blob: d96e223de775378f78d7597d8c2d87a9e24e70a2 [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 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800577static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800578 .batchcount = 1,
579 .limit = BOOT_CPUCACHE_ENTRIES,
580 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800581 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800582 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583};
584
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700585#define BAD_ALIEN_MAGIC 0x01020304ul
586
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 * chicken and egg problem: delay the per-cpu array allocation
589 * until the general caches are up.
590 */
591static enum {
592 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700593 PARTIAL_AC,
594 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300595 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 FULL
597} g_cpucache_up;
598
Mike Kravetz39d24e62006-05-15 09:44:13 -0700599/*
600 * used by boot code to determine if it can use slab based allocator
601 */
602int slab_is_available(void)
603{
Pekka Enberg8429db52009-06-12 15:58:59 +0300604 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700605}
606
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200607#ifdef CONFIG_LOCKDEP
608
609/*
610 * Slab sometimes uses the kmalloc slabs to store the slab headers
611 * for other slabs "off slab".
612 * The locking for this is tricky in that it nests within the locks
613 * of all other slabs in a few places; to deal with this special
614 * locking we put on-slab caches into a separate lock-class.
615 *
616 * We set lock class for alien array caches which are up during init.
617 * The lock annotation will be lost if all cpus of a node goes down and
618 * then comes back up during hotplug
619 */
620static struct lock_class_key on_slab_l3_key;
621static struct lock_class_key on_slab_alc_key;
622
623static void init_node_lock_keys(int q)
624{
625 struct cache_sizes *s = malloc_sizes;
626
627 if (g_cpucache_up != FULL)
628 return;
629
630 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
631 struct array_cache **alc;
632 struct kmem_list3 *l3;
633 int r;
634
635 l3 = s->cs_cachep->nodelists[q];
636 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200637 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200638 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
639 alc = l3->alien;
640 /*
641 * FIXME: This check for BAD_ALIEN_MAGIC
642 * should go away when common slab code is taught to
643 * work even without alien caches.
644 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
645 * for alloc_alien_cache,
646 */
647 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
Pekka Enberg00afa752009-12-27 14:33:14 +0200648 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200649 for_each_node(r) {
650 if (alc[r])
651 lockdep_set_class(&alc[r]->lock,
652 &on_slab_alc_key);
653 }
654 }
655}
656
657static inline void init_lock_keys(void)
658{
659 int node;
660
661 for_each_node(node)
662 init_node_lock_keys(node);
663}
664#else
665static void init_node_lock_keys(int q)
666{
667}
668
669static inline void init_lock_keys(void)
670{
671}
672#endif
673
674/*
675 * Guard access to the cache-chain.
676 */
677static DEFINE_MUTEX(cache_chain_mutex);
678static struct list_head cache_chain;
679
Tejun Heo1871e522009-10-29 22:34:13 +0900680static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
Pekka Enberg343e0d72006-02-01 03:05:50 -0800682static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683{
684 return cachep->array[smp_processor_id()];
685}
686
Andrew Mortona737b3e2006-03-22 00:08:11 -0800687static inline struct kmem_cache *__find_general_cachep(size_t size,
688 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
690 struct cache_sizes *csizep = malloc_sizes;
691
692#if DEBUG
693 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
696 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700699 if (!size)
700 return ZERO_SIZE_PTR;
701
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 while (size > csizep->cs_size)
703 csizep++;
704
705 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700706 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 * has cs_{dma,}cachep==NULL. Thus no special case
708 * for large kmalloc calls required.
709 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800710#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 if (unlikely(gfpflags & GFP_DMA))
712 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800713#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 return csizep->cs_cachep;
715}
716
Adrian Bunkb2213852006-09-25 23:31:02 -0700717static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700718{
719 return __find_general_cachep(size, gfpflags);
720}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700721
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800722static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800724 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
725}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726
Andrew Mortona737b3e2006-03-22 00:08:11 -0800727/*
728 * Calculate the number of objects and left-over bytes for a given buffer size.
729 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800730static void cache_estimate(unsigned long gfporder, size_t buffer_size,
731 size_t align, int flags, size_t *left_over,
732 unsigned int *num)
733{
734 int nr_objs;
735 size_t mgmt_size;
736 size_t slab_size = PAGE_SIZE << gfporder;
737
738 /*
739 * The slab management structure can be either off the slab or
740 * on it. For the latter case, the memory allocated for a
741 * slab is used for:
742 *
743 * - The struct slab
744 * - One kmem_bufctl_t for each object
745 * - Padding to respect alignment of @align
746 * - @buffer_size bytes for each object
747 *
748 * If the slab management structure is off the slab, then the
749 * alignment will already be calculated into the size. Because
750 * the slabs are all pages aligned, the objects will be at the
751 * correct alignment when allocated.
752 */
753 if (flags & CFLGS_OFF_SLAB) {
754 mgmt_size = 0;
755 nr_objs = slab_size / buffer_size;
756
757 if (nr_objs > SLAB_LIMIT)
758 nr_objs = SLAB_LIMIT;
759 } else {
760 /*
761 * Ignore padding for the initial guess. The padding
762 * is at most @align-1 bytes, and @buffer_size is at
763 * least @align. In the worst case, this result will
764 * be one greater than the number of objects that fit
765 * into the memory allocation when taking the padding
766 * into account.
767 */
768 nr_objs = (slab_size - sizeof(struct slab)) /
769 (buffer_size + sizeof(kmem_bufctl_t));
770
771 /*
772 * This calculated number will be either the right
773 * amount, or one greater than what we want.
774 */
775 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
776 > slab_size)
777 nr_objs--;
778
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
781
782 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800784 *num = nr_objs;
785 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786}
787
Harvey Harrisond40cee22008-04-30 00:55:07 -0700788#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
Andrew Mortona737b3e2006-03-22 00:08:11 -0800790static void __slab_error(const char *function, struct kmem_cache *cachep,
791 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792{
793 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800794 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 dump_stack();
796}
797
Paul Menage3395ee02006-12-06 20:32:16 -0800798/*
799 * By default on NUMA we use alien caches to stage the freeing of
800 * objects allocated from other nodes. This causes massive memory
801 * inefficiencies when using fake NUMA setup to split memory into a
802 * large number of small nodes, so it can be disabled on the command
803 * line
804 */
805
806static int use_alien_caches __read_mostly = 1;
807static int __init noaliencache_setup(char *s)
808{
809 use_alien_caches = 0;
810 return 1;
811}
812__setup("noaliencache", noaliencache_setup);
813
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800814#ifdef CONFIG_NUMA
815/*
816 * Special reaping functions for NUMA systems called from cache_reap().
817 * These take care of doing round robin flushing of alien caches (containing
818 * objects freed on different nodes from which they were allocated) and the
819 * flushing of remote pcps by calling drain_node_pages.
820 */
Tejun Heo1871e522009-10-29 22:34:13 +0900821static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800822
823static void init_reap_node(int cpu)
824{
825 int node;
826
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700827 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800828 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800829 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800830
Tejun Heo1871e522009-10-29 22:34:13 +0900831 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800832}
833
834static void next_reap_node(void)
835{
Christoph Lameter909ea962010-12-08 16:22:55 +0100836 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800837
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800838 node = next_node(node, node_online_map);
839 if (unlikely(node >= MAX_NUMNODES))
840 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100841 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800842}
843
844#else
845#define init_reap_node(cpu) do { } while (0)
846#define next_reap_node(void) do { } while (0)
847#endif
848
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849/*
850 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
851 * via the workqueue/eventd.
852 * Add the CPU number into the expiration time to minimize the possibility of
853 * the CPUs getting into lockstep and contending for the global cache chain
854 * lock.
855 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700856static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
Tejun Heo1871e522009-10-29 22:34:13 +0900858 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859
860 /*
861 * When this gets called from do_initcalls via cpucache_init(),
862 * init_workqueues() has already run, so keventd will be setup
863 * at that time.
864 */
David Howells52bad642006-11-22 14:54:01 +0000865 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800866 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700867 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800868 schedule_delayed_work_on(cpu, reap_work,
869 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 }
871}
872
Christoph Lametere498be72005-09-09 13:03:32 -0700873static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300874 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800876 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 struct array_cache *nc = NULL;
878
Pekka Enberg83b519e2009-06-10 19:40:04 +0300879 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100880 /*
881 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300882 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100883 * cache the pointers are not cleared and they could be counted as
884 * valid references during a kmemleak scan. Therefore, kmemleak must
885 * not scan such objects.
886 */
887 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 if (nc) {
889 nc->avail = 0;
890 nc->limit = entries;
891 nc->batchcount = batchcount;
892 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700893 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 }
895 return nc;
896}
897
Christoph Lameter3ded1752006-03-25 03:06:44 -0800898/*
899 * Transfer objects in one arraycache to another.
900 * Locking must be handled by the caller.
901 *
902 * Return the number of entries transferred.
903 */
904static int transfer_objects(struct array_cache *to,
905 struct array_cache *from, unsigned int max)
906{
907 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700908 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800909
910 if (!nr)
911 return 0;
912
913 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
914 sizeof(void *) *nr);
915
916 from->avail -= nr;
917 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800918 return nr;
919}
920
Christoph Lameter765c4502006-09-27 01:50:08 -0700921#ifndef CONFIG_NUMA
922
923#define drain_alien_cache(cachep, alien) do { } while (0)
924#define reap_alien(cachep, l3) do { } while (0)
925
Pekka Enberg83b519e2009-06-10 19:40:04 +0300926static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700927{
928 return (struct array_cache **)BAD_ALIEN_MAGIC;
929}
930
931static inline void free_alien_cache(struct array_cache **ac_ptr)
932{
933}
934
935static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
936{
937 return 0;
938}
939
940static inline void *alternate_node_alloc(struct kmem_cache *cachep,
941 gfp_t flags)
942{
943 return NULL;
944}
945
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800946static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700947 gfp_t flags, int nodeid)
948{
949 return NULL;
950}
951
952#else /* CONFIG_NUMA */
953
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800954static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800955static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800956
Pekka Enberg83b519e2009-06-10 19:40:04 +0300957static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700958{
959 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800960 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700961 int i;
962
963 if (limit > 1)
964 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800965 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700966 if (ac_ptr) {
967 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800968 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700969 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300970 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700971 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800972 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700973 kfree(ac_ptr[i]);
974 kfree(ac_ptr);
975 return NULL;
976 }
977 }
978 }
979 return ac_ptr;
980}
981
Pekka Enberg5295a742006-02-01 03:05:48 -0800982static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700983{
984 int i;
985
986 if (!ac_ptr)
987 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700988 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800989 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700990 kfree(ac_ptr);
991}
992
Pekka Enberg343e0d72006-02-01 03:05:50 -0800993static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800994 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700995{
996 struct kmem_list3 *rl3 = cachep->nodelists[node];
997
998 if (ac->avail) {
999 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001000 /*
1001 * Stuff objects into the remote nodes shared array first.
1002 * That way we could avoid the overhead of putting the objects
1003 * into the free lists and getting them back later.
1004 */
shin, jacob693f7d32006-04-28 10:54:37 -05001005 if (rl3->shared)
1006 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001007
Christoph Lameterff694162005-09-22 21:44:02 -07001008 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001009 ac->avail = 0;
1010 spin_unlock(&rl3->list_lock);
1011 }
1012}
1013
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001014/*
1015 * Called from cache_reap() to regularly drain alien caches round robin.
1016 */
1017static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1018{
Christoph Lameter909ea962010-12-08 16:22:55 +01001019 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001020
1021 if (l3->alien) {
1022 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001023
1024 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001025 __drain_alien_cache(cachep, ac, node);
1026 spin_unlock_irq(&ac->lock);
1027 }
1028 }
1029}
1030
Andrew Mortona737b3e2006-03-22 00:08:11 -08001031static void drain_alien_cache(struct kmem_cache *cachep,
1032 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001033{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001034 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001035 struct array_cache *ac;
1036 unsigned long flags;
1037
1038 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001039 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001040 if (ac) {
1041 spin_lock_irqsave(&ac->lock, flags);
1042 __drain_alien_cache(cachep, ac, i);
1043 spin_unlock_irqrestore(&ac->lock, flags);
1044 }
1045 }
1046}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001047
Ingo Molnar873623d2006-07-13 14:44:38 +02001048static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001049{
1050 struct slab *slabp = virt_to_slab(objp);
1051 int nodeid = slabp->nodeid;
1052 struct kmem_list3 *l3;
1053 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001054 int node;
1055
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001056 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001057
1058 /*
1059 * Make sure we are not freeing a object from another node to the array
1060 * cache on this cpu.
