blob: e901a36e2520c2b7aa5123ddcd0553415b764397 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120#include <asm/cacheflush.h>
121#include <asm/tlbflush.h>
122#include <asm/page.h>
123
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500124#include <trace/events/kmem.h>
125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700127 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * STATS - 1 to collect stats for /proc/slabinfo.
131 * 0 for faster, smaller code (especially in the critical paths).
132 *
133 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
134 */
135
136#ifdef CONFIG_DEBUG_SLAB
137#define DEBUG 1
138#define STATS 1
139#define FORCED_DEBUG 1
140#else
141#define DEBUG 0
142#define STATS 0
143#define FORCED_DEBUG 0
144#endif
145
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146/* Shouldn't this be in a header file somewhere? */
147#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400148#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150#ifndef ARCH_KMALLOC_FLAGS
151#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
152#endif
153
154/* Legal flag mask for kmem_cache_create(). */
155#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700156# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800158 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700159 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700161 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200162 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800164# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700165 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700167 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200168 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169#endif
170
171/*
172 * kmem_bufctl_t:
173 *
174 * Bufctl's are used for linking objs within a slab
175 * linked offsets.
176 *
177 * This implementation relies on "struct page" for locating the cache &
178 * slab an object belongs to.
179 * This allows the bufctl structure to be small (one int), but limits
180 * the number of objects a slab (not a cache) can contain when off-slab
181 * bufctls are used. The limit is the size of the largest general cache
182 * that does not use off-slab slabs.
183 * For 32bit archs with 4 kB pages, is this 56.
184 * This is not serious, as it is only for large objects, when it is unwise
185 * to have too many per slab.
186 * Note: This limit can be raised by introducing a general cache whose size
187 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
188 */
189
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700190typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
192#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800193#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
194#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197 * struct slab_rcu
198 *
199 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
200 * arrange for kmem_freepages to be called via RCU. This is useful if
201 * we need to approach a kernel structure obliquely, from its address
202 * obtained without the usual locking. We can lock the structure to
203 * stabilize it and check it's still at the given address, only if we
204 * can be sure that the memory has not been meanwhile reused for some
205 * other kind of object (which our subsystem's lock might corrupt).
206 *
207 * rcu_read_lock before reading the address, then rcu_read_unlock after
208 * taking the spinlock within the structure expected at that address.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 */
210struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800211 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800212 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800213 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214};
215
216/*
Lai Jiangshan5bfe53a2011-03-10 15:22:24 +0800217 * struct slab
218 *
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 */
223struct slab {
224 union {
225 struct {
226 struct list_head list;
227 unsigned long colouroff;
228 void *s_mem; /* including colour offset */
229 unsigned int inuse; /* num of objs active in slab */
230 kmem_bufctl_t free;
231 unsigned short nodeid;
232 };
233 struct slab_rcu __slab_cover_slab_rcu;
234 };
235};
236
237/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 * struct array_cache
239 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 * Purpose:
241 * - LIFO ordering, to hand out cache-warm objects from _alloc
242 * - reduce the number of linked list operations
243 * - reduce spinlock operations
244 *
245 * The limit is stored in the per-cpu structure to reduce the data cache
246 * footprint.
247 *
248 */
249struct array_cache {
250 unsigned int avail;
251 unsigned int limit;
252 unsigned int batchcount;
253 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700254 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700255 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800256 * Must have this definition in here for the proper
257 * alignment of array_cache. Also simplifies accessing
258 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800259 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260};
261
Andrew Mortona737b3e2006-03-22 00:08:11 -0800262/*
263 * bootstrap: The caches do not work without cpuarrays anymore, but the
264 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 */
266#define BOOT_CPUCACHE_ENTRIES 1
267struct arraycache_init {
268 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800269 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270};
271
272/*
Christoph Lametere498be72005-09-09 13:03:32 -0700273 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 */
275struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800276 struct list_head slabs_partial; /* partial list first, better asm code */
277 struct list_head slabs_full;
278 struct list_head slabs_free;
279 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800280 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800281 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800282 spinlock_t list_lock;
283 struct array_cache *shared; /* shared per node */
284 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800285 unsigned long next_reap; /* updated without locking */
286 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287};
288
Christoph Lametere498be72005-09-09 13:03:32 -0700289/*
290 * Need this for bootstrapping a per node allocator.
291 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200292#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
H Hartley Sweeten68a1b192011-01-11 17:49:32 -0600293static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700294#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200295#define SIZE_AC MAX_NUMNODES
296#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297
Christoph Lametered11d9e2006-06-30 01:55:45 -0700298static int drain_freelist(struct kmem_cache *cache,
299 struct kmem_list3 *l3, int tofree);
300static void free_block(struct kmem_cache *cachep, void **objpp, int len,
301 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300302static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000303static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700304
Christoph Lametere498be72005-09-09 13:03:32 -0700305/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800306 * This function must be completely optimized away if a constant is passed to
307 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700308 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700309static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700310{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800311 extern void __bad_size(void);
312
Christoph Lametere498be72005-09-09 13:03:32 -0700313 if (__builtin_constant_p(size)) {
314 int i = 0;
315
316#define CACHE(x) \
317 if (size <=x) \
318 return i; \
319 else \
320 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800321#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700322#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800323 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800325 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700326 return 0;
327}
328
Ingo Molnare0a42722006-06-23 02:03:46 -0700329static int slab_early_init = 1;
330
Christoph Lametere498be72005-09-09 13:03:32 -0700331#define INDEX_AC index_of(sizeof(struct arraycache_init))
332#define INDEX_L3 index_of(sizeof(struct kmem_list3))
333
Pekka Enberg5295a742006-02-01 03:05:48 -0800334static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700335{
336 INIT_LIST_HEAD(&parent->slabs_full);
337 INIT_LIST_HEAD(&parent->slabs_partial);
338 INIT_LIST_HEAD(&parent->slabs_free);
339 parent->shared = NULL;
340 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800341 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700342 spin_lock_init(&parent->list_lock);
343 parent->free_objects = 0;
344 parent->free_touched = 0;
345}
346
Andrew Mortona737b3e2006-03-22 00:08:11 -0800347#define MAKE_LIST(cachep, listp, slab, nodeid) \
348 do { \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700351 } while (0)
352
Andrew Mortona737b3e2006-03-22 00:08:11 -0800353#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
354 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
358 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360#define CFLGS_OFF_SLAB (0x80000000UL)
361#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
362
363#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800364/*
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100368 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 * which could lock up otherwise freeable slabs.
370 */
371#define REAPTIMEOUT_CPUC (2*HZ)
372#define REAPTIMEOUT_LIST3 (4*HZ)
373
374#if STATS
375#define STATS_INC_ACTIVE(x) ((x)->num_active++)
376#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700379#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800380#define STATS_SET_HIGH(x) \
381 do { \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
384 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385#define STATS_INC_ERR(x) ((x)->errors++)
386#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700387#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700388#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800389#define STATS_SET_FREEABLE(x, i) \
390 do { \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
393 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
398#else
399#define STATS_INC_ACTIVE(x) do { } while (0)
400#define STATS_DEC_ACTIVE(x) do { } while (0)
401#define STATS_INC_ALLOCED(x) do { } while (0)
402#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700403#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404#define STATS_SET_HIGH(x) do { } while (0)
405#define STATS_INC_ERR(x) do { } while (0)
406#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700407#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700408#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800409#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410#define STATS_INC_ALLOCHIT(x) do { } while (0)
411#define STATS_INC_ALLOCMISS(x) do { } while (0)
412#define STATS_INC_FREEHIT(x) do { } while (0)
413#define STATS_INC_FREEMISS(x) do { } while (0)
414#endif
415
416#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
Andrew Mortona737b3e2006-03-22 00:08:11 -0800418/*
419 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800426 * cachep->obj_offset: The real object.
427 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800428 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800431static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800433 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
Pekka Enberg343e0d72006-02-01 03:05:50 -0800436static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800438 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439}
440
David Woodhouseb46b8f12007-05-08 00:22:59 -0700441static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442{
443 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700444 return (unsigned long long*) (objp + obj_offset(cachep) -
445 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446}
447
David Woodhouseb46b8f12007-05-08 00:22:59 -0700448static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
450 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
451 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700452 return (unsigned long long *)(objp + cachep->buffer_size -
453 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400454 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700455 return (unsigned long long *) (objp + cachep->buffer_size -
456 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
Pekka Enberg343e0d72006-02-01 03:05:50 -0800459static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800462 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463}
464
465#else
466
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800467#define obj_offset(x) 0
468#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700469#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
470#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
472
473#endif
474
Li Zefan0f24f122009-12-11 15:45:30 +0800475#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300476size_t slab_buffer_size(struct kmem_cache *cachep)
477{
478 return cachep->buffer_size;
479}
480EXPORT_SYMBOL(slab_buffer_size);
481#endif
482
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700484 * Do not go above this order unless 0 objects fit into the slab or
485 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 */
David Rientjes543585c2011-10-18 22:09:24 -0700487#define SLAB_MAX_ORDER_HI 1
488#define SLAB_MAX_ORDER_LO 0
489static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700490static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491
Andrew Mortona737b3e2006-03-22 00:08:11 -0800492/*
493 * Functions for storing/retrieving the cachep and or slab from the page
494 * allocator. These are used to find the slab an obj belongs to. With kfree(),
495 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800497static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
498{
499 page->lru.next = (struct list_head *)cache;
500}
501
502static inline struct kmem_cache *page_get_cache(struct page *page)
503{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700504 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700505 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800506 return (struct kmem_cache *)page->lru.next;
507}
508
509static inline void page_set_slab(struct page *page, struct slab *slab)
510{
511 page->lru.prev = (struct list_head *)slab;
512}
513
514static inline struct slab *page_get_slab(struct page *page)
515{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700516 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800517 return (struct slab *)page->lru.prev;
518}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800520static inline struct kmem_cache *virt_to_cache(const void *obj)
521{
Christoph Lameterb49af682007-05-06 14:49:41 -0700522 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800523 return page_get_cache(page);
524}
525
526static inline struct slab *virt_to_slab(const void *obj)
527{
Christoph Lameterb49af682007-05-06 14:49:41 -0700528 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800529 return page_get_slab(page);
530}
531
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800532static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
533 unsigned int idx)
534{
535 return slab->s_mem + cache->buffer_size * idx;
536}
537
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800538/*
539 * We want to avoid an expensive divide : (offset / cache->buffer_size)
540 * Using the fact that buffer_size is a constant for a particular cache,
541 * we can replace (offset / cache->buffer_size) by
542 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
543 */
544static inline unsigned int obj_to_index(const struct kmem_cache *cache,
545 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800546{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800547 u32 offset = (obj - slab->s_mem);
548 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800549}
550
Andrew Mortona737b3e2006-03-22 00:08:11 -0800551/*
552 * These are the default caches for kmalloc. Custom caches can have other sizes.
553 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554struct cache_sizes malloc_sizes[] = {
555#define CACHE(x) { .cs_size = (x) },
556#include <linux/kmalloc_sizes.h>
557 CACHE(ULONG_MAX)
558#undef CACHE
559};
560EXPORT_SYMBOL(malloc_sizes);
561
562/* Must match cache_sizes above. Out of line to keep cache footprint low. */
563struct cache_names {
564 char *name;
565 char *name_dma;
566};
567
568static struct cache_names __initdata cache_names[] = {
569#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
570#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800571 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572#undef CACHE
573};
574
575static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800576 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800578 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580/* internal cache of cache description objs */
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200581static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
Pekka Enberg343e0d72006-02-01 03:05:50 -0800582static struct kmem_cache cache_cache = {
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200583 .nodelists = cache_cache_nodelists,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800584 .batchcount = 1,
585 .limit = BOOT_CPUCACHE_ENTRIES,
586 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800587 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800588 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589};
590
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700591#define BAD_ALIEN_MAGIC 0x01020304ul
592
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 * chicken and egg problem: delay the per-cpu array allocation
595 * until the general caches are up.
596 */
597static enum {
598 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700599 PARTIAL_AC,
600 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300601 EARLY,
Peter Zijlstra52cef182011-11-28 21:12:40 +0100602 LATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 FULL
604} g_cpucache_up;
605
Mike Kravetz39d24e62006-05-15 09:44:13 -0700606/*
607 * used by boot code to determine if it can use slab based allocator
608 */
609int slab_is_available(void)
610{
Pekka Enberg8429db52009-06-12 15:58:59 +0300611 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700612}
613
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200614#ifdef CONFIG_LOCKDEP
615
616/*
617 * Slab sometimes uses the kmalloc slabs to store the slab headers
618 * for other slabs "off slab".
619 * The locking for this is tricky in that it nests within the locks
620 * of all other slabs in a few places; to deal with this special
621 * locking we put on-slab caches into a separate lock-class.
622 *
623 * We set lock class for alien array caches which are up during init.
624 * The lock annotation will be lost if all cpus of a node goes down and
625 * then comes back up during hotplug
626 */
627static struct lock_class_key on_slab_l3_key;
628static struct lock_class_key on_slab_alc_key;
629
Peter Zijlstra83835b32011-07-22 15:26:05 +0200630static struct lock_class_key debugobj_l3_key;
631static struct lock_class_key debugobj_alc_key;
632
633static void slab_set_lock_classes(struct kmem_cache *cachep,
634 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
635 int q)
636{
637 struct array_cache **alc;
638 struct kmem_list3 *l3;
639 int r;
640
641 l3 = cachep->nodelists[q];
642 if (!l3)
643 return;
644
645 lockdep_set_class(&l3->list_lock, l3_key);
646 alc = l3->alien;
647 /*
648 * FIXME: This check for BAD_ALIEN_MAGIC
649 * should go away when common slab code is taught to
650 * work even without alien caches.
651 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
652 * for alloc_alien_cache,
653 */
654 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
655 return;
656 for_each_node(r) {
657 if (alc[r])
658 lockdep_set_class(&alc[r]->lock, alc_key);
659 }
660}
661
662static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
663{
664 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
665}
666
667static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
668{
669 int node;
670
671 for_each_online_node(node)
672 slab_set_debugobj_lock_classes_node(cachep, node);
673}
674
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200675static void init_node_lock_keys(int q)
676{
677 struct cache_sizes *s = malloc_sizes;
678
Peter Zijlstra52cef182011-11-28 21:12:40 +0100679 if (g_cpucache_up < LATE)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200680 return;
681
682 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200683 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200684
685 l3 = s->cs_cachep->nodelists[q];
686 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200687 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200688
689 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
690 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200691 }
692}
693
694static inline void init_lock_keys(void)
695{
696 int node;
697
698 for_each_node(node)
699 init_node_lock_keys(node);
700}
701#else
702static void init_node_lock_keys(int q)
703{
704}
705
706static inline void init_lock_keys(void)
707{
708}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200709
710static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
711{
712}
713
714static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
715{
716}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200717#endif
718
719/*
720 * Guard access to the cache-chain.
