blob: 708efe886154da626cfe01f4a2b6eaf055ada914 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
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 Enberg6ed5eb22006-02-01 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 Enberg6ed5eb22006-02-01 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 Enberg6ed5eb22006-02-01 03:05:49 -0800525 return page_get_slab(page);
526}
527
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800528static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529 unsigned int idx)
530{
531 return slab->s_mem + cache->buffer_size * idx;
532}
533
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800534/*
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 */
540static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800542{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800545}
546
Andrew Mortona737b3e2006-03-22 00:08:11 -0800547/*
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550struct cache_sizes malloc_sizes[] = {
551#define CACHE(x) { .cs_size = (x) },
552#include <linux/kmalloc_sizes.h>
553 CACHE(ULONG_MAX)
554#undef CACHE
555};
556EXPORT_SYMBOL(malloc_sizes);
557
558/* Must match cache_sizes above. Out of line to keep cache footprint low. */
559struct cache_names {
560 char *name;
561 char *name_dma;
562};
563
564static struct cache_names __initdata cache_names[] = {
565#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800567 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568#undef CACHE
569};
570
571static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575
576/* internal cache of cache description objs */
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200577static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
Pekka Enberg343e0d72006-02-01 03:05:50 -0800578static struct kmem_cache cache_cache = {
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200579 .nodelists = cache_cache_nodelists,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800580 .batchcount = 1,
581 .limit = BOOT_CPUCACHE_ENTRIES,
582 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800583 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800584 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585};
586
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700587#define BAD_ALIEN_MAGIC 0x01020304ul
588
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 * chicken and egg problem: delay the per-cpu array allocation
591 * until the general caches are up.
592 */
593static enum {
594 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700595 PARTIAL_AC,
596 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300597 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 FULL
599} g_cpucache_up;
600
Mike Kravetz39d24e62006-05-15 09:44:13 -0700601/*
602 * used by boot code to determine if it can use slab based allocator
603 */
604int slab_is_available(void)
605{
Pekka Enberg8429db52009-06-12 15:58:59 +0300606 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700607}
608
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200609#ifdef CONFIG_LOCKDEP
610
611/*
612 * Slab sometimes uses the kmalloc slabs to store the slab headers
613 * for other slabs "off slab".
614 * The locking for this is tricky in that it nests within the locks
615 * of all other slabs in a few places; to deal with this special
616 * locking we put on-slab caches into a separate lock-class.
617 *
618 * We set lock class for alien array caches which are up during init.
619 * The lock annotation will be lost if all cpus of a node goes down and
620 * then comes back up during hotplug
621 */
622static struct lock_class_key on_slab_l3_key;
623static struct lock_class_key on_slab_alc_key;
624
Peter Zijlstra83835b32011-07-22 15:26:05 +0200625static struct lock_class_key debugobj_l3_key;
626static struct lock_class_key debugobj_alc_key;
627
628static void slab_set_lock_classes(struct kmem_cache *cachep,
629 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
630 int q)
631{
632 struct array_cache **alc;
633 struct kmem_list3 *l3;
634 int r;
635
636 l3 = cachep->nodelists[q];
637 if (!l3)
638 return;
639
640 lockdep_set_class(&l3->list_lock, l3_key);
641 alc = l3->alien;
642 /*
643 * FIXME: This check for BAD_ALIEN_MAGIC
644 * should go away when common slab code is taught to
645 * work even without alien caches.
646 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
647 * for alloc_alien_cache,
648 */
649 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
650 return;
651 for_each_node(r) {
652 if (alc[r])
653 lockdep_set_class(&alc[r]->lock, alc_key);
654 }
655}
656
657static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
658{
659 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
660}
661
662static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
663{
664 int node;
665
666 for_each_online_node(node)
667 slab_set_debugobj_lock_classes_node(cachep, node);
668}
669
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200670static void init_node_lock_keys(int q)
671{
672 struct cache_sizes *s = malloc_sizes;
673
674 if (g_cpucache_up != FULL)
675 return;
676
677 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200678 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200679
680 l3 = s->cs_cachep->nodelists[q];
681 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200682 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200683
684 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
685 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200686 }
687}
688
689static inline void init_lock_keys(void)
690{
691 int node;
692
693 for_each_node(node)
694 init_node_lock_keys(node);
695}
696#else
697static void init_node_lock_keys(int q)
698{
699}
700
701static inline void init_lock_keys(void)
702{
703}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200704
705static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
706{
707}
708
709static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
710{
711}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200712#endif
713
714/*
715 * Guard access to the cache-chain.
716 */
717static DEFINE_MUTEX(cache_chain_mutex);
718static struct list_head cache_chain;
719
Tejun Heo1871e522009-10-29 22:34:13 +0900720static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721
Pekka Enberg343e0d72006-02-01 03:05:50 -0800722static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723{
724 return cachep->array[smp_processor_id()];
725}
726
Andrew Mortona737b3e2006-03-22 00:08:11 -0800727static inline struct kmem_cache *__find_general_cachep(size_t size,
728 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729{
730 struct cache_sizes *csizep = malloc_sizes;
731
732#if DEBUG
733 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800734 * kmem_cache_create(), or __kmalloc(), before
735 * the generic caches are initialized.
736 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700737 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700739 if (!size)
740 return ZERO_SIZE_PTR;
741
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742 while (size > csizep->cs_size)
743 csizep++;
744
745 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700746 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 * has cs_{dma,}cachep==NULL. Thus no special case
748 * for large kmalloc calls required.
749 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800750#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 if (unlikely(gfpflags & GFP_DMA))
752 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800753#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 return csizep->cs_cachep;
755}
756
Adrian Bunkb2213852006-09-25 23:31:02 -0700757static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700758{
759 return __find_general_cachep(size, gfpflags);
760}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700761
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800762static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800764 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
765}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
Andrew Mortona737b3e2006-03-22 00:08:11 -0800767/*
768 * Calculate the number of objects and left-over bytes for a given buffer size.
769 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800770static void cache_estimate(unsigned long gfporder, size_t buffer_size,
771 size_t align, int flags, size_t *left_over,
772 unsigned int *num)
773{
774 int nr_objs;
775 size_t mgmt_size;
776 size_t slab_size = PAGE_SIZE << gfporder;
777
778 /*
779 * The slab management structure can be either off the slab or
780 * on it. For the latter case, the memory allocated for a
781 * slab is used for:
782 *
783 * - The struct slab
784 * - One kmem_bufctl_t for each object
785 * - Padding to respect alignment of @align
786 * - @buffer_size bytes for each object
787 *
788 * If the slab management structure is off the slab, then the
789 * alignment will already be calculated into the size. Because
790 * the slabs are all pages aligned, the objects will be at the
791 * correct alignment when allocated.
792 */
793 if (flags & CFLGS_OFF_SLAB) {
794 mgmt_size = 0;
795 nr_objs = slab_size / buffer_size;
796
797 if (nr_objs > SLAB_LIMIT)
798 nr_objs = SLAB_LIMIT;
799 } else {
800 /*
801 * Ignore padding for the initial guess. The padding
802 * is at most @align-1 bytes, and @buffer_size is at
803 * least @align. In the worst case, this result will
804 * be one greater than the number of objects that fit
805 * into the memory allocation when taking the padding
806 * into account.
807 */
808 nr_objs = (slab_size - sizeof(struct slab)) /
809 (buffer_size + sizeof(kmem_bufctl_t));
810
811 /*
812 * This calculated number will be either the right
813 * amount, or one greater than what we want.
814 */
815 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
816 > slab_size)
817 nr_objs--;
818
819 if (nr_objs > SLAB_LIMIT)
820 nr_objs = SLAB_LIMIT;
821
822 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800824 *num = nr_objs;
825 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826}
827
Harvey Harrisond40cee22008-04-30 00:55:07 -0700828#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829
Andrew Mortona737b3e2006-03-22 00:08:11 -0800830static void __slab_error(const char *function, struct kmem_cache *cachep,
831 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832{
833 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800834 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835 dump_stack();
836}
837
Paul Menage3395ee02006-12-06 20:32:16 -0800838/*
839 * By default on NUMA we use alien caches to stage the freeing of
840 * objects allocated from other nodes. This causes massive memory
841 * inefficiencies when using fake NUMA setup to split memory into a
842 * large number of small nodes, so it can be disabled on the command
843 * line
844 */
845
846static int use_alien_caches __read_mostly = 1;
847static int __init noaliencache_setup(char *s)
848{
849 use_alien_caches = 0;
850 return 1;
851}
852__setup("noaliencache", noaliencache_setup);
853
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800854#ifdef CONFIG_NUMA
855/*
856 * Special reaping functions for NUMA systems called from cache_reap().
857 * These take care of doing round robin flushing of alien caches (containing
858 * objects freed on different nodes from which they were allocated) and the
859 * flushing of remote pcps by calling drain_node_pages.
860 */
Tejun Heo1871e522009-10-29 22:34:13 +0900861static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800862
863static void init_reap_node(int cpu)
864{
865 int node;
866
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700867 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800868 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800869 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800870
Tejun Heo1871e522009-10-29 22:34:13 +0900871 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800872}
873
874static void next_reap_node(void)
875{
Christoph Lameter909ea962010-12-08 16:22:55 +0100876 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800877
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878 node = next_node(node, node_online_map);
879 if (unlikely(node >= MAX_NUMNODES))
880 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100881 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800882}
883
884#else
885#define init_reap_node(cpu) do { } while (0)
886#define next_reap_node(void) do { } while (0)
887#endif
888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889/*
890 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
891 * via the workqueue/eventd.
892 * Add the CPU number into the expiration time to minimize the possibility of
893 * the CPUs getting into lockstep and contending for the global cache chain
894 * lock.
895 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700896static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
Tejun Heo1871e522009-10-29 22:34:13 +0900898 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899
900 /*
901 * When this gets called from do_initcalls via cpucache_init(),
902 * init_workqueues() has already run, so keventd will be setup
903 * at that time.
904 */
David Howells52bad642006-11-22 14:54:01 +0000905 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800906 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700907 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800908 schedule_delayed_work_on(cpu, reap_work,
909 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 }
911}
912
Christoph Lametere498be72005-09-09 13:03:32 -0700913static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300914 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800916 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 struct array_cache *nc = NULL;
918
Pekka Enberg83b519e2009-06-10 19:40:04 +0300919 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100920 /*
921 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300922 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100923 * cache the pointers are not cleared and they could be counted as
924 * valid references during a kmemleak scan. Therefore, kmemleak must
925 * not scan such objects.
926 */
927 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 if (nc) {
929 nc->avail = 0;
930 nc->limit = entries;
931 nc->batchcount = batchcount;
932 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700933 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934 }
935 return nc;
936}
937
Christoph Lameter3ded1752006-03-25 03:06:44 -0800938/*
939 * Transfer objects in one arraycache to another.
940 * Locking must be handled by the caller.
941 *
942 * Return the number of entries transferred.
943 */
944static int transfer_objects(struct array_cache *to,
945 struct array_cache *from, unsigned int max)
946{
947 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700948 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800949
950 if (!nr)
951 return 0;
952
953 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
954 sizeof(void *) *nr);
955
956 from->avail -= nr;
957 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800958 return nr;
959}
960
Christoph Lameter765c4502006-09-27 01:50:08 -0700961#ifndef CONFIG_NUMA
962
963#define drain_alien_cache(cachep, alien) do { } while (0)
964#define reap_alien(cachep, l3) do { } while (0)
965
Pekka Enberg83b519e2009-06-10 19:40:04 +0300966static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700967{
968 return (struct array_cache **)BAD_ALIEN_MAGIC;
969}
970
971static inline void free_alien_cache(struct array_cache **ac_ptr)
972{
973}
974
975static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
976{
977 return 0;
978}
979
980static inline void *alternate_node_alloc(struct kmem_cache *cachep,
981 gfp_t flags)
982{
983 return NULL;
984}
985
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800986static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700987 gfp_t flags, int nodeid)
988{
989 return NULL;
990}
991
992#else /* CONFIG_NUMA */
993
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800994static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800995static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800996
Pekka Enberg83b519e2009-06-10 19:40:04 +0300997static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700998{
999 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001000 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001001 int i;
1002
1003 if (limit > 1)
1004 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001005 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001006 if (ac_ptr) {
1007 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001008 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001009 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001010 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001011 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001012 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001013 kfree(ac_ptr[i]);
1014 kfree(ac_ptr);
1015 return NULL;
1016 }
1017 }
1018 }
1019 return ac_ptr;
1020}
1021
Pekka Enberg5295a742006-02-01 03:05:48 -08001022static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001023{
1024 int i;
1025
1026 if (!ac_ptr)
1027 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001028 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001029 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001030 kfree(ac_ptr);
1031}
1032
Pekka Enberg343e0d72006-02-01 03:05:50 -08001033static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001034 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001035{
1036 struct kmem_list3 *rl3 = cachep->nodelists[node];
1037
1038 if (ac->avail) {
1039 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001040 /*
1041 * Stuff objects into the remote nodes shared array first.