1061 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001062 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001063 return 0;
1064
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001065 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001066 STATS_INC_NODEFREES(cachep);
1067 if (l3->alien && l3->alien[nodeid]) {
1068 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001069 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001070 if (unlikely(alien->avail == alien->limit)) {
1071 STATS_INC_ACOVERFLOW(cachep);
1072 __drain_alien_cache(cachep, alien, nodeid);
1073 }
1074 alien->entry[alien->avail++] = objp;
1075 spin_unlock(&alien->lock);
1076 } else {
1077 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1078 free_block(cachep, &objp, 1, nodeid);
1079 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1080 }
1081 return 1;
1082}
Christoph Lametere498be72005-09-09 13:03:32 -07001083#endif
1084
David Rientjes8f9f8d92010-03-27 19:40:47 -07001085/*
1086 * Allocates and initializes nodelists for a node on each slab cache, used for
1087 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1088 * will be allocated off-node since memory is not yet online for the new node.
1089 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1090 * already in use.
1091 *
1092 * Must hold cache_chain_mutex.
1093 */
1094static int init_cache_nodelists_node(int node)
1095{
1096 struct kmem_cache *cachep;
1097 struct kmem_list3 *l3;
1098 const int memsize = sizeof(struct kmem_list3);
1099
1100 list_for_each_entry(cachep, &cache_chain, next) {
1101 /*
1102 * Set up the size64 kmemlist for cpu before we can
1103 * begin anything. Make sure some other cpu on this
1104 * node has not already allocated this
1105 */
1106 if (!cachep->nodelists[node]) {
1107 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1108 if (!l3)
1109 return -ENOMEM;
1110 kmem_list3_init(l3);
1111 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1112 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1113
1114 /*
1115 * The l3s don't come and go as CPUs come and
1116 * go. cache_chain_mutex is sufficient
1117 * protection here.
1118 */
1119 cachep->nodelists[node] = l3;
1120 }
1121
1122 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1123 cachep->nodelists[node]->free_limit =
1124 (1 + nr_cpus_node(node)) *
1125 cachep->batchcount + cachep->num;
1126 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1127 }
1128 return 0;
1129}
1130
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001131static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001133 struct kmem_cache *cachep;
1134 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001135 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301136 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001137
1138 list_for_each_entry(cachep, &cache_chain, next) {
1139 struct array_cache *nc;
1140 struct array_cache *shared;
1141 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001142
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001143 /* cpu is dead; no one can alloc from it. */
1144 nc = cachep->array[cpu];
1145 cachep->array[cpu] = NULL;
1146 l3 = cachep->nodelists[node];
1147
1148 if (!l3)
1149 goto free_array_cache;
1150
1151 spin_lock_irq(&l3->list_lock);
1152
1153 /* Free limit for this kmem_list3 */
1154 l3->free_limit -= cachep->batchcount;
1155 if (nc)
1156 free_block(cachep, nc->entry, nc->avail, node);
1157
Rusty Russell58463c12009-12-17 11:43:12 -06001158 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001159 spin_unlock_irq(&l3->list_lock);
1160 goto free_array_cache;
1161 }
1162
1163 shared = l3->shared;
1164 if (shared) {
1165 free_block(cachep, shared->entry,
1166 shared->avail, node);
1167 l3->shared = NULL;
1168 }
1169
1170 alien = l3->alien;
1171 l3->alien = NULL;
1172
1173 spin_unlock_irq(&l3->list_lock);
1174
1175 kfree(shared);
1176 if (alien) {
1177 drain_alien_cache(cachep, alien);
1178 free_alien_cache(alien);
1179 }
1180free_array_cache:
1181 kfree(nc);
1182 }
1183 /*
1184 * In the previous loop, all the objects were freed to
1185 * the respective cache's slabs, now we can go ahead and
1186 * shrink each nodelist to its limit.
1187 */
1188 list_for_each_entry(cachep, &cache_chain, next) {
1189 l3 = cachep->nodelists[node];
1190 if (!l3)
1191 continue;
1192 drain_freelist(cachep, l3, l3->free_objects);
1193 }
1194}
1195
1196static int __cpuinit cpuup_prepare(long cpu)
1197{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001198 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001199 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001200 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001201 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001203 /*
1204 * We need to do this right in the beginning since
1205 * alloc_arraycache's are going to use this list.
1206 * kmalloc_node allows us to add the slab to the right
1207 * kmem_list3 and not this cpu's kmem_list3
1208 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001209 err = init_cache_nodelists_node(node);
1210 if (err < 0)
1211 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001212
1213 /*
1214 * Now we can go ahead with allocating the shared arrays and
1215 * array caches
1216 */
1217 list_for_each_entry(cachep, &cache_chain, next) {
1218 struct array_cache *nc;
1219 struct array_cache *shared = NULL;
1220 struct array_cache **alien = NULL;
1221
1222 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001223 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001224 if (!nc)
1225 goto bad;
1226 if (cachep->shared) {
1227 shared = alloc_arraycache(node,
1228 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001229 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001230 if (!shared) {
1231 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001232 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001233 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001234 }
1235 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001236 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001237 if (!alien) {
1238 kfree(shared);
1239 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001240 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001241 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001242 }
1243 cachep->array[cpu] = nc;
1244 l3 = cachep->nodelists[node];
1245 BUG_ON(!l3);
1246
1247 spin_lock_irq(&l3->list_lock);
1248 if (!l3->shared) {
1249 /*
1250 * We are serialised from CPU_DEAD or
1251 * CPU_UP_CANCELLED by the cpucontrol lock
1252 */
1253 l3->shared = shared;
1254 shared = NULL;
1255 }
1256#ifdef CONFIG_NUMA
1257 if (!l3->alien) {
1258 l3->alien = alien;
1259 alien = NULL;
1260 }
1261#endif
1262 spin_unlock_irq(&l3->list_lock);
1263 kfree(shared);
1264 free_alien_cache(alien);
1265 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001266 init_node_lock_keys(node);
1267
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001268 return 0;
1269bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001270 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001271 return -ENOMEM;
1272}
1273
1274static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1275 unsigned long action, void *hcpu)
1276{
1277 long cpu = (long)hcpu;
1278 int err = 0;
1279
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001281 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001282 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001283 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001284 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001285 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 break;
1287 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001288 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 start_cpu_timer(cpu);
1290 break;
1291#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001292 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001293 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001294 /*
1295 * Shutdown cache reaper. Note that the cache_chain_mutex is
1296 * held so that if cache_reap() is invoked it cannot do
1297 * anything expensive but will only modify reap_work
1298 * and reschedule the timer.
1299 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001300 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001301 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001302 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001303 break;
1304 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001305 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001306 start_cpu_timer(cpu);
1307 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001309 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001310 /*
1311 * Even if all the cpus of a node are down, we don't free the
1312 * kmem_list3 of any cache. This to avoid a race between
1313 * cpu_down, and a kmalloc allocation from another cpu for
1314 * memory from the node of the cpu going down. The list3
1315 * structure is usually allocated from kmem_cache_create() and
1316 * gets destroyed at kmem_cache_destroy().
1317 */
Simon Arlott183ff222007-10-20 01:27:18 +02001318 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001319#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001321 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001322 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001323 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001324 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001327 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328}
1329
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001330static struct notifier_block __cpuinitdata cpucache_notifier = {
1331 &cpuup_callback, NULL, 0
1332};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333
David Rientjes8f9f8d92010-03-27 19:40:47 -07001334#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1335/*
1336 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1337 * Returns -EBUSY if all objects cannot be drained so that the node is not
1338 * removed.
1339 *
1340 * Must hold cache_chain_mutex.
1341 */
1342static int __meminit drain_cache_nodelists_node(int node)
1343{
1344 struct kmem_cache *cachep;
1345 int ret = 0;
1346
1347 list_for_each_entry(cachep, &cache_chain, next) {
1348 struct kmem_list3 *l3;
1349
1350 l3 = cachep->nodelists[node];
1351 if (!l3)
1352 continue;
1353
1354 drain_freelist(cachep, l3, l3->free_objects);
1355
1356 if (!list_empty(&l3->slabs_full) ||
1357 !list_empty(&l3->slabs_partial)) {
1358 ret = -EBUSY;
1359 break;
1360 }
1361 }
1362 return ret;
1363}
1364
1365static int __meminit slab_memory_callback(struct notifier_block *self,
1366 unsigned long action, void *arg)
1367{
1368 struct memory_notify *mnb = arg;
1369 int ret = 0;
1370 int nid;
1371
1372 nid = mnb->status_change_nid;
1373 if (nid < 0)
1374 goto out;
1375
1376 switch (action) {
1377 case MEM_GOING_ONLINE:
1378 mutex_lock(&cache_chain_mutex);
1379 ret = init_cache_nodelists_node(nid);
1380 mutex_unlock(&cache_chain_mutex);
1381 break;
1382 case MEM_GOING_OFFLINE:
1383 mutex_lock(&cache_chain_mutex);
1384 ret = drain_cache_nodelists_node(nid);
1385 mutex_unlock(&cache_chain_mutex);
1386 break;
1387 case MEM_ONLINE:
1388 case MEM_OFFLINE:
1389 case MEM_CANCEL_ONLINE:
1390 case MEM_CANCEL_OFFLINE:
1391 break;
1392 }
1393out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001394 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001395}
1396#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1397
Christoph Lametere498be72005-09-09 13:03:32 -07001398/*
1399 * swap the static kmem_list3 with kmalloced memory
1400 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001401static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1402 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001403{
1404 struct kmem_list3 *ptr;
1405
Pekka Enberg83b519e2009-06-10 19:40:04 +03001406 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001407 BUG_ON(!ptr);
1408
Christoph Lametere498be72005-09-09 13:03:32 -07001409 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001410 /*
1411 * Do not assume that spinlocks can be initialized via memcpy:
1412 */
1413 spin_lock_init(&ptr->list_lock);
1414
Christoph Lametere498be72005-09-09 13:03:32 -07001415 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1416 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001417}
1418
Andrew Mortona737b3e2006-03-22 00:08:11 -08001419/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001420 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1421 * size of kmem_list3.
1422 */
1423static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1424{
1425 int node;
1426
1427 for_each_online_node(node) {
1428 cachep->nodelists[node] = &initkmem_list3[index + node];
1429 cachep->nodelists[node]->next_reap = jiffies +
1430 REAPTIMEOUT_LIST3 +
1431 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1432 }
1433}
1434
1435/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001436 * Initialisation. Called after the page allocator have been initialised and
1437 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 */
1439void __init kmem_cache_init(void)
1440{
1441 size_t left_over;
1442 struct cache_sizes *sizes;
1443 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001444 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001445 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001446 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001447
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001448 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001449 use_alien_caches = 0;
1450
Christoph Lametere498be72005-09-09 13:03:32 -07001451 for (i = 0; i < NUM_INIT_LISTS; i++) {
1452 kmem_list3_init(&initkmem_list3[i]);
1453 if (i < MAX_NUMNODES)
1454 cache_cache.nodelists[i] = NULL;
1455 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001456 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
1458 /*
1459 * Fragmentation resistance on low memory - only use bigger
1460 * page orders on machines with more than 32MB of memory.
1461 */
Jan Beulich44813742009-09-21 17:03:05 -07001462 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1464
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 /* Bootstrap is tricky, because several objects are allocated
1466 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001467 * 1) initialize the cache_cache cache: it contains the struct
1468 * kmem_cache structures of all caches, except cache_cache itself:
1469 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001470 * Initially an __init data area is used for the head array and the
1471 * kmem_list3 structures, it's replaced with a kmalloc allocated
1472 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001474 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001475 * An __init data area is used for the head array.
1476 * 3) Create the remaining kmalloc caches, with minimally sized
1477 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 * 4) Replace the __init data head arrays for cache_cache and the first
1479 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001480 * 5) Replace the __init data for kmem_list3 for cache_cache and
1481 * the other cache's with kmalloc allocated memory.
1482 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 */
1484
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001485 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 INIT_LIST_HEAD(&cache_chain);
1489 list_add(&cache_cache.next, &cache_chain);
1490 cache_cache.colour_off = cache_line_size();
1491 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001492 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
Eric Dumazet8da34302007-05-06 14:49:29 -07001494 /*
1495 * struct kmem_cache size depends on nr_node_ids, which
1496 * can be less than MAX_NUMNODES.
