blob: b1e40dafbab3cc6326a6913acf17155cbcd8e7f0 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119#include <asm/cacheflush.h>
120#include <asm/tlbflush.h>
121#include <asm/page.h>
122
123/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 */
132
133#ifdef CONFIG_DEBUG_SLAB
134#define DEBUG 1
135#define STATS 1
136#define FORCED_DEBUG 1
137#else
138#define DEBUG 0
139#define STATS 0
140#define FORCED_DEBUG 0
141#endif
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/* Shouldn't this be in a header file somewhere? */
144#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400145#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147#ifndef ARCH_KMALLOC_FLAGS
148#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
149#endif
150
151/* Legal flag mask for kmem_cache_create(). */
152#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700153# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800155 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700156 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700158 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200159 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800161# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700162 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700164 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200165 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166#endif
167
168/*
169 * kmem_bufctl_t:
170 *
171 * Bufctl's are used for linking objs within a slab
172 * linked offsets.
173 *
174 * This implementation relies on "struct page" for locating the cache &
175 * slab an object belongs to.
176 * This allows the bufctl structure to be small (one int), but limits
177 * the number of objects a slab (not a cache) can contain when off-slab
178 * bufctls are used. The limit is the size of the largest general cache
179 * that does not use off-slab slabs.
180 * For 32bit archs with 4 kB pages, is this 56.
181 * This is not serious, as it is only for large objects, when it is unwise
182 * to have too many per slab.
183 * Note: This limit can be raised by introducing a general cache whose size
184 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 */
186
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700187typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
189#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800190#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
191#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193/*
194 * struct slab
195 *
196 * Manages the objs in a slab. Placed either at the beginning of mem allocated
197 * for a slab, or allocated from an general cache.
198 * Slabs are chained into three list: fully used, partial, fully free slabs.
199 */
200struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800201 struct list_head list;
202 unsigned long colouroff;
203 void *s_mem; /* including colour offset */
204 unsigned int inuse; /* num of objs active in slab */
205 kmem_bufctl_t free;
206 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207};
208
209/*
210 * struct slab_rcu
211 *
212 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
213 * arrange for kmem_freepages to be called via RCU. This is useful if
214 * we need to approach a kernel structure obliquely, from its address
215 * obtained without the usual locking. We can lock the structure to
216 * stabilize it and check it's still at the given address, only if we
217 * can be sure that the memory has not been meanwhile reused for some
218 * other kind of object (which our subsystem's lock might corrupt).
219 *
220 * rcu_read_lock before reading the address, then rcu_read_unlock after
221 * taking the spinlock within the structure expected at that address.
222 *
223 * We assume struct slab_rcu can overlay struct slab when destroying.
224 */
225struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800226 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800227 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800228 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229};
230
231/*
232 * struct array_cache
233 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 * Purpose:
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
238 *
239 * The limit is stored in the per-cpu structure to reduce the data cache
240 * footprint.
241 *
242 */
243struct array_cache {
244 unsigned int avail;
245 unsigned int limit;
246 unsigned int batchcount;
247 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700248 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700249 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
252 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800253 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254};
255
Andrew Mortona737b3e2006-03-22 00:08:11 -0800256/*
257 * bootstrap: The caches do not work without cpuarrays anymore, but the
258 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 */
260#define BOOT_CPUCACHE_ENTRIES 1
261struct arraycache_init {
262 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800263 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264};
265
266/*
Christoph Lametere498be72005-09-09 13:03:32 -0700267 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 */
269struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800270 struct list_head slabs_partial; /* partial list first, better asm code */
271 struct list_head slabs_full;
272 struct list_head slabs_free;
273 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800274 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800275 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800276 spinlock_t list_lock;
277 struct array_cache *shared; /* shared per node */
278 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800279 unsigned long next_reap; /* updated without locking */
280 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281};
282
Christoph Lametere498be72005-09-09 13:03:32 -0700283/*
284 * Need this for bootstrapping a per node allocator.
285 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200286#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700287struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
288#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200289#define SIZE_AC MAX_NUMNODES
290#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291
Christoph Lametered11d9e2006-06-30 01:55:45 -0700292static int drain_freelist(struct kmem_cache *cache,
293 struct kmem_list3 *l3, int tofree);
294static void free_block(struct kmem_cache *cachep, void **objpp, int len,
295 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300296static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000297static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700298
Christoph Lametere498be72005-09-09 13:03:32 -0700299/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800300 * This function must be completely optimized away if a constant is passed to
301 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700302 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700303static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700304{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800305 extern void __bad_size(void);
306
Christoph Lametere498be72005-09-09 13:03:32 -0700307 if (__builtin_constant_p(size)) {
308 int i = 0;
309
310#define CACHE(x) \
311 if (size <=x) \
312 return i; \
313 else \
314 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800315#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700316#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800317 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700318 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800319 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700320 return 0;
321}
322
Ingo Molnare0a42722006-06-23 02:03:46 -0700323static int slab_early_init = 1;
324
Christoph Lametere498be72005-09-09 13:03:32 -0700325#define INDEX_AC index_of(sizeof(struct arraycache_init))
326#define INDEX_L3 index_of(sizeof(struct kmem_list3))
327
Pekka Enberg5295a742006-02-01 03:05:48 -0800328static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700329{
330 INIT_LIST_HEAD(&parent->slabs_full);
331 INIT_LIST_HEAD(&parent->slabs_partial);
332 INIT_LIST_HEAD(&parent->slabs_free);
333 parent->shared = NULL;
334 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800335 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700336 spin_lock_init(&parent->list_lock);
337 parent->free_objects = 0;
338 parent->free_touched = 0;
339}
340
Andrew Mortona737b3e2006-03-22 00:08:11 -0800341#define MAKE_LIST(cachep, listp, slab, nodeid) \
342 do { \
343 INIT_LIST_HEAD(listp); \
344 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700345 } while (0)
346
Andrew Mortona737b3e2006-03-22 00:08:11 -0800347#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
348 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700349 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
350 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
351 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
352 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354#define CFLGS_OFF_SLAB (0x80000000UL)
355#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
356
357#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800358/*
359 * Optimization question: fewer reaps means less probability for unnessary
360 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100362 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 * which could lock up otherwise freeable slabs.
364 */
365#define REAPTIMEOUT_CPUC (2*HZ)
366#define REAPTIMEOUT_LIST3 (4*HZ)
367
368#if STATS
369#define STATS_INC_ACTIVE(x) ((x)->num_active++)
370#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
371#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
372#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700373#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800374#define STATS_SET_HIGH(x) \
375 do { \
376 if ((x)->num_active > (x)->high_mark) \
377 (x)->high_mark = (x)->num_active; \
378 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379#define STATS_INC_ERR(x) ((x)->errors++)
380#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700381#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700382#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800383#define STATS_SET_FREEABLE(x, i) \
384 do { \
385 if ((x)->max_freeable < i) \
386 (x)->max_freeable = i; \
387 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
389#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
390#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
391#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
392#else
393#define STATS_INC_ACTIVE(x) do { } while (0)
394#define STATS_DEC_ACTIVE(x) do { } while (0)
395#define STATS_INC_ALLOCED(x) do { } while (0)
396#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700397#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398#define STATS_SET_HIGH(x) do { } while (0)
399#define STATS_INC_ERR(x) do { } while (0)
400#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700401#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700402#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800403#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404#define STATS_INC_ALLOCHIT(x) do { } while (0)
405#define STATS_INC_ALLOCMISS(x) do { } while (0)
406#define STATS_INC_FREEHIT(x) do { } while (0)
407#define STATS_INC_FREEMISS(x) do { } while (0)
408#endif
409
410#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Andrew Mortona737b3e2006-03-22 00:08:11 -0800412/*
413 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800415 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 * the end of an object is aligned with the end of the real
417 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800418 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800420 * cachep->obj_offset: The real object.
421 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800422 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
423 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800425static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800427 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428}
429
Pekka Enberg343e0d72006-02-01 03:05:50 -0800430static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800432 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433}
434
David Woodhouseb46b8f12007-05-08 00:22:59 -0700435static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436{
437 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700438 return (unsigned long long*) (objp + obj_offset(cachep) -
439 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440}
441
David Woodhouseb46b8f12007-05-08 00:22:59 -0700442static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443{
444 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
445 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700446 return (unsigned long long *)(objp + cachep->buffer_size -
447 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400448 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700449 return (unsigned long long *) (objp + cachep->buffer_size -
450 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451}
452
Pekka Enberg343e0d72006-02-01 03:05:50 -0800453static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454{
455 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800456 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
459#else
460
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800461#define obj_offset(x) 0
462#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700463#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
466
467#endif
468
Li Zefan0f24f122009-12-11 15:45:30 +0800469#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300470size_t slab_buffer_size(struct kmem_cache *cachep)
471{
472 return cachep->buffer_size;
473}
474EXPORT_SYMBOL(slab_buffer_size);
475#endif
476
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 * Do not go above this order unless 0 objects fit into the slab.
479 */
480#define BREAK_GFP_ORDER_HI 1
481#define BREAK_GFP_ORDER_LO 0
482static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
483
Andrew Mortona737b3e2006-03-22 00:08:11 -0800484/*
485 * Functions for storing/retrieving the cachep and or slab from the page
486 * allocator. These are used to find the slab an obj belongs to. With kfree(),
487 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800489static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
490{
491 page->lru.next = (struct list_head *)cache;
492}
493
494static inline struct kmem_cache *page_get_cache(struct page *page)
495{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700496 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700497 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800498 return (struct kmem_cache *)page->lru.next;
499}
500
501static inline void page_set_slab(struct page *page, struct slab *slab)
502{
503 page->lru.prev = (struct list_head *)slab;
504}
505
506static inline struct slab *page_get_slab(struct page *page)
507{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700508 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800509 return (struct slab *)page->lru.prev;
510}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800512static inline struct kmem_cache *virt_to_cache(const void *obj)
513{
Christoph Lameterb49af682007-05-06 14:49:41 -0700514 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800515 return page_get_cache(page);
516}
517
518static inline struct slab *virt_to_slab(const void *obj)
519{
Christoph Lameterb49af682007-05-06 14:49:41 -0700520 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800521 return page_get_slab(page);
522}
523
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800524static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
525 unsigned int idx)
526{
527 return slab->s_mem + cache->buffer_size * idx;
528}
529
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800530/*
531 * We want to avoid an expensive divide : (offset / cache->buffer_size)
532 * Using the fact that buffer_size is a constant for a particular cache,
533 * we can replace (offset / cache->buffer_size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
535 */
536static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800538{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800539 u32 offset = (obj - slab->s_mem);
540 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800541}
542
Andrew Mortona737b3e2006-03-22 00:08:11 -0800543/*
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
545 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546struct cache_sizes malloc_sizes[] = {
547#define CACHE(x) { .cs_size = (x) },
548#include <linux/kmalloc_sizes.h>
549 CACHE(ULONG_MAX)
550#undef CACHE
551};
552EXPORT_SYMBOL(malloc_sizes);
553
554/* Must match cache_sizes above. Out of line to keep cache footprint low. */
555struct cache_names {
556 char *name;
557 char *name_dma;
558};
559
560static struct cache_names __initdata cache_names[] = {
561#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800563 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564#undef CACHE
565};
566
567static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800568 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800570 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
572/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800573static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800574 .batchcount = 1,
575 .limit = BOOT_CPUCACHE_ENTRIES,
576 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800577 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800578 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579};
580
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700581#define BAD_ALIEN_MAGIC 0x01020304ul
582
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 * chicken and egg problem: delay the per-cpu array allocation
585 * until the general caches are up.
586 */
587static enum {
588 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700589 PARTIAL_AC,
590 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300591 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 FULL
593} g_cpucache_up;
594
Mike Kravetz39d24e62006-05-15 09:44:13 -0700595/*
596 * used by boot code to determine if it can use slab based allocator
597 */
598int slab_is_available(void)
599{
Pekka Enberg8429db52009-06-12 15:58:59 +0300600 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700601}
602
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200603#ifdef CONFIG_LOCKDEP
604
605/*
606 * Slab sometimes uses the kmalloc slabs to store the slab headers
607 * for other slabs "off slab".
608 * The locking for this is tricky in that it nests within the locks
609 * of all other slabs in a few places; to deal with this special
610 * locking we put on-slab caches into a separate lock-class.
611 *
612 * We set lock class for alien array caches which are up during init.
613 * The lock annotation will be lost if all cpus of a node goes down and
614 * then comes back up during hotplug
615 */
616static struct lock_class_key on_slab_l3_key;
617static struct lock_class_key on_slab_alc_key;
618
619static void init_node_lock_keys(int q)
620{
621 struct cache_sizes *s = malloc_sizes;
622
623 if (g_cpucache_up != FULL)
624 return;
625
626 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
627 struct array_cache **alc;
628 struct kmem_list3 *l3;
629 int r;
630
631 l3 = s->cs_cachep->nodelists[q];
632 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200633 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200634 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
635 alc = l3->alien;
636 /*
637 * FIXME: This check for BAD_ALIEN_MAGIC
638 * should go away when common slab code is taught to
639 * work even without alien caches.
640 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
641 * for alloc_alien_cache,
642 */
643 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
Pekka Enberg00afa752009-12-27 14:33:14 +0200644 continue;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200645 for_each_node(r) {
646 if (alc[r])
647 lockdep_set_class(&alc[r]->lock,
648 &on_slab_alc_key);
649 }
650 }
651}
652
653static inline void init_lock_keys(void)
654{
655 int node;
656
657 for_each_node(node)
658 init_node_lock_keys(node);
659}
660#else
661static void init_node_lock_keys(int q)
662{
663}
664
665static inline void init_lock_keys(void)
666{
667}
668#endif
669
670/*
671 * Guard access to the cache-chain.