1042 * That way we could avoid the overhead of putting the objects
1043 * into the free lists and getting them back later.
1044 */
shin, jacob693f7d32006-04-28 10:54:37 -05001045 if (rl3->shared)
1046 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001047
Christoph Lameterff694162005-09-22 21:44:02 -07001048 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001049 ac->avail = 0;
1050 spin_unlock(&rl3->list_lock);
1051 }
1052}
1053
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001054/*
1055 * Called from cache_reap() to regularly drain alien caches round robin.
1056 */
1057static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1058{
Christoph Lameter909ea962010-12-08 16:22:55 +01001059 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001060
1061 if (l3->alien) {
1062 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001063
1064 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001065 __drain_alien_cache(cachep, ac, node);
1066 spin_unlock_irq(&ac->lock);
1067 }
1068 }
1069}
1070
Andrew Mortona737b3e2006-03-22 00:08:11 -08001071static void drain_alien_cache(struct kmem_cache *cachep,
1072 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001073{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001074 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001075 struct array_cache *ac;
1076 unsigned long flags;
1077
1078 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001079 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001080 if (ac) {
1081 spin_lock_irqsave(&ac->lock, flags);
1082 __drain_alien_cache(cachep, ac, i);
1083 spin_unlock_irqrestore(&ac->lock, flags);
1084 }
1085 }
1086}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001087
Ingo Molnar873623d2006-07-13 14:44:38 +02001088static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001089{
1090 struct slab *slabp = virt_to_slab(objp);
1091 int nodeid = slabp->nodeid;
1092 struct kmem_list3 *l3;
1093 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001094 int node;
1095
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001096 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001097
1098 /*
1099 * Make sure we are not freeing a object from another node to the array
1100 * cache on this cpu.
1101 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001102 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001103 return 0;
1104
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001105 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001106 STATS_INC_NODEFREES(cachep);
1107 if (l3->alien && l3->alien[nodeid]) {
1108 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001109 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001110 if (unlikely(alien->avail == alien->limit)) {
1111 STATS_INC_ACOVERFLOW(cachep);
1112 __drain_alien_cache(cachep, alien, nodeid);
1113 }
1114 alien->entry[alien->avail++] = objp;
1115 spin_unlock(&alien->lock);
1116 } else {
1117 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1118 free_block(cachep, &objp, 1, nodeid);
1119 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1120 }
1121 return 1;
1122}
Christoph Lametere498be72005-09-09 13:03:32 -07001123#endif
1124
David Rientjes8f9f8d92010-03-27 19:40:47 -07001125/*
1126 * Allocates and initializes nodelists for a node on each slab cache, used for
1127 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1128 * will be allocated off-node since memory is not yet online for the new node.
1129 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1130 * already in use.
1131 *
1132 * Must hold cache_chain_mutex.
1133 */
1134static int init_cache_nodelists_node(int node)
1135{
1136 struct kmem_cache *cachep;
1137 struct kmem_list3 *l3;
1138 const int memsize = sizeof(struct kmem_list3);
1139
1140 list_for_each_entry(cachep, &cache_chain, next) {
1141 /*
1142 * Set up the size64 kmemlist for cpu before we can
1143 * begin anything. Make sure some other cpu on this
1144 * node has not already allocated this
1145 */
1146 if (!cachep->nodelists[node]) {
1147 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1148 if (!l3)
1149 return -ENOMEM;
1150 kmem_list3_init(l3);
1151 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1152 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1153
1154 /*
1155 * The l3s don't come and go as CPUs come and
1156 * go. cache_chain_mutex is sufficient
1157 * protection here.
1158 */
1159 cachep->nodelists[node] = l3;
1160 }
1161
1162 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1163 cachep->nodelists[node]->free_limit =
1164 (1 + nr_cpus_node(node)) *
1165 cachep->batchcount + cachep->num;
1166 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1167 }
1168 return 0;
1169}
1170
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001171static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001173 struct kmem_cache *cachep;
1174 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001175 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301176 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001177
1178 list_for_each_entry(cachep, &cache_chain, next) {
1179 struct array_cache *nc;
1180 struct array_cache *shared;
1181 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001182
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001183 /* cpu is dead; no one can alloc from it. */
1184 nc = cachep->array[cpu];
1185 cachep->array[cpu] = NULL;
1186 l3 = cachep->nodelists[node];
1187
1188 if (!l3)
1189 goto free_array_cache;
1190
1191 spin_lock_irq(&l3->list_lock);
1192
1193 /* Free limit for this kmem_list3 */
1194 l3->free_limit -= cachep->batchcount;
1195 if (nc)
1196 free_block(cachep, nc->entry, nc->avail, node);
1197
Rusty Russell58463c12009-12-17 11:43:12 -06001198 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 spin_unlock_irq(&l3->list_lock);
1200 goto free_array_cache;
1201 }
1202
1203 shared = l3->shared;
1204 if (shared) {
1205 free_block(cachep, shared->entry,
1206 shared->avail, node);
1207 l3->shared = NULL;
1208 }
1209
1210 alien = l3->alien;
1211 l3->alien = NULL;
1212
1213 spin_unlock_irq(&l3->list_lock);
1214
1215 kfree(shared);
1216 if (alien) {
1217 drain_alien_cache(cachep, alien);
1218 free_alien_cache(alien);
1219 }
1220free_array_cache:
1221 kfree(nc);
1222 }
1223 /*
1224 * In the previous loop, all the objects were freed to
1225 * the respective cache's slabs, now we can go ahead and
1226 * shrink each nodelist to its limit.
1227 */
1228 list_for_each_entry(cachep, &cache_chain, next) {
1229 l3 = cachep->nodelists[node];
1230 if (!l3)
1231 continue;
1232 drain_freelist(cachep, l3, l3->free_objects);
1233 }
1234}
1235
1236static int __cpuinit cpuup_prepare(long cpu)
1237{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001238 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001239 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001240 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001241 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001243 /*
1244 * We need to do this right in the beginning since
1245 * alloc_arraycache's are going to use this list.
1246 * kmalloc_node allows us to add the slab to the right
1247 * kmem_list3 and not this cpu's kmem_list3
1248 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001249 err = init_cache_nodelists_node(node);
1250 if (err < 0)
1251 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001252
1253 /*
1254 * Now we can go ahead with allocating the shared arrays and
1255 * array caches
1256 */
1257 list_for_each_entry(cachep, &cache_chain, next) {
1258 struct array_cache *nc;
1259 struct array_cache *shared = NULL;
1260 struct array_cache **alien = NULL;
1261
1262 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001263 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001264 if (!nc)
1265 goto bad;
1266 if (cachep->shared) {
1267 shared = alloc_arraycache(node,
1268 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001269 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001270 if (!shared) {
1271 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001272 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001273 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001274 }
1275 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001276 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001277 if (!alien) {
1278 kfree(shared);
1279 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001281 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001282 }
1283 cachep->array[cpu] = nc;
1284 l3 = cachep->nodelists[node];
1285 BUG_ON(!l3);
1286
1287 spin_lock_irq(&l3->list_lock);
1288 if (!l3->shared) {
1289 /*
1290 * We are serialised from CPU_DEAD or
1291 * CPU_UP_CANCELLED by the cpucontrol lock
1292 */
1293 l3->shared = shared;
1294 shared = NULL;
1295 }
1296#ifdef CONFIG_NUMA
1297 if (!l3->alien) {
1298 l3->alien = alien;
1299 alien = NULL;
1300 }
1301#endif
1302 spin_unlock_irq(&l3->list_lock);
1303 kfree(shared);
1304 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001305 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1306 slab_set_debugobj_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001307 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001308 init_node_lock_keys(node);
1309
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001310 return 0;
1311bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001312 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001313 return -ENOMEM;
1314}
1315
1316static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1317 unsigned long action, void *hcpu)
1318{
1319 long cpu = (long)hcpu;
1320 int err = 0;
1321
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001323 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001324 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001325 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001326 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001327 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 break;
1329 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001330 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 start_cpu_timer(cpu);
1332 break;
1333#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001334 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001335 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001336 /*
1337 * Shutdown cache reaper. Note that the cache_chain_mutex is
1338 * held so that if cache_reap() is invoked it cannot do
1339 * anything expensive but will only modify reap_work
1340 * and reschedule the timer.
1341 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001342 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001343 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001344 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001345 break;
1346 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001347 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001348 start_cpu_timer(cpu);
1349 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001351 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001352 /*
1353 * Even if all the cpus of a node are down, we don't free the
1354 * kmem_list3 of any cache. This to avoid a race between
1355 * cpu_down, and a kmalloc allocation from another cpu for
1356 * memory from the node of the cpu going down. The list3
1357 * structure is usually allocated from kmem_cache_create() and
1358 * gets destroyed at kmem_cache_destroy().
1359 */
Simon Arlott183ff222007-10-20 01:27:18 +02001360 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001361#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001363 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001364 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001365 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001366 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001369 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370}
1371
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001372static struct notifier_block __cpuinitdata cpucache_notifier = {
1373 &cpuup_callback, NULL, 0
1374};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375
David Rientjes8f9f8d92010-03-27 19:40:47 -07001376#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1377/*
1378 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1379 * Returns -EBUSY if all objects cannot be drained so that the node is not
1380 * removed.
1381 *
1382 * Must hold cache_chain_mutex.
1383 */
1384static int __meminit drain_cache_nodelists_node(int node)
1385{
1386 struct kmem_cache *cachep;
1387 int ret = 0;
1388
1389 list_for_each_entry(cachep, &cache_chain, next) {
1390 struct kmem_list3 *l3;
1391
1392 l3 = cachep->nodelists[node];
1393 if (!l3)
1394 continue;
1395
1396 drain_freelist(cachep, l3, l3->free_objects);
1397
1398 if (!list_empty(&l3->slabs_full) ||
1399 !list_empty(&l3->slabs_partial)) {
1400 ret = -EBUSY;
1401 break;
1402 }
1403 }
1404 return ret;
1405}
1406
1407static int __meminit slab_memory_callback(struct notifier_block *self,
1408 unsigned long action, void *arg)
1409{
1410 struct memory_notify *mnb = arg;
1411 int ret = 0;
1412 int nid;
1413
1414 nid = mnb->status_change_nid;
1415 if (nid < 0)
1416 goto out;
1417
1418 switch (action) {
1419 case MEM_GOING_ONLINE:
1420 mutex_lock(&cache_chain_mutex);
1421 ret = init_cache_nodelists_node(nid);
1422 mutex_unlock(&cache_chain_mutex);
1423 break;
1424 case MEM_GOING_OFFLINE:
1425 mutex_lock(&cache_chain_mutex);
1426 ret = drain_cache_nodelists_node(nid);
1427 mutex_unlock(&cache_chain_mutex);
1428 break;
1429 case MEM_ONLINE:
1430 case MEM_OFFLINE:
1431 case MEM_CANCEL_ONLINE:
1432 case MEM_CANCEL_OFFLINE:
1433 break;
1434 }
1435out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001436 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001437}
1438#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1439
Christoph Lametere498be72005-09-09 13:03:32 -07001440/*
1441 * swap the static kmem_list3 with kmalloced memory
1442 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001443static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1444 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001445{
1446 struct kmem_list3 *ptr;
1447
Pekka Enberg83b519e2009-06-10 19:40:04 +03001448 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001449 BUG_ON(!ptr);
1450
Christoph Lametere498be72005-09-09 13:03:32 -07001451 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001452 /*
1453 * Do not assume that spinlocks can be initialized via memcpy:
1454 */
1455 spin_lock_init(&ptr->list_lock);
1456
Christoph Lametere498be72005-09-09 13:03:32 -07001457 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1458 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001459}
1460
Andrew Mortona737b3e2006-03-22 00:08:11 -08001461/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001462 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1463 * size of kmem_list3.
1464 */
1465static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1466{
1467 int node;
1468
1469 for_each_online_node(node) {
1470 cachep->nodelists[node] = &initkmem_list3[index + node];
1471 cachep->nodelists[node]->next_reap = jiffies +
1472 REAPTIMEOUT_LIST3 +
1473 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1474 }
1475}
1476
1477/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001478 * Initialisation. Called after the page allocator have been initialised and
1479 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 */
1481void __init kmem_cache_init(void)
1482{
1483 size_t left_over;
1484 struct cache_sizes *sizes;
1485 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001486 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001487 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001488 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001489
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001490 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001491 use_alien_caches = 0;
1492
Christoph Lametere498be72005-09-09 13:03:32 -07001493 for (i = 0; i < NUM_INIT_LISTS; i++) {
1494 kmem_list3_init(&initkmem_list3[i]);
1495 if (i < MAX_NUMNODES)
1496 cache_cache.nodelists[i] = NULL;
1497 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001498 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
1500 /*
1501 * Fragmentation resistance on low memory - only use bigger
1502 * page orders on machines with more than 32MB of memory.
