blob: 856e4a192d25c73954771b1b0893458f0d9209bb [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
Christoph Lameter18004c52012-07-06 15:25:12 -050071 * The global cache-chain is protected by the mutex 'slab_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
Mel Gorman381760e2012-07-31 16:44:30 -0700120#include <net/sock.h>
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500126#include <trace/events/kmem.h>
127
Mel Gorman072bb0a2012-07-31 16:43:58 -0700128#include "internal.h"
129
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800130#include "slab.h"
131
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
Mel Gorman072bb0a2012-07-31 16:43:58 -0700160/*
161 * true if a page was allocated from pfmemalloc reserves for network-based
162 * swap
163 */
164static bool pfmemalloc_active __read_mostly;
165
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166/*
167 * kmem_bufctl_t:
168 *
169 * Bufctl's are used for linking objs within a slab
170 * linked offsets.
171 *
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
183 */
184
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700185typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800188#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 * struct slab_rcu
193 *
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
201 *
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 */
205struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800206 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800207 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800208 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209};
210
211/*
Lai Jiangshan5bfe53a2011-03-10 15:22:24 +0800212 * struct slab
213 *
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
217 */
218struct slab {
219 union {
220 struct {
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
227 };
228 struct slab_rcu __slab_cover_slab_rcu;
229 };
230};
231
232/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 * struct array_cache
234 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 * Purpose:
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
239 *
240 * The limit is stored in the per-cpu structure to reduce the data cache
241 * footprint.
242 *
243 */
244struct array_cache {
245 unsigned int avail;
246 unsigned int limit;
247 unsigned int batchcount;
248 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700249 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700250 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
253 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700254 *
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800258 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259};
260
Mel Gorman072bb0a2012-07-31 16:43:58 -0700261#define SLAB_OBJ_PFMEMALLOC 1
262static inline bool is_obj_pfmemalloc(void *objp)
263{
264 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
265}
266
267static inline void set_obj_pfmemalloc(void **objp)
268{
269 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
270 return;
271}
272
273static inline void clear_obj_pfmemalloc(void **objp)
274{
275 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
276}
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278/*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282#define BOOT_CPUCACHE_ENTRIES 1
283struct arraycache_init {
284 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800285 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286};
287
288/*
Christoph Lametere498be72005-09-09 13:03:32 -0700289 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 */
291struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800297 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303};
304
Christoph Lametere498be72005-09-09 13:03:32 -0700305/*
306 * Need this for bootstrapping a per node allocator.
307 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200308#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
H Hartley Sweeten68a1b192011-01-11 17:49:32 -0600309static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700310#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200311#define SIZE_AC MAX_NUMNODES
312#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Christoph Lametered11d9e2006-06-30 01:55:45 -0700314static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300318static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000319static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700320
Christoph Lametere498be72005-09-09 13:03:32 -0700321/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700324 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700325static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700326{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800327 extern void __bad_size(void);
328
Christoph Lametere498be72005-09-09 13:03:32 -0700329 if (__builtin_constant_p(size)) {
330 int i = 0;
331
332#define CACHE(x) \
333 if (size <=x) \
334 return i; \
335 else \
336 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800337#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700338#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700340 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700342 return 0;
343}
344
Ingo Molnare0a42722006-06-23 02:03:46 -0700345static int slab_early_init = 1;
346
Christoph Lametere498be72005-09-09 13:03:32 -0700347#define INDEX_AC index_of(sizeof(struct arraycache_init))
348#define INDEX_L3 index_of(sizeof(struct kmem_list3))
349
Pekka Enberg5295a742006-02-01 03:05:48 -0800350static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700351{
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800357 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
361}
362
Andrew Mortona737b3e2006-03-22 00:08:11 -0800363#define MAKE_LIST(cachep, listp, slab, nodeid) \
364 do { \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700367 } while (0)
368
Andrew Mortona737b3e2006-03-22 00:08:11 -0800369#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376#define CFLGS_OFF_SLAB (0x80000000UL)
377#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
378
379#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800380/*
381 * Optimization question: fewer reaps means less probability for unnessary
382 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100384 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 * which could lock up otherwise freeable slabs.
386 */
387#define REAPTIMEOUT_CPUC (2*HZ)
388#define REAPTIMEOUT_LIST3 (4*HZ)
389
390#if STATS
391#define STATS_INC_ACTIVE(x) ((x)->num_active++)
392#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
393#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
394#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700395#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800396#define STATS_SET_HIGH(x) \
397 do { \
398 if ((x)->num_active > (x)->high_mark) \
399 (x)->high_mark = (x)->num_active; \
400 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401#define STATS_INC_ERR(x) ((x)->errors++)
402#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700403#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700404#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800405#define STATS_SET_FREEABLE(x, i) \
406 do { \
407 if ((x)->max_freeable < i) \
408 (x)->max_freeable = i; \
409 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
411#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
412#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
413#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
414#else
415#define STATS_INC_ACTIVE(x) do { } while (0)
416#define STATS_DEC_ACTIVE(x) do { } while (0)
417#define STATS_INC_ALLOCED(x) do { } while (0)
418#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700419#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420#define STATS_SET_HIGH(x) do { } while (0)
421#define STATS_INC_ERR(x) do { } while (0)
422#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700423#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700424#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800425#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#define STATS_INC_ALLOCHIT(x) do { } while (0)
427#define STATS_INC_ALLOCMISS(x) do { } while (0)
428#define STATS_INC_FREEHIT(x) do { } while (0)
429#define STATS_INC_FREEMISS(x) do { } while (0)
430#endif
431
432#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Andrew Mortona737b3e2006-03-22 00:08:11 -0800434/*
435 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800437 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 * the end of an object is aligned with the end of the real
439 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800440 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800442 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500443 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
444 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800445 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800447static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800449 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450}
451
David Woodhouseb46b8f12007-05-08 00:22:59 -0700452static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700455 return (unsigned long long*) (objp + obj_offset(cachep) -
456 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
David Woodhouseb46b8f12007-05-08 00:22:59 -0700459static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
462 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500463 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700464 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400465 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500466 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700467 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468}
469
Pekka Enberg343e0d72006-02-01 03:05:50 -0800470static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471{
472 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500473 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474}
475
476#else
477
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800478#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700479#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
480#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
482
483#endif
484
485/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700486 * Do not go above this order unless 0 objects fit into the slab or
487 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 */
David Rientjes543585c2011-10-18 22:09:24 -0700489#define SLAB_MAX_ORDER_HI 1
490#define SLAB_MAX_ORDER_LO 0
491static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700492static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800494static inline struct kmem_cache *virt_to_cache(const void *obj)
495{
Christoph Lameterb49af682007-05-06 14:49:41 -0700496 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500497 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800498}
499
500static inline struct slab *virt_to_slab(const void *obj)
501{
Christoph Lameterb49af682007-05-06 14:49:41 -0700502 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500503
504 VM_BUG_ON(!PageSlab(page));
505 return page->slab_page;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800506}
507
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800508static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
509 unsigned int idx)
510{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500511 return slab->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800512}
513
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800514/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500515 * We want to avoid an expensive divide : (offset / cache->size)
516 * Using the fact that size is a constant for a particular cache,
517 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800518 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
519 */
520static inline unsigned int obj_to_index(const struct kmem_cache *cache,
521 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800522{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800523 u32 offset = (obj - slab->s_mem);
524 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800525}
526
Andrew Mortona737b3e2006-03-22 00:08:11 -0800527/*
528 * These are the default caches for kmalloc. Custom caches can have other sizes.
529 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530struct cache_sizes malloc_sizes[] = {
531#define CACHE(x) { .cs_size = (x) },
532#include <linux/kmalloc_sizes.h>
533 CACHE(ULONG_MAX)
534#undef CACHE
535};
536EXPORT_SYMBOL(malloc_sizes);
537
538/* Must match cache_sizes above. Out of line to keep cache footprint low. */
539struct cache_names {
540 char *name;
541 char *name_dma;
542};
543
544static struct cache_names __initdata cache_names[] = {
545#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
546#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800547 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548#undef CACHE
549};
550
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800552 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553
554/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000555static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800556 .batchcount = 1,
557 .limit = BOOT_CPUCACHE_ENTRIES,
558 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500559 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800560 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561};
562
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700563#define BAD_ALIEN_MAGIC 0x01020304ul
564
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200565#ifdef CONFIG_LOCKDEP
566
567/*
568 * Slab sometimes uses the kmalloc slabs to store the slab headers
569 * for other slabs "off slab".
570 * The locking for this is tricky in that it nests within the locks
571 * of all other slabs in a few places; to deal with this special
572 * locking we put on-slab caches into a separate lock-class.
573 *
574 * We set lock class for alien array caches which are up during init.
575 * The lock annotation will be lost if all cpus of a node goes down and
576 * then comes back up during hotplug
577 */
578static struct lock_class_key on_slab_l3_key;
579static struct lock_class_key on_slab_alc_key;
580
Peter Zijlstra83835b32011-07-22 15:26:05 +0200581static struct lock_class_key debugobj_l3_key;
582static struct lock_class_key debugobj_alc_key;
583
584static void slab_set_lock_classes(struct kmem_cache *cachep,
585 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
586 int q)
587{
588 struct array_cache **alc;
589 struct kmem_list3 *l3;
590 int r;
591
592 l3 = cachep->nodelists[q];
593 if (!l3)
594 return;
595
596 lockdep_set_class(&l3->list_lock, l3_key);
597 alc = l3->alien;
598 /*
599 * FIXME: This check for BAD_ALIEN_MAGIC
600 * should go away when common slab code is taught to
601 * work even without alien caches.
602 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
603 * for alloc_alien_cache,
604 */
605 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
606 return;
607 for_each_node(r) {
608 if (alc[r])
609 lockdep_set_class(&alc[r]->lock, alc_key);
610 }
611}
612
613static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
614{
615 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
616}
617
618static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
619{
620 int node;
621
622 for_each_online_node(node)
623 slab_set_debugobj_lock_classes_node(cachep, node);
624}
625
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200626static void init_node_lock_keys(int q)
627{
628 struct cache_sizes *s = malloc_sizes;
629
Christoph Lameter97d06602012-07-06 15:25:11 -0500630 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200631 return;
632
633 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200634 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200635
636 l3 = s->cs_cachep->nodelists[q];
637 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200638 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200639
640 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
641 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200642 }
643}
644
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800645static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
646{
647 struct kmem_list3 *l3;
648 l3 = cachep->nodelists[q];
649 if (!l3)
650 return;
651
652 slab_set_lock_classes(cachep, &on_slab_l3_key,
653 &on_slab_alc_key, q);
654}
655
656static inline void on_slab_lock_classes(struct kmem_cache *cachep)
657{
658 int node;
659
660 VM_BUG_ON(OFF_SLAB(cachep));
661 for_each_node(node)
662 on_slab_lock_classes_node(cachep, node);
663}
664
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200665static inline void init_lock_keys(void)
666{
667 int node;
668
669 for_each_node(node)
670 init_node_lock_keys(node);
671}
672#else
673static void init_node_lock_keys(int q)
674{
675}
676
677static inline void init_lock_keys(void)
678{
679}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200680
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800681static inline void on_slab_lock_classes(struct kmem_cache *cachep)
682{
683}
684
685static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
686{
687}
688
Peter Zijlstra83835b32011-07-22 15:26:05 +0200689static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
690{
691}
692
693static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
694{
695}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200696#endif
697
Tejun Heo1871e522009-10-29 22:34:13 +0900698static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
Pekka Enberg343e0d72006-02-01 03:05:50 -0800700static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701{
702 return cachep->array[smp_processor_id()];
703}
704
Andrew Mortona737b3e2006-03-22 00:08:11 -0800705static inline struct kmem_cache *__find_general_cachep(size_t size,
706 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707{
708 struct cache_sizes *csizep = malloc_sizes;
709
710#if DEBUG
711 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800712 * kmem_cache_create(), or __kmalloc(), before
713 * the generic caches are initialized.
