blob: e49f8f46f46d6878dba427dae10fc9b5fb153a9f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Zhaolei02af61b2009-04-10 14:26:18 +0800105#include <linux/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200117#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700118#include <linux/memory.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/*
195 * struct slab
196 *
197 * Manages the objs in a slab. Placed either at the beginning of mem allocated
198 * for a slab, or allocated from an general cache.
199 * Slabs are chained into three list: fully used, partial, fully free slabs.
200 */
201struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800202 struct list_head list;
203 unsigned long colouroff;
204 void *s_mem; /* including colour offset */
205 unsigned int inuse; /* num of objs active in slab */
206 kmem_bufctl_t free;
207 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208};
209
210/*
211 * struct slab_rcu
212 *
213 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
214 * arrange for kmem_freepages to be called via RCU. This is useful if
215 * we need to approach a kernel structure obliquely, from its address
216 * obtained without the usual locking. We can lock the structure to
217 * stabilize it and check it's still at the given address, only if we
218 * can be sure that the memory has not been meanwhile reused for some
219 * other kind of object (which our subsystem's lock might corrupt).
220 *
221 * rcu_read_lock before reading the address, then rcu_read_unlock after
222 * taking the spinlock within the structure expected at that address.
223 *
224 * We assume struct slab_rcu can overlay struct slab when destroying.
225 */
226struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800227 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800228 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800229 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230};
231
232/*
233 * struct array_cache
234 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 * Purpose:
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
239 *
240 * The limit is stored in the per-cpu structure to reduce the data cache
241 * footprint.
242 *
243 */
244struct array_cache {
245 unsigned int avail;
246 unsigned int limit;
247 unsigned int batchcount;
248 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700249 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700250 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
253 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800254 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255};
256
Andrew Mortona737b3e2006-03-22 00:08:11 -0800257/*
258 * bootstrap: The caches do not work without cpuarrays anymore, but the
259 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 */
261#define BOOT_CPUCACHE_ENTRIES 1
262struct arraycache_init {
263 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800264 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265};
266
267/*
Christoph Lametere498be72005-09-09 13:03:32 -0700268 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 */
270struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800271 struct list_head slabs_partial; /* partial list first, better asm code */
272 struct list_head slabs_full;
273 struct list_head slabs_free;
274 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800275 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800276 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800277 spinlock_t list_lock;
278 struct array_cache *shared; /* shared per node */
279 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800280 unsigned long next_reap; /* updated without locking */
281 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282};
283
Christoph Lametere498be72005-09-09 13:03:32 -0700284/*
285 * Need this for bootstrapping a per node allocator.
286 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200287#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700288struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
289#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200290#define SIZE_AC MAX_NUMNODES
291#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Christoph Lametered11d9e2006-06-30 01:55:45 -0700293static int drain_freelist(struct kmem_cache *cache,
294 struct kmem_list3 *l3, int tofree);
295static void free_block(struct kmem_cache *cachep, void **objpp, int len,
296 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300297static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000298static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700299
Christoph Lametere498be72005-09-09 13:03:32 -0700300/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800301 * This function must be completely optimized away if a constant is passed to
302 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700303 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700304static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700305{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800306 extern void __bad_size(void);
307
Christoph Lametere498be72005-09-09 13:03:32 -0700308 if (__builtin_constant_p(size)) {
309 int i = 0;
310
311#define CACHE(x) \
312 if (size <=x) \
313 return i; \
314 else \
315 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800316#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700317#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800318 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700319 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800320 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700321 return 0;
322}
323
Ingo Molnare0a42722006-06-23 02:03:46 -0700324static int slab_early_init = 1;
325
Christoph Lametere498be72005-09-09 13:03:32 -0700326#define INDEX_AC index_of(sizeof(struct arraycache_init))
327#define INDEX_L3 index_of(sizeof(struct kmem_list3))
328
Pekka Enberg5295a742006-02-01 03:05:48 -0800329static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700330{
331 INIT_LIST_HEAD(&parent->slabs_full);
332 INIT_LIST_HEAD(&parent->slabs_partial);
333 INIT_LIST_HEAD(&parent->slabs_free);
334 parent->shared = NULL;
335 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800336 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700337 spin_lock_init(&parent->list_lock);
338 parent->free_objects = 0;
339 parent->free_touched = 0;
340}
341
Andrew Mortona737b3e2006-03-22 00:08:11 -0800342#define MAKE_LIST(cachep, listp, slab, nodeid) \
343 do { \
344 INIT_LIST_HEAD(listp); \
345 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700346 } while (0)
347
Andrew Mortona737b3e2006-03-22 00:08:11 -0800348#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
349 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700350 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
351 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
352 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
353 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355#define CFLGS_OFF_SLAB (0x80000000UL)
356#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
357
358#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800359/*
360 * Optimization question: fewer reaps means less probability for unnessary
361 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100363 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 * which could lock up otherwise freeable slabs.
365 */
366#define REAPTIMEOUT_CPUC (2*HZ)
367#define REAPTIMEOUT_LIST3 (4*HZ)
368
369#if STATS
370#define STATS_INC_ACTIVE(x) ((x)->num_active++)
371#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
372#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
373#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700374#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800375#define STATS_SET_HIGH(x) \
376 do { \
377 if ((x)->num_active > (x)->high_mark) \
378 (x)->high_mark = (x)->num_active; \
379 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380#define STATS_INC_ERR(x) ((x)->errors++)
381#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700382#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700383#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800384#define STATS_SET_FREEABLE(x, i) \
385 do { \
386 if ((x)->max_freeable < i) \
387 (x)->max_freeable = i; \
388 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
390#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
391#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
392#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
393#else
394#define STATS_INC_ACTIVE(x) do { } while (0)
395#define STATS_DEC_ACTIVE(x) do { } while (0)
396#define STATS_INC_ALLOCED(x) do { } while (0)
397#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700398#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399#define STATS_SET_HIGH(x) do { } while (0)
400#define STATS_INC_ERR(x) do { } while (0)
401#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700402#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700403#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405#define STATS_INC_ALLOCHIT(x) do { } while (0)
406#define STATS_INC_ALLOCMISS(x) do { } while (0)
407#define STATS_INC_FREEHIT(x) do { } while (0)
408#define STATS_INC_FREEMISS(x) do { } while (0)
409#endif
410
411#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Andrew Mortona737b3e2006-03-22 00:08:11 -0800413/*
414 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800416 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 * the end of an object is aligned with the end of the real
418 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800419 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800421 * cachep->obj_offset: The real object.
422 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800423 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
424 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800426static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800428 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429}
430
Pekka Enberg343e0d72006-02-01 03:05:50 -0800431static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800433 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
David Woodhouseb46b8f12007-05-08 00:22:59 -0700436static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
438 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700439 return (unsigned long long*) (objp + obj_offset(cachep) -
440 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441}
442
David Woodhouseb46b8f12007-05-08 00:22:59 -0700443static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444{
445 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
446 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700447 return (unsigned long long *)(objp + cachep->buffer_size -
448 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400449 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700450 return (unsigned long long *) (objp + cachep->buffer_size -
451 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
Pekka Enberg343e0d72006-02-01 03:05:50 -0800454static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455{
456 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800457 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458}
459
460#else
461
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800462#define obj_offset(x) 0
463#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700464#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
467
468#endif
469
Li Zefan0f24f122009-12-11 15:45:30 +0800470#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300471size_t slab_buffer_size(struct kmem_cache *cachep)
472{
473 return cachep->buffer_size;
474}
475EXPORT_SYMBOL(slab_buffer_size);
476#endif
477
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 * Do not go above this order unless 0 objects fit into the slab.
480 */
481#define BREAK_GFP_ORDER_HI 1
482#define BREAK_GFP_ORDER_LO 0
483static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
484
Andrew Mortona737b3e2006-03-22 00:08:11 -0800485/*
486 * Functions for storing/retrieving the cachep and or slab from the page
487 * allocator. These are used to find the slab an obj belongs to. With kfree(),
488 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800490static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
491{
492 page->lru.next = (struct list_head *)cache;
493}
494
495static inline struct kmem_cache *page_get_cache(struct page *page)
496{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700497 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700498 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800499 return (struct kmem_cache *)page->lru.next;
500}
501
502static inline void page_set_slab(struct page *page, struct slab *slab)
503{
504 page->lru.prev = (struct list_head *)slab;
505}
506
507static inline struct slab *page_get_slab(struct page *page)
508{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700509 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800510 return (struct slab *)page->lru.prev;
511}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800513static inline struct kmem_cache *virt_to_cache(const void *obj)
514{
Christoph Lameterb49af682007-05-06 14:49:41 -0700515 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800516 return page_get_cache(page);
517}
518
519static inline struct slab *virt_to_slab(const void *obj)
520{
Christoph Lameterb49af682007-05-06 14:49:41 -0700521 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800522 return page_get_slab(page);
523}
524
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800525static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
526 unsigned int idx)
527{
528 return slab->s_mem + cache->buffer_size * idx;
529}
530
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800531/*
532 * We want to avoid an expensive divide : (offset / cache->buffer_size)
533 * Using the fact that buffer_size is a constant for a particular cache,
534 * we can replace (offset / cache->buffer_size) by
535 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 */
537static inline unsigned int obj_to_index(const struct kmem_cache *cache,
538 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800539{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800540 u32 offset = (obj - slab->s_mem);
541 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800542}
543
Andrew Mortona737b3e2006-03-22 00:08:11 -0800544/*
545 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547struct cache_sizes malloc_sizes[] = {
548#define CACHE(x) { .cs_size = (x) },
549#include <linux/kmalloc_sizes.h>
550 CACHE(ULONG_MAX)
551#undef CACHE
552};
553EXPORT_SYMBOL(malloc_sizes);
554
555/* Must match cache_sizes above. Out of line to keep cache footprint low. */
556struct cache_names {
557 char *name;
558 char *name_dma;
559};
560
561static struct cache_names __initdata cache_names[] = {
562#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
563#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800564 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565#undef CACHE
566};
567
568static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800569 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800571 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
573/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800574static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800575 .batchcount = 1,
576 .limit = BOOT_CPUCACHE_ENTRIES,
577 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800578 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800579 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580};
581
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700582#define BAD_ALIEN_MAGIC 0x01020304ul
583
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 * chicken and egg problem: delay the per-cpu array allocation
586 * until the general caches are up.
587 */
588static enum {
589 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700590 PARTIAL_AC,
591 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300592 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 FULL
594} g_cpucache_up;
595
Mike Kravetz39d24e62006-05-15 09:44:13 -0700596/*
597 * used by boot code to determine if it can use slab based allocator
598 */
599int slab_is_available(void)
600{
Pekka Enberg8429db52009-06-12 15:58:59 +0300601 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700602}
603
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200604#ifdef CONFIG_LOCKDEP
605
606/*
607 * Slab sometimes uses the kmalloc slabs to store the slab headers
608 * for other slabs "off slab".
609 * The locking for this is tricky in that it nests within the locks
610 * of all other slabs in a few places; to deal with this special
611 * locking we put on-slab caches into a separate lock-class.
612 *
613 * We set lock class for alien array caches which are up during init.
614 * The lock annotation will be lost if all cpus of a node goes down and
615 * then comes back up during hotplug
616 */
617static struct lock_class_key on_slab_l3_key;
618static struct lock_class_key on_slab_alc_key;
619
620static void init_node_lock_keys(int q)
621{
622 struct cache_sizes *s = malloc_sizes;
623
624 if (g_cpucache_up != FULL)
625 return;
626
627 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
628 struct array_cache **alc;
629 struct kmem_list3 *l3;
630 int r;
631
632 l3 = s->cs_cachep->nodelists[q];
633 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200634 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200635 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
636 alc = l3->alien;
637 /*
638 * FIXME: This check for BAD_ALIEN_MAGIC
639 * should go away when common slab code is taught to
640 * work even without alien caches.
641 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
642 * for alloc_alien_cache,
643 */
644 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
Pekka Enberg00afa752009-12-27 14:33:14 +0200645 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200646 for_each_node(r) {
647 if (alc[r])
648 lockdep_set_class(&alc[r]->lock,
649 &on_slab_alc_key);
650 }
651 }
652}
653
654static inline void init_lock_keys(void)
655{
656 int node;
657
658 for_each_node(node)
659 init_node_lock_keys(node);
660}
661#else
662static void init_node_lock_keys(int q)
663{
664}
665
666static inline void init_lock_keys(void)
667{
668}
669#endif
670
671/*
672 * Guard access to the cache-chain.
673 */
674static DEFINE_MUTEX(cache_chain_mutex);
675static struct list_head cache_chain;
676
Tejun Heo1871e522009-10-29 22:34:13 +0900677static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Pekka Enberg343e0d72006-02-01 03:05:50 -0800679static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680{
681 return cachep->array[smp_processor_id()];
682}
683
Andrew Mortona737b3e2006-03-22 00:08:11 -0800684static inline struct kmem_cache *__find_general_cachep(size_t size,
685 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686{
687 struct cache_sizes *csizep = malloc_sizes;
688
689#if DEBUG
690 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800691 * kmem_cache_create(), or __kmalloc(), before
692 * the generic caches are initialized.
693 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700694 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700696 if (!size)
697 return ZERO_SIZE_PTR;
698
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 while (size > csizep->cs_size)
700 csizep++;
701
702 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700703 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 * has cs_{dma,}cachep==NULL. Thus no special case
705 * for large kmalloc calls required.
706 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800707#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 if (unlikely(gfpflags & GFP_DMA))
709 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800710#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 return csizep->cs_cachep;
712}
713
Adrian Bunkb2213852006-09-25 23:31:02 -0700714static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700715{
716 return __find_general_cachep(size, gfpflags);
717}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700718
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800719static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800721 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
722}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Andrew Mortona737b3e2006-03-22 00:08:11 -0800724/*
725 * Calculate the number of objects and left-over bytes for a given buffer size.