672 */
673static DEFINE_MUTEX(cache_chain_mutex);
674static struct list_head cache_chain;
675
Tejun Heo1871e522009-10-29 22:34:13 +0900676static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677
Pekka Enberg343e0d72006-02-01 03:05:50 -0800678static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679{
680 return cachep->array[smp_processor_id()];
681}
682
Andrew Mortona737b3e2006-03-22 00:08:11 -0800683static inline struct kmem_cache *__find_general_cachep(size_t size,
684 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685{
686 struct cache_sizes *csizep = malloc_sizes;
687
688#if DEBUG
689 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800690 * kmem_cache_create(), or __kmalloc(), before
691 * the generic caches are initialized.
692 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700693 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700695 if (!size)
696 return ZERO_SIZE_PTR;
697
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 while (size > csizep->cs_size)
699 csizep++;
700
701 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700702 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 * has cs_{dma,}cachep==NULL. Thus no special case
704 * for large kmalloc calls required.
705 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800706#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 if (unlikely(gfpflags & GFP_DMA))
708 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800709#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 return csizep->cs_cachep;
711}
712
Adrian Bunkb2213852006-09-25 23:31:02 -0700713static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700714{
715 return __find_general_cachep(size, gfpflags);
716}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700717
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800718static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800720 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
721}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
Andrew Mortona737b3e2006-03-22 00:08:11 -0800723/*
724 * Calculate the number of objects and left-over bytes for a given buffer size.
725 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800726static void cache_estimate(unsigned long gfporder, size_t buffer_size,
727 size_t align, int flags, size_t *left_over,
728 unsigned int *num)
729{
730 int nr_objs;
731 size_t mgmt_size;
732 size_t slab_size = PAGE_SIZE << gfporder;
733
734 /*
735 * The slab management structure can be either off the slab or
736 * on it. For the latter case, the memory allocated for a
737 * slab is used for:
738 *
739 * - The struct slab
740 * - One kmem_bufctl_t for each object
741 * - Padding to respect alignment of @align
742 * - @buffer_size bytes for each object
743 *
744 * If the slab management structure is off the slab, then the
745 * alignment will already be calculated into the size. Because
746 * the slabs are all pages aligned, the objects will be at the
747 * correct alignment when allocated.
748 */
749 if (flags & CFLGS_OFF_SLAB) {
750 mgmt_size = 0;
751 nr_objs = slab_size / buffer_size;
752
753 if (nr_objs > SLAB_LIMIT)
754 nr_objs = SLAB_LIMIT;
755 } else {
756 /*
757 * Ignore padding for the initial guess. The padding
758 * is at most @align-1 bytes, and @buffer_size is at
759 * least @align. In the worst case, this result will
760 * be one greater than the number of objects that fit
761 * into the memory allocation when taking the padding
762 * into account.
763 */
764 nr_objs = (slab_size - sizeof(struct slab)) /
765 (buffer_size + sizeof(kmem_bufctl_t));
766
767 /*
768 * This calculated number will be either the right
769 * amount, or one greater than what we want.
770 */
771 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
772 > slab_size)
773 nr_objs--;
774
775 if (nr_objs > SLAB_LIMIT)
776 nr_objs = SLAB_LIMIT;
777
778 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800780 *num = nr_objs;
781 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782}
783
Harvey Harrisond40cee22008-04-30 00:55:07 -0700784#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785
Andrew Mortona737b3e2006-03-22 00:08:11 -0800786static void __slab_error(const char *function, struct kmem_cache *cachep,
787 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788{
789 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800790 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 dump_stack();
792}
793
Paul Menage3395ee02006-12-06 20:32:16 -0800794/*
795 * By default on NUMA we use alien caches to stage the freeing of
796 * objects allocated from other nodes. This causes massive memory
797 * inefficiencies when using fake NUMA setup to split memory into a
798 * large number of small nodes, so it can be disabled on the command
799 * line
800 */
801
802static int use_alien_caches __read_mostly = 1;
803static int __init noaliencache_setup(char *s)
804{
805 use_alien_caches = 0;
806 return 1;
807}
808__setup("noaliencache", noaliencache_setup);
809
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800810#ifdef CONFIG_NUMA
811/*
812 * Special reaping functions for NUMA systems called from cache_reap().
813 * These take care of doing round robin flushing of alien caches (containing
814 * objects freed on different nodes from which they were allocated) and the
815 * flushing of remote pcps by calling drain_node_pages.
816 */
Tejun Heo1871e522009-10-29 22:34:13 +0900817static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800818
819static void init_reap_node(int cpu)
820{
821 int node;
822
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700823 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800824 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800825 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800826
Tejun Heo1871e522009-10-29 22:34:13 +0900827 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800828}
829
830static void next_reap_node(void)
831{
Tejun Heo1871e522009-10-29 22:34:13 +0900832 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800833
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800834 node = next_node(node, node_online_map);
835 if (unlikely(node >= MAX_NUMNODES))
836 node = first_node(node_online_map);
Tejun Heo1871e522009-10-29 22:34:13 +0900837 __get_cpu_var(slab_reap_node) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800838}
839
840#else
841#define init_reap_node(cpu) do { } while (0)
842#define next_reap_node(void) do { } while (0)
843#endif
844
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845/*
846 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
847 * via the workqueue/eventd.
848 * Add the CPU number into the expiration time to minimize the possibility of
849 * the CPUs getting into lockstep and contending for the global cache chain
850 * lock.
851 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700852static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853{
Tejun Heo1871e522009-10-29 22:34:13 +0900854 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
856 /*
857 * When this gets called from do_initcalls via cpucache_init(),
858 * init_workqueues() has already run, so keventd will be setup
859 * at that time.
860 */
David Howells52bad642006-11-22 14:54:01 +0000861 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800862 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700863 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800864 schedule_delayed_work_on(cpu, reap_work,
865 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 }
867}
868
Christoph Lametere498be72005-09-09 13:03:32 -0700869static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300870 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800872 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 struct array_cache *nc = NULL;
874
Pekka Enberg83b519e2009-06-10 19:40:04 +0300875 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100876 /*
877 * The array_cache structures contain pointers to free object.
878 * However, when such objects are allocated or transfered to another
879 * cache the pointers are not cleared and they could be counted as
880 * valid references during a kmemleak scan. Therefore, kmemleak must
881 * not scan such objects.
882 */
883 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 if (nc) {
885 nc->avail = 0;
886 nc->limit = entries;
887 nc->batchcount = batchcount;
888 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700889 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 }
891 return nc;
892}
893
Christoph Lameter3ded1752006-03-25 03:06:44 -0800894/*
895 * Transfer objects in one arraycache to another.
896 * Locking must be handled by the caller.
897 *
898 * Return the number of entries transferred.
899 */
900static int transfer_objects(struct array_cache *to,
901 struct array_cache *from, unsigned int max)
902{
903 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700904 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800905
906 if (!nr)
907 return 0;
908
909 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
910 sizeof(void *) *nr);
911
912 from->avail -= nr;
913 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800914 return nr;
915}
916
Christoph Lameter765c4502006-09-27 01:50:08 -0700917#ifndef CONFIG_NUMA
918
919#define drain_alien_cache(cachep, alien) do { } while (0)
920#define reap_alien(cachep, l3) do { } while (0)
921
Pekka Enberg83b519e2009-06-10 19:40:04 +0300922static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700923{
924 return (struct array_cache **)BAD_ALIEN_MAGIC;
925}
926
927static inline void free_alien_cache(struct array_cache **ac_ptr)
928{
929}
930
931static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
932{
933 return 0;
934}
935
936static inline void *alternate_node_alloc(struct kmem_cache *cachep,
937 gfp_t flags)
938{
939 return NULL;
940}
941
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800942static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700943 gfp_t flags, int nodeid)
944{
945 return NULL;
946}
947
948#else /* CONFIG_NUMA */
949
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800950static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800951static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800952
Pekka Enberg83b519e2009-06-10 19:40:04 +0300953static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700954{
955 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800956 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700957 int i;
958
959 if (limit > 1)
960 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800961 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700962 if (ac_ptr) {
963 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800964 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700965 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300966 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700967 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800968 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700969 kfree(ac_ptr[i]);
970 kfree(ac_ptr);
971 return NULL;
972 }
973 }
974 }
975 return ac_ptr;
976}
977
Pekka Enberg5295a742006-02-01 03:05:48 -0800978static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700979{
980 int i;
981
982 if (!ac_ptr)
983 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700984 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800985 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700986 kfree(ac_ptr);
987}
988
Pekka Enberg343e0d72006-02-01 03:05:50 -0800989static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800990 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700991{
992 struct kmem_list3 *rl3 = cachep->nodelists[node];
993
994 if (ac->avail) {
995 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800996 /*
997 * Stuff objects into the remote nodes shared array first.
998 * That way we could avoid the overhead of putting the objects
999 * into the free lists and getting them back later.
1000 */
shin, jacob693f7d32006-04-28 10:54:37 -05001001 if (rl3->shared)
1002 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001003
Christoph Lameterff694162005-09-22 21:44:02 -07001004 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001005 ac->avail = 0;
1006 spin_unlock(&rl3->list_lock);
1007 }
1008}
1009
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001010/*
1011 * Called from cache_reap() to regularly drain alien caches round robin.
1012 */
1013static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1014{
Tejun Heo1871e522009-10-29 22:34:13 +09001015 int node = __get_cpu_var(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001016
1017 if (l3->alien) {
1018 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001019
1020 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001021 __drain_alien_cache(cachep, ac, node);
1022 spin_unlock_irq(&ac->lock);
1023 }
1024 }
1025}
1026
Andrew Mortona737b3e2006-03-22 00:08:11 -08001027static void drain_alien_cache(struct kmem_cache *cachep,
1028 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001029{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001030 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001031 struct array_cache *ac;
1032 unsigned long flags;
1033
1034 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001035 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001036 if (ac) {
1037 spin_lock_irqsave(&ac->lock, flags);
1038 __drain_alien_cache(cachep, ac, i);
1039 spin_unlock_irqrestore(&ac->lock, flags);
1040 }
1041 }
1042}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001043
Ingo Molnar873623d2006-07-13 14:44:38 +02001044static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001045{
1046 struct slab *slabp = virt_to_slab(objp);
1047 int nodeid = slabp->nodeid;
1048 struct kmem_list3 *l3;
1049 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001050 int node;
1051
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001052 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001053
1054 /*
1055 * Make sure we are not freeing a object from another node to the array
1056 * cache on this cpu.
1057 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001058 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001059 return 0;
1060
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001061 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001062 STATS_INC_NODEFREES(cachep);
1063 if (l3->alien && l3->alien[nodeid]) {
1064 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001065 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001066 if (unlikely(alien->avail == alien->limit)) {
1067 STATS_INC_ACOVERFLOW(cachep);
1068 __drain_alien_cache(cachep, alien, nodeid);
1069 }
1070 alien->entry[alien->avail++] = objp;
1071 spin_unlock(&alien->lock);
1072 } else {
1073 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1074 free_block(cachep, &objp, 1, nodeid);
1075 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1076 }
1077 return 1;
1078}
Christoph Lametere498be72005-09-09 13:03:32 -07001079#endif
1080
David Rientjes8f9f8d92010-03-27 19:40:47 -07001081/*
1082 * Allocates and initializes nodelists for a node on each slab cache, used for
1083 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1084 * will be allocated off-node since memory is not yet online for the new node.
1085 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1086 * already in use.
1087 *
1088 * Must hold cache_chain_mutex.
1089 */
1090static int init_cache_nodelists_node(int node)
1091{
1092 struct kmem_cache *cachep;
1093 struct kmem_list3 *l3;
1094 const int memsize = sizeof(struct kmem_list3);
1095
1096 list_for_each_entry(cachep, &cache_chain, next) {
1097 /*
1098 * Set up the size64 kmemlist for cpu before we can
1099 * begin anything. Make sure some other cpu on this
1100 * node has not already allocated this
1101 */
1102 if (!cachep->nodelists[node]) {
1103 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1104 if (!l3)
1105 return -ENOMEM;
1106 kmem_list3_init(l3);
1107 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1108 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1109
1110 /*
1111 * The l3s don't come and go as CPUs come and
1112 * go. cache_chain_mutex is sufficient
1113 * protection here.
1114 */
1115 cachep->nodelists[node] = l3;
1116 }
1117
1118 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1119 cachep->nodelists[node]->free_limit =
1120 (1 + nr_cpus_node(node)) *
1121 cachep->batchcount + cachep->num;
1122 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1123 }
1124 return 0;
1125}
1126
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001127static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001129 struct kmem_cache *cachep;
1130 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001131 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301132 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001133
1134 list_for_each_entry(cachep, &cache_chain, next) {
1135 struct array_cache *nc;
1136 struct array_cache *shared;
1137 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001138
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001139 /* cpu is dead; no one can alloc from it. */
1140 nc = cachep->array[cpu];
1141 cachep->array[cpu] = NULL;
1142 l3 = cachep->nodelists[node];
1143
1144 if (!l3)
1145 goto free_array_cache;
1146
1147 spin_lock_irq(&l3->list_lock);
1148
1149 /* Free limit for this kmem_list3 */
1150 l3->free_limit -= cachep->batchcount;
1151 if (nc)
1152 free_block(cachep, nc->entry, nc->avail, node);
1153
Rusty Russell58463c12009-12-17 11:43:12 -06001154 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001155 spin_unlock_irq(&l3->list_lock);
1156 goto free_array_cache;
1157 }
1158
1159 shared = l3->shared;
1160 if (shared) {
1161 free_block(cachep, shared->entry,
1162 shared->avail, node);
1163 l3->shared = NULL;
1164 }
1165
1166 alien = l3->alien;
1167 l3->alien = NULL;
1168
1169 spin_unlock_irq(&l3->list_lock);
1170
1171 kfree(shared);
1172 if (alien) {
1173 drain_alien_cache(cachep, alien);
1174 free_alien_cache(alien);
1175 }
1176free_array_cache:
1177 kfree(nc);
1178 }
1179 /*
1180 * In the previous loop, all the objects were freed to
1181 * the respective cache's slabs, now we can go ahead and
1182 * shrink each nodelist to its limit.