714 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700715 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700717 if (!size)
718 return ZERO_SIZE_PTR;
719
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 while (size > csizep->cs_size)
721 csizep++;
722
723 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700724 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 * has cs_{dma,}cachep==NULL. Thus no special case
726 * for large kmalloc calls required.
727 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800728#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 if (unlikely(gfpflags & GFP_DMA))
730 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800731#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 return csizep->cs_cachep;
733}
734
Adrian Bunkb2213852006-09-25 23:31:02 -0700735static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700736{
737 return __find_general_cachep(size, gfpflags);
738}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700739
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800740static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800742 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
743}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744
Andrew Mortona737b3e2006-03-22 00:08:11 -0800745/*
746 * Calculate the number of objects and left-over bytes for a given buffer size.
747 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800748static void cache_estimate(unsigned long gfporder, size_t buffer_size,
749 size_t align, int flags, size_t *left_over,
750 unsigned int *num)
751{
752 int nr_objs;
753 size_t mgmt_size;
754 size_t slab_size = PAGE_SIZE << gfporder;
755
756 /*
757 * The slab management structure can be either off the slab or
758 * on it. For the latter case, the memory allocated for a
759 * slab is used for:
760 *
761 * - The struct slab
762 * - One kmem_bufctl_t for each object
763 * - Padding to respect alignment of @align
764 * - @buffer_size bytes for each object
765 *
766 * If the slab management structure is off the slab, then the
767 * alignment will already be calculated into the size. Because
768 * the slabs are all pages aligned, the objects will be at the
769 * correct alignment when allocated.
770 */
771 if (flags & CFLGS_OFF_SLAB) {
772 mgmt_size = 0;
773 nr_objs = slab_size / buffer_size;
774
775 if (nr_objs > SLAB_LIMIT)
776 nr_objs = SLAB_LIMIT;
777 } else {
778 /*
779 * Ignore padding for the initial guess. The padding
780 * is at most @align-1 bytes, and @buffer_size is at
781 * least @align. In the worst case, this result will
782 * be one greater than the number of objects that fit
783 * into the memory allocation when taking the padding
784 * into account.
785 */
786 nr_objs = (slab_size - sizeof(struct slab)) /
787 (buffer_size + sizeof(kmem_bufctl_t));
788
789 /*
790 * This calculated number will be either the right
791 * amount, or one greater than what we want.
792 */
793 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
794 > slab_size)
795 nr_objs--;
796
797 if (nr_objs > SLAB_LIMIT)
798 nr_objs = SLAB_LIMIT;
799
800 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800802 *num = nr_objs;
803 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804}
805
Christoph Lameterf28510d2012-09-11 19:49:38 +0000806#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700807#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
Andrew Mortona737b3e2006-03-22 00:08:11 -0800809static void __slab_error(const char *function, struct kmem_cache *cachep,
810 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811{
812 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800813 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030815 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000817#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818
Paul Menage3395ee02006-12-06 20:32:16 -0800819/*
820 * By default on NUMA we use alien caches to stage the freeing of
821 * objects allocated from other nodes. This causes massive memory
822 * inefficiencies when using fake NUMA setup to split memory into a
823 * large number of small nodes, so it can be disabled on the command
824 * line
825 */
826
827static int use_alien_caches __read_mostly = 1;
828static int __init noaliencache_setup(char *s)
829{
830 use_alien_caches = 0;
831 return 1;
832}
833__setup("noaliencache", noaliencache_setup);
834
David Rientjes3df1ccc2011-10-18 22:09:28 -0700835static int __init slab_max_order_setup(char *str)
836{
837 get_option(&str, &slab_max_order);
838 slab_max_order = slab_max_order < 0 ? 0 :
839 min(slab_max_order, MAX_ORDER - 1);
840 slab_max_order_set = true;
841
842 return 1;
843}
844__setup("slab_max_order=", slab_max_order_setup);
845
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800846#ifdef CONFIG_NUMA
847/*
848 * Special reaping functions for NUMA systems called from cache_reap().
849 * These take care of doing round robin flushing of alien caches (containing
850 * objects freed on different nodes from which they were allocated) and the
851 * flushing of remote pcps by calling drain_node_pages.
852 */
Tejun Heo1871e522009-10-29 22:34:13 +0900853static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800854
855static void init_reap_node(int cpu)
856{
857 int node;
858
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700859 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800860 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800861 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800862
Tejun Heo1871e522009-10-29 22:34:13 +0900863 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800864}
865
866static void next_reap_node(void)
867{
Christoph Lameter909ea962010-12-08 16:22:55 +0100868 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800869
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800870 node = next_node(node, node_online_map);
871 if (unlikely(node >= MAX_NUMNODES))
872 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100873 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800874}
875
876#else
877#define init_reap_node(cpu) do { } while (0)
878#define next_reap_node(void) do { } while (0)
879#endif
880
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881/*
882 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
883 * via the workqueue/eventd.
884 * Add the CPU number into the expiration time to minimize the possibility of
885 * the CPUs getting into lockstep and contending for the global cache chain
886 * lock.
887 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700888static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
Tejun Heo1871e522009-10-29 22:34:13 +0900890 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891
892 /*
893 * When this gets called from do_initcalls via cpucache_init(),
894 * init_workqueues() has already run, so keventd will be setup
895 * at that time.
896 */
David Howells52bad642006-11-22 14:54:01 +0000897 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800898 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700899 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800900 schedule_delayed_work_on(cpu, reap_work,
901 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903}
904
Christoph Lametere498be72005-09-09 13:03:32 -0700905static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300906 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800908 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 struct array_cache *nc = NULL;
910
Pekka Enberg83b519e2009-06-10 19:40:04 +0300911 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100912 /*
913 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300914 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100915 * cache the pointers are not cleared and they could be counted as
916 * valid references during a kmemleak scan. Therefore, kmemleak must
917 * not scan such objects.
918 */
919 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 if (nc) {
921 nc->avail = 0;
922 nc->limit = entries;
923 nc->batchcount = batchcount;
924 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700925 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 }
927 return nc;
928}
929
Mel Gorman072bb0a2012-07-31 16:43:58 -0700930static inline bool is_slab_pfmemalloc(struct slab *slabp)
931{
932 struct page *page = virt_to_page(slabp->s_mem);
933
934 return PageSlabPfmemalloc(page);
935}
936
937/* Clears pfmemalloc_active if no slabs have pfmalloc set */
938static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
939 struct array_cache *ac)
940{
941 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
942 struct slab *slabp;
943 unsigned long flags;
944
945 if (!pfmemalloc_active)
946 return;
947
948 spin_lock_irqsave(&l3->list_lock, flags);
949 list_for_each_entry(slabp, &l3->slabs_full, list)
950 if (is_slab_pfmemalloc(slabp))
951 goto out;
952
953 list_for_each_entry(slabp, &l3->slabs_partial, list)
954 if (is_slab_pfmemalloc(slabp))
955 goto out;
956
957 list_for_each_entry(slabp, &l3->slabs_free, list)
958 if (is_slab_pfmemalloc(slabp))
959 goto out;
960
961 pfmemalloc_active = false;
962out:
963 spin_unlock_irqrestore(&l3->list_lock, flags);
964}
965
Mel Gorman381760e2012-07-31 16:44:30 -0700966static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700967 gfp_t flags, bool force_refill)
968{
969 int i;
970 void *objp = ac->entry[--ac->avail];
971
972 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
973 if (unlikely(is_obj_pfmemalloc(objp))) {
974 struct kmem_list3 *l3;
975
976 if (gfp_pfmemalloc_allowed(flags)) {
977 clear_obj_pfmemalloc(&objp);
978 return objp;
979 }
980
981 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700982 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700983 /* If a !PFMEMALLOC object is found, swap them */
984 if (!is_obj_pfmemalloc(ac->entry[i])) {
985 objp = ac->entry[i];
986 ac->entry[i] = ac->entry[ac->avail];
987 ac->entry[ac->avail] = objp;
988 return objp;
989 }
990 }
991
992 /*
993 * If there are empty slabs on the slabs_free list and we are
994 * being forced to refill the cache, mark this one !pfmemalloc.
995 */
996 l3 = cachep->nodelists[numa_mem_id()];
997 if (!list_empty(&l3->slabs_free) && force_refill) {
998 struct slab *slabp = virt_to_slab(objp);
Mel Gorman30c29be2012-09-17 14:09:03 -0700999 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
Mel Gorman072bb0a2012-07-31 16:43:58 -07001000 clear_obj_pfmemalloc(&objp);
1001 recheck_pfmemalloc_active(cachep, ac);
1002 return objp;
1003 }
1004
1005 /* No !PFMEMALLOC objects available */
1006 ac->avail++;
1007 objp = NULL;
1008 }
1009
1010 return objp;
1011}
1012
Mel Gorman381760e2012-07-31 16:44:30 -07001013static inline void *ac_get_obj(struct kmem_cache *cachep,
1014 struct array_cache *ac, gfp_t flags, bool force_refill)
1015{
1016 void *objp;
1017
1018 if (unlikely(sk_memalloc_socks()))
1019 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1020 else
1021 objp = ac->entry[--ac->avail];
1022
1023 return objp;
1024}
1025
1026static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -07001027 void *objp)
1028{
1029 if (unlikely(pfmemalloc_active)) {
1030 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -07001031 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001032 if (PageSlabPfmemalloc(page))
1033 set_obj_pfmemalloc(&objp);
1034 }
1035
Mel Gorman381760e2012-07-31 16:44:30 -07001036 return objp;
1037}
1038
1039static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1040 void *objp)
1041{
1042 if (unlikely(sk_memalloc_socks()))
1043 objp = __ac_put_obj(cachep, ac, objp);
1044
Mel Gorman072bb0a2012-07-31 16:43:58 -07001045 ac->entry[ac->avail++] = objp;
1046}
1047
Christoph Lameter3ded1752006-03-25 03:06:44 -08001048/*
1049 * Transfer objects in one arraycache to another.
1050 * Locking must be handled by the caller.
1051 *
1052 * Return the number of entries transferred.
1053 */
1054static int transfer_objects(struct array_cache *to,
1055 struct array_cache *from, unsigned int max)
1056{
1057 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -07001058 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -08001059
1060 if (!nr)
1061 return 0;
1062
1063 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1064 sizeof(void *) *nr);
1065
1066 from->avail -= nr;
1067 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -08001068 return nr;
1069}
1070
Christoph Lameter765c4502006-09-27 01:50:08 -07001071#ifndef CONFIG_NUMA
1072
1073#define drain_alien_cache(cachep, alien) do { } while (0)
1074#define reap_alien(cachep, l3) do { } while (0)
1075
Pekka Enberg83b519e2009-06-10 19:40:04 +03001076static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -07001077{
1078 return (struct array_cache **)BAD_ALIEN_MAGIC;
1079}
1080
1081static inline void free_alien_cache(struct array_cache **ac_ptr)
1082{
1083}
1084
1085static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1086{
1087 return 0;
1088}
1089
1090static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1091 gfp_t flags)
1092{
1093 return NULL;
1094}
1095
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001096static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001097 gfp_t flags, int nodeid)
1098{
1099 return NULL;
1100}
1101
1102#else /* CONFIG_NUMA */
1103
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001104static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001105static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001106
Pekka Enberg83b519e2009-06-10 19:40:04 +03001107static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001108{
1109 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001110 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001111 int i;
1112
1113 if (limit > 1)
1114 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001115 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001116 if (ac_ptr) {
1117 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001118 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001119 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001120 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001121 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001122 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001123 kfree(ac_ptr[i]);
1124 kfree(ac_ptr);
1125 return NULL;
1126 }
1127 }
1128 }
1129 return ac_ptr;
1130}
1131
Pekka Enberg5295a742006-02-01 03:05:48 -08001132static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001133{
1134 int i;
1135
1136 if (!ac_ptr)
1137 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001138 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001139 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001140 kfree(ac_ptr);
1141}
1142
Pekka Enberg343e0d72006-02-01 03:05:50 -08001143static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001144 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001145{
1146 struct kmem_list3 *rl3 = cachep->nodelists[node];
1147
1148 if (ac->avail) {
1149 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001150 /*
1151 * Stuff objects into the remote nodes shared array first.