726 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800727static void cache_estimate(unsigned long gfporder, size_t buffer_size,
728 size_t align, int flags, size_t *left_over,
729 unsigned int *num)
730{
731 int nr_objs;
732 size_t mgmt_size;
733 size_t slab_size = PAGE_SIZE << gfporder;
734
735 /*
736 * The slab management structure can be either off the slab or
737 * on it. For the latter case, the memory allocated for a
738 * slab is used for:
739 *
740 * - The struct slab
741 * - One kmem_bufctl_t for each object
742 * - Padding to respect alignment of @align
743 * - @buffer_size bytes for each object
744 *
745 * If the slab management structure is off the slab, then the
746 * alignment will already be calculated into the size. Because
747 * the slabs are all pages aligned, the objects will be at the
748 * correct alignment when allocated.
749 */
750 if (flags & CFLGS_OFF_SLAB) {
751 mgmt_size = 0;
752 nr_objs = slab_size / buffer_size;
753
754 if (nr_objs > SLAB_LIMIT)
755 nr_objs = SLAB_LIMIT;
756 } else {
757 /*
758 * Ignore padding for the initial guess. The padding
759 * is at most @align-1 bytes, and @buffer_size is at
760 * least @align. In the worst case, this result will
761 * be one greater than the number of objects that fit
762 * into the memory allocation when taking the padding
763 * into account.
764 */
765 nr_objs = (slab_size - sizeof(struct slab)) /
766 (buffer_size + sizeof(kmem_bufctl_t));
767
768 /*
769 * This calculated number will be either the right
770 * amount, or one greater than what we want.
771 */
772 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
773 > slab_size)
774 nr_objs--;
775
776 if (nr_objs > SLAB_LIMIT)
777 nr_objs = SLAB_LIMIT;
778
779 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800781 *num = nr_objs;
782 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783}
784
Harvey Harrisond40cee22008-04-30 00:55:07 -0700785#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786
Andrew Mortona737b3e2006-03-22 00:08:11 -0800787static void __slab_error(const char *function, struct kmem_cache *cachep,
788 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789{
790 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800791 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 dump_stack();
793}
794
Paul Menage3395ee02006-12-06 20:32:16 -0800795/*
796 * By default on NUMA we use alien caches to stage the freeing of
797 * objects allocated from other nodes. This causes massive memory
798 * inefficiencies when using fake NUMA setup to split memory into a
799 * large number of small nodes, so it can be disabled on the command
800 * line
801 */
802
803static int use_alien_caches __read_mostly = 1;
804static int __init noaliencache_setup(char *s)
805{
806 use_alien_caches = 0;
807 return 1;
808}
809__setup("noaliencache", noaliencache_setup);
810
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800811#ifdef CONFIG_NUMA
812/*
813 * Special reaping functions for NUMA systems called from cache_reap().
814 * These take care of doing round robin flushing of alien caches (containing
815 * objects freed on different nodes from which they were allocated) and the
816 * flushing of remote pcps by calling drain_node_pages.
817 */
Tejun Heo1871e522009-10-29 22:34:13 +0900818static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800819
820static void init_reap_node(int cpu)
821{
822 int node;
823
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700824 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800825 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800826 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800827
Tejun Heo1871e522009-10-29 22:34:13 +0900828 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800829}
830
831static void next_reap_node(void)
832{
Tejun Heo1871e522009-10-29 22:34:13 +0900833 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800834
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800835 node = next_node(node, node_online_map);
836 if (unlikely(node >= MAX_NUMNODES))
837 node = first_node(node_online_map);
Tejun Heo1871e522009-10-29 22:34:13 +0900838 __get_cpu_var(slab_reap_node) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800839}
840
841#else
842#define init_reap_node(cpu) do { } while (0)
843#define next_reap_node(void) do { } while (0)
844#endif
845
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846/*
847 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
848 * via the workqueue/eventd.
849 * Add the CPU number into the expiration time to minimize the possibility of
850 * the CPUs getting into lockstep and contending for the global cache chain
851 * lock.
852 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700853static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854{
Tejun Heo1871e522009-10-29 22:34:13 +0900855 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
857 /*
858 * When this gets called from do_initcalls via cpucache_init(),
859 * init_workqueues() has already run, so keventd will be setup
860 * at that time.
861 */
David Howells52bad642006-11-22 14:54:01 +0000862 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800863 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000864 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800865 schedule_delayed_work_on(cpu, reap_work,
866 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 }
868}
869
Christoph Lametere498be72005-09-09 13:03:32 -0700870static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300871 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800873 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 struct array_cache *nc = NULL;
875
Pekka Enberg83b519e2009-06-10 19:40:04 +0300876 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100877 /*
878 * The array_cache structures contain pointers to free object.
879 * However, when such objects are allocated or transfered to another
880 * cache the pointers are not cleared and they could be counted as
881 * valid references during a kmemleak scan. Therefore, kmemleak must
882 * not scan such objects.
883 */
884 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 if (nc) {
886 nc->avail = 0;
887 nc->limit = entries;
888 nc->batchcount = batchcount;
889 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700890 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 }
892 return nc;
893}
894
Christoph Lameter3ded1752006-03-25 03:06:44 -0800895/*
896 * Transfer objects in one arraycache to another.
897 * Locking must be handled by the caller.
898 *
899 * Return the number of entries transferred.
900 */
901static int transfer_objects(struct array_cache *to,
902 struct array_cache *from, unsigned int max)
903{
904 /* Figure out how many entries to transfer */
905 int nr = min(min(from->avail, max), to->limit - to->avail);
906
907 if (!nr)
908 return 0;
909
910 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
911 sizeof(void *) *nr);
912
913 from->avail -= nr;
914 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800915 return nr;
916}
917
Christoph Lameter765c4502006-09-27 01:50:08 -0700918#ifndef CONFIG_NUMA
919
920#define drain_alien_cache(cachep, alien) do { } while (0)
921#define reap_alien(cachep, l3) do { } while (0)
922
Pekka Enberg83b519e2009-06-10 19:40:04 +0300923static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700924{
925 return (struct array_cache **)BAD_ALIEN_MAGIC;
926}
927
928static inline void free_alien_cache(struct array_cache **ac_ptr)
929{
930}
931
932static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
933{
934 return 0;
935}
936
937static inline void *alternate_node_alloc(struct kmem_cache *cachep,
938 gfp_t flags)
939{
940 return NULL;
941}
942
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800943static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700944 gfp_t flags, int nodeid)
945{
946 return NULL;
947}
948
949#else /* CONFIG_NUMA */
950
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800951static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800952static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800953
Pekka Enberg83b519e2009-06-10 19:40:04 +0300954static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700955{
956 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800957 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700958 int i;
959
960 if (limit > 1)
961 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800962 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700963 if (ac_ptr) {
964 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800965 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700966 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300967 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700968 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800969 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700970 kfree(ac_ptr[i]);
971 kfree(ac_ptr);
972 return NULL;
973 }
974 }
975 }
976 return ac_ptr;
977}
978
Pekka Enberg5295a742006-02-01 03:05:48 -0800979static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700980{
981 int i;
982
983 if (!ac_ptr)
984 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700985 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800986 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700987 kfree(ac_ptr);
988}
989
Pekka Enberg343e0d72006-02-01 03:05:50 -0800990static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800991 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700992{
993 struct kmem_list3 *rl3 = cachep->nodelists[node];
994
995 if (ac->avail) {
996 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800997 /*
998 * Stuff objects into the remote nodes shared array first.
999 * That way we could avoid the overhead of putting the objects
1000 * into the free lists and getting them back later.
1001 */
shin, jacob693f7d32006-04-28 10:54:37 -05001002 if (rl3->shared)
1003 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001004
Christoph Lameterff694162005-09-22 21:44:02 -07001005 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001006 ac->avail = 0;
1007 spin_unlock(&rl3->list_lock);
1008 }
1009}
1010
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001011/*
1012 * Called from cache_reap() to regularly drain alien caches round robin.
1013 */
1014static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1015{
Tejun Heo1871e522009-10-29 22:34:13 +09001016 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001017
1018 if (l3->alien) {
1019 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001020
1021 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001022 __drain_alien_cache(cachep, ac, node);
1023 spin_unlock_irq(&ac->lock);
1024 }
1025 }
1026}
1027
Andrew Mortona737b3e2006-03-22 00:08:11 -08001028static void drain_alien_cache(struct kmem_cache *cachep,
1029 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001030{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001031 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001032 struct array_cache *ac;
1033 unsigned long flags;
1034
1035 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001036 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001037 if (ac) {
1038 spin_lock_irqsave(&ac->lock, flags);
1039 __drain_alien_cache(cachep, ac, i);
1040 spin_unlock_irqrestore(&ac->lock, flags);
1041 }
1042 }
1043}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001044
Ingo Molnar873623d2006-07-13 14:44:38 +02001045static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001046{
1047 struct slab *slabp = virt_to_slab(objp);
1048 int nodeid = slabp->nodeid;
1049 struct kmem_list3 *l3;
1050 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001051 int node;
1052
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001053 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001054
1055 /*
1056 * Make sure we are not freeing a object from another node to the array
1057 * cache on this cpu.
1058 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001059 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001060 return 0;
1061
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001062 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001063 STATS_INC_NODEFREES(cachep);
1064 if (l3->alien && l3->alien[nodeid]) {
1065 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001066 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001067 if (unlikely(alien->avail == alien->limit)) {
1068 STATS_INC_ACOVERFLOW(cachep);
1069 __drain_alien_cache(cachep, alien, nodeid);
1070 }
1071 alien->entry[alien->avail++] = objp;
1072 spin_unlock(&alien->lock);
1073 } else {
1074 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1075 free_block(cachep, &objp, 1, nodeid);
1076 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1077 }
1078 return 1;
1079}
Christoph Lametere498be72005-09-09 13:03:32 -07001080#endif
1081
David Rientjes8f9f8d92010-03-27 19:40:47 -07001082/*
1083 * Allocates and initializes nodelists for a node on each slab cache, used for
1084 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1085 * will be allocated off-node since memory is not yet online for the new node.
1086 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1087 * already in use.
1088 *
1089 * Must hold cache_chain_mutex.
1090 */
1091static int init_cache_nodelists_node(int node)
1092{
1093 struct kmem_cache *cachep;
1094 struct kmem_list3 *l3;
1095 const int memsize = sizeof(struct kmem_list3);
1096
1097 list_for_each_entry(cachep, &cache_chain, next) {
1098 /*
1099 * Set up the size64 kmemlist for cpu before we can
1100 * begin anything. Make sure some other cpu on this
1101 * node has not already allocated this
1102 */
1103 if (!cachep->nodelists[node]) {
1104 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1105 if (!l3)
1106 return -ENOMEM;
1107 kmem_list3_init(l3);
1108 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1109 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1110
1111 /*
1112 * The l3s don't come and go as CPUs come and
1113 * go. cache_chain_mutex is sufficient
1114 * protection here.
1115 */
1116 cachep->nodelists[node] = l3;
1117 }
1118
1119 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1120 cachep->nodelists[node]->free_limit =
1121 (1 + nr_cpus_node(node)) *
1122 cachep->batchcount + cachep->num;
1123 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1124 }
1125 return 0;
1126}
1127
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001128static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001130 struct kmem_cache *cachep;
1131 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001132 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301133 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134
1135 list_for_each_entry(cachep, &cache_chain, next) {
1136 struct array_cache *nc;
1137 struct array_cache *shared;
1138 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001139
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001140 /* cpu is dead; no one can alloc from it. */
1141 nc = cachep->array[cpu];
1142 cachep->array[cpu] = NULL;
1143 l3 = cachep->nodelists[node];
1144
1145 if (!l3)
1146 goto free_array_cache;
1147
1148 spin_lock_irq(&l3->list_lock);
1149
1150 /* Free limit for this kmem_list3 */
1151 l3->free_limit -= cachep->batchcount;
1152 if (nc)
1153 free_block(cachep, nc->entry, nc->avail, node);
1154
Rusty Russell58463c12009-12-17 11:43:12 -06001155 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001156 spin_unlock_irq(&l3->list_lock);
1157 goto free_array_cache;
1158 }
1159
1160 shared = l3->shared;
1161 if (shared) {
1162 free_block(cachep, shared->entry,
1163 shared->avail, node);
1164 l3->shared = NULL;
1165 }
1166
1167 alien = l3->alien;
1168 l3->alien = NULL;
1169
1170 spin_unlock_irq(&l3->list_lock);
1171
1172 kfree(shared);
1173 if (alien) {
1174 drain_alien_cache(cachep, alien);
1175 free_alien_cache(alien);
1176 }
1177free_array_cache:
1178 kfree(nc);
1179 }
1180 /*
1181 * In the previous loop, all the objects were freed to
1182 * the respective cache's slabs, now we can go ahead and
1183 * shrink each nodelist to its limit.
1184 */
1185 list_for_each_entry(cachep, &cache_chain, next) {
1186 l3 = cachep->nodelists[node];
1187 if (!l3)
1188 continue;
1189 drain_freelist(cachep, l3, l3->free_objects);
1190 }
1191}
1192
1193static int __cpuinit cpuup_prepare(long cpu)
1194{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001195 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001196 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001197 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001198 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001200 /*
1201 * We need to do this right in the beginning since
1202 * alloc_arraycache's are going to use this list.