721 */
722static DEFINE_MUTEX(cache_chain_mutex);
723static struct list_head cache_chain;
724
Tejun Heo1871e522009-10-29 22:34:13 +0900725static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726
Pekka Enberg343e0d72006-02-01 03:05:50 -0800727static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728{
729 return cachep->array[smp_processor_id()];
730}
731
Andrew Mortona737b3e2006-03-22 00:08:11 -0800732static inline struct kmem_cache *__find_general_cachep(size_t size,
733 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734{
735 struct cache_sizes *csizep = malloc_sizes;
736
737#if DEBUG
738 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800739 * kmem_cache_create(), or __kmalloc(), before
740 * the generic caches are initialized.
741 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700742 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700744 if (!size)
745 return ZERO_SIZE_PTR;
746
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 while (size > csizep->cs_size)
748 csizep++;
749
750 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700751 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 * has cs_{dma,}cachep==NULL. Thus no special case
753 * for large kmalloc calls required.
754 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800755#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 if (unlikely(gfpflags & GFP_DMA))
757 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800758#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 return csizep->cs_cachep;
760}
761
Adrian Bunkb2213852006-09-25 23:31:02 -0700762static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700763{
764 return __find_general_cachep(size, gfpflags);
765}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700766
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800767static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800769 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
770}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771
Andrew Mortona737b3e2006-03-22 00:08:11 -0800772/*
773 * Calculate the number of objects and left-over bytes for a given buffer size.
774 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800775static void cache_estimate(unsigned long gfporder, size_t buffer_size,
776 size_t align, int flags, size_t *left_over,
777 unsigned int *num)
778{
779 int nr_objs;
780 size_t mgmt_size;
781 size_t slab_size = PAGE_SIZE << gfporder;
782
783 /*
784 * The slab management structure can be either off the slab or
785 * on it. For the latter case, the memory allocated for a
786 * slab is used for:
787 *
788 * - The struct slab
789 * - One kmem_bufctl_t for each object
790 * - Padding to respect alignment of @align
791 * - @buffer_size bytes for each object
792 *
793 * If the slab management structure is off the slab, then the
794 * alignment will already be calculated into the size. Because
795 * the slabs are all pages aligned, the objects will be at the
796 * correct alignment when allocated.
797 */
798 if (flags & CFLGS_OFF_SLAB) {
799 mgmt_size = 0;
800 nr_objs = slab_size / buffer_size;
801
802 if (nr_objs > SLAB_LIMIT)
803 nr_objs = SLAB_LIMIT;
804 } else {
805 /*
806 * Ignore padding for the initial guess. The padding
807 * is at most @align-1 bytes, and @buffer_size is at
808 * least @align. In the worst case, this result will
809 * be one greater than the number of objects that fit
810 * into the memory allocation when taking the padding
811 * into account.
812 */
813 nr_objs = (slab_size - sizeof(struct slab)) /
814 (buffer_size + sizeof(kmem_bufctl_t));
815
816 /*
817 * This calculated number will be either the right
818 * amount, or one greater than what we want.
819 */
820 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
821 > slab_size)
822 nr_objs--;
823
824 if (nr_objs > SLAB_LIMIT)
825 nr_objs = SLAB_LIMIT;
826
827 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800829 *num = nr_objs;
830 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831}
832
Harvey Harrisond40cee22008-04-30 00:55:07 -0700833#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834
Andrew Mortona737b3e2006-03-22 00:08:11 -0800835static void __slab_error(const char *function, struct kmem_cache *cachep,
836 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837{
838 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800839 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840 dump_stack();
841}
842
Paul Menage3395ee02006-12-06 20:32:16 -0800843/*
844 * By default on NUMA we use alien caches to stage the freeing of
845 * objects allocated from other nodes. This causes massive memory
846 * inefficiencies when using fake NUMA setup to split memory into a
847 * large number of small nodes, so it can be disabled on the command
848 * line
849 */
850
851static int use_alien_caches __read_mostly = 1;
852static int __init noaliencache_setup(char *s)
853{
854 use_alien_caches = 0;
855 return 1;
856}
857__setup("noaliencache", noaliencache_setup);
858
David Rientjes3df1ccc2011-10-18 22:09:28 -0700859static int __init slab_max_order_setup(char *str)
860{
861 get_option(&str, &slab_max_order);
862 slab_max_order = slab_max_order < 0 ? 0 :
863 min(slab_max_order, MAX_ORDER - 1);
864 slab_max_order_set = true;
865
866 return 1;
867}
868__setup("slab_max_order=", slab_max_order_setup);
869
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800870#ifdef CONFIG_NUMA
871/*
872 * Special reaping functions for NUMA systems called from cache_reap().
873 * These take care of doing round robin flushing of alien caches (containing
874 * objects freed on different nodes from which they were allocated) and the
875 * flushing of remote pcps by calling drain_node_pages.
876 */
Tejun Heo1871e522009-10-29 22:34:13 +0900877static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878
879static void init_reap_node(int cpu)
880{
881 int node;
882
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700883 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800884 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800885 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800886
Tejun Heo1871e522009-10-29 22:34:13 +0900887 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800888}
889
890static void next_reap_node(void)
891{
Christoph Lameter909ea962010-12-08 16:22:55 +0100892 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800893
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800894 node = next_node(node, node_online_map);
895 if (unlikely(node >= MAX_NUMNODES))
896 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100897 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800898}
899
900#else
901#define init_reap_node(cpu) do { } while (0)
902#define next_reap_node(void) do { } while (0)
903#endif
904
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905/*
906 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
907 * via the workqueue/eventd.
908 * Add the CPU number into the expiration time to minimize the possibility of
909 * the CPUs getting into lockstep and contending for the global cache chain
910 * lock.
911 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700912static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913{
Tejun Heo1871e522009-10-29 22:34:13 +0900914 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915
916 /*
917 * When this gets called from do_initcalls via cpucache_init(),
918 * init_workqueues() has already run, so keventd will be setup
919 * at that time.
920 */
David Howells52bad642006-11-22 14:54:01 +0000921 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800922 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700923 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800924 schedule_delayed_work_on(cpu, reap_work,
925 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 }
927}
928
Christoph Lametere498be72005-09-09 13:03:32 -0700929static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300930 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800932 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 struct array_cache *nc = NULL;
934
Pekka Enberg83b519e2009-06-10 19:40:04 +0300935 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100936 /*
937 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300938 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100939 * cache the pointers are not cleared and they could be counted as
940 * valid references during a kmemleak scan. Therefore, kmemleak must
941 * not scan such objects.
942 */
943 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 if (nc) {
945 nc->avail = 0;
946 nc->limit = entries;
947 nc->batchcount = batchcount;
948 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700949 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950 }
951 return nc;
952}
953
Christoph Lameter3ded1752006-03-25 03:06:44 -0800954/*
955 * Transfer objects in one arraycache to another.
956 * Locking must be handled by the caller.
957 *
958 * Return the number of entries transferred.
959 */
960static int transfer_objects(struct array_cache *to,
961 struct array_cache *from, unsigned int max)
962{
963 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700964 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800965
966 if (!nr)
967 return 0;
968
969 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
970 sizeof(void *) *nr);
971
972 from->avail -= nr;
973 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800974 return nr;
975}
976
Christoph Lameter765c4502006-09-27 01:50:08 -0700977#ifndef CONFIG_NUMA
978
979#define drain_alien_cache(cachep, alien) do { } while (0)
980#define reap_alien(cachep, l3) do { } while (0)
981
Pekka Enberg83b519e2009-06-10 19:40:04 +0300982static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700983{
984 return (struct array_cache **)BAD_ALIEN_MAGIC;
985}
986
987static inline void free_alien_cache(struct array_cache **ac_ptr)
988{
989}
990
991static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
992{
993 return 0;
994}
995
996static inline void *alternate_node_alloc(struct kmem_cache *cachep,
997 gfp_t flags)
998{
999 return NULL;
1000}
1001
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001002static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001003 gfp_t flags, int nodeid)
1004{
1005 return NULL;
1006}
1007
1008#else /* CONFIG_NUMA */
1009
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001010static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001011static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001012
Pekka Enberg83b519e2009-06-10 19:40:04 +03001013static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001014{
1015 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001016 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001017 int i;
1018
1019 if (limit > 1)
1020 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001021 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001022 if (ac_ptr) {
1023 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001024 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001025 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001026 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001027 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001028 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001029 kfree(ac_ptr[i]);
1030 kfree(ac_ptr);
1031 return NULL;
1032 }
1033 }
1034 }
1035 return ac_ptr;
1036}
1037
Pekka Enberg5295a742006-02-01 03:05:48 -08001038static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001039{
1040 int i;
1041
1042 if (!ac_ptr)
1043 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001044 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001045 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001046 kfree(ac_ptr);
1047}
1048
Pekka Enberg343e0d72006-02-01 03:05:50 -08001049static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001050 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001051{
1052 struct kmem_list3 *rl3 = cachep->nodelists[node];
1053
1054 if (ac->avail) {
1055 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001056 /*
1057 * Stuff objects into the remote nodes shared array first.
1058 * That way we could avoid the overhead of putting the objects
1059 * into the free lists and getting them back later.
1060 */
shin, jacob693f7d32006-04-28 10:54:37 -05001061 if (rl3->shared)
1062 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001063
Christoph Lameterff694162005-09-22 21:44:02 -07001064 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001065 ac->avail = 0;
1066 spin_unlock(&rl3->list_lock);
1067 }
1068}
1069
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001070/*
1071 * Called from cache_reap() to regularly drain alien caches round robin.
1072 */
1073static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1074{
Christoph Lameter909ea962010-12-08 16:22:55 +01001075 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001076
1077 if (l3->alien) {
1078 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001079
1080 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001081 __drain_alien_cache(cachep, ac, node);
1082 spin_unlock_irq(&ac->lock);
1083 }
1084 }
1085}
1086
Andrew Mortona737b3e2006-03-22 00:08:11 -08001087static void drain_alien_cache(struct kmem_cache *cachep,
1088 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001089{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001090 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001091 struct array_cache *ac;
1092 unsigned long flags;
1093
1094 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001095 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001096 if (ac) {
1097 spin_lock_irqsave(&ac->lock, flags);
1098 __drain_alien_cache(cachep, ac, i);
1099 spin_unlock_irqrestore(&ac->lock, flags);
1100 }
1101 }
1102}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001103
Ingo Molnar873623d2006-07-13 14:44:38 +02001104static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001105{
1106 struct slab *slabp = virt_to_slab(objp);
1107 int nodeid = slabp->nodeid;
1108 struct kmem_list3 *l3;
1109 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001110 int node;
1111
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001112 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001113
1114 /*
1115 * Make sure we are not freeing a object from another node to the array
1116 * cache on this cpu.
1117 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001118 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001119 return 0;
1120
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001121 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001122 STATS_INC_NODEFREES(cachep);
1123 if (l3->alien && l3->alien[nodeid]) {
1124 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001125 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001126 if (unlikely(alien->avail == alien->limit)) {
1127 STATS_INC_ACOVERFLOW(cachep);
1128 __drain_alien_cache(cachep, alien, nodeid);
1129 }
1130 alien->entry[alien->avail++] = objp;
1131 spin_unlock(&alien->lock);
1132 } else {
1133 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1134 free_block(cachep, &objp, 1, nodeid);
1135 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1136 }
1137 return 1;
1138}
Christoph Lametere498be72005-09-09 13:03:32 -07001139#endif
1140
David Rientjes8f9f8d92010-03-27 19:40:47 -07001141/*
1142 * Allocates and initializes nodelists for a node on each slab cache, used for
1143 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1144 * will be allocated off-node since memory is not yet online for the new node.
1145 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1146 * already in use.
1147 *
1148 * Must hold cache_chain_mutex.
1149 */
1150static int init_cache_nodelists_node(int node)
1151{
1152 struct kmem_cache *cachep;
1153 struct kmem_list3 *l3;
1154 const int memsize = sizeof(struct kmem_list3);
1155
1156 list_for_each_entry(cachep, &cache_chain, next) {
1157 /*
1158 * Set up the size64 kmemlist for cpu before we can
1159 * begin anything. Make sure some other cpu on this
1160 * node has not already allocated this
1161 */
1162 if (!cachep->nodelists[node]) {
1163 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1164 if (!l3)
1165 return -ENOMEM;
1166 kmem_list3_init(l3);
1167 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1168 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1169
1170 /*
1171 * The l3s don't come and go as CPUs come and
1172 * go. cache_chain_mutex is sufficient
1173 * protection here.
1174 */
1175 cachep->nodelists[node] = l3;
1176 }
1177
1178 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1179 cachep->nodelists[node]->free_limit =
1180 (1 + nr_cpus_node(node)) *
1181 cachep->batchcount + cachep->num;
1182 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1183 }
1184 return 0;
1185}
1186
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001189 struct kmem_cache *cachep;
1190 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001191 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301192 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193
1194 list_for_each_entry(cachep, &cache_chain, next) {
1195 struct array_cache *nc;
1196 struct array_cache *shared;
1197 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001198
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 /* cpu is dead; no one can alloc from it. */
1200 nc = cachep->array[cpu];
1201 cachep->array[cpu] = NULL;
1202 l3 = cachep->nodelists[node];
1203
1204 if (!l3)
1205 goto free_array_cache;
1206
1207 spin_lock_irq(&l3->list_lock);
1208
1209 /* Free limit for this kmem_list3 */
1210 l3->free_limit -= cachep->batchcount;
1211 if (nc)
1212 free_block(cachep, nc->entry, nc->avail, node);
1213
Rusty Russell58463c12009-12-17 11:43:12 -06001214 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001215 spin_unlock_irq(&l3->list_lock);
1216 goto free_array_cache;
1217 }
1218
1219 shared = l3->shared;
1220 if (shared) {
1221 free_block(cachep, shared->entry,
1222 shared->avail, node);
1223 l3->shared = NULL;
1224 }
1225
1226 alien = l3->alien;
1227 l3->alien = NULL;
1228
1229 spin_unlock_irq(&l3->list_lock);
1230
1231 kfree(shared);
1232 if (alien) {
1233 drain_alien_cache(cachep, alien);
1234 free_alien_cache(alien);
1235 }
1236free_array_cache:
1237 kfree(nc);
1238 }
1239 /*
1240 * In the previous loop, all the objects were freed to
1241 * the respective cache's slabs, now we can go ahead and
1242 * shrink each nodelist to its limit.