1497 */
1498 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1499 nr_node_ids * sizeof(struct kmem_list3 *);
1500#if DEBUG
1501 cache_cache.obj_size = cache_cache.buffer_size;
1502#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001503 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1504 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001505 cache_cache.reciprocal_buffer_size =
1506 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Jack Steiner07ed76b2006-03-07 21:55:46 -08001508 for (order = 0; order < MAX_ORDER; order++) {
1509 cache_estimate(order, cache_cache.buffer_size,
1510 cache_line_size(), 0, &left_over, &cache_cache.num);
1511 if (cache_cache.num)
1512 break;
1513 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001514 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001515 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001516 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001517 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1518 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519
1520 /* 2+3) create the kmalloc caches */
1521 sizes = malloc_sizes;
1522 names = cache_names;
1523
Andrew Mortona737b3e2006-03-22 00:08:11 -08001524 /*
1525 * Initialize the caches that provide memory for the array cache and the
1526 * kmem_list3 structures first. Without this, further allocations will
1527 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001528 */
1529
1530 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001531 sizes[INDEX_AC].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001534 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001535
Andrew Mortona737b3e2006-03-22 00:08:11 -08001536 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001537 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001538 kmem_cache_create(names[INDEX_L3].name,
1539 sizes[INDEX_L3].cs_size,
1540 ARCH_KMALLOC_MINALIGN,
1541 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001542 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001543 }
Christoph Lametere498be72005-09-09 13:03:32 -07001544
Ingo Molnare0a42722006-06-23 02:03:46 -07001545 slab_early_init = 0;
1546
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001548 /*
1549 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 * This should be particularly beneficial on SMP boxes, as it
1551 * eliminates "false sharing".
1552 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001553 * allow tighter packing of the smaller caches.
1554 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001555 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001556 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 sizes->cs_size,
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001560 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001561 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001562#ifdef CONFIG_ZONE_DMA
1563 sizes->cs_dmacachep = kmem_cache_create(
1564 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001565 sizes->cs_size,
1566 ARCH_KMALLOC_MINALIGN,
1567 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1568 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001569 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001570#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 sizes++;
1572 names++;
1573 }
1574 /* 4) Replace the bootstrap head arrays */
1575 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001576 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001577
Pekka Enberg83b519e2009-06-10 19:40:04 +03001578 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001579
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001580 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1581 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001582 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001583 /*
1584 * Do not assume that spinlocks can be initialized via memcpy:
1585 */
1586 spin_lock_init(&ptr->lock);
1587
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001589
Pekka Enberg83b519e2009-06-10 19:40:04 +03001590 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001591
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001592 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001593 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001594 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001595 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001596 /*
1597 * Do not assume that spinlocks can be initialized via memcpy:
1598 */
1599 spin_lock_init(&ptr->lock);
1600
Christoph Lametere498be72005-09-09 13:03:32 -07001601 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001602 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 }
Christoph Lametere498be72005-09-09 13:03:32 -07001604 /* 5) Replace the bootstrap kmem_list3's */
1605 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001606 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607
Mel Gorman9c09a952008-01-24 05:49:54 -08001608 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001609 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001610
Christoph Lametere498be72005-09-09 13:03:32 -07001611 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001612 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001613
1614 if (INDEX_AC != INDEX_L3) {
1615 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001616 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001617 }
1618 }
1619 }
1620
Pekka Enberg8429db52009-06-12 15:58:59 +03001621 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001622}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001623
Pekka Enberg8429db52009-06-12 15:58:59 +03001624void __init kmem_cache_init_late(void)
1625{
1626 struct kmem_cache *cachep;
1627
Pekka Enberg8429db52009-06-12 15:58:59 +03001628 /* 6) resize the head arrays to their final sizes */
1629 mutex_lock(&cache_chain_mutex);
1630 list_for_each_entry(cachep, &cache_chain, next)
1631 if (enable_cpucache(cachep, GFP_NOWAIT))
1632 BUG();
1633 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001634
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 /* Done! */
1636 g_cpucache_up = FULL;
1637
Pekka Enbergec5a36f2009-06-29 09:57:10 +03001638 /* Annotate slab for lockdep -- annotate the malloc caches */
1639 init_lock_keys();
1640
Andrew Mortona737b3e2006-03-22 00:08:11 -08001641 /*
1642 * Register a cpu startup notifier callback that initializes
1643 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 */
1645 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646
David Rientjes8f9f8d92010-03-27 19:40:47 -07001647#ifdef CONFIG_NUMA
1648 /*
1649 * Register a memory hotplug callback that initializes and frees
1650 * nodelists.
1651 */
1652 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1653#endif
1654
Andrew Mortona737b3e2006-03-22 00:08:11 -08001655 /*
1656 * The reap timers are started later, with a module init call: That part
1657 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 */
1659}
1660
1661static int __init cpucache_init(void)
1662{
1663 int cpu;
1664
Andrew Mortona737b3e2006-03-22 00:08:11 -08001665 /*
1666 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 */
Christoph Lametere498be72005-09-09 13:03:32 -07001668 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001669 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 return 0;
1671}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672__initcall(cpucache_init);
1673
1674/*
1675 * Interface to system's page allocator. No need to hold the cache-lock.
1676 *
1677 * If we requested dmaable memory, we will get it. Even if we
1678 * did not request dmaable memory, we might get it, but that
1679 * would be relatively rare and ignorable.
1680 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001681static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682{
1683 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001684 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 int i;
1686
Luke Yangd6fef9d2006-04-10 22:52:56 -07001687#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001688 /*
1689 * Nommu uses slab's for process anonymous memory allocations, and thus
1690 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001691 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001692 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001693#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001694
Christoph Lameter3c517a62006-12-06 20:33:29 -08001695 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001696 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1697 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001698
Linus Torvalds517d0862009-06-16 19:50:13 -07001699 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 if (!page)
1701 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001703 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001705 add_zone_page_state(page_zone(page),
1706 NR_SLAB_RECLAIMABLE, nr_pages);
1707 else
1708 add_zone_page_state(page_zone(page),
1709 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001710 for (i = 0; i < nr_pages; i++)
1711 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001712
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001713 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1714 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1715
1716 if (cachep->ctor)
1717 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1718 else
1719 kmemcheck_mark_unallocated_pages(page, nr_pages);
1720 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001721
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001722 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723}
1724
1725/*
1726 * Interface to system's page release.
1727 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001728static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001730 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 struct page *page = virt_to_page(addr);
1732 const unsigned long nr_freed = i;
1733
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001734 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001735
Christoph Lameter972d1a72006-09-25 23:31:51 -07001736 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1737 sub_zone_page_state(page_zone(page),
1738 NR_SLAB_RECLAIMABLE, nr_freed);
1739 else
1740 sub_zone_page_state(page_zone(page),
1741 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001743 BUG_ON(!PageSlab(page));
1744 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 page++;
1746 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 if (current->reclaim_state)
1748 current->reclaim_state->reclaimed_slab += nr_freed;
1749 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750}
1751
1752static void kmem_rcu_free(struct rcu_head *head)
1753{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001754 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001755 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
1757 kmem_freepages(cachep, slab_rcu->addr);
1758 if (OFF_SLAB(cachep))
1759 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1760}
1761
1762#if DEBUG
1763
1764#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001765static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001766 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001768 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001770 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001772 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 return;
1774
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001775 *addr++ = 0x12345678;
1776 *addr++ = caller;
1777 *addr++ = smp_processor_id();
1778 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 {
1780 unsigned long *sptr = &caller;
1781 unsigned long svalue;
1782
1783 while (!kstack_end(sptr)) {
1784 svalue = *sptr++;
1785 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001786 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 size -= sizeof(unsigned long);
1788 if (size <= sizeof(unsigned long))
1789 break;
1790 }
1791 }
1792
1793 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795}
1796#endif
1797
Pekka Enberg343e0d72006-02-01 03:05:50 -08001798static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001800 int size = obj_size(cachep);
1801 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802
1803 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001804 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805}
1806
1807static void dump_line(char *data, int offset, int limit)
1808{
1809 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001810 unsigned char error = 0;
1811 int bad_count = 0;
1812
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001814 for (i = 0; i < limit; i++) {
1815 if (data[offset + i] != POISON_FREE) {
1816 error = data[offset + i];
1817 bad_count++;
1818 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001819 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001820 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001822
1823 if (bad_count == 1) {
1824 error ^= POISON_FREE;
1825 if (!(error & (error - 1))) {
1826 printk(KERN_ERR "Single bit error detected. Probably "
1827 "bad RAM.\n");
1828#ifdef CONFIG_X86
1829 printk(KERN_ERR "Run memtest86+ or a similar memory "
1830 "test tool.\n");
1831#else
1832 printk(KERN_ERR "Run a memory test tool.\n");
1833#endif
1834 }
1835 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836}
1837#endif
1838
1839#if DEBUG
1840
Pekka Enberg343e0d72006-02-01 03:05:50 -08001841static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842{
1843 int i, size;
1844 char *realobj;
1845
1846 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001847 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001848 *dbg_redzone1(cachep, objp),
1849 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 }
1851
1852 if (cachep->flags & SLAB_STORE_USER) {
1853 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001854 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001856 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 printk("\n");
1858 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001859 realobj = (char *)objp + obj_offset(cachep);
1860 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 int limit;
1863 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001864 if (i + limit > size)
1865 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 dump_line(realobj, i, limit);
1867 }
1868}
1869
Pekka Enberg343e0d72006-02-01 03:05:50 -08001870static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871{
1872 char *realobj;
1873 int size, i;
1874 int lines = 0;
1875
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001876 realobj = (char *)objp + obj_offset(cachep);
1877 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001879 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001881 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 exp = POISON_END;
1883 if (realobj[i] != exp) {
1884 int limit;
1885 /* Mismatch ! */
1886 /* Print header */
1887 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001888 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001889 "Slab corruption: %s start=%p, len=%d\n",
1890 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 print_objinfo(cachep, objp, 0);
1892 }
1893 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001894 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001896 if (i + limit > size)
1897 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 dump_line(realobj, i, limit);
1899 i += 16;
1900 lines++;
1901 /* Limit to 5 lines */
1902 if (lines > 5)
1903 break;
1904 }
1905 }
1906 if (lines != 0) {
1907 /* Print some data about the neighboring objects, if they
1908 * exist:
1909 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001910 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001911 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001913 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001915 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001916 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001918 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 print_objinfo(cachep, objp, 2);
1920 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001922 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001923 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001925 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 print_objinfo(cachep, objp, 2);
1927 }
1928 }
1929}
1930#endif
1931
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301933static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001934{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 int i;
1936 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001937 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
1939 if (cachep->flags & SLAB_POISON) {
1940#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001941 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1942 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001943 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001944 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 else
1946 check_poison_obj(cachep, objp);
1947#else
1948 check_poison_obj(cachep, objp);
1949#endif
1950 }
1951 if (cachep->flags & SLAB_RED_ZONE) {
1952 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1953 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001954 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1956 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001957 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001960}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301962static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001963{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001964}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965#endif
1966
Randy Dunlap911851e2006-03-22 00:08:14 -08001967/**
1968 * slab_destroy - destroy and release all objects in a slab
1969 * @cachep: cache pointer being destroyed
1970 * @slabp: slab pointer being destroyed
1971 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001972 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001973 * Before calling the slab must have been unlinked from the cache. The
1974 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001975 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001976static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001977{
1978 void *addr = slabp->s_mem - slabp->colouroff;
1979
Rabin Vincente79aec22008-07-04 00:40:32 +05301980 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1982 struct slab_rcu *slab_rcu;
1983
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001984 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 slab_rcu->cachep = cachep;
1986 slab_rcu->addr = addr;
1987 call_rcu(&slab_rcu->head, kmem_rcu_free);
1988 } else {
1989 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001990 if (OFF_SLAB(cachep))
1991 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 }
1993}
1994
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001995static void __kmem_cache_destroy(struct kmem_cache *cachep)
1996{
1997 int i;
1998 struct kmem_list3 *l3;
1999
2000 for_each_online_cpu(i)
2001 kfree(cachep->array[i]);
2002
2003 /* NUMA: free the list3 structures */
2004 for_each_online_node(i) {
2005 l3 = cachep->nodelists[i];
2006 if (l3) {
2007 kfree(l3->shared);
2008 free_alien_cache(l3->alien);
2009 kfree(l3);
2010 }
2011 }
2012 kmem_cache_free(&cache_cache, cachep);
2013}
2014
2015
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002017 * calculate_slab_order - calculate size (page order) of slabs
2018 * @cachep: pointer to the cache that is being created
2019 * @size: size of objects to be created in this cache.