1203 * kmalloc_node allows us to add the slab to the right
1204 * kmem_list3 and not this cpu's kmem_list3
1205 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001206 err = init_cache_nodelists_node(node);
1207 if (err < 0)
1208 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001209
1210 /*
1211 * Now we can go ahead with allocating the shared arrays and
1212 * array caches
1213 */
1214 list_for_each_entry(cachep, &cache_chain, next) {
1215 struct array_cache *nc;
1216 struct array_cache *shared = NULL;
1217 struct array_cache **alien = NULL;
1218
1219 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001220 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001221 if (!nc)
1222 goto bad;
1223 if (cachep->shared) {
1224 shared = alloc_arraycache(node,
1225 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001226 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001227 if (!shared) {
1228 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001229 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001230 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001231 }
1232 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001233 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001234 if (!alien) {
1235 kfree(shared);
1236 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001237 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001238 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001239 }
1240 cachep->array[cpu] = nc;
1241 l3 = cachep->nodelists[node];
1242 BUG_ON(!l3);
1243
1244 spin_lock_irq(&l3->list_lock);
1245 if (!l3->shared) {
1246 /*
1247 * We are serialised from CPU_DEAD or
1248 * CPU_UP_CANCELLED by the cpucontrol lock
1249 */
1250 l3->shared = shared;
1251 shared = NULL;
1252 }
1253#ifdef CONFIG_NUMA
1254 if (!l3->alien) {
1255 l3->alien = alien;
1256 alien = NULL;
1257 }
1258#endif
1259 spin_unlock_irq(&l3->list_lock);
1260 kfree(shared);
1261 free_alien_cache(alien);
1262 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001263 init_node_lock_keys(node);
1264
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001265 return 0;
1266bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001267 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001268 return -ENOMEM;
1269}
1270
1271static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1272 unsigned long action, void *hcpu)
1273{
1274 long cpu = (long)hcpu;
1275 int err = 0;
1276
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001278 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001279 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001280 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001281 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001282 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 break;
1284 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001285 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 start_cpu_timer(cpu);
1287 break;
1288#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001289 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001290 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001291 /*
1292 * Shutdown cache reaper. Note that the cache_chain_mutex is
1293 * held so that if cache_reap() is invoked it cannot do
1294 * anything expensive but will only modify reap_work
1295 * and reschedule the timer.
1296 */
Tejun Heo1871e522009-10-29 22:34:13 +09001297 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001298 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001299 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001300 break;
1301 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001302 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001303 start_cpu_timer(cpu);
1304 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001306 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001307 /*
1308 * Even if all the cpus of a node are down, we don't free the
1309 * kmem_list3 of any cache. This to avoid a race between
1310 * cpu_down, and a kmalloc allocation from another cpu for
1311 * memory from the node of the cpu going down. The list3
1312 * structure is usually allocated from kmem_cache_create() and
1313 * gets destroyed at kmem_cache_destroy().
1314 */
Simon Arlott183ff222007-10-20 01:27:18 +02001315 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001316#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001318 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001319 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001320 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001321 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001324 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325}
1326
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001327static struct notifier_block __cpuinitdata cpucache_notifier = {
1328 &cpuup_callback, NULL, 0
1329};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
David Rientjes8f9f8d92010-03-27 19:40:47 -07001331#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1332/*
1333 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1334 * Returns -EBUSY if all objects cannot be drained so that the node is not
1335 * removed.
1336 *
1337 * Must hold cache_chain_mutex.
1338 */
1339static int __meminit drain_cache_nodelists_node(int node)
1340{
1341 struct kmem_cache *cachep;
1342 int ret = 0;
1343
1344 list_for_each_entry(cachep, &cache_chain, next) {
1345 struct kmem_list3 *l3;
1346
1347 l3 = cachep->nodelists[node];
1348 if (!l3)
1349 continue;
1350
1351 drain_freelist(cachep, l3, l3->free_objects);
1352
1353 if (!list_empty(&l3->slabs_full) ||
1354 !list_empty(&l3->slabs_partial)) {
1355 ret = -EBUSY;
1356 break;
1357 }
1358 }
1359 return ret;
1360}
1361
1362static int __meminit slab_memory_callback(struct notifier_block *self,
1363 unsigned long action, void *arg)
1364{
1365 struct memory_notify *mnb = arg;
1366 int ret = 0;
1367 int nid;
1368
1369 nid = mnb->status_change_nid;
1370 if (nid < 0)
1371 goto out;
1372
1373 switch (action) {
1374 case MEM_GOING_ONLINE:
1375 mutex_lock(&cache_chain_mutex);
1376 ret = init_cache_nodelists_node(nid);
1377 mutex_unlock(&cache_chain_mutex);
1378 break;
1379 case MEM_GOING_OFFLINE:
1380 mutex_lock(&cache_chain_mutex);
1381 ret = drain_cache_nodelists_node(nid);
1382 mutex_unlock(&cache_chain_mutex);
1383 break;
1384 case MEM_ONLINE:
1385 case MEM_OFFLINE:
1386 case MEM_CANCEL_ONLINE:
1387 case MEM_CANCEL_OFFLINE:
1388 break;
1389 }
1390out:
1391 return ret ? notifier_from_errno(ret) : NOTIFY_OK;
1392}
1393#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1394
Christoph Lametere498be72005-09-09 13:03:32 -07001395/*
1396 * swap the static kmem_list3 with kmalloced memory
1397 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001398static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1399 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001400{
1401 struct kmem_list3 *ptr;
1402
Pekka Enberg83b519e2009-06-10 19:40:04 +03001403 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001404 BUG_ON(!ptr);
1405
Christoph Lametere498be72005-09-09 13:03:32 -07001406 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001407 /*
1408 * Do not assume that spinlocks can be initialized via memcpy:
1409 */
1410 spin_lock_init(&ptr->list_lock);
1411
Christoph Lametere498be72005-09-09 13:03:32 -07001412 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1413 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001414}
1415
Andrew Mortona737b3e2006-03-22 00:08:11 -08001416/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001417 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1418 * size of kmem_list3.
1419 */
1420static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1421{
1422 int node;
1423
1424 for_each_online_node(node) {
1425 cachep->nodelists[node] = &initkmem_list3[index + node];
1426 cachep->nodelists[node]->next_reap = jiffies +
1427 REAPTIMEOUT_LIST3 +
1428 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1429 }
1430}
1431
1432/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001433 * Initialisation. Called after the page allocator have been initialised and
1434 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 */
1436void __init kmem_cache_init(void)
1437{
1438 size_t left_over;
1439 struct cache_sizes *sizes;
1440 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001441 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001442 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001443 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001444
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001445 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001446 use_alien_caches = 0;
1447
Christoph Lametere498be72005-09-09 13:03:32 -07001448 for (i = 0; i < NUM_INIT_LISTS; i++) {
1449 kmem_list3_init(&initkmem_list3[i]);
1450 if (i < MAX_NUMNODES)
1451 cache_cache.nodelists[i] = NULL;
1452 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001453 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454
1455 /*
1456 * Fragmentation resistance on low memory - only use bigger
1457 * page orders on machines with more than 32MB of memory.
1458 */
Jan Beulich44813742009-09-21 17:03:05 -07001459 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1461
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 /* Bootstrap is tricky, because several objects are allocated
1463 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001464 * 1) initialize the cache_cache cache: it contains the struct
1465 * kmem_cache structures of all caches, except cache_cache itself:
1466 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001467 * Initially an __init data area is used for the head array and the
1468 * kmem_list3 structures, it's replaced with a kmalloc allocated
1469 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001471 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001472 * An __init data area is used for the head array.
1473 * 3) Create the remaining kmalloc caches, with minimally sized
1474 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 * 4) Replace the __init data head arrays for cache_cache and the first
1476 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001477 * 5) Replace the __init data for kmem_list3 for cache_cache and
1478 * the other cache's with kmalloc allocated memory.
1479 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 */
1481
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001482 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001483
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 INIT_LIST_HEAD(&cache_chain);
1486 list_add(&cache_cache.next, &cache_chain);
1487 cache_cache.colour_off = cache_line_size();
1488 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001489 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490
Eric Dumazet8da34302007-05-06 14:49:29 -07001491 /*
1492 * struct kmem_cache size depends on nr_node_ids, which
1493 * can be less than MAX_NUMNODES.
1494 */
1495 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1496 nr_node_ids * sizeof(struct kmem_list3 *);
1497#if DEBUG
1498 cache_cache.obj_size = cache_cache.buffer_size;
1499#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001500 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1501 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001502 cache_cache.reciprocal_buffer_size =
1503 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Jack Steiner07ed76b2006-03-07 21:55:46 -08001505 for (order = 0; order < MAX_ORDER; order++) {
1506 cache_estimate(order, cache_cache.buffer_size,
1507 cache_line_size(), 0, &left_over, &cache_cache.num);
1508 if (cache_cache.num)
1509 break;
1510 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001511 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001512 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001513 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001514 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1515 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516
1517 /* 2+3) create the kmalloc caches */
1518 sizes = malloc_sizes;
1519 names = cache_names;
1520
Andrew Mortona737b3e2006-03-22 00:08:11 -08001521 /*
1522 * Initialize the caches that provide memory for the array cache and the
1523 * kmem_list3 structures first. Without this, further allocations will
1524 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001525 */
1526
1527 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001528 sizes[INDEX_AC].cs_size,
1529 ARCH_KMALLOC_MINALIGN,
1530 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001531 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001532
Andrew Mortona737b3e2006-03-22 00:08:11 -08001533 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001534 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001535 kmem_cache_create(names[INDEX_L3].name,
1536 sizes[INDEX_L3].cs_size,
1537 ARCH_KMALLOC_MINALIGN,
1538 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001539 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001540 }
Christoph Lametere498be72005-09-09 13:03:32 -07001541
Ingo Molnare0a42722006-06-23 02:03:46 -07001542 slab_early_init = 0;
1543
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001545 /*
1546 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 * This should be particularly beneficial on SMP boxes, as it
1548 * eliminates "false sharing".
1549 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001550 * allow tighter packing of the smaller caches.
1551 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001552 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001553 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001554 sizes->cs_size,
1555 ARCH_KMALLOC_MINALIGN,
1556 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001557 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001558 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001559#ifdef CONFIG_ZONE_DMA
1560 sizes->cs_dmacachep = kmem_cache_create(
1561 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001562 sizes->cs_size,
1563 ARCH_KMALLOC_MINALIGN,
1564 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1565 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001566 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001567#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568 sizes++;
1569 names++;
1570 }
1571 /* 4) Replace the bootstrap head arrays */
1572 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001573 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001574
Pekka Enberg83b519e2009-06-10 19:40:04 +03001575 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001576
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001577 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1578 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001579 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001580 /*
1581 * Do not assume that spinlocks can be initialized via memcpy:
1582 */
1583 spin_lock_init(&ptr->lock);
1584
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001586
Pekka Enberg83b519e2009-06-10 19:40:04 +03001587 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001588
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001589 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001590 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001591 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001592 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001593 /*
1594 * Do not assume that spinlocks can be initialized via memcpy:
1595 */
1596 spin_lock_init(&ptr->lock);
1597
Christoph Lametere498be72005-09-09 13:03:32 -07001598 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001599 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 }
Christoph Lametere498be72005-09-09 13:03:32 -07001601 /* 5) Replace the bootstrap kmem_list3's */
1602 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001603 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604
Mel Gorman9c09a952008-01-24 05:49:54 -08001605 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001606 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001607
Christoph Lametere498be72005-09-09 13:03:32 -07001608 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001609 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001610
1611 if (INDEX_AC != INDEX_L3) {
1612 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001613 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001614 }
1615 }
1616 }
1617
Pekka Enberg8429db52009-06-12 15:58:59 +03001618 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001619}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001620
Pekka Enberg8429db52009-06-12 15:58:59 +03001621void __init kmem_cache_init_late(void)
1622{
1623 struct kmem_cache *cachep;
1624
Pekka Enberg8429db52009-06-12 15:58:59 +03001625 /* 6) resize the head arrays to their final sizes */
1626 mutex_lock(&cache_chain_mutex);
1627 list_for_each_entry(cachep, &cache_chain, next)
1628 if (enable_cpucache(cachep, GFP_NOWAIT))
1629 BUG();
1630 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001631
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 /* Done! */
1633 g_cpucache_up = FULL;
1634
Pekka Enbergec5a36f2009-06-29 09:57:10 +03001635 /* Annotate slab for lockdep -- annotate the malloc caches */
1636 init_lock_keys();
1637
Andrew Mortona737b3e2006-03-22 00:08:11 -08001638 /*
1639 * Register a cpu startup notifier callback that initializes
1640 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 */
1642 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643
David Rientjes8f9f8d92010-03-27 19:40:47 -07001644#ifdef CONFIG_NUMA
1645 /*
1646 * Register a memory hotplug callback that initializes and frees
1647 * nodelists.
1648 */
1649 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1650#endif
1651
Andrew Mortona737b3e2006-03-22 00:08:11 -08001652 /*
1653 * The reap timers are started later, with a module init call: That part
1654 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 */
1656}
1657
1658static int __init cpucache_init(void)
1659{
1660 int cpu;
1661
Andrew Mortona737b3e2006-03-22 00:08:11 -08001662 /*
1663 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 */
Christoph Lametere498be72005-09-09 13:03:32 -07001665 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001666 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 return 0;
1668}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669__initcall(cpucache_init);
1670
1671/*
1672 * Interface to system's page allocator. No need to hold the cache-lock.
1673 *
1674 * If we requested dmaable memory, we will get it. Even if we
1675 * did not request dmaable memory, we might get it, but that
1676 * would be relatively rare and ignorable.
1677 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001678static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
1680 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001681 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 int i;
1683
Luke Yangd6fef9d2006-04-10 22:52:56 -07001684#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001685 /*
1686 * Nommu uses slab's for process anonymous memory allocations, and thus
1687 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001688 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001689 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001690#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001691
Christoph Lameter3c517a62006-12-06 20:33:29 -08001692 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001693 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1694 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001695
Linus Torvalds517d0862009-06-16 19:50:13 -07001696 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 if (!page)
1698 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001700 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001702 add_zone_page_state(page_zone(page),
1703 NR_SLAB_RECLAIMABLE, nr_pages);
1704 else
1705 add_zone_page_state(page_zone(page),
1706 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001707 for (i = 0; i < nr_pages; i++)
1708 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001709
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001710 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1711 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1712
1713 if (cachep->ctor)
1714 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1715 else
1716 kmemcheck_mark_unallocated_pages(page, nr_pages);
1717 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001718
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001719 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720}
1721
1722/*
1723 * Interface to system's page release.