1243 */
1244 list_for_each_entry(cachep, &cache_chain, next) {
1245 l3 = cachep->nodelists[node];
1246 if (!l3)
1247 continue;
1248 drain_freelist(cachep, l3, l3->free_objects);
1249 }
1250}
1251
1252static int __cpuinit cpuup_prepare(long cpu)
1253{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001254 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001255 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001256 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001257 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001259 /*
1260 * We need to do this right in the beginning since
1261 * alloc_arraycache's are going to use this list.
1262 * kmalloc_node allows us to add the slab to the right
1263 * kmem_list3 and not this cpu's kmem_list3
1264 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001265 err = init_cache_nodelists_node(node);
1266 if (err < 0)
1267 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001268
1269 /*
1270 * Now we can go ahead with allocating the shared arrays and
1271 * array caches
1272 */
1273 list_for_each_entry(cachep, &cache_chain, next) {
1274 struct array_cache *nc;
1275 struct array_cache *shared = NULL;
1276 struct array_cache **alien = NULL;
1277
1278 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001279 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 if (!nc)
1281 goto bad;
1282 if (cachep->shared) {
1283 shared = alloc_arraycache(node,
1284 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001285 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001286 if (!shared) {
1287 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001288 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001289 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001290 }
1291 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001292 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001293 if (!alien) {
1294 kfree(shared);
1295 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001296 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001297 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001298 }
1299 cachep->array[cpu] = nc;
1300 l3 = cachep->nodelists[node];
1301 BUG_ON(!l3);
1302
1303 spin_lock_irq(&l3->list_lock);
1304 if (!l3->shared) {
1305 /*
1306 * We are serialised from CPU_DEAD or
1307 * CPU_UP_CANCELLED by the cpucontrol lock
1308 */
1309 l3->shared = shared;
1310 shared = NULL;
1311 }
1312#ifdef CONFIG_NUMA
1313 if (!l3->alien) {
1314 l3->alien = alien;
1315 alien = NULL;
1316 }
1317#endif
1318 spin_unlock_irq(&l3->list_lock);
1319 kfree(shared);
1320 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001321 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1322 slab_set_debugobj_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001323 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001324 init_node_lock_keys(node);
1325
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001326 return 0;
1327bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001328 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001329 return -ENOMEM;
1330}
1331
1332static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1333 unsigned long action, void *hcpu)
1334{
1335 long cpu = (long)hcpu;
1336 int err = 0;
1337
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001339 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001340 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001341 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001342 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001343 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 break;
1345 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001346 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 start_cpu_timer(cpu);
1348 break;
1349#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001350 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001351 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001352 /*
1353 * Shutdown cache reaper. Note that the cache_chain_mutex is
1354 * held so that if cache_reap() is invoked it cannot do
1355 * anything expensive but will only modify reap_work
1356 * and reschedule the timer.
1357 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001358 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001359 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001360 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001361 break;
1362 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001363 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001364 start_cpu_timer(cpu);
1365 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001367 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001368 /*
1369 * Even if all the cpus of a node are down, we don't free the
1370 * kmem_list3 of any cache. This to avoid a race between
1371 * cpu_down, and a kmalloc allocation from another cpu for
1372 * memory from the node of the cpu going down. The list3
1373 * structure is usually allocated from kmem_cache_create() and
1374 * gets destroyed at kmem_cache_destroy().
1375 */
Simon Arlott183ff222007-10-20 01:27:18 +02001376 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001377#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001379 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001380 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001381 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001382 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001385 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386}
1387
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001388static struct notifier_block __cpuinitdata cpucache_notifier = {
1389 &cpuup_callback, NULL, 0
1390};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
David Rientjes8f9f8d92010-03-27 19:40:47 -07001392#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1393/*
1394 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1395 * Returns -EBUSY if all objects cannot be drained so that the node is not
1396 * removed.
1397 *
1398 * Must hold cache_chain_mutex.
1399 */
1400static int __meminit drain_cache_nodelists_node(int node)
1401{
1402 struct kmem_cache *cachep;
1403 int ret = 0;
1404
1405 list_for_each_entry(cachep, &cache_chain, next) {
1406 struct kmem_list3 *l3;
1407
1408 l3 = cachep->nodelists[node];
1409 if (!l3)
1410 continue;
1411
1412 drain_freelist(cachep, l3, l3->free_objects);
1413
1414 if (!list_empty(&l3->slabs_full) ||
1415 !list_empty(&l3->slabs_partial)) {
1416 ret = -EBUSY;
1417 break;
1418 }
1419 }
1420 return ret;
1421}
1422
1423static int __meminit slab_memory_callback(struct notifier_block *self,
1424 unsigned long action, void *arg)
1425{
1426 struct memory_notify *mnb = arg;
1427 int ret = 0;
1428 int nid;
1429
1430 nid = mnb->status_change_nid;
1431 if (nid < 0)
1432 goto out;
1433
1434 switch (action) {
1435 case MEM_GOING_ONLINE:
1436 mutex_lock(&cache_chain_mutex);
1437 ret = init_cache_nodelists_node(nid);
1438 mutex_unlock(&cache_chain_mutex);
1439 break;
1440 case MEM_GOING_OFFLINE:
1441 mutex_lock(&cache_chain_mutex);
1442 ret = drain_cache_nodelists_node(nid);
1443 mutex_unlock(&cache_chain_mutex);
1444 break;
1445 case MEM_ONLINE:
1446 case MEM_OFFLINE:
1447 case MEM_CANCEL_ONLINE:
1448 case MEM_CANCEL_OFFLINE:
1449 break;
1450 }
1451out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001452 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001453}
1454#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1455
Christoph Lametere498be72005-09-09 13:03:32 -07001456/*
1457 * swap the static kmem_list3 with kmalloced memory
1458 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001459static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1460 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001461{
1462 struct kmem_list3 *ptr;
1463
Pekka Enberg83b519e2009-06-10 19:40:04 +03001464 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001465 BUG_ON(!ptr);
1466
Christoph Lametere498be72005-09-09 13:03:32 -07001467 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001468 /*
1469 * Do not assume that spinlocks can be initialized via memcpy:
1470 */
1471 spin_lock_init(&ptr->list_lock);
1472
Christoph Lametere498be72005-09-09 13:03:32 -07001473 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1474 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001475}
1476
Andrew Mortona737b3e2006-03-22 00:08:11 -08001477/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001478 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1479 * size of kmem_list3.
1480 */
1481static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1482{
1483 int node;
1484
1485 for_each_online_node(node) {
1486 cachep->nodelists[node] = &initkmem_list3[index + node];
1487 cachep->nodelists[node]->next_reap = jiffies +
1488 REAPTIMEOUT_LIST3 +
1489 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1490 }
1491}
1492
1493/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001494 * Initialisation. Called after the page allocator have been initialised and
1495 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 */
1497void __init kmem_cache_init(void)
1498{
1499 size_t left_over;
1500 struct cache_sizes *sizes;
1501 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001502 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001503 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001504 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001505
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001506 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001507 use_alien_caches = 0;
1508
Christoph Lametere498be72005-09-09 13:03:32 -07001509 for (i = 0; i < NUM_INIT_LISTS; i++) {
1510 kmem_list3_init(&initkmem_list3[i]);
1511 if (i < MAX_NUMNODES)
1512 cache_cache.nodelists[i] = NULL;
1513 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001514 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515
1516 /*
1517 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001518 * page orders on machines with more than 32MB of memory if
1519 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001521 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001522 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 /* Bootstrap is tricky, because several objects are allocated
1525 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001526 * 1) initialize the cache_cache cache: it contains the struct
1527 * kmem_cache structures of all caches, except cache_cache itself:
1528 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001529 * Initially an __init data area is used for the head array and the
1530 * kmem_list3 structures, it's replaced with a kmalloc allocated
1531 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001533 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001534 * An __init data area is used for the head array.
1535 * 3) Create the remaining kmalloc caches, with minimally sized
1536 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 * 4) Replace the __init data head arrays for cache_cache and the first
1538 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001539 * 5) Replace the __init data for kmem_list3 for cache_cache and
1540 * the other cache's with kmalloc allocated memory.
1541 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 */
1543
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001544 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001545
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 INIT_LIST_HEAD(&cache_chain);
1548 list_add(&cache_cache.next, &cache_chain);
1549 cache_cache.colour_off = cache_line_size();
1550 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001551 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
Eric Dumazet8da34302007-05-06 14:49:29 -07001553 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001554 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001555 */
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001556 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1557 nr_node_ids * sizeof(struct kmem_list3 *);
Eric Dumazet8da34302007-05-06 14:49:29 -07001558#if DEBUG
1559 cache_cache.obj_size = cache_cache.buffer_size;
1560#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001561 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1562 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001563 cache_cache.reciprocal_buffer_size =
1564 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565
Jack Steiner07ed76b2006-03-07 21:55:46 -08001566 for (order = 0; order < MAX_ORDER; order++) {
1567 cache_estimate(order, cache_cache.buffer_size,
1568 cache_line_size(), 0, &left_over, &cache_cache.num);
1569 if (cache_cache.num)
1570 break;
1571 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001572 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001573 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001574 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001575 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1576 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577
1578 /* 2+3) create the kmalloc caches */
1579 sizes = malloc_sizes;
1580 names = cache_names;
1581
Andrew Mortona737b3e2006-03-22 00:08:11 -08001582 /*
1583 * Initialize the caches that provide memory for the array cache and the
1584 * kmem_list3 structures first. Without this, further allocations will
1585 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001586 */
1587
1588 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001589 sizes[INDEX_AC].cs_size,
1590 ARCH_KMALLOC_MINALIGN,
1591 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001592 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001593
Andrew Mortona737b3e2006-03-22 00:08:11 -08001594 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001595 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 kmem_cache_create(names[INDEX_L3].name,
1597 sizes[INDEX_L3].cs_size,
1598 ARCH_KMALLOC_MINALIGN,
1599 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001600 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001601 }
Christoph Lametere498be72005-09-09 13:03:32 -07001602
Ingo Molnare0a42722006-06-23 02:03:46 -07001603 slab_early_init = 0;
1604
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001606 /*
1607 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 * This should be particularly beneficial on SMP boxes, as it
1609 * eliminates "false sharing".
1610 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001611 * allow tighter packing of the smaller caches.
1612 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001613 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001614 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001615 sizes->cs_size,
1616 ARCH_KMALLOC_MINALIGN,
1617 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001618 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001619 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001620#ifdef CONFIG_ZONE_DMA
1621 sizes->cs_dmacachep = kmem_cache_create(
1622 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001623 sizes->cs_size,
1624 ARCH_KMALLOC_MINALIGN,
1625 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1626 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001627 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001628#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 sizes++;
1630 names++;
1631 }
1632 /* 4) Replace the bootstrap head arrays */
1633 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001634 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001635
Pekka Enberg83b519e2009-06-10 19:40:04 +03001636 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001637
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001638 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1639 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001640 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001641 /*
1642 * Do not assume that spinlocks can be initialized via memcpy:
1643 */
1644 spin_lock_init(&ptr->lock);
1645
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001647
Pekka Enberg83b519e2009-06-10 19:40:04 +03001648 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001649
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001650 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001652 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001653 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001654 /*
1655 * Do not assume that spinlocks can be initialized via memcpy:
1656 */
1657 spin_lock_init(&ptr->lock);
1658
Christoph Lametere498be72005-09-09 13:03:32 -07001659 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001660 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 }
Christoph Lametere498be72005-09-09 13:03:32 -07001662 /* 5) Replace the bootstrap kmem_list3's */
1663 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001664 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Mel Gorman9c09a952008-01-24 05:49:54 -08001666 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001667 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001668
Christoph Lametere498be72005-09-09 13:03:32 -07001669 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001670 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001671
1672 if (INDEX_AC != INDEX_L3) {
1673 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001674 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001675 }
1676 }
1677 }
1678
Pekka Enberg8429db52009-06-12 15:58:59 +03001679 g_cpucache_up = EARLY;
Pekka Enberg8429db52009-06-12 15:58:59 +03001680}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001681
Pekka Enberg8429db52009-06-12 15:58:59 +03001682void __init kmem_cache_init_late(void)
1683{
1684 struct kmem_cache *cachep;
1685
Peter Zijlstra52cef182011-11-28 21:12:40 +01001686 g_cpucache_up = LATE;
1687
Peter Zijlstra30765b92011-07-28 23:22:56 +02001688 /* Annotate slab for lockdep -- annotate the malloc caches */
1689 init_lock_keys();
1690
Pekka Enberg8429db52009-06-12 15:58:59 +03001691 /* 6) resize the head arrays to their final sizes */
1692 mutex_lock(&cache_chain_mutex);
1693 list_for_each_entry(cachep, &cache_chain, next)
1694 if (enable_cpucache(cachep, GFP_NOWAIT))
1695 BUG();
1696 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001697
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 /* Done! */
1699 g_cpucache_up = FULL;
1700
Andrew Mortona737b3e2006-03-22 00:08:11 -08001701 /*
1702 * Register a cpu startup notifier callback that initializes
1703 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 */
1705 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706
David Rientjes8f9f8d92010-03-27 19:40:47 -07001707#ifdef CONFIG_NUMA
1708 /*
1709 * Register a memory hotplug callback that initializes and frees
1710 * nodelists.