1503 */
Jan Beulich44813742009-09-21 17:03:05 -07001504 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1506
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 /* Bootstrap is tricky, because several objects are allocated
1508 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001509 * 1) initialize the cache_cache cache: it contains the struct
1510 * kmem_cache structures of all caches, except cache_cache itself:
1511 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001512 * Initially an __init data area is used for the head array and the
1513 * kmem_list3 structures, it's replaced with a kmalloc allocated
1514 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001516 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001517 * An __init data area is used for the head array.
1518 * 3) Create the remaining kmalloc caches, with minimally sized
1519 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 * 4) Replace the __init data head arrays for cache_cache and the first
1521 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001522 * 5) Replace the __init data for kmem_list3 for cache_cache and
1523 * the other cache's with kmalloc allocated memory.
1524 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 */
1526
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001527 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001528
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 INIT_LIST_HEAD(&cache_chain);
1531 list_add(&cache_cache.next, &cache_chain);
1532 cache_cache.colour_off = cache_line_size();
1533 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001534 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535
Eric Dumazet8da34302007-05-06 14:49:29 -07001536 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001537 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001538 */
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001539 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1540 nr_node_ids * sizeof(struct kmem_list3 *);
Eric Dumazet8da34302007-05-06 14:49:29 -07001541#if DEBUG
1542 cache_cache.obj_size = cache_cache.buffer_size;
1543#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001544 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1545 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001546 cache_cache.reciprocal_buffer_size =
1547 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548
Jack Steiner07ed76b2006-03-07 21:55:46 -08001549 for (order = 0; order < MAX_ORDER; order++) {
1550 cache_estimate(order, cache_cache.buffer_size,
1551 cache_line_size(), 0, &left_over, &cache_cache.num);
1552 if (cache_cache.num)
1553 break;
1554 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001555 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001556 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001557 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001558 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1559 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560
1561 /* 2+3) create the kmalloc caches */
1562 sizes = malloc_sizes;
1563 names = cache_names;
1564
Andrew Mortona737b3e2006-03-22 00:08:11 -08001565 /*
1566 * Initialize the caches that provide memory for the array cache and the
1567 * kmem_list3 structures first. Without this, further allocations will
1568 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001569 */
1570
1571 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001572 sizes[INDEX_AC].cs_size,
1573 ARCH_KMALLOC_MINALIGN,
1574 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001575 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001576
Andrew Mortona737b3e2006-03-22 00:08:11 -08001577 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001578 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001579 kmem_cache_create(names[INDEX_L3].name,
1580 sizes[INDEX_L3].cs_size,
1581 ARCH_KMALLOC_MINALIGN,
1582 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001583 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001584 }
Christoph Lametere498be72005-09-09 13:03:32 -07001585
Ingo Molnare0a42722006-06-23 02:03:46 -07001586 slab_early_init = 0;
1587
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001589 /*
1590 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591 * This should be particularly beneficial on SMP boxes, as it
1592 * eliminates "false sharing".
1593 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001594 * allow tighter packing of the smaller caches.
1595 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001597 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001598 sizes->cs_size,
1599 ARCH_KMALLOC_MINALIGN,
1600 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001601 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001602 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001603#ifdef CONFIG_ZONE_DMA
1604 sizes->cs_dmacachep = kmem_cache_create(
1605 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001606 sizes->cs_size,
1607 ARCH_KMALLOC_MINALIGN,
1608 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1609 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001610 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001611#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 sizes++;
1613 names++;
1614 }
1615 /* 4) Replace the bootstrap head arrays */
1616 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001617 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001618
Pekka Enberg83b519e2009-06-10 19:40:04 +03001619 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001620
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001621 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1622 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001623 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001624 /*
1625 * Do not assume that spinlocks can be initialized via memcpy:
1626 */
1627 spin_lock_init(&ptr->lock);
1628
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001630
Pekka Enberg83b519e2009-06-10 19:40:04 +03001631 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001632
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001633 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001634 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001635 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001636 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001637 /*
1638 * Do not assume that spinlocks can be initialized via memcpy:
1639 */
1640 spin_lock_init(&ptr->lock);
1641
Christoph Lametere498be72005-09-09 13:03:32 -07001642 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001643 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 }
Christoph Lametere498be72005-09-09 13:03:32 -07001645 /* 5) Replace the bootstrap kmem_list3's */
1646 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001647 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
Mel Gorman9c09a952008-01-24 05:49:54 -08001649 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001650 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001651
Christoph Lametere498be72005-09-09 13:03:32 -07001652 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001653 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001654
1655 if (INDEX_AC != INDEX_L3) {
1656 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001657 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001658 }
1659 }
1660 }
1661
Pekka Enberg8429db52009-06-12 15:58:59 +03001662 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001663}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001664
Pekka Enberg8429db52009-06-12 15:58:59 +03001665void __init kmem_cache_init_late(void)
1666{
1667 struct kmem_cache *cachep;
1668
Peter Zijlstra30765b92011-07-28 23:22:56 +02001669 /* Annotate slab for lockdep -- annotate the malloc caches */
1670 init_lock_keys();
1671
Pekka Enberg8429db52009-06-12 15:58:59 +03001672 /* 6) resize the head arrays to their final sizes */
1673 mutex_lock(&cache_chain_mutex);
1674 list_for_each_entry(cachep, &cache_chain, next)
1675 if (enable_cpucache(cachep, GFP_NOWAIT))
1676 BUG();
1677 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001678
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 /* Done! */
1680 g_cpucache_up = FULL;
1681
Andrew Mortona737b3e2006-03-22 00:08:11 -08001682 /*
1683 * Register a cpu startup notifier callback that initializes
1684 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 */
1686 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
David Rientjes8f9f8d92010-03-27 19:40:47 -07001688#ifdef CONFIG_NUMA
1689 /*
1690 * Register a memory hotplug callback that initializes and frees
1691 * nodelists.
1692 */
1693 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1694#endif
1695
Andrew Mortona737b3e2006-03-22 00:08:11 -08001696 /*
1697 * The reap timers are started later, with a module init call: That part
1698 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 */
1700}
1701
1702static int __init cpucache_init(void)
1703{
1704 int cpu;
1705
Andrew Mortona737b3e2006-03-22 00:08:11 -08001706 /*
1707 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 */
Christoph Lametere498be72005-09-09 13:03:32 -07001709 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001710 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 return 0;
1712}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713__initcall(cpucache_init);
1714
1715/*
1716 * Interface to system's page allocator. No need to hold the cache-lock.
1717 *
1718 * If we requested dmaable memory, we will get it. Even if we
1719 * did not request dmaable memory, we might get it, but that
1720 * would be relatively rare and ignorable.
1721 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001722static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723{
1724 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001725 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 int i;
1727
Luke Yangd6fef9d2006-04-10 22:52:56 -07001728#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001729 /*
1730 * Nommu uses slab's for process anonymous memory allocations, and thus
1731 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001732 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001733 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001734#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001735
Christoph Lameter3c517a62006-12-06 20:33:29 -08001736 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001737 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1738 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001739
Linus Torvalds517d0862009-06-16 19:50:13 -07001740 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 if (!page)
1742 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001744 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001746 add_zone_page_state(page_zone(page),
1747 NR_SLAB_RECLAIMABLE, nr_pages);
1748 else
1749 add_zone_page_state(page_zone(page),
1750 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001751 for (i = 0; i < nr_pages; i++)
1752 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001753
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001754 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1755 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1756
1757 if (cachep->ctor)
1758 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1759 else
1760 kmemcheck_mark_unallocated_pages(page, nr_pages);
1761 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001762
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001763 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764}
1765
1766/*
1767 * Interface to system's page release.
1768 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001769static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 struct page *page = virt_to_page(addr);
1773 const unsigned long nr_freed = i;
1774
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001775 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001776
Christoph Lameter972d1a72006-09-25 23:31:51 -07001777 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1778 sub_zone_page_state(page_zone(page),
1779 NR_SLAB_RECLAIMABLE, nr_freed);
1780 else
1781 sub_zone_page_state(page_zone(page),
1782 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001784 BUG_ON(!PageSlab(page));
1785 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 page++;
1787 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 if (current->reclaim_state)
1789 current->reclaim_state->reclaimed_slab += nr_freed;
1790 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791}
1792
1793static void kmem_rcu_free(struct rcu_head *head)
1794{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001796 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797
1798 kmem_freepages(cachep, slab_rcu->addr);
1799 if (OFF_SLAB(cachep))
1800 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1801}
1802
1803#if DEBUG
1804
1805#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001806static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001807 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001809 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001811 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001813 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 return;
1815
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001816 *addr++ = 0x12345678;
1817 *addr++ = caller;
1818 *addr++ = smp_processor_id();
1819 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 {
1821 unsigned long *sptr = &caller;
1822 unsigned long svalue;
1823
1824 while (!kstack_end(sptr)) {
1825 svalue = *sptr++;
1826 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 size -= sizeof(unsigned long);
1829 if (size <= sizeof(unsigned long))
1830 break;
1831 }
1832 }
1833
1834 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001835 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836}
1837#endif
1838
Pekka Enberg343e0d72006-02-01 03:05:50 -08001839static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001841 int size = obj_size(cachep);
1842 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843
1844 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001845 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846}
1847
1848static void dump_line(char *data, int offset, int limit)
1849{
1850 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001851 unsigned char error = 0;
1852 int bad_count = 0;
1853
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001854 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001855 for (i = 0; i < limit; i++) {
1856 if (data[offset + i] != POISON_FREE) {
1857 error = data[offset + i];
1858 bad_count++;
1859 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001860 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001861 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1862 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001863
1864 if (bad_count == 1) {
1865 error ^= POISON_FREE;
1866 if (!(error & (error - 1))) {
1867 printk(KERN_ERR "Single bit error detected. Probably "
1868 "bad RAM.\n");
1869#ifdef CONFIG_X86
1870 printk(KERN_ERR "Run memtest86+ or a similar memory "
1871 "test tool.\n");
1872#else
1873 printk(KERN_ERR "Run a memory test tool.\n");
1874#endif
1875 }
1876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877}
1878#endif
1879
1880#if DEBUG
1881
Pekka Enberg343e0d72006-02-01 03:05:50 -08001882static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883{
1884 int i, size;
1885 char *realobj;
1886
1887 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001888 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001889 *dbg_redzone1(cachep, objp),
1890 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 }
1892
1893 if (cachep->flags & SLAB_STORE_USER) {
1894 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001895 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001897 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 printk("\n");
1899 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001900 realobj = (char *)objp + obj_offset(cachep);
1901 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001902 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 int limit;
1904 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001905 if (i + limit > size)
1906 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 dump_line(realobj, i, limit);
1908 }
1909}
1910
Pekka Enberg343e0d72006-02-01 03:05:50 -08001911static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912{
1913 char *realobj;
1914 int size, i;
1915 int lines = 0;
1916
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001917 realobj = (char *)objp + obj_offset(cachep);
1918 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001920 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 exp = POISON_END;
1924 if (realobj[i] != exp) {
1925 int limit;
1926 /* Mismatch ! */
1927 /* Print header */
1928 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001929 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001930 "Slab corruption: %s start=%p, len=%d\n",
1931 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 print_objinfo(cachep, objp, 0);
1933 }
1934 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001937 if (i + limit > size)
1938 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 dump_line(realobj, i, limit);
1940 i += 16;
1941 lines++;
1942 /* Limit to 5 lines */
1943 if (lines > 5)
1944 break;
1945 }
1946 }
1947 if (lines != 0) {
1948 /* Print some data about the neighboring objects, if they
1949 * exist:
1950 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001951 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001952 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001954 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001956 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001957 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001959 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 print_objinfo(cachep, objp, 2);
1961 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001962 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001963 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001964 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001966 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 print_objinfo(cachep, objp, 2);
1968 }
1969 }
1970}
1971#endif
1972
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301974static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001975{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 int i;
1977 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001978 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979
1980 if (cachep->flags & SLAB_POISON) {
1981#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001982 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1983 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001984 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001985 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 else
1987 check_poison_obj(cachep, objp);
1988#else
1989 check_poison_obj(cachep, objp);
1990#endif
1991 }
1992 if (cachep->flags & SLAB_RED_ZONE) {
1993 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1994 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001995 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1997 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001998 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002001}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302003static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002004{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002005}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006#endif
2007
Randy Dunlap911851e2006-03-22 00:08:14 -08002008/**
2009 * slab_destroy - destroy and release all objects in a slab
2010 * @cachep: cache pointer being destroyed
2011 * @slabp: slab pointer being destroyed
2012 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002013 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002014 * Before calling the slab must have been unlinked from the cache. The
2015 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002016 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002017static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002018{
2019 void *addr = slabp->s_mem - slabp->colouroff;
2020
Rabin Vincente79aec22008-07-04 00:40:32 +05302021 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2023 struct slab_rcu *slab_rcu;
2024
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002025 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 slab_rcu->cachep = cachep;
2027 slab_rcu->addr = addr;
2028 call_rcu(&slab_rcu->head, kmem_rcu_free);
2029 } else {
2030 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002031 if (OFF_SLAB(cachep))
2032 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 }
2034}
2035
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002036static void __kmem_cache_destroy(struct kmem_cache *cachep)
2037{
2038 int i;
2039 struct kmem_list3 *l3;
2040
2041 for_each_online_cpu(i)
2042 kfree(cachep->array[i]);
2043
2044 /* NUMA: free the list3 structures */
2045 for_each_online_node(i) {
2046 l3 = cachep->nodelists[i];
2047 if (l3) {
2048 kfree(l3->shared);
2049 free_alien_cache(l3->alien);
2050 kfree(l3);
2051 }
2052 }
2053 kmem_cache_free(&cache_cache, cachep);
2054}
2055
2056
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002058 * calculate_slab_order - calculate size (page order) of slabs
2059 * @cachep: pointer to the cache that is being created
2060 * @size: size of objects to be created in this cache.