1724 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001725static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001727 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 struct page *page = virt_to_page(addr);
1729 const unsigned long nr_freed = i;
1730
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001731 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001732
Christoph Lameter972d1a72006-09-25 23:31:51 -07001733 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1734 sub_zone_page_state(page_zone(page),
1735 NR_SLAB_RECLAIMABLE, nr_freed);
1736 else
1737 sub_zone_page_state(page_zone(page),
1738 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001740 BUG_ON(!PageSlab(page));
1741 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 page++;
1743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 if (current->reclaim_state)
1745 current->reclaim_state->reclaimed_slab += nr_freed;
1746 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747}
1748
1749static void kmem_rcu_free(struct rcu_head *head)
1750{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001751 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001752 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753
1754 kmem_freepages(cachep, slab_rcu->addr);
1755 if (OFF_SLAB(cachep))
1756 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1757}
1758
1759#if DEBUG
1760
1761#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001762static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001763 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001765 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001767 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001769 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 return;
1771
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001772 *addr++ = 0x12345678;
1773 *addr++ = caller;
1774 *addr++ = smp_processor_id();
1775 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 {
1777 unsigned long *sptr = &caller;
1778 unsigned long svalue;
1779
1780 while (!kstack_end(sptr)) {
1781 svalue = *sptr++;
1782 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001783 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 size -= sizeof(unsigned long);
1785 if (size <= sizeof(unsigned long))
1786 break;
1787 }
1788 }
1789
1790 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792}
1793#endif
1794
Pekka Enberg343e0d72006-02-01 03:05:50 -08001795static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001797 int size = obj_size(cachep);
1798 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
1800 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001801 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802}
1803
1804static void dump_line(char *data, int offset, int limit)
1805{
1806 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001807 unsigned char error = 0;
1808 int bad_count = 0;
1809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001811 for (i = 0; i < limit; i++) {
1812 if (data[offset + i] != POISON_FREE) {
1813 error = data[offset + i];
1814 bad_count++;
1815 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001816 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001817 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001819
1820 if (bad_count == 1) {
1821 error ^= POISON_FREE;
1822 if (!(error & (error - 1))) {
1823 printk(KERN_ERR "Single bit error detected. Probably "
1824 "bad RAM.\n");
1825#ifdef CONFIG_X86
1826 printk(KERN_ERR "Run memtest86+ or a similar memory "
1827 "test tool.\n");
1828#else
1829 printk(KERN_ERR "Run a memory test tool.\n");
1830#endif
1831 }
1832 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833}
1834#endif
1835
1836#if DEBUG
1837
Pekka Enberg343e0d72006-02-01 03:05:50 -08001838static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839{
1840 int i, size;
1841 char *realobj;
1842
1843 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001844 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001845 *dbg_redzone1(cachep, objp),
1846 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 }
1848
1849 if (cachep->flags & SLAB_STORE_USER) {
1850 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001851 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001853 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 printk("\n");
1855 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001856 realobj = (char *)objp + obj_offset(cachep);
1857 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 int limit;
1860 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 if (i + limit > size)
1862 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 dump_line(realobj, i, limit);
1864 }
1865}
1866
Pekka Enberg343e0d72006-02-01 03:05:50 -08001867static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868{
1869 char *realobj;
1870 int size, i;
1871 int lines = 0;
1872
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001873 realobj = (char *)objp + obj_offset(cachep);
1874 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001876 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001878 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 exp = POISON_END;
1880 if (realobj[i] != exp) {
1881 int limit;
1882 /* Mismatch ! */
1883 /* Print header */
1884 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001885 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001886 "Slab corruption: %s start=%p, len=%d\n",
1887 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 print_objinfo(cachep, objp, 0);
1889 }
1890 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001891 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001893 if (i + limit > size)
1894 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 dump_line(realobj, i, limit);
1896 i += 16;
1897 lines++;
1898 /* Limit to 5 lines */
1899 if (lines > 5)
1900 break;
1901 }
1902 }
1903 if (lines != 0) {
1904 /* Print some data about the neighboring objects, if they
1905 * exist:
1906 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001907 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001908 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001910 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001912 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001913 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 print_objinfo(cachep, objp, 2);
1917 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001918 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001919 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001920 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001922 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 print_objinfo(cachep, objp, 2);
1924 }
1925 }
1926}
1927#endif
1928
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301930static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001931{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 int i;
1933 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001934 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935
1936 if (cachep->flags & SLAB_POISON) {
1937#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001938 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1939 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001940 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001941 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 else
1943 check_poison_obj(cachep, objp);
1944#else
1945 check_poison_obj(cachep, objp);
1946#endif
1947 }
1948 if (cachep->flags & SLAB_RED_ZONE) {
1949 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1950 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001951 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1953 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001954 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001957}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301959static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001960{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001961}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962#endif
1963
Randy Dunlap911851e2006-03-22 00:08:14 -08001964/**
1965 * slab_destroy - destroy and release all objects in a slab
1966 * @cachep: cache pointer being destroyed
1967 * @slabp: slab pointer being destroyed
1968 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001969 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001970 * Before calling the slab must have been unlinked from the cache. The
1971 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001972 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001973static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001974{
1975 void *addr = slabp->s_mem - slabp->colouroff;
1976
Rabin Vincente79aec22008-07-04 00:40:32 +05301977 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1979 struct slab_rcu *slab_rcu;
1980
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001981 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 slab_rcu->cachep = cachep;
1983 slab_rcu->addr = addr;
1984 call_rcu(&slab_rcu->head, kmem_rcu_free);
1985 } else {
1986 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001987 if (OFF_SLAB(cachep))
1988 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 }
1990}
1991
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001992static void __kmem_cache_destroy(struct kmem_cache *cachep)
1993{
1994 int i;
1995 struct kmem_list3 *l3;
1996
1997 for_each_online_cpu(i)
1998 kfree(cachep->array[i]);
1999
2000 /* NUMA: free the list3 structures */
2001 for_each_online_node(i) {
2002 l3 = cachep->nodelists[i];
2003 if (l3) {
2004 kfree(l3->shared);
2005 free_alien_cache(l3->alien);
2006 kfree(l3);
2007 }
2008 }
2009 kmem_cache_free(&cache_cache, cachep);
2010}
2011
2012
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002014 * calculate_slab_order - calculate size (page order) of slabs
2015 * @cachep: pointer to the cache that is being created
2016 * @size: size of objects to be created in this cache.
2017 * @align: required alignment for the objects.
2018 * @flags: slab allocation flags
2019 *
2020 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002021 *
2022 * This could be made much more intelligent. For now, try to avoid using
2023 * high order pages for slabs. When the gfp() functions are more friendly
2024 * towards high-order requests, this should be changed.
2025 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002026static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002027 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002028{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002029 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002030 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002033 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002034 unsigned int num;
2035 size_t remainder;
2036
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002037 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002038 if (!num)
2039 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002040
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002041 if (flags & CFLGS_OFF_SLAB) {
2042 /*
2043 * Max number of objs-per-slab for caches which
2044 * use off-slab slabs. Needed to avoid a possible
2045 * looping condition in cache_grow().
2046 */
2047 offslab_limit = size - sizeof(struct slab);
2048 offslab_limit /= sizeof(kmem_bufctl_t);
2049
2050 if (num > offslab_limit)
2051 break;
2052 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002053
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002054 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002055 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002056 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057 left_over = remainder;
2058
2059 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002060 * A VFS-reclaimable slab tends to have most allocations
2061 * as GFP_NOFS and we really don't want to have to be allocating
2062 * higher-order pages when we are unable to shrink dcache.
2063 */
2064 if (flags & SLAB_RECLAIM_ACCOUNT)
2065 break;
2066
2067 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002068 * Large number of objects is good, but very large slabs are
2069 * currently bad for the gfp()s.
2070 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002071 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002072 break;
2073
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002074 /*
2075 * Acceptable internal fragmentation?
2076 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002077 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002078 break;
2079 }
2080 return left_over;
2081}
2082
Pekka Enberg83b519e2009-06-10 19:40:04 +03002083static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002084{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002085 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002086 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002087
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002088 if (g_cpucache_up == NONE) {
2089 /*
2090 * Note: the first kmem_cache_create must create the cache
2091 * that's used by kmalloc(24), otherwise the creation of
2092 * further caches will BUG().
2093 */
2094 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2095
2096 /*
2097 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2098 * the first cache, then we need to set up all its list3s,
2099 * otherwise the creation of further caches will BUG().
2100 */
2101 set_up_list3s(cachep, SIZE_AC);
2102 if (INDEX_AC == INDEX_L3)
2103 g_cpucache_up = PARTIAL_L3;
2104 else
2105 g_cpucache_up = PARTIAL_AC;
2106 } else {
2107 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002108 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002109
2110 if (g_cpucache_up == PARTIAL_AC) {
2111 set_up_list3s(cachep, SIZE_L3);
2112 g_cpucache_up = PARTIAL_L3;
2113 } else {
2114 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002115 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002116 cachep->nodelists[node] =
2117 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002118 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002119 BUG_ON(!cachep->nodelists[node]);
2120 kmem_list3_init(cachep->nodelists[node]);
2121 }
2122 }
2123 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002124 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002125 jiffies + REAPTIMEOUT_LIST3 +
2126 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2127
2128 cpu_cache_get(cachep)->avail = 0;
2129 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2130 cpu_cache_get(cachep)->batchcount = 1;
2131 cpu_cache_get(cachep)->touched = 0;
2132 cachep->batchcount = 1;
2133 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002134 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002135}
2136
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002137/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 * kmem_cache_create - Create a cache.
2139 * @name: A string which is used in /proc/slabinfo to identify this cache.
2140 * @size: The size of objects to be created in this cache.
2141 * @align: The required alignment for the objects.
2142 * @flags: SLAB flags
2143 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 *
2145 * Returns a ptr to the cache on success, NULL on failure.
2146 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002147 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 *
2149 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002150 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002151 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2152 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002153 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154 * The flags are
2155 *
2156 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2157 * to catch references to uninitialised memory.
2158 *
2159 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2160 * for buffer overruns.
2161 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2163 * cacheline. This can be beneficial if you're counting cycles as closely
2164 * as davem.
2165 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002166struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002168 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169{
2170 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002171 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002172 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173
2174 /*
2175 * Sanity checks... these are all serious usage bugs.
2176 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002177 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002178 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002179 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002180 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002181 BUG();
2182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002184 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002185 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302186 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002187 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002188 if (slab_is_available()) {
2189 get_online_cpus();
2190 mutex_lock(&cache_chain_mutex);
2191 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002192
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002193 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002194 char tmp;
2195 int res;
2196
2197 /*
2198 * This happens when the module gets unloaded and doesn't
2199 * destroy its slab cache and no-one else reuses the vmalloc
2200 * area of the module. Print a warning.
2201 */
Andrew Morton138ae662006-12-06 20:36:41 -08002202 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002203 if (res) {
matzeb4169522007-05-06 14:49:52 -07002204 printk(KERN_ERR
2205 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002206 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002207 continue;
2208 }
2209
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002210 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002211 printk(KERN_ERR
2212 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002213 dump_stack();
2214 goto oops;
2215 }
2216 }
2217
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218#if DEBUG
2219 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220#if FORCED_DEBUG
2221 /*
2222 * Enable redzoning and last user accounting, except for caches with
2223 * large objects, if the increased size would increase the object size
2224 * above the next power of two: caches with object sizes just above a
2225 * power of two have a significant amount of internal fragmentation.
2226 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002227 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2228 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002229 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 if (!(flags & SLAB_DESTROY_BY_RCU))
2231 flags |= SLAB_POISON;
2232#endif
2233 if (flags & SLAB_DESTROY_BY_RCU)
2234 BUG_ON(flags & SLAB_POISON);
2235#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002237 * Always checks flags, a caller might be expecting debug support which
2238 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002240 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241
Andrew Mortona737b3e2006-03-22 00:08:11 -08002242 /*
2243 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 * unaligned accesses for some archs when redzoning is used, and makes
2245 * sure any on-slab bufctl's are also correctly aligned.
2246 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002247 if (size & (BYTES_PER_WORD - 1)) {
2248 size += (BYTES_PER_WORD - 1);
2249 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 }
2251
Andrew Mortona737b3e2006-03-22 00:08:11 -08002252 /* calculate the final buffer alignment: */
2253
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 /* 1) arch recommendation: can be overridden for debug */
2255 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002256 /*
2257 * Default alignment: as specified by the arch code. Except if
2258 * an object is really small, then squeeze multiple objects into
2259 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 */
2261 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002262 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 ralign /= 2;
2264 } else {
2265 ralign = BYTES_PER_WORD;
2266 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002267
2268 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002269 * Redzoning and user store require word alignment or possibly larger.