1711 */
1712 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1713#endif
1714
Andrew Mortona737b3e2006-03-22 00:08:11 -08001715 /*
1716 * The reap timers are started later, with a module init call: That part
1717 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 */
1719}
1720
1721static int __init cpucache_init(void)
1722{
1723 int cpu;
1724
Andrew Mortona737b3e2006-03-22 00:08:11 -08001725 /*
1726 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 */
Christoph Lametere498be72005-09-09 13:03:32 -07001728 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001729 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 return 0;
1731}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732__initcall(cpucache_init);
1733
Rafael Aquini8bdec192012-03-09 17:27:27 -03001734static noinline void
1735slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1736{
1737 struct kmem_list3 *l3;
1738 struct slab *slabp;
1739 unsigned long flags;
1740 int node;
1741
1742 printk(KERN_WARNING
1743 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1744 nodeid, gfpflags);
1745 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1746 cachep->name, cachep->buffer_size, cachep->gfporder);
1747
1748 for_each_online_node(node) {
1749 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1750 unsigned long active_slabs = 0, num_slabs = 0;
1751
1752 l3 = cachep->nodelists[node];
1753 if (!l3)
1754 continue;
1755
1756 spin_lock_irqsave(&l3->list_lock, flags);
1757 list_for_each_entry(slabp, &l3->slabs_full, list) {
1758 active_objs += cachep->num;
1759 active_slabs++;
1760 }
1761 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1762 active_objs += slabp->inuse;
1763 active_slabs++;
1764 }
1765 list_for_each_entry(slabp, &l3->slabs_free, list)
1766 num_slabs++;
1767
1768 free_objects += l3->free_objects;
1769 spin_unlock_irqrestore(&l3->list_lock, flags);
1770
1771 num_slabs += active_slabs;
1772 num_objs = num_slabs * cachep->num;
1773 printk(KERN_WARNING
1774 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1775 node, active_slabs, num_slabs, active_objs, num_objs,
1776 free_objects);
1777 }
1778}
1779
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780/*
1781 * Interface to system's page allocator. No need to hold the cache-lock.
1782 *
1783 * If we requested dmaable memory, we will get it. Even if we
1784 * did not request dmaable memory, we might get it, but that
1785 * would be relatively rare and ignorable.
1786 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001787static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788{
1789 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001790 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 int i;
1792
Luke Yangd6fef9d2006-04-10 22:52:56 -07001793#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001794 /*
1795 * Nommu uses slab's for process anonymous memory allocations, and thus
1796 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001797 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001798 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001799#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001800
Christoph Lameter3c517a62006-12-06 20:33:29 -08001801 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001802 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1803 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001804
Linus Torvalds517d0862009-06-16 19:50:13 -07001805 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001806 if (!page) {
1807 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1808 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001812 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001814 add_zone_page_state(page_zone(page),
1815 NR_SLAB_RECLAIMABLE, nr_pages);
1816 else
1817 add_zone_page_state(page_zone(page),
1818 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001819 for (i = 0; i < nr_pages; i++)
1820 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001821
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001822 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1823 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1824
1825 if (cachep->ctor)
1826 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1827 else
1828 kmemcheck_mark_unallocated_pages(page, nr_pages);
1829 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001830
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001831 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832}
1833
1834/*
1835 * Interface to system's page release.
1836 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001837static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001839 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 struct page *page = virt_to_page(addr);
1841 const unsigned long nr_freed = i;
1842
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001843 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001844
Christoph Lameter972d1a72006-09-25 23:31:51 -07001845 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1846 sub_zone_page_state(page_zone(page),
1847 NR_SLAB_RECLAIMABLE, nr_freed);
1848 else
1849 sub_zone_page_state(page_zone(page),
1850 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001852 BUG_ON(!PageSlab(page));
1853 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 page++;
1855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 if (current->reclaim_state)
1857 current->reclaim_state->reclaimed_slab += nr_freed;
1858 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859}
1860
1861static void kmem_rcu_free(struct rcu_head *head)
1862{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001864 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865
1866 kmem_freepages(cachep, slab_rcu->addr);
1867 if (OFF_SLAB(cachep))
1868 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1869}
1870
1871#if DEBUG
1872
1873#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001874static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001877 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001879 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001881 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 return;
1883
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001884 *addr++ = 0x12345678;
1885 *addr++ = caller;
1886 *addr++ = smp_processor_id();
1887 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 {
1889 unsigned long *sptr = &caller;
1890 unsigned long svalue;
1891
1892 while (!kstack_end(sptr)) {
1893 svalue = *sptr++;
1894 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 size -= sizeof(unsigned long);
1897 if (size <= sizeof(unsigned long))
1898 break;
1899 }
1900 }
1901
1902 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001903 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904}
1905#endif
1906
Pekka Enberg343e0d72006-02-01 03:05:50 -08001907static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001909 int size = obj_size(cachep);
1910 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
1912 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001913 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914}
1915
1916static void dump_line(char *data, int offset, int limit)
1917{
1918 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001919 unsigned char error = 0;
1920 int bad_count = 0;
1921
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001922 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001923 for (i = 0; i < limit; i++) {
1924 if (data[offset + i] != POISON_FREE) {
1925 error = data[offset + i];
1926 bad_count++;
1927 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001928 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001929 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1930 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001931
1932 if (bad_count == 1) {
1933 error ^= POISON_FREE;
1934 if (!(error & (error - 1))) {
1935 printk(KERN_ERR "Single bit error detected. Probably "
1936 "bad RAM.\n");
1937#ifdef CONFIG_X86
1938 printk(KERN_ERR "Run memtest86+ or a similar memory "
1939 "test tool.\n");
1940#else
1941 printk(KERN_ERR "Run a memory test tool.\n");
1942#endif
1943 }
1944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945}
1946#endif
1947
1948#if DEBUG
1949
Pekka Enberg343e0d72006-02-01 03:05:50 -08001950static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951{
1952 int i, size;
1953 char *realobj;
1954
1955 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001956 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001957 *dbg_redzone1(cachep, objp),
1958 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 }
1960
1961 if (cachep->flags & SLAB_STORE_USER) {
1962 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001963 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001965 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 printk("\n");
1967 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001968 realobj = (char *)objp + obj_offset(cachep);
1969 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001970 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 int limit;
1972 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001973 if (i + limit > size)
1974 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 dump_line(realobj, i, limit);
1976 }
1977}
1978
Pekka Enberg343e0d72006-02-01 03:05:50 -08001979static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
1981 char *realobj;
1982 int size, i;
1983 int lines = 0;
1984
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001985 realobj = (char *)objp + obj_offset(cachep);
1986 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001988 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001990 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 exp = POISON_END;
1992 if (realobj[i] != exp) {
1993 int limit;
1994 /* Mismatch ! */
1995 /* Print header */
1996 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001997 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001998 "Slab corruption (%s): %s start=%p, len=%d\n",
1999 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 print_objinfo(cachep, objp, 0);
2001 }
2002 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002003 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002005 if (i + limit > size)
2006 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 dump_line(realobj, i, limit);
2008 i += 16;
2009 lines++;
2010 /* Limit to 5 lines */
2011 if (lines > 5)
2012 break;
2013 }
2014 }
2015 if (lines != 0) {
2016 /* Print some data about the neighboring objects, if they
2017 * exist:
2018 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002019 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002020 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002022 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002024 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002025 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002027 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 print_objinfo(cachep, objp, 2);
2029 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002030 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002031 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002032 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002034 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 print_objinfo(cachep, objp, 2);
2036 }
2037 }
2038}
2039#endif
2040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05302042static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002043{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 int i;
2045 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002046 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047
2048 if (cachep->flags & SLAB_POISON) {
2049#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002050 if (cachep->buffer_size % PAGE_SIZE == 0 &&
2051 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002052 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08002053 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054 else
2055 check_poison_obj(cachep, objp);
2056#else
2057 check_poison_obj(cachep, objp);
2058#endif
2059 }
2060 if (cachep->flags & SLAB_RED_ZONE) {
2061 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2062 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002063 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2065 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002066 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002069}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302071static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002072{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002073}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074#endif
2075
Randy Dunlap911851e2006-03-22 00:08:14 -08002076/**
2077 * slab_destroy - destroy and release all objects in a slab
2078 * @cachep: cache pointer being destroyed
2079 * @slabp: slab pointer being destroyed
2080 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002081 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002082 * Before calling the slab must have been unlinked from the cache. The
2083 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002084 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002085static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002086{
2087 void *addr = slabp->s_mem - slabp->colouroff;
2088
Rabin Vincente79aec22008-07-04 00:40:32 +05302089 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2091 struct slab_rcu *slab_rcu;
2092
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002093 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 slab_rcu->cachep = cachep;
2095 slab_rcu->addr = addr;
2096 call_rcu(&slab_rcu->head, kmem_rcu_free);
2097 } else {
2098 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002099 if (OFF_SLAB(cachep))
2100 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 }
2102}
2103
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002104static void __kmem_cache_destroy(struct kmem_cache *cachep)
2105{
2106 int i;
2107 struct kmem_list3 *l3;
2108
2109 for_each_online_cpu(i)
2110 kfree(cachep->array[i]);
2111
2112 /* NUMA: free the list3 structures */
2113 for_each_online_node(i) {
2114 l3 = cachep->nodelists[i];
2115 if (l3) {
2116 kfree(l3->shared);
2117 free_alien_cache(l3->alien);
2118 kfree(l3);
2119 }
2120 }
2121 kmem_cache_free(&cache_cache, cachep);
2122}
2123
2124
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002126 * calculate_slab_order - calculate size (page order) of slabs
2127 * @cachep: pointer to the cache that is being created
2128 * @size: size of objects to be created in this cache.
2129 * @align: required alignment for the objects.
2130 * @flags: slab allocation flags
2131 *
2132 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002133 *
2134 * This could be made much more intelligent. For now, try to avoid using
2135 * high order pages for slabs. When the gfp() functions are more friendly
2136 * towards high-order requests, this should be changed.
2137 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002138static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002139 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002140{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002141 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002142 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002143 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002144
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002145 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002146 unsigned int num;
2147 size_t remainder;
2148
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002149 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002150 if (!num)
2151 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002152
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002153 if (flags & CFLGS_OFF_SLAB) {
2154 /*
2155 * Max number of objs-per-slab for caches which
2156 * use off-slab slabs. Needed to avoid a possible
2157 * looping condition in cache_grow().
2158 */
2159 offslab_limit = size - sizeof(struct slab);
2160 offslab_limit /= sizeof(kmem_bufctl_t);
2161
2162 if (num > offslab_limit)
2163 break;
2164 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002165
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002166 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002167 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002168 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002169 left_over = remainder;
2170
2171 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002172 * A VFS-reclaimable slab tends to have most allocations
2173 * as GFP_NOFS and we really don't want to have to be allocating
2174 * higher-order pages when we are unable to shrink dcache.
2175 */
2176 if (flags & SLAB_RECLAIM_ACCOUNT)
2177 break;
2178
2179 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002180 * Large number of objects is good, but very large slabs are
2181 * currently bad for the gfp()s.
2182 */
David Rientjes543585c2011-10-18 22:09:24 -07002183 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002184 break;
2185
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002186 /*
2187 * Acceptable internal fragmentation?
2188 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002189 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002190 break;
2191 }
2192 return left_over;
2193}
2194
Pekka Enberg83b519e2009-06-10 19:40:04 +03002195static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002196{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002197 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002198 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002199
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002200 if (g_cpucache_up == NONE) {
2201 /*
2202 * Note: the first kmem_cache_create must create the cache
2203 * that's used by kmalloc(24), otherwise the creation of
2204 * further caches will BUG().
2205 */
2206 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2207
2208 /*
2209 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2210 * the first cache, then we need to set up all its list3s,
2211 * otherwise the creation of further caches will BUG().
2212 */
2213 set_up_list3s(cachep, SIZE_AC);
2214 if (INDEX_AC == INDEX_L3)
2215 g_cpucache_up = PARTIAL_L3;
2216 else
2217 g_cpucache_up = PARTIAL_AC;
2218 } else {
2219 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002220 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002221
2222 if (g_cpucache_up == PARTIAL_AC) {
2223 set_up_list3s(cachep, SIZE_L3);
2224 g_cpucache_up = PARTIAL_L3;
2225 } else {
2226 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002227 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002228 cachep->nodelists[node] =
2229 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002230 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002231 BUG_ON(!cachep->nodelists[node]);
2232 kmem_list3_init(cachep->nodelists[node]);
2233 }
2234 }
2235 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002236 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002237 jiffies + REAPTIMEOUT_LIST3 +
2238 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2239
2240 cpu_cache_get(cachep)->avail = 0;
2241 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2242 cpu_cache_get(cachep)->batchcount = 1;
2243 cpu_cache_get(cachep)->touched = 0;
2244 cachep->batchcount = 1;
2245 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002246 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002247}
2248
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002249/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 * kmem_cache_create - Create a cache.
2251 * @name: A string which is used in /proc/slabinfo to identify this cache.
2252 * @size: The size of objects to be created in this cache.
2253 * @align: The required alignment for the objects.
2254 * @flags: SLAB flags
2255 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 *
2257 * Returns a ptr to the cache on success, NULL on failure.
2258 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002259 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 *
2261 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002262 * the module calling this has to destroy the cache before getting unloaded.
2263 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 * The flags are
2265 *
2266 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2267 * to catch references to uninitialised memory.
2268 *
2269 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2270 * for buffer overruns.
2271 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2273 * cacheline. This can be beneficial if you're counting cycles as closely
2274 * as davem.
2275 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002276struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002278 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279{
2280 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002281 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002282 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
2284 /*
2285 * Sanity checks... these are all serious usage bugs.
2286 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002287 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002288 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002289 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002290 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002291 BUG();
2292 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002294 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002295 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302296 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002297 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002298 if (slab_is_available()) {
2299 get_online_cpus();
2300 mutex_lock(&cache_chain_mutex);
2301 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002302
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002303 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002304 char tmp;
2305 int res;
2306
2307 /*
2308 * This happens when the module gets unloaded and doesn't
2309 * destroy its slab cache and no-one else reuses the vmalloc
2310 * area of the module. Print a warning.
2311 */
Andrew Morton138ae662006-12-06 20:36:41 -08002312 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002313 if (res) {
matzeb4169522007-05-06 14:49:52 -07002314 printk(KERN_ERR
2315 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002316 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002317 continue;
2318 }
2319
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002320 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002321 printk(KERN_ERR
2322 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002323 dump_stack();
2324 goto oops;
2325 }
2326 }
2327
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328#if DEBUG
2329 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330#if FORCED_DEBUG
2331 /*
2332 * Enable redzoning and last user accounting, except for caches with
2333 * large objects, if the increased size would increase the object size
2334 * above the next power of two: caches with object sizes just above a
2335 * power of two have a significant amount of internal fragmentation.