1183 */
1184 list_for_each_entry(cachep, &cache_chain, next) {
1185 l3 = cachep->nodelists[node];
1186 if (!l3)
1187 continue;
1188 drain_freelist(cachep, l3, l3->free_objects);
1189 }
1190}
1191
1192static int __cpuinit cpuup_prepare(long cpu)
1193{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001194 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001195 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001196 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001197 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 /*
1200 * We need to do this right in the beginning since
1201 * alloc_arraycache's are going to use this list.
1202 * kmalloc_node allows us to add the slab to the right
1203 * kmem_list3 and not this cpu's kmem_list3
1204 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001205 err = init_cache_nodelists_node(node);
1206 if (err < 0)
1207 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001208
1209 /*
1210 * Now we can go ahead with allocating the shared arrays and
1211 * array caches
1212 */
1213 list_for_each_entry(cachep, &cache_chain, next) {
1214 struct array_cache *nc;
1215 struct array_cache *shared = NULL;
1216 struct array_cache **alien = NULL;
1217
1218 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001219 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001220 if (!nc)
1221 goto bad;
1222 if (cachep->shared) {
1223 shared = alloc_arraycache(node,
1224 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001225 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001226 if (!shared) {
1227 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001228 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001229 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001230 }
1231 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001232 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001233 if (!alien) {
1234 kfree(shared);
1235 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001236 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001237 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001238 }
1239 cachep->array[cpu] = nc;
1240 l3 = cachep->nodelists[node];
1241 BUG_ON(!l3);
1242
1243 spin_lock_irq(&l3->list_lock);
1244 if (!l3->shared) {
1245 /*
1246 * We are serialised from CPU_DEAD or
1247 * CPU_UP_CANCELLED by the cpucontrol lock
1248 */
1249 l3->shared = shared;
1250 shared = NULL;
1251 }
1252#ifdef CONFIG_NUMA
1253 if (!l3->alien) {
1254 l3->alien = alien;
1255 alien = NULL;
1256 }
1257#endif
1258 spin_unlock_irq(&l3->list_lock);
1259 kfree(shared);
1260 free_alien_cache(alien);
1261 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001262 init_node_lock_keys(node);
1263
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001264 return 0;
1265bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001266 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001267 return -ENOMEM;
1268}
1269
1270static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1271 unsigned long action, void *hcpu)
1272{
1273 long cpu = (long)hcpu;
1274 int err = 0;
1275
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001277 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001279 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001281 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 break;
1283 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001284 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 start_cpu_timer(cpu);
1286 break;
1287#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001288 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001289 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001290 /*
1291 * Shutdown cache reaper. Note that the cache_chain_mutex is
1292 * held so that if cache_reap() is invoked it cannot do
1293 * anything expensive but will only modify reap_work
1294 * and reschedule the timer.
1295 */
Tejun Heo1871e522009-10-29 22:34:13 +09001296 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001297 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001298 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001299 break;
1300 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001301 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001302 start_cpu_timer(cpu);
1303 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001305 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001306 /*
1307 * Even if all the cpus of a node are down, we don't free the
1308 * kmem_list3 of any cache. This to avoid a race between
1309 * cpu_down, and a kmalloc allocation from another cpu for
1310 * memory from the node of the cpu going down. The list3
1311 * structure is usually allocated from kmem_cache_create() and
1312 * gets destroyed at kmem_cache_destroy().
1313 */
Simon Arlott183ff222007-10-20 01:27:18 +02001314 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001315#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001317 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001318 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001319 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001320 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001323 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324}
1325
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001326static struct notifier_block __cpuinitdata cpucache_notifier = {
1327 &cpuup_callback, NULL, 0
1328};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329
David Rientjes8f9f8d92010-03-27 19:40:47 -07001330#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1331/*
1332 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1333 * Returns -EBUSY if all objects cannot be drained so that the node is not
1334 * removed.
1335 *
1336 * Must hold cache_chain_mutex.
1337 */
1338static int __meminit drain_cache_nodelists_node(int node)
1339{
1340 struct kmem_cache *cachep;
1341 int ret = 0;
1342
1343 list_for_each_entry(cachep, &cache_chain, next) {
1344 struct kmem_list3 *l3;
1345
1346 l3 = cachep->nodelists[node];
1347 if (!l3)
1348 continue;
1349
1350 drain_freelist(cachep, l3, l3->free_objects);
1351
1352 if (!list_empty(&l3->slabs_full) ||
1353 !list_empty(&l3->slabs_partial)) {
1354 ret = -EBUSY;
1355 break;
1356 }
1357 }
1358 return ret;
1359}
1360
1361static int __meminit slab_memory_callback(struct notifier_block *self,
1362 unsigned long action, void *arg)
1363{
1364 struct memory_notify *mnb = arg;
1365 int ret = 0;
1366 int nid;
1367
1368 nid = mnb->status_change_nid;
1369 if (nid < 0)
1370 goto out;
1371
1372 switch (action) {
1373 case MEM_GOING_ONLINE:
1374 mutex_lock(&cache_chain_mutex);
1375 ret = init_cache_nodelists_node(nid);
1376 mutex_unlock(&cache_chain_mutex);
1377 break;
1378 case MEM_GOING_OFFLINE:
1379 mutex_lock(&cache_chain_mutex);
1380 ret = drain_cache_nodelists_node(nid);
1381 mutex_unlock(&cache_chain_mutex);
1382 break;
1383 case MEM_ONLINE:
1384 case MEM_OFFLINE:
1385 case MEM_CANCEL_ONLINE:
1386 case MEM_CANCEL_OFFLINE:
1387 break;
1388 }
1389out:
1390 return ret ? notifier_from_errno(ret) : NOTIFY_OK;
1391}
1392#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1393
Christoph Lametere498be72005-09-09 13:03:32 -07001394/*
1395 * swap the static kmem_list3 with kmalloced memory
1396 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001397static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1398 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001399{
1400 struct kmem_list3 *ptr;
1401
Pekka Enberg83b519e2009-06-10 19:40:04 +03001402 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001403 BUG_ON(!ptr);
1404
Christoph Lametere498be72005-09-09 13:03:32 -07001405 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001406 /*
1407 * Do not assume that spinlocks can be initialized via memcpy:
1408 */
1409 spin_lock_init(&ptr->list_lock);
1410
Christoph Lametere498be72005-09-09 13:03:32 -07001411 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1412 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001413}
1414
Andrew Mortona737b3e2006-03-22 00:08:11 -08001415/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001416 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1417 * size of kmem_list3.
1418 */
1419static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1420{
1421 int node;
1422
1423 for_each_online_node(node) {
1424 cachep->nodelists[node] = &initkmem_list3[index + node];
1425 cachep->nodelists[node]->next_reap = jiffies +
1426 REAPTIMEOUT_LIST3 +
1427 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1428 }
1429}
1430
1431/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001432 * Initialisation. Called after the page allocator have been initialised and
1433 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 */
1435void __init kmem_cache_init(void)
1436{
1437 size_t left_over;
1438 struct cache_sizes *sizes;
1439 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001440 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001441 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001442 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001443
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001444 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001445 use_alien_caches = 0;
1446
Christoph Lametere498be72005-09-09 13:03:32 -07001447 for (i = 0; i < NUM_INIT_LISTS; i++) {
1448 kmem_list3_init(&initkmem_list3[i]);
1449 if (i < MAX_NUMNODES)
1450 cache_cache.nodelists[i] = NULL;
1451 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001452 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
1454 /*
1455 * Fragmentation resistance on low memory - only use bigger
1456 * page orders on machines with more than 32MB of memory.
1457 */
Jan Beulich44813742009-09-21 17:03:05 -07001458 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1460
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 /* Bootstrap is tricky, because several objects are allocated
1462 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001463 * 1) initialize the cache_cache cache: it contains the struct
1464 * kmem_cache structures of all caches, except cache_cache itself:
1465 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001466 * Initially an __init data area is used for the head array and the
1467 * kmem_list3 structures, it's replaced with a kmalloc allocated
1468 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001470 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001471 * An __init data area is used for the head array.
1472 * 3) Create the remaining kmalloc caches, with minimally sized
1473 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 * 4) Replace the __init data head arrays for cache_cache and the first
1475 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001476 * 5) Replace the __init data for kmem_list3 for cache_cache and
1477 * the other cache's with kmalloc allocated memory.
1478 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 */
1480
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001481 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001482
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 INIT_LIST_HEAD(&cache_chain);
1485 list_add(&cache_cache.next, &cache_chain);
1486 cache_cache.colour_off = cache_line_size();
1487 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001488 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
Eric Dumazet8da34302007-05-06 14:49:29 -07001490 /*
1491 * struct kmem_cache size depends on nr_node_ids, which
1492 * can be less than MAX_NUMNODES.
1493 */
1494 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1495 nr_node_ids * sizeof(struct kmem_list3 *);
1496#if DEBUG
1497 cache_cache.obj_size = cache_cache.buffer_size;
1498#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001499 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1500 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001501 cache_cache.reciprocal_buffer_size =
1502 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503
Jack Steiner07ed76b2006-03-07 21:55:46 -08001504 for (order = 0; order < MAX_ORDER; order++) {
1505 cache_estimate(order, cache_cache.buffer_size,
1506 cache_line_size(), 0, &left_over, &cache_cache.num);
1507 if (cache_cache.num)
1508 break;
1509 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001510 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001511 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001512 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001513 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1514 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515
1516 /* 2+3) create the kmalloc caches */
1517 sizes = malloc_sizes;
1518 names = cache_names;
1519
Andrew Mortona737b3e2006-03-22 00:08:11 -08001520 /*
1521 * Initialize the caches that provide memory for the array cache and the
1522 * kmem_list3 structures first. Without this, further allocations will
1523 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001524 */
1525
1526 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001527 sizes[INDEX_AC].cs_size,
1528 ARCH_KMALLOC_MINALIGN,
1529 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001530 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001531
Andrew Mortona737b3e2006-03-22 00:08:11 -08001532 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001533 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001534 kmem_cache_create(names[INDEX_L3].name,
1535 sizes[INDEX_L3].cs_size,
1536 ARCH_KMALLOC_MINALIGN,
1537 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001538 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001539 }
Christoph Lametere498be72005-09-09 13:03:32 -07001540
Ingo Molnare0a42722006-06-23 02:03:46 -07001541 slab_early_init = 0;
1542
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001544 /*
1545 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 * This should be particularly beneficial on SMP boxes, as it
1547 * eliminates "false sharing".
1548 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001549 * allow tighter packing of the smaller caches.
1550 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001551 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001552 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001553 sizes->cs_size,
1554 ARCH_KMALLOC_MINALIGN,
1555 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001556 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001558#ifdef CONFIG_ZONE_DMA
1559 sizes->cs_dmacachep = kmem_cache_create(
1560 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001561 sizes->cs_size,
1562 ARCH_KMALLOC_MINALIGN,
1563 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1564 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001565 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001566#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 sizes++;
1568 names++;
1569 }
1570 /* 4) Replace the bootstrap head arrays */
1571 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001572 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001573
Pekka Enberg83b519e2009-06-10 19:40:04 +03001574 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001575
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001576 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1577 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001578 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001579 /*
1580 * Do not assume that spinlocks can be initialized via memcpy:
1581 */
1582 spin_lock_init(&ptr->lock);
1583
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001585
Pekka Enberg83b519e2009-06-10 19:40:04 +03001586 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001587
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001588 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001590 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001591 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001592 /*
1593 * Do not assume that spinlocks can be initialized via memcpy:
1594 */
1595 spin_lock_init(&ptr->lock);
1596
Christoph Lametere498be72005-09-09 13:03:32 -07001597 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001598 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 }
Christoph Lametere498be72005-09-09 13:03:32 -07001600 /* 5) Replace the bootstrap kmem_list3's */
1601 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001602 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603
Mel Gorman9c09a952008-01-24 05:49:54 -08001604 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001605 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001606
Christoph Lametere498be72005-09-09 13:03:32 -07001607 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001608 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001609
1610 if (INDEX_AC != INDEX_L3) {
1611 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001612 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001613 }
1614 }
1615 }
1616
Pekka Enberg8429db52009-06-12 15:58:59 +03001617 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001618}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001619
Pekka Enberg8429db52009-06-12 15:58:59 +03001620void __init kmem_cache_init_late(void)
1621{
1622 struct kmem_cache *cachep;
1623
Pekka Enberg8429db52009-06-12 15:58:59 +03001624 /* 6) resize the head arrays to their final sizes */
1625 mutex_lock(&cache_chain_mutex);
1626 list_for_each_entry(cachep, &cache_chain, next)
1627 if (enable_cpucache(cachep, GFP_NOWAIT))
1628 BUG();
1629 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001630
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 /* Done! */
1632 g_cpucache_up = FULL;
1633
Pekka Enbergec5a36f2009-06-29 09:57:10 +03001634 /* Annotate slab for lockdep -- annotate the malloc caches */
1635 init_lock_keys();
1636
Andrew Mortona737b3e2006-03-22 00:08:11 -08001637 /*
1638 * Register a cpu startup notifier callback that initializes
1639 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 */
1641 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
David Rientjes8f9f8d92010-03-27 19:40:47 -07001643#ifdef CONFIG_NUMA
1644 /*
1645 * Register a memory hotplug callback that initializes and frees
1646 * nodelists.
1647 */
1648 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1649#endif
1650
Andrew Mortona737b3e2006-03-22 00:08:11 -08001651 /*
1652 * The reap timers are started later, with a module init call: That part
1653 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 */
1655}
1656
1657static int __init cpucache_init(void)
1658{
1659 int cpu;
1660
Andrew Mortona737b3e2006-03-22 00:08:11 -08001661 /*
1662 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 */
Christoph Lametere498be72005-09-09 13:03:32 -07001664 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001665 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 return 0;
1667}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668__initcall(cpucache_init);
1669
1670/*
1671 * Interface to system's page allocator. No need to hold the cache-lock.