2270 * Note this will be overridden by architecture or caller mandated
2271 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002272 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002273 if (flags & SLAB_STORE_USER)
2274 ralign = BYTES_PER_WORD;
2275
2276 if (flags & SLAB_RED_ZONE) {
2277 ralign = REDZONE_ALIGN;
2278 /* If redzoning, ensure that the second redzone is suitably
2279 * aligned, by adjusting the object size accordingly. */
2280 size += REDZONE_ALIGN - 1;
2281 size &= ~(REDZONE_ALIGN - 1);
2282 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002283
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002284 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 if (ralign < ARCH_SLAB_MINALIGN) {
2286 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002288 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 if (ralign < align) {
2290 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 }
Shiyong Li5c5e3b32010-04-12 13:48:21 +08002292 /* disable debug if not aligning with REDZONE_ALIGN */
2293 if (ralign & (__alignof__(unsigned long long) - 1))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002294 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002295 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002296 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 */
2298 align = ralign;
2299
Pekka Enberg83b519e2009-06-10 19:40:04 +03002300 if (slab_is_available())
2301 gfp = GFP_KERNEL;
2302 else
2303 gfp = GFP_NOWAIT;
2304
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002306 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002308 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309
2310#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002311 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
Pekka Enbergca5f9702006-09-25 23:31:25 -07002313 /*
2314 * Both debugging options require word-alignment which is calculated
2315 * into align above.
2316 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 /* add space for red zone words */
Shiyong Li5c5e3b32010-04-12 13:48:21 +08002319 cachep->obj_offset += align;
2320 size += align + sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 }
2322 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002323 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002324 * the real object. But if the second red zone needs to be
2325 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002327 if (flags & SLAB_RED_ZONE)
2328 size += REDZONE_ALIGN;
2329 else
2330 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 }
2332#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002333 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002334 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2335 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 size = PAGE_SIZE;
2337 }
2338#endif
2339#endif
2340
Ingo Molnare0a42722006-06-23 02:03:46 -07002341 /*
2342 * Determine if the slab management is 'on' or 'off' slab.
2343 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002344 * it too early on. Always use on-slab management when
2345 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002346 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002347 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2348 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 /*
2350 * Size is large, assume best to place the slab management obj
2351 * off-slab (should allow better packing of objs).
2352 */
2353 flags |= CFLGS_OFF_SLAB;
2354
2355 size = ALIGN(size, align);
2356
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002357 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
2359 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002360 printk(KERN_ERR
2361 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 kmem_cache_free(&cache_cache, cachep);
2363 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002364 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002366 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2367 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
2369 /*
2370 * If the slab has been placed off-slab, and we have enough space then
2371 * move it on-slab. This is at the expense of any extra colouring.
2372 */
2373 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2374 flags &= ~CFLGS_OFF_SLAB;
2375 left_over -= slab_size;
2376 }
2377
2378 if (flags & CFLGS_OFF_SLAB) {
2379 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002380 slab_size =
2381 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302382
2383#ifdef CONFIG_PAGE_POISONING
2384 /* If we're going to use the generic kernel_map_pages()
2385 * poisoning, then it's going to smash the contents of
2386 * the redzone and userword anyhow, so switch them off.
2387 */
2388 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2389 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2390#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 }
2392
2393 cachep->colour_off = cache_line_size();
2394 /* Offset must be a multiple of the alignment. */
2395 if (cachep->colour_off < align)
2396 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002397 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 cachep->slab_size = slab_size;
2399 cachep->flags = flags;
2400 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002401 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002403 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002404 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002406 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002407 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002408 /*
2409 * This is a possibility for one of the malloc_sizes caches.
2410 * But since we go off slab only for object size greater than
2411 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2412 * this should not happen at all.
2413 * But leave a BUG_ON for some lucky dude.
2414 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002415 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 cachep->name = name;
2419
Pekka Enberg83b519e2009-06-10 19:40:04 +03002420 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002421 __kmem_cache_destroy(cachep);
2422 cachep = NULL;
2423 goto oops;
2424 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 /* cache setup completed, link it into the list */
2427 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002428oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 if (!cachep && (flags & SLAB_PANIC))
2430 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002431 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002432 if (slab_is_available()) {
2433 mutex_unlock(&cache_chain_mutex);
2434 put_online_cpus();
2435 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 return cachep;
2437}
2438EXPORT_SYMBOL(kmem_cache_create);
2439
2440#if DEBUG
2441static void check_irq_off(void)
2442{
2443 BUG_ON(!irqs_disabled());
2444}
2445
2446static void check_irq_on(void)
2447{
2448 BUG_ON(irqs_disabled());
2449}
2450
Pekka Enberg343e0d72006-02-01 03:05:50 -08002451static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452{
2453#ifdef CONFIG_SMP
2454 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002455 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456#endif
2457}
Christoph Lametere498be72005-09-09 13:03:32 -07002458
Pekka Enberg343e0d72006-02-01 03:05:50 -08002459static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002460{
2461#ifdef CONFIG_SMP
2462 check_irq_off();
2463 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2464#endif
2465}
2466
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467#else
2468#define check_irq_off() do { } while(0)
2469#define check_irq_on() do { } while(0)
2470#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002471#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472#endif
2473
Christoph Lameteraab22072006-03-22 00:09:06 -08002474static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2475 struct array_cache *ac,
2476 int force, int node);
2477
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478static void do_drain(void *arg)
2479{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002480 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002482 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483
2484 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002485 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002486 spin_lock(&cachep->nodelists[node]->list_lock);
2487 free_block(cachep, ac->entry, ac->avail, node);
2488 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 ac->avail = 0;
2490}
2491
Pekka Enberg343e0d72006-02-01 03:05:50 -08002492static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493{
Christoph Lametere498be72005-09-09 13:03:32 -07002494 struct kmem_list3 *l3;
2495 int node;
2496
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002497 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002499 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002500 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002501 if (l3 && l3->alien)
2502 drain_alien_cache(cachep, l3->alien);
2503 }
2504
2505 for_each_online_node(node) {
2506 l3 = cachep->nodelists[node];
2507 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002508 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002509 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510}
2511
Christoph Lametered11d9e2006-06-30 01:55:45 -07002512/*
2513 * Remove slabs from the list of free slabs.
2514 * Specify the number of slabs to drain in tofree.
2515 *
2516 * Returns the actual number of slabs released.
2517 */
2518static int drain_freelist(struct kmem_cache *cache,
2519 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002521 struct list_head *p;
2522 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524
Christoph Lametered11d9e2006-06-30 01:55:45 -07002525 nr_freed = 0;
2526 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527
Christoph Lametered11d9e2006-06-30 01:55:45 -07002528 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002529 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002530 if (p == &l3->slabs_free) {
2531 spin_unlock_irq(&l3->list_lock);
2532 goto out;
2533 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534
Christoph Lametered11d9e2006-06-30 01:55:45 -07002535 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002537 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538#endif
2539 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002540 /*
2541 * Safe to drop the lock. The slab is no longer linked
2542 * to the cache.
2543 */
2544 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002545 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002546 slab_destroy(cache, slabp);
2547 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002549out:
2550 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551}
2552
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002553/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002554static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002555{
2556 int ret = 0, i = 0;
2557 struct kmem_list3 *l3;
2558
2559 drain_cpu_caches(cachep);
2560
2561 check_irq_on();
2562 for_each_online_node(i) {
2563 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002564 if (!l3)
2565 continue;
2566
2567 drain_freelist(cachep, l3, l3->free_objects);
2568
2569 ret += !list_empty(&l3->slabs_full) ||
2570 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002571 }
2572 return (ret ? 1 : 0);
2573}
2574
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575/**
2576 * kmem_cache_shrink - Shrink a cache.
2577 * @cachep: The cache to shrink.
2578 *
2579 * Releases as many slabs as possible for a cache.
2580 * To help debugging, a zero exit status indicates all slabs were released.
2581 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002582int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002584 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002585 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002587 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002588 mutex_lock(&cache_chain_mutex);
2589 ret = __cache_shrink(cachep);
2590 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002591 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002592 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593}
2594EXPORT_SYMBOL(kmem_cache_shrink);
2595
2596/**
2597 * kmem_cache_destroy - delete a cache
2598 * @cachep: the cache to destroy
2599 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002600 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 *
2602 * It is expected this function will be called by a module when it is
2603 * unloaded. This will remove the cache completely, and avoid a duplicate
2604 * cache being allocated each time a module is loaded and unloaded, if the
2605 * module doesn't have persistent in-kernel storage across loads and unloads.
2606 *
2607 * The cache must be empty before calling this function.
2608 *
2609 * The caller must guarantee that noone will allocate memory from the cache
2610 * during the kmem_cache_destroy().
2611 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002612void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002614 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002617 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002618 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 /*
2620 * the chain is never empty, cache_cache is never destroyed
2621 */
2622 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 if (__cache_shrink(cachep)) {
2624 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002625 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002626 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002627 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002628 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 }
2630
2631 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002632 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002634 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002635 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002636 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637}
2638EXPORT_SYMBOL(kmem_cache_destroy);
2639
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002640/*
2641 * Get the memory for a slab management obj.
2642 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2643 * always come from malloc_sizes caches. The slab descriptor cannot
2644 * come from the same cache which is getting created because,
2645 * when we are searching for an appropriate cache for these
2646 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2647 * If we are creating a malloc_sizes cache here it would not be visible to
2648 * kmem_find_general_cachep till the initialization is complete.
2649 * Hence we cannot have slabp_cache same as the original cache.
2650 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002651static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002652 int colour_off, gfp_t local_flags,
2653 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654{
2655 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002656
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 if (OFF_SLAB(cachep)) {
2658 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002659 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002660 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002661 /*
2662 * If the first object in the slab is leaked (it's allocated
2663 * but no one has a reference to it), we want to make sure
2664 * kmemleak does not treat the ->s_mem pointer as a reference
2665 * to the object. Otherwise we will not report the leak.
2666 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002667 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2668 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 if (!slabp)
2670 return NULL;
2671 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002672 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 colour_off += cachep->slab_size;
2674 }
2675 slabp->inuse = 0;
2676 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002677 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002678 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002679 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 return slabp;
2681}
2682
2683static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2684{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002685 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686}
2687
Pekka Enberg343e0d72006-02-01 03:05:50 -08002688static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002689 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690{
2691 int i;
2692
2693 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002694 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695#if DEBUG
2696 /* need to poison the objs? */
2697 if (cachep->flags & SLAB_POISON)
2698 poison_obj(cachep, objp, POISON_FREE);
2699 if (cachep->flags & SLAB_STORE_USER)
2700 *dbg_userword(cachep, objp) = NULL;
2701
2702 if (cachep->flags & SLAB_RED_ZONE) {
2703 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2704 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2705 }
2706 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002707 * Constructors are not allowed to allocate memory from the same
2708 * cache which they are a constructor for. Otherwise, deadlock.
2709 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 */
2711 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002712 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713
2714 if (cachep->flags & SLAB_RED_ZONE) {
2715 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2716 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002717 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2719 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002720 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002722 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2723 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002724 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002725 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726#else
2727 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002728 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002730 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002732 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733}
2734
Pekka Enberg343e0d72006-02-01 03:05:50 -08002735static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002737 if (CONFIG_ZONE_DMA_FLAG) {
2738 if (flags & GFP_DMA)
2739 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2740 else
2741 BUG_ON(cachep->gfpflags & GFP_DMA);
2742 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743}
2744
Andrew Mortona737b3e2006-03-22 00:08:11 -08002745static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2746 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002747{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002748 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002749 kmem_bufctl_t next;
2750
2751 slabp->inuse++;
2752 next = slab_bufctl(slabp)[slabp->free];
2753#if DEBUG
2754 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2755 WARN_ON(slabp->nodeid != nodeid);
2756#endif
2757 slabp->free = next;
2758
2759 return objp;
2760}
2761
Andrew Mortona737b3e2006-03-22 00:08:11 -08002762static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2763 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002764{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002765 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002766
2767#if DEBUG
2768 /* Verify that the slab belongs to the intended node */
2769 WARN_ON(slabp->nodeid != nodeid);
2770
Al Viro871751e2006-03-25 03:06:39 -08002771 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002772 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002773 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002774 BUG();
2775 }
2776#endif
2777 slab_bufctl(slabp)[objnr] = slabp->free;
2778 slabp->free = objnr;
2779 slabp->inuse--;
2780}
2781
Pekka Enberg47768742006-06-23 02:03:07 -07002782/*
2783 * Map pages beginning at addr to the given cache and slab. This is required
2784 * for the slab allocator to be able to lookup the cache and slab of a
2785 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2786 */
2787static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2788 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789{
Pekka Enberg47768742006-06-23 02:03:07 -07002790 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 struct page *page;
2792
Pekka Enberg47768742006-06-23 02:03:07 -07002793 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002794
Pekka Enberg47768742006-06-23 02:03:07 -07002795 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002796 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002797 nr_pages <<= cache->gfporder;
2798
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002800 page_set_cache(page, cache);
2801 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002803 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804}
2805
2806/*
2807 * Grow (by 1) the number of slabs within a cache. This is called by
2808 * kmem_cache_alloc() when there are no active objs left in a cache.
2809 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002810static int cache_grow(struct kmem_cache *cachep,
2811 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002813 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002814 size_t offset;
2815 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002816 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817
Andrew Mortona737b3e2006-03-22 00:08:11 -08002818 /*
2819 * Be lazy and only check for valid flags here, keeping it out of the
2820 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002822 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2823 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002825 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002827 l3 = cachep->nodelists[nodeid];
2828 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829
2830 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002831 offset = l3->colour_next;
2832 l3->colour_next++;
2833 if (l3->colour_next >= cachep->colour)
2834 l3->colour_next = 0;
2835 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002837 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838
2839 if (local_flags & __GFP_WAIT)
2840 local_irq_enable();
2841
2842 /*
2843 * The test for missing atomic flag is performed here, rather than
2844 * the more obvious place, simply to reduce the critical path length
2845 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2846 * will eventually be caught here (where it matters).