2061 * @align: required alignment for the objects.
2062 * @flags: slab allocation flags
2063 *
2064 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002065 *
2066 * This could be made much more intelligent. For now, try to avoid using
2067 * high order pages for slabs. When the gfp() functions are more friendly
2068 * towards high-order requests, this should be changed.
2069 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002070static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002071 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002072{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002073 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002074 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002075 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002076
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002077 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002078 unsigned int num;
2079 size_t remainder;
2080
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002081 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002082 if (!num)
2083 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002084
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002085 if (flags & CFLGS_OFF_SLAB) {
2086 /*
2087 * Max number of objs-per-slab for caches which
2088 * use off-slab slabs. Needed to avoid a possible
2089 * looping condition in cache_grow().
2090 */
2091 offslab_limit = size - sizeof(struct slab);
2092 offslab_limit /= sizeof(kmem_bufctl_t);
2093
2094 if (num > offslab_limit)
2095 break;
2096 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002097
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002098 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002099 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002100 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002101 left_over = remainder;
2102
2103 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002104 * A VFS-reclaimable slab tends to have most allocations
2105 * as GFP_NOFS and we really don't want to have to be allocating
2106 * higher-order pages when we are unable to shrink dcache.
2107 */
2108 if (flags & SLAB_RECLAIM_ACCOUNT)
2109 break;
2110
2111 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002112 * Large number of objects is good, but very large slabs are
2113 * currently bad for the gfp()s.
2114 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002115 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002116 break;
2117
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002118 /*
2119 * Acceptable internal fragmentation?
2120 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002121 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002122 break;
2123 }
2124 return left_over;
2125}
2126
Pekka Enberg83b519e2009-06-10 19:40:04 +03002127static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002128{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002129 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002130 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002131
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002132 if (g_cpucache_up == NONE) {
2133 /*
2134 * Note: the first kmem_cache_create must create the cache
2135 * that's used by kmalloc(24), otherwise the creation of
2136 * further caches will BUG().
2137 */
2138 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2139
2140 /*
2141 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2142 * the first cache, then we need to set up all its list3s,
2143 * otherwise the creation of further caches will BUG().
2144 */
2145 set_up_list3s(cachep, SIZE_AC);
2146 if (INDEX_AC == INDEX_L3)
2147 g_cpucache_up = PARTIAL_L3;
2148 else
2149 g_cpucache_up = PARTIAL_AC;
2150 } else {
2151 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002152 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002153
2154 if (g_cpucache_up == PARTIAL_AC) {
2155 set_up_list3s(cachep, SIZE_L3);
2156 g_cpucache_up = PARTIAL_L3;
2157 } else {
2158 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002159 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002160 cachep->nodelists[node] =
2161 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002162 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002163 BUG_ON(!cachep->nodelists[node]);
2164 kmem_list3_init(cachep->nodelists[node]);
2165 }
2166 }
2167 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002168 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002169 jiffies + REAPTIMEOUT_LIST3 +
2170 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2171
2172 cpu_cache_get(cachep)->avail = 0;
2173 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2174 cpu_cache_get(cachep)->batchcount = 1;
2175 cpu_cache_get(cachep)->touched = 0;
2176 cachep->batchcount = 1;
2177 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002178 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002179}
2180
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002181/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 * kmem_cache_create - Create a cache.
2183 * @name: A string which is used in /proc/slabinfo to identify this cache.
2184 * @size: The size of objects to be created in this cache.
2185 * @align: The required alignment for the objects.
2186 * @flags: SLAB flags
2187 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 *
2189 * Returns a ptr to the cache on success, NULL on failure.
2190 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002191 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 *
2193 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002194 * the module calling this has to destroy the cache before getting unloaded.
2195 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 * The flags are
2197 *
2198 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2199 * to catch references to uninitialised memory.
2200 *
2201 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2202 * for buffer overruns.
2203 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2205 * cacheline. This can be beneficial if you're counting cycles as closely
2206 * as davem.
2207 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002208struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002210 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211{
2212 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002213 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002214 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215
2216 /*
2217 * Sanity checks... these are all serious usage bugs.
2218 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002219 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002220 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002221 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002222 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002223 BUG();
2224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002226 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002227 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302228 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002229 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002230 if (slab_is_available()) {
2231 get_online_cpus();
2232 mutex_lock(&cache_chain_mutex);
2233 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002234
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002235 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002236 char tmp;
2237 int res;
2238
2239 /*
2240 * This happens when the module gets unloaded and doesn't
2241 * destroy its slab cache and no-one else reuses the vmalloc
2242 * area of the module. Print a warning.
2243 */
Andrew Morton138ae662006-12-06 20:36:41 -08002244 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002245 if (res) {
matzeb4169522007-05-06 14:49:52 -07002246 printk(KERN_ERR
2247 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002248 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002249 continue;
2250 }
2251
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002252 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002253 printk(KERN_ERR
2254 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002255 dump_stack();
2256 goto oops;
2257 }
2258 }
2259
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260#if DEBUG
2261 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262#if FORCED_DEBUG
2263 /*
2264 * Enable redzoning and last user accounting, except for caches with
2265 * large objects, if the increased size would increase the object size
2266 * above the next power of two: caches with object sizes just above a
2267 * power of two have a significant amount of internal fragmentation.
2268 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002269 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2270 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002271 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 if (!(flags & SLAB_DESTROY_BY_RCU))
2273 flags |= SLAB_POISON;
2274#endif
2275 if (flags & SLAB_DESTROY_BY_RCU)
2276 BUG_ON(flags & SLAB_POISON);
2277#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002279 * Always checks flags, a caller might be expecting debug support which
2280 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002282 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Andrew Mortona737b3e2006-03-22 00:08:11 -08002284 /*
2285 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 * unaligned accesses for some archs when redzoning is used, and makes
2287 * sure any on-slab bufctl's are also correctly aligned.
2288 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002289 if (size & (BYTES_PER_WORD - 1)) {
2290 size += (BYTES_PER_WORD - 1);
2291 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 }
2293
Andrew Mortona737b3e2006-03-22 00:08:11 -08002294 /* calculate the final buffer alignment: */
2295
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 /* 1) arch recommendation: can be overridden for debug */
2297 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002298 /*
2299 * Default alignment: as specified by the arch code. Except if
2300 * an object is really small, then squeeze multiple objects into
2301 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 */
2303 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002304 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 ralign /= 2;
2306 } else {
2307 ralign = BYTES_PER_WORD;
2308 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002309
2310 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002311 * Redzoning and user store require word alignment or possibly larger.
2312 * Note this will be overridden by architecture or caller mandated
2313 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002314 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002315 if (flags & SLAB_STORE_USER)
2316 ralign = BYTES_PER_WORD;
2317
2318 if (flags & SLAB_RED_ZONE) {
2319 ralign = REDZONE_ALIGN;
2320 /* If redzoning, ensure that the second redzone is suitably
2321 * aligned, by adjusting the object size accordingly. */
2322 size += REDZONE_ALIGN - 1;
2323 size &= ~(REDZONE_ALIGN - 1);
2324 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002325
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002326 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 if (ralign < ARCH_SLAB_MINALIGN) {
2328 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002330 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 if (ralign < align) {
2332 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002334 /* disable debug if necessary */
2335 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002336 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002337 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002338 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 */
2340 align = ralign;
2341
Pekka Enberg83b519e2009-06-10 19:40:04 +03002342 if (slab_is_available())
2343 gfp = GFP_KERNEL;
2344 else
2345 gfp = GFP_NOWAIT;
2346
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002348 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002350 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002352 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002354 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355
Pekka Enbergca5f9702006-09-25 23:31:25 -07002356 /*
2357 * Both debugging options require word-alignment which is calculated
2358 * into align above.
2359 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002362 cachep->obj_offset += sizeof(unsigned long long);
2363 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 }
2365 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002366 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002367 * the real object. But if the second red zone needs to be
2368 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002370 if (flags & SLAB_RED_ZONE)
2371 size += REDZONE_ALIGN;
2372 else
2373 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 }
2375#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002376 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002377 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2378 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 size = PAGE_SIZE;
2380 }
2381#endif
2382#endif
2383
Ingo Molnare0a42722006-06-23 02:03:46 -07002384 /*
2385 * Determine if the slab management is 'on' or 'off' slab.
2386 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002387 * it too early on. Always use on-slab management when
2388 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002389 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002390 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2391 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 /*
2393 * Size is large, assume best to place the slab management obj
2394 * off-slab (should allow better packing of objs).
2395 */
2396 flags |= CFLGS_OFF_SLAB;
2397
2398 size = ALIGN(size, align);
2399
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002400 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
2402 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002403 printk(KERN_ERR
2404 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 kmem_cache_free(&cache_cache, cachep);
2406 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002407 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002409 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2410 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
2412 /*
2413 * If the slab has been placed off-slab, and we have enough space then
2414 * move it on-slab. This is at the expense of any extra colouring.
2415 */
2416 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2417 flags &= ~CFLGS_OFF_SLAB;
2418 left_over -= slab_size;
2419 }
2420
2421 if (flags & CFLGS_OFF_SLAB) {
2422 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002423 slab_size =
2424 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302425
2426#ifdef CONFIG_PAGE_POISONING
2427 /* If we're going to use the generic kernel_map_pages()
2428 * poisoning, then it's going to smash the contents of
2429 * the redzone and userword anyhow, so switch them off.
2430 */
2431 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2432 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2433#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 }
2435
2436 cachep->colour_off = cache_line_size();
2437 /* Offset must be a multiple of the alignment. */
2438 if (cachep->colour_off < align)
2439 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002440 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 cachep->slab_size = slab_size;
2442 cachep->flags = flags;
2443 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002444 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002446 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002447 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002449 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002450 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002451 /*
2452 * This is a possibility for one of the malloc_sizes caches.
2453 * But since we go off slab only for object size greater than
2454 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2455 * this should not happen at all.
2456 * But leave a BUG_ON for some lucky dude.
2457 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002458 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002459 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 cachep->name = name;
2462
Pekka Enberg83b519e2009-06-10 19:40:04 +03002463 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002464 __kmem_cache_destroy(cachep);
2465 cachep = NULL;
2466 goto oops;
2467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Peter Zijlstra83835b32011-07-22 15:26:05 +02002469 if (flags & SLAB_DEBUG_OBJECTS) {
2470 /*
2471 * Would deadlock through slab_destroy()->call_rcu()->
2472 * debug_object_activate()->kmem_cache_alloc().
