this CPU.
4085 * - return freeable pages to the main free memory pool.
4086 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004087 * If we cannot acquire the cache chain mutex then just give up - we'll try
4088 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004090static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004092 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004093 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004094 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004095 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004097 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004099 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004101 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102 check_irq_on();
4103
Christoph Lameter35386e32006-03-22 00:09:05 -08004104 /*
4105 * We only take the l3 lock if absolutely necessary and we
4106 * have established with reasonable certainty that
4107 * we can do some work if the lock was obtained.
4108 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004109 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004110
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004111 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Christoph Lameteraab22072006-03-22 00:09:06 -08004113 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114
Christoph Lameter35386e32006-03-22 00:09:05 -08004115 /*
4116 * These are racy checks but it does not matter
4117 * if we skip one check or scan twice.
4118 */
Christoph Lametere498be72005-09-09 13:03:32 -07004119 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004120 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Christoph Lametere498be72005-09-09 13:03:32 -07004122 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123
Christoph Lameteraab22072006-03-22 00:09:06 -08004124 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125
Christoph Lametered11d9e2006-06-30 01:55:45 -07004126 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004127 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004128 else {
4129 int freed;
4130
4131 freed = drain_freelist(searchp, l3, (l3->free_limit +
4132 5 * searchp->num - 1) / (5 * searchp->num));
4133 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004135next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136 cond_resched();
4137 }
4138 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004139 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004140 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004141out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004142 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004143 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144}
4145
Linus Torvalds158a9622008-01-02 13:04:48 -08004146#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147
Pekka Enberg85289f92006-01-08 01:00:36 -08004148static void print_slabinfo_header(struct seq_file *m)
4149{
4150 /*
4151 * Output format version, so at least we can change it
4152 * without _too_ many complaints.
4153 */
4154#if STATS
4155 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4156#else
4157 seq_puts(m, "slabinfo - version: 2.1\n");
4158#endif
4159 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4160 "<objperslab> <pagesperslab>");
4161 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4162 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4163#if STATS
4164 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004165 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004166 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4167#endif
4168 seq_putc(m, '\n');
4169}
4170
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171static void *s_start(struct seq_file *m, loff_t *pos)
4172{
4173 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004175 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004176 if (!n)
4177 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004178
4179 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180}
4181
4182static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4183{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004184 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185}
4186
4187static void s_stop(struct seq_file *m, void *p)
4188{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004189 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190}
4191
4192static int s_show(struct seq_file *m, void *p)
4193{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004194 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004195 struct slab *slabp;
4196 unsigned long active_objs;
4197 unsigned long num_objs;
4198 unsigned long active_slabs = 0;
4199 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004200 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004202 int node;
4203 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 active_objs = 0;
4206 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004207 for_each_online_node(node) {
4208 l3 = cachep->nodelists[node];
4209 if (!l3)
4210 continue;
4211
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004212 check_irq_on();
4213 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004214
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004215 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004216 if (slabp->inuse != cachep->num && !error)
4217 error = "slabs_full accounting error";
4218 active_objs += cachep->num;
4219 active_slabs++;
4220 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004221 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004222 if (slabp->inuse == cachep->num && !error)
4223 error = "slabs_partial inuse accounting error";
4224 if (!slabp->inuse && !error)
4225 error = "slabs_partial/inuse accounting error";