2336 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002337 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2338 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002339 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 if (!(flags & SLAB_DESTROY_BY_RCU))
2341 flags |= SLAB_POISON;
2342#endif
2343 if (flags & SLAB_DESTROY_BY_RCU)
2344 BUG_ON(flags & SLAB_POISON);
2345#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002347 * Always checks flags, a caller might be expecting debug support which
2348 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002350 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351
Andrew Mortona737b3e2006-03-22 00:08:11 -08002352 /*
2353 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 * unaligned accesses for some archs when redzoning is used, and makes
2355 * sure any on-slab bufctl's are also correctly aligned.
2356 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002357 if (size & (BYTES_PER_WORD - 1)) {
2358 size += (BYTES_PER_WORD - 1);
2359 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 }
2361
Andrew Mortona737b3e2006-03-22 00:08:11 -08002362 /* calculate the final buffer alignment: */
2363
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 /* 1) arch recommendation: can be overridden for debug */
2365 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002366 /*
2367 * Default alignment: as specified by the arch code. Except if
2368 * an object is really small, then squeeze multiple objects into
2369 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 */
2371 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002372 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 ralign /= 2;
2374 } else {
2375 ralign = BYTES_PER_WORD;
2376 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002377
2378 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002379 * Redzoning and user store require word alignment or possibly larger.
2380 * Note this will be overridden by architecture or caller mandated
2381 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002382 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002383 if (flags & SLAB_STORE_USER)
2384 ralign = BYTES_PER_WORD;
2385
2386 if (flags & SLAB_RED_ZONE) {
2387 ralign = REDZONE_ALIGN;
2388 /* If redzoning, ensure that the second redzone is suitably
2389 * aligned, by adjusting the object size accordingly. */
2390 size += REDZONE_ALIGN - 1;
2391 size &= ~(REDZONE_ALIGN - 1);
2392 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002393
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002394 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 if (ralign < ARCH_SLAB_MINALIGN) {
2396 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002398 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 if (ralign < align) {
2400 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002402 /* disable debug if necessary */
2403 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002404 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002405 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002406 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 */
2408 align = ralign;
2409
Pekka Enberg83b519e2009-06-10 19:40:04 +03002410 if (slab_is_available())
2411 gfp = GFP_KERNEL;
2412 else
2413 gfp = GFP_NOWAIT;
2414
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002416 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002418 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002420 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002422 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423
Pekka Enbergca5f9702006-09-25 23:31:25 -07002424 /*
2425 * Both debugging options require word-alignment which is calculated
2426 * into align above.
2427 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002430 cachep->obj_offset += sizeof(unsigned long long);
2431 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 }
2433 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002434 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002435 * the real object. But if the second red zone needs to be
2436 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002438 if (flags & SLAB_RED_ZONE)
2439 size += REDZONE_ALIGN;
2440 else
2441 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 }
2443#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002444 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Carsten Otte1ab335d2010-08-06 18:19:22 +02002445 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2446 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 size = PAGE_SIZE;
2448 }
2449#endif
2450#endif
2451
Ingo Molnare0a42722006-06-23 02:03:46 -07002452 /*
2453 * Determine if the slab management is 'on' or 'off' slab.
2454 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002455 * it too early on. Always use on-slab management when
2456 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002457 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002458 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2459 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 /*
2461 * Size is large, assume best to place the slab management obj
2462 * off-slab (should allow better packing of objs).
2463 */
2464 flags |= CFLGS_OFF_SLAB;
2465
2466 size = ALIGN(size, align);
2467
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002468 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
2470 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002471 printk(KERN_ERR
2472 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 kmem_cache_free(&cache_cache, cachep);
2474 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002475 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002477 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2478 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479
2480 /*
2481 * If the slab has been placed off-slab, and we have enough space then
2482 * move it on-slab. This is at the expense of any extra colouring.
2483 */
2484 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2485 flags &= ~CFLGS_OFF_SLAB;
2486 left_over -= slab_size;
2487 }
2488
2489 if (flags & CFLGS_OFF_SLAB) {
2490 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002491 slab_size =
2492 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302493
2494#ifdef CONFIG_PAGE_POISONING
2495 /* If we're going to use the generic kernel_map_pages()
2496 * poisoning, then it's going to smash the contents of
2497 * the redzone and userword anyhow, so switch them off.
2498 */
2499 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2500 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2501#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 }
2503
2504 cachep->colour_off = cache_line_size();
2505 /* Offset must be a multiple of the alignment. */
2506 if (cachep->colour_off < align)
2507 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002508 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 cachep->slab_size = slab_size;
2510 cachep->flags = flags;
2511 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002512 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002514 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002515 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002517 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002518 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002519 /*
2520 * This is a possibility for one of the malloc_sizes caches.
2521 * But since we go off slab only for object size greater than
2522 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2523 * this should not happen at all.
2524 * But leave a BUG_ON for some lucky dude.
2525 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002526 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 cachep->name = name;
2530
Pekka Enberg83b519e2009-06-10 19:40:04 +03002531 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002532 __kmem_cache_destroy(cachep);
2533 cachep = NULL;
2534 goto oops;
2535 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536
Peter Zijlstra83835b32011-07-22 15:26:05 +02002537 if (flags & SLAB_DEBUG_OBJECTS) {
2538 /*
2539 * Would deadlock through slab_destroy()->call_rcu()->
2540 * debug_object_activate()->kmem_cache_alloc().
2541 */
2542 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2543
2544 slab_set_debugobj_lock_classes(cachep);
2545 }
2546
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 /* cache setup completed, link it into the list */
2548 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002549oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 if (!cachep && (flags & SLAB_PANIC))
2551 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002552 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002553 if (slab_is_available()) {
2554 mutex_unlock(&cache_chain_mutex);
2555 put_online_cpus();
2556 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 return cachep;
2558}
2559EXPORT_SYMBOL(kmem_cache_create);
2560
2561#if DEBUG
2562static void check_irq_off(void)
2563{
2564 BUG_ON(!irqs_disabled());
2565}
2566
2567static void check_irq_on(void)
2568{
2569 BUG_ON(irqs_disabled());
2570}
2571
Pekka Enberg343e0d72006-02-01 03:05:50 -08002572static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573{
2574#ifdef CONFIG_SMP
2575 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002576 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577#endif
2578}
Christoph Lametere498be72005-09-09 13:03:32 -07002579
Pekka Enberg343e0d72006-02-01 03:05:50 -08002580static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002581{
2582#ifdef CONFIG_SMP
2583 check_irq_off();
2584 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2585#endif
2586}
2587
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588#else
2589#define check_irq_off() do { } while(0)
2590#define check_irq_on() do { } while(0)
2591#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002592#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593#endif
2594
Christoph Lameteraab22072006-03-22 00:09:06 -08002595static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2596 struct array_cache *ac,
2597 int force, int node);
2598
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599static void do_drain(void *arg)
2600{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002601 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002603 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604
2605 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002606 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002607 spin_lock(&cachep->nodelists[node]->list_lock);
2608 free_block(cachep, ac->entry, ac->avail, node);
2609 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 ac->avail = 0;
2611}
2612
Pekka Enberg343e0d72006-02-01 03:05:50 -08002613static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614{
Christoph Lametere498be72005-09-09 13:03:32 -07002615 struct kmem_list3 *l3;
2616 int node;
2617
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002618 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002620 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002621 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002622 if (l3 && l3->alien)
2623 drain_alien_cache(cachep, l3->alien);
2624 }
2625
2626 for_each_online_node(node) {
2627 l3 = cachep->nodelists[node];
2628 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002629 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002630 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631}
2632
Christoph Lametered11d9e2006-06-30 01:55:45 -07002633/*
2634 * Remove slabs from the list of free slabs.
2635 * Specify the number of slabs to drain in tofree.
2636 *
2637 * Returns the actual number of slabs released.
2638 */
2639static int drain_freelist(struct kmem_cache *cache,
2640 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002642 struct list_head *p;
2643 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645
Christoph Lametered11d9e2006-06-30 01:55:45 -07002646 nr_freed = 0;
2647 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648
Christoph Lametered11d9e2006-06-30 01:55:45 -07002649 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002650 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002651 if (p == &l3->slabs_free) {
2652 spin_unlock_irq(&l3->list_lock);
2653 goto out;
2654 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655
Christoph Lametered11d9e2006-06-30 01:55:45 -07002656 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002658 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659#endif
2660 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002661 /*
2662 * Safe to drop the lock. The slab is no longer linked
2663 * to the cache.
2664 */
2665 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002666 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002667 slab_destroy(cache, slabp);
2668 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002670out:
2671 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672}
2673
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002674/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002675static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002676{
2677 int ret = 0, i = 0;
2678 struct kmem_list3 *l3;
2679
2680 drain_cpu_caches(cachep);
2681
2682 check_irq_on();
2683 for_each_online_node(i) {
2684 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002685 if (!l3)
2686 continue;
2687
2688 drain_freelist(cachep, l3, l3->free_objects);
2689
2690 ret += !list_empty(&l3->slabs_full) ||
2691 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002692 }
2693 return (ret ? 1 : 0);
2694}
2695
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696/**
2697 * kmem_cache_shrink - Shrink a cache.
2698 * @cachep: The cache to shrink.
2699 *
2700 * Releases as many slabs as possible for a cache.
2701 * To help debugging, a zero exit status indicates all slabs were released.
2702 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002703int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002705 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002706 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002708 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002709 mutex_lock(&cache_chain_mutex);
2710 ret = __cache_shrink(cachep);
2711 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002712 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002713 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714}
2715EXPORT_SYMBOL(kmem_cache_shrink);
2716
2717/**
2718 * kmem_cache_destroy - delete a cache
2719 * @cachep: the cache to destroy
2720 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002721 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 *
2723 * It is expected this function will be called by a module when it is
2724 * unloaded. This will remove the cache completely, and avoid a duplicate
2725 * cache being allocated each time a module is loaded and unloaded, if the
2726 * module doesn't have persistent in-kernel storage across loads and unloads.
2727 *
2728 * The cache must be empty before calling this function.
2729 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002730 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 * during the kmem_cache_destroy().
2732 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002733void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002735 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002738 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002739 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 /*
2741 * the chain is never empty, cache_cache is never destroyed
2742 */
2743 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 if (__cache_shrink(cachep)) {
2745 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002746 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002747 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002748 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002749 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 }
2751
2752 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002753 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002755 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002756 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002757 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759EXPORT_SYMBOL(kmem_cache_destroy);
2760
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002761/*
2762 * Get the memory for a slab management obj.
2763 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2764 * always come from malloc_sizes caches. The slab descriptor cannot
2765 * come from the same cache which is getting created because,
2766 * when we are searching for an appropriate cache for these
2767 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2768 * If we are creating a malloc_sizes cache here it would not be visible to
2769 * kmem_find_general_cachep till the initialization is complete.
2770 * Hence we cannot have slabp_cache same as the original cache.
2771 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002772static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002773 int colour_off, gfp_t local_flags,
2774 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775{
2776 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002777
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 if (OFF_SLAB(cachep)) {
2779 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002780 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002781 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002782 /*
2783 * If the first object in the slab is leaked (it's allocated
2784 * but no one has a reference to it), we want to make sure
2785 * kmemleak does not treat the ->s_mem pointer as a reference
2786 * to the object. Otherwise we will not report the leak.
2787 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002788 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2789 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 if (!slabp)
2791 return NULL;
2792 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002793 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 colour_off += cachep->slab_size;
2795 }
2796 slabp->inuse = 0;
2797 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002798 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002799 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002800 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 return slabp;
2802}
2803
2804static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2805{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002806 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807}
2808
Pekka Enberg343e0d72006-02-01 03:05:50 -08002809static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002810 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811{
2812 int i;
2813
2814 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002815 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816#if DEBUG
2817 /* need to poison the objs? */
2818 if (cachep->flags & SLAB_POISON)
2819 poison_obj(cachep, objp, POISON_FREE);
2820 if (cachep->flags & SLAB_STORE_USER)
2821 *dbg_userword(cachep, objp) = NULL;
2822
2823 if (cachep->flags & SLAB_RED_ZONE) {
2824 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2825 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2826 }
2827 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002828 * Constructors are not allowed to allocate memory from the same
2829 * cache which they are a constructor for. Otherwise, deadlock.
2830 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 */
2832 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002833 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834
2835 if (cachep->flags & SLAB_RED_ZONE) {
2836 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2837 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002838 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2840 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002841 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002843 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2844 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002845 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002846 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847#else
2848 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002849 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002851 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002853 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854}
2855
Pekka Enberg343e0d72006-02-01 03:05:50 -08002856static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002858 if (CONFIG_ZONE_DMA_FLAG) {
2859 if (flags & GFP_DMA)
2860 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2861 else
2862 BUG_ON(cachep->gfpflags & GFP_DMA);
2863 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864}
2865
Andrew Mortona737b3e2006-03-22 00:08:11 -08002866static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2867 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002868{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002869 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002870 kmem_bufctl_t next;
2871
2872 slabp->inuse++;
2873 next = slab_bufctl(slabp)[slabp->free];
2874#if DEBUG
2875 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2876 WARN_ON(slabp->nodeid != nodeid);
2877#endif
2878 slabp->free = next;
2879
2880 return objp;
2881}
2882
Andrew Mortona737b3e2006-03-22 00:08:11 -08002883static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2884 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002885{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002886 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002887
2888#if DEBUG
2889 /* Verify that the slab belongs to the intended node */
2890 WARN_ON(slabp->nodeid != nodeid);
2891
Al Viro871751e2006-03-25 03:06:39 -08002892 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002893 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002894 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002895 BUG();
2896 }
2897#endif
2898 slab_bufctl(slabp)[objnr] = slabp->free;
2899 slabp->free = objnr;
2900 slabp->inuse--;
2901}
2902
Pekka Enberg47768742006-06-23 02:03:07 -07002903/*
2904 * Map pages beginning at addr to the given cache and slab. This is required
2905 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002906 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002907 */
2908static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2909 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910{
Pekka Enberg47768742006-06-23 02:03:07 -07002911 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 struct page *page;
2913
Pekka Enberg47768742006-06-23 02:03:07 -07002914 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002915
Pekka Enberg47768742006-06-23 02:03:07 -07002916 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002917 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002918 nr_pages <<= cache->gfporder;
2919
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002921 page_set_cache(page, cache);
2922 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002924 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925}
2926
2927/*
2928 * Grow (by 1) the number of slabs within a cache. This is called by
2929 * kmem_cache_alloc() when there are no active objs left in a cache.