2020 * @align: required alignment for the objects.
2021 * @flags: slab allocation flags
2022 *
2023 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002024 *
2025 * This could be made much more intelligent. For now, try to avoid using
2026 * high order pages for slabs. When the gfp() functions are more friendly
2027 * towards high-order requests, this should be changed.
2028 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002029static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002030 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002031{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002032 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002033 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002034 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002035
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002036 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037 unsigned int num;
2038 size_t remainder;
2039
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002040 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002041 if (!num)
2042 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002043
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002044 if (flags & CFLGS_OFF_SLAB) {
2045 /*
2046 * Max number of objs-per-slab for caches which
2047 * use off-slab slabs. Needed to avoid a possible
2048 * looping condition in cache_grow().
2049 */
2050 offslab_limit = size - sizeof(struct slab);
2051 offslab_limit /= sizeof(kmem_bufctl_t);
2052
2053 if (num > offslab_limit)
2054 break;
2055 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002056
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002057 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002058 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002059 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002060 left_over = remainder;
2061
2062 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002063 * A VFS-reclaimable slab tends to have most allocations
2064 * as GFP_NOFS and we really don't want to have to be allocating
2065 * higher-order pages when we are unable to shrink dcache.
2066 */
2067 if (flags & SLAB_RECLAIM_ACCOUNT)
2068 break;
2069
2070 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002071 * Large number of objects is good, but very large slabs are
2072 * currently bad for the gfp()s.
2073 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002074 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002075 break;
2076
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002077 /*
2078 * Acceptable internal fragmentation?
2079 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002080 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002081 break;
2082 }
2083 return left_over;
2084}
2085
Pekka Enberg83b519e2009-06-10 19:40:04 +03002086static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002087{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002088 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002089 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002090
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002091 if (g_cpucache_up == NONE) {
2092 /*
2093 * Note: the first kmem_cache_create must create the cache
2094 * that's used by kmalloc(24), otherwise the creation of
2095 * further caches will BUG().
2096 */
2097 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2098
2099 /*
2100 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2101 * the first cache, then we need to set up all its list3s,
2102 * otherwise the creation of further caches will BUG().
2103 */
2104 set_up_list3s(cachep, SIZE_AC);
2105 if (INDEX_AC == INDEX_L3)
2106 g_cpucache_up = PARTIAL_L3;
2107 else
2108 g_cpucache_up = PARTIAL_AC;
2109 } else {
2110 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002111 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002112
2113 if (g_cpucache_up == PARTIAL_AC) {
2114 set_up_list3s(cachep, SIZE_L3);
2115 g_cpucache_up = PARTIAL_L3;
2116 } else {
2117 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002118 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002119 cachep->nodelists[node] =
2120 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002121 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002122 BUG_ON(!cachep->nodelists[node]);
2123 kmem_list3_init(cachep->nodelists[node]);
2124 }
2125 }
2126 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002127 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002128 jiffies + REAPTIMEOUT_LIST3 +
2129 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2130
2131 cpu_cache_get(cachep)->avail = 0;
2132 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2133 cpu_cache_get(cachep)->batchcount = 1;
2134 cpu_cache_get(cachep)->touched = 0;
2135 cachep->batchcount = 1;
2136 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002137 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002138}
2139
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002140/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 * kmem_cache_create - Create a cache.
2142 * @name: A string which is used in /proc/slabinfo to identify this cache.
2143 * @size: The size of objects to be created in this cache.
2144 * @align: The required alignment for the objects.
2145 * @flags: SLAB flags
2146 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 *
2148 * Returns a ptr to the cache on success, NULL on failure.
2149 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002150 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 *
2152 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002153 * the module calling this has to destroy the cache before getting unloaded.
2154 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 * The flags are
2156 *
2157 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2158 * to catch references to uninitialised memory.
2159 *
2160 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2161 * for buffer overruns.
2162 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2164 * cacheline. This can be beneficial if you're counting cycles as closely
2165 * as davem.
2166 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002167struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002169 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170{
2171 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002172 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002173 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
2175 /*
2176 * Sanity checks... these are all serious usage bugs.
2177 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002178 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002179 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002180 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002181 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002182 BUG();
2183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002185 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002186 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302187 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002188 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002189 if (slab_is_available()) {
2190 get_online_cpus();
2191 mutex_lock(&cache_chain_mutex);
2192 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002193
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002194 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002195 char tmp;
2196 int res;
2197
2198 /*
2199 * This happens when the module gets unloaded and doesn't
2200 * destroy its slab cache and no-one else reuses the vmalloc
2201 * area of the module. Print a warning.
2202 */
Andrew Morton138ae662006-12-06 20:36:41 -08002203 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002204 if (res) {
matzeb4169522007-05-06 14:49:52 -07002205 printk(KERN_ERR
2206 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002207 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002208 continue;
2209 }
2210
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002211 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002212 printk(KERN_ERR
2213 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002214 dump_stack();
2215 goto oops;
2216 }
2217 }
2218
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219#if DEBUG
2220 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221#if FORCED_DEBUG
2222 /*
2223 * Enable redzoning and last user accounting, except for caches with
2224 * large objects, if the increased size would increase the object size
2225 * above the next power of two: caches with object sizes just above a
2226 * power of two have a significant amount of internal fragmentation.
2227 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002228 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2229 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002230 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 if (!(flags & SLAB_DESTROY_BY_RCU))
2232 flags |= SLAB_POISON;
2233#endif
2234 if (flags & SLAB_DESTROY_BY_RCU)
2235 BUG_ON(flags & SLAB_POISON);
2236#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002238 * Always checks flags, a caller might be expecting debug support which
2239 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002241 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242
Andrew Mortona737b3e2006-03-22 00:08:11 -08002243 /*
2244 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 * unaligned accesses for some archs when redzoning is used, and makes
2246 * sure any on-slab bufctl's are also correctly aligned.
2247 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002248 if (size & (BYTES_PER_WORD - 1)) {
2249 size += (BYTES_PER_WORD - 1);
2250 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 }
2252
Andrew Mortona737b3e2006-03-22 00:08:11 -08002253 /* calculate the final buffer alignment: */
2254
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 /* 1) arch recommendation: can be overridden for debug */
2256 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002257 /*
2258 * Default alignment: as specified by the arch code. Except if
2259 * an object is really small, then squeeze multiple objects into
2260 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 */
2262 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002263 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 ralign /= 2;
2265 } else {
2266 ralign = BYTES_PER_WORD;
2267 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002268
2269 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002270 * Redzoning and user store require word alignment or possibly larger.
2271 * Note this will be overridden by architecture or caller mandated
2272 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002273 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002274 if (flags & SLAB_STORE_USER)
2275 ralign = BYTES_PER_WORD;
2276
2277 if (flags & SLAB_RED_ZONE) {
2278 ralign = REDZONE_ALIGN;
2279 /* If redzoning, ensure that the second redzone is suitably
2280 * aligned, by adjusting the object size accordingly. */
2281 size += REDZONE_ALIGN - 1;
2282 size &= ~(REDZONE_ALIGN - 1);
2283 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002284
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002285 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 if (ralign < ARCH_SLAB_MINALIGN) {
2287 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002289 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 if (ralign < align) {
2291 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002293 /* disable debug if necessary */
2294 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002295 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002296 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002297 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 */
2299 align = ralign;
2300
Pekka Enberg83b519e2009-06-10 19:40:04 +03002301 if (slab_is_available())
2302 gfp = GFP_KERNEL;
2303 else
2304 gfp = GFP_NOWAIT;
2305
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002307 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002309 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310
2311#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002312 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313
Pekka Enbergca5f9702006-09-25 23:31:25 -07002314 /*
2315 * Both debugging options require word-alignment which is calculated
2316 * into align above.
2317 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002320 cachep->obj_offset += sizeof(unsigned long long);
2321 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 }
2323 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002324 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002325 * the real object. But if the second red zone needs to be
2326 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002328 if (flags & SLAB_RED_ZONE)
2329 size += REDZONE_ALIGN;
2330 else
2331 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 }
2333#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002334 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002335 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2336 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 size = PAGE_SIZE;
2338 }
2339#endif
2340#endif
2341
Ingo Molnare0a42722006-06-23 02:03:46 -07002342 /*
2343 * Determine if the slab management is 'on' or 'off' slab.
2344 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002345 * it too early on. Always use on-slab management when
2346 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002347 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002348 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2349 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 /*
2351 * Size is large, assume best to place the slab management obj
2352 * off-slab (should allow better packing of objs).
2353 */
2354 flags |= CFLGS_OFF_SLAB;
2355
2356 size = ALIGN(size, align);
2357
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002358 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359
2360 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002361 printk(KERN_ERR
2362 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 kmem_cache_free(&cache_cache, cachep);
2364 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002365 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002367 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2368 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369
2370 /*
2371 * If the slab has been placed off-slab, and we have enough space then
2372 * move it on-slab. This is at the expense of any extra colouring.
2373 */
2374 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2375 flags &= ~CFLGS_OFF_SLAB;
2376 left_over -= slab_size;
2377 }
2378
2379 if (flags & CFLGS_OFF_SLAB) {
2380 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002381 slab_size =
2382 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302383
2384#ifdef CONFIG_PAGE_POISONING
2385 /* If we're going to use the generic kernel_map_pages()
2386 * poisoning, then it's going to smash the contents of
2387 * the redzone and userword anyhow, so switch them off.
2388 */
2389 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2390 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2391#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 }
2393
2394 cachep->colour_off = cache_line_size();
2395 /* Offset must be a multiple of the alignment. */
2396 if (cachep->colour_off < align)
2397 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002398 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 cachep->slab_size = slab_size;
2400 cachep->flags = flags;
2401 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002402 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002404 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002405 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002407 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002408 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002409 /*
2410 * This is a possibility for one of the malloc_sizes caches.
2411 * But since we go off slab only for object size greater than
2412 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2413 * this should not happen at all.
2414 * But leave a BUG_ON for some lucky dude.
2415 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002416 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002417 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 cachep->name = name;
2420
Pekka Enberg83b519e2009-06-10 19:40:04 +03002421 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002422 __kmem_cache_destroy(cachep);
2423 cachep = NULL;
2424 goto oops;
2425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 /* cache setup completed, link it into the list */
2428 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002429oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 if (!cachep && (flags & SLAB_PANIC))
2431 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002432 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002433 if (slab_is_available()) {
2434 mutex_unlock(&cache_chain_mutex);
2435 put_online_cpus();
2436 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 return cachep;
2438}
2439EXPORT_SYMBOL(kmem_cache_create);
2440
2441#if DEBUG
2442static void check_irq_off(void)
2443{
2444 BUG_ON(!irqs_disabled());
2445}
2446
2447static void check_irq_on(void)
2448{
2449 BUG_ON(irqs_disabled());
2450}
2451
Pekka Enberg343e0d72006-02-01 03:05:50 -08002452static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453{
2454#ifdef CONFIG_SMP
2455 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002456 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457#endif
2458}
Christoph Lametere498be72005-09-09 13:03:32 -07002459
Pekka Enberg343e0d72006-02-01 03:05:50 -08002460static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002461{
2462#ifdef CONFIG_SMP
2463 check_irq_off();
2464 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2465#endif
2466}
2467
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468#else
2469#define check_irq_off() do { } while(0)
2470#define check_irq_on() do { } while(0)
2471#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002472#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473#endif
2474
Christoph Lameteraab22072006-03-22 00:09:06 -08002475static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2476 struct array_cache *ac,
2477 int force, int node);
2478
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479static void do_drain(void *arg)
2480{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002481 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002483 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484
2485 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002486 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002487 spin_lock(&cachep->nodelists[node]->list_lock);
2488 free_block(cachep, ac->entry, ac->avail, node);
2489 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 ac->avail = 0;
2491}
2492
Pekka Enberg343e0d72006-02-01 03:05:50 -08002493static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494{
Christoph Lametere498be72005-09-09 13:03:32 -07002495 struct kmem_list3 *l3;
2496 int node;
2497
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002498 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002500 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002501 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002502 if (l3 && l3->alien)
2503 drain_alien_cache(cachep, l3->alien);
2504 }
2505
2506 for_each_online_node(node) {
2507 l3 = cachep->nodelists[node];
2508 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002509 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511}
2512
Christoph Lametered11d9e2006-06-30 01:55:45 -07002513/*
2514 * Remove slabs from the list of free slabs.