4226 active_objs += slabp->inuse;
4227 active_slabs++;
4228 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004229 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004230 if (slabp->inuse && !error)
4231 error = "slabs_free/inuse accounting error";
4232 num_slabs++;
4233 }
4234 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004235 if (l3->shared)
4236 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004237
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004238 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004240 num_slabs += active_slabs;
4241 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004242 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243 error = "free_objects accounting error";
4244
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004245 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246 if (error)
4247 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4248
4249 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004250 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004251 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004253 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004254 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004255 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004257 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258 unsigned long high = cachep->high_mark;
4259 unsigned long allocs = cachep->num_allocations;
4260 unsigned long grown = cachep->grown;
4261 unsigned long reaped = cachep->reaped;
4262 unsigned long errors = cachep->errors;
4263 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004265 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004266 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267
Joe Perchese92dd4f2010-03-26 19:27:58 -07004268 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4269 "%4lu %4lu %4lu %4lu %4lu",
4270 allocs, high, grown,
4271 reaped, errors, max_freeable, node_allocs,
4272 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273 }
4274 /* cpu stats */
4275 {
4276 unsigned long allochit = atomic_read(&cachep->allochit);
4277 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4278 unsigned long freehit = atomic_read(&cachep->freehit);
4279 unsigned long freemiss = atomic_read(&cachep->freemiss);
4280
4281 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004282 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 }
4284#endif
4285 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 return 0;
4287}
4288
4289/*
4290 * slabinfo_op - iterator that generates /proc/slabinfo
4291 *
4292 * Output layout:
4293 * cache-name
4294 * num-active-objs
4295 * total-objs
4296 * object size
4297 * num-active-slabs
4298 * total-slabs
4299 * num-pages-per-slab
4300 * + further values on SMP and with statistics enabled
4301 */
4302
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004303static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004304 .start = s_start,
4305 .next = s_next,
4306 .stop = s_stop,
4307 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308};
4309
4310#define MAX_SLABINFO_WRITE 128
4311/**
4312 * slabinfo_write - Tuning for the slab allocator
4313 * @file: unused
4314 * @buffer: user buffer
4315 * @count: data length
4316 * @ppos: unused
4317 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004318static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004319 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004321 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004323 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004324
Linus Torvalds1da177e2005-04-16 15:20:36 -07004325 if (count > MAX_SLABINFO_WRITE)
4326 return -EINVAL;
4327 if (copy_from_user(&kbuf, buffer, count))
4328 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004329 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004330
4331 tmp = strchr(kbuf, ' ');
4332 if (!tmp)
4333 return -EINVAL;
4334 *tmp = '\0';
4335 tmp++;
4336 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4337 return -EINVAL;
4338
4339 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004340 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004341 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004342 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004343 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004344 if (limit < 1 || batchcount < 1 ||
4345 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004346 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004347 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004348 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004349 batchcount, shared,
4350 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351 }
4352 break;
4353 }
4354 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004355 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 if (res >= 0)
4357 res = count;
4358 return res;
4359}
Al Viro871751e2006-03-25 03:06:39 -08004360
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004361static int slabinfo_open(struct inode *inode, struct file *file)
4362{
4363 return seq_open(file, &slabinfo_op);
4364}
4365
4366static const struct file_operations proc_slabinfo_operations = {
4367 .open = slabinfo_open,
4368 .read = seq_read,
4369 .write = slabinfo_write,
4370 .llseek = seq_lseek,
4371 .release = seq_release,
4372};
4373
Al Viro871751e2006-03-25 03:06:39 -08004374#ifdef CONFIG_DEBUG_SLAB_LEAK
4375
4376static void *leaks_start(struct seq_file *m, loff_t *pos)
4377{
Al Viro871751e2006-03-25 03:06:39 -08004378 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004379 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004380}
4381
4382static inline int add_caller(unsigned long *n, unsigned long v)
4383{