1152 * That way we could avoid the overhead of putting the objects
1153 * into the free lists and getting them back later.
1154 */
shin, jacob693f7d32006-04-28 10:54:37 -05001155 if (rl3->shared)
1156 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001157
Christoph Lameterff694162005-09-22 21:44:02 -07001158 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001159 ac->avail = 0;
1160 spin_unlock(&rl3->list_lock);
1161 }
1162}
1163
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001164/*
1165 * Called from cache_reap() to regularly drain alien caches round robin.
1166 */
1167static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1168{
Christoph Lameter909ea962010-12-08 16:22:55 +01001169 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001170
1171 if (l3->alien) {
1172 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001173
1174 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001175 __drain_alien_cache(cachep, ac, node);
1176 spin_unlock_irq(&ac->lock);
1177 }
1178 }
1179}
1180
Andrew Mortona737b3e2006-03-22 00:08:11 -08001181static void drain_alien_cache(struct kmem_cache *cachep,
1182 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001183{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001184 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001185 struct array_cache *ac;
1186 unsigned long flags;
1187
1188 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001189 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001190 if (ac) {
1191 spin_lock_irqsave(&ac->lock, flags);
1192 __drain_alien_cache(cachep, ac, i);
1193 spin_unlock_irqrestore(&ac->lock, flags);
1194 }
1195 }
1196}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001197
Ingo Molnar873623d2006-07-13 14:44:38 +02001198static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001199{
1200 struct slab *slabp = virt_to_slab(objp);
1201 int nodeid = slabp->nodeid;
1202 struct kmem_list3 *l3;
1203 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001204 int node;
1205
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001206 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001207
1208 /*
1209 * Make sure we are not freeing a object from another node to the array
1210 * cache on this cpu.
1211 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001212 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001213 return 0;
1214
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001215 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001216 STATS_INC_NODEFREES(cachep);
1217 if (l3->alien && l3->alien[nodeid]) {
1218 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001219 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001220 if (unlikely(alien->avail == alien->limit)) {
1221 STATS_INC_ACOVERFLOW(cachep);
1222 __drain_alien_cache(cachep, alien, nodeid);
1223 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001224 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001225 spin_unlock(&alien->lock);
1226 } else {
1227 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1228 free_block(cachep, &objp, 1, nodeid);
1229 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1230 }
1231 return 1;
1232}
Christoph Lametere498be72005-09-09 13:03:32 -07001233#endif
1234
David Rientjes8f9f8d92010-03-27 19:40:47 -07001235/*
1236 * Allocates and initializes nodelists for a node on each slab cache, used for
1237 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1238 * will be allocated off-node since memory is not yet online for the new node.
1239 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1240 * already in use.
1241 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001242 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001243 */
1244static int init_cache_nodelists_node(int node)
1245{
1246 struct kmem_cache *cachep;
1247 struct kmem_list3 *l3;
1248 const int memsize = sizeof(struct kmem_list3);
1249
Christoph Lameter18004c52012-07-06 15:25:12 -05001250 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001251 /*
1252 * Set up the size64 kmemlist for cpu before we can
1253 * begin anything. Make sure some other cpu on this
1254 * node has not already allocated this
1255 */
1256 if (!cachep->nodelists[node]) {
1257 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1258 if (!l3)
1259 return -ENOMEM;
1260 kmem_list3_init(l3);
1261 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1262 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1263
1264 /*
1265 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001266 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001267 * protection here.
1268 */
1269 cachep->nodelists[node] = l3;
1270 }
1271
1272 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1273 cachep->nodelists[node]->free_limit =
1274 (1 + nr_cpus_node(node)) *
1275 cachep->batchcount + cachep->num;
1276 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1277 }
1278 return 0;
1279}
1280
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001281static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001283 struct kmem_cache *cachep;
1284 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001285 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301286 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001287
Christoph Lameter18004c52012-07-06 15:25:12 -05001288 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001289 struct array_cache *nc;
1290 struct array_cache *shared;
1291 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001292
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001293 /* cpu is dead; no one can alloc from it. */
1294 nc = cachep->array[cpu];
1295 cachep->array[cpu] = NULL;
1296 l3 = cachep->nodelists[node];
1297
1298 if (!l3)
1299 goto free_array_cache;
1300
1301 spin_lock_irq(&l3->list_lock);
1302
1303 /* Free limit for this kmem_list3 */
1304 l3->free_limit -= cachep->batchcount;
1305 if (nc)
1306 free_block(cachep, nc->entry, nc->avail, node);
1307
Rusty Russell58463c12009-12-17 11:43:12 -06001308 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001309 spin_unlock_irq(&l3->list_lock);
1310 goto free_array_cache;
1311 }
1312
1313 shared = l3->shared;
1314 if (shared) {
1315 free_block(cachep, shared->entry,
1316 shared->avail, node);
1317 l3->shared = NULL;
1318 }
1319
1320 alien = l3->alien;
1321 l3->alien = NULL;
1322
1323 spin_unlock_irq(&l3->list_lock);
1324
1325 kfree(shared);
1326 if (alien) {
1327 drain_alien_cache(cachep, alien);
1328 free_alien_cache(alien);
1329 }
1330free_array_cache:
1331 kfree(nc);
1332 }
1333 /*
1334 * In the previous loop, all the objects were freed to
1335 * the respective cache's slabs, now we can go ahead and
1336 * shrink each nodelist to its limit.
1337 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001338 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001339 l3 = cachep->nodelists[node];
1340 if (!l3)
1341 continue;
1342 drain_freelist(cachep, l3, l3->free_objects);
1343 }
1344}
1345
1346static int __cpuinit cpuup_prepare(long cpu)
1347{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001348 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001349 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001350 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001351 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001353 /*
1354 * We need to do this right in the beginning since
1355 * alloc_arraycache's are going to use this list.
1356 * kmalloc_node allows us to add the slab to the right
1357 * kmem_list3 and not this cpu's kmem_list3
1358 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001359 err = init_cache_nodelists_node(node);
1360 if (err < 0)
1361 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001362
1363 /*
1364 * Now we can go ahead with allocating the shared arrays and
1365 * array caches
1366 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001367 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001368 struct array_cache *nc;
1369 struct array_cache *shared = NULL;
1370 struct array_cache **alien = NULL;
1371
1372 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001373 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001374 if (!nc)
1375 goto bad;
1376 if (cachep->shared) {
1377 shared = alloc_arraycache(node,
1378 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001379 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001380 if (!shared) {
1381 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001382 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001383 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001384 }
1385 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001386 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001387 if (!alien) {
1388 kfree(shared);
1389 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001390 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001391 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001392 }
1393 cachep->array[cpu] = nc;
1394 l3 = cachep->nodelists[node];
1395 BUG_ON(!l3);
1396
1397 spin_lock_irq(&l3->list_lock);
1398 if (!l3->shared) {
1399 /*
1400 * We are serialised from CPU_DEAD or
1401 * CPU_UP_CANCELLED by the cpucontrol lock
1402 */
1403 l3->shared = shared;
1404 shared = NULL;
1405 }
1406#ifdef CONFIG_NUMA
1407 if (!l3->alien) {
1408 l3->alien = alien;
1409 alien = NULL;
1410 }
1411#endif
1412 spin_unlock_irq(&l3->list_lock);
1413 kfree(shared);
1414 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001415 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1416 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001417 else if (!OFF_SLAB(cachep) &&
1418 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1419 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001420 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001421 init_node_lock_keys(node);
1422
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001423 return 0;
1424bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001425 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001426 return -ENOMEM;
1427}
1428
1429static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1430 unsigned long action, void *hcpu)
1431{
1432 long cpu = (long)hcpu;
1433 int err = 0;
1434
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001436 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001437 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001438 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001439 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001440 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 break;
1442 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001443 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 start_cpu_timer(cpu);
1445 break;
1446#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001447 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001448 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001449 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001450 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001451 * held so that if cache_reap() is invoked it cannot do
1452 * anything expensive but will only modify reap_work
1453 * and reschedule the timer.
1454 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001455 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001456 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001457 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001458 break;
1459 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001460 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001461 start_cpu_timer(cpu);
1462 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001464 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001465 /*
1466 * Even if all the cpus of a node are down, we don't free the
1467 * kmem_list3 of any cache. This to avoid a race between
1468 * cpu_down, and a kmalloc allocation from another cpu for
1469 * memory from the node of the cpu going down. The list3
1470 * structure is usually allocated from kmem_cache_create() and
1471 * gets destroyed at kmem_cache_destroy().
1472 */
Simon Arlott183ff222007-10-20 01:27:18 +02001473 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001474#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001476 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001477 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001478 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001479 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001482 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483}
1484
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001485static struct notifier_block __cpuinitdata cpucache_notifier = {
1486 &cpuup_callback, NULL, 0
1487};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488
David Rientjes8f9f8d92010-03-27 19:40:47 -07001489#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1490/*
1491 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1492 * Returns -EBUSY if all objects cannot be drained so that the node is not
1493 * removed.
1494 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001495 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001496 */
1497static int __meminit drain_cache_nodelists_node(int node)
1498{
1499 struct kmem_cache *cachep;
1500 int ret = 0;
1501
Christoph Lameter18004c52012-07-06 15:25:12 -05001502 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001503 struct kmem_list3 *l3;
1504
1505 l3 = cachep->nodelists[node];
1506 if (!l3)
1507 continue;
1508
1509 drain_freelist(cachep, l3, l3->free_objects);
1510
1511 if (!list_empty(&l3->slabs_full) ||
1512 !list_empty(&l3->slabs_partial)) {
1513 ret = -EBUSY;
1514 break;
1515 }
1516 }
1517 return ret;
1518}
1519
1520static int __meminit slab_memory_callback(struct notifier_block *self,
1521 unsigned long action, void *arg)
1522{
1523 struct memory_notify *mnb = arg;
1524 int ret = 0;
1525 int nid;
1526
1527 nid = mnb->status_change_nid;
1528 if (nid < 0)
1529 goto out;
1530
1531 switch (action) {
1532 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001533 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001534 ret = init_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001535 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001536 break;
1537 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001538 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001539 ret = drain_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001540 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001541 break;
1542 case MEM_ONLINE:
1543 case MEM_OFFLINE:
1544 case MEM_CANCEL_ONLINE:
1545 case MEM_CANCEL_OFFLINE:
1546 break;
1547 }
1548out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001549 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001550}
1551#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1552
Christoph Lametere498be72005-09-09 13:03:32 -07001553/*
1554 * swap the static kmem_list3 with kmalloced memory
1555 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001556static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1557 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001558{
1559 struct kmem_list3 *ptr;
1560
Pekka Enberg83b519e2009-06-10 19:40:04 +03001561 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001562 BUG_ON(!ptr);
1563
Christoph Lametere498be72005-09-09 13:03:32 -07001564 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001565 /*
1566 * Do not assume that spinlocks can be initialized via memcpy:
1567 */
1568 spin_lock_init(&ptr->list_lock);
1569
Christoph Lametere498be72005-09-09 13:03:32 -07001570 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1571 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001572}
1573
Andrew Mortona737b3e2006-03-22 00:08:11 -08001574/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001575 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1576 * size of kmem_list3.