2515 * Specify the number of slabs to drain in tofree.
2516 *
2517 * Returns the actual number of slabs released.
2518 */
2519static int drain_freelist(struct kmem_cache *cache,
2520 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002522 struct list_head *p;
2523 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Christoph Lametered11d9e2006-06-30 01:55:45 -07002526 nr_freed = 0;
2527 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528
Christoph Lametered11d9e2006-06-30 01:55:45 -07002529 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002530 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002531 if (p == &l3->slabs_free) {
2532 spin_unlock_irq(&l3->list_lock);
2533 goto out;
2534 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535
Christoph Lametered11d9e2006-06-30 01:55:45 -07002536 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002538 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539#endif
2540 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002541 /*
2542 * Safe to drop the lock. The slab is no longer linked
2543 * to the cache.
2544 */
2545 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002546 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002547 slab_destroy(cache, slabp);
2548 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002550out:
2551 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552}
2553
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002554/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002555static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002556{
2557 int ret = 0, i = 0;
2558 struct kmem_list3 *l3;
2559
2560 drain_cpu_caches(cachep);
2561
2562 check_irq_on();
2563 for_each_online_node(i) {
2564 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002565 if (!l3)
2566 continue;
2567
2568 drain_freelist(cachep, l3, l3->free_objects);
2569
2570 ret += !list_empty(&l3->slabs_full) ||
2571 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002572 }
2573 return (ret ? 1 : 0);
2574}
2575
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576/**
2577 * kmem_cache_shrink - Shrink a cache.
2578 * @cachep: The cache to shrink.
2579 *
2580 * Releases as many slabs as possible for a cache.
2581 * To help debugging, a zero exit status indicates all slabs were released.
2582 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002583int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002585 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002586 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002588 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002589 mutex_lock(&cache_chain_mutex);
2590 ret = __cache_shrink(cachep);
2591 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002592 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002593 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594}
2595EXPORT_SYMBOL(kmem_cache_shrink);
2596
2597/**
2598 * kmem_cache_destroy - delete a cache
2599 * @cachep: the cache to destroy
2600 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002601 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 *
2603 * It is expected this function will be called by a module when it is
2604 * unloaded. This will remove the cache completely, and avoid a duplicate
2605 * cache being allocated each time a module is loaded and unloaded, if the
2606 * module doesn't have persistent in-kernel storage across loads and unloads.
2607 *
2608 * The cache must be empty before calling this function.
2609 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002610 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 * during the kmem_cache_destroy().
2612 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002613void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002615 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002618 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002619 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 /*
2621 * the chain is never empty, cache_cache is never destroyed
2622 */
2623 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 if (__cache_shrink(cachep)) {
2625 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002627 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002628 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002629 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 }
2631
2632 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002633 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002635 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002636 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002637 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638}
2639EXPORT_SYMBOL(kmem_cache_destroy);
2640
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002641/*
2642 * Get the memory for a slab management obj.
2643 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2644 * always come from malloc_sizes caches. The slab descriptor cannot
2645 * come from the same cache which is getting created because,
2646 * when we are searching for an appropriate cache for these
2647 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2648 * If we are creating a malloc_sizes cache here it would not be visible to
2649 * kmem_find_general_cachep till the initialization is complete.
2650 * Hence we cannot have slabp_cache same as the original cache.
2651 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002652static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002653 int colour_off, gfp_t local_flags,
2654 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655{
2656 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002657
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 if (OFF_SLAB(cachep)) {
2659 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002660 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002661 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002662 /*
2663 * If the first object in the slab is leaked (it's allocated
2664 * but no one has a reference to it), we want to make sure
2665 * kmemleak does not treat the ->s_mem pointer as a reference
2666 * to the object. Otherwise we will not report the leak.
2667 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002668 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2669 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 if (!slabp)
2671 return NULL;
2672 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 colour_off += cachep->slab_size;
2675 }
2676 slabp->inuse = 0;
2677 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002679 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002680 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 return slabp;
2682}
2683
2684static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2685{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002686 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687}
2688
Pekka Enberg343e0d72006-02-01 03:05:50 -08002689static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002690 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691{
2692 int i;
2693
2694 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002695 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696#if DEBUG
2697 /* need to poison the objs? */
2698 if (cachep->flags & SLAB_POISON)
2699 poison_obj(cachep, objp, POISON_FREE);
2700 if (cachep->flags & SLAB_STORE_USER)
2701 *dbg_userword(cachep, objp) = NULL;
2702
2703 if (cachep->flags & SLAB_RED_ZONE) {
2704 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2705 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2706 }
2707 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002708 * Constructors are not allowed to allocate memory from the same
2709 * cache which they are a constructor for. Otherwise, deadlock.
2710 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 */
2712 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002713 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714
2715 if (cachep->flags & SLAB_RED_ZONE) {
2716 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2717 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002718 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2720 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002721 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002723 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2724 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002725 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002726 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727#else
2728 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002729 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002731 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002733 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734}
2735
Pekka Enberg343e0d72006-02-01 03:05:50 -08002736static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002738 if (CONFIG_ZONE_DMA_FLAG) {
2739 if (flags & GFP_DMA)
2740 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2741 else
2742 BUG_ON(cachep->gfpflags & GFP_DMA);
2743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744}
2745
Andrew Mortona737b3e2006-03-22 00:08:11 -08002746static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2747 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002748{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002749 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002750 kmem_bufctl_t next;
2751
2752 slabp->inuse++;
2753 next = slab_bufctl(slabp)[slabp->free];
2754#if DEBUG
2755 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2756 WARN_ON(slabp->nodeid != nodeid);
2757#endif
2758 slabp->free = next;
2759
2760 return objp;
2761}
2762
Andrew Mortona737b3e2006-03-22 00:08:11 -08002763static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2764 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002765{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002766 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002767
2768#if DEBUG
2769 /* Verify that the slab belongs to the intended node */
2770 WARN_ON(slabp->nodeid != nodeid);
2771
Al Viro871751e2006-03-25 03:06:39 -08002772 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002773 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002775 BUG();
2776 }
2777#endif
2778 slab_bufctl(slabp)[objnr] = slabp->free;
2779 slabp->free = objnr;
2780 slabp->inuse--;
2781}
2782
Pekka Enberg47768742006-06-23 02:03:07 -07002783/*
2784 * Map pages beginning at addr to the given cache and slab. This is required
2785 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002786 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002787 */
2788static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2789 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790{
Pekka Enberg47768742006-06-23 02:03:07 -07002791 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 struct page *page;
2793
Pekka Enberg47768742006-06-23 02:03:07 -07002794 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002795
Pekka Enberg47768742006-06-23 02:03:07 -07002796 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002797 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002798 nr_pages <<= cache->gfporder;
2799
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002801 page_set_cache(page, cache);
2802 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002804 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805}
2806
2807/*
2808 * Grow (by 1) the number of slabs within a cache. This is called by
2809 * kmem_cache_alloc() when there are no active objs left in a cache.
2810 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002811static int cache_grow(struct kmem_cache *cachep,
2812 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002814 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002815 size_t offset;
2816 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002817 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818
Andrew Mortona737b3e2006-03-22 00:08:11 -08002819 /*
2820 * Be lazy and only check for valid flags here, keeping it out of the
2821 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002823 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2824 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002826 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002828 l3 = cachep->nodelists[nodeid];
2829 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
2831 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002832 offset = l3->colour_next;
2833 l3->colour_next++;
2834 if (l3->colour_next >= cachep->colour)
2835 l3->colour_next = 0;
2836 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002838 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839
2840 if (local_flags & __GFP_WAIT)
2841 local_irq_enable();
2842
2843 /*
2844 * The test for missing atomic flag is performed here, rather than
2845 * the more obvious place, simply to reduce the critical path length
2846 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2847 * will eventually be caught here (where it matters).
2848 */
2849 kmem_flagcheck(cachep, flags);
2850
Andrew Mortona737b3e2006-03-22 00:08:11 -08002851 /*
2852 * Get mem for the objs. Attempt to allocate a physical page from
2853 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002854 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002855 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002856 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002857 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 goto failed;
2859
2860 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002861 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002862 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002863 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 goto opps1;
2865
Pekka Enberg47768742006-06-23 02:03:07 -07002866 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
Christoph Lametera35afb82007-05-16 22:10:57 -07002868 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869
2870 if (local_flags & __GFP_WAIT)
2871 local_irq_disable();
2872 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002873 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874
2875 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002876 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002878 l3->free_objects += cachep->num;
2879 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002881opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002883failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 if (local_flags & __GFP_WAIT)
2885 local_irq_disable();
2886 return 0;
2887}
2888
2889#if DEBUG
2890
2891/*
2892 * Perform extra freeing checks:
2893 * - detect bad pointers.
2894 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 */
2896static void kfree_debugcheck(const void *objp)
2897{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 if (!virt_addr_valid(objp)) {
2899 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002900 (unsigned long)objp);
2901 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903}
2904
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002905static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2906{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002907 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002908
2909 redzone1 = *dbg_redzone1(cache, obj);
2910 redzone2 = *dbg_redzone2(cache, obj);
2911
2912 /*
2913 * Redzone is ok.
2914 */
2915 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2916 return;
2917
2918 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2919 slab_error(cache, "double free detected");
2920 else
2921 slab_error(cache, "memory outside object was overwritten");
2922
David Woodhouseb46b8f12007-05-08 00:22:59 -07002923 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002924 obj, redzone1, redzone2);
2925}
2926
Pekka Enberg343e0d72006-02-01 03:05:50 -08002927static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002928 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929{
2930 struct page *page;
2931 unsigned int objnr;
2932 struct slab *slabp;
2933
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002934 BUG_ON(virt_to_cache(objp) != cachep);
2935
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002936 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002938 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939
Pekka Enberg065d41c2005-11-13 16:06:46 -08002940 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
2942 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002943 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2945 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2946 }
2947 if (cachep->flags & SLAB_STORE_USER)
2948 *dbg_userword(cachep, objp) = caller;
2949
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002950 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951
2952 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002953 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954
Al Viro871751e2006-03-25 03:06:39 -08002955#ifdef CONFIG_DEBUG_SLAB_LEAK
2956 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2957#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 if (cachep->flags & SLAB_POISON) {
2959#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002960 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002962 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002963 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 } else {
2965 poison_obj(cachep, objp, POISON_FREE);
2966 }
2967#else
2968 poison_obj(cachep, objp, POISON_FREE);
2969#endif
2970 }
2971 return objp;
2972}
2973
Pekka Enberg343e0d72006-02-01 03:05:50 -08002974static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975{
2976 kmem_bufctl_t i;
2977 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002978
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 /* Check slab's freelist to see if this obj is there. */
2980 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2981 entries++;
2982 if (entries > cachep->num || i >= cachep->num)
2983 goto bad;
2984 }
2985 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002986bad:
2987 printk(KERN_ERR "slab: Internal list corruption detected in "
2988 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2989 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002990 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002991 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002992 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002993 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002995 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 }
2997 printk("\n");
2998 BUG();
2999 }
3000}
3001#else
3002#define kfree_debugcheck(x) do { } while(0)
3003#define cache_free_debugcheck(x,objp,z) (objp)
3004#define check_slabp(x,y) do { } while(0)
3005#endif
3006
Pekka Enberg343e0d72006-02-01 03:05:50 -08003007static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008{
3009 int batchcount;
3010 struct kmem_list3 *l3;
3011 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003012 int node;
3013
Andrew Mortona737b3e2006-03-22 00:08:11 -08003014retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003015 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003016 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003017 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 batchcount = ac->batchcount;
3019 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003020 /*
3021 * If there was little recent activity on this cache, then
3022 * perform only a partial refill. Otherwise we could generate
3023 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 */
3025 batchcount = BATCHREFILL_LIMIT;
3026 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003027 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028
Christoph Lametere498be72005-09-09 13:03:32 -07003029 BUG_ON(ac->avail > 0 || !l3);
3030 spin_lock(&l3->list_lock);
3031
Christoph Lameter3ded1752006-03-25 03:06:44 -08003032 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003033 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3034 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003035 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003036 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003037
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 while (batchcount > 0) {
3039 struct list_head *entry;
3040 struct slab *slabp;
3041 /* Get slab alloc is to come from. */
3042 entry = l3->slabs_partial.next;
3043 if (entry == &l3->slabs_partial) {
3044 l3->free_touched = 1;
3045 entry = l3->slabs_free.next;
3046 if (entry == &l3->slabs_free)
3047 goto must_grow;
3048 }
3049
3050 slabp = list_entry(entry, struct slab, list);
3051 check_slabp(cachep, slabp);
3052 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003053
3054 /*
3055 * The slab was either on partial or free list so
3056 * there must be at least one object available for
3057 * allocation.