1672 *
1673 * If we requested dmaable memory, we will get it. Even if we
1674 * did not request dmaable memory, we might get it, but that
1675 * would be relatively rare and ignorable.
1676 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001677static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678{
1679 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001680 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 int i;
1682
Luke Yangd6fef9d2006-04-10 22:52:56 -07001683#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001684 /*
1685 * Nommu uses slab's for process anonymous memory allocations, and thus
1686 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001687 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001688 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001689#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001690
Christoph Lameter3c517a62006-12-06 20:33:29 -08001691 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001692 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1693 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001694
Linus Torvalds517d0862009-06-16 19:50:13 -07001695 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 if (!page)
1697 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001699 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001701 add_zone_page_state(page_zone(page),
1702 NR_SLAB_RECLAIMABLE, nr_pages);
1703 else
1704 add_zone_page_state(page_zone(page),
1705 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001706 for (i = 0; i < nr_pages; i++)
1707 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001708
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001709 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1710 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1711
1712 if (cachep->ctor)
1713 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1714 else
1715 kmemcheck_mark_unallocated_pages(page, nr_pages);
1716 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001717
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001718 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719}
1720
1721/*
1722 * Interface to system's page release.
1723 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001724static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001726 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 struct page *page = virt_to_page(addr);
1728 const unsigned long nr_freed = i;
1729
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001730 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001731
Christoph Lameter972d1a72006-09-25 23:31:51 -07001732 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1733 sub_zone_page_state(page_zone(page),
1734 NR_SLAB_RECLAIMABLE, nr_freed);
1735 else
1736 sub_zone_page_state(page_zone(page),
1737 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001739 BUG_ON(!PageSlab(page));
1740 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 page++;
1742 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 if (current->reclaim_state)
1744 current->reclaim_state->reclaimed_slab += nr_freed;
1745 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746}
1747
1748static void kmem_rcu_free(struct rcu_head *head)
1749{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001750 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001751 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
1753 kmem_freepages(cachep, slab_rcu->addr);
1754 if (OFF_SLAB(cachep))
1755 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1756}
1757
1758#if DEBUG
1759
1760#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001761static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001762 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001764 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001766 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001768 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 return;
1770
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 *addr++ = 0x12345678;
1772 *addr++ = caller;
1773 *addr++ = smp_processor_id();
1774 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 {
1776 unsigned long *sptr = &caller;
1777 unsigned long svalue;
1778
1779 while (!kstack_end(sptr)) {
1780 svalue = *sptr++;
1781 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001782 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 size -= sizeof(unsigned long);
1784 if (size <= sizeof(unsigned long))
1785 break;
1786 }
1787 }
1788
1789 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001790 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791}
1792#endif
1793
Pekka Enberg343e0d72006-02-01 03:05:50 -08001794static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001796 int size = obj_size(cachep);
1797 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
1799 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801}
1802
1803static void dump_line(char *data, int offset, int limit)
1804{
1805 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001806 unsigned char error = 0;
1807 int bad_count = 0;
1808
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001810 for (i = 0; i < limit; i++) {
1811 if (data[offset + i] != POISON_FREE) {
1812 error = data[offset + i];
1813 bad_count++;
1814 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001815 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001816 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001818
1819 if (bad_count == 1) {
1820 error ^= POISON_FREE;
1821 if (!(error & (error - 1))) {
1822 printk(KERN_ERR "Single bit error detected. Probably "
1823 "bad RAM.\n");
1824#ifdef CONFIG_X86
1825 printk(KERN_ERR "Run memtest86+ or a similar memory "
1826 "test tool.\n");
1827#else
1828 printk(KERN_ERR "Run a memory test tool.\n");
1829#endif
1830 }
1831 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832}
1833#endif
1834
1835#if DEBUG
1836
Pekka Enberg343e0d72006-02-01 03:05:50 -08001837static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
1839 int i, size;
1840 char *realobj;
1841
1842 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001843 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001844 *dbg_redzone1(cachep, objp),
1845 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 }
1847
1848 if (cachep->flags & SLAB_STORE_USER) {
1849 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001850 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001852 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 printk("\n");
1854 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001855 realobj = (char *)objp + obj_offset(cachep);
1856 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001857 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 int limit;
1859 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 if (i + limit > size)
1861 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 dump_line(realobj, i, limit);
1863 }
1864}
1865
Pekka Enberg343e0d72006-02-01 03:05:50 -08001866static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867{
1868 char *realobj;
1869 int size, i;
1870 int lines = 0;
1871
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001872 realobj = (char *)objp + obj_offset(cachep);
1873 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 exp = POISON_END;
1879 if (realobj[i] != exp) {
1880 int limit;
1881 /* Mismatch ! */
1882 /* Print header */
1883 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001884 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001885 "Slab corruption: %s start=%p, len=%d\n",
1886 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 print_objinfo(cachep, objp, 0);
1888 }
1889 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001892 if (i + limit > size)
1893 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 dump_line(realobj, i, limit);
1895 i += 16;
1896 lines++;
1897 /* Limit to 5 lines */
1898 if (lines > 5)
1899 break;
1900 }
1901 }
1902 if (lines != 0) {
1903 /* Print some data about the neighboring objects, if they
1904 * exist:
1905 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001906 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001907 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001909 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001911 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001912 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001914 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 print_objinfo(cachep, objp, 2);
1916 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001917 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001918 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001919 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 print_objinfo(cachep, objp, 2);
1923 }
1924 }
1925}
1926#endif
1927
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301929static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001930{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 int i;
1932 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001933 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934
1935 if (cachep->flags & SLAB_POISON) {
1936#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001937 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1938 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001939 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001940 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 else
1942 check_poison_obj(cachep, objp);
1943#else
1944 check_poison_obj(cachep, objp);
1945#endif
1946 }
1947 if (cachep->flags & SLAB_RED_ZONE) {
1948 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1949 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001950 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1952 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001953 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301958static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001959{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001960}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961#endif
1962
Randy Dunlap911851e2006-03-22 00:08:14 -08001963/**
1964 * slab_destroy - destroy and release all objects in a slab
1965 * @cachep: cache pointer being destroyed
1966 * @slabp: slab pointer being destroyed
1967 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001968 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001969 * Before calling the slab must have been unlinked from the cache. The
1970 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001971 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001972static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001973{
1974 void *addr = slabp->s_mem - slabp->colouroff;
1975
Rabin Vincente79aec22008-07-04 00:40:32 +05301976 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1978 struct slab_rcu *slab_rcu;
1979
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001980 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 slab_rcu->cachep = cachep;
1982 slab_rcu->addr = addr;
1983 call_rcu(&slab_rcu->head, kmem_rcu_free);
1984 } else {
1985 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001986 if (OFF_SLAB(cachep))
1987 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 }
1989}
1990
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001991static void __kmem_cache_destroy(struct kmem_cache *cachep)
1992{
1993 int i;
1994 struct kmem_list3 *l3;
1995
1996 for_each_online_cpu(i)
1997 kfree(cachep->array[i]);
1998
1999 /* NUMA: free the list3 structures */
2000 for_each_online_node(i) {
2001 l3 = cachep->nodelists[i];
2002 if (l3) {
2003 kfree(l3->shared);
2004 free_alien_cache(l3->alien);
2005 kfree(l3);
2006 }
2007 }
2008 kmem_cache_free(&cache_cache, cachep);
2009}
2010
2011
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002013 * calculate_slab_order - calculate size (page order) of slabs
2014 * @cachep: pointer to the cache that is being created
2015 * @size: size of objects to be created in this cache.
2016 * @align: required alignment for the objects.
2017 * @flags: slab allocation flags
2018 *
2019 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002020 *
2021 * This could be made much more intelligent. For now, try to avoid using
2022 * high order pages for slabs. When the gfp() functions are more friendly
2023 * towards high-order requests, this should be changed.
2024 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002025static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002026 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002027{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002028 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002029 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002030 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002031
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002032 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002033 unsigned int num;
2034 size_t remainder;
2035
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002036 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037 if (!num)
2038 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002039
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002040 if (flags & CFLGS_OFF_SLAB) {
2041 /*
2042 * Max number of objs-per-slab for caches which
2043 * use off-slab slabs. Needed to avoid a possible
2044 * looping condition in cache_grow().
2045 */
2046 offslab_limit = size - sizeof(struct slab);
2047 offslab_limit /= sizeof(kmem_bufctl_t);
2048
2049 if (num > offslab_limit)
2050 break;
2051 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002052
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002053 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002054 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002055 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002056 left_over = remainder;
2057
2058 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002059 * A VFS-reclaimable slab tends to have most allocations
2060 * as GFP_NOFS and we really don't want to have to be allocating
2061 * higher-order pages when we are unable to shrink dcache.
2062 */
2063 if (flags & SLAB_RECLAIM_ACCOUNT)
2064 break;
2065
2066 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002067 * Large number of objects is good, but very large slabs are
2068 * currently bad for the gfp()s.
2069 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002070 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002071 break;
2072
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002073 /*
2074 * Acceptable internal fragmentation?
2075 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002076 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002077 break;
2078 }
2079 return left_over;
2080}
2081
Pekka Enberg83b519e2009-06-10 19:40:04 +03002082static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002083{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002084 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002085 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002086
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002087 if (g_cpucache_up == NONE) {
2088 /*
2089 * Note: the first kmem_cache_create must create the cache
2090 * that's used by kmalloc(24), otherwise the creation of
2091 * further caches will BUG().
2092 */
2093 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2094
2095 /*
2096 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2097 * the first cache, then we need to set up all its list3s,
2098 * otherwise the creation of further caches will BUG().
2099 */
2100 set_up_list3s(cachep, SIZE_AC);
2101 if (INDEX_AC == INDEX_L3)
2102 g_cpucache_up = PARTIAL_L3;
2103 else
2104 g_cpucache_up = PARTIAL_AC;
2105 } else {
2106 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002107 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002108
2109 if (g_cpucache_up == PARTIAL_AC) {
2110 set_up_list3s(cachep, SIZE_L3);
2111 g_cpucache_up = PARTIAL_L3;
2112 } else {
2113 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002114 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002115 cachep->nodelists[node] =
2116 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002117 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002118 BUG_ON(!cachep->nodelists[node]);
2119 kmem_list3_init(cachep->nodelists[node]);
2120 }
2121 }
2122 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002123 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002124 jiffies + REAPTIMEOUT_LIST3 +
2125 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2126
2127 cpu_cache_get(cachep)->avail = 0;
2128 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2129 cpu_cache_get(cachep)->batchcount = 1;
2130 cpu_cache_get(cachep)->touched = 0;
2131 cachep->batchcount = 1;
2132 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002133 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002134}
2135
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002136/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 * kmem_cache_create - Create a cache.
2138 * @name: A string which is used in /proc/slabinfo to identify this cache.
2139 * @size: The size of objects to be created in this cache.
2140 * @align: The required alignment for the objects.
2141 * @flags: SLAB flags
2142 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 *
2144 * Returns a ptr to the cache on success, NULL on failure.
2145 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002146 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 *
2148 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002149 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002150 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2151 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002152 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 * The flags are
2154 *
2155 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2156 * to catch references to uninitialised memory.
2157 *
2158 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2159 * for buffer overruns.
2160 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2162 * cacheline. This can be beneficial if you're counting cycles as closely
2163 * as davem.
2164 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002165struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002167 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168{
2169 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002170 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002171 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
2173 /*
2174 * Sanity checks... these are all serious usage bugs.
2175 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002177 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002178 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002179 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002180 BUG();
2181 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002183 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002184 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302185 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002186 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002187 if (slab_is_available()) {
2188 get_online_cpus();
2189 mutex_lock(&cache_chain_mutex);
2190 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002191
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002192 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002193 char tmp;
2194 int res;
2195
2196 /*
2197 * This happens when the module gets unloaded and doesn't
2198 * destroy its slab cache and no-one else reuses the vmalloc
2199 * area of the module. Print a warning.
2200 */
Andrew Morton138ae662006-12-06 20:36:41 -08002201 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002202 if (res) {
matzeb4169522007-05-06 14:49:52 -07002203 printk(KERN_ERR
2204 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002205 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002206 continue;
2207 }
2208
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002209 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002210 printk(KERN_ERR
2211 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002212 dump_stack();
2213 goto oops;
2214 }
2215 }
2216
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217#if DEBUG
2218 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219#if FORCED_DEBUG
2220 /*
2221 * Enable redzoning and last user accounting, except for caches with
2222 * large objects, if the increased size would increase the object size
2223 * above the next power of two: caches with object sizes just above a
2224 * power of two have a significant amount of internal fragmentation.
2225 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002226 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2227 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002228 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 if (!(flags & SLAB_DESTROY_BY_RCU))
2230 flags |= SLAB_POISON;
2231#endif
2232 if (flags & SLAB_DESTROY_BY_RCU)
2233 BUG_ON(flags & SLAB_POISON);
2234#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002236 * Always checks flags, a caller might be expecting debug support which
2237 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002239 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240
Andrew Mortona737b3e2006-03-22 00:08:11 -08002241 /*
2242 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 * unaligned accesses for some archs when redzoning is used, and makes
2244 * sure any on-slab bufctl's are also correctly aligned.
2245 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002246 if (size & (BYTES_PER_WORD - 1)) {
2247 size += (BYTES_PER_WORD - 1);
2248 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
2250
Andrew Mortona737b3e2006-03-22 00:08:11 -08002251 /* calculate the final buffer alignment: */
2252
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 /* 1) arch recommendation: can be overridden for debug */
2254 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002255 /*
2256 * Default alignment: as specified by the arch code. Except if
2257 * an object is really small, then squeeze multiple objects into
2258 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 */
2260 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002261 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 ralign /= 2;
2263 } else {
2264 ralign = BYTES_PER_WORD;
2265 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002266
2267 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002268 * Redzoning and user store require word alignment or possibly larger.