1577 */
1578static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1579{
1580 int node;
1581
1582 for_each_online_node(node) {
1583 cachep->nodelists[node] = &initkmem_list3[index + node];
1584 cachep->nodelists[node]->next_reap = jiffies +
1585 REAPTIMEOUT_LIST3 +
1586 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1587 }
1588}
1589
1590/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001591 * The memory after the last cpu cache pointer is used for the
1592 * the nodelists pointer.
1593 */
1594static void setup_nodelists_pointer(struct kmem_cache *cachep)
1595{
1596 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
1597}
1598
1599/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001600 * Initialisation. Called after the page allocator have been initialised and
1601 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 */
1603void __init kmem_cache_init(void)
1604{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 struct cache_sizes *sizes;
1606 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001607 int i;
1608
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001609 kmem_cache = &kmem_cache_boot;
Christoph Lameter3c583462012-11-28 16:23:01 +00001610 setup_nodelists_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001611
Mel Gormanb6e68bc12009-06-16 15:32:16 -07001612 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001613 use_alien_caches = 0;
1614
Christoph Lameter3c583462012-11-28 16:23:01 +00001615 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lametere498be72005-09-09 13:03:32 -07001616 kmem_list3_init(&initkmem_list3[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001617
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001618 set_up_list3s(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619
1620 /*
1621 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001622 * page orders on machines with more than 32MB of memory if
1623 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001625 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001626 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 /* Bootstrap is tricky, because several objects are allocated
1629 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001630 * 1) initialize the kmem_cache cache: it contains the struct
1631 * kmem_cache structures of all caches, except kmem_cache itself:
1632 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001633 * Initially an __init data area is used for the head array and the
1634 * kmem_list3 structures, it's replaced with a kmalloc allocated
1635 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001637 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001638 * An __init data area is used for the head array.
1639 * 3) Create the remaining kmalloc caches, with minimally sized
1640 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001641 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001643 * 5) Replace the __init data for kmem_list3 for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001644 * the other cache's with kmalloc allocated memory.
1645 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 */
1647
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001648 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Eric Dumazet8da34302007-05-06 14:49:29 -07001650 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001651 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001652 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001653 create_boot_cache(kmem_cache, "kmem_cache",
1654 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1655 nr_node_ids * sizeof(struct kmem_list3 *),
1656 SLAB_HWCACHE_ALIGN);
1657 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658
1659 /* 2+3) create the kmalloc caches */
1660 sizes = malloc_sizes;
1661 names = cache_names;
1662
Andrew Mortona737b3e2006-03-22 00:08:11 -08001663 /*
1664 * Initialize the caches that provide memory for the array cache and the
1665 * kmem_list3 structures first. Without this, further allocations will
1666 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001667 */
1668
Christoph Lameter45530c42012-11-28 16:23:07 +00001669 sizes[INDEX_AC].cs_cachep = create_kmalloc_cache(names[INDEX_AC].name,
1670 sizes[INDEX_AC].cs_size, ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001671
Christoph Lameter45530c42012-11-28 16:23:07 +00001672 if (INDEX_AC != INDEX_L3)
1673 sizes[INDEX_L3].cs_cachep =
1674 create_kmalloc_cache(names[INDEX_L3].name,
1675 sizes[INDEX_L3].cs_size, ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001676
Ingo Molnare0a42722006-06-23 02:03:46 -07001677 slab_early_init = 0;
1678
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001680 /*
1681 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 * This should be particularly beneficial on SMP boxes, as it
1683 * eliminates "false sharing".
1684 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001685 * allow tighter packing of the smaller caches.
1686 */
Christoph Lameter45530c42012-11-28 16:23:07 +00001687 if (!sizes->cs_cachep)
1688 sizes->cs_cachep = create_kmalloc_cache(names->name,
1689 sizes->cs_size, ARCH_KMALLOC_FLAGS);
1690
Christoph Lameter4b51d662007-02-10 01:43:10 -08001691#ifdef CONFIG_ZONE_DMA
Christoph Lameter45530c42012-11-28 16:23:07 +00001692 sizes->cs_dmacachep = create_kmalloc_cache(
1693 names->name_dma, sizes->cs_size,
1694 SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001695#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 sizes++;
1697 names++;
1698 }
1699 /* 4) Replace the bootstrap head arrays */
1700 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001701 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001702
Pekka Enberg83b519e2009-06-10 19:40:04 +03001703 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001704
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001705 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001706 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001707 /*
1708 * Do not assume that spinlocks can be initialized via memcpy:
1709 */
1710 spin_lock_init(&ptr->lock);
1711
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001712 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001713
Pekka Enberg83b519e2009-06-10 19:40:04 +03001714 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001715
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001716 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001717 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001718 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001719 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001720 /*
1721 * Do not assume that spinlocks can be initialized via memcpy:
1722 */
1723 spin_lock_init(&ptr->lock);
1724
Christoph Lametere498be72005-09-09 13:03:32 -07001725 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001726 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 }
Christoph Lametere498be72005-09-09 13:03:32 -07001728 /* 5) Replace the bootstrap kmem_list3's */
1729 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001730 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731
Mel Gorman9c09a952008-01-24 05:49:54 -08001732 for_each_online_node(nid) {
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001733 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001734
Christoph Lametere498be72005-09-09 13:03:32 -07001735 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001736 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001737
1738 if (INDEX_AC != INDEX_L3) {
1739 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001740 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001741 }
1742 }
1743 }
1744
Christoph Lameter97d06602012-07-06 15:25:11 -05001745 slab_state = UP;
Pekka Enberg8429db52009-06-12 15:58:59 +03001746}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001747
Pekka Enberg8429db52009-06-12 15:58:59 +03001748void __init kmem_cache_init_late(void)
1749{
1750 struct kmem_cache *cachep;
1751
Christoph Lameter97d06602012-07-06 15:25:11 -05001752 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001753
Pekka Enberg8429db52009-06-12 15:58:59 +03001754 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001755 mutex_lock(&slab_mutex);
1756 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001757 if (enable_cpucache(cachep, GFP_NOWAIT))
1758 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001759 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001760
Michael Wang947ca182012-09-05 10:33:18 +08001761 /* Annotate slab for lockdep -- annotate the malloc caches */
1762 init_lock_keys();
1763
Christoph Lameter97d06602012-07-06 15:25:11 -05001764 /* Done! */
1765 slab_state = FULL;
1766
Andrew Mortona737b3e2006-03-22 00:08:11 -08001767 /*
1768 * Register a cpu startup notifier callback that initializes
1769 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 */
1771 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
David Rientjes8f9f8d92010-03-27 19:40:47 -07001773#ifdef CONFIG_NUMA
1774 /*
1775 * Register a memory hotplug callback that initializes and frees
1776 * nodelists.
1777 */
1778 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1779#endif
1780
Andrew Mortona737b3e2006-03-22 00:08:11 -08001781 /*
1782 * The reap timers are started later, with a module init call: That part
1783 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 */
1785}
1786
1787static int __init cpucache_init(void)
1788{
1789 int cpu;
1790
Andrew Mortona737b3e2006-03-22 00:08:11 -08001791 /*
1792 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 */
Christoph Lametere498be72005-09-09 13:03:32 -07001794 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001795 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001796
1797 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001798 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 return 0;
1800}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801__initcall(cpucache_init);
1802
Rafael Aquini8bdec192012-03-09 17:27:27 -03001803static noinline void
1804slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1805{
1806 struct kmem_list3 *l3;
1807 struct slab *slabp;
1808 unsigned long flags;
1809 int node;
1810
1811 printk(KERN_WARNING
1812 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1813 nodeid, gfpflags);
1814 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001815 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001816
1817 for_each_online_node(node) {
1818 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1819 unsigned long active_slabs = 0, num_slabs = 0;
1820
1821 l3 = cachep->nodelists[node];
1822 if (!l3)
1823 continue;
1824
1825 spin_lock_irqsave(&l3->list_lock, flags);
1826 list_for_each_entry(slabp, &l3->slabs_full, list) {
1827 active_objs += cachep->num;
1828 active_slabs++;
1829 }
1830 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1831 active_objs += slabp->inuse;
1832 active_slabs++;
1833 }
1834 list_for_each_entry(slabp, &l3->slabs_free, list)
1835 num_slabs++;
1836
1837 free_objects += l3->free_objects;
1838 spin_unlock_irqrestore(&l3->list_lock, flags);
1839
1840 num_slabs += active_slabs;
1841 num_objs = num_slabs * cachep->num;
1842 printk(KERN_WARNING
1843 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1844 node, active_slabs, num_slabs, active_objs, num_objs,
1845 free_objects);
1846 }
1847}
1848
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849/*
1850 * Interface to system's page allocator. No need to hold the cache-lock.
1851 *
1852 * If we requested dmaable memory, we will get it. Even if we
1853 * did not request dmaable memory, we might get it, but that
1854 * would be relatively rare and ignorable.
1855 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001856static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857{
1858 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001859 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 int i;
1861
Luke Yangd6fef9d2006-04-10 22:52:56 -07001862#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001863 /*
1864 * Nommu uses slab's for process anonymous memory allocations, and thus
1865 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001866 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001867 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001868#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001869
Glauber Costaa618e892012-06-14 16:17:21 +04001870 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001871 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1872 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001873
Linus Torvalds517d0862009-06-16 19:50:13 -07001874 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001875 if (!page) {
1876 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1877 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001879 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001881 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001882 if (unlikely(page->pfmemalloc))
1883 pfmemalloc_active = true;
1884
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001885 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001887 add_zone_page_state(page_zone(page),
1888 NR_SLAB_RECLAIMABLE, nr_pages);
1889 else
1890 add_zone_page_state(page_zone(page),
1891 NR_SLAB_UNRECLAIMABLE, nr_pages);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001892 for (i = 0; i < nr_pages; i++) {
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001893 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001894
Mel Gorman072bb0a2012-07-31 16:43:58 -07001895 if (page->pfmemalloc)
1896 SetPageSlabPfmemalloc(page + i);
1897 }
Glauber Costa1f458cb2012-12-18 14:22:50 -08001898 memcg_bind_pages(cachep, cachep->gfporder);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001899
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001900 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1901 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1902
1903 if (cachep->ctor)
1904 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1905 else
1906 kmemcheck_mark_unallocated_pages(page, nr_pages);
1907 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001908
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001909 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910}
1911
1912/*
1913 * Interface to system's page release.