2930 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002931static int cache_grow(struct kmem_cache *cachep,
2932 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002934 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002935 size_t offset;
2936 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002937 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938
Andrew Mortona737b3e2006-03-22 00:08:11 -08002939 /*
2940 * Be lazy and only check for valid flags here, keeping it out of the
2941 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002943 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2944 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002946 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002948 l3 = cachep->nodelists[nodeid];
2949 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950
2951 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002952 offset = l3->colour_next;
2953 l3->colour_next++;
2954 if (l3->colour_next >= cachep->colour)
2955 l3->colour_next = 0;
2956 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002958 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959
2960 if (local_flags & __GFP_WAIT)
2961 local_irq_enable();
2962
2963 /*
2964 * The test for missing atomic flag is performed here, rather than
2965 * the more obvious place, simply to reduce the critical path length
2966 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2967 * will eventually be caught here (where it matters).
2968 */
2969 kmem_flagcheck(cachep, flags);
2970
Andrew Mortona737b3e2006-03-22 00:08:11 -08002971 /*
2972 * Get mem for the objs. Attempt to allocate a physical page from
2973 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002974 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002975 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002976 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002977 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 goto failed;
2979
2980 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002981 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002982 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002983 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 goto opps1;
2985
Pekka Enberg47768742006-06-23 02:03:07 -07002986 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987
Christoph Lametera35afb82007-05-16 22:10:57 -07002988 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989
2990 if (local_flags & __GFP_WAIT)
2991 local_irq_disable();
2992 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002993 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994
2995 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002996 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002998 l3->free_objects += cachep->num;
2999 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003001opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003003failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004 if (local_flags & __GFP_WAIT)
3005 local_irq_disable();
3006 return 0;
3007}
3008
3009#if DEBUG
3010
3011/*
3012 * Perform extra freeing checks:
3013 * - detect bad pointers.
3014 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 */
3016static void kfree_debugcheck(const void *objp)
3017{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 if (!virt_addr_valid(objp)) {
3019 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003020 (unsigned long)objp);
3021 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023}
3024
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003025static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3026{
David Woodhouseb46b8f12007-05-08 00:22:59 -07003027 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003028
3029 redzone1 = *dbg_redzone1(cache, obj);
3030 redzone2 = *dbg_redzone2(cache, obj);
3031
3032 /*
3033 * Redzone is ok.
3034 */
3035 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3036 return;
3037
3038 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3039 slab_error(cache, "double free detected");
3040 else
3041 slab_error(cache, "memory outside object was overwritten");
3042
David Woodhouseb46b8f12007-05-08 00:22:59 -07003043 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003044 obj, redzone1, redzone2);
3045}
3046
Pekka Enberg343e0d72006-02-01 03:05:50 -08003047static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003048 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049{
3050 struct page *page;
3051 unsigned int objnr;
3052 struct slab *slabp;
3053
Matthew Wilcox80cbd912007-11-29 12:05:13 -07003054 BUG_ON(virt_to_cache(objp) != cachep);
3055
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003056 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07003058 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059
Pekka Enberg065d41c2005-11-13 16:06:46 -08003060 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061
3062 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003063 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3065 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3066 }
3067 if (cachep->flags & SLAB_STORE_USER)
3068 *dbg_userword(cachep, objp) = caller;
3069
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003070 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071
3072 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003073 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074
Al Viro871751e2006-03-25 03:06:39 -08003075#ifdef CONFIG_DEBUG_SLAB_LEAK
3076 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3077#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 if (cachep->flags & SLAB_POISON) {
3079#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08003080 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003082 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003083 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 } else {
3085 poison_obj(cachep, objp, POISON_FREE);
3086 }
3087#else
3088 poison_obj(cachep, objp, POISON_FREE);
3089#endif
3090 }
3091 return objp;
3092}
3093
Pekka Enberg343e0d72006-02-01 03:05:50 -08003094static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095{
3096 kmem_bufctl_t i;
3097 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003098
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 /* Check slab's freelist to see if this obj is there. */
3100 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3101 entries++;
3102 if (entries > cachep->num || i >= cachep->num)
3103 goto bad;
3104 }
3105 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003106bad:
3107 printk(KERN_ERR "slab: Internal list corruption detected in "
Dave Jonesface37f2011-11-15 15:03:52 -08003108 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3109 cachep->name, cachep->num, slabp, slabp->inuse,
3110 print_tainted());
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003111 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3112 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3113 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114 BUG();
3115 }
3116}
3117#else
3118#define kfree_debugcheck(x) do { } while(0)
3119#define cache_free_debugcheck(x,objp,z) (objp)
3120#define check_slabp(x,y) do { } while(0)
3121#endif
3122
Pekka Enberg343e0d72006-02-01 03:05:50 -08003123static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124{
3125 int batchcount;
3126 struct kmem_list3 *l3;
3127 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003128 int node;
3129
Andrew Mortona737b3e2006-03-22 00:08:11 -08003130retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003131 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003132 node = numa_mem_id();
Joe Korty6d2144d2008-03-05 15:04:59 -08003133 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 batchcount = ac->batchcount;
3135 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003136 /*
3137 * If there was little recent activity on this cache, then
3138 * perform only a partial refill. Otherwise we could generate
3139 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 */
3141 batchcount = BATCHREFILL_LIMIT;
3142 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003143 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144
Christoph Lametere498be72005-09-09 13:03:32 -07003145 BUG_ON(ac->avail > 0 || !l3);
3146 spin_lock(&l3->list_lock);
3147
Christoph Lameter3ded1752006-03-25 03:06:44 -08003148 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003149 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3150 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003151 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003152 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003153
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154 while (batchcount > 0) {
3155 struct list_head *entry;
3156 struct slab *slabp;
3157 /* Get slab alloc is to come from. */
3158 entry = l3->slabs_partial.next;
3159 if (entry == &l3->slabs_partial) {
3160 l3->free_touched = 1;
3161 entry = l3->slabs_free.next;
3162 if (entry == &l3->slabs_free)
3163 goto must_grow;
3164 }
3165
3166 slabp = list_entry(entry, struct slab, list);
3167 check_slabp(cachep, slabp);
3168 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003169
3170 /*
3171 * The slab was either on partial or free list so
3172 * there must be at least one object available for
3173 * allocation.
3174 */
roel kluin249b9f32008-10-29 17:18:07 -04003175 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003176
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178 STATS_INC_ALLOCED(cachep);
3179 STATS_INC_ACTIVE(cachep);
3180 STATS_SET_HIGH(cachep);
3181
Matthew Dobson78d382d2006-02-01 03:05:47 -08003182 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003183 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 }
3185 check_slabp(cachep, slabp);
3186
3187 /* move slabp to correct slabp list: */
3188 list_del(&slabp->list);
3189 if (slabp->free == BUFCTL_END)
3190 list_add(&slabp->list, &l3->slabs_full);
3191 else
3192 list_add(&slabp->list, &l3->slabs_partial);
3193 }
3194
Andrew Mortona737b3e2006-03-22 00:08:11 -08003195must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003197alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003198 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199
3200 if (unlikely(!ac->avail)) {
3201 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003202 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003203
Andrew Mortona737b3e2006-03-22 00:08:11 -08003204 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003205 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003206 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 return NULL;
3208
Andrew Mortona737b3e2006-03-22 00:08:11 -08003209 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 goto retry;
3211 }
3212 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003213 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214}
3215
Andrew Mortona737b3e2006-03-22 00:08:11 -08003216static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3217 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218{
3219 might_sleep_if(flags & __GFP_WAIT);
3220#if DEBUG
3221 kmem_flagcheck(cachep, flags);
3222#endif
3223}
3224
3225#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003226static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3227 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003229 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003231 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003233 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003234 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003235 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 else
3237 check_poison_obj(cachep, objp);
3238#else
3239 check_poison_obj(cachep, objp);
3240#endif
3241 poison_obj(cachep, objp, POISON_INUSE);
3242 }
3243 if (cachep->flags & SLAB_STORE_USER)
3244 *dbg_userword(cachep, objp) = caller;
3245
3246 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003247 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3248 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3249 slab_error(cachep, "double free, or memory outside"
3250 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003251 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003252 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003253 objp, *dbg_redzone1(cachep, objp),
3254 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 }
3256 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3257 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3258 }
Al Viro871751e2006-03-25 03:06:39 -08003259#ifdef CONFIG_DEBUG_SLAB_LEAK
3260 {
3261 struct slab *slabp;
3262 unsigned objnr;
3263
Christoph Lameterb49af682007-05-06 14:49:41 -07003264 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003265 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3266 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3267 }
3268#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003269 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003270 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003271 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003272 if (ARCH_SLAB_MINALIGN &&
3273 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003274 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003275 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003276 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277 return objp;
3278}
3279#else
3280#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3281#endif
3282
Akinobu Mita773ff602008-12-23 19:37:01 +09003283static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003284{
3285 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003286 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003287
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03003288 return should_failslab(obj_size(cachep), flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003289}
3290
Pekka Enberg343e0d72006-02-01 03:05:50 -08003291static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003293 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 struct array_cache *ac;
3295
Alok N Kataria5c382302005-09-27 21:45:46 -07003296 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003297
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003298 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 if (likely(ac->avail)) {
3300 STATS_INC_ALLOCHIT(cachep);
3301 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003302 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 } else {
3304 STATS_INC_ALLOCMISS(cachep);
3305 objp = cache_alloc_refill(cachep, flags);
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003306 /*
3307 * the 'ac' may be updated by cache_alloc_refill(),
3308 * and kmemleak_erase() requires its correct value.
3309 */
3310 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003312 /*
3313 * To avoid a false negative, if an object that is in one of the
3314 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3315 * treat the array pointers as a reference to the object.
3316 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003317 if (objp)
3318 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003319 return objp;
3320}
3321
Christoph Lametere498be72005-09-09 13:03:32 -07003322#ifdef CONFIG_NUMA
3323/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003324 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003325 *
3326 * If we are in_interrupt, then process context, including cpusets and
3327 * mempolicy, may not apply and should not be used for allocation policy.
3328 */
3329static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3330{
3331 int nid_alloc, nid_here;
3332
Christoph Lameter765c4502006-09-27 01:50:08 -07003333 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003334 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003335 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003336 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003337 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003338 else if (current->mempolicy)
3339 nid_alloc = slab_node(current->mempolicy);
3340 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003341 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003342 return NULL;
3343}
3344
3345/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003346 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003347 * certain node and fall back is permitted. First we scan all the
3348 * available nodelists for available objects. If that fails then we
3349 * perform an allocation without specifying a node. This allows the page
3350 * allocator to do its reclaim / fallback magic. We then insert the
3351 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003352 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003353static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003354{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003355 struct zonelist *zonelist;
3356 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003357 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003358 struct zone *zone;
3359 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003360 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003361 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003362 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003363
3364 if (flags & __GFP_THISNODE)
3365 return NULL;
3366
Christoph Lameter6cb06222007-10-16 01:25:41 -07003367 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003368
Mel Gormancc9a6c82012-03-21 16:34:11 -07003369retry_cpuset:
3370 cpuset_mems_cookie = get_mems_allowed();
3371 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3372
Christoph Lameter3c517a62006-12-06 20:33:29 -08003373retry:
3374 /*
3375 * Look through allowed nodes for objects available
3376 * from existing per node queues.
3377 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003378 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3379 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003380
Mel Gorman54a6eb52008-04-28 02:12:16 -07003381 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003382 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003383 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003384 obj = ____cache_alloc_node(cache,
3385 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003386 if (obj)
3387 break;
3388 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003389 }
3390
Christoph Lametercfce6602007-05-06 14:50:17 -07003391 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003392 /*
3393 * This allocation will be performed within the constraints
3394 * of the current cpuset / memory policy requirements.
3395 * We may trigger various forms of reclaim on the allowed
3396 * set and go into memory reserves if necessary.
3397 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003398 if (local_flags & __GFP_WAIT)
3399 local_irq_enable();
3400 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003401 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003402 if (local_flags & __GFP_WAIT)
3403 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003404 if (obj) {
3405 /*
3406 * Insert into the appropriate per node queues
3407 */
3408 nid = page_to_nid(virt_to_page(obj));
3409 if (cache_grow(cache, flags, nid, obj)) {
3410 obj = ____cache_alloc_node(cache,
3411 flags | GFP_THISNODE, nid);
3412 if (!obj)
3413 /*
3414 * Another processor may allocate the
3415 * objects in the slab since we are
3416 * not holding any locks.
3417 */
3418 goto retry;
3419 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003420 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003421 obj = NULL;
3422 }
3423 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003424 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003425
3426 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3427 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003428 return obj;
3429}
3430
3431/*
Christoph Lametere498be72005-09-09 13:03:32 -07003432 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003434static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003435 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003436{
3437 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003438 struct slab *slabp;
3439 struct kmem_list3 *l3;
3440 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003441 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003443 l3 = cachep->nodelists[nodeid];
3444 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003445
Andrew Mortona737b3e2006-03-22 00:08:11 -08003446retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003447 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003448 spin_lock(&l3->list_lock);
3449 entry = l3->slabs_partial.next;
3450 if (entry == &l3->slabs_partial) {
3451 l3->free_touched = 1;
3452 entry = l3->slabs_free.next;
3453 if (entry == &l3->slabs_free)
3454 goto must_grow;
3455 }
Christoph Lametere498be72005-09-09 13:03:32 -07003456
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003457 slabp = list_entry(entry, struct slab, list);
3458 check_spinlock_acquired_node(cachep, nodeid);
3459 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003460
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 STATS_INC_NODEALLOCS(cachep);
3462 STATS_INC_ACTIVE(cachep);
3463 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003464
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003465 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003466
Matthew Dobson78d382d2006-02-01 03:05:47 -08003467 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003468 check_slabp(cachep, slabp);
3469 l3->free_objects--;
3470 /* move slabp to correct slabp list: */
3471 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003472
Andrew Mortona737b3e2006-03-22 00:08:11 -08003473 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003474 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003475 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003476 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003477
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003478 spin_unlock(&l3->list_lock);
3479 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003480
Andrew Mortona737b3e2006-03-22 00:08:11 -08003481must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003482 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003483 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003484 if (x)
3485 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003486
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003487 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003488
Andrew Mortona737b3e2006-03-22 00:08:11 -08003489done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003490 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003491}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003492
3493/**
3494 * kmem_cache_alloc_node - Allocate an object on the specified node
3495 * @cachep: The cache to allocate from.