2473 */
2474 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2475
2476 slab_set_debugobj_lock_classes(cachep);
2477 }
2478
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 /* cache setup completed, link it into the list */
2480 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002481oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 if (!cachep && (flags & SLAB_PANIC))
2483 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002484 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002485 if (slab_is_available()) {
2486 mutex_unlock(&cache_chain_mutex);
2487 put_online_cpus();
2488 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 return cachep;
2490}
2491EXPORT_SYMBOL(kmem_cache_create);
2492
2493#if DEBUG
2494static void check_irq_off(void)
2495{
2496 BUG_ON(!irqs_disabled());
2497}
2498
2499static void check_irq_on(void)
2500{
2501 BUG_ON(irqs_disabled());
2502}
2503
Pekka Enberg343e0d72006-02-01 03:05:50 -08002504static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505{
2506#ifdef CONFIG_SMP
2507 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002508 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509#endif
2510}
Christoph Lametere498be72005-09-09 13:03:32 -07002511
Pekka Enberg343e0d72006-02-01 03:05:50 -08002512static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002513{
2514#ifdef CONFIG_SMP
2515 check_irq_off();
2516 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2517#endif
2518}
2519
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520#else
2521#define check_irq_off() do { } while(0)
2522#define check_irq_on() do { } while(0)
2523#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002524#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525#endif
2526
Christoph Lameteraab22072006-03-22 00:09:06 -08002527static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2528 struct array_cache *ac,
2529 int force, int node);
2530
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531static void do_drain(void *arg)
2532{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002533 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002535 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
2537 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002538 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002539 spin_lock(&cachep->nodelists[node]->list_lock);
2540 free_block(cachep, ac->entry, ac->avail, node);
2541 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 ac->avail = 0;
2543}
2544
Pekka Enberg343e0d72006-02-01 03:05:50 -08002545static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
Christoph Lametere498be72005-09-09 13:03:32 -07002547 struct kmem_list3 *l3;
2548 int node;
2549
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002550 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002552 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002553 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002554 if (l3 && l3->alien)
2555 drain_alien_cache(cachep, l3->alien);
2556 }
2557
2558 for_each_online_node(node) {
2559 l3 = cachep->nodelists[node];
2560 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002561 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002562 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563}
2564
Christoph Lametered11d9e2006-06-30 01:55:45 -07002565/*
2566 * Remove slabs from the list of free slabs.
2567 * Specify the number of slabs to drain in tofree.
2568 *
2569 * Returns the actual number of slabs released.
2570 */
2571static int drain_freelist(struct kmem_cache *cache,
2572 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002574 struct list_head *p;
2575 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577
Christoph Lametered11d9e2006-06-30 01:55:45 -07002578 nr_freed = 0;
2579 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
Christoph Lametered11d9e2006-06-30 01:55:45 -07002581 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002582 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002583 if (p == &l3->slabs_free) {
2584 spin_unlock_irq(&l3->list_lock);
2585 goto out;
2586 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Christoph Lametered11d9e2006-06-30 01:55:45 -07002588 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002590 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591#endif
2592 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002593 /*
2594 * Safe to drop the lock. The slab is no longer linked
2595 * to the cache.
2596 */
2597 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002598 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002599 slab_destroy(cache, slabp);
2600 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002602out:
2603 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604}
2605
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002606/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002607static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002608{
2609 int ret = 0, i = 0;
2610 struct kmem_list3 *l3;
2611
2612 drain_cpu_caches(cachep);
2613
2614 check_irq_on();
2615 for_each_online_node(i) {
2616 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002617 if (!l3)
2618 continue;
2619
2620 drain_freelist(cachep, l3, l3->free_objects);
2621
2622 ret += !list_empty(&l3->slabs_full) ||
2623 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002624 }
2625 return (ret ? 1 : 0);
2626}
2627
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628/**
2629 * kmem_cache_shrink - Shrink a cache.
2630 * @cachep: The cache to shrink.
2631 *
2632 * Releases as many slabs as possible for a cache.
2633 * To help debugging, a zero exit status indicates all slabs were released.
2634 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002635int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002637 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002638 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002640 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002641 mutex_lock(&cache_chain_mutex);
2642 ret = __cache_shrink(cachep);
2643 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002644 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002645 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646}
2647EXPORT_SYMBOL(kmem_cache_shrink);
2648
2649/**
2650 * kmem_cache_destroy - delete a cache
2651 * @cachep: the cache to destroy
2652 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002653 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 *
2655 * It is expected this function will be called by a module when it is
2656 * unloaded. This will remove the cache completely, and avoid a duplicate
2657 * cache being allocated each time a module is loaded and unloaded, if the
2658 * module doesn't have persistent in-kernel storage across loads and unloads.
2659 *
2660 * The cache must be empty before calling this function.
2661 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002662 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 * during the kmem_cache_destroy().
2664 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002665void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002667 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002670 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002671 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 /*
2673 * the chain is never empty, cache_cache is never destroyed
2674 */
2675 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 if (__cache_shrink(cachep)) {
2677 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002679 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002680 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002681 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682 }
2683
2684 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002685 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002687 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002688 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002689 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690}
2691EXPORT_SYMBOL(kmem_cache_destroy);
2692
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002693/*
2694 * Get the memory for a slab management obj.
2695 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2696 * always come from malloc_sizes caches. The slab descriptor cannot
2697 * come from the same cache which is getting created because,
2698 * when we are searching for an appropriate cache for these
2699 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2700 * If we are creating a malloc_sizes cache here it would not be visible to
2701 * kmem_find_general_cachep till the initialization is complete.
2702 * Hence we cannot have slabp_cache same as the original cache.
2703 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002704static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002705 int colour_off, gfp_t local_flags,
2706 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707{
2708 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002709
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 if (OFF_SLAB(cachep)) {
2711 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002712 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002713 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002714 /*
2715 * If the first object in the slab is leaked (it's allocated
2716 * but no one has a reference to it), we want to make sure
2717 * kmemleak does not treat the ->s_mem pointer as a reference
2718 * to the object. Otherwise we will not report the leak.
2719 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002720 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2721 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 if (!slabp)
2723 return NULL;
2724 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002725 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 colour_off += cachep->slab_size;
2727 }
2728 slabp->inuse = 0;
2729 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002730 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002731 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002732 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 return slabp;
2734}
2735
2736static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2737{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002738 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739}
2740
Pekka Enberg343e0d72006-02-01 03:05:50 -08002741static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002742 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743{
2744 int i;
2745
2746 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002747 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748#if DEBUG
2749 /* need to poison the objs? */
2750 if (cachep->flags & SLAB_POISON)
2751 poison_obj(cachep, objp, POISON_FREE);
2752 if (cachep->flags & SLAB_STORE_USER)
2753 *dbg_userword(cachep, objp) = NULL;
2754
2755 if (cachep->flags & SLAB_RED_ZONE) {
2756 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2757 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2758 }
2759 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002760 * Constructors are not allowed to allocate memory from the same
2761 * cache which they are a constructor for. Otherwise, deadlock.
2762 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 */
2764 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002765 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767 if (cachep->flags & SLAB_RED_ZONE) {
2768 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2769 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002770 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002773 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002775 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2776 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002777 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002778 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779#else
2780 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002781 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002783 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002785 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786}
2787
Pekka Enberg343e0d72006-02-01 03:05:50 -08002788static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002790 if (CONFIG_ZONE_DMA_FLAG) {
2791 if (flags & GFP_DMA)
2792 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2793 else
2794 BUG_ON(cachep->gfpflags & GFP_DMA);
2795 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796}
2797
Andrew Mortona737b3e2006-03-22 00:08:11 -08002798static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2799 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002800{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002801 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002802 kmem_bufctl_t next;
2803
2804 slabp->inuse++;
2805 next = slab_bufctl(slabp)[slabp->free];
2806#if DEBUG
2807 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2808 WARN_ON(slabp->nodeid != nodeid);
2809#endif
2810 slabp->free = next;
2811
2812 return objp;
2813}
2814
Andrew Mortona737b3e2006-03-22 00:08:11 -08002815static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2816 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002817{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002818 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002819
2820#if DEBUG
2821 /* Verify that the slab belongs to the intended node */
2822 WARN_ON(slabp->nodeid != nodeid);
2823
Al Viro871751e2006-03-25 03:06:39 -08002824 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002825 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002826 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002827 BUG();
2828 }
2829#endif
2830 slab_bufctl(slabp)[objnr] = slabp->free;
2831 slabp->free = objnr;
2832 slabp->inuse--;
2833}
2834
Pekka Enberg47768742006-06-23 02:03:07 -07002835/*
2836 * Map pages beginning at addr to the given cache and slab. This is required
2837 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002838 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002839 */
2840static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2841 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842{
Pekka Enberg47768742006-06-23 02:03:07 -07002843 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844 struct page *page;
2845
Pekka Enberg47768742006-06-23 02:03:07 -07002846 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002847
Pekka Enberg47768742006-06-23 02:03:07 -07002848 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002849 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002850 nr_pages <<= cache->gfporder;
2851
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002853 page_set_cache(page, cache);
2854 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002856 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857}
2858
2859/*
2860 * Grow (by 1) the number of slabs within a cache. This is called by
2861 * kmem_cache_alloc() when there are no active objs left in a cache.
2862 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002863static int cache_grow(struct kmem_cache *cachep,
2864 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002866 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002867 size_t offset;
2868 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002869 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870
Andrew Mortona737b3e2006-03-22 00:08:11 -08002871 /*
2872 * Be lazy and only check for valid flags here, keeping it out of the
2873 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002875 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2876 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002878 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002880 l3 = cachep->nodelists[nodeid];
2881 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882
2883 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002884 offset = l3->colour_next;
2885 l3->colour_next++;
2886 if (l3->colour_next >= cachep->colour)
2887 l3->colour_next = 0;
2888 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002890 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891
2892 if (local_flags & __GFP_WAIT)
2893 local_irq_enable();
2894
2895 /*
2896 * The test for missing atomic flag is performed here, rather than
2897 * the more obvious place, simply to reduce the critical path length
2898 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2899 * will eventually be caught here (where it matters).
2900 */
2901 kmem_flagcheck(cachep, flags);
2902
Andrew Mortona737b3e2006-03-22 00:08:11 -08002903 /*
2904 * Get mem for the objs. Attempt to allocate a physical page from
2905 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002906 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002907 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002908 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002909 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 goto failed;
2911
2912 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002913 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002914 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002915 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 goto opps1;
2917
Pekka Enberg47768742006-06-23 02:03:07 -07002918 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919
Christoph Lametera35afb82007-05-16 22:10:57 -07002920 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921
2922 if (local_flags & __GFP_WAIT)
2923 local_irq_disable();
2924 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002925 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926
2927 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002928 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002930 l3->free_objects += cachep->num;
2931 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002933opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002935failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 if (local_flags & __GFP_WAIT)
2937 local_irq_disable();
2938 return 0;
2939}
2940
2941#if DEBUG
2942
2943/*
2944 * Perform extra freeing checks:
2945 * - detect bad pointers.
2946 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 */
2948static void kfree_debugcheck(const void *objp)
2949{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 if (!virt_addr_valid(objp)) {
2951 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002952 (unsigned long)objp);
2953 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955}
2956
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002957static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2958{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002959 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002960
2961 redzone1 = *dbg_redzone1(cache, obj);
2962 redzone2 = *dbg_redzone2(cache, obj);
2963
2964 /*
2965 * Redzone is ok.
2966 */
2967 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2968 return;
2969
2970 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2971 slab_error(cache, "double free detected");
2972 else
2973 slab_error(cache, "memory outside object was overwritten");
2974
David Woodhouseb46b8f12007-05-08 00:22:59 -07002975 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002976 obj, redzone1, redzone2);
2977}
2978
Pekka Enberg343e0d72006-02-01 03:05:50 -08002979static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002980 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981{
2982 struct page *page;
2983 unsigned int objnr;
2984 struct slab *slabp;
2985
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002986 BUG_ON(virt_to_cache(objp) != cachep);
2987
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002988 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002990 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
Pekka Enberg065d41c2005-11-13 16:06:46 -08002992 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993
2994 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002995 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2998 }
2999 if (cachep->flags & SLAB_STORE_USER)
3000 *dbg_userword(cachep, objp) = caller;
3001
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003002 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
3004 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003005 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006
Al Viro871751e2006-03-25 03:06:39 -08003007#ifdef CONFIG_DEBUG_SLAB_LEAK
3008 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3009#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 if (cachep->flags & SLAB_POISON) {
3011#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003014 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003015 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 } else {
3017 poison_obj(cachep, objp, POISON_FREE);
3018 }
3019#else
3020 poison_obj(cachep, objp, POISON_FREE);
3021#endif
3022 }
3023 return objp;
3024}
3025
Pekka Enberg343e0d72006-02-01 03:05:50 -08003026static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027{
3028 kmem_bufctl_t i;
3029 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003030
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 /* Check slab's freelist to see if this obj is there. */
3032 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3033 entries++;
3034 if (entries > cachep->num || i >= cachep->num)
3035 goto bad;
3036 }
3037 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003038bad:
3039 printk(KERN_ERR "slab: Internal list corruption detected in "
3040 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3041 cachep->name, cachep->num, slabp, slabp->inuse);
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003042 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3043 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3044 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 BUG();
3046 }
3047}
3048#else
3049#define kfree_debugcheck(x) do { } while(0)
3050#define cache_free_debugcheck(x,objp,z) (objp)
3051#define check_slabp(x,y) do { } while(0)
3052#endif
3053
Pekka Enberg343e0d72006-02-01 03:05:50 -08003054static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055{
3056 int batchcount;
3057 struct kmem_list3 *l3;
3058 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003059 int node;
3060
Andrew Mortona737b3e2006-03-22 00:08:11 -08003061retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003062 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003063 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003064 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065 batchcount = ac->batchcount;
3066 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003067 /*
3068 * If there was little recent activity on this cache, then
3069 * perform only a partial refill. Otherwise we could generate
3070 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 */
3072 batchcount = BATCHREFILL_LIMIT;
3073 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003074 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075
Christoph Lametere498be72005-09-09 13:03:32 -07003076 BUG_ON(ac->avail > 0 || !l3);
3077 spin_lock(&l3->list_lock);
3078
Christoph Lameter3ded1752006-03-25 03:06:44 -08003079 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003080 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3081 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003082 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003083 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003084
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 while (batchcount > 0) {
3086 struct list_head *entry;
3087 struct slab *slabp;
3088 /* Get slab alloc is to come from. */
3089 entry = l3->slabs_partial.next;
3090 if (entry == &l3->slabs_partial) {
3091 l3->free_touched = 1;
3092 entry = l3->slabs_free.next;
3093 if (entry == &l3->slabs_free)
3094 goto must_grow;
3095 }
3096
3097 slabp = list_entry(entry, struct slab, list);
3098 check_slabp(cachep, slabp);
3099 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003100
3101 /*
3102 * The slab was either on partial or free list so
3103 * there must be at least one object available for
3104 * allocation.