1914 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001915static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001917 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 struct page *page = virt_to_page(addr);
1919 const unsigned long nr_freed = i;
1920
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001921 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001922
Christoph Lameter972d1a72006-09-25 23:31:51 -07001923 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1924 sub_zone_page_state(page_zone(page),
1925 NR_SLAB_RECLAIMABLE, nr_freed);
1926 else
1927 sub_zone_page_state(page_zone(page),
1928 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001930 BUG_ON(!PageSlab(page));
Mel Gorman072bb0a2012-07-31 16:43:58 -07001931 __ClearPageSlabPfmemalloc(page);
Nick Pigginf205b2f2006-03-22 00:08:02 -08001932 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 page++;
1934 }
Glauber Costa1f458cb2012-12-18 14:22:50 -08001935
1936 memcg_release_pages(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 if (current->reclaim_state)
1938 current->reclaim_state->reclaimed_slab += nr_freed;
Glauber Costad79923f2012-12-18 14:22:48 -08001939 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940}
1941
1942static void kmem_rcu_free(struct rcu_head *head)
1943{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001944 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001945 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946
1947 kmem_freepages(cachep, slab_rcu->addr);
1948 if (OFF_SLAB(cachep))
1949 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1950}
1951
1952#if DEBUG
1953
1954#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001955static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001956 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001958 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001960 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001962 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 return;
1964
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 *addr++ = 0x12345678;
1966 *addr++ = caller;
1967 *addr++ = smp_processor_id();
1968 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 {
1970 unsigned long *sptr = &caller;
1971 unsigned long svalue;
1972
1973 while (!kstack_end(sptr)) {
1974 svalue = *sptr++;
1975 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001976 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 size -= sizeof(unsigned long);
1978 if (size <= sizeof(unsigned long))
1979 break;
1980 }
1981 }
1982
1983 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001984 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985}
1986#endif
1987
Pekka Enberg343e0d72006-02-01 03:05:50 -08001988static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001990 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001991 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992
1993 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001994 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995}
1996
1997static void dump_line(char *data, int offset, int limit)
1998{
1999 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07002000 unsigned char error = 0;
2001 int bad_count = 0;
2002
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002003 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002004 for (i = 0; i < limit; i++) {
2005 if (data[offset + i] != POISON_FREE) {
2006 error = data[offset + i];
2007 bad_count++;
2008 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07002009 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002010 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2011 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002012
2013 if (bad_count == 1) {
2014 error ^= POISON_FREE;
2015 if (!(error & (error - 1))) {
2016 printk(KERN_ERR "Single bit error detected. Probably "
2017 "bad RAM.\n");
2018#ifdef CONFIG_X86
2019 printk(KERN_ERR "Run memtest86+ or a similar memory "
2020 "test tool.\n");
2021#else
2022 printk(KERN_ERR "Run a memory test tool.\n");
2023#endif
2024 }
2025 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026}
2027#endif
2028
2029#if DEBUG
2030
Pekka Enberg343e0d72006-02-01 03:05:50 -08002031static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032{
2033 int i, size;
2034 char *realobj;
2035
2036 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07002037 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002038 *dbg_redzone1(cachep, objp),
2039 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 }
2041
2042 if (cachep->flags & SLAB_STORE_USER) {
2043 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002044 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002046 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 printk("\n");
2048 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002049 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002050 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002051 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 int limit;
2053 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002054 if (i + limit > size)
2055 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 dump_line(realobj, i, limit);
2057 }
2058}
2059
Pekka Enberg343e0d72006-02-01 03:05:50 -08002060static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061{
2062 char *realobj;
2063 int size, i;
2064 int lines = 0;
2065
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002066 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002067 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002069 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002071 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 exp = POISON_END;
2073 if (realobj[i] != exp) {
2074 int limit;
2075 /* Mismatch ! */
2076 /* Print header */
2077 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002078 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08002079 "Slab corruption (%s): %s start=%p, len=%d\n",
2080 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 print_objinfo(cachep, objp, 0);
2082 }
2083 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002084 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002086 if (i + limit > size)
2087 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 dump_line(realobj, i, limit);
2089 i += 16;
2090 lines++;
2091 /* Limit to 5 lines */
2092 if (lines > 5)
2093 break;
2094 }
2095 }
2096 if (lines != 0) {
2097 /* Print some data about the neighboring objects, if they
2098 * exist:
2099 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002100 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002101 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002103 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002105 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002106 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002108 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 print_objinfo(cachep, objp, 2);
2110 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002111 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002112 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002113 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002115 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 print_objinfo(cachep, objp, 2);
2117 }
2118 }
2119}
2120#endif
2121
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05302123static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002124{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 int i;
2126 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002127 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128
2129 if (cachep->flags & SLAB_POISON) {
2130#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002131 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002132 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002133 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002134 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 else
2136 check_poison_obj(cachep, objp);
2137#else
2138 check_poison_obj(cachep, objp);
2139#endif
2140 }
2141 if (cachep->flags & SLAB_RED_ZONE) {
2142 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2143 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002144 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2146 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002147 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002150}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302152static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002153{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002154}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155#endif
2156
Randy Dunlap911851e2006-03-22 00:08:14 -08002157/**
2158 * slab_destroy - destroy and release all objects in a slab
2159 * @cachep: cache pointer being destroyed
2160 * @slabp: slab pointer being destroyed
2161 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002162 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002163 * Before calling the slab must have been unlinked from the cache. The
2164 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002165 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002166static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002167{
2168 void *addr = slabp->s_mem - slabp->colouroff;
2169
Rabin Vincente79aec22008-07-04 00:40:32 +05302170 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2172 struct slab_rcu *slab_rcu;
2173
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002174 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 slab_rcu->cachep = cachep;
2176 slab_rcu->addr = addr;
2177 call_rcu(&slab_rcu->head, kmem_rcu_free);
2178 } else {
2179 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002180 if (OFF_SLAB(cachep))
2181 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 }
2183}
2184
2185/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002186 * calculate_slab_order - calculate size (page order) of slabs
2187 * @cachep: pointer to the cache that is being created
2188 * @size: size of objects to be created in this cache.
2189 * @align: required alignment for the objects.
2190 * @flags: slab allocation flags
2191 *
2192 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002193 *
2194 * This could be made much more intelligent. For now, try to avoid using
2195 * high order pages for slabs. When the gfp() functions are more friendly
2196 * towards high-order requests, this should be changed.
2197 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002198static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002199 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002200{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002201 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002202 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002203 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002204
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002205 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002206 unsigned int num;
2207 size_t remainder;
2208
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002209 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002210 if (!num)
2211 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002212
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002213 if (flags & CFLGS_OFF_SLAB) {
2214 /*
2215 * Max number of objs-per-slab for caches which
2216 * use off-slab slabs. Needed to avoid a possible
2217 * looping condition in cache_grow().
2218 */
2219 offslab_limit = size - sizeof(struct slab);
2220 offslab_limit /= sizeof(kmem_bufctl_t);
2221
2222 if (num > offslab_limit)
2223 break;
2224 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002225
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002226 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002227 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002228 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002229 left_over = remainder;
2230
2231 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002232 * A VFS-reclaimable slab tends to have most allocations
2233 * as GFP_NOFS and we really don't want to have to be allocating
2234 * higher-order pages when we are unable to shrink dcache.
2235 */
2236 if (flags & SLAB_RECLAIM_ACCOUNT)
2237 break;
2238
2239 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002240 * Large number of objects is good, but very large slabs are
2241 * currently bad for the gfp()s.
2242 */
David Rientjes543585c2011-10-18 22:09:24 -07002243 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002244 break;
2245
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002246 /*
2247 * Acceptable internal fragmentation?
2248 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002249 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002250 break;
2251 }
2252 return left_over;
2253}
2254
Pekka Enberg83b519e2009-06-10 19:40:04 +03002255static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002256{
Christoph Lameter97d06602012-07-06 15:25:11 -05002257 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002258 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002259
Christoph Lameter97d06602012-07-06 15:25:11 -05002260 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002261 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002262 * Note: Creation of first cache (kmem_cache).
2263 * The setup_list3s is taken care
2264 * of by the caller of __kmem_cache_create
2265 */
2266 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2267 slab_state = PARTIAL;
2268 } else if (slab_state == PARTIAL) {
2269 /*
2270 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002271 * that's used by kmalloc(24), otherwise the creation of
2272 * further caches will BUG().
2273 */
2274 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2275
2276 /*
2277 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002278 * the second cache, then we need to set up all its list3s,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002279 * otherwise the creation of further caches will BUG().
2280 */
2281 set_up_list3s(cachep, SIZE_AC);
2282 if (INDEX_AC == INDEX_L3)
Christoph Lameter97d06602012-07-06 15:25:11 -05002283 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002284 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002285 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002286 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002287 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002288 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002289 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002290
Christoph Lameter97d06602012-07-06 15:25:11 -05002291 if (slab_state == PARTIAL_ARRAYCACHE) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002292 set_up_list3s(cachep, SIZE_L3);
Christoph Lameter97d06602012-07-06 15:25:11 -05002293 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002294 } else {
2295 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002296 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002297 cachep->nodelists[node] =
2298 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002299 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002300 BUG_ON(!cachep->nodelists[node]);
2301 kmem_list3_init(cachep->nodelists[node]);
2302 }
2303 }
2304 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002305 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002306 jiffies + REAPTIMEOUT_LIST3 +
2307 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2308
2309 cpu_cache_get(cachep)->avail = 0;
2310 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2311 cpu_cache_get(cachep)->batchcount = 1;
2312 cpu_cache_get(cachep)->touched = 0;
2313 cachep->batchcount = 1;
2314 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002315 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002316}
2317
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002318/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002319 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002320 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 *
2323 * Returns a ptr to the cache on success, NULL on failure.
2324 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002325 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 * The flags are
2328 *
2329 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2330 * to catch references to uninitialised memory.
2331 *
2332 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2333 * for buffer overruns.
2334 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2336 * cacheline. This can be beneficial if you're counting cycles as closely
2337 * as davem.
2338 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002339int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002340__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341{
2342 size_t left_over, slab_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002343 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002344 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002345 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348#if FORCED_DEBUG
2349 /*
2350 * Enable redzoning and last user accounting, except for caches with
2351 * large objects, if the increased size would increase the object size
2352 * above the next power of two: caches with object sizes just above a
2353 * power of two have a significant amount of internal fragmentation.
2354 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002355 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2356 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002357 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 if (!(flags & SLAB_DESTROY_BY_RCU))
2359 flags |= SLAB_POISON;
2360#endif
2361 if (flags & SLAB_DESTROY_BY_RCU)
2362 BUG_ON(flags & SLAB_POISON);
2363#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
Andrew Mortona737b3e2006-03-22 00:08:11 -08002365 /*
2366 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 * unaligned accesses for some archs when redzoning is used, and makes
2368 * sure any on-slab bufctl's are also correctly aligned.
2369 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002370 if (size & (BYTES_PER_WORD - 1)) {
2371 size += (BYTES_PER_WORD - 1);
2372 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 }
2374
Pekka Enbergca5f9702006-09-25 23:31:25 -07002375 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002376 * Redzoning and user store require word alignment or possibly larger.
2377 * Note this will be overridden by architecture or caller mandated
2378 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002379 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002380 if (flags & SLAB_STORE_USER)
2381 ralign = BYTES_PER_WORD;
2382
2383 if (flags & SLAB_RED_ZONE) {
2384 ralign = REDZONE_ALIGN;
2385 /* If redzoning, ensure that the second redzone is suitably
2386 * aligned, by adjusting the object size accordingly. */
2387 size += REDZONE_ALIGN - 1;
2388 size &= ~(REDZONE_ALIGN - 1);
2389 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002390
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002391 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002392 if (ralign < cachep->align) {
2393 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002395 /* disable debug if necessary */
2396 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002397 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002398 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002399 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002401 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402
Pekka Enberg83b519e2009-06-10 19:40:04 +03002403 if (slab_is_available())
2404 gfp = GFP_KERNEL;
2405 else
2406 gfp = GFP_NOWAIT;
2407
Christoph Lameter3c583462012-11-28 16:23:01 +00002408 setup_nodelists_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410
Pekka Enbergca5f9702006-09-25 23:31:25 -07002411 /*
2412 * Both debugging options require word-alignment which is calculated
2413 * into align above.
2414 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002417 cachep->obj_offset += sizeof(unsigned long long);
2418 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 }
2420 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002421 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002422 * the real object. But if the second red zone needs to be
2423 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002425 if (flags & SLAB_RED_ZONE)
2426 size += REDZONE_ALIGN;
2427 else
2428 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 }
2430#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002431 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002432 && cachep->object_size > cache_line_size()
2433 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2434 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 size = PAGE_SIZE;
2436 }
2437#endif
2438#endif
2439
Ingo Molnare0a42722006-06-23 02:03:46 -07002440 /*
2441 * Determine if the slab management is 'on' or 'off' slab.
2442 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002443 * it too early on. Always use on-slab management when
2444 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002445 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002446 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2447 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 /*
2449 * Size is large, assume best to place the slab management obj
2450 * off-slab (should allow better packing of objs).