3496 * @flags: See kmalloc().
3497 * @nodeid: node number of the target node.
3498 * @caller: return address of caller, used for debug information
3499 *
3500 * Identical to kmem_cache_alloc but it will allocate memory on the given
3501 * node, which can improve the performance for cpu bound structures.
3502 *
3503 * Fallback to other node is possible if __GFP_THISNODE is not set.
3504 */
3505static __always_inline void *
3506__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3507 void *caller)
3508{
3509 unsigned long save_flags;
3510 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003511 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003512
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003513 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003514
Nick Piggincf40bd12009-01-21 08:12:39 +01003515 lockdep_trace_alloc(flags);
3516
Akinobu Mita773ff602008-12-23 19:37:01 +09003517 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003518 return NULL;
3519
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003520 cache_alloc_debugcheck_before(cachep, flags);
3521 local_irq_save(save_flags);
3522
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003523 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003524 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003525
3526 if (unlikely(!cachep->nodelists[nodeid])) {
3527 /* Node not bootstrapped yet */
3528 ptr = fallback_alloc(cachep, flags);
3529 goto out;
3530 }
3531
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003532 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003533 /*
3534 * Use the locally cached objects if possible.
3535 * However ____cache_alloc does not allow fallback
3536 * to other nodes. It may fail while we still have
3537 * objects on other nodes available.
3538 */
3539 ptr = ____cache_alloc(cachep, flags);
3540 if (ptr)
3541 goto out;
3542 }
3543 /* ___cache_alloc_node can fall back to other nodes */
3544 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3545 out:
3546 local_irq_restore(save_flags);
3547 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003548 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3549 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003550
Pekka Enbergc175eea2008-05-09 20:35:53 +02003551 if (likely(ptr))
3552 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3553
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003554 if (unlikely((flags & __GFP_ZERO) && ptr))
3555 memset(ptr, 0, obj_size(cachep));
3556
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003557 return ptr;
3558}
3559
3560static __always_inline void *
3561__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3562{
3563 void *objp;
3564
3565 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3566 objp = alternate_node_alloc(cache, flags);
3567 if (objp)
3568 goto out;
3569 }
3570 objp = ____cache_alloc(cache, flags);
3571
3572 /*
3573 * We may just have run out of memory on the local node.
3574 * ____cache_alloc_node() knows how to locate memory on other nodes
3575 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003576 if (!objp)
3577 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003578
3579 out:
3580 return objp;
3581}
3582#else
3583
3584static __always_inline void *
3585__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3586{
3587 return ____cache_alloc(cachep, flags);
3588}
3589
3590#endif /* CONFIG_NUMA */
3591
3592static __always_inline void *
3593__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3594{
3595 unsigned long save_flags;
3596 void *objp;
3597
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003598 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003599
Nick Piggincf40bd12009-01-21 08:12:39 +01003600 lockdep_trace_alloc(flags);
3601
Akinobu Mita773ff602008-12-23 19:37:01 +09003602 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003603 return NULL;
3604
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003605 cache_alloc_debugcheck_before(cachep, flags);
3606 local_irq_save(save_flags);
3607 objp = __do_cache_alloc(cachep, flags);
3608 local_irq_restore(save_flags);
3609 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003610 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3611 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003612 prefetchw(objp);
3613
Pekka Enbergc175eea2008-05-09 20:35:53 +02003614 if (likely(objp))
3615 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3616
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003617 if (unlikely((flags & __GFP_ZERO) && objp))
3618 memset(objp, 0, obj_size(cachep));
3619
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003620 return objp;
3621}
Christoph Lametere498be72005-09-09 13:03:32 -07003622
3623/*
3624 * Caller needs to acquire correct kmem_list's list_lock
3625 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003626static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003627 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628{
3629 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003630 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631
3632 for (i = 0; i < nr_objects; i++) {
3633 void *objp = objpp[i];
3634 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003636 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003637 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003639 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003641 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003643 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 check_slabp(cachep, slabp);
3645
3646 /* fixup slab chains */
3647 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003648 if (l3->free_objects > l3->free_limit) {
3649 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003650 /* No need to drop any previously held
3651 * lock here, even if we have a off-slab slab
3652 * descriptor it is guaranteed to come from
3653 * a different cache, refer to comments before
3654 * alloc_slabmgmt.
3655 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656 slab_destroy(cachep, slabp);
3657 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003658 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 }
3660 } else {
3661 /* Unconditionally move a slab to the end of the
3662 * partial list on free - maximum time for the
3663 * other objects to be freed, too.
3664 */
Christoph Lametere498be72005-09-09 13:03:32 -07003665 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 }
3667 }
3668}
3669
Pekka Enberg343e0d72006-02-01 03:05:50 -08003670static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671{
3672 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003673 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003674 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675
3676 batchcount = ac->batchcount;
3677#if DEBUG
3678 BUG_ON(!batchcount || batchcount > ac->avail);
3679#endif
3680 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003681 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003682 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003683 if (l3->shared) {
3684 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003685 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 if (max) {
3687 if (batchcount > max)
3688 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003689 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003690 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 shared_array->avail += batchcount;
3692 goto free_done;
3693 }
3694 }
3695
Christoph Lameterff694162005-09-22 21:44:02 -07003696 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003697free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698#if STATS
3699 {
3700 int i = 0;
3701 struct list_head *p;
3702
Christoph Lametere498be72005-09-09 13:03:32 -07003703 p = l3->slabs_free.next;
3704 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 struct slab *slabp;
3706
3707 slabp = list_entry(p, struct slab, list);
3708 BUG_ON(slabp->inuse);
3709
3710 i++;
3711 p = p->next;
3712 }
3713 STATS_SET_FREEABLE(cachep, i);
3714 }
3715#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003716 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003718 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719}
3720
3721/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003722 * Release an obj back to its cache. If the obj has a constructed state, it must
3723 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003725static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3726 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003728 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729
3730 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003731 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003732 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733
Pekka Enbergc175eea2008-05-09 20:35:53 +02003734 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3735
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003736 /*
3737 * Skip calling cache_free_alien() when the platform is not numa.
3738 * This will avoid cache misses that happen while accessing slabp (which
3739 * is per page memory reference) to get nodeid. Instead use a global
3740 * variable to skip the call, which is mostly likely to be present in
3741 * the cache.
3742 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003743 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003744 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003745
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746 if (likely(ac->avail < ac->limit)) {
3747 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748 } else {
3749 STATS_INC_FREEMISS(cachep);
3750 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003752
3753 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754}
3755
3756/**
3757 * kmem_cache_alloc - Allocate an object
3758 * @cachep: The cache to allocate from.
3759 * @flags: See kmalloc().
3760 *
3761 * Allocate an object from this cache. The flags are only relevant
3762 * if the cache has no available objects.
3763 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003764void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003766 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3767
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003768 trace_kmem_cache_alloc(_RET_IP_, ret,
3769 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003770
3771 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772}
3773EXPORT_SYMBOL(kmem_cache_alloc);
3774
Li Zefan0f24f122009-12-11 15:45:30 +08003775#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003776void *
3777kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003778{
Steven Rostedt85beb582010-11-24 16:23:34 -05003779 void *ret;
3780
3781 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3782
3783 trace_kmalloc(_RET_IP_, ret,
3784 size, slab_buffer_size(cachep), flags);
3785 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003786}
Steven Rostedt85beb582010-11-24 16:23:34 -05003787EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003788#endif
3789
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003791void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3792{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003793 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3794 __builtin_return_address(0));
3795
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003796 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3797 obj_size(cachep), cachep->buffer_size,
3798 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003799
3800 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003801}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802EXPORT_SYMBOL(kmem_cache_alloc_node);
3803
Li Zefan0f24f122009-12-11 15:45:30 +08003804#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003805void *kmem_cache_alloc_node_trace(size_t size,
3806 struct kmem_cache *cachep,
3807 gfp_t flags,
3808 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003809{
Steven Rostedt85beb582010-11-24 16:23:34 -05003810 void *ret;
3811
3812 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003813 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003814 trace_kmalloc_node(_RET_IP_, ret,
3815 size, slab_buffer_size(cachep),
3816 flags, nodeid);
3817 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003818}
Steven Rostedt85beb582010-11-24 16:23:34 -05003819EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003820#endif
3821
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003822static __always_inline void *
3823__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003824{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003825 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003826
3827 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003828 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3829 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003830 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003831}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003832
Li Zefan0bb38a52009-12-11 15:45:50 +08003833#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003834void *__kmalloc_node(size_t size, gfp_t flags, int node)
3835{
3836 return __do_kmalloc_node(size, flags, node,
3837 __builtin_return_address(0));
3838}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003839EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003840
3841void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003842 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003843{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003844 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003845}
3846EXPORT_SYMBOL(__kmalloc_node_track_caller);
3847#else
3848void *__kmalloc_node(size_t size, gfp_t flags, int node)
3849{
3850 return __do_kmalloc_node(size, flags, node, NULL);
3851}
3852EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003853#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003854#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855
3856/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003857 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003859 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003860 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003862static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3863 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003865 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003866 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003868 /* If you want to save a few bytes .text space: replace
3869 * __ with kmem_.
3870 * Then kmalloc uses the uninlined functions instead of the inline
3871 * functions.
3872 */
3873 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003874 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3875 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003876 ret = __cache_alloc(cachep, flags, caller);
3877
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003878 trace_kmalloc((unsigned long) caller, ret,
3879 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003880
3881 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003882}
3883
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003884
Li Zefan0bb38a52009-12-11 15:45:50 +08003885#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003886void *__kmalloc(size_t size, gfp_t flags)
3887{
Al Viro871751e2006-03-25 03:06:39 -08003888 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889}
3890EXPORT_SYMBOL(__kmalloc);
3891
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003892void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003893{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003894 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003895}
3896EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003897
3898#else
3899void *__kmalloc(size_t size, gfp_t flags)
3900{
3901 return __do_kmalloc(size, flags, NULL);
3902}
3903EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003904#endif
3905
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906/**
3907 * kmem_cache_free - Deallocate an object
3908 * @cachep: The cache the allocation was from.
3909 * @objp: The previously allocated object.
3910 *
3911 * Free an object which was previously allocated from this
3912 * cache.
3913 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003914void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915{
3916 unsigned long flags;
3917
3918 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003919 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003920 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3921 debug_check_no_obj_freed(objp, obj_size(cachep));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003922 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003924
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003925 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926}
3927EXPORT_SYMBOL(kmem_cache_free);
3928
3929/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 * kfree - free previously allocated memory
3931 * @objp: pointer returned by kmalloc.
3932 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003933 * If @objp is NULL, no operation is performed.
3934 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935 * Don't free memory not originally allocated by kmalloc()
3936 * or you will run into trouble.
3937 */
3938void kfree(const void *objp)
3939{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003940 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 unsigned long flags;
3942
Pekka Enberg2121db72009-03-25 11:05:57 +02003943 trace_kfree(_RET_IP_, objp);
3944
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003945 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946 return;
3947 local_irq_save(flags);
3948 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003949 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003950 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003951 debug_check_no_obj_freed(objp, obj_size(c));
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003952 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 local_irq_restore(flags);
3954}
3955EXPORT_SYMBOL(kfree);
3956
Pekka Enberg343e0d72006-02-01 03:05:50 -08003957unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003959 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960}
3961EXPORT_SYMBOL(kmem_cache_size);
3962
Christoph Lametere498be72005-09-09 13:03:32 -07003963/*
Simon Arlott183ff222007-10-20 01:27:18 +02003964 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003965 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003966static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003967{
3968 int node;
3969 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003970 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003971 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003972
Mel Gorman9c09a952008-01-24 05:49:54 -08003973 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003974
Paul Menage3395ee02006-12-06 20:32:16 -08003975 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003976 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003977 if (!new_alien)
3978 goto fail;
3979 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003980
Eric Dumazet63109842007-05-06 14:49:28 -07003981 new_shared = NULL;
3982 if (cachep->shared) {
3983 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003984 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003985 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003986 if (!new_shared) {
3987 free_alien_cache(new_alien);
3988 goto fail;
3989 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003990 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003991
Andrew Mortona737b3e2006-03-22 00:08:11 -08003992 l3 = cachep->nodelists[node];
3993 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003994 struct array_cache *shared = l3->shared;
3995
Christoph Lametere498be72005-09-09 13:03:32 -07003996 spin_lock_irq(&l3->list_lock);
3997
Christoph Lametercafeb022006-03-25 03:06:46 -08003998 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003999 free_block(cachep, shared->entry,
4000 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07004001
Christoph Lametercafeb022006-03-25 03:06:46 -08004002 l3->shared = new_shared;
4003 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07004004 l3->alien = new_alien;
4005 new_alien = NULL;
4006 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004007 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004008 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004009 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08004010 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004011 free_alien_cache(new_alien);
4012 continue;
4013 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03004014 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08004015 if (!l3) {
4016 free_alien_cache(new_alien);
4017 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004018 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004019 }
Christoph Lametere498be72005-09-09 13:03:32 -07004020
4021 kmem_list3_init(l3);
4022 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08004023 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08004024 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07004025 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004026 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004027 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004028 cachep->nodelists[node] = l3;
4029 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004030 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004031
Andrew Mortona737b3e2006-03-22 00:08:11 -08004032fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08004033 if (!cachep->next.next) {
4034 /* Cache is not active yet. Roll back what we did */
4035 node--;
4036 while (node >= 0) {
4037 if (cachep->nodelists[node]) {
4038 l3 = cachep->nodelists[node];
4039
4040 kfree(l3->shared);
4041 free_alien_cache(l3->alien);
4042 kfree(l3);
4043 cachep->nodelists[node] = NULL;
4044 }
4045 node--;
4046 }
4047 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004048 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07004049}
4050
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08004052 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004053 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054};
4055
4056static void do_ccupdate_local(void *info)
4057{
Andrew Mortona737b3e2006-03-22 00:08:11 -08004058 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 struct array_cache *old;
4060
4061 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08004062 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07004063
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4065 new->new[smp_processor_id()] = old;
4066}
4067
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004068/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08004069static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004070 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004072 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004073 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004075 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4076 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004077 if (!new)
4078 return -ENOMEM;
4079
Christoph Lametere498be72005-09-09 13:03:32 -07004080 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004081 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004082 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004083 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004084 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004085 kfree(new->new[i]);
4086 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004087 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 }
4089 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004090 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004092 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004093
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 cachep->batchcount = batchcount;
4096 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004097 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
Christoph Lametere498be72005-09-09 13:03:32 -07004099 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004100 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 if (!ccold)
4102 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004103 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4104 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4105 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 kfree(ccold);
4107 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004108 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004109 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110}
4111
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004112/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004113static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114{
4115 int err;
4116 int limit, shared;
4117
Andrew Mortona737b3e2006-03-22 00:08:11 -08004118 /*
4119 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120 * - create a LIFO ordering, i.e. return objects that are cache-warm
4121 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004122 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 * bufctl chains: array operations are cheaper.