3105 */
roel kluin249b9f32008-10-29 17:18:07 -04003106 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07003107
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 STATS_INC_ALLOCED(cachep);
3110 STATS_INC_ACTIVE(cachep);
3111 STATS_SET_HIGH(cachep);
3112
Matthew Dobson78d382d2006-02-01 03:05:47 -08003113 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003114 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 }
3116 check_slabp(cachep, slabp);
3117
3118 /* move slabp to correct slabp list: */
3119 list_del(&slabp->list);
3120 if (slabp->free == BUFCTL_END)
3121 list_add(&slabp->list, &l3->slabs_full);
3122 else
3123 list_add(&slabp->list, &l3->slabs_partial);
3124 }
3125
Andrew Mortona737b3e2006-03-22 00:08:11 -08003126must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003128alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003129 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130
3131 if (unlikely(!ac->avail)) {
3132 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003133 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003134
Andrew Mortona737b3e2006-03-22 00:08:11 -08003135 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003136 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003137 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 return NULL;
3139
Andrew Mortona737b3e2006-03-22 00:08:11 -08003140 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 goto retry;
3142 }
3143 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003144 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145}
3146
Andrew Mortona737b3e2006-03-22 00:08:11 -08003147static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3148 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149{
3150 might_sleep_if(flags & __GFP_WAIT);
3151#if DEBUG
3152 kmem_flagcheck(cachep, flags);
3153#endif
3154}
3155
3156#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003157static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3158 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003160 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003162 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003164 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003165 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003166 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 else
3168 check_poison_obj(cachep, objp);
3169#else
3170 check_poison_obj(cachep, objp);
3171#endif
3172 poison_obj(cachep, objp, POISON_INUSE);
3173 }
3174 if (cachep->flags & SLAB_STORE_USER)
3175 *dbg_userword(cachep, objp) = caller;
3176
3177 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003178 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3179 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3180 slab_error(cachep, "double free, or memory outside"
3181 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003182 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003183 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003184 objp, *dbg_redzone1(cachep, objp),
3185 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 }
3187 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3188 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3189 }
Al Viro871751e2006-03-25 03:06:39 -08003190#ifdef CONFIG_DEBUG_SLAB_LEAK
3191 {
3192 struct slab *slabp;
3193 unsigned objnr;
3194
Christoph Lameterb49af682007-05-06 14:49:41 -07003195 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003196 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3197 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3198 }
3199#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003200 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003201 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003202 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003203 if (ARCH_SLAB_MINALIGN &&
3204 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003205 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003206 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003207 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 return objp;
3209}
3210#else
3211#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3212#endif
3213
Akinobu Mita773ff602008-12-23 19:37:01 +09003214static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003215{
3216 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003217 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003218
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003219 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003220}
3221
Pekka Enberg343e0d72006-02-01 03:05:50 -08003222static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003224 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 struct array_cache *ac;
3226
Alok N Kataria5c382302005-09-27 21:45:46 -07003227 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003228
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003229 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 if (likely(ac->avail)) {
3231 STATS_INC_ALLOCHIT(cachep);
3232 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003233 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 } else {
3235 STATS_INC_ALLOCMISS(cachep);
3236 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003237 /*
3238 * the 'ac' may be updated by cache_alloc_refill(),
3239 * and kmemleak_erase() requires its correct value.
3240 */
3241 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003243 /*
3244 * To avoid a false negative, if an object that is in one of the
3245 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3246 * treat the array pointers as a reference to the object.
3247 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003248 if (objp)
3249 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003250 return objp;
3251}
3252
Christoph Lametere498be72005-09-09 13:03:32 -07003253#ifdef CONFIG_NUMA
3254/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003255 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003256 *
3257 * If we are in_interrupt, then process context, including cpusets and
3258 * mempolicy, may not apply and should not be used for allocation policy.
3259 */
3260static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3261{
3262 int nid_alloc, nid_here;
3263
Christoph Lameter765c4502006-09-27 01:50:08 -07003264 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003265 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003266 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003267 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003268 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003269 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003270 else if (current->mempolicy)
3271 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003272 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003273 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003274 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003275 return NULL;
3276}
3277
3278/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003279 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003280 * certain node and fall back is permitted. First we scan all the
3281 * available nodelists for available objects. If that fails then we
3282 * perform an allocation without specifying a node. This allows the page
3283 * allocator to do its reclaim / fallback magic. We then insert the
3284 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003285 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003286static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003287{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003288 struct zonelist *zonelist;
3289 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003290 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003291 struct zone *zone;
3292 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003293 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003294 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003295
3296 if (flags & __GFP_THISNODE)
3297 return NULL;
3298
Miao Xiec0ff7452010-05-24 14:32:08 -07003299 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003300 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003301 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003302
Christoph Lameter3c517a62006-12-06 20:33:29 -08003303retry:
3304 /*
3305 * Look through allowed nodes for objects available
3306 * from existing per node queues.
3307 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003308 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3309 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003310
Mel Gorman54a6eb52008-04-28 02:12:16 -07003311 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003312 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003313 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003314 obj = ____cache_alloc_node(cache,
3315 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003316 if (obj)
3317 break;
3318 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003319 }
3320
Christoph Lametercfce6602007-05-06 14:50:17 -07003321 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003322 /*
3323 * This allocation will be performed within the constraints
3324 * of the current cpuset / memory policy requirements.
3325 * We may trigger various forms of reclaim on the allowed
3326 * set and go into memory reserves if necessary.
3327 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003328 if (local_flags & __GFP_WAIT)
3329 local_irq_enable();
3330 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003331 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003332 if (local_flags & __GFP_WAIT)
3333 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003334 if (obj) {
3335 /*
3336 * Insert into the appropriate per node queues
3337 */
3338 nid = page_to_nid(virt_to_page(obj));
3339 if (cache_grow(cache, flags, nid, obj)) {
3340 obj = ____cache_alloc_node(cache,
3341 flags | GFP_THISNODE, nid);
3342 if (!obj)
3343 /*
3344 * Another processor may allocate the
3345 * objects in the slab since we are
3346 * not holding any locks.
3347 */
3348 goto retry;
3349 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003350 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003351 obj = NULL;
3352 }
3353 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003354 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003355 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003356 return obj;
3357}
3358
3359/*
Christoph Lametere498be72005-09-09 13:03:32 -07003360 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003362static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003363 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003364{
3365 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003366 struct slab *slabp;
3367 struct kmem_list3 *l3;
3368 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003369 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003371 l3 = cachep->nodelists[nodeid];
3372 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003373
Andrew Mortona737b3e2006-03-22 00:08:11 -08003374retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003375 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003376 spin_lock(&l3->list_lock);
3377 entry = l3->slabs_partial.next;
3378 if (entry == &l3->slabs_partial) {
3379 l3->free_touched = 1;
3380 entry = l3->slabs_free.next;
3381 if (entry == &l3->slabs_free)
3382 goto must_grow;
3383 }
Christoph Lametere498be72005-09-09 13:03:32 -07003384
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003385 slabp = list_entry(entry, struct slab, list);
3386 check_spinlock_acquired_node(cachep, nodeid);
3387 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003388
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003389 STATS_INC_NODEALLOCS(cachep);
3390 STATS_INC_ACTIVE(cachep);
3391 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003392
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003393 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003394
Matthew Dobson78d382d2006-02-01 03:05:47 -08003395 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003396 check_slabp(cachep, slabp);
3397 l3->free_objects--;
3398 /* move slabp to correct slabp list: */
3399 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003400
Andrew Mortona737b3e2006-03-22 00:08:11 -08003401 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003402 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003403 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003404 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003405
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003406 spin_unlock(&l3->list_lock);
3407 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003408
Andrew Mortona737b3e2006-03-22 00:08:11 -08003409must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003410 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003411 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003412 if (x)
3413 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003414
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003415 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003416
Andrew Mortona737b3e2006-03-22 00:08:11 -08003417done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003418 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003419}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003420
3421/**
3422 * kmem_cache_alloc_node - Allocate an object on the specified node
3423 * @cachep: The cache to allocate from.
3424 * @flags: See kmalloc().
3425 * @nodeid: node number of the target node.
3426 * @caller: return address of caller, used for debug information
3427 *
3428 * Identical to kmem_cache_alloc but it will allocate memory on the given
3429 * node, which can improve the performance for cpu bound structures.
3430 *
3431 * Fallback to other node is possible if __GFP_THISNODE is not set.
3432 */
3433static __always_inline void *
3434__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3435 void *caller)
3436{
3437 unsigned long save_flags;
3438 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003439 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003440
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003441 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003442
Nick Piggincf40bd12009-01-21 08:12:39 +01003443 lockdep_trace_alloc(flags);
3444
Akinobu Mita773ff602008-12-23 19:37:01 +09003445 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003446 return NULL;
3447
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003448 cache_alloc_debugcheck_before(cachep, flags);
3449 local_irq_save(save_flags);
3450
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003451 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003452 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003453
3454 if (unlikely(!cachep->nodelists[nodeid])) {
3455 /* Node not bootstrapped yet */
3456 ptr = fallback_alloc(cachep, flags);
3457 goto out;
3458 }
3459
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003460 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003461 /*
3462 * Use the locally cached objects if possible.
3463 * However ____cache_alloc does not allow fallback
3464 * to other nodes. It may fail while we still have
3465 * objects on other nodes available.
3466 */
3467 ptr = ____cache_alloc(cachep, flags);
3468 if (ptr)
3469 goto out;
3470 }
3471 /* ___cache_alloc_node can fall back to other nodes */
3472 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3473 out:
3474 local_irq_restore(save_flags);
3475 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003476 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3477 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003478
Pekka Enbergc175eea2008-05-09 20:35:53 +02003479 if (likely(ptr))
3480 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3481
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003482 if (unlikely((flags & __GFP_ZERO) && ptr))
3483 memset(ptr, 0, obj_size(cachep));
3484
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003485 return ptr;
3486}
3487
3488static __always_inline void *
3489__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3490{
3491 void *objp;
3492
3493 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3494 objp = alternate_node_alloc(cache, flags);
3495 if (objp)
3496 goto out;
3497 }
3498 objp = ____cache_alloc(cache, flags);
3499
3500 /*
3501 * We may just have run out of memory on the local node.