2451 */
2452 flags |= CFLGS_OFF_SLAB;
2453
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002454 size = ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002456 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002458 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002459 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002460
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002461 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002462 + sizeof(struct slab), cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
2464 /*
2465 * If the slab has been placed off-slab, and we have enough space then
2466 * move it on-slab. This is at the expense of any extra colouring.
2467 */
2468 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2469 flags &= ~CFLGS_OFF_SLAB;
2470 left_over -= slab_size;
2471 }
2472
2473 if (flags & CFLGS_OFF_SLAB) {
2474 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002475 slab_size =
2476 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302477
2478#ifdef CONFIG_PAGE_POISONING
2479 /* If we're going to use the generic kernel_map_pages()
2480 * poisoning, then it's going to smash the contents of
2481 * the redzone and userword anyhow, so switch them off.
2482 */
2483 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2484 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2485#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 }
2487
2488 cachep->colour_off = cache_line_size();
2489 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002490 if (cachep->colour_off < cachep->align)
2491 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002492 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 cachep->slab_size = slab_size;
2494 cachep->flags = flags;
Glauber Costaa618e892012-06-14 16:17:21 +04002495 cachep->allocflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002496 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002497 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002498 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002499 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002501 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002502 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002503 /*
2504 * This is a possibility for one of the malloc_sizes caches.
2505 * But since we go off slab only for object size greater than
2506 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2507 * this should not happen at all.
2508 * But leave a BUG_ON for some lucky dude.
2509 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002510 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002511 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002513 err = setup_cpu_cache(cachep, gfp);
2514 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002515 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002516 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
Peter Zijlstra83835b32011-07-22 15:26:05 +02002519 if (flags & SLAB_DEBUG_OBJECTS) {
2520 /*
2521 * Would deadlock through slab_destroy()->call_rcu()->
2522 * debug_object_activate()->kmem_cache_alloc().
2523 */
2524 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2525
2526 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002527 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2528 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002529
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002530 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532
2533#if DEBUG
2534static void check_irq_off(void)
2535{
2536 BUG_ON(!irqs_disabled());
2537}
2538
2539static void check_irq_on(void)
2540{
2541 BUG_ON(irqs_disabled());
2542}
2543
Pekka Enberg343e0d72006-02-01 03:05:50 -08002544static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545{
2546#ifdef CONFIG_SMP
2547 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002548 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549#endif
2550}
Christoph Lametere498be72005-09-09 13:03:32 -07002551
Pekka Enberg343e0d72006-02-01 03:05:50 -08002552static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002553{
2554#ifdef CONFIG_SMP
2555 check_irq_off();
2556 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2557#endif
2558}
2559
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560#else
2561#define check_irq_off() do { } while(0)
2562#define check_irq_on() do { } while(0)
2563#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002564#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565#endif
2566
Christoph Lameteraab22072006-03-22 00:09:06 -08002567static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2568 struct array_cache *ac,
2569 int force, int node);
2570
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571static void do_drain(void *arg)
2572{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002573 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002575 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576
2577 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002578 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002579 spin_lock(&cachep->nodelists[node]->list_lock);
2580 free_block(cachep, ac->entry, ac->avail, node);
2581 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 ac->avail = 0;
2583}
2584
Pekka Enberg343e0d72006-02-01 03:05:50 -08002585static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586{
Christoph Lametere498be72005-09-09 13:03:32 -07002587 struct kmem_list3 *l3;
2588 int node;
2589
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002590 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002592 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002593 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002594 if (l3 && l3->alien)
2595 drain_alien_cache(cachep, l3->alien);
2596 }
2597
2598 for_each_online_node(node) {
2599 l3 = cachep->nodelists[node];
2600 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002601 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002602 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603}
2604
Christoph Lametered11d9e2006-06-30 01:55:45 -07002605/*
2606 * Remove slabs from the list of free slabs.
2607 * Specify the number of slabs to drain in tofree.
2608 *
2609 * Returns the actual number of slabs released.
2610 */
2611static int drain_freelist(struct kmem_cache *cache,
2612 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002614 struct list_head *p;
2615 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617
Christoph Lametered11d9e2006-06-30 01:55:45 -07002618 nr_freed = 0;
2619 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620
Christoph Lametered11d9e2006-06-30 01:55:45 -07002621 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002622 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002623 if (p == &l3->slabs_free) {
2624 spin_unlock_irq(&l3->list_lock);
2625 goto out;
2626 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627
Christoph Lametered11d9e2006-06-30 01:55:45 -07002628 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002630 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631#endif
2632 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002633 /*
2634 * Safe to drop the lock. The slab is no longer linked
2635 * to the cache.
2636 */
2637 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002638 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002639 slab_destroy(cache, slabp);
2640 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002642out:
2643 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644}
2645
Christoph Lameter18004c52012-07-06 15:25:12 -05002646/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002647static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002648{
2649 int ret = 0, i = 0;
2650 struct kmem_list3 *l3;
2651
2652 drain_cpu_caches(cachep);
2653
2654 check_irq_on();
2655 for_each_online_node(i) {
2656 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002657 if (!l3)
2658 continue;
2659
2660 drain_freelist(cachep, l3, l3->free_objects);
2661
2662 ret += !list_empty(&l3->slabs_full) ||
2663 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002664 }
2665 return (ret ? 1 : 0);
2666}
2667
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668/**
2669 * kmem_cache_shrink - Shrink a cache.
2670 * @cachep: The cache to shrink.
2671 *
2672 * Releases as many slabs as possible for a cache.
2673 * To help debugging, a zero exit status indicates all slabs were released.
2674 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002675int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002677 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002678 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002680 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002681 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002682 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002683 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002684 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002685 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686}
2687EXPORT_SYMBOL(kmem_cache_shrink);
2688
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002689int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690{
Christoph Lameter12c36672012-09-04 23:38:33 +00002691 int i;
2692 struct kmem_list3 *l3;
2693 int rc = __cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694
Christoph Lameter12c36672012-09-04 23:38:33 +00002695 if (rc)
2696 return rc;
2697
2698 for_each_online_cpu(i)
2699 kfree(cachep->array[i]);
2700
2701 /* NUMA: free the list3 structures */
2702 for_each_online_node(i) {
2703 l3 = cachep->nodelists[i];
2704 if (l3) {
2705 kfree(l3->shared);
2706 free_alien_cache(l3->alien);
2707 kfree(l3);
2708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002710 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002713/*
2714 * Get the memory for a slab management obj.
2715 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2716 * always come from malloc_sizes caches. The slab descriptor cannot
2717 * come from the same cache which is getting created because,
2718 * when we are searching for an appropriate cache for these
2719 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2720 * If we are creating a malloc_sizes cache here it would not be visible to
2721 * kmem_find_general_cachep till the initialization is complete.
2722 * Hence we cannot have slabp_cache same as the original cache.
2723 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002724static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002725 int colour_off, gfp_t local_flags,
2726 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727{
2728 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002729
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 if (OFF_SLAB(cachep)) {
2731 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002732 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002733 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002734 /*
2735 * If the first object in the slab is leaked (it's allocated
2736 * but no one has a reference to it), we want to make sure
2737 * kmemleak does not treat the ->s_mem pointer as a reference
2738 * to the object. Otherwise we will not report the leak.
2739 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002740 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2741 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 if (!slabp)
2743 return NULL;
2744 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002745 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 colour_off += cachep->slab_size;
2747 }
2748 slabp->inuse = 0;
2749 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002750 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002751 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002752 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 return slabp;
2754}
2755
2756static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2757{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002758 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759}
2760
Pekka Enberg343e0d72006-02-01 03:05:50 -08002761static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002762 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
2764 int i;
2765
2766 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002767 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768#if DEBUG
2769 /* need to poison the objs? */
2770 if (cachep->flags & SLAB_POISON)
2771 poison_obj(cachep, objp, POISON_FREE);
2772 if (cachep->flags & SLAB_STORE_USER)
2773 *dbg_userword(cachep, objp) = NULL;
2774
2775 if (cachep->flags & SLAB_RED_ZONE) {
2776 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2777 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2778 }
2779 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780 * Constructors are not allowed to allocate memory from the same
2781 * cache which they are a constructor for. Otherwise, deadlock.
2782 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 */
2784 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002785 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786
2787 if (cachep->flags & SLAB_RED_ZONE) {
2788 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2789 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002790 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2792 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002793 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002795 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002796 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002797 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002798 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799#else
2800 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002801 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002803 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002805 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806}
2807
Pekka Enberg343e0d72006-02-01 03:05:50 -08002808static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002810 if (CONFIG_ZONE_DMA_FLAG) {
2811 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002812 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002813 else
Glauber Costaa618e892012-06-14 16:17:21 +04002814 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816}
2817
Andrew Mortona737b3e2006-03-22 00:08:11 -08002818static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2819 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002820{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002821 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002822 kmem_bufctl_t next;
2823
2824 slabp->inuse++;
2825 next = slab_bufctl(slabp)[slabp->free];
2826#if DEBUG
2827 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2828 WARN_ON(slabp->nodeid != nodeid);
2829#endif
2830 slabp->free = next;
2831
2832 return objp;
2833}
2834
Andrew Mortona737b3e2006-03-22 00:08:11 -08002835static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2836 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002837{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002838 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002839
2840#if DEBUG
2841 /* Verify that the slab belongs to the intended node */
2842 WARN_ON(slabp->nodeid != nodeid);
2843
Al Viro871751e2006-03-25 03:06:39 -08002844 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002845 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002846 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002847 BUG();
2848 }
2849#endif
2850 slab_bufctl(slabp)[objnr] = slabp->free;
2851 slabp->free = objnr;
2852 slabp->inuse--;
2853}
2854
Pekka Enberg47768742006-06-23 02:03:07 -07002855/*
2856 * Map pages beginning at addr to the given cache and slab. This is required
2857 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002858 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002859 */
2860static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2861 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862{
Pekka Enberg47768742006-06-23 02:03:07 -07002863 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 struct page *page;
2865
Pekka Enberg47768742006-06-23 02:03:07 -07002866 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002867
Pekka Enberg47768742006-06-23 02:03:07 -07002868 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002869 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002870 nr_pages <<= cache->gfporder;
2871
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 do {
Christoph Lameter35026082012-06-13 10:24:56 -05002873 page->slab_cache = cache;
2874 page->slab_page = slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002876 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877}
2878
2879/*
2880 * Grow (by 1) the number of slabs within a cache. This is called by
2881 * kmem_cache_alloc() when there are no active objs left in a cache.
2882 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002883static int cache_grow(struct kmem_cache *cachep,
2884 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002886 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002887 size_t offset;
2888 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002889 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890
Andrew Mortona737b3e2006-03-22 00:08:11 -08002891 /*
2892 * Be lazy and only check for valid flags here, keeping it out of the
2893 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002895 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2896 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002898 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002900 l3 = cachep->nodelists[nodeid];
2901 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902
2903 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002904 offset = l3->colour_next;
2905 l3->colour_next++;
2906 if (l3->colour_next >= cachep->colour)
2907 l3->colour_next = 0;
2908 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002910 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911
2912 if (local_flags & __GFP_WAIT)
2913 local_irq_enable();
2914
2915 /*
2916 * The test for missing atomic flag is performed here, rather than
2917 * the more obvious place, simply to reduce the critical path length
2918 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2919 * will eventually be caught here (where it matters).