2269 * Note this will be overridden by architecture or caller mandated
2270 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002271 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002272 if (flags & SLAB_STORE_USER)
2273 ralign = BYTES_PER_WORD;
2274
2275 if (flags & SLAB_RED_ZONE) {
2276 ralign = REDZONE_ALIGN;
2277 /* If redzoning, ensure that the second redzone is suitably
2278 * aligned, by adjusting the object size accordingly. */
2279 size += REDZONE_ALIGN - 1;
2280 size &= ~(REDZONE_ALIGN - 1);
2281 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002282
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002283 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 if (ralign < ARCH_SLAB_MINALIGN) {
2285 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002287 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 if (ralign < align) {
2289 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 }
Shiyong Li5c5e3b32010-04-12 13:48:21 +08002291 /* disable debug if not aligning with REDZONE_ALIGN */
2292 if (ralign & (__alignof__(unsigned long long) - 1))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002293 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002294 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002295 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 */
2297 align = ralign;
2298
Pekka Enberg83b519e2009-06-10 19:40:04 +03002299 if (slab_is_available())
2300 gfp = GFP_KERNEL;
2301 else
2302 gfp = GFP_NOWAIT;
2303
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002305 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002307 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308
2309#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002310 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
Pekka Enbergca5f9702006-09-25 23:31:25 -07002312 /*
2313 * Both debugging options require word-alignment which is calculated
2314 * into align above.
2315 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 /* add space for red zone words */
Shiyong Li5c5e3b32010-04-12 13:48:21 +08002318 cachep->obj_offset += align;
2319 size += align + sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 }
2321 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002322 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002323 * the real object. But if the second red zone needs to be
2324 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002326 if (flags & SLAB_RED_ZONE)
2327 size += REDZONE_ALIGN;
2328 else
2329 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 }
2331#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002332 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002333 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2334 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 size = PAGE_SIZE;
2336 }
2337#endif
2338#endif
2339
Ingo Molnare0a42722006-06-23 02:03:46 -07002340 /*
2341 * Determine if the slab management is 'on' or 'off' slab.
2342 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002343 * it too early on. Always use on-slab management when
2344 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002345 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002346 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2347 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 /*
2349 * Size is large, assume best to place the slab management obj
2350 * off-slab (should allow better packing of objs).
2351 */
2352 flags |= CFLGS_OFF_SLAB;
2353
2354 size = ALIGN(size, align);
2355
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002356 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357
2358 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002359 printk(KERN_ERR
2360 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 kmem_cache_free(&cache_cache, cachep);
2362 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002363 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002365 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2366 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
2368 /*
2369 * If the slab has been placed off-slab, and we have enough space then
2370 * move it on-slab. This is at the expense of any extra colouring.
2371 */
2372 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2373 flags &= ~CFLGS_OFF_SLAB;
2374 left_over -= slab_size;
2375 }
2376
2377 if (flags & CFLGS_OFF_SLAB) {
2378 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002379 slab_size =
2380 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302381
2382#ifdef CONFIG_PAGE_POISONING
2383 /* If we're going to use the generic kernel_map_pages()
2384 * poisoning, then it's going to smash the contents of
2385 * the redzone and userword anyhow, so switch them off.
2386 */
2387 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2388 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2389#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 }
2391
2392 cachep->colour_off = cache_line_size();
2393 /* Offset must be a multiple of the alignment. */
2394 if (cachep->colour_off < align)
2395 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 cachep->slab_size = slab_size;
2398 cachep->flags = flags;
2399 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002400 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002402 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002403 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002405 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002406 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002407 /*
2408 * This is a possibility for one of the malloc_sizes caches.
2409 * But since we go off slab only for object size greater than
2410 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2411 * this should not happen at all.
2412 * But leave a BUG_ON for some lucky dude.
2413 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002414 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002415 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 cachep->name = name;
2418
Pekka Enberg83b519e2009-06-10 19:40:04 +03002419 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002420 __kmem_cache_destroy(cachep);
2421 cachep = NULL;
2422 goto oops;
2423 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 /* cache setup completed, link it into the list */
2426 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002427oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 if (!cachep && (flags & SLAB_PANIC))
2429 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002430 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002431 if (slab_is_available()) {
2432 mutex_unlock(&cache_chain_mutex);
2433 put_online_cpus();
2434 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 return cachep;
2436}
2437EXPORT_SYMBOL(kmem_cache_create);
2438
2439#if DEBUG
2440static void check_irq_off(void)
2441{
2442 BUG_ON(!irqs_disabled());
2443}
2444
2445static void check_irq_on(void)
2446{
2447 BUG_ON(irqs_disabled());
2448}
2449
Pekka Enberg343e0d72006-02-01 03:05:50 -08002450static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451{
2452#ifdef CONFIG_SMP
2453 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002454 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455#endif
2456}
Christoph Lametere498be72005-09-09 13:03:32 -07002457
Pekka Enberg343e0d72006-02-01 03:05:50 -08002458static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002459{
2460#ifdef CONFIG_SMP
2461 check_irq_off();
2462 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2463#endif
2464}
2465
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466#else
2467#define check_irq_off() do { } while(0)
2468#define check_irq_on() do { } while(0)
2469#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002470#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471#endif
2472
Christoph Lameteraab22072006-03-22 00:09:06 -08002473static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2474 struct array_cache *ac,
2475 int force, int node);
2476
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477static void do_drain(void *arg)
2478{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002479 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002481 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482
2483 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002484 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002485 spin_lock(&cachep->nodelists[node]->list_lock);
2486 free_block(cachep, ac->entry, ac->avail, node);
2487 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 ac->avail = 0;
2489}
2490
Pekka Enberg343e0d72006-02-01 03:05:50 -08002491static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492{
Christoph Lametere498be72005-09-09 13:03:32 -07002493 struct kmem_list3 *l3;
2494 int node;
2495
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002496 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002498 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002499 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002500 if (l3 && l3->alien)
2501 drain_alien_cache(cachep, l3->alien);
2502 }
2503
2504 for_each_online_node(node) {
2505 l3 = cachep->nodelists[node];
2506 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002507 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002508 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509}
2510
Christoph Lametered11d9e2006-06-30 01:55:45 -07002511/*
2512 * Remove slabs from the list of free slabs.
2513 * Specify the number of slabs to drain in tofree.
2514 *
2515 * Returns the actual number of slabs released.
2516 */
2517static int drain_freelist(struct kmem_cache *cache,
2518 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002520 struct list_head *p;
2521 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523
Christoph Lametered11d9e2006-06-30 01:55:45 -07002524 nr_freed = 0;
2525 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526
Christoph Lametered11d9e2006-06-30 01:55:45 -07002527 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002528 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002529 if (p == &l3->slabs_free) {
2530 spin_unlock_irq(&l3->list_lock);
2531 goto out;
2532 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
Christoph Lametered11d9e2006-06-30 01:55:45 -07002534 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002536 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537#endif
2538 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002539 /*
2540 * Safe to drop the lock. The slab is no longer linked
2541 * to the cache.
2542 */
2543 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002544 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002545 slab_destroy(cache, slabp);
2546 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002548out:
2549 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550}
2551
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002552/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002553static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002554{
2555 int ret = 0, i = 0;
2556 struct kmem_list3 *l3;
2557
2558 drain_cpu_caches(cachep);
2559
2560 check_irq_on();
2561 for_each_online_node(i) {
2562 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002563 if (!l3)
2564 continue;
2565
2566 drain_freelist(cachep, l3, l3->free_objects);
2567
2568 ret += !list_empty(&l3->slabs_full) ||
2569 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002570 }
2571 return (ret ? 1 : 0);
2572}
2573
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574/**
2575 * kmem_cache_shrink - Shrink a cache.
2576 * @cachep: The cache to shrink.
2577 *
2578 * Releases as many slabs as possible for a cache.
2579 * To help debugging, a zero exit status indicates all slabs were released.
2580 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002581int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002583 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002584 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002586 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002587 mutex_lock(&cache_chain_mutex);
2588 ret = __cache_shrink(cachep);
2589 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002590 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002591 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592}
2593EXPORT_SYMBOL(kmem_cache_shrink);
2594
2595/**
2596 * kmem_cache_destroy - delete a cache
2597 * @cachep: the cache to destroy
2598 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002599 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 *
2601 * It is expected this function will be called by a module when it is
2602 * unloaded. This will remove the cache completely, and avoid a duplicate
2603 * cache being allocated each time a module is loaded and unloaded, if the
2604 * module doesn't have persistent in-kernel storage across loads and unloads.
2605 *
2606 * The cache must be empty before calling this function.
2607 *
2608 * The caller must guarantee that noone will allocate memory from the cache
2609 * during the kmem_cache_destroy().
2610 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002611void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002613 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002616 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002617 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 /*
2619 * the chain is never empty, cache_cache is never destroyed
2620 */
2621 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 if (__cache_shrink(cachep)) {
2623 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002624 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002625 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002626 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002627 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 }
2629
2630 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002631 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002633 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002634 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002635 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636}
2637EXPORT_SYMBOL(kmem_cache_destroy);
2638
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002639/*
2640 * Get the memory for a slab management obj.
2641 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2642 * always come from malloc_sizes caches. The slab descriptor cannot
2643 * come from the same cache which is getting created because,
2644 * when we are searching for an appropriate cache for these
2645 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2646 * If we are creating a malloc_sizes cache here it would not be visible to
2647 * kmem_find_general_cachep till the initialization is complete.
2648 * Hence we cannot have slabp_cache same as the original cache.
2649 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002650static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002651 int colour_off, gfp_t local_flags,
2652 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653{
2654 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002655
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 if (OFF_SLAB(cachep)) {
2657 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002658 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002659 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002660 /*
2661 * If the first object in the slab is leaked (it's allocated
2662 * but no one has a reference to it), we want to make sure
2663 * kmemleak does not treat the ->s_mem pointer as a reference
2664 * to the object. Otherwise we will not report the leak.
2665 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002666 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2667 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 if (!slabp)
2669 return NULL;
2670 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002671 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 colour_off += cachep->slab_size;
2673 }
2674 slabp->inuse = 0;
2675 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002677 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002678 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 return slabp;
2680}
2681
2682static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2683{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002684 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685}
2686
Pekka Enberg343e0d72006-02-01 03:05:50 -08002687static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002688 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689{
2690 int i;
2691
2692 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002693 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694#if DEBUG
2695 /* need to poison the objs? */
2696 if (cachep->flags & SLAB_POISON)
2697 poison_obj(cachep, objp, POISON_FREE);
2698 if (cachep->flags & SLAB_STORE_USER)
2699 *dbg_userword(cachep, objp) = NULL;
2700
2701 if (cachep->flags & SLAB_RED_ZONE) {
2702 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2703 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2704 }
2705 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002706 * Constructors are not allowed to allocate memory from the same
2707 * cache which they are a constructor for. Otherwise, deadlock.
2708 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 */
2710 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002711 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712
2713 if (cachep->flags & SLAB_RED_ZONE) {
2714 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2715 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002716 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2718 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002719 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002721 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2722 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002723 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002724 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725#else
2726 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002727 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002729 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002731 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733
Pekka Enberg343e0d72006-02-01 03:05:50 -08002734static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002736 if (CONFIG_ZONE_DMA_FLAG) {
2737 if (flags & GFP_DMA)
2738 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2739 else
2740 BUG_ON(cachep->gfpflags & GFP_DMA);
2741 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742}
2743
Andrew Mortona737b3e2006-03-22 00:08:11 -08002744static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2745 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002746{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002747 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002748 kmem_bufctl_t next;
2749
2750 slabp->inuse++;
2751 next = slab_bufctl(slabp)[slabp->free];
2752#if DEBUG
2753 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2754 WARN_ON(slabp->nodeid != nodeid);
2755#endif
2756 slabp->free = next;
2757
2758 return objp;
2759}
2760
Andrew Mortona737b3e2006-03-22 00:08:11 -08002761static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2762 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002763{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002764 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002765
2766#if DEBUG
2767 /* Verify that the slab belongs to the intended node */
2768 WARN_ON(slabp->nodeid != nodeid);
2769
Al Viro871751e2006-03-25 03:06:39 -08002770 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002771 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002773 BUG();
2774 }
2775#endif
2776 slab_bufctl(slabp)[objnr] = slabp->free;
2777 slabp->free = objnr;
2778 slabp->inuse--;
2779}
2780
Pekka Enberg47768742006-06-23 02:03:07 -07002781/*
2782 * Map pages beginning at addr to the given cache and slab. This is required
2783 * for the slab allocator to be able to lookup the cache and slab of a
2784 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2785 */
2786static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2787 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788{
Pekka Enberg47768742006-06-23 02:03:07 -07002789 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 struct page *page;
2791
Pekka Enberg47768742006-06-23 02:03:07 -07002792 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002793
Pekka Enberg47768742006-06-23 02:03:07 -07002794 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002795 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002796 nr_pages <<= cache->gfporder;
2797
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002799 page_set_cache(page, cache);
2800 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002802 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803}
2804
2805/*
2806 * Grow (by 1) the number of slabs within a cache. This is called by
2807 * kmem_cache_alloc() when there are no active objs left in a cache.
2808 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002809static int cache_grow(struct kmem_cache *cachep,
2810 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002812 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002813 size_t offset;
2814 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002815 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816
Andrew Mortona737b3e2006-03-22 00:08:11 -08002817 /*
2818 * Be lazy and only check for valid flags here, keeping it out of the
2819 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002821 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2822 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002824 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002826 l3 = cachep->nodelists[nodeid];
2827 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
2829 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002830 offset = l3->colour_next;
2831 l3->colour_next++;
2832 if (l3->colour_next >= cachep->colour)
2833 l3->colour_next = 0;
2834 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002836 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837
2838 if (local_flags & __GFP_WAIT)
2839 local_irq_enable();
2840
2841 /*
2842 * The test for missing atomic flag is performed here, rather than
2843 * the more obvious place, simply to reduce the critical path length
2844 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2845 * will eventually be caught here (where it matters).