3502 * ____cache_alloc_node() knows how to locate memory on other nodes
3503 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003504 if (!objp)
3505 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003506
3507 out:
3508 return objp;
3509}
3510#else
3511
3512static __always_inline void *
3513__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3514{
3515 return ____cache_alloc(cachep, flags);
3516}
3517
3518#endif /* CONFIG_NUMA */
3519
3520static __always_inline void *
3521__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3522{
3523 unsigned long save_flags;
3524 void *objp;
3525
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003526 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003527
Nick Piggincf40bd12009-01-21 08:12:39 +01003528 lockdep_trace_alloc(flags);
3529
Akinobu Mita773ff602008-12-23 19:37:01 +09003530 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003531 return NULL;
3532
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003533 cache_alloc_debugcheck_before(cachep, flags);
3534 local_irq_save(save_flags);
3535 objp = __do_cache_alloc(cachep, flags);
3536 local_irq_restore(save_flags);
3537 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003538 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3539 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003540 prefetchw(objp);
3541
Pekka Enbergc175eea2008-05-09 20:35:53 +02003542 if (likely(objp))
3543 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3544
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003545 if (unlikely((flags & __GFP_ZERO) && objp))
3546 memset(objp, 0, obj_size(cachep));
3547
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003548 return objp;
3549}
Christoph Lametere498be72005-09-09 13:03:32 -07003550
3551/*
3552 * Caller needs to acquire correct kmem_list's list_lock
3553 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003554static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003555 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556{
3557 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003558 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559
3560 for (i = 0; i < nr_objects; i++) {
3561 void *objp = objpp[i];
3562 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003564 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003565 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003567 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003569 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003571 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 check_slabp(cachep, slabp);
3573
3574 /* fixup slab chains */
3575 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003576 if (l3->free_objects > l3->free_limit) {
3577 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003578 /* No need to drop any previously held
3579 * lock here, even if we have a off-slab slab
3580 * descriptor it is guaranteed to come from
3581 * a different cache, refer to comments before
3582 * alloc_slabmgmt.
3583 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 slab_destroy(cachep, slabp);
3585 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003586 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 }
3588 } else {
3589 /* Unconditionally move a slab to the end of the
3590 * partial list on free - maximum time for the
3591 * other objects to be freed, too.
3592 */
Christoph Lametere498be72005-09-09 13:03:32 -07003593 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 }
3595 }
3596}
3597
Pekka Enberg343e0d72006-02-01 03:05:50 -08003598static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599{
3600 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003601 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003602 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603
3604 batchcount = ac->batchcount;
3605#if DEBUG
3606 BUG_ON(!batchcount || batchcount > ac->avail);
3607#endif
3608 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003609 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003610 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003611 if (l3->shared) {
3612 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003613 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 if (max) {
3615 if (batchcount > max)
3616 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003617 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003618 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 shared_array->avail += batchcount;
3620 goto free_done;
3621 }
3622 }
3623
Christoph Lameterff694162005-09-22 21:44:02 -07003624 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003625free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626#if STATS
3627 {
3628 int i = 0;
3629 struct list_head *p;
3630
Christoph Lametere498be72005-09-09 13:03:32 -07003631 p = l3->slabs_free.next;
3632 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633 struct slab *slabp;
3634
3635 slabp = list_entry(p, struct slab, list);
3636 BUG_ON(slabp->inuse);
3637
3638 i++;
3639 p = p->next;
3640 }
3641 STATS_SET_FREEABLE(cachep, i);
3642 }
3643#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003644 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003646 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647}
3648
3649/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003650 * Release an obj back to its cache. If the obj has a constructed state, it must
3651 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003653static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3654 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003656 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657
3658 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003659 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003660 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661
Pekka Enbergc175eea2008-05-09 20:35:53 +02003662 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3663
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003664 /*
3665 * Skip calling cache_free_alien() when the platform is not numa.
3666 * This will avoid cache misses that happen while accessing slabp (which
3667 * is per page memory reference) to get nodeid. Instead use a global
3668 * variable to skip the call, which is mostly likely to be present in
3669 * the cache.
3670 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003671 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003672 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003673
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 if (likely(ac->avail < ac->limit)) {
3675 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003676 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 return;
3678 } else {
3679 STATS_INC_FREEMISS(cachep);
3680 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003681 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682 }
3683}
3684
3685/**
3686 * kmem_cache_alloc - Allocate an object
3687 * @cachep: The cache to allocate from.
3688 * @flags: See kmalloc().
3689 *
3690 * Allocate an object from this cache. The flags are only relevant
3691 * if the cache has no available objects.
3692 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003693void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003695 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3696
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003697 trace_kmem_cache_alloc(_RET_IP_, ret,
3698 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003699
3700 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701}
3702EXPORT_SYMBOL(kmem_cache_alloc);
3703
Li Zefan0f24f122009-12-11 15:45:30 +08003704#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003705void *
3706kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003707{
Steven Rostedt85beb582010-11-24 16:23:34 -05003708 void *ret;
3709
3710 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3711
3712 trace_kmalloc(_RET_IP_, ret,
3713 size, slab_buffer_size(cachep), flags);
3714 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003715}
Steven Rostedt85beb582010-11-24 16:23:34 -05003716EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003717#endif
3718
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003720void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3721{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003722 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3723 __builtin_return_address(0));
3724
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003725 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3726 obj_size(cachep), cachep->buffer_size,
3727 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003728
3729 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003730}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731EXPORT_SYMBOL(kmem_cache_alloc_node);
3732
Li Zefan0f24f122009-12-11 15:45:30 +08003733#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003734void *kmem_cache_alloc_node_trace(size_t size,
3735 struct kmem_cache *cachep,
3736 gfp_t flags,
3737 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003738{
Steven Rostedt85beb582010-11-24 16:23:34 -05003739 void *ret;
3740
3741 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003742 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003743 trace_kmalloc_node(_RET_IP_, ret,
3744 size, slab_buffer_size(cachep),
3745 flags, nodeid);
3746 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003747}
Steven Rostedt85beb582010-11-24 16:23:34 -05003748EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003749#endif
3750
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003751static __always_inline void *
3752__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003753{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003754 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003755
3756 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003757 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3758 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003759 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003760}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003761
Li Zefan0bb38a52009-12-11 15:45:50 +08003762#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003763void *__kmalloc_node(size_t size, gfp_t flags, int node)
3764{
3765 return __do_kmalloc_node(size, flags, node,
3766 __builtin_return_address(0));
3767}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003768EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003769
3770void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003771 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003772{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003773 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003774}
3775EXPORT_SYMBOL(__kmalloc_node_track_caller);
3776#else
3777void *__kmalloc_node(size_t size, gfp_t flags, int node)
3778{
3779 return __do_kmalloc_node(size, flags, node, NULL);
3780}
3781EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003782#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003783#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784
3785/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003786 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003788 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003789 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003791static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3792 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003794 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003795 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003797 /* If you want to save a few bytes .text space: replace
3798 * __ with kmem_.
3799 * Then kmalloc uses the uninlined functions instead of the inline
3800 * functions.
3801 */
3802 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003803 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3804 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003805 ret = __cache_alloc(cachep, flags, caller);
3806
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003807 trace_kmalloc((unsigned long) caller, ret,
3808 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003809
3810 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003811}
3812
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003813
Li Zefan0bb38a52009-12-11 15:45:50 +08003814#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003815void *__kmalloc(size_t size, gfp_t flags)
3816{
Al Viro871751e2006-03-25 03:06:39 -08003817 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818}
3819EXPORT_SYMBOL(__kmalloc);
3820
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003821void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003822{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003823 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003824}
3825EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003826
3827#else
3828void *__kmalloc(size_t size, gfp_t flags)
3829{
3830 return __do_kmalloc(size, flags, NULL);
3831}
3832EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003833#endif
3834
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835/**
3836 * kmem_cache_free - Deallocate an object
3837 * @cachep: The cache the allocation was from.
3838 * @objp: The previously allocated object.
3839 *
3840 * Free an object which was previously allocated from this
3841 * cache.
3842 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003843void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844{
3845 unsigned long flags;
3846
3847 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003848 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003849 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3850 debug_check_no_obj_freed(objp, obj_size(cachep));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003851 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003853
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003854 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855}
3856EXPORT_SYMBOL(kmem_cache_free);
3857
3858/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003859 * kfree - free previously allocated memory
3860 * @objp: pointer returned by kmalloc.
3861 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003862 * If @objp is NULL, no operation is performed.
3863 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 * Don't free memory not originally allocated by kmalloc()
3865 * or you will run into trouble.
3866 */
3867void kfree(const void *objp)
3868{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003869 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 unsigned long flags;
3871
Pekka Enberg2121db72009-03-25 11:05:57 +02003872 trace_kfree(_RET_IP_, objp);
3873
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003874 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 return;
3876 local_irq_save(flags);
3877 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003878 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003879 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003880 debug_check_no_obj_freed(objp, obj_size(c));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003881 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882 local_irq_restore(flags);
3883}
3884EXPORT_SYMBOL(kfree);
3885
Pekka Enberg343e0d72006-02-01 03:05:50 -08003886unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003888 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889}
3890EXPORT_SYMBOL(kmem_cache_size);
3891
Christoph Lametere498be72005-09-09 13:03:32 -07003892/*
Simon Arlott183ff222007-10-20 01:27:18 +02003893 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003894 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003895static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003896{
3897 int node;
3898 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003899 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003900 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003901
Mel Gorman9c09a952008-01-24 05:49:54 -08003902 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003903
Paul Menage3395ee02006-12-06 20:32:16 -08003904 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003905 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003906 if (!new_alien)
3907 goto fail;
3908 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003909
Eric Dumazet63109842007-05-06 14:49:28 -07003910 new_shared = NULL;
3911 if (cachep->shared) {
3912 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003913 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003914 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003915 if (!new_shared) {
3916 free_alien_cache(new_alien);
3917 goto fail;
3918 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003919 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003920
Andrew Mortona737b3e2006-03-22 00:08:11 -08003921 l3 = cachep->nodelists[node];
3922 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003923 struct array_cache *shared = l3->shared;
3924
Christoph Lametere498be72005-09-09 13:03:32 -07003925 spin_lock_irq(&l3->list_lock);
3926
Christoph Lametercafeb022006-03-25 03:06:46 -08003927 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003928 free_block(cachep, shared->entry,
3929 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003930
Christoph Lametercafeb022006-03-25 03:06:46 -08003931 l3->shared = new_shared;
3932 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003933 l3->alien = new_alien;
3934 new_alien = NULL;
3935 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003936 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003937 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003938 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003939 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003940 free_alien_cache(new_alien);
3941 continue;
3942 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003943 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003944 if (!l3) {
3945 free_alien_cache(new_alien);
3946 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003947 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003948 }
Christoph Lametere498be72005-09-09 13:03:32 -07003949
3950 kmem_list3_init(l3);
3951 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003952 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003953 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003954 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003955 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003956 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003957 cachep->nodelists[node] = l3;
3958 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003959 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003960
Andrew Mortona737b3e2006-03-22 00:08:11 -08003961fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003962 if (!cachep->next.next) {
3963 /* Cache is not active yet. Roll back what we did */
3964 node--;
3965 while (node >= 0) {
3966 if (cachep->nodelists[node]) {
3967 l3 = cachep->nodelists[node];
3968
3969 kfree(l3->shared);
3970 free_alien_cache(l3->alien);
3971 kfree(l3);
3972 cachep->nodelists[node] = NULL;
3973 }
3974 node--;
3975 }
3976 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003977 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003978}
3979
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003981 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003982 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983};
3984
3985static void do_ccupdate_local(void *info)
3986{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003987 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 struct array_cache *old;
3989
3990 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003991 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003992
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3994 new->new[smp_processor_id()] = old;
3995}
3996
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003997/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003998static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003999 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004001 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004002 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004004 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4005 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004006 if (!new)
4007 return -ENOMEM;
4008
Christoph Lametere498be72005-09-09 13:03:32 -07004009 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004010 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004011 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004012 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004013 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004014 kfree(new->new[i]);
4015 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004016 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 }
4018 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004019 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004021 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004022
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024 cachep->batchcount = batchcount;
4025 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004026 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
Christoph Lametere498be72005-09-09 13:03:32 -07004028 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004029 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 if (!ccold)
4031 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004032 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4033 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4034 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 kfree(ccold);
4036 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004037 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004038 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039}
4040
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004041/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004042static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043{
4044 int err;
4045 int limit, shared;
4046
Andrew Mortona737b3e2006-03-22 00:08:11 -08004047 /*
4048 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 * - create a LIFO ordering, i.e. return objects that are cache-warm
4050 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004051 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052 * bufctl chains: array operations are cheaper.
4053 * The numbers are guessed, we should auto-tune as described by
4054 * Bonwick.
4055 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004056 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004058 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004060 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004062 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 limit = 54;
4064 else
4065 limit = 120;
4066
Andrew Mortona737b3e2006-03-22 00:08:11 -08004067 /*
4068 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069 * allocation behaviour: Most allocs on one cpu, most free operations
4070 * on another cpu. For these cases, an efficient object passing between
4071 * cpus is necessary. This is provided by a shared array. The array
4072 * replaces Bonwick's magazine layer.
4073 * On uniprocessor, it's functionally equivalent (but less efficient)
4074 * to a larger limit. Thus disabled by default.