4384 unsigned long *p;
4385 int l;
4386 if (!v)
4387 return 1;
4388 l = n[1];
4389 p = n + 2;
4390 while (l) {
4391 int i = l/2;
4392 unsigned long *q = p + 2 * i;
4393 if (*q == v) {
4394 q[1]++;
4395 return 1;
4396 }
4397 if (*q > v) {
4398 l = i;
4399 } else {
4400 p = q + 2;
4401 l -= i + 1;
4402 }
4403 }
4404 if (++n[1] == n[0])
4405 return 0;
4406 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4407 p[0] = v;
4408 p[1] = 1;
4409 return 1;
4410}
4411
4412static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4413{
4414 void *p;
4415 int i;
4416 if (n[0] == n[1])
4417 return;
4418 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4419 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4420 continue;
4421 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4422 return;
4423 }
4424}
4425
4426static void show_symbol(struct seq_file *m, unsigned long address)
4427{
4428#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004429 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004430 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004431
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004432 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004433 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004434 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004435 seq_printf(m, " [%s]", modname);
4436 return;
4437 }
4438#endif
4439 seq_printf(m, "%p", (void *)address);
4440}
4441
4442static int leaks_show(struct seq_file *m, void *p)
4443{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004444 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004445 struct slab *slabp;
4446 struct kmem_list3 *l3;
4447 const char *name;
4448 unsigned long *n = m->private;
4449 int node;
4450 int i;
4451
4452 if (!(cachep->flags & SLAB_STORE_USER))
4453 return 0;
4454 if (!(cachep->flags & SLAB_RED_ZONE))
4455 return 0;
4456
4457 /* OK, we can do it */
4458
4459 n[1] = 0;
4460
4461 for_each_online_node(node) {
4462 l3 = cachep->nodelists[node];
4463 if (!l3)
4464 continue;
4465
4466 check_irq_on();
4467 spin_lock_irq(&l3->list_lock);
4468
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004469 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004470 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004471 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004472 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004473 spin_unlock_irq(&l3->list_lock);
4474 }
4475 name = cachep->name;
4476 if (n[0] == n[1]) {
4477 /* Increase the buffer size */
4478 mutex_unlock(&cache_chain_mutex);
4479 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4480 if (!m->private) {
4481 /* Too bad, we are really out */
4482 m->private = n;
4483 mutex_lock(&cache_chain_mutex);
4484 return -ENOMEM;
4485 }
4486 *(unsigned long *)m->private = n[0] * 2;
4487 kfree(n);
4488 mutex_lock(&cache_chain_mutex);
4489 /* Now make sure this entry will be retried */
4490 m->count = m->size;
4491 return 0;
4492 }
4493 for (i = 0; i < n[1]; i++) {
4494 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4495 show_symbol(m, n[2*i+2]);
4496 seq_putc(m, '\n');
4497 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004498
Al Viro871751e2006-03-25 03:06:39 -08004499 return 0;
4500}
4501
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004502static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004503 .start = leaks_start,
4504 .next = s_next,
4505 .stop = s_stop,
4506 .show = leaks_show,
4507};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004508
4509static int slabstats_open(struct inode *inode, struct file *file)
4510{
4511 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4512 int ret = -ENOMEM;
4513 if (n) {
4514 ret = seq_open(file, &slabstats_op);
4515 if (!ret) {
4516 struct seq_file *m = file->private_data;
4517 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4518 m->private = n;
4519 n = NULL;
4520 }
4521 kfree(n);
4522 }
4523 return ret;
4524}
4525
4526static const struct file_operations proc_slabstats_operations = {
4527 .open = slabstats_open,
4528 .read = seq_read,
4529 .llseek = seq_lseek,
4530 .release = seq_release_private,
4531};
Al Viro871751e2006-03-25 03:06:39 -08004532#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004533
4534static int __init slab_proc_init(void)
4535{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004536 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004537#ifdef CONFIG_DEBUG_SLAB_LEAK
4538 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4539#endif
4540 return 0;
4541}
4542module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004543#endif
4544
Manfred Spraul00e145b2005-09-03 15:55:07 -07004545/**
4546 * ksize - get the actual amount of memory allocated for a given object
4547 * @objp: Pointer to the object
4548 *
4549 * kmalloc may internally round up allocations and return more memory
4550 * than requested. ksize() can be used to determine the actual amount of
4551 * memory allocated. The caller may use this additional memory, even though
4552 * a smaller amount of memory was initially specified with the kmalloc call.
4553 * The caller must guarantee that objp points to a valid object previously
4554 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4555 * must not be freed during the duration of the call.
4556 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004557size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004559 BUG_ON(!objp);
4560 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004561 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004562
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004563 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004564}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004565EXPORT_SYMBOL(ksize);