3058 */
roel kluin249b9f32008-10-29 17:18:07 -04003059 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003060
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 STATS_INC_ALLOCED(cachep);
3063 STATS_INC_ACTIVE(cachep);
3064 STATS_SET_HIGH(cachep);
3065
Matthew Dobson78d382d2006-02-01 03:05:47 -08003066 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003067 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 }
3069 check_slabp(cachep, slabp);
3070
3071 /* move slabp to correct slabp list: */
3072 list_del(&slabp->list);
3073 if (slabp->free == BUFCTL_END)
3074 list_add(&slabp->list, &l3->slabs_full);
3075 else
3076 list_add(&slabp->list, &l3->slabs_partial);
3077 }
3078
Andrew Mortona737b3e2006-03-22 00:08:11 -08003079must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003081alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003082 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083
3084 if (unlikely(!ac->avail)) {
3085 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003086 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003087
Andrew Mortona737b3e2006-03-22 00:08:11 -08003088 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003089 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 return NULL;
3092
Andrew Mortona737b3e2006-03-22 00:08:11 -08003093 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 goto retry;
3095 }
3096 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003097 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098}
3099
Andrew Mortona737b3e2006-03-22 00:08:11 -08003100static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3101 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102{
3103 might_sleep_if(flags & __GFP_WAIT);
3104#if DEBUG
3105 kmem_flagcheck(cachep, flags);
3106#endif
3107}
3108
3109#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003110static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3111 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003113 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003115 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003117 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003118 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003119 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120 else
3121 check_poison_obj(cachep, objp);
3122#else
3123 check_poison_obj(cachep, objp);
3124#endif
3125 poison_obj(cachep, objp, POISON_INUSE);
3126 }
3127 if (cachep->flags & SLAB_STORE_USER)
3128 *dbg_userword(cachep, objp) = caller;
3129
3130 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003131 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3132 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3133 slab_error(cachep, "double free, or memory outside"
3134 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003135 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003136 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003137 objp, *dbg_redzone1(cachep, objp),
3138 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 }
3140 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3141 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3142 }
Al Viro871751e2006-03-25 03:06:39 -08003143#ifdef CONFIG_DEBUG_SLAB_LEAK
3144 {
3145 struct slab *slabp;
3146 unsigned objnr;
3147
Christoph Lameterb49af682007-05-06 14:49:41 -07003148 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003149 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3150 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3151 }
3152#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003153 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003154 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003155 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003156#if ARCH_SLAB_MINALIGN
3157 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3158 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3159 objp, ARCH_SLAB_MINALIGN);
3160 }
3161#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162 return objp;
3163}
3164#else
3165#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3166#endif
3167
Akinobu Mita773ff602008-12-23 19:37:01 +09003168static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003169{
3170 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003171 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003172
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003173 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003174}
3175
Pekka Enberg343e0d72006-02-01 03:05:50 -08003176static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003178 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 struct array_cache *ac;
3180
Alok N Kataria5c382302005-09-27 21:45:46 -07003181 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003182
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003183 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 if (likely(ac->avail)) {
3185 STATS_INC_ALLOCHIT(cachep);
3186 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003187 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188 } else {
3189 STATS_INC_ALLOCMISS(cachep);
3190 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003191 /*
3192 * the 'ac' may be updated by cache_alloc_refill(),
3193 * and kmemleak_erase() requires its correct value.
3194 */
3195 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003197 /*
3198 * To avoid a false negative, if an object that is in one of the
3199 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3200 * treat the array pointers as a reference to the object.
3201 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003202 if (objp)
3203 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003204 return objp;
3205}
3206
Christoph Lametere498be72005-09-09 13:03:32 -07003207#ifdef CONFIG_NUMA
3208/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003209 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003210 *
3211 * If we are in_interrupt, then process context, including cpusets and
3212 * mempolicy, may not apply and should not be used for allocation policy.
3213 */
3214static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3215{
3216 int nid_alloc, nid_here;
3217
Christoph Lameter765c4502006-09-27 01:50:08 -07003218 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003219 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003220 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003221 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003222 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003223 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003224 else if (current->mempolicy)
3225 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003226 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003227 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003228 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003229 return NULL;
3230}
3231
3232/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003233 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003234 * certain node and fall back is permitted. First we scan all the
3235 * available nodelists for available objects. If that fails then we
3236 * perform an allocation without specifying a node. This allows the page
3237 * allocator to do its reclaim / fallback magic. We then insert the
3238 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003239 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003240static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003241{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003242 struct zonelist *zonelist;
3243 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003244 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003245 struct zone *zone;
3246 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003247 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003248 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003249
3250 if (flags & __GFP_THISNODE)
3251 return NULL;
3252
Miao Xiec0ff7452010-05-24 14:32:08 -07003253 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003254 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003255 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003256
Christoph Lameter3c517a62006-12-06 20:33:29 -08003257retry:
3258 /*
3259 * Look through allowed nodes for objects available
3260 * from existing per node queues.
3261 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003262 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3263 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003264
Mel Gorman54a6eb52008-04-28 02:12:16 -07003265 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003266 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003267 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003268 obj = ____cache_alloc_node(cache,
3269 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003270 if (obj)
3271 break;
3272 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003273 }
3274
Christoph Lametercfce6602007-05-06 14:50:17 -07003275 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003276 /*
3277 * This allocation will be performed within the constraints
3278 * of the current cpuset / memory policy requirements.
3279 * We may trigger various forms of reclaim on the allowed
3280 * set and go into memory reserves if necessary.
3281 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003282 if (local_flags & __GFP_WAIT)
3283 local_irq_enable();
3284 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003285 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003286 if (local_flags & __GFP_WAIT)
3287 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003288 if (obj) {
3289 /*
3290 * Insert into the appropriate per node queues
3291 */
3292 nid = page_to_nid(virt_to_page(obj));
3293 if (cache_grow(cache, flags, nid, obj)) {
3294 obj = ____cache_alloc_node(cache,
3295 flags | GFP_THISNODE, nid);
3296 if (!obj)
3297 /*
3298 * Another processor may allocate the
3299 * objects in the slab since we are
3300 * not holding any locks.
3301 */
3302 goto retry;
3303 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003304 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003305 obj = NULL;
3306 }
3307 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003308 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003309 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003310 return obj;
3311}
3312
3313/*
Christoph Lametere498be72005-09-09 13:03:32 -07003314 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003316static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003317 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003318{
3319 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003320 struct slab *slabp;
3321 struct kmem_list3 *l3;
3322 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 l3 = cachep->nodelists[nodeid];
3326 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003327
Andrew Mortona737b3e2006-03-22 00:08:11 -08003328retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003329 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003330 spin_lock(&l3->list_lock);
3331 entry = l3->slabs_partial.next;
3332 if (entry == &l3->slabs_partial) {
3333 l3->free_touched = 1;
3334 entry = l3->slabs_free.next;
3335 if (entry == &l3->slabs_free)
3336 goto must_grow;
3337 }
Christoph Lametere498be72005-09-09 13:03:32 -07003338
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003339 slabp = list_entry(entry, struct slab, list);
3340 check_spinlock_acquired_node(cachep, nodeid);
3341 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003342
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003343 STATS_INC_NODEALLOCS(cachep);
3344 STATS_INC_ACTIVE(cachep);
3345 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003347 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003348
Matthew Dobson78d382d2006-02-01 03:05:47 -08003349 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003350 check_slabp(cachep, slabp);
3351 l3->free_objects--;
3352 /* move slabp to correct slabp list: */
3353 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003354
Andrew Mortona737b3e2006-03-22 00:08:11 -08003355 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003356 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003357 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003359
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003360 spin_unlock(&l3->list_lock);
3361 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003362
Andrew Mortona737b3e2006-03-22 00:08:11 -08003363must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003364 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003365 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003366 if (x)
3367 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003368
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003369 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003370
Andrew Mortona737b3e2006-03-22 00:08:11 -08003371done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003372 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003373}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003374
3375/**
3376 * kmem_cache_alloc_node - Allocate an object on the specified node
3377 * @cachep: The cache to allocate from.
3378 * @flags: See kmalloc().
3379 * @nodeid: node number of the target node.
3380 * @caller: return address of caller, used for debug information
3381 *
3382 * Identical to kmem_cache_alloc but it will allocate memory on the given
3383 * node, which can improve the performance for cpu bound structures.
3384 *
3385 * Fallback to other node is possible if __GFP_THISNODE is not set.
3386 */
3387static __always_inline void *
3388__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3389 void *caller)
3390{
3391 unsigned long save_flags;
3392 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003393 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003394
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003395 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003396
Nick Piggincf40bd12009-01-21 08:12:39 +01003397 lockdep_trace_alloc(flags);
3398
Akinobu Mita773ff602008-12-23 19:37:01 +09003399 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003400 return NULL;
3401
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003402 cache_alloc_debugcheck_before(cachep, flags);
3403 local_irq_save(save_flags);
3404
Tim Blechmann8e15b792009-11-30 18:59:34 +01003405 if (nodeid == -1)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003406 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003407
3408 if (unlikely(!cachep->nodelists[nodeid])) {
3409 /* Node not bootstrapped yet */
3410 ptr = fallback_alloc(cachep, flags);
3411 goto out;
3412 }
3413
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003414 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003415 /*
3416 * Use the locally cached objects if possible.
3417 * However ____cache_alloc does not allow fallback
3418 * to other nodes. It may fail while we still have
3419 * objects on other nodes available.
3420 */
3421 ptr = ____cache_alloc(cachep, flags);
3422 if (ptr)
3423 goto out;
3424 }
3425 /* ___cache_alloc_node can fall back to other nodes */
3426 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3427 out:
3428 local_irq_restore(save_flags);
3429 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003430 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3431 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003432
Pekka Enbergc175eea2008-05-09 20:35:53 +02003433 if (likely(ptr))
3434 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3435
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003436 if (unlikely((flags & __GFP_ZERO) && ptr))
3437 memset(ptr, 0, obj_size(cachep));
3438
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003439 return ptr;
3440}
3441
3442static __always_inline void *
3443__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3444{
3445 void *objp;
3446
3447 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3448 objp = alternate_node_alloc(cache, flags);
3449 if (objp)
3450 goto out;
3451 }
3452 objp = ____cache_alloc(cache, flags);
3453
3454 /*
3455 * We may just have run out of memory on the local node.
3456 * ____cache_alloc_node() knows how to locate memory on other nodes
3457 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003458 if (!objp)
3459 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003460
3461 out:
3462 return objp;
3463}
3464#else
3465
3466static __always_inline void *
3467__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3468{
3469 return ____cache_alloc(cachep, flags);
3470}
3471
3472#endif /* CONFIG_NUMA */
3473
3474static __always_inline void *
3475__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3476{
3477 unsigned long save_flags;
3478 void *objp;
3479
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003480 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003481
Nick Piggincf40bd12009-01-21 08:12:39 +01003482 lockdep_trace_alloc(flags);
3483
Akinobu Mita773ff602008-12-23 19:37:01 +09003484 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003485 return NULL;
3486
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003487 cache_alloc_debugcheck_before(cachep, flags);
3488 local_irq_save(save_flags);
3489 objp = __do_cache_alloc(cachep, flags);
3490 local_irq_restore(save_flags);
3491 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003492 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3493 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003494 prefetchw(objp);
3495
Pekka Enbergc175eea2008-05-09 20:35:53 +02003496 if (likely(objp))
3497 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3498
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003499 if (unlikely((flags & __GFP_ZERO) && objp))
3500 memset(objp, 0, obj_size(cachep));
3501
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003502 return objp;
3503}
Christoph Lametere498be72005-09-09 13:03:32 -07003504
3505/*
3506 * Caller needs to acquire correct kmem_list's list_lock
3507 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003508static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003509 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510{
3511 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003512 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513
3514 for (i = 0; i < nr_objects; i++) {
3515 void *objp = objpp[i];
3516 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003518 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003519 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003521 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003523 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003525 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 check_slabp(cachep, slabp);
3527
3528 /* fixup slab chains */
3529 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003530 if (l3->free_objects > l3->free_limit) {
3531 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003532 /* No need to drop any previously held
3533 * lock here, even if we have a off-slab slab
3534 * descriptor it is guaranteed to come from
3535 * a different cache, refer to comments before
3536 * alloc_slabmgmt.