2920 */
2921 kmem_flagcheck(cachep, flags);
2922
Andrew Mortona737b3e2006-03-22 00:08:11 -08002923 /*
2924 * Get mem for the objs. Attempt to allocate a physical page from
2925 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002926 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002927 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002928 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002929 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 goto failed;
2931
2932 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002933 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002934 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002935 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 goto opps1;
2937
Pekka Enberg47768742006-06-23 02:03:07 -07002938 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939
Christoph Lametera35afb82007-05-16 22:10:57 -07002940 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
2942 if (local_flags & __GFP_WAIT)
2943 local_irq_disable();
2944 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002945 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946
2947 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002948 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002950 l3->free_objects += cachep->num;
2951 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002953opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002955failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 if (local_flags & __GFP_WAIT)
2957 local_irq_disable();
2958 return 0;
2959}
2960
2961#if DEBUG
2962
2963/*
2964 * Perform extra freeing checks:
2965 * - detect bad pointers.
2966 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 */
2968static void kfree_debugcheck(const void *objp)
2969{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 if (!virt_addr_valid(objp)) {
2971 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002972 (unsigned long)objp);
2973 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975}
2976
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002977static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2978{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002979 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002980
2981 redzone1 = *dbg_redzone1(cache, obj);
2982 redzone2 = *dbg_redzone2(cache, obj);
2983
2984 /*
2985 * Redzone is ok.
2986 */
2987 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2988 return;
2989
2990 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2991 slab_error(cache, "double free detected");
2992 else
2993 slab_error(cache, "memory outside object was overwritten");
2994
David Woodhouseb46b8f12007-05-08 00:22:59 -07002995 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002996 obj, redzone1, redzone2);
2997}
2998
Pekka Enberg343e0d72006-02-01 03:05:50 -08002999static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003000 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001{
3002 struct page *page;
3003 unsigned int objnr;
3004 struct slab *slabp;
3005
Matthew Wilcox80cbd912007-11-29 12:05:13 -07003006 BUG_ON(virt_to_cache(objp) != cachep);
3007
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003008 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07003010 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011
Christoph Lameter35026082012-06-13 10:24:56 -05003012 slabp = page->slab_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013
3014 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003015 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3017 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3018 }
3019 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003020 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003022 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023
3024 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003025 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Al Viro871751e2006-03-25 03:06:39 -08003027#ifdef CONFIG_DEBUG_SLAB_LEAK
3028 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3029#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 if (cachep->flags & SLAB_POISON) {
3031#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003032 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003033 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003034 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003035 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 } else {
3037 poison_obj(cachep, objp, POISON_FREE);
3038 }
3039#else
3040 poison_obj(cachep, objp, POISON_FREE);
3041#endif
3042 }
3043 return objp;
3044}
3045
Pekka Enberg343e0d72006-02-01 03:05:50 -08003046static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047{
3048 kmem_bufctl_t i;
3049 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003050
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 /* Check slab's freelist to see if this obj is there. */
3052 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3053 entries++;
3054 if (entries > cachep->num || i >= cachep->num)
3055 goto bad;
3056 }
3057 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003058bad:
3059 printk(KERN_ERR "slab: Internal list corruption detected in "
Dave Jonesface37f2011-11-15 15:03:52 -08003060 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3061 cachep->name, cachep->num, slabp, slabp->inuse,
3062 print_tainted());
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003063 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3064 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3065 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 BUG();
3067 }
3068}
3069#else
3070#define kfree_debugcheck(x) do { } while(0)
3071#define cache_free_debugcheck(x,objp,z) (objp)
3072#define check_slabp(x,y) do { } while(0)
3073#endif
3074
Mel Gorman072bb0a2012-07-31 16:43:58 -07003075static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3076 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077{
3078 int batchcount;
3079 struct kmem_list3 *l3;
3080 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003081 int node;
3082
Joe Korty6d2144d2008-03-05 15:04:59 -08003083 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003084 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07003085 if (unlikely(force_refill))
3086 goto force_grow;
3087retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003088 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 batchcount = ac->batchcount;
3090 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003091 /*
3092 * If there was little recent activity on this cache, then
3093 * perform only a partial refill. Otherwise we could generate
3094 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095 */
3096 batchcount = BATCHREFILL_LIMIT;
3097 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003098 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099
Christoph Lametere498be72005-09-09 13:03:32 -07003100 BUG_ON(ac->avail > 0 || !l3);
3101 spin_lock(&l3->list_lock);
3102
Christoph Lameter3ded1752006-03-25 03:06:44 -08003103 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003104 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3105 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003106 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003107 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003108
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 while (batchcount > 0) {
3110 struct list_head *entry;
3111 struct slab *slabp;
3112 /* Get slab alloc is to come from. */
3113 entry = l3->slabs_partial.next;
3114 if (entry == &l3->slabs_partial) {
3115 l3->free_touched = 1;
3116 entry = l3->slabs_free.next;
3117 if (entry == &l3->slabs_free)
3118 goto must_grow;
3119 }
3120
3121 slabp = list_entry(entry, struct slab, list);
3122 check_slabp(cachep, slabp);
3123 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003124
3125 /*
3126 * The slab was either on partial or free list so
3127 * there must be at least one object available for
3128 * allocation.
3129 */
roel kluin249b9f32008-10-29 17:18:07 -04003130 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003131
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 STATS_INC_ALLOCED(cachep);
3134 STATS_INC_ACTIVE(cachep);
3135 STATS_SET_HIGH(cachep);
3136
Mel Gorman072bb0a2012-07-31 16:43:58 -07003137 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3138 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 }
3140 check_slabp(cachep, slabp);
3141
3142 /* move slabp to correct slabp list: */
3143 list_del(&slabp->list);
3144 if (slabp->free == BUFCTL_END)
3145 list_add(&slabp->list, &l3->slabs_full);
3146 else
3147 list_add(&slabp->list, &l3->slabs_partial);
3148 }
3149
Andrew Mortona737b3e2006-03-22 00:08:11 -08003150must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003152alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003153 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154
3155 if (unlikely(!ac->avail)) {
3156 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003157force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08003158 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003159
Andrew Mortona737b3e2006-03-22 00:08:11 -08003160 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003161 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07003162 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07003163
3164 /* no objects in sight? abort */
3165 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 return NULL;
3167
Andrew Mortona737b3e2006-03-22 00:08:11 -08003168 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 goto retry;
3170 }
3171 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003172
3173 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174}
3175
Andrew Mortona737b3e2006-03-22 00:08:11 -08003176static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3177 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178{
3179 might_sleep_if(flags & __GFP_WAIT);
3180#if DEBUG
3181 kmem_flagcheck(cachep, flags);
3182#endif
3183}
3184
3185#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003186static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003187 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003189 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003191 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003193 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003194 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003195 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 else
3197 check_poison_obj(cachep, objp);
3198#else
3199 check_poison_obj(cachep, objp);
3200#endif
3201 poison_obj(cachep, objp, POISON_INUSE);
3202 }
3203 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003204 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205
3206 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003207 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3208 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3209 slab_error(cachep, "double free, or memory outside"
3210 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003211 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003212 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003213 objp, *dbg_redzone1(cachep, objp),
3214 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 }
3216 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3217 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3218 }
Al Viro871751e2006-03-25 03:06:39 -08003219#ifdef CONFIG_DEBUG_SLAB_LEAK
3220 {
3221 struct slab *slabp;
3222 unsigned objnr;
3223
Christoph Lameter35026082012-06-13 10:24:56 -05003224 slabp = virt_to_head_page(objp)->slab_page;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003225 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
Al Viro871751e2006-03-25 03:06:39 -08003226 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3227 }
3228#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003229 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003230 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003231 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003232 if (ARCH_SLAB_MINALIGN &&
3233 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003234 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003235 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003236 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 return objp;
3238}
3239#else
3240#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3241#endif
3242
Akinobu Mita773ff602008-12-23 19:37:01 +09003243static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003244{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003245 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003246 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003247
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003248 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003249}
3250
Pekka Enberg343e0d72006-02-01 03:05:50 -08003251static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003253 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003255 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256
Alok N Kataria5c382302005-09-27 21:45:46 -07003257 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003258
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003259 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003262 objp = ac_get_obj(cachep, ac, flags, false);
3263
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003264 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003265 * Allow for the possibility all avail objects are not allowed
3266 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003267 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003268 if (objp) {
3269 STATS_INC_ALLOCHIT(cachep);
3270 goto out;
3271 }
3272 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003274
3275 STATS_INC_ALLOCMISS(cachep);
3276 objp = cache_alloc_refill(cachep, flags, force_refill);
3277 /*
3278 * the 'ac' may be updated by cache_alloc_refill(),
3279 * and kmemleak_erase() requires its correct value.
3280 */
3281 ac = cpu_cache_get(cachep);
3282
3283out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003284 /*
3285 * To avoid a false negative, if an object that is in one of the
3286 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3287 * treat the array pointers as a reference to the object.
3288 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003289 if (objp)
3290 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003291 return objp;
3292}
3293
Christoph Lametere498be72005-09-09 13:03:32 -07003294#ifdef CONFIG_NUMA
3295/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003296 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003297 *
3298 * If we are in_interrupt, then process context, including cpusets and
3299 * mempolicy, may not apply and should not be used for allocation policy.
3300 */
3301static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3302{
3303 int nid_alloc, nid_here;
3304
Christoph Lameter765c4502006-09-27 01:50:08 -07003305 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003306 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003307 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003308 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003309 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003310 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003311 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003312 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003313 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003314 return NULL;
3315}
3316
3317/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003318 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003319 * certain node and fall back is permitted. First we scan all the
3320 * available nodelists for available objects. If that fails then we
3321 * perform an allocation without specifying a node. This allows the page
3322 * allocator to do its reclaim / fallback magic. We then insert the
3323 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003324 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003325static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003326{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003327 struct zonelist *zonelist;
3328 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003329 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003330 struct zone *zone;
3331 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003332 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003333 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003334 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003335
3336 if (flags & __GFP_THISNODE)
3337 return NULL;
3338
Christoph Lameter6cb06222007-10-16 01:25:41 -07003339 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003340
Mel Gormancc9a6c82012-03-21 16:34:11 -07003341retry_cpuset:
3342 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003343 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003344
Christoph Lameter3c517a62006-12-06 20:33:29 -08003345retry:
3346 /*
3347 * Look through allowed nodes for objects available
3348 * from existing per node queues.
3349 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003350 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3351 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003352
Mel Gorman54a6eb52008-04-28 02:12:16 -07003353 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003354 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003355 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003356 obj = ____cache_alloc_node(cache,
3357 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003358 if (obj)
3359 break;
3360 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003361 }
3362
Christoph Lametercfce6602007-05-06 14:50:17 -07003363 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003364 /*
3365 * This allocation will be performed within the constraints
3366 * of the current cpuset / memory policy requirements.
3367 * We may trigger various forms of reclaim on the allowed
3368 * set and go into memory reserves if necessary.
3369 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003370 if (local_flags & __GFP_WAIT)
3371 local_irq_enable();
3372 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003373 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003374 if (local_flags & __GFP_WAIT)
3375 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003376 if (obj) {
3377 /*
3378 * Insert into the appropriate per node queues
3379 */
3380 nid = page_to_nid(virt_to_page(obj));
3381 if (cache_grow(cache, flags, nid, obj)) {
3382 obj = ____cache_alloc_node(cache,
3383 flags | GFP_THISNODE, nid);
3384 if (!obj)
3385 /*
3386 * Another processor may allocate the
3387 * objects in the slab since we are
3388 * not holding any locks.