4075 */
4076 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004077 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079
4080#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004081 /*
4082 * With debugging enabled, large batchcount lead to excessively long
4083 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084 */
4085 if (limit > 32)
4086 limit = 32;
4087#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004088 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 if (err)
4090 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004091 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004092 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004093}
4094
Christoph Lameter1b552532006-03-22 00:09:07 -08004095/*
4096 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004097 * necessary. Note that the l3 listlock also protects the array_cache
4098 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004099 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004100static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004101 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102{
4103 int tofree;
4104
Christoph Lameter1b552532006-03-22 00:09:07 -08004105 if (!ac || !ac->avail)
4106 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 if (ac->touched && !force) {
4108 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004109 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004110 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004111 if (ac->avail) {
4112 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4113 if (tofree > ac->avail)
4114 tofree = (ac->avail + 1) / 2;
4115 free_block(cachep, ac->entry, tofree, node);
4116 ac->avail -= tofree;
4117 memmove(ac->entry, &(ac->entry[tofree]),
4118 sizeof(void *) * ac->avail);
4119 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004120 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121 }
4122}
4123
4124/**
4125 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004126 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 *
4128 * Called from workqueue/eventd every few seconds.
4129 * Purpose:
4130 * - clear the per-cpu caches for this CPU.
4131 * - return freeable pages to the main free memory pool.
4132 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004133 * If we cannot acquire the cache chain mutex then just give up - we'll try
4134 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004136static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004138 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004139 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004140 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004141 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004143 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004145 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004147 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 check_irq_on();
4149
Christoph Lameter35386e32006-03-22 00:09:05 -08004150 /*
4151 * We only take the l3 lock if absolutely necessary and we
4152 * have established with reasonable certainty that
4153 * we can do some work if the lock was obtained.
4154 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004155 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004156
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004157 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158
Christoph Lameteraab22072006-03-22 00:09:06 -08004159 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160
Christoph Lameter35386e32006-03-22 00:09:05 -08004161 /*
4162 * These are racy checks but it does not matter
4163 * if we skip one check or scan twice.
4164 */
Christoph Lametere498be72005-09-09 13:03:32 -07004165 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004166 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
Christoph Lametere498be72005-09-09 13:03:32 -07004168 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169
Christoph Lameteraab22072006-03-22 00:09:06 -08004170 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171
Christoph Lametered11d9e2006-06-30 01:55:45 -07004172 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004173 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004174 else {
4175 int freed;
4176
4177 freed = drain_freelist(searchp, l3, (l3->free_limit +
4178 5 * searchp->num - 1) / (5 * searchp->num));
4179 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004181next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182 cond_resched();
4183 }
4184 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004185 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004186 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004187out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004188 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004189 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190}
4191
Linus Torvalds158a9622008-01-02 13:04:48 -08004192#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193
Pekka Enberg85289f92006-01-08 01:00:36 -08004194static void print_slabinfo_header(struct seq_file *m)
4195{
4196 /*
4197 * Output format version, so at least we can change it
4198 * without _too_ many complaints.
4199 */
4200#if STATS
4201 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4202#else
4203 seq_puts(m, "slabinfo - version: 2.1\n");
4204#endif
4205 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4206 "<objperslab> <pagesperslab>");
4207 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4208 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4209#if STATS
4210 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004211 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004212 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4213#endif
4214 seq_putc(m, '\n');
4215}
4216
Linus Torvalds1da177e2005-04-16 15:20:36 -07004217static void *s_start(struct seq_file *m, loff_t *pos)
4218{
4219 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004221 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004222 if (!n)
4223 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004224
4225 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226}
4227
4228static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4229{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004230 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231}
4232
4233static void s_stop(struct seq_file *m, void *p)
4234{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004235 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236}
4237
4238static int s_show(struct seq_file *m, void *p)
4239{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004240 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004241 struct slab *slabp;
4242 unsigned long active_objs;
4243 unsigned long num_objs;
4244 unsigned long active_slabs = 0;
4245 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004246 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004248 int node;
4249 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 active_objs = 0;
4252 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004253 for_each_online_node(node) {
4254 l3 = cachep->nodelists[node];
4255 if (!l3)
4256 continue;
4257
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004258 check_irq_on();
4259 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004260
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004261 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004262 if (slabp->inuse != cachep->num && !error)
4263 error = "slabs_full accounting error";
4264 active_objs += cachep->num;
4265 active_slabs++;
4266 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004267 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004268 if (slabp->inuse == cachep->num && !error)
4269 error = "slabs_partial inuse accounting error";
4270 if (!slabp->inuse && !error)
4271 error = "slabs_partial/inuse accounting error";
4272 active_objs += slabp->inuse;
4273 active_slabs++;
4274 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004275 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004276 if (slabp->inuse && !error)
4277 error = "slabs_free/inuse accounting error";
4278 num_slabs++;
4279 }
4280 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004281 if (l3->shared)
4282 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004283
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004284 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004286 num_slabs += active_slabs;
4287 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004288 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289 error = "free_objects accounting error";
4290
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004291 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292 if (error)
4293 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4294
4295 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004296 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004297 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004299 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004300 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004301 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004303 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304 unsigned long high = cachep->high_mark;
4305 unsigned long allocs = cachep->num_allocations;
4306 unsigned long grown = cachep->grown;
4307 unsigned long reaped = cachep->reaped;
4308 unsigned long errors = cachep->errors;
4309 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004311 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004312 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313
Joe Perchese92dd4f2010-03-26 19:27:58 -07004314 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4315 "%4lu %4lu %4lu %4lu %4lu",
4316 allocs, high, grown,
4317 reaped, errors, max_freeable, node_allocs,
4318 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004319 }
4320 /* cpu stats */
4321 {
4322 unsigned long allochit = atomic_read(&cachep->allochit);
4323 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4324 unsigned long freehit = atomic_read(&cachep->freehit);
4325 unsigned long freemiss = atomic_read(&cachep->freemiss);
4326
4327 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004328 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 }
4330#endif
4331 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332 return 0;
4333}
4334
4335/*
4336 * slabinfo_op - iterator that generates /proc/slabinfo
4337 *
4338 * Output layout:
4339 * cache-name
4340 * num-active-objs
4341 * total-objs
4342 * object size
4343 * num-active-slabs
4344 * total-slabs
4345 * num-pages-per-slab
4346 * + further values on SMP and with statistics enabled
4347 */
4348
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004349static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004350 .start = s_start,
4351 .next = s_next,
4352 .stop = s_stop,
4353 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354};
4355
4356#define MAX_SLABINFO_WRITE 128
4357/**
4358 * slabinfo_write - Tuning for the slab allocator
4359 * @file: unused
4360 * @buffer: user buffer
4361 * @count: data length
4362 * @ppos: unused
4363 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004364static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004365 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004366{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004367 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004369 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004370
Linus Torvalds1da177e2005-04-16 15:20:36 -07004371 if (count > MAX_SLABINFO_WRITE)
4372 return -EINVAL;
4373 if (copy_from_user(&kbuf, buffer, count))
4374 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004375 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004376
4377 tmp = strchr(kbuf, ' ');
4378 if (!tmp)
4379 return -EINVAL;
4380 *tmp = '\0';
4381 tmp++;
4382 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4383 return -EINVAL;
4384
4385 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004386 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004387 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004388 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004389 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004390 if (limit < 1 || batchcount < 1 ||
4391 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004392 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004393 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004394 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004395 batchcount, shared,
4396 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004397 }
4398 break;
4399 }
4400 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004401 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004402 if (res >= 0)
4403 res = count;
4404 return res;
4405}
Al Viro871751e2006-03-25 03:06:39 -08004406
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004407static int slabinfo_open(struct inode *inode, struct file *file)
4408{
4409 return seq_open(file, &slabinfo_op);
4410}
4411
4412static const struct file_operations proc_slabinfo_operations = {
4413 .open = slabinfo_open,
4414 .read = seq_read,
4415 .write = slabinfo_write,
4416 .llseek = seq_lseek,
4417 .release = seq_release,
4418};
4419
Al Viro871751e2006-03-25 03:06:39 -08004420#ifdef CONFIG_DEBUG_SLAB_LEAK
4421
4422static void *leaks_start(struct seq_file *m, loff_t *pos)
4423{
Al Viro871751e2006-03-25 03:06:39 -08004424 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004425 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004426}
4427
4428static inline int add_caller(unsigned long *n, unsigned long v)
4429{
4430 unsigned long *p;
4431 int l;
4432 if (!v)
4433 return 1;
4434 l = n[1];
4435 p = n + 2;
4436 while (l) {
4437 int i = l/2;
4438 unsigned long *q = p + 2 * i;
4439 if (*q == v) {
4440 q[1]++;
4441 return 1;
4442 }
4443 if (*q > v) {
4444 l = i;
4445 } else {
4446 p = q + 2;
4447 l -= i + 1;
4448 }
4449 }
4450 if (++n[1] == n[0])
4451 return 0;
4452 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4453 p[0] = v;
4454 p[1] = 1;
4455 return 1;
4456}
4457
4458static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4459{
4460 void *p;
4461 int i;
4462 if (n[0] == n[1])
4463 return;
4464 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4465 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4466 continue;
4467 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4468 return;
4469 }
4470}
4471
4472static void show_symbol(struct seq_file *m, unsigned long address)
4473{
4474#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004475 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004476 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004477
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004478 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004479 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004480 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004481 seq_printf(m, " [%s]", modname);
4482 return;
4483 }
4484#endif
4485 seq_printf(m, "%p", (void *)address);
4486}
4487
4488static int leaks_show(struct seq_file *m, void *p)
4489{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004490 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004491 struct slab *slabp;
4492 struct kmem_list3 *l3;
4493 const char *name;
4494 unsigned long *n = m->private;
4495 int node;
4496 int i;
4497
4498 if (!(cachep->flags & SLAB_STORE_USER))
4499 return 0;
4500 if (!(cachep->flags & SLAB_RED_ZONE))
4501 return 0;
4502
4503 /* OK, we can do it */
4504
4505 n[1] = 0;
4506
4507 for_each_online_node(node) {
4508 l3 = cachep->nodelists[node];
4509 if (!l3)
4510 continue;
4511
4512 check_irq_on();
4513 spin_lock_irq(&l3->list_lock);
4514
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004515 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004516 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004517 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004518 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004519 spin_unlock_irq(&l3->list_lock);
4520 }
4521 name = cachep->name;
4522 if (n[0] == n[1]) {
4523 /* Increase the buffer size */
4524 mutex_unlock(&cache_chain_mutex);
4525 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4526 if (!m->private) {
4527 /* Too bad, we are really out */
4528 m->private = n;
4529 mutex_lock(&cache_chain_mutex);
4530 return -ENOMEM;
4531 }
4532 *(unsigned long *)m->private = n[0] * 2;
4533 kfree(n);
4534 mutex_lock(&cache_chain_mutex);
4535 /* Now make sure this entry will be retried */
4536 m->count = m->size;
4537 return 0;
4538 }
4539 for (i = 0; i < n[1]; i++) {
4540 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4541 show_symbol(m, n[2*i+2]);
4542 seq_putc(m, '\n');
4543 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004544
Al Viro871751e2006-03-25 03:06:39 -08004545 return 0;
4546}
4547
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004548static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004549 .start = leaks_start,
4550 .next = s_next,
4551 .stop = s_stop,
4552 .show = leaks_show,
4553};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004554
4555static int slabstats_open(struct inode *inode, struct file *file)
4556{
4557 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4558 int ret = -ENOMEM;
4559 if (n) {
4560 ret = seq_open(file, &slabstats_op);
4561 if (!ret) {
4562 struct seq_file *m = file->private_data;
4563 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4564 m->private = n;
4565 n = NULL;
4566 }
4567 kfree(n);
4568 }
4569 return ret;
4570}
4571
4572static const struct file_operations proc_slabstats_operations = {
4573 .open = slabstats_open,
4574 .read = seq_read,
4575 .llseek = seq_lseek,
4576 .release = seq_release_private,
4577};
Al Viro871751e2006-03-25 03:06:39 -08004578#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004579
4580static int __init slab_proc_init(void)
4581{
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004582 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004583#ifdef CONFIG_DEBUG_SLAB_LEAK
4584 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4585#endif
4586 return 0;
4587}
4588module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589#endif
4590
Manfred Spraul00e145b2005-09-03 15:55:07 -07004591/**
4592 * ksize - get the actual amount of memory allocated for a given object
4593 * @objp: Pointer to the object
4594 *
4595 * kmalloc may internally round up allocations and return more memory
4596 * than requested. ksize() can be used to determine the actual amount of
4597 * memory allocated. The caller may use this additional memory, even though
4598 * a smaller amount of memory was initially specified with the kmalloc call.
4599 * The caller must guarantee that objp points to a valid object previously
4600 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4601 * must not be freed during the duration of the call.
4602 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004603size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004604{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004605 BUG_ON(!objp);
4606 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004607 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004608
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004609 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004610}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004611EXPORT_SYMBOL(ksize);