3537 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 slab_destroy(cachep, slabp);
3539 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003540 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 }
3542 } else {
3543 /* Unconditionally move a slab to the end of the
3544 * partial list on free - maximum time for the
3545 * other objects to be freed, too.
3546 */
Christoph Lametere498be72005-09-09 13:03:32 -07003547 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 }
3549 }
3550}
3551
Pekka Enberg343e0d72006-02-01 03:05:50 -08003552static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553{
3554 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003555 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003556 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557
3558 batchcount = ac->batchcount;
3559#if DEBUG
3560 BUG_ON(!batchcount || batchcount > ac->avail);
3561#endif
3562 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003563 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003564 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003565 if (l3->shared) {
3566 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003567 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 if (max) {
3569 if (batchcount > max)
3570 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003571 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003572 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 shared_array->avail += batchcount;
3574 goto free_done;
3575 }
3576 }
3577
Christoph Lameterff694162005-09-22 21:44:02 -07003578 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003579free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580#if STATS
3581 {
3582 int i = 0;
3583 struct list_head *p;
3584
Christoph Lametere498be72005-09-09 13:03:32 -07003585 p = l3->slabs_free.next;
3586 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 struct slab *slabp;
3588
3589 slabp = list_entry(p, struct slab, list);
3590 BUG_ON(slabp->inuse);
3591
3592 i++;
3593 p = p->next;
3594 }
3595 STATS_SET_FREEABLE(cachep, i);
3596 }
3597#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003598 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003600 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601}
3602
3603/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003604 * Release an obj back to its cache. If the obj has a constructed state, it must
3605 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003607static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3608 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003610 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611
3612 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003613 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003614 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615
Pekka Enbergc175eea2008-05-09 20:35:53 +02003616 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3617
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003618 /*
3619 * Skip calling cache_free_alien() when the platform is not numa.
3620 * This will avoid cache misses that happen while accessing slabp (which
3621 * is per page memory reference) to get nodeid. Instead use a global
3622 * variable to skip the call, which is mostly likely to be present in
3623 * the cache.
3624 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003625 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003626 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003627
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 if (likely(ac->avail < ac->limit)) {
3629 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003630 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 return;
3632 } else {
3633 STATS_INC_FREEMISS(cachep);
3634 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003635 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 }
3637}
3638
3639/**
3640 * kmem_cache_alloc - Allocate an object
3641 * @cachep: The cache to allocate from.
3642 * @flags: See kmalloc().
3643 *
3644 * Allocate an object from this cache. The flags are only relevant
3645 * if the cache has no available objects.
3646 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003647void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003649 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3650
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003651 trace_kmem_cache_alloc(_RET_IP_, ret,
3652 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003653
3654 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655}
3656EXPORT_SYMBOL(kmem_cache_alloc);
3657
Li Zefan0f24f122009-12-11 15:45:30 +08003658#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003659void *
3660kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003661{
Steven Rostedt85beb582010-11-24 16:23:34 -05003662 void *ret;
3663
3664 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3665
3666 trace_kmalloc(_RET_IP_, ret,
3667 size, slab_buffer_size(cachep), flags);
3668 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003669}
Steven Rostedt85beb582010-11-24 16:23:34 -05003670EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003671#endif
3672
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003674void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3675{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003676 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3677 __builtin_return_address(0));
3678
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003679 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3680 obj_size(cachep), cachep->buffer_size,
3681 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003682
3683 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003684}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685EXPORT_SYMBOL(kmem_cache_alloc_node);
3686
Li Zefan0f24f122009-12-11 15:45:30 +08003687#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003688void *kmem_cache_alloc_node_trace(size_t size,
3689 struct kmem_cache *cachep,
3690 gfp_t flags,
3691 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003692{
Steven Rostedt85beb582010-11-24 16:23:34 -05003693 void *ret;
3694
3695 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003696 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003697 trace_kmalloc_node(_RET_IP_, ret,
3698 size, slab_buffer_size(cachep),
3699 flags, nodeid);
3700 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003701}
Steven Rostedt85beb582010-11-24 16:23:34 -05003702EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003703#endif
3704
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705static __always_inline void *
3706__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003707{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003708 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003709
3710 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003711 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3712 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003713 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003714}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003715
Li Zefan0bb38a52009-12-11 15:45:50 +08003716#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003717void *__kmalloc_node(size_t size, gfp_t flags, int node)
3718{
3719 return __do_kmalloc_node(size, flags, node,
3720 __builtin_return_address(0));
3721}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003722EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003723
3724void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003725 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003726{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003727 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003728}
3729EXPORT_SYMBOL(__kmalloc_node_track_caller);
3730#else
3731void *__kmalloc_node(size_t size, gfp_t flags, int node)
3732{
3733 return __do_kmalloc_node(size, flags, node, NULL);
3734}
3735EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003736#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003737#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738
3739/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003740 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003742 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003743 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003745static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3746 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003748 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003749 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003751 /* If you want to save a few bytes .text space: replace
3752 * __ with kmem_.
3753 * Then kmalloc uses the uninlined functions instead of the inline
3754 * functions.
3755 */
3756 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003757 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3758 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003759 ret = __cache_alloc(cachep, flags, caller);
3760
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003761 trace_kmalloc((unsigned long) caller, ret,
3762 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003763
3764 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003765}
3766
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003767
Li Zefan0bb38a52009-12-11 15:45:50 +08003768#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003769void *__kmalloc(size_t size, gfp_t flags)
3770{
Al Viro871751e2006-03-25 03:06:39 -08003771 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772}
3773EXPORT_SYMBOL(__kmalloc);
3774
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003775void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003776{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003777 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003778}
3779EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003780
3781#else
3782void *__kmalloc(size_t size, gfp_t flags)
3783{
3784 return __do_kmalloc(size, flags, NULL);
3785}
3786EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003787#endif
3788
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789/**
3790 * kmem_cache_free - Deallocate an object
3791 * @cachep: The cache the allocation was from.
3792 * @objp: The previously allocated object.
3793 *
3794 * Free an object which was previously allocated from this
3795 * cache.
3796 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003797void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798{
3799 unsigned long flags;
3800
3801 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003802 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003803 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3804 debug_check_no_obj_freed(objp, obj_size(cachep));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003805 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003807
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003808 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809}
3810EXPORT_SYMBOL(kmem_cache_free);
3811
3812/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 * kfree - free previously allocated memory
3814 * @objp: pointer returned by kmalloc.
3815 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003816 * If @objp is NULL, no operation is performed.
3817 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818 * Don't free memory not originally allocated by kmalloc()
3819 * or you will run into trouble.
3820 */
3821void kfree(const void *objp)
3822{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003823 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824 unsigned long flags;
3825
Pekka Enberg2121db72009-03-25 11:05:57 +02003826 trace_kfree(_RET_IP_, objp);
3827
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003828 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 return;
3830 local_irq_save(flags);
3831 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003832 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003833 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003834 debug_check_no_obj_freed(objp, obj_size(c));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003835 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 local_irq_restore(flags);
3837}
3838EXPORT_SYMBOL(kfree);
3839
Pekka Enberg343e0d72006-02-01 03:05:50 -08003840unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003842 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843}
3844EXPORT_SYMBOL(kmem_cache_size);
3845
Christoph Lametere498be72005-09-09 13:03:32 -07003846/*
Simon Arlott183ff222007-10-20 01:27:18 +02003847 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003848 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003849static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003850{
3851 int node;
3852 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003853 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003854 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003855
Mel Gorman9c09a952008-01-24 05:49:54 -08003856 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003857
Paul Menage3395ee02006-12-06 20:32:16 -08003858 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003859 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003860 if (!new_alien)
3861 goto fail;
3862 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003863
Eric Dumazet63109842007-05-06 14:49:28 -07003864 new_shared = NULL;
3865 if (cachep->shared) {
3866 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003867 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003868 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003869 if (!new_shared) {
3870 free_alien_cache(new_alien);
3871 goto fail;
3872 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003873 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003874
Andrew Mortona737b3e2006-03-22 00:08:11 -08003875 l3 = cachep->nodelists[node];
3876 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003877 struct array_cache *shared = l3->shared;
3878
Christoph Lametere498be72005-09-09 13:03:32 -07003879 spin_lock_irq(&l3->list_lock);
3880
Christoph Lametercafeb022006-03-25 03:06:46 -08003881 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003882 free_block(cachep, shared->entry,
3883 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003884
Christoph Lametercafeb022006-03-25 03:06:46 -08003885 l3->shared = new_shared;
3886 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003887 l3->alien = new_alien;
3888 new_alien = NULL;
3889 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003890 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003891 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003892 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003893 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003894 free_alien_cache(new_alien);
3895 continue;
3896 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003897 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003898 if (!l3) {
3899 free_alien_cache(new_alien);
3900 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003901 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003902 }
Christoph Lametere498be72005-09-09 13:03:32 -07003903
3904 kmem_list3_init(l3);
3905 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003906 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003907 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003908 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003909 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003910 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003911 cachep->nodelists[node] = l3;
3912 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003913 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003914
Andrew Mortona737b3e2006-03-22 00:08:11 -08003915fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003916 if (!cachep->next.next) {
3917 /* Cache is not active yet. Roll back what we did */
3918 node--;
3919 while (node >= 0) {
3920 if (cachep->nodelists[node]) {
3921 l3 = cachep->nodelists[node];
3922
3923 kfree(l3->shared);
3924 free_alien_cache(l3->alien);
3925 kfree(l3);
3926 cachep->nodelists[node] = NULL;
3927 }
3928 node--;
3929 }
3930 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003931 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003932}
3933
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003935 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 struct array_cache *new[NR_CPUS];
3937};
3938
3939static void do_ccupdate_local(void *info)
3940{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003941 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 struct array_cache *old;
3943
3944 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003945 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003946
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3948 new->new[smp_processor_id()] = old;
3949}
3950
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003951/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003952static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003953 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003955 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003956 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957
Pekka Enberg83b519e2009-06-10 19:40:04 +03003958 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003959 if (!new)
3960 return -ENOMEM;
3961
Christoph Lametere498be72005-09-09 13:03:32 -07003962 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003963 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003964 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003965 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003966 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003967 kfree(new->new[i]);
3968 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003969 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 }
3971 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003972 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003974 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003975
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 cachep->batchcount = batchcount;
3978 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003979 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980
Christoph Lametere498be72005-09-09 13:03:32 -07003981 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003982 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 if (!ccold)
3984 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003985 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3986 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3987 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 kfree(ccold);
3989 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003990 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003991 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992}
3993
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003994/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003995static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996{
3997 int err;
3998 int limit, shared;
3999
Andrew Mortona737b3e2006-03-22 00:08:11 -08004000 /*
4001 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 * - create a LIFO ordering, i.e. return objects that are cache-warm
4003 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004004 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 * bufctl chains: array operations are cheaper.
4006 * The numbers are guessed, we should auto-tune as described by
4007 * Bonwick.
4008 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004009 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004011 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004013 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004015 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 limit = 54;
4017 else
4018 limit = 120;
4019
Andrew Mortona737b3e2006-03-22 00:08:11 -08004020 /*
4021 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 * allocation behaviour: Most allocs on one cpu, most free operations
4023 * on another cpu. For these cases, an efficient object passing between
4024 * cpus is necessary. This is provided by a shared array. The array
4025 * replaces Bonwick's magazine layer.
4026 * On uniprocessor, it's functionally equivalent (but less efficient)
4027 * to a larger limit. Thus disabled by default.