2846 */
2847 kmem_flagcheck(cachep, flags);
2848
Andrew Mortona737b3e2006-03-22 00:08:11 -08002849 /*
2850 * Get mem for the objs. Attempt to allocate a physical page from
2851 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002852 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002853 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002854 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002855 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 goto failed;
2857
2858 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002859 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002860 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002861 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 goto opps1;
2863
Pekka Enberg47768742006-06-23 02:03:07 -07002864 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865
Christoph Lametera35afb82007-05-16 22:10:57 -07002866 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
2868 if (local_flags & __GFP_WAIT)
2869 local_irq_disable();
2870 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002871 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872
2873 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002874 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002876 l3->free_objects += cachep->num;
2877 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002879opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002881failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 if (local_flags & __GFP_WAIT)
2883 local_irq_disable();
2884 return 0;
2885}
2886
2887#if DEBUG
2888
2889/*
2890 * Perform extra freeing checks:
2891 * - detect bad pointers.
2892 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 */
2894static void kfree_debugcheck(const void *objp)
2895{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 if (!virt_addr_valid(objp)) {
2897 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002898 (unsigned long)objp);
2899 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901}
2902
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002903static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2904{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002905 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002906
2907 redzone1 = *dbg_redzone1(cache, obj);
2908 redzone2 = *dbg_redzone2(cache, obj);
2909
2910 /*
2911 * Redzone is ok.
2912 */
2913 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2914 return;
2915
2916 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2917 slab_error(cache, "double free detected");
2918 else
2919 slab_error(cache, "memory outside object was overwritten");
2920
David Woodhouseb46b8f12007-05-08 00:22:59 -07002921 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002922 obj, redzone1, redzone2);
2923}
2924
Pekka Enberg343e0d72006-02-01 03:05:50 -08002925static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002926 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927{
2928 struct page *page;
2929 unsigned int objnr;
2930 struct slab *slabp;
2931
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002932 BUG_ON(virt_to_cache(objp) != cachep);
2933
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002934 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002936 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
Pekka Enberg065d41c2005-11-13 16:06:46 -08002938 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939
2940 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002941 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2943 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2944 }
2945 if (cachep->flags & SLAB_STORE_USER)
2946 *dbg_userword(cachep, objp) = caller;
2947
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002948 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949
2950 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002951 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952
Al Viro871751e2006-03-25 03:06:39 -08002953#ifdef CONFIG_DEBUG_SLAB_LEAK
2954 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2955#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 if (cachep->flags & SLAB_POISON) {
2957#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002958 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002960 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002961 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 } else {
2963 poison_obj(cachep, objp, POISON_FREE);
2964 }
2965#else
2966 poison_obj(cachep, objp, POISON_FREE);
2967#endif
2968 }
2969 return objp;
2970}
2971
Pekka Enberg343e0d72006-02-01 03:05:50 -08002972static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973{
2974 kmem_bufctl_t i;
2975 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002976
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 /* Check slab's freelist to see if this obj is there. */
2978 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2979 entries++;
2980 if (entries > cachep->num || i >= cachep->num)
2981 goto bad;
2982 }
2983 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002984bad:
2985 printk(KERN_ERR "slab: Internal list corruption detected in "
2986 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2987 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002988 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002989 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002990 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002991 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002993 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 }
2995 printk("\n");
2996 BUG();
2997 }
2998}
2999#else
3000#define kfree_debugcheck(x) do { } while(0)
3001#define cache_free_debugcheck(x,objp,z) (objp)
3002#define check_slabp(x,y) do { } while(0)
3003#endif
3004
Pekka Enberg343e0d72006-02-01 03:05:50 -08003005static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006{
3007 int batchcount;
3008 struct kmem_list3 *l3;
3009 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003010 int node;
3011
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003013 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003014 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003015 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 batchcount = ac->batchcount;
3017 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003018 /*
3019 * If there was little recent activity on this cache, then
3020 * perform only a partial refill. Otherwise we could generate
3021 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 */
3023 batchcount = BATCHREFILL_LIMIT;
3024 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003025 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Christoph Lametere498be72005-09-09 13:03:32 -07003027 BUG_ON(ac->avail > 0 || !l3);
3028 spin_lock(&l3->list_lock);
3029
Christoph Lameter3ded1752006-03-25 03:06:44 -08003030 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003031 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3032 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003033 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003034 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003035
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 while (batchcount > 0) {
3037 struct list_head *entry;
3038 struct slab *slabp;
3039 /* Get slab alloc is to come from. */
3040 entry = l3->slabs_partial.next;
3041 if (entry == &l3->slabs_partial) {
3042 l3->free_touched = 1;
3043 entry = l3->slabs_free.next;
3044 if (entry == &l3->slabs_free)
3045 goto must_grow;
3046 }
3047
3048 slabp = list_entry(entry, struct slab, list);
3049 check_slabp(cachep, slabp);
3050 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003051
3052 /*
3053 * The slab was either on partial or free list so
3054 * there must be at least one object available for
3055 * allocation.
3056 */
roel kluin249b9f32008-10-29 17:18:07 -04003057 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07003058
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 STATS_INC_ALLOCED(cachep);
3061 STATS_INC_ACTIVE(cachep);
3062 STATS_SET_HIGH(cachep);
3063
Matthew Dobson78d382d2006-02-01 03:05:47 -08003064 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003065 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 }
3067 check_slabp(cachep, slabp);
3068
3069 /* move slabp to correct slabp list: */
3070 list_del(&slabp->list);
3071 if (slabp->free == BUFCTL_END)
3072 list_add(&slabp->list, &l3->slabs_full);
3073 else
3074 list_add(&slabp->list, &l3->slabs_partial);
3075 }
3076
Andrew Mortona737b3e2006-03-22 00:08:11 -08003077must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003079alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003080 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081
3082 if (unlikely(!ac->avail)) {
3083 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003084 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003085
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003087 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003088 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 return NULL;
3090
Andrew Mortona737b3e2006-03-22 00:08:11 -08003091 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 goto retry;
3093 }
3094 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003095 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096}
3097
Andrew Mortona737b3e2006-03-22 00:08:11 -08003098static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3099 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100{
3101 might_sleep_if(flags & __GFP_WAIT);
3102#if DEBUG
3103 kmem_flagcheck(cachep, flags);
3104#endif
3105}
3106
3107#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003108static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3109 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003111 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003113 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003115 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003116 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003117 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 else
3119 check_poison_obj(cachep, objp);
3120#else
3121 check_poison_obj(cachep, objp);
3122#endif
3123 poison_obj(cachep, objp, POISON_INUSE);
3124 }
3125 if (cachep->flags & SLAB_STORE_USER)
3126 *dbg_userword(cachep, objp) = caller;
3127
3128 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003129 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3130 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3131 slab_error(cachep, "double free, or memory outside"
3132 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003133 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003134 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003135 objp, *dbg_redzone1(cachep, objp),
3136 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 }
3138 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3139 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3140 }
Al Viro871751e2006-03-25 03:06:39 -08003141#ifdef CONFIG_DEBUG_SLAB_LEAK
3142 {
3143 struct slab *slabp;
3144 unsigned objnr;
3145
Christoph Lameterb49af682007-05-06 14:49:41 -07003146 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003147 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3148 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3149 }
3150#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003151 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003152 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003153 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003154#if ARCH_SLAB_MINALIGN
3155 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3156 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3157 objp, ARCH_SLAB_MINALIGN);
3158 }
3159#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 return objp;
3161}
3162#else
3163#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3164#endif
3165
Akinobu Mita773ff602008-12-23 19:37:01 +09003166static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003167{
3168 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003169 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003170
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003171 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003172}
3173
Pekka Enberg343e0d72006-02-01 03:05:50 -08003174static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003176 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 struct array_cache *ac;
3178
Alok N Kataria5c382302005-09-27 21:45:46 -07003179 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003180
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003181 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 if (likely(ac->avail)) {
3183 STATS_INC_ALLOCHIT(cachep);
3184 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003185 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 } else {
3187 STATS_INC_ALLOCMISS(cachep);
3188 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003189 /*
3190 * the 'ac' may be updated by cache_alloc_refill(),
3191 * and kmemleak_erase() requires its correct value.
3192 */
3193 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003195 /*
3196 * To avoid a false negative, if an object that is in one of the
3197 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3198 * treat the array pointers as a reference to the object.
3199 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003200 if (objp)
3201 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003202 return objp;
3203}
3204
Christoph Lametere498be72005-09-09 13:03:32 -07003205#ifdef CONFIG_NUMA
3206/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003207 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003208 *
3209 * If we are in_interrupt, then process context, including cpusets and
3210 * mempolicy, may not apply and should not be used for allocation policy.
3211 */
3212static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3213{
3214 int nid_alloc, nid_here;
3215
Christoph Lameter765c4502006-09-27 01:50:08 -07003216 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003217 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003218 nid_alloc = nid_here = numa_mem_id();
Miao Xiec0ff7452010-05-24 14:32:08 -07003219 get_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003220 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003221 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003222 else if (current->mempolicy)
3223 nid_alloc = slab_node(current->mempolicy);
Miao Xiec0ff7452010-05-24 14:32:08 -07003224 put_mems_allowed();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003225 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003226 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003227 return NULL;
3228}
3229
3230/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003231 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003232 * certain node and fall back is permitted. First we scan all the
3233 * available nodelists for available objects. If that fails then we
3234 * perform an allocation without specifying a node. This allows the page
3235 * allocator to do its reclaim / fallback magic. We then insert the
3236 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003237 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003238static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003239{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003240 struct zonelist *zonelist;
3241 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003242 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003243 struct zone *zone;
3244 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003245 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003246 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003247
3248 if (flags & __GFP_THISNODE)
3249 return NULL;
3250
Miao Xiec0ff7452010-05-24 14:32:08 -07003251 get_mems_allowed();
Mel Gorman0e884602008-04-28 02:12:14 -07003252 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003253 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003254
Christoph Lameter3c517a62006-12-06 20:33:29 -08003255retry:
3256 /*
3257 * Look through allowed nodes for objects available
3258 * from existing per node queues.
3259 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003260 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3261 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003262
Mel Gorman54a6eb52008-04-28 02:12:16 -07003263 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003264 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003265 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003266 obj = ____cache_alloc_node(cache,
3267 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003268 if (obj)
3269 break;
3270 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003271 }
3272
Christoph Lametercfce6602007-05-06 14:50:17 -07003273 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003274 /*
3275 * This allocation will be performed within the constraints
3276 * of the current cpuset / memory policy requirements.
3277 * We may trigger various forms of reclaim on the allowed
3278 * set and go into memory reserves if necessary.
3279 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003280 if (local_flags & __GFP_WAIT)
3281 local_irq_enable();
3282 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003283 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003284 if (local_flags & __GFP_WAIT)
3285 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003286 if (obj) {
3287 /*
3288 * Insert into the appropriate per node queues
3289 */
3290 nid = page_to_nid(virt_to_page(obj));
3291 if (cache_grow(cache, flags, nid, obj)) {
3292 obj = ____cache_alloc_node(cache,
3293 flags | GFP_THISNODE, nid);
3294 if (!obj)
3295 /*
3296 * Another processor may allocate the
3297 * objects in the slab since we are
3298 * not holding any locks.
3299 */
3300 goto retry;
3301 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003302 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003303 obj = NULL;
3304 }
3305 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003306 }
Miao Xiec0ff7452010-05-24 14:32:08 -07003307 put_mems_allowed();
Christoph Lameter765c4502006-09-27 01:50:08 -07003308 return obj;
3309}
3310
3311/*
Christoph Lametere498be72005-09-09 13:03:32 -07003312 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003314static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003315 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003316{
3317 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 struct slab *slabp;
3319 struct kmem_list3 *l3;
3320 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 l3 = cachep->nodelists[nodeid];
3324 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Andrew Mortona737b3e2006-03-22 00:08:11 -08003326retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003327 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 spin_lock(&l3->list_lock);
3329 entry = l3->slabs_partial.next;
3330 if (entry == &l3->slabs_partial) {
3331 l3->free_touched = 1;
3332 entry = l3->slabs_free.next;
3333 if (entry == &l3->slabs_free)
3334 goto must_grow;
3335 }
Christoph Lametere498be72005-09-09 13:03:32 -07003336
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003337 slabp = list_entry(entry, struct slab, list);
3338 check_spinlock_acquired_node(cachep, nodeid);
3339 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003341 STATS_INC_NODEALLOCS(cachep);
3342 STATS_INC_ACTIVE(cachep);
3343 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003344
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003345 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Matthew Dobson78d382d2006-02-01 03:05:47 -08003347 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003348 check_slabp(cachep, slabp);
3349 l3->free_objects--;
3350 /* move slabp to correct slabp list: */
3351 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003352
Andrew Mortona737b3e2006-03-22 00:08:11 -08003353 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003354 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003355 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003356 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003357
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 spin_unlock(&l3->list_lock);
3359 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003360
Andrew Mortona737b3e2006-03-22 00:08:11 -08003361must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003362 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003363 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003364 if (x)
3365 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003366
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003367 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003368
Andrew Mortona737b3e2006-03-22 00:08:11 -08003369done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003370 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003371}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003372
3373/**
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3379 *
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3382 *
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3384 */
3385static __always_inline void *
3386__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3387 void *caller)
3388{
3389 unsigned long save_flags;
3390 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003391 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003392
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003393 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003394
Nick Piggincf40bd12009-01-21 08:12:39 +01003395 lockdep_trace_alloc(flags);
3396
Akinobu Mita773ff602008-12-23 19:37:01 +09003397 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003398 return NULL;
3399
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003400 cache_alloc_debugcheck_before(cachep, flags);
3401 local_irq_save(save_flags);
3402
Tim Blechmann8e15b792009-11-30 18:59:34 +01003403 if (nodeid == -1)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003404 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003405
3406 if (unlikely(!cachep->nodelists[nodeid])) {
3407 /* Node not bootstrapped yet */
3408 ptr = fallback_alloc(cachep, flags);
3409 goto out;
3410 }
3411
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003412 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003413 /*
3414 * Use the locally cached objects if possible.