2847 */
2848 kmem_flagcheck(cachep, flags);
2849
Andrew Mortona737b3e2006-03-22 00:08:11 -08002850 /*
2851 * Get mem for the objs. Attempt to allocate a physical page from
2852 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002853 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002854 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002855 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002856 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 goto failed;
2858
2859 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002860 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002861 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002862 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 goto opps1;
2864
Pekka Enberg47768742006-06-23 02:03:07 -07002865 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866
Christoph Lametera35afb82007-05-16 22:10:57 -07002867 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868
2869 if (local_flags & __GFP_WAIT)
2870 local_irq_disable();
2871 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002872 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
2874 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002875 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002877 l3->free_objects += cachep->num;
2878 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002880opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002882failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 if (local_flags & __GFP_WAIT)
2884 local_irq_disable();
2885 return 0;
2886}
2887
2888#if DEBUG
2889
2890/*
2891 * Perform extra freeing checks:
2892 * - detect bad pointers.
2893 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 */
2895static void kfree_debugcheck(const void *objp)
2896{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 if (!virt_addr_valid(objp)) {
2898 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002899 (unsigned long)objp);
2900 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902}
2903
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002904static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2905{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002906 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002907
2908 redzone1 = *dbg_redzone1(cache, obj);
2909 redzone2 = *dbg_redzone2(cache, obj);
2910
2911 /*
2912 * Redzone is ok.
2913 */
2914 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2915 return;
2916
2917 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2918 slab_error(cache, "double free detected");
2919 else
2920 slab_error(cache, "memory outside object was overwritten");
2921
David Woodhouseb46b8f12007-05-08 00:22:59 -07002922 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002923 obj, redzone1, redzone2);
2924}
2925
Pekka Enberg343e0d72006-02-01 03:05:50 -08002926static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002927 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928{
2929 struct page *page;
2930 unsigned int objnr;
2931 struct slab *slabp;
2932
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002933 BUG_ON(virt_to_cache(objp) != cachep);
2934
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002935 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002937 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938
Pekka Enberg065d41c2005-11-13 16:06:46 -08002939 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940
2941 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002942 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2944 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2945 }
2946 if (cachep->flags & SLAB_STORE_USER)
2947 *dbg_userword(cachep, objp) = caller;
2948
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002949 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950
2951 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002952 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953
Al Viro871751e2006-03-25 03:06:39 -08002954#ifdef CONFIG_DEBUG_SLAB_LEAK
2955 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2956#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 if (cachep->flags & SLAB_POISON) {
2958#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002959 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002961 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002962 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 } else {
2964 poison_obj(cachep, objp, POISON_FREE);
2965 }
2966#else
2967 poison_obj(cachep, objp, POISON_FREE);
2968#endif
2969 }
2970 return objp;
2971}
2972
Pekka Enberg343e0d72006-02-01 03:05:50 -08002973static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974{
2975 kmem_bufctl_t i;
2976 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002977
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 /* Check slab's freelist to see if this obj is there. */
2979 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2980 entries++;
2981 if (entries > cachep->num || i >= cachep->num)
2982 goto bad;
2983 }
2984 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002985bad:
2986 printk(KERN_ERR "slab: Internal list corruption detected in "
2987 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2988 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002989 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002990 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002991 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002992 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002994 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 }
2996 printk("\n");
2997 BUG();
2998 }
2999}
3000#else
3001#define kfree_debugcheck(x) do { } while(0)
3002#define cache_free_debugcheck(x,objp,z) (objp)
3003#define check_slabp(x,y) do { } while(0)
3004#endif
3005
Pekka Enberg343e0d72006-02-01 03:05:50 -08003006static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007{
3008 int batchcount;
3009 struct kmem_list3 *l3;
3010 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003011 int node;
3012
Andrew Mortona737b3e2006-03-22 00:08:11 -08003013retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003014 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003015 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003016 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 batchcount = ac->batchcount;
3018 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003019 /*
3020 * If there was little recent activity on this cache, then
3021 * perform only a partial refill. Otherwise we could generate
3022 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 */
3024 batchcount = BATCHREFILL_LIMIT;
3025 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003026 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027
Christoph Lametere498be72005-09-09 13:03:32 -07003028 BUG_ON(ac->avail > 0 || !l3);
3029 spin_lock(&l3->list_lock);
3030
Christoph Lameter3ded1752006-03-25 03:06:44 -08003031 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003032 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3033 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003034 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003035 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003036
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 while (batchcount > 0) {
3038 struct list_head *entry;
3039 struct slab *slabp;
3040 /* Get slab alloc is to come from. */
3041 entry = l3->slabs_partial.next;
3042 if (entry == &l3->slabs_partial) {
3043 l3->free_touched = 1;
3044 entry = l3->slabs_free.next;
3045 if (entry == &l3->slabs_free)
3046 goto must_grow;
3047 }
3048
3049 slabp = list_entry(entry, struct slab, list);
3050 check_slabp(cachep, slabp);
3051 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003052
3053 /*
3054 * The slab was either on partial or free list so
3055 * there must be at least one object available for
3056 * allocation.
3057 */
roel kluin249b9f32008-10-29 17:18:07 -04003058 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003059
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 STATS_INC_ALLOCED(cachep);
3062 STATS_INC_ACTIVE(cachep);
3063 STATS_SET_HIGH(cachep);
3064
Matthew Dobson78d382d2006-02-01 03:05:47 -08003065 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003066 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067 }
3068 check_slabp(cachep, slabp);
3069
3070 /* move slabp to correct slabp list: */
3071 list_del(&slabp->list);
3072 if (slabp->free == BUFCTL_END)
3073 list_add(&slabp->list, &l3->slabs_full);
3074 else
3075 list_add(&slabp->list, &l3->slabs_partial);
3076 }
3077
Andrew Mortona737b3e2006-03-22 00:08:11 -08003078must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003080alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003081 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082
3083 if (unlikely(!ac->avail)) {
3084 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003085 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003086
Andrew Mortona737b3e2006-03-22 00:08:11 -08003087 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003088 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003089 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 return NULL;
3091
Andrew Mortona737b3e2006-03-22 00:08:11 -08003092 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 goto retry;
3094 }
3095 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003096 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097}
3098
Andrew Mortona737b3e2006-03-22 00:08:11 -08003099static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3100 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101{
3102 might_sleep_if(flags & __GFP_WAIT);
3103#if DEBUG
3104 kmem_flagcheck(cachep, flags);
3105#endif
3106}
3107
3108#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003109static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3110 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003112 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003114 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003116 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003117 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003118 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 else
3120 check_poison_obj(cachep, objp);
3121#else
3122 check_poison_obj(cachep, objp);
3123#endif
3124 poison_obj(cachep, objp, POISON_INUSE);
3125 }
3126 if (cachep->flags & SLAB_STORE_USER)
3127 *dbg_userword(cachep, objp) = caller;
3128
3129 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003130 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3131 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3132 slab_error(cachep, "double free, or memory outside"
3133 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003134 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003135 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003136 objp, *dbg_redzone1(cachep, objp),
3137 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 }
3139 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3140 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3141 }
Al Viro871751e2006-03-25 03:06:39 -08003142#ifdef CONFIG_DEBUG_SLAB_LEAK
3143 {
3144 struct slab *slabp;
3145 unsigned objnr;
3146
Christoph Lameterb49af682007-05-06 14:49:41 -07003147 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003148 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3149 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3150 }
3151#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003152 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003153 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003154 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003155#if ARCH_SLAB_MINALIGN
3156 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3157 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3158 objp, ARCH_SLAB_MINALIGN);
3159 }
3160#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 return objp;
3162}
3163#else
3164#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3165#endif
3166
Akinobu Mita773ff602008-12-23 19:37:01 +09003167static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003168{
3169 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003170 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003171
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003172 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003173}
3174
Pekka Enberg343e0d72006-02-01 03:05:50 -08003175static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003177 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178 struct array_cache *ac;
3179
Alok N Kataria5c382302005-09-27 21:45:46 -07003180 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003181
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003182 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 if (likely(ac->avail)) {
3184 STATS_INC_ALLOCHIT(cachep);
3185 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003186 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 } else {
3188 STATS_INC_ALLOCMISS(cachep);
3189 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003190 /*
3191 * the 'ac' may be updated by cache_alloc_refill(),
3192 * and kmemleak_erase() requires its correct value.
3193 */
3194 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003196 /*
3197 * To avoid a false negative, if an object that is in one of the
3198 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3199 * treat the array pointers as a reference to the object.
3200 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003201 if (objp)
3202 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003203 return objp;
3204}
3205
Christoph Lametere498be72005-09-09 13:03:32 -07003206#ifdef CONFIG_NUMA
3207/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003208 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003209 *
3210 * If we are in_interrupt, then process context, including cpusets and
3211 * mempolicy, may not apply and should not be used for allocation policy.
3212 */
3213static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3214{
3215 int nid_alloc, nid_here;
3216
Christoph Lameter765c4502006-09-27 01:50:08 -07003217 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003218 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003219 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003220 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003221 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003222 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003223 else if (current->mempolicy)
3224 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003225 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003226 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003227 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003228 return NULL;
3229}
3230
3231/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003232 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003233 * certain node and fall back is permitted. First we scan all the
3234 * available nodelists for available objects. If that fails then we
3235 * perform an allocation without specifying a node. This allows the page
3236 * allocator to do its reclaim / fallback magic. We then insert the
3237 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003238 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003239static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003240{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003241 struct zonelist *zonelist;
3242 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003243 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003244 struct zone *zone;
3245 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003246 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003247 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003248
3249 if (flags & __GFP_THISNODE)
3250 return NULL;
3251
Miao Xiec0ff7452010-05-24 14:32:08 -07003252 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003253 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003254 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003255
Christoph Lameter3c517a62006-12-06 20:33:29 -08003256retry:
3257 /*
3258 * Look through allowed nodes for objects available
3259 * from existing per node queues.
3260 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003261 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3262 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003263
Mel Gorman54a6eb52008-04-28 02:12:16 -07003264 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003265 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003266 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003267 obj = ____cache_alloc_node(cache,
3268 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003269 if (obj)
3270 break;
3271 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003272 }
3273
Christoph Lametercfce6602007-05-06 14:50:17 -07003274 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003275 /*
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3280 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003281 if (local_flags & __GFP_WAIT)
3282 local_irq_enable();
3283 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003284 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003285 if (local_flags & __GFP_WAIT)
3286 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003287 if (obj) {
3288 /*
3289 * Insert into the appropriate per node queues
3290 */
3291 nid = page_to_nid(virt_to_page(obj));
3292 if (cache_grow(cache, flags, nid, obj)) {
3293 obj = ____cache_alloc_node(cache,
3294 flags | GFP_THISNODE, nid);
3295 if (!obj)
3296 /*
3297 * Another processor may allocate the
3298 * objects in the slab since we are
3299 * not holding any locks.
3300 */
3301 goto retry;
3302 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003303 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003304 obj = NULL;
3305 }
3306 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003307 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003308 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003309 return obj;
3310}
3311
3312/*
Christoph Lametere498be72005-09-09 13:03:32 -07003313 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003315static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003316 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003317{
3318 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003319 struct slab *slabp;
3320 struct kmem_list3 *l3;
3321 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003322 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003324 l3 = cachep->nodelists[nodeid];
3325 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003326
Andrew Mortona737b3e2006-03-22 00:08:11 -08003327retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003328 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003329 spin_lock(&l3->list_lock);
3330 entry = l3->slabs_partial.next;
3331 if (entry == &l3->slabs_partial) {
3332 l3->free_touched = 1;
3333 entry = l3->slabs_free.next;
3334 if (entry == &l3->slabs_free)
3335 goto must_grow;
3336 }
Christoph Lametere498be72005-09-09 13:03:32 -07003337
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 slabp = list_entry(entry, struct slab, list);
3339 check_spinlock_acquired_node(cachep, nodeid);
3340 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003341
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 STATS_INC_NODEALLOCS(cachep);
3343 STATS_INC_ACTIVE(cachep);
3344 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003345
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003346 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003347
Matthew Dobson78d382d2006-02-01 03:05:47 -08003348 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003349 check_slabp(cachep, slabp);
3350 l3->free_objects--;
3351 /* move slabp to correct slabp list: */
3352 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003353
Andrew Mortona737b3e2006-03-22 00:08:11 -08003354 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003355 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003356 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003357 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003358
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003359 spin_unlock(&l3->list_lock);
3360 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003361
Andrew Mortona737b3e2006-03-22 00:08:11 -08003362must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003363 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003364 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003365 if (x)
3366 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003367
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003368 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003369
Andrew Mortona737b3e2006-03-22 00:08:11 -08003370done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003371 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003372}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003373
3374/**
3375 * kmem_cache_alloc_node - Allocate an object on the specified node
3376 * @cachep: The cache to allocate from.
3377 * @flags: See kmalloc().
3378 * @nodeid: node number of the target node.
3379 * @caller: return address of caller, used for debug information
3380 *
3381 * Identical to kmem_cache_alloc but it will allocate memory on the given
3382 * node, which can improve the performance for cpu bound structures.
3383 *
3384 * Fallback to other node is possible if __GFP_THISNODE is not set.
3385 */
3386static __always_inline void *
3387__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3388 void *caller)
3389{
3390 unsigned long save_flags;
3391 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003392 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003393
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003394 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003395
Nick Piggincf40bd12009-01-21 08:12:39 +01003396 lockdep_trace_alloc(flags);
3397
Akinobu Mita773ff602008-12-23 19:37:01 +09003398 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003399 return NULL;
3400
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003401 cache_alloc_debugcheck_before(cachep, flags);
3402 local_irq_save(save_flags);
3403
Tim Blechmann8e15b792009-11-30 18:59:34 +01003404 if (nodeid == -1)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003405 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003406
3407 if (unlikely(!cachep->nodelists[nodeid])) {
3408 /* Node not bootstrapped yet */
3409 ptr = fallback_alloc(cachep, flags);
3410 goto out;
3411 }
3412
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003413 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003414 /*
3415 * Use the locally cached objects if possible.
3416 * However ____cache_alloc does not allow fallback
3417 * to other nodes. It may fail while we still have
3418 * objects on other nodes available.