4028 */
4029 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004030 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032
4033#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004034 /*
4035 * With debugging enabled, large batchcount lead to excessively long
4036 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 */
4038 if (limit > 32)
4039 limit = 32;
4040#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004041 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 if (err)
4043 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004044 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004045 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046}
4047
Christoph Lameter1b552532006-03-22 00:09:07 -08004048/*
4049 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004050 * necessary. Note that the l3 listlock also protects the array_cache
4051 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004052 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004053static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004054 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055{
4056 int tofree;
4057
Christoph Lameter1b552532006-03-22 00:09:07 -08004058 if (!ac || !ac->avail)
4059 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 if (ac->touched && !force) {
4061 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004062 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004063 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004064 if (ac->avail) {
4065 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4066 if (tofree > ac->avail)
4067 tofree = (ac->avail + 1) / 2;
4068 free_block(cachep, ac->entry, tofree, node);
4069 ac->avail -= tofree;
4070 memmove(ac->entry, &(ac->entry[tofree]),
4071 sizeof(void *) * ac->avail);
4072 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004073 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 }
4075}
4076
4077/**
4078 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004079 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080 *
4081 * Called from workqueue/eventd every few seconds.
4082 * Purpose:
4083 * - clear the per-cpu caches for this CPU.
4084 * - return freeable pages to the main free memory pool.
4085 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004086 * If we cannot acquire the cache chain mutex then just give up - we'll try
4087 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004089static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004091 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004092 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004093 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004094 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004096 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004098 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004100 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 check_irq_on();
4102
Christoph Lameter35386e32006-03-22 00:09:05 -08004103 /*
4104 * We only take the l3 lock if absolutely necessary and we
4105 * have established with reasonable certainty that
4106 * we can do some work if the lock was obtained.
4107 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004108 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004109
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004110 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111
Christoph Lameteraab22072006-03-22 00:09:06 -08004112 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113
Christoph Lameter35386e32006-03-22 00:09:05 -08004114 /*
4115 * These are racy checks but it does not matter
4116 * if we skip one check or scan twice.
4117 */
Christoph Lametere498be72005-09-09 13:03:32 -07004118 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004119 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
Christoph Lametere498be72005-09-09 13:03:32 -07004121 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122
Christoph Lameteraab22072006-03-22 00:09:06 -08004123 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124
Christoph Lametered11d9e2006-06-30 01:55:45 -07004125 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004126 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004127 else {
4128 int freed;
4129
4130 freed = drain_freelist(searchp, l3, (l3->free_limit +
4131 5 * searchp->num - 1) / (5 * searchp->num));
4132 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004134next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 cond_resched();
4136 }
4137 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004138 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004139 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004140out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004141 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004142 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143}
4144
Linus Torvalds158a9622008-01-02 13:04:48 -08004145#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146
Pekka Enberg85289f92006-01-08 01:00:36 -08004147static void print_slabinfo_header(struct seq_file *m)
4148{
4149 /*
4150 * Output format version, so at least we can change it
4151 * without _too_ many complaints.
4152 */
4153#if STATS
4154 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4155#else
4156 seq_puts(m, "slabinfo - version: 2.1\n");
4157#endif
4158 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4159 "<objperslab> <pagesperslab>");
4160 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4161 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4162#if STATS
4163 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004164 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004165 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4166#endif
4167 seq_putc(m, '\n');
4168}
4169
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170static void *s_start(struct seq_file *m, loff_t *pos)
4171{
4172 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004174 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004175 if (!n)
4176 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004177
4178 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179}
4180
4181static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4182{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004183 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184}
4185
4186static void s_stop(struct seq_file *m, void *p)
4187{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004188 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189}
4190
4191static int s_show(struct seq_file *m, void *p)
4192{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004193 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004194 struct slab *slabp;
4195 unsigned long active_objs;
4196 unsigned long num_objs;
4197 unsigned long active_slabs = 0;
4198 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004199 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004201 int node;
4202 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004203
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 active_objs = 0;
4205 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004206 for_each_online_node(node) {
4207 l3 = cachep->nodelists[node];
4208 if (!l3)
4209 continue;
4210
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004211 check_irq_on();
4212 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004213
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004214 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004215 if (slabp->inuse != cachep->num && !error)
4216 error = "slabs_full accounting error";
4217 active_objs += cachep->num;
4218 active_slabs++;
4219 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004220 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004221 if (slabp->inuse == cachep->num && !error)
4222 error = "slabs_partial inuse accounting error";
4223 if (!slabp->inuse && !error)
4224 error = "slabs_partial/inuse accounting error";
4225 active_objs += slabp->inuse;
4226 active_slabs++;
4227 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004228 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004229 if (slabp->inuse && !error)
4230 error = "slabs_free/inuse accounting error";
4231 num_slabs++;
4232 }
4233 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004234 if (l3->shared)
4235 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004236
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004237 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004239 num_slabs += active_slabs;
4240 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004241 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004242 error = "free_objects accounting error";
4243
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 if (error)
4246 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4247
4248 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004249 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004250 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004252 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004253 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004254 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004255#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004256 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004257 unsigned long high = cachep->high_mark;
4258 unsigned long allocs = cachep->num_allocations;
4259 unsigned long grown = cachep->grown;
4260 unsigned long reaped = cachep->reaped;
4261 unsigned long errors = cachep->errors;
4262 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004264 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004265 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266
Joe Perchese92dd4f2010-03-26 19:27:58 -07004267 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4268 "%4lu %4lu %4lu %4lu %4lu",
4269 allocs, high, grown,
4270 reaped, errors, max_freeable, node_allocs,
4271 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 }
4273 /* cpu stats */
4274 {
4275 unsigned long allochit = atomic_read(&cachep->allochit);
4276 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4277 unsigned long freehit = atomic_read(&cachep->freehit);
4278 unsigned long freemiss = atomic_read(&cachep->freemiss);
4279
4280 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004281 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282 }
4283#endif
4284 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285 return 0;
4286}
4287
4288/*
4289 * slabinfo_op - iterator that generates /proc/slabinfo
4290 *
4291 * Output layout:
4292 * cache-name
4293 * num-active-objs
4294 * total-objs
4295 * object size
4296 * num-active-slabs
4297 * total-slabs
4298 * num-pages-per-slab
4299 * + further values on SMP and with statistics enabled
4300 */
4301
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004302static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004303 .start = s_start,
4304 .next = s_next,
4305 .stop = s_stop,
4306 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307};
4308
4309#define MAX_SLABINFO_WRITE 128
4310/**
4311 * slabinfo_write - Tuning for the slab allocator
4312 * @file: unused
4313 * @buffer: user buffer
4314 * @count: data length
4315 * @ppos: unused
4316 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004317static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004318 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004319{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004320 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004322 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004323
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324 if (count > MAX_SLABINFO_WRITE)
4325 return -EINVAL;
4326 if (copy_from_user(&kbuf, buffer, count))
4327 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004328 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329
4330 tmp = strchr(kbuf, ' ');
4331 if (!tmp)
4332 return -EINVAL;
4333 *tmp = '\0';
4334 tmp++;
4335 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4336 return -EINVAL;
4337
4338 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004339 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004341 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004342 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004343 if (limit < 1 || batchcount < 1 ||
4344 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004345 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004346 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004347 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004348 batchcount, shared,
4349 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 }
4351 break;
4352 }
4353 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004354 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355 if (res >= 0)
4356 res = count;
4357 return res;
4358}
Al Viro871751e2006-03-25 03:06:39 -08004359
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004360static int slabinfo_open(struct inode *inode, struct file *file)
4361{
4362 return seq_open(file, &slabinfo_op);
4363}
4364
4365static const struct file_operations proc_slabinfo_operations = {
4366 .open = slabinfo_open,
4367 .read = seq_read,
4368 .write = slabinfo_write,
4369 .llseek = seq_lseek,
4370 .release = seq_release,
4371};
4372
Al Viro871751e2006-03-25 03:06:39 -08004373#ifdef CONFIG_DEBUG_SLAB_LEAK
4374
4375static void *leaks_start(struct seq_file *m, loff_t *pos)
4376{
Al Viro871751e2006-03-25 03:06:39 -08004377 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004378 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004379}
4380
4381static inline int add_caller(unsigned long *n, unsigned long v)
4382{
4383 unsigned long *p;
4384 int l;
4385 if (!v)
4386 return 1;
4387 l = n[1];
4388 p = n + 2;
4389 while (l) {
4390 int i = l/2;
4391 unsigned long *q = p + 2 * i;
4392 if (*q == v) {
4393 q[1]++;
4394 return 1;
4395 }
4396 if (*q > v) {
4397 l = i;
4398 } else {
4399 p = q + 2;
4400 l -= i + 1;
4401 }
4402 }
4403 if (++n[1] == n[0])
4404 return 0;
4405 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4406 p[0] = v;
4407 p[1] = 1;
4408 return 1;
4409}
4410
4411static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4412{
4413 void *p;
4414 int i;
4415 if (n[0] == n[1])
4416 return;
4417 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4418 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4419 continue;
4420 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4421 return;
4422 }
4423}
4424
4425static void show_symbol(struct seq_file *m, unsigned long address)
4426{
4427#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004428 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004429 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004430
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004431 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004432 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004433 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004434 seq_printf(m, " [%s]", modname);
4435 return;
4436 }
4437#endif
4438 seq_printf(m, "%p", (void *)address);
4439}
4440
4441static int leaks_show(struct seq_file *m, void *p)
4442{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004443 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004444 struct slab *slabp;
4445 struct kmem_list3 *l3;
4446 const char *name;
4447 unsigned long *n = m->private;
4448 int node;
4449 int i;
4450
4451 if (!(cachep->flags & SLAB_STORE_USER))
4452 return 0;
4453 if (!(cachep->flags & SLAB_RED_ZONE))
4454 return 0;
4455
4456 /* OK, we can do it */
4457
4458 n[1] = 0;
4459
4460 for_each_online_node(node) {
4461 l3 = cachep->nodelists[node];
4462 if (!l3)
4463 continue;
4464
4465 check_irq_on();
4466 spin_lock_irq(&l3->list_lock);
4467
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004468 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004469 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004470 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004471 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004472 spin_unlock_irq(&l3->list_lock);
4473 }
4474 name = cachep->name;
4475 if (n[0] == n[1]) {
4476 /* Increase the buffer size */
4477 mutex_unlock(&cache_chain_mutex);
4478 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4479 if (!m->private) {
4480 /* Too bad, we are really out */
4481 m->private = n;
4482 mutex_lock(&cache_chain_mutex);
4483 return -ENOMEM;
4484 }
4485 *(unsigned long *)m->private = n[0] * 2;
4486 kfree(n);
4487 mutex_lock(&cache_chain_mutex);
4488 /* Now make sure this entry will be retried */
4489 m->count = m->size;
4490 return 0;
4491 }
4492 for (i = 0; i < n[1]; i++) {
4493 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4494 show_symbol(m, n[2*i+2]);
4495 seq_putc(m, '\n');
4496 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004497
Al Viro871751e2006-03-25 03:06:39 -08004498 return 0;
4499}
4500
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004501static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004502 .start = leaks_start,
4503 .next = s_next,
4504 .stop = s_stop,
4505 .show = leaks_show,
4506};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004507
4508static int slabstats_open(struct inode *inode, struct file *file)
4509{
4510 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4511 int ret = -ENOMEM;
4512 if (n) {
4513 ret = seq_open(file, &slabstats_op);
4514 if (!ret) {
4515 struct seq_file *m = file->private_data;
4516 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4517 m->private = n;
4518 n = NULL;
4519 }
4520 kfree(n);
4521 }
4522 return ret;
4523}
4524
4525static const struct file_operations proc_slabstats_operations = {
4526 .open = slabstats_open,
4527 .read = seq_read,
4528 .llseek = seq_lseek,
4529 .release = seq_release_private,
4530};
Al Viro871751e2006-03-25 03:06:39 -08004531#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004532
4533static int __init slab_proc_init(void)
4534{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004535 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004536#ifdef CONFIG_DEBUG_SLAB_LEAK
4537 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4538#endif
4539 return 0;
4540}
4541module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004542#endif
4543
Manfred Spraul00e145b2005-09-03 15:55:07 -07004544/**
4545 * ksize - get the actual amount of memory allocated for a given object
4546 * @objp: Pointer to the object
4547 *
4548 * kmalloc may internally round up allocations and return more memory
4549 * than requested. ksize() can be used to determine the actual amount of
4550 * memory allocated. The caller may use this additional memory, even though
4551 * a smaller amount of memory was initially specified with the kmalloc call.
4552 * The caller must guarantee that objp points to a valid object previously
4553 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4554 * must not be freed during the duration of the call.
4555 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004556size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004557{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004558 BUG_ON(!objp);
4559 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004560 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004561
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004562 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004563}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004564EXPORT_SYMBOL(ksize);