3415 * However ____cache_alloc does not allow fallback
3416 * to other nodes. It may fail while we still have
3417 * objects on other nodes available.
3418 */
3419 ptr = ____cache_alloc(cachep, flags);
3420 if (ptr)
3421 goto out;
3422 }
3423 /* ___cache_alloc_node can fall back to other nodes */
3424 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3425 out:
3426 local_irq_restore(save_flags);
3427 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003428 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3429 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003430
Pekka Enbergc175eea2008-05-09 20:35:53 +02003431 if (likely(ptr))
3432 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3433
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003434 if (unlikely((flags & __GFP_ZERO) && ptr))
3435 memset(ptr, 0, obj_size(cachep));
3436
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003437 return ptr;
3438}
3439
3440static __always_inline void *
3441__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3442{
3443 void *objp;
3444
3445 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3446 objp = alternate_node_alloc(cache, flags);
3447 if (objp)
3448 goto out;
3449 }
3450 objp = ____cache_alloc(cache, flags);
3451
3452 /*
3453 * We may just have run out of memory on the local node.
3454 * ____cache_alloc_node() knows how to locate memory on other nodes
3455 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003456 if (!objp)
3457 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003458
3459 out:
3460 return objp;
3461}
3462#else
3463
3464static __always_inline void *
3465__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3466{
3467 return ____cache_alloc(cachep, flags);
3468}
3469
3470#endif /* CONFIG_NUMA */
3471
3472static __always_inline void *
3473__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3474{
3475 unsigned long save_flags;
3476 void *objp;
3477
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003478 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003479
Nick Piggincf40bd12009-01-21 08:12:39 +01003480 lockdep_trace_alloc(flags);
3481
Akinobu Mita773ff602008-12-23 19:37:01 +09003482 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003483 return NULL;
3484
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003485 cache_alloc_debugcheck_before(cachep, flags);
3486 local_irq_save(save_flags);
3487 objp = __do_cache_alloc(cachep, flags);
3488 local_irq_restore(save_flags);
3489 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003490 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3491 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003492 prefetchw(objp);
3493
Pekka Enbergc175eea2008-05-09 20:35:53 +02003494 if (likely(objp))
3495 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3496
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003497 if (unlikely((flags & __GFP_ZERO) && objp))
3498 memset(objp, 0, obj_size(cachep));
3499
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003500 return objp;
3501}
Christoph Lametere498be72005-09-09 13:03:32 -07003502
3503/*
3504 * Caller needs to acquire correct kmem_list's list_lock
3505 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003506static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003507 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508{
3509 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003510 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511
3512 for (i = 0; i < nr_objects; i++) {
3513 void *objp = objpp[i];
3514 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003516 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003517 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003519 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003521 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003523 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 check_slabp(cachep, slabp);
3525
3526 /* fixup slab chains */
3527 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003528 if (l3->free_objects > l3->free_limit) {
3529 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003530 /* No need to drop any previously held
3531 * lock here, even if we have a off-slab slab
3532 * descriptor it is guaranteed to come from
3533 * a different cache, refer to comments before
3534 * alloc_slabmgmt.
3535 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 slab_destroy(cachep, slabp);
3537 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003538 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 }
3540 } else {
3541 /* Unconditionally move a slab to the end of the
3542 * partial list on free - maximum time for the
3543 * other objects to be freed, too.
3544 */
Christoph Lametere498be72005-09-09 13:03:32 -07003545 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546 }
3547 }
3548}
3549
Pekka Enberg343e0d72006-02-01 03:05:50 -08003550static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551{
3552 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003553 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003554 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555
3556 batchcount = ac->batchcount;
3557#if DEBUG
3558 BUG_ON(!batchcount || batchcount > ac->avail);
3559#endif
3560 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003561 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003562 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003563 if (l3->shared) {
3564 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003565 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 if (max) {
3567 if (batchcount > max)
3568 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003569 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003570 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 shared_array->avail += batchcount;
3572 goto free_done;
3573 }
3574 }
3575
Christoph Lameterff694162005-09-22 21:44:02 -07003576 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003577free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578#if STATS
3579 {
3580 int i = 0;
3581 struct list_head *p;
3582
Christoph Lametere498be72005-09-09 13:03:32 -07003583 p = l3->slabs_free.next;
3584 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 struct slab *slabp;
3586
3587 slabp = list_entry(p, struct slab, list);
3588 BUG_ON(slabp->inuse);
3589
3590 i++;
3591 p = p->next;
3592 }
3593 STATS_SET_FREEABLE(cachep, i);
3594 }
3595#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003596 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003598 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599}
3600
3601/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003602 * Release an obj back to its cache. If the obj has a constructed state, it must
3603 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003605static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003607 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608
3609 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003610 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3612
Pekka Enbergc175eea2008-05-09 20:35:53 +02003613 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3614
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003615 /*
3616 * Skip calling cache_free_alien() when the platform is not numa.
3617 * This will avoid cache misses that happen while accessing slabp (which
3618 * is per page memory reference) to get nodeid. Instead use a global
3619 * variable to skip the call, which is mostly likely to be present in
3620 * the cache.
3621 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003622 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003623 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003624
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 if (likely(ac->avail < ac->limit)) {
3626 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003627 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 return;
3629 } else {
3630 STATS_INC_FREEMISS(cachep);
3631 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003632 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633 }
3634}
3635
3636/**
3637 * kmem_cache_alloc - Allocate an object
3638 * @cachep: The cache to allocate from.
3639 * @flags: See kmalloc().
3640 *
3641 * Allocate an object from this cache. The flags are only relevant
3642 * if the cache has no available objects.
3643 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003644void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003646 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3647
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003648 trace_kmem_cache_alloc(_RET_IP_, ret,
3649 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003650
3651 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652}
3653EXPORT_SYMBOL(kmem_cache_alloc);
3654
Li Zefan0f24f122009-12-11 15:45:30 +08003655#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003656void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3657{
3658 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3659}
3660EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3661#endif
3662
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663/**
Randy Dunlap76824862008-03-19 17:00:40 -07003664 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 * @cachep: the cache we're checking against
3666 * @ptr: pointer to validate
3667 *
Randy Dunlap76824862008-03-19 17:00:40 -07003668 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 * it is _not_ a guarantee that the pointer is actually
3670 * part of the slab cache in question, but it at least
3671 * validates that the pointer can be dereferenced and
3672 * looks half-way sane.
3673 *
3674 * Currently only used for dentry validation.
3675 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003676int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003678 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 struct page *page;
3680
Pekka Enbergfc1c1832010-04-07 19:23:40 +03003681 if (unlikely(!kern_ptr_validate(ptr, size)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682 goto out;
3683 page = virt_to_page(ptr);
3684 if (unlikely(!PageSlab(page)))
3685 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003686 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687 goto out;
3688 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003689out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690 return 0;
3691}
3692
3693#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003694void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3695{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003696 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3697 __builtin_return_address(0));
3698
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003699 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3700 obj_size(cachep), cachep->buffer_size,
3701 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003702
3703 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003704}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705EXPORT_SYMBOL(kmem_cache_alloc_node);
3706
Li Zefan0f24f122009-12-11 15:45:30 +08003707#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003708void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3709 gfp_t flags,
3710 int nodeid)
3711{
3712 return __cache_alloc_node(cachep, flags, nodeid,
3713 __builtin_return_address(0));
3714}
3715EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3716#endif
3717
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003718static __always_inline void *
3719__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003720{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003721 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003722 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003723
3724 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003725 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3726 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003727 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3728
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003729 trace_kmalloc_node((unsigned long) caller, ret,
3730 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003731
3732 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003733}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003734
Li Zefan0bb38a52009-12-11 15:45:50 +08003735#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003736void *__kmalloc_node(size_t size, gfp_t flags, int node)
3737{
3738 return __do_kmalloc_node(size, flags, node,
3739 __builtin_return_address(0));
3740}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003741EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003742
3743void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003744 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003745{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003746 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003747}
3748EXPORT_SYMBOL(__kmalloc_node_track_caller);
3749#else
3750void *__kmalloc_node(size_t size, gfp_t flags, int node)
3751{
3752 return __do_kmalloc_node(size, flags, node, NULL);
3753}
3754EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003755#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003756#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757
3758/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003759 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003761 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003762 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003764static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3765 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003767 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003768 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003770 /* If you want to save a few bytes .text space: replace
3771 * __ with kmem_.
3772 * Then kmalloc uses the uninlined functions instead of the inline
3773 * functions.
3774 */
3775 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003776 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3777 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003778 ret = __cache_alloc(cachep, flags, caller);
3779
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003780 trace_kmalloc((unsigned long) caller, ret,
3781 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003782
3783 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003784}
3785
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003786
Li Zefan0bb38a52009-12-11 15:45:50 +08003787#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003788void *__kmalloc(size_t size, gfp_t flags)
3789{
Al Viro871751e2006-03-25 03:06:39 -08003790 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791}
3792EXPORT_SYMBOL(__kmalloc);
3793
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003794void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003795{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003796 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003797}
3798EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003799
3800#else
3801void *__kmalloc(size_t size, gfp_t flags)
3802{
3803 return __do_kmalloc(size, flags, NULL);
3804}
3805EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003806#endif
3807
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808/**
3809 * kmem_cache_free - Deallocate an object
3810 * @cachep: The cache the allocation was from.
3811 * @objp: The previously allocated object.
3812 *
3813 * Free an object which was previously allocated from this
3814 * cache.
3815 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003816void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817{
3818 unsigned long flags;
3819
3820 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003821 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003822 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3823 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003824 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003826
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003827 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828}
3829EXPORT_SYMBOL(kmem_cache_free);
3830
3831/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 * kfree - free previously allocated memory
3833 * @objp: pointer returned by kmalloc.
3834 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003835 * If @objp is NULL, no operation is performed.
3836 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837 * Don't free memory not originally allocated by kmalloc()
3838 * or you will run into trouble.
3839 */
3840void kfree(const void *objp)
3841{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003842 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843 unsigned long flags;
3844
Pekka Enberg2121db72009-03-25 11:05:57 +02003845 trace_kfree(_RET_IP_, objp);
3846
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003847 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848 return;
3849 local_irq_save(flags);
3850 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003851 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003852 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003853 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003854 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 local_irq_restore(flags);
3856}
3857EXPORT_SYMBOL(kfree);
3858
Pekka Enberg343e0d72006-02-01 03:05:50 -08003859unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003861 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862}
3863EXPORT_SYMBOL(kmem_cache_size);
3864
Pekka Enberg343e0d72006-02-01 03:05:50 -08003865const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003866{
3867 return cachep->name;
3868}
3869EXPORT_SYMBOL_GPL(kmem_cache_name);
3870
Christoph Lametere498be72005-09-09 13:03:32 -07003871/*
Simon Arlott183ff222007-10-20 01:27:18 +02003872 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003873 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003874static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003875{
3876 int node;
3877 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003878 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003879 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003880
Mel Gorman9c09a952008-01-24 05:49:54 -08003881 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003882
Paul Menage3395ee02006-12-06 20:32:16 -08003883 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003884 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003885 if (!new_alien)
3886 goto fail;
3887 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003888
Eric Dumazet63109842007-05-06 14:49:28 -07003889 new_shared = NULL;
3890 if (cachep->shared) {
3891 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003892 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003893 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003894 if (!new_shared) {
3895 free_alien_cache(new_alien);
3896 goto fail;
3897 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003898 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003899
Andrew Mortona737b3e2006-03-22 00:08:11 -08003900 l3 = cachep->nodelists[node];
3901 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003902 struct array_cache *shared = l3->shared;
3903
Christoph Lametere498be72005-09-09 13:03:32 -07003904 spin_lock_irq(&l3->list_lock);
3905
Christoph Lametercafeb022006-03-25 03:06:46 -08003906 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003907 free_block(cachep, shared->entry,
3908 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003909
Christoph Lametercafeb022006-03-25 03:06:46 -08003910 l3->shared = new_shared;
3911 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003912 l3->alien = new_alien;
3913 new_alien = NULL;
3914 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003915 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003916 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003917 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003918 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003919 free_alien_cache(new_alien);
3920 continue;
3921 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003922 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003923 if (!l3) {
3924 free_alien_cache(new_alien);
3925 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003926 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003927 }
Christoph Lametere498be72005-09-09 13:03:32 -07003928
3929 kmem_list3_init(l3);
3930 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003931 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003932 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003933 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003934 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003935 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003936 cachep->nodelists[node] = l3;
3937 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003938 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003939
Andrew Mortona737b3e2006-03-22 00:08:11 -08003940fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003941 if (!cachep->next.next) {
3942 /* Cache is not active yet. Roll back what we did */
3943 node--;
3944 while (node >= 0) {
3945 if (cachep->nodelists[node]) {
3946 l3 = cachep->nodelists[node];
3947
3948 kfree(l3->shared);
3949 free_alien_cache(l3->alien);
3950 kfree(l3);
3951 cachep->nodelists[node] = NULL;
3952 }
3953 node--;
3954 }
3955 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003956 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003957}
3958
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003960 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 struct array_cache *new[NR_CPUS];
3962};
3963
3964static void do_ccupdate_local(void *info)
3965{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003966 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 struct array_cache *old;
3968
3969 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003970 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003971
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3973 new->new[smp_processor_id()] = old;
3974}
3975
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003976/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003977static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003978 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003980 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003981 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982
Pekka Enberg83b519e2009-06-10 19:40:04 +03003983 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003984 if (!new)
3985 return -ENOMEM;
3986
Christoph Lametere498be72005-09-09 13:03:32 -07003987 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003988 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003989 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003990 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003991 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003992 kfree(new->new[i]);
3993 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003994 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 }
3996 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003997 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003999 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004000
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 cachep->batchcount = batchcount;
4003 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004004 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005
Christoph Lametere498be72005-09-09 13:03:32 -07004006 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004007 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 if (!ccold)
4009 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004010 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4011 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4012 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 kfree(ccold);
4014 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004015 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004016 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017}
4018
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004019/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004020static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021{
4022 int err;
4023 int limit, shared;
4024
Andrew Mortona737b3e2006-03-22 00:08:11 -08004025 /*
4026 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027 * - create a LIFO ordering, i.e. return objects that are cache-warm
4028 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004029 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 * bufctl chains: array operations are cheaper.