3419 */
3420 ptr = ____cache_alloc(cachep, flags);
3421 if (ptr)
3422 goto out;
3423 }
3424 /* ___cache_alloc_node can fall back to other nodes */
3425 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3426 out:
3427 local_irq_restore(save_flags);
3428 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003429 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3430 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003431
Pekka Enbergc175eea2008-05-09 20:35:53 +02003432 if (likely(ptr))
3433 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3434
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003435 if (unlikely((flags & __GFP_ZERO) && ptr))
3436 memset(ptr, 0, obj_size(cachep));
3437
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003438 return ptr;
3439}
3440
3441static __always_inline void *
3442__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3443{
3444 void *objp;
3445
3446 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3447 objp = alternate_node_alloc(cache, flags);
3448 if (objp)
3449 goto out;
3450 }
3451 objp = ____cache_alloc(cache, flags);
3452
3453 /*
3454 * We may just have run out of memory on the local node.
3455 * ____cache_alloc_node() knows how to locate memory on other nodes
3456 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003457 if (!objp)
3458 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003459
3460 out:
3461 return objp;
3462}
3463#else
3464
3465static __always_inline void *
3466__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3467{
3468 return ____cache_alloc(cachep, flags);
3469}
3470
3471#endif /* CONFIG_NUMA */
3472
3473static __always_inline void *
3474__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3475{
3476 unsigned long save_flags;
3477 void *objp;
3478
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003479 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003480
Nick Piggincf40bd12009-01-21 08:12:39 +01003481 lockdep_trace_alloc(flags);
3482
Akinobu Mita773ff602008-12-23 19:37:01 +09003483 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003484 return NULL;
3485
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003486 cache_alloc_debugcheck_before(cachep, flags);
3487 local_irq_save(save_flags);
3488 objp = __do_cache_alloc(cachep, flags);
3489 local_irq_restore(save_flags);
3490 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003491 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3492 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003493 prefetchw(objp);
3494
Pekka Enbergc175eea2008-05-09 20:35:53 +02003495 if (likely(objp))
3496 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3497
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003498 if (unlikely((flags & __GFP_ZERO) && objp))
3499 memset(objp, 0, obj_size(cachep));
3500
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003501 return objp;
3502}
Christoph Lametere498be72005-09-09 13:03:32 -07003503
3504/*
3505 * Caller needs to acquire correct kmem_list's list_lock
3506 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003507static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003508 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509{
3510 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003511 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512
3513 for (i = 0; i < nr_objects; i++) {
3514 void *objp = objpp[i];
3515 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003517 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003518 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003520 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003522 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003524 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 check_slabp(cachep, slabp);
3526
3527 /* fixup slab chains */
3528 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003529 if (l3->free_objects > l3->free_limit) {
3530 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003531 /* No need to drop any previously held
3532 * lock here, even if we have a off-slab slab
3533 * descriptor it is guaranteed to come from
3534 * a different cache, refer to comments before
3535 * alloc_slabmgmt.
3536 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 slab_destroy(cachep, slabp);
3538 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003539 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 }
3541 } else {
3542 /* Unconditionally move a slab to the end of the
3543 * partial list on free - maximum time for the
3544 * other objects to be freed, too.
3545 */
Christoph Lametere498be72005-09-09 13:03:32 -07003546 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547 }
3548 }
3549}
3550
Pekka Enberg343e0d72006-02-01 03:05:50 -08003551static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552{
3553 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003554 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003555 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556
3557 batchcount = ac->batchcount;
3558#if DEBUG
3559 BUG_ON(!batchcount || batchcount > ac->avail);
3560#endif
3561 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003562 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003563 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003564 if (l3->shared) {
3565 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003566 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 if (max) {
3568 if (batchcount > max)
3569 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003570 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003571 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 shared_array->avail += batchcount;
3573 goto free_done;
3574 }
3575 }
3576
Christoph Lameterff694162005-09-22 21:44:02 -07003577 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003578free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579#if STATS
3580 {
3581 int i = 0;
3582 struct list_head *p;
3583
Christoph Lametere498be72005-09-09 13:03:32 -07003584 p = l3->slabs_free.next;
3585 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 struct slab *slabp;
3587
3588 slabp = list_entry(p, struct slab, list);
3589 BUG_ON(slabp->inuse);
3590
3591 i++;
3592 p = p->next;
3593 }
3594 STATS_SET_FREEABLE(cachep, i);
3595 }
3596#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003597 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003599 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600}
3601
3602/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003603 * Release an obj back to its cache. If the obj has a constructed state, it must
3604 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003606static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003608 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609
3610 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003611 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3613
Pekka Enbergc175eea2008-05-09 20:35:53 +02003614 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3615
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003616 /*
3617 * Skip calling cache_free_alien() when the platform is not numa.
3618 * This will avoid cache misses that happen while accessing slabp (which
3619 * is per page memory reference) to get nodeid. Instead use a global
3620 * variable to skip the call, which is mostly likely to be present in
3621 * the cache.
3622 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003623 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003624 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003625
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 if (likely(ac->avail < ac->limit)) {
3627 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003628 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 return;
3630 } else {
3631 STATS_INC_FREEMISS(cachep);
3632 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003633 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 }
3635}
3636
3637/**
3638 * kmem_cache_alloc - Allocate an object
3639 * @cachep: The cache to allocate from.
3640 * @flags: See kmalloc().
3641 *
3642 * Allocate an object from this cache. The flags are only relevant
3643 * if the cache has no available objects.
3644 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003645void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003647 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3648
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003649 trace_kmem_cache_alloc(_RET_IP_, ret,
3650 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003651
3652 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653}
3654EXPORT_SYMBOL(kmem_cache_alloc);
3655
Li Zefan0f24f122009-12-11 15:45:30 +08003656#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003657void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3658{
3659 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3660}
3661EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3662#endif
3663
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664/**
Randy Dunlap76824862008-03-19 17:00:40 -07003665 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 * @cachep: the cache we're checking against
3667 * @ptr: pointer to validate
3668 *
Randy Dunlap76824862008-03-19 17:00:40 -07003669 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 * it is _not_ a guarantee that the pointer is actually
3671 * part of the slab cache in question, but it at least
3672 * validates that the pointer can be dereferenced and
3673 * looks half-way sane.
3674 *
3675 * Currently only used for dentry validation.
3676 */
Christoph Lameterb7f869a22006-12-22 01:06:44 -08003677int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003679 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 struct page *page;
3681
Pekka Enbergfc1c1832010-04-07 19:23:40 +03003682 if (unlikely(!kern_ptr_validate(ptr, size)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 goto out;
3684 page = virt_to_page(ptr);
3685 if (unlikely(!PageSlab(page)))
3686 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003687 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688 goto out;
3689 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003690out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 return 0;
3692}
3693
3694#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003695void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3696{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003697 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3698 __builtin_return_address(0));
3699
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003700 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3701 obj_size(cachep), cachep->buffer_size,
3702 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003703
3704 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706EXPORT_SYMBOL(kmem_cache_alloc_node);
3707
Li Zefan0f24f122009-12-11 15:45:30 +08003708#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003709void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3710 gfp_t flags,
3711 int nodeid)
3712{
3713 return __cache_alloc_node(cachep, flags, nodeid,
3714 __builtin_return_address(0));
3715}
3716EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3717#endif
3718
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003719static __always_inline void *
3720__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003721{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003722 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003723 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003724
3725 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003726 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3727 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003728 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3729
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003730 trace_kmalloc_node((unsigned long) caller, ret,
3731 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003732
3733 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003734}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003735
Li Zefan0bb38a52009-12-11 15:45:50 +08003736#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003737void *__kmalloc_node(size_t size, gfp_t flags, int node)
3738{
3739 return __do_kmalloc_node(size, flags, node,
3740 __builtin_return_address(0));
3741}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003742EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003743
3744void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003745 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003746{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003747 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003748}
3749EXPORT_SYMBOL(__kmalloc_node_track_caller);
3750#else
3751void *__kmalloc_node(size_t size, gfp_t flags, int node)
3752{
3753 return __do_kmalloc_node(size, flags, node, NULL);
3754}
3755EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003756#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003757#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758
3759/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003760 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003762 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003763 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003765static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3766 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003768 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003769 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003770
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003771 /* If you want to save a few bytes .text space: replace
3772 * __ with kmem_.
3773 * Then kmalloc uses the uninlined functions instead of the inline
3774 * functions.
3775 */
3776 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003777 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3778 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003779 ret = __cache_alloc(cachep, flags, caller);
3780
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003781 trace_kmalloc((unsigned long) caller, ret,
3782 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003783
3784 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003785}
3786
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003787
Li Zefan0bb38a52009-12-11 15:45:50 +08003788#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003789void *__kmalloc(size_t size, gfp_t flags)
3790{
Al Viro871751e2006-03-25 03:06:39 -08003791 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792}
3793EXPORT_SYMBOL(__kmalloc);
3794
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003795void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003796{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003797 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003798}
3799EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003800
3801#else
3802void *__kmalloc(size_t size, gfp_t flags)
3803{
3804 return __do_kmalloc(size, flags, NULL);
3805}
3806EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003807#endif
3808
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809/**
3810 * kmem_cache_free - Deallocate an object
3811 * @cachep: The cache the allocation was from.
3812 * @objp: The previously allocated object.
3813 *
3814 * Free an object which was previously allocated from this
3815 * cache.
3816 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003817void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818{
3819 unsigned long flags;
3820
3821 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003822 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003823 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3824 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003825 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003827
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003828 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829}
3830EXPORT_SYMBOL(kmem_cache_free);
3831
3832/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 * kfree - free previously allocated memory
3834 * @objp: pointer returned by kmalloc.
3835 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003836 * If @objp is NULL, no operation is performed.
3837 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838 * Don't free memory not originally allocated by kmalloc()
3839 * or you will run into trouble.
3840 */
3841void kfree(const void *objp)
3842{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003843 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 unsigned long flags;
3845
Pekka Enberg2121db72009-03-25 11:05:57 +02003846 trace_kfree(_RET_IP_, objp);
3847
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003848 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849 return;
3850 local_irq_save(flags);
3851 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003852 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003853 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003854 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003855 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 local_irq_restore(flags);
3857}
3858EXPORT_SYMBOL(kfree);
3859
Pekka Enberg343e0d72006-02-01 03:05:50 -08003860unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003862 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863}
3864EXPORT_SYMBOL(kmem_cache_size);
3865
Pekka Enberg343e0d72006-02-01 03:05:50 -08003866const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003867{
3868 return cachep->name;
3869}
3870EXPORT_SYMBOL_GPL(kmem_cache_name);
3871
Christoph Lametere498be72005-09-09 13:03:32 -07003872/*
Simon Arlott183ff222007-10-20 01:27:18 +02003873 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003874 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003875static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003876{
3877 int node;
3878 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003879 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003880 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003881
Mel Gorman9c09a952008-01-24 05:49:54 -08003882 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003883
Paul Menage3395ee02006-12-06 20:32:16 -08003884 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003885 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003886 if (!new_alien)
3887 goto fail;
3888 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003889
Eric Dumazet63109842007-05-06 14:49:28 -07003890 new_shared = NULL;
3891 if (cachep->shared) {
3892 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003893 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003894 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003895 if (!new_shared) {
3896 free_alien_cache(new_alien);
3897 goto fail;
3898 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003899 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003900
Andrew Mortona737b3e2006-03-22 00:08:11 -08003901 l3 = cachep->nodelists[node];
3902 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003903 struct array_cache *shared = l3->shared;
3904
Christoph Lametere498be72005-09-09 13:03:32 -07003905 spin_lock_irq(&l3->list_lock);
3906
Christoph Lametercafeb022006-03-25 03:06:46 -08003907 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003908 free_block(cachep, shared->entry,
3909 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003910
Christoph Lametercafeb022006-03-25 03:06:46 -08003911 l3->shared = new_shared;
3912 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003913 l3->alien = new_alien;
3914 new_alien = NULL;
3915 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003916 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003917 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003918 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003919 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003920 free_alien_cache(new_alien);
3921 continue;
3922 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003923 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003924 if (!l3) {
3925 free_alien_cache(new_alien);
3926 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003927 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003928 }
Christoph Lametere498be72005-09-09 13:03:32 -07003929
3930 kmem_list3_init(l3);
3931 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003932 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003933 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003934 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003935 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003936 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003937 cachep->nodelists[node] = l3;
3938 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003939 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003940
Andrew Mortona737b3e2006-03-22 00:08:11 -08003941fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003942 if (!cachep->next.next) {
3943 /* Cache is not active yet. Roll back what we did */
3944 node--;
3945 while (node >= 0) {
3946 if (cachep->nodelists[node]) {
3947 l3 = cachep->nodelists[node];
3948
3949 kfree(l3->shared);
3950 free_alien_cache(l3->alien);
3951 kfree(l3);
3952 cachep->nodelists[node] = NULL;
3953 }
3954 node--;
3955 }
3956 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003957 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003958}
3959
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003961 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 struct array_cache *new[NR_CPUS];
3963};
3964
3965static void do_ccupdate_local(void *info)
3966{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003967 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 struct array_cache *old;
3969
3970 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003971 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003972
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3974 new->new[smp_processor_id()] = old;
3975}
3976
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003977/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003978static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003979 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003981 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003982 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983
Pekka Enberg83b519e2009-06-10 19:40:04 +03003984 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003985 if (!new)
3986 return -ENOMEM;
3987
Christoph Lametere498be72005-09-09 13:03:32 -07003988 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003989 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003990 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003991 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003992 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003993 kfree(new->new[i]);
3994 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003995 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996 }
3997 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003998 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004000 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004001
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 cachep->batchcount = batchcount;
4004 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004005 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
Christoph Lametere498be72005-09-09 13:03:32 -07004007 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004008 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009 if (!ccold)
4010 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004011 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4012 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4013 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014 kfree(ccold);
4015 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004016 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004017 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018}
4019
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004020/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004021static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022{
4023 int err;
4024 int limit, shared;
4025
Andrew Mortona737b3e2006-03-22 00:08:11 -08004026 /*
4027 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 * - create a LIFO ordering, i.e. return objects that are cache-warm
4029 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004030 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 * bufctl chains: array operations are cheaper.