4124 * The numbers are guessed, we should auto-tune as described by
4125 * Bonwick.
4126 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004127 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004129 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004131 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004133 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134 limit = 54;
4135 else
4136 limit = 120;
4137
Andrew Mortona737b3e2006-03-22 00:08:11 -08004138 /*
4139 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 * allocation behaviour: Most allocs on one cpu, most free operations
4141 * on another cpu. For these cases, an efficient object passing between
4142 * cpus is necessary. This is provided by a shared array. The array
4143 * replaces Bonwick's magazine layer.
4144 * On uniprocessor, it's functionally equivalent (but less efficient)
4145 * to a larger limit. Thus disabled by default.
4146 */
4147 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004148 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150
4151#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004152 /*
4153 * With debugging enabled, large batchcount lead to excessively long
4154 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 */
4156 if (limit > 32)
4157 limit = 32;
4158#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004159 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 if (err)
4161 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004162 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004163 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164}
4165
Christoph Lameter1b552532006-03-22 00:09:07 -08004166/*
4167 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004168 * necessary. Note that the l3 listlock also protects the array_cache
4169 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004170 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004171static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004172 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173{
4174 int tofree;
4175
Christoph Lameter1b552532006-03-22 00:09:07 -08004176 if (!ac || !ac->avail)
4177 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 if (ac->touched && !force) {
4179 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004180 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004181 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004182 if (ac->avail) {
4183 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4184 if (tofree > ac->avail)
4185 tofree = (ac->avail + 1) / 2;
4186 free_block(cachep, ac->entry, tofree, node);
4187 ac->avail -= tofree;
4188 memmove(ac->entry, &(ac->entry[tofree]),
4189 sizeof(void *) * ac->avail);
4190 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004191 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 }
4193}
4194
4195/**
4196 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004197 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198 *
4199 * Called from workqueue/eventd every few seconds.
4200 * Purpose:
4201 * - clear the per-cpu caches for this CPU.
4202 * - return freeable pages to the main free memory pool.
4203 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004204 * If we cannot acquire the cache chain mutex then just give up - we'll try
4205 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004207static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004209 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004210 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004211 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004212 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004214 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004216 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004217
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004218 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219 check_irq_on();
4220
Christoph Lameter35386e32006-03-22 00:09:05 -08004221 /*
4222 * We only take the l3 lock if absolutely necessary and we
4223 * have established with reasonable certainty that
4224 * we can do some work if the lock was obtained.
4225 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004226 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004227
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004228 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229
Christoph Lameteraab22072006-03-22 00:09:06 -08004230 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231
Christoph Lameter35386e32006-03-22 00:09:05 -08004232 /*
4233 * These are racy checks but it does not matter
4234 * if we skip one check or scan twice.
4235 */
Christoph Lametere498be72005-09-09 13:03:32 -07004236 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004237 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238
Christoph Lametere498be72005-09-09 13:03:32 -07004239 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240
Christoph Lameteraab22072006-03-22 00:09:06 -08004241 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004242
Christoph Lametered11d9e2006-06-30 01:55:45 -07004243 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004244 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004245 else {
4246 int freed;
4247
4248 freed = drain_freelist(searchp, l3, (l3->free_limit +
4249 5 * searchp->num - 1) / (5 * searchp->num));
4250 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004252next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253 cond_resched();
4254 }
4255 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004256 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004257 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004258out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004259 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004260 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261}
4262
Linus Torvalds158a9622008-01-02 13:04:48 -08004263#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264
Pekka Enberg85289f92006-01-08 01:00:36 -08004265static void print_slabinfo_header(struct seq_file *m)
4266{
4267 /*
4268 * Output format version, so at least we can change it
4269 * without _too_ many complaints.
4270 */
4271#if STATS
4272 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4273#else
4274 seq_puts(m, "slabinfo - version: 2.1\n");
4275#endif
4276 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4277 "<objperslab> <pagesperslab>");
4278 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4279 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4280#if STATS
4281 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004282 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004283 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4284#endif
4285 seq_putc(m, '\n');
4286}
4287
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288static void *s_start(struct seq_file *m, loff_t *pos)
4289{
4290 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004292 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004293 if (!n)
4294 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004295
4296 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297}
4298
4299static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4300{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004301 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302}
4303
4304static void s_stop(struct seq_file *m, void *p)
4305{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004306 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307}
4308
4309static int s_show(struct seq_file *m, void *p)
4310{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004311 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004312 struct slab *slabp;
4313 unsigned long active_objs;
4314 unsigned long num_objs;
4315 unsigned long active_slabs = 0;
4316 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004317 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004318 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004319 int node;
4320 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321
Linus Torvalds1da177e2005-04-16 15:20:36 -07004322 active_objs = 0;
4323 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004324 for_each_online_node(node) {
4325 l3 = cachep->nodelists[node];
4326 if (!l3)
4327 continue;
4328
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004329 check_irq_on();
4330 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004331
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004332 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004333 if (slabp->inuse != cachep->num && !error)
4334 error = "slabs_full accounting error";
4335 active_objs += cachep->num;
4336 active_slabs++;
4337 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004338 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004339 if (slabp->inuse == cachep->num && !error)
4340 error = "slabs_partial inuse accounting error";
4341 if (!slabp->inuse && !error)
4342 error = "slabs_partial/inuse accounting error";
4343 active_objs += slabp->inuse;
4344 active_slabs++;
4345 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004346 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004347 if (slabp->inuse && !error)
4348 error = "slabs_free/inuse accounting error";
4349 num_slabs++;
4350 }
4351 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004352 if (l3->shared)
4353 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004354
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004355 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004357 num_slabs += active_slabs;
4358 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004359 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004360 error = "free_objects accounting error";
4361
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004362 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004363 if (error)
4364 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4365
4366 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004367 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004368 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004369 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004370 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004371 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004372 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004374 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375 unsigned long high = cachep->high_mark;
4376 unsigned long allocs = cachep->num_allocations;
4377 unsigned long grown = cachep->grown;
4378 unsigned long reaped = cachep->reaped;
4379 unsigned long errors = cachep->errors;
4380 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004382 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004383 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004384
Joe Perchese92dd4f2010-03-26 19:27:58 -07004385 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4386 "%4lu %4lu %4lu %4lu %4lu",
4387 allocs, high, grown,
4388 reaped, errors, max_freeable, node_allocs,
4389 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004390 }
4391 /* cpu stats */
4392 {
4393 unsigned long allochit = atomic_read(&cachep->allochit);
4394 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4395 unsigned long freehit = atomic_read(&cachep->freehit);
4396 unsigned long freemiss = atomic_read(&cachep->freemiss);
4397
4398 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004399 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004400 }
4401#endif
4402 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004403 return 0;
4404}
4405
4406/*
4407 * slabinfo_op - iterator that generates /proc/slabinfo
4408 *
4409 * Output layout:
4410 * cache-name
4411 * num-active-objs
4412 * total-objs
4413 * object size
4414 * num-active-slabs
4415 * total-slabs
4416 * num-pages-per-slab
4417 * + further values on SMP and with statistics enabled
4418 */
4419
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004420static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004421 .start = s_start,
4422 .next = s_next,
4423 .stop = s_stop,
4424 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004425};
4426
4427#define MAX_SLABINFO_WRITE 128
4428/**
4429 * slabinfo_write - Tuning for the slab allocator
4430 * @file: unused
4431 * @buffer: user buffer
4432 * @count: data length
4433 * @ppos: unused
4434 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004435static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004436 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004437{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004438 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004439 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004440 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004441
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442 if (count > MAX_SLABINFO_WRITE)
4443 return -EINVAL;
4444 if (copy_from_user(&kbuf, buffer, count))
4445 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004446 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004447
4448 tmp = strchr(kbuf, ' ');
4449 if (!tmp)
4450 return -EINVAL;
4451 *tmp = '\0';
4452 tmp++;
4453 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4454 return -EINVAL;
4455
4456 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004457 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004459 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004461 if (limit < 1 || batchcount < 1 ||
4462 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004463 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004464 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004465 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004466 batchcount, shared,
4467 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004468 }
4469 break;
4470 }
4471 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004472 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004473 if (res >= 0)
4474 res = count;
4475 return res;
4476}
Al Viro871751e2006-03-25 03:06:39 -08004477
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004478static int slabinfo_open(struct inode *inode, struct file *file)
4479{
4480 return seq_open(file, &slabinfo_op);
4481}
4482
4483static const struct file_operations proc_slabinfo_operations = {
4484 .open = slabinfo_open,
4485 .read = seq_read,
4486 .write = slabinfo_write,
4487 .llseek = seq_lseek,
4488 .release = seq_release,
4489};
4490
Al Viro871751e2006-03-25 03:06:39 -08004491#ifdef CONFIG_DEBUG_SLAB_LEAK
4492
4493static void *leaks_start(struct seq_file *m, loff_t *pos)
4494{
Al Viro871751e2006-03-25 03:06:39 -08004495 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004496 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004497}
4498
4499static inline int add_caller(unsigned long *n, unsigned long v)
4500{
4501 unsigned long *p;
4502 int l;
4503 if (!v)
4504 return 1;
4505 l = n[1];
4506 p = n + 2;
4507 while (l) {
4508 int i = l/2;
4509 unsigned long *q = p + 2 * i;
4510 if (*q == v) {
4511 q[1]++;
4512 return 1;
4513 }
4514 if (*q > v) {
4515 l = i;
4516 } else {
4517 p = q + 2;
4518 l -= i + 1;
4519 }
4520 }
4521 if (++n[1] == n[0])
4522 return 0;
4523 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4524 p[0] = v;
4525 p[1] = 1;
4526 return 1;
4527}
4528
4529static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4530{
4531 void *p;
4532 int i;
4533 if (n[0] == n[1])
4534 return;
4535 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4536 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4537 continue;
4538 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4539 return;
4540 }
4541}
4542
4543static void show_symbol(struct seq_file *m, unsigned long address)
4544{
4545#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004546 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004547 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004548
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004549 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004550 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004551 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004552 seq_printf(m, " [%s]", modname);
4553 return;
4554 }
4555#endif
4556 seq_printf(m, "%p", (void *)address);
4557}
4558
4559static int leaks_show(struct seq_file *m, void *p)
4560{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004561 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004562 struct slab *slabp;
4563 struct kmem_list3 *l3;
4564 const char *name;
4565 unsigned long *n = m->private;
4566 int node;
4567 int i;
4568
4569 if (!(cachep->flags & SLAB_STORE_USER))
4570 return 0;
4571 if (!(cachep->flags & SLAB_RED_ZONE))
4572 return 0;
4573
4574 /* OK, we can do it */
4575
4576 n[1] = 0;
4577
4578 for_each_online_node(node) {
4579 l3 = cachep->nodelists[node];
4580 if (!l3)
4581 continue;
4582
4583 check_irq_on();
4584 spin_lock_irq(&l3->list_lock);
4585
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004586 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004587 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004588 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004589 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004590 spin_unlock_irq(&l3->list_lock);
4591 }
4592 name = cachep->name;
4593 if (n[0] == n[1]) {
4594 /* Increase the buffer size */
4595 mutex_unlock(&cache_chain_mutex);
4596 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4597 if (!m->private) {
4598 /* Too bad, we are really out */
4599 m->private = n;
4600 mutex_lock(&cache_chain_mutex);
4601 return -ENOMEM;
4602 }
4603 *(unsigned long *)m->private = n[0] * 2;
4604 kfree(n);
4605 mutex_lock(&cache_chain_mutex);
4606 /* Now make sure this entry will be retried */
4607 m->count = m->size;
4608 return 0;
4609 }
4610 for (i = 0; i < n[1]; i++) {
4611 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4612 show_symbol(m, n[2*i+2]);
4613 seq_putc(m, '\n');
4614 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004615
Al Viro871751e2006-03-25 03:06:39 -08004616 return 0;
4617}
4618
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004619static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004620 .start = leaks_start,
4621 .next = s_next,
4622 .stop = s_stop,
4623 .show = leaks_show,
4624};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004625
4626static int slabstats_open(struct inode *inode, struct file *file)
4627{
4628 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4629 int ret = -ENOMEM;
4630 if (n) {
4631 ret = seq_open(file, &slabstats_op);
4632 if (!ret) {
4633 struct seq_file *m = file->private_data;
4634 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4635 m->private = n;
4636 n = NULL;
4637 }
4638 kfree(n);
4639 }
4640 return ret;
4641}
4642
4643static const struct file_operations proc_slabstats_operations = {
4644 .open = slabstats_open,
4645 .read = seq_read,
4646 .llseek = seq_lseek,
4647 .release = seq_release_private,
4648};
Al Viro871751e2006-03-25 03:06:39 -08004649#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004650
4651static int __init slab_proc_init(void)
4652{
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004653 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004654#ifdef CONFIG_DEBUG_SLAB_LEAK
4655 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4656#endif
4657 return 0;
4658}
4659module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004660#endif
4661
Manfred Spraul00e145b2005-09-03 15:55:07 -07004662/**
4663 * ksize - get the actual amount of memory allocated for a given object
4664 * @objp: Pointer to the object
4665 *
4666 * kmalloc may internally round up allocations and return more memory
4667 * than requested. ksize() can be used to determine the actual amount of
4668 * memory allocated. The caller may use this additional memory, even though
4669 * a smaller amount of memory was initially specified with the kmalloc call.
4670 * The caller must guarantee that objp points to a valid object previously
4671 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4672 * must not be freed during the duration of the call.
4673 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004674size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004675{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004676 BUG_ON(!objp);
4677 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004678 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004679
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004680 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004681}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004682EXPORT_SYMBOL(ksize);