4031 * The numbers are guessed, we should auto-tune as described by
4032 * Bonwick.
4033 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004034 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004036 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004038 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004040 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041 limit = 54;
4042 else
4043 limit = 120;
4044
Andrew Mortona737b3e2006-03-22 00:08:11 -08004045 /*
4046 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047 * allocation behaviour: Most allocs on one cpu, most free operations
4048 * on another cpu. For these cases, an efficient object passing between
4049 * cpus is necessary. This is provided by a shared array. The array
4050 * replaces Bonwick's magazine layer.
4051 * On uniprocessor, it's functionally equivalent (but less efficient)
4052 * to a larger limit. Thus disabled by default.
4053 */
4054 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004055 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057
4058#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004059 /*
4060 * With debugging enabled, large batchcount lead to excessively long
4061 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062 */
4063 if (limit > 32)
4064 limit = 32;
4065#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004066 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 if (err)
4068 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004069 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004070 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071}
4072
Christoph Lameter1b552532006-03-22 00:09:07 -08004073/*
4074 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004075 * necessary. Note that the l3 listlock also protects the array_cache
4076 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004077 */
4078void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4079 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080{
4081 int tofree;
4082
Christoph Lameter1b552532006-03-22 00:09:07 -08004083 if (!ac || !ac->avail)
4084 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085 if (ac->touched && !force) {
4086 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004087 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004088 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004089 if (ac->avail) {
4090 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4091 if (tofree > ac->avail)
4092 tofree = (ac->avail + 1) / 2;
4093 free_block(cachep, ac->entry, tofree, node);
4094 ac->avail -= tofree;
4095 memmove(ac->entry, &(ac->entry[tofree]),
4096 sizeof(void *) * ac->avail);
4097 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004098 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099 }
4100}
4101
4102/**
4103 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004104 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105 *
4106 * Called from workqueue/eventd every few seconds.
4107 * Purpose:
4108 * - clear the per-cpu caches for this CPU.
4109 * - return freeable pages to the main free memory pool.
4110 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004111 * If we cannot acquire the cache chain mutex then just give up - we'll try
4112 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004114static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004116 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004117 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004118 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004119 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004121 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004123 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004125 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126 check_irq_on();
4127
Christoph Lameter35386e32006-03-22 00:09:05 -08004128 /*
4129 * We only take the l3 lock if absolutely necessary and we
4130 * have established with reasonable certainty that
4131 * we can do some work if the lock was obtained.
4132 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004133 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004134
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004135 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136
Christoph Lameteraab22072006-03-22 00:09:06 -08004137 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138
Christoph Lameter35386e32006-03-22 00:09:05 -08004139 /*
4140 * These are racy checks but it does not matter
4141 * if we skip one check or scan twice.
4142 */
Christoph Lametere498be72005-09-09 13:03:32 -07004143 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004144 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
Christoph Lametere498be72005-09-09 13:03:32 -07004146 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147
Christoph Lameteraab22072006-03-22 00:09:06 -08004148 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149
Christoph Lametered11d9e2006-06-30 01:55:45 -07004150 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004151 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004152 else {
4153 int freed;
4154
4155 freed = drain_freelist(searchp, l3, (l3->free_limit +
4156 5 * searchp->num - 1) / (5 * searchp->num));
4157 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004159next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 cond_resched();
4161 }
4162 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004163 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004164 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004165out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004166 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004167 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004168}
4169
Linus Torvalds158a9622008-01-02 13:04:48 -08004170#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171
Pekka Enberg85289f92006-01-08 01:00:36 -08004172static void print_slabinfo_header(struct seq_file *m)
4173{
4174 /*
4175 * Output format version, so at least we can change it
4176 * without _too_ many complaints.
4177 */
4178#if STATS
4179 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4180#else
4181 seq_puts(m, "slabinfo - version: 2.1\n");
4182#endif
4183 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4184 "<objperslab> <pagesperslab>");
4185 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4186 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4187#if STATS
4188 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004189 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004190 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4191#endif
4192 seq_putc(m, '\n');
4193}
4194
Linus Torvalds1da177e2005-04-16 15:20:36 -07004195static void *s_start(struct seq_file *m, loff_t *pos)
4196{
4197 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004199 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004200 if (!n)
4201 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004202
4203 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204}
4205
4206static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4207{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004208 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209}
4210
4211static void s_stop(struct seq_file *m, void *p)
4212{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004213 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214}
4215
4216static int s_show(struct seq_file *m, void *p)
4217{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004218 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004219 struct slab *slabp;
4220 unsigned long active_objs;
4221 unsigned long num_objs;
4222 unsigned long active_slabs = 0;
4223 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004224 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004226 int node;
4227 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004228
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 active_objs = 0;
4230 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004231 for_each_online_node(node) {
4232 l3 = cachep->nodelists[node];
4233 if (!l3)
4234 continue;
4235
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004236 check_irq_on();
4237 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004238
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004239 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004240 if (slabp->inuse != cachep->num && !error)
4241 error = "slabs_full accounting error";
4242 active_objs += cachep->num;
4243 active_slabs++;
4244 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004245 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004246 if (slabp->inuse == cachep->num && !error)
4247 error = "slabs_partial inuse accounting error";
4248 if (!slabp->inuse && !error)
4249 error = "slabs_partial/inuse accounting error";
4250 active_objs += slabp->inuse;
4251 active_slabs++;
4252 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004253 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004254 if (slabp->inuse && !error)
4255 error = "slabs_free/inuse accounting error";
4256 num_slabs++;
4257 }
4258 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004259 if (l3->shared)
4260 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004261
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004262 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004264 num_slabs += active_slabs;
4265 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004266 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267 error = "free_objects accounting error";
4268
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004269 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270 if (error)
4271 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4272
4273 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004274 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004275 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004276 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004277 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004278 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004279 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004281 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282 unsigned long high = cachep->high_mark;
4283 unsigned long allocs = cachep->num_allocations;
4284 unsigned long grown = cachep->grown;
4285 unsigned long reaped = cachep->reaped;
4286 unsigned long errors = cachep->errors;
4287 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004289 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004290 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291
Joe Perchese92dd4f2010-03-26 19:27:58 -07004292 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4293 "%4lu %4lu %4lu %4lu %4lu",
4294 allocs, high, grown,
4295 reaped, errors, max_freeable, node_allocs,
4296 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297 }
4298 /* cpu stats */
4299 {
4300 unsigned long allochit = atomic_read(&cachep->allochit);
4301 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4302 unsigned long freehit = atomic_read(&cachep->freehit);
4303 unsigned long freemiss = atomic_read(&cachep->freemiss);
4304
4305 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004306 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307 }
4308#endif
4309 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 return 0;
4311}
4312
4313/*
4314 * slabinfo_op - iterator that generates /proc/slabinfo
4315 *
4316 * Output layout:
4317 * cache-name
4318 * num-active-objs
4319 * total-objs
4320 * object size
4321 * num-active-slabs
4322 * total-slabs
4323 * num-pages-per-slab
4324 * + further values on SMP and with statistics enabled
4325 */
4326
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004327static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004328 .start = s_start,
4329 .next = s_next,
4330 .stop = s_stop,
4331 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332};
4333
4334#define MAX_SLABINFO_WRITE 128
4335/**
4336 * slabinfo_write - Tuning for the slab allocator
4337 * @file: unused
4338 * @buffer: user buffer
4339 * @count: data length
4340 * @ppos: unused
4341 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004342ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4343 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004344{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004345 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004346 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004347 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004348
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349 if (count > MAX_SLABINFO_WRITE)
4350 return -EINVAL;
4351 if (copy_from_user(&kbuf, buffer, count))
4352 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004353 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354
4355 tmp = strchr(kbuf, ' ');
4356 if (!tmp)
4357 return -EINVAL;
4358 *tmp = '\0';
4359 tmp++;
4360 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4361 return -EINVAL;
4362
4363 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004364 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004365 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004366 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004368 if (limit < 1 || batchcount < 1 ||
4369 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004370 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004371 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004372 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004373 batchcount, shared,
4374 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375 }
4376 break;
4377 }
4378 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004379 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004380 if (res >= 0)
4381 res = count;
4382 return res;
4383}
Al Viro871751e2006-03-25 03:06:39 -08004384
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004385static int slabinfo_open(struct inode *inode, struct file *file)
4386{
4387 return seq_open(file, &slabinfo_op);
4388}
4389
4390static const struct file_operations proc_slabinfo_operations = {
4391 .open = slabinfo_open,
4392 .read = seq_read,
4393 .write = slabinfo_write,
4394 .llseek = seq_lseek,
4395 .release = seq_release,
4396};
4397
Al Viro871751e2006-03-25 03:06:39 -08004398#ifdef CONFIG_DEBUG_SLAB_LEAK
4399
4400static void *leaks_start(struct seq_file *m, loff_t *pos)
4401{
Al Viro871751e2006-03-25 03:06:39 -08004402 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004403 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004404}
4405
4406static inline int add_caller(unsigned long *n, unsigned long v)
4407{
4408 unsigned long *p;
4409 int l;
4410 if (!v)
4411 return 1;
4412 l = n[1];
4413 p = n + 2;
4414 while (l) {
4415 int i = l/2;
4416 unsigned long *q = p + 2 * i;
4417 if (*q == v) {
4418 q[1]++;
4419 return 1;
4420 }
4421 if (*q > v) {
4422 l = i;
4423 } else {
4424 p = q + 2;
4425 l -= i + 1;
4426 }
4427 }
4428 if (++n[1] == n[0])
4429 return 0;
4430 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4431 p[0] = v;
4432 p[1] = 1;
4433 return 1;
4434}
4435
4436static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4437{
4438 void *p;
4439 int i;
4440 if (n[0] == n[1])
4441 return;
4442 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4443 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4444 continue;
4445 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4446 return;
4447 }
4448}
4449
4450static void show_symbol(struct seq_file *m, unsigned long address)
4451{
4452#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004453 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004454 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004455
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004456 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004457 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004458 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004459 seq_printf(m, " [%s]", modname);
4460 return;
4461 }
4462#endif
4463 seq_printf(m, "%p", (void *)address);
4464}
4465
4466static int leaks_show(struct seq_file *m, void *p)
4467{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004468 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004469 struct slab *slabp;
4470 struct kmem_list3 *l3;
4471 const char *name;
4472 unsigned long *n = m->private;
4473 int node;
4474 int i;
4475
4476 if (!(cachep->flags & SLAB_STORE_USER))
4477 return 0;
4478 if (!(cachep->flags & SLAB_RED_ZONE))
4479 return 0;
4480
4481 /* OK, we can do it */
4482
4483 n[1] = 0;
4484
4485 for_each_online_node(node) {
4486 l3 = cachep->nodelists[node];
4487 if (!l3)
4488 continue;
4489
4490 check_irq_on();
4491 spin_lock_irq(&l3->list_lock);
4492
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004493 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004494 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004495 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004496 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004497 spin_unlock_irq(&l3->list_lock);
4498 }
4499 name = cachep->name;
4500 if (n[0] == n[1]) {
4501 /* Increase the buffer size */
4502 mutex_unlock(&cache_chain_mutex);
4503 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4504 if (!m->private) {
4505 /* Too bad, we are really out */
4506 m->private = n;
4507 mutex_lock(&cache_chain_mutex);
4508 return -ENOMEM;
4509 }
4510 *(unsigned long *)m->private = n[0] * 2;
4511 kfree(n);
4512 mutex_lock(&cache_chain_mutex);
4513 /* Now make sure this entry will be retried */
4514 m->count = m->size;
4515 return 0;
4516 }
4517 for (i = 0; i < n[1]; i++) {
4518 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4519 show_symbol(m, n[2*i+2]);
4520 seq_putc(m, '\n');
4521 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004522
Al Viro871751e2006-03-25 03:06:39 -08004523 return 0;
4524}
4525
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004526static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004527 .start = leaks_start,
4528 .next = s_next,
4529 .stop = s_stop,
4530 .show = leaks_show,
4531};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004532
4533static int slabstats_open(struct inode *inode, struct file *file)
4534{
4535 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4536 int ret = -ENOMEM;
4537 if (n) {
4538 ret = seq_open(file, &slabstats_op);
4539 if (!ret) {
4540 struct seq_file *m = file->private_data;
4541 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4542 m->private = n;
4543 n = NULL;
4544 }
4545 kfree(n);
4546 }
4547 return ret;
4548}
4549
4550static const struct file_operations proc_slabstats_operations = {
4551 .open = slabstats_open,
4552 .read = seq_read,
4553 .llseek = seq_lseek,
4554 .release = seq_release_private,
4555};
Al Viro871751e2006-03-25 03:06:39 -08004556#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004557
4558static int __init slab_proc_init(void)
4559{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004560 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004561#ifdef CONFIG_DEBUG_SLAB_LEAK
4562 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4563#endif
4564 return 0;
4565}
4566module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004567#endif
4568
Manfred Spraul00e145b2005-09-03 15:55:07 -07004569/**
4570 * ksize - get the actual amount of memory allocated for a given object
4571 * @objp: Pointer to the object
4572 *
4573 * kmalloc may internally round up allocations and return more memory
4574 * than requested. ksize() can be used to determine the actual amount of
4575 * memory allocated. The caller may use this additional memory, even though
4576 * a smaller amount of memory was initially specified with the kmalloc call.
4577 * The caller must guarantee that objp points to a valid object previously
4578 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4579 * must not be freed during the duration of the call.
4580 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004581size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004582{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004583 BUG_ON(!objp);
4584 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004585 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004586
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004587 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004588}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004589EXPORT_SYMBOL(ksize);