4032 * The numbers are guessed, we should auto-tune as described by
4033 * Bonwick.
4034 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004035 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004037 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004039 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004041 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 limit = 54;
4043 else
4044 limit = 120;
4045
Andrew Mortona737b3e2006-03-22 00:08:11 -08004046 /*
4047 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 * allocation behaviour: Most allocs on one cpu, most free operations
4049 * on another cpu. For these cases, an efficient object passing between
4050 * cpus is necessary. This is provided by a shared array. The array
4051 * replaces Bonwick's magazine layer.
4052 * On uniprocessor, it's functionally equivalent (but less efficient)
4053 * to a larger limit. Thus disabled by default.
4054 */
4055 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004056 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058
4059#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004060 /*
4061 * With debugging enabled, large batchcount lead to excessively long
4062 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 */
4064 if (limit > 32)
4065 limit = 32;
4066#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004067 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 if (err)
4069 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004070 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004071 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072}
4073
Christoph Lameter1b552532006-03-22 00:09:07 -08004074/*
4075 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004076 * necessary. Note that the l3 listlock also protects the array_cache
4077 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004078 */
4079void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4080 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081{
4082 int tofree;
4083
Christoph Lameter1b552532006-03-22 00:09:07 -08004084 if (!ac || !ac->avail)
4085 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 if (ac->touched && !force) {
4087 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004088 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004089 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004090 if (ac->avail) {
4091 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4092 if (tofree > ac->avail)
4093 tofree = (ac->avail + 1) / 2;
4094 free_block(cachep, ac->entry, tofree, node);
4095 ac->avail -= tofree;
4096 memmove(ac->entry, &(ac->entry[tofree]),
4097 sizeof(void *) * ac->avail);
4098 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004099 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 }
4101}
4102
4103/**
4104 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004105 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 *
4107 * Called from workqueue/eventd every few seconds.
4108 * Purpose:
4109 * - clear the per-cpu caches for this CPU.
4110 * - return freeable pages to the main free memory pool.
4111 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004112 * If we cannot acquire the cache chain mutex then just give up - we'll try
4113 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004115static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004117 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004118 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004119 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004120 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004122 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004124 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004126 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 check_irq_on();
4128
Christoph Lameter35386e32006-03-22 00:09:05 -08004129 /*
4130 * We only take the l3 lock if absolutely necessary and we
4131 * have established with reasonable certainty that
4132 * we can do some work if the lock was obtained.
4133 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004134 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004135
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004136 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137
Christoph Lameteraab22072006-03-22 00:09:06 -08004138 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139
Christoph Lameter35386e32006-03-22 00:09:05 -08004140 /*
4141 * These are racy checks but it does not matter
4142 * if we skip one check or scan twice.
4143 */
Christoph Lametere498be72005-09-09 13:03:32 -07004144 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004145 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146
Christoph Lametere498be72005-09-09 13:03:32 -07004147 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148
Christoph Lameteraab22072006-03-22 00:09:06 -08004149 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150
Christoph Lametered11d9e2006-06-30 01:55:45 -07004151 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004152 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004153 else {
4154 int freed;
4155
4156 freed = drain_freelist(searchp, l3, (l3->free_limit +
4157 5 * searchp->num - 1) / (5 * searchp->num));
4158 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004160next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 cond_resched();
4162 }
4163 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004164 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004165 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004166out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004167 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004168 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169}
4170
Linus Torvalds158a9622008-01-02 13:04:48 -08004171#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172
Pekka Enberg85289f92006-01-08 01:00:36 -08004173static void print_slabinfo_header(struct seq_file *m)
4174{
4175 /*
4176 * Output format version, so at least we can change it
4177 * without _too_ many complaints.
4178 */
4179#if STATS
4180 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4181#else
4182 seq_puts(m, "slabinfo - version: 2.1\n");
4183#endif
4184 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4185 "<objperslab> <pagesperslab>");
4186 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4187 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4188#if STATS
4189 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004190 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004191 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4192#endif
4193 seq_putc(m, '\n');
4194}
4195
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196static void *s_start(struct seq_file *m, loff_t *pos)
4197{
4198 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004200 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004201 if (!n)
4202 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004203
4204 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205}
4206
4207static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4208{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004209 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210}
4211
4212static void s_stop(struct seq_file *m, void *p)
4213{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004214 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215}
4216
4217static int s_show(struct seq_file *m, void *p)
4218{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004219 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004220 struct slab *slabp;
4221 unsigned long active_objs;
4222 unsigned long num_objs;
4223 unsigned long active_slabs = 0;
4224 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004225 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004227 int node;
4228 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230 active_objs = 0;
4231 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004232 for_each_online_node(node) {
4233 l3 = cachep->nodelists[node];
4234 if (!l3)
4235 continue;
4236
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004237 check_irq_on();
4238 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004239
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004240 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004241 if (slabp->inuse != cachep->num && !error)
4242 error = "slabs_full accounting error";
4243 active_objs += cachep->num;
4244 active_slabs++;
4245 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004246 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004247 if (slabp->inuse == cachep->num && !error)
4248 error = "slabs_partial inuse accounting error";
4249 if (!slabp->inuse && !error)
4250 error = "slabs_partial/inuse accounting error";
4251 active_objs += slabp->inuse;
4252 active_slabs++;
4253 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004254 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004255 if (slabp->inuse && !error)
4256 error = "slabs_free/inuse accounting error";
4257 num_slabs++;
4258 }
4259 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004260 if (l3->shared)
4261 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004262
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004263 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004265 num_slabs += active_slabs;
4266 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004267 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004268 error = "free_objects accounting error";
4269
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004270 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004271 if (error)
4272 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4273
4274 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004275 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004276 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004278 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004279 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004280 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004282 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 unsigned long high = cachep->high_mark;
4284 unsigned long allocs = cachep->num_allocations;
4285 unsigned long grown = cachep->grown;
4286 unsigned long reaped = cachep->reaped;
4287 unsigned long errors = cachep->errors;
4288 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004290 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004291 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004292
Joe Perchese92dd4f2010-03-26 19:27:58 -07004293 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4294 "%4lu %4lu %4lu %4lu %4lu",
4295 allocs, high, grown,
4296 reaped, errors, max_freeable, node_allocs,
4297 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298 }
4299 /* cpu stats */
4300 {
4301 unsigned long allochit = atomic_read(&cachep->allochit);
4302 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4303 unsigned long freehit = atomic_read(&cachep->freehit);
4304 unsigned long freemiss = atomic_read(&cachep->freemiss);
4305
4306 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004307 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 }
4309#endif
4310 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311 return 0;
4312}
4313
4314/*
4315 * slabinfo_op - iterator that generates /proc/slabinfo
4316 *
4317 * Output layout:
4318 * cache-name
4319 * num-active-objs
4320 * total-objs
4321 * object size
4322 * num-active-slabs
4323 * total-slabs
4324 * num-pages-per-slab
4325 * + further values on SMP and with statistics enabled
4326 */
4327
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004328static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004329 .start = s_start,
4330 .next = s_next,
4331 .stop = s_stop,
4332 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333};
4334
4335#define MAX_SLABINFO_WRITE 128
4336/**
4337 * slabinfo_write - Tuning for the slab allocator
4338 * @file: unused
4339 * @buffer: user buffer
4340 * @count: data length
4341 * @ppos: unused
4342 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004343ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4344 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004345{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004346 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004347 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004348 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004349
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 if (count > MAX_SLABINFO_WRITE)
4351 return -EINVAL;
4352 if (copy_from_user(&kbuf, buffer, count))
4353 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004354 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355
4356 tmp = strchr(kbuf, ' ');
4357 if (!tmp)
4358 return -EINVAL;
4359 *tmp = '\0';
4360 tmp++;
4361 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4362 return -EINVAL;
4363
4364 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004365 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004366 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004367 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004369 if (limit < 1 || batchcount < 1 ||
4370 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004371 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004372 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004373 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004374 batchcount, shared,
4375 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004376 }
4377 break;
4378 }
4379 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004380 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381 if (res >= 0)
4382 res = count;
4383 return res;
4384}
Al Viro871751e2006-03-25 03:06:39 -08004385
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004386static int slabinfo_open(struct inode *inode, struct file *file)
4387{
4388 return seq_open(file, &slabinfo_op);
4389}
4390
4391static const struct file_operations proc_slabinfo_operations = {
4392 .open = slabinfo_open,
4393 .read = seq_read,
4394 .write = slabinfo_write,
4395 .llseek = seq_lseek,
4396 .release = seq_release,
4397};
4398
Al Viro871751e2006-03-25 03:06:39 -08004399#ifdef CONFIG_DEBUG_SLAB_LEAK
4400
4401static void *leaks_start(struct seq_file *m, loff_t *pos)
4402{
Al Viro871751e2006-03-25 03:06:39 -08004403 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004404 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004405}
4406
4407static inline int add_caller(unsigned long *n, unsigned long v)
4408{
4409 unsigned long *p;
4410 int l;
4411 if (!v)
4412 return 1;
4413 l = n[1];
4414 p = n + 2;
4415 while (l) {
4416 int i = l/2;
4417 unsigned long *q = p + 2 * i;
4418 if (*q == v) {
4419 q[1]++;
4420 return 1;
4421 }
4422 if (*q > v) {
4423 l = i;
4424 } else {
4425 p = q + 2;
4426 l -= i + 1;
4427 }
4428 }
4429 if (++n[1] == n[0])
4430 return 0;
4431 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4432 p[0] = v;
4433 p[1] = 1;
4434 return 1;
4435}
4436
4437static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4438{
4439 void *p;
4440 int i;
4441 if (n[0] == n[1])
4442 return;
4443 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4444 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4445 continue;
4446 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4447 return;
4448 }
4449}
4450
4451static void show_symbol(struct seq_file *m, unsigned long address)
4452{
4453#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004454 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004455 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004456
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004457 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004458 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004459 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004460 seq_printf(m, " [%s]", modname);
4461 return;
4462 }
4463#endif
4464 seq_printf(m, "%p", (void *)address);
4465}
4466
4467static int leaks_show(struct seq_file *m, void *p)
4468{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004469 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004470 struct slab *slabp;
4471 struct kmem_list3 *l3;
4472 const char *name;
4473 unsigned long *n = m->private;
4474 int node;
4475 int i;
4476
4477 if (!(cachep->flags & SLAB_STORE_USER))
4478 return 0;
4479 if (!(cachep->flags & SLAB_RED_ZONE))
4480 return 0;
4481
4482 /* OK, we can do it */
4483
4484 n[1] = 0;
4485
4486 for_each_online_node(node) {
4487 l3 = cachep->nodelists[node];
4488 if (!l3)
4489 continue;
4490
4491 check_irq_on();
4492 spin_lock_irq(&l3->list_lock);
4493
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004494 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004495 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004496 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004497 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004498 spin_unlock_irq(&l3->list_lock);
4499 }
4500 name = cachep->name;
4501 if (n[0] == n[1]) {
4502 /* Increase the buffer size */
4503 mutex_unlock(&cache_chain_mutex);
4504 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4505 if (!m->private) {
4506 /* Too bad, we are really out */
4507 m->private = n;
4508 mutex_lock(&cache_chain_mutex);
4509 return -ENOMEM;
4510 }
4511 *(unsigned long *)m->private = n[0] * 2;
4512 kfree(n);
4513 mutex_lock(&cache_chain_mutex);
4514 /* Now make sure this entry will be retried */
4515 m->count = m->size;
4516 return 0;
4517 }
4518 for (i = 0; i < n[1]; i++) {
4519 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4520 show_symbol(m, n[2*i+2]);
4521 seq_putc(m, '\n');
4522 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004523
Al Viro871751e2006-03-25 03:06:39 -08004524 return 0;
4525}
4526
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004527static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004528 .start = leaks_start,
4529 .next = s_next,
4530 .stop = s_stop,
4531 .show = leaks_show,
4532};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004533
4534static int slabstats_open(struct inode *inode, struct file *file)
4535{
4536 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4537 int ret = -ENOMEM;
4538 if (n) {
4539 ret = seq_open(file, &slabstats_op);
4540 if (!ret) {
4541 struct seq_file *m = file->private_data;
4542 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4543 m->private = n;
4544 n = NULL;
4545 }
4546 kfree(n);
4547 }
4548 return ret;
4549}
4550
4551static const struct file_operations proc_slabstats_operations = {
4552 .open = slabstats_open,
4553 .read = seq_read,
4554 .llseek = seq_lseek,
4555 .release = seq_release_private,
4556};
Al Viro871751e2006-03-25 03:06:39 -08004557#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004558
4559static int __init slab_proc_init(void)
4560{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004561 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004562#ifdef CONFIG_DEBUG_SLAB_LEAK
4563 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4564#endif
4565 return 0;
4566}
4567module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004568#endif
4569
Manfred Spraul00e145b2005-09-03 15:55:07 -07004570/**
4571 * ksize - get the actual amount of memory allocated for a given object
4572 * @objp: Pointer to the object
4573 *
4574 * kmalloc may internally round up allocations and return more memory
4575 * than requested. ksize() can be used to determine the actual amount of
4576 * memory allocated. The caller may use this additional memory, even though
4577 * a smaller amount of memory was initially specified with the kmalloc call.
4578 * The caller must guarantee that objp points to a valid object previously
4579 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4580 * must not be freed during the duration of the call.
4581 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004582size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004584 BUG_ON(!objp);
4585 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004586 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004587
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004588 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004590EXPORT_SYMBOL(ksize);