3389 */
3390 goto retry;
3391 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003392 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003393 obj = NULL;
3394 }
3395 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003396 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003397
3398 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3399 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003400 return obj;
3401}
3402
3403/*
Christoph Lametere498be72005-09-09 13:03:32 -07003404 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003406static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003407 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003408{
3409 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003410 struct slab *slabp;
3411 struct kmem_list3 *l3;
3412 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003413 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003415 l3 = cachep->nodelists[nodeid];
3416 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003417
Andrew Mortona737b3e2006-03-22 00:08:11 -08003418retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003419 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003420 spin_lock(&l3->list_lock);
3421 entry = l3->slabs_partial.next;
3422 if (entry == &l3->slabs_partial) {
3423 l3->free_touched = 1;
3424 entry = l3->slabs_free.next;
3425 if (entry == &l3->slabs_free)
3426 goto must_grow;
3427 }
Christoph Lametere498be72005-09-09 13:03:32 -07003428
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003429 slabp = list_entry(entry, struct slab, list);
3430 check_spinlock_acquired_node(cachep, nodeid);
3431 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003432
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003433 STATS_INC_NODEALLOCS(cachep);
3434 STATS_INC_ACTIVE(cachep);
3435 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003436
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003437 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003438
Matthew Dobson78d382d2006-02-01 03:05:47 -08003439 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003440 check_slabp(cachep, slabp);
3441 l3->free_objects--;
3442 /* move slabp to correct slabp list: */
3443 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003444
Andrew Mortona737b3e2006-03-22 00:08:11 -08003445 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003446 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003447 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003448 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003449
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003450 spin_unlock(&l3->list_lock);
3451 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003452
Andrew Mortona737b3e2006-03-22 00:08:11 -08003453must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003454 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003455 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003456 if (x)
3457 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003458
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003459 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003460
Andrew Mortona737b3e2006-03-22 00:08:11 -08003461done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003462 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003463}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003464
3465/**
3466 * kmem_cache_alloc_node - Allocate an object on the specified node
3467 * @cachep: The cache to allocate from.
3468 * @flags: See kmalloc().
3469 * @nodeid: node number of the target node.
3470 * @caller: return address of caller, used for debug information
3471 *
3472 * Identical to kmem_cache_alloc but it will allocate memory on the given
3473 * node, which can improve the performance for cpu bound structures.
3474 *
3475 * Fallback to other node is possible if __GFP_THISNODE is not set.
3476 */
3477static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003478slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003479 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003480{
3481 unsigned long save_flags;
3482 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003483 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003484
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003485 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003486
Nick Piggincf40bd12009-01-21 08:12:39 +01003487 lockdep_trace_alloc(flags);
3488
Akinobu Mita773ff602008-12-23 19:37:01 +09003489 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003490 return NULL;
3491
Glauber Costad79923f2012-12-18 14:22:48 -08003492 cachep = memcg_kmem_get_cache(cachep, flags);
3493
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003494 cache_alloc_debugcheck_before(cachep, flags);
3495 local_irq_save(save_flags);
3496
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003497 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003498 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003499
3500 if (unlikely(!cachep->nodelists[nodeid])) {
3501 /* Node not bootstrapped yet */
3502 ptr = fallback_alloc(cachep, flags);
3503 goto out;
3504 }
3505
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003506 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003507 /*
3508 * Use the locally cached objects if possible.
3509 * However ____cache_alloc does not allow fallback
3510 * to other nodes. It may fail while we still have
3511 * objects on other nodes available.
3512 */
3513 ptr = ____cache_alloc(cachep, flags);
3514 if (ptr)
3515 goto out;
3516 }
3517 /* ___cache_alloc_node can fall back to other nodes */
3518 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3519 out:
3520 local_irq_restore(save_flags);
3521 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003522 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003523 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003524
Pekka Enbergc175eea2008-05-09 20:35:53 +02003525 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003526 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003527
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003528 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003529 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003530
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003531 return ptr;
3532}
3533
3534static __always_inline void *
3535__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3536{
3537 void *objp;
3538
3539 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3540 objp = alternate_node_alloc(cache, flags);
3541 if (objp)
3542 goto out;
3543 }
3544 objp = ____cache_alloc(cache, flags);
3545
3546 /*
3547 * We may just have run out of memory on the local node.
3548 * ____cache_alloc_node() knows how to locate memory on other nodes
3549 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003550 if (!objp)
3551 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003552
3553 out:
3554 return objp;
3555}
3556#else
3557
3558static __always_inline void *
3559__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3560{
3561 return ____cache_alloc(cachep, flags);
3562}
3563
3564#endif /* CONFIG_NUMA */
3565
3566static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003567slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003568{
3569 unsigned long save_flags;
3570 void *objp;
3571
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003572 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003573
Nick Piggincf40bd12009-01-21 08:12:39 +01003574 lockdep_trace_alloc(flags);
3575
Akinobu Mita773ff602008-12-23 19:37:01 +09003576 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003577 return NULL;
3578
Glauber Costad79923f2012-12-18 14:22:48 -08003579 cachep = memcg_kmem_get_cache(cachep, flags);
3580
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003581 cache_alloc_debugcheck_before(cachep, flags);
3582 local_irq_save(save_flags);
3583 objp = __do_cache_alloc(cachep, flags);
3584 local_irq_restore(save_flags);
3585 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003586 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003587 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003588 prefetchw(objp);
3589
Pekka Enbergc175eea2008-05-09 20:35:53 +02003590 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003591 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003592
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003593 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003594 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003595
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003596 return objp;
3597}
Christoph Lametere498be72005-09-09 13:03:32 -07003598
3599/*
3600 * Caller needs to acquire correct kmem_list's list_lock
3601 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003602static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003603 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604{
3605 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003606 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607
3608 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003609 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611
Mel Gorman072bb0a2012-07-31 16:43:58 -07003612 clear_obj_pfmemalloc(&objpp[i]);
3613 objp = objpp[i];
3614
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003615 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003616 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003618 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003620 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003622 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 check_slabp(cachep, slabp);
3624
3625 /* fixup slab chains */
3626 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003627 if (l3->free_objects > l3->free_limit) {
3628 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003629 /* No need to drop any previously held
3630 * lock here, even if we have a off-slab slab
3631 * descriptor it is guaranteed to come from
3632 * a different cache, refer to comments before
3633 * alloc_slabmgmt.
3634 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 slab_destroy(cachep, slabp);
3636 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003637 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 }
3639 } else {
3640 /* Unconditionally move a slab to the end of the
3641 * partial list on free - maximum time for the
3642 * other objects to be freed, too.
3643 */
Christoph Lametere498be72005-09-09 13:03:32 -07003644 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 }
3646 }
3647}
3648
Pekka Enberg343e0d72006-02-01 03:05:50 -08003649static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650{
3651 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003652 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003653 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654
3655 batchcount = ac->batchcount;
3656#if DEBUG
3657 BUG_ON(!batchcount || batchcount > ac->avail);
3658#endif
3659 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003660 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003661 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003662 if (l3->shared) {
3663 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003664 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 if (max) {
3666 if (batchcount > max)
3667 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003668 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003669 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 shared_array->avail += batchcount;
3671 goto free_done;
3672 }
3673 }
3674
Christoph Lameterff694162005-09-22 21:44:02 -07003675 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003676free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677#if STATS
3678 {
3679 int i = 0;
3680 struct list_head *p;
3681
Christoph Lametere498be72005-09-09 13:03:32 -07003682 p = l3->slabs_free.next;
3683 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684 struct slab *slabp;
3685
3686 slabp = list_entry(p, struct slab, list);
3687 BUG_ON(slabp->inuse);
3688
3689 i++;
3690 p = p->next;
3691 }
3692 STATS_SET_FREEABLE(cachep, i);
3693 }
3694#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003695 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003697 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698}
3699
3700/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003701 * Release an obj back to its cache. If the obj has a constructed state, it must
3702 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003704static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003705 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003707 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708
3709 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003710 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003711 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003713 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003714
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003715 /*
3716 * Skip calling cache_free_alien() when the platform is not numa.
3717 * This will avoid cache misses that happen while accessing slabp (which
3718 * is per page memory reference) to get nodeid. Instead use a global
3719 * variable to skip the call, which is mostly likely to be present in
3720 * the cache.
3721 */
Mel Gormanb6e68bc12009-06-16 15:32:16 -07003722 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003723 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003724
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 if (likely(ac->avail < ac->limit)) {
3726 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 } else {
3728 STATS_INC_FREEMISS(cachep);
3729 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003731
Mel Gorman072bb0a2012-07-31 16:43:58 -07003732 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733}
3734
3735/**
3736 * kmem_cache_alloc - Allocate an object
3737 * @cachep: The cache to allocate from.
3738 * @flags: See kmalloc().
3739 *
3740 * Allocate an object from this cache. The flags are only relevant
3741 * if the cache has no available objects.
3742 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003743void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003745 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003746
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003747 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003748 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003749
3750 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751}
3752EXPORT_SYMBOL(kmem_cache_alloc);
3753
Li Zefan0f24f122009-12-11 15:45:30 +08003754#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003755void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003756kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003757{
Steven Rostedt85beb582010-11-24 16:23:34 -05003758 void *ret;
3759
Ezequiel Garcia48356302012-09-08 17:47:57 -03003760 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003761
3762 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003763 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003764 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003765}
Steven Rostedt85beb582010-11-24 16:23:34 -05003766EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003767#endif
3768
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003770void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3771{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003772 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003773
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003774 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003775 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003776 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003777
3778 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003779}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780EXPORT_SYMBOL(kmem_cache_alloc_node);
3781
Li Zefan0f24f122009-12-11 15:45:30 +08003782#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003783void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003784 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003785 int nodeid,
3786 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003787{
Steven Rostedt85beb582010-11-24 16:23:34 -05003788 void *ret;
3789
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003790 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003791
Steven Rostedt85beb582010-11-24 16:23:34 -05003792 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003793 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003794 flags, nodeid);
3795 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003796}
Steven Rostedt85beb582010-11-24 16:23:34 -05003797EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003798#endif
3799
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003800static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003801__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003802{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003803 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003804
3805 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003806 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3807 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003808 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003809}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003810
Li Zefan0bb38a52009-12-11 15:45:50 +08003811#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003812void *__kmalloc_node(size_t size, gfp_t flags, int node)
3813{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003814 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003815}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003816EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003817
3818void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003819 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003820{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003821 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003822}
3823EXPORT_SYMBOL(__kmalloc_node_track_caller);
3824#else
3825void *__kmalloc_node(size_t size, gfp_t flags, int node)
3826{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003827 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003828}
3829EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003830#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003831#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832
3833/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003834 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003836 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003837 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003839static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003840 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003842 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003843 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003845 /* If you want to save a few bytes .text space: replace
3846 * __ with kmem_.
3847 * Then kmalloc uses the uninlined functions instead of the inline
3848 * functions.
3849 */
3850 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003851 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3852 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003853 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003854
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003855 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003856 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003857
3858 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003859}
3860
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003861
Li Zefan0bb38a52009-12-11 15:45:50 +08003862#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003863void *__kmalloc(size_t size, gfp_t flags)
3864{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003865 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866}
3867EXPORT_SYMBOL(__kmalloc);
3868
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003869void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003870{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003871 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003872}
3873EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003874
3875#else
3876void *__kmalloc(size_t size, gfp_t flags)
3877{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003878 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003879}
3880EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003881#endif
3882
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883/**
3884 * kmem_cache_free - Deallocate an object
3885 * @cachep: The cache the allocation was from.
3886 * @objp: The previously allocated object.
3887 *
3888 * Free an object which was previously allocated from this
3889 * cache.
3890 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003891void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892{
3893 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003894 cachep = cache_from_obj(cachep, objp);
3895 if (!cachep)
3896 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897
3898 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003899 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003900 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003901 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003902 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003904
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003905 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906}
3907EXPORT_SYMBOL(kmem_cache_free);
3908
3909/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910 * kfree - free previously allocated memory
3911 * @objp: pointer returned by kmalloc.
3912 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003913 * If @objp is NULL, no operation is performed.
3914 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915 * Don't free memory not originally allocated by kmalloc()
3916 * or you will run into trouble.
3917 */
3918void kfree(const void *objp)
3919{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003920 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921 unsigned long flags;
3922
Pekka Enberg2121db72009-03-25 11:05:57 +02003923 trace_kfree(_RET_IP_, objp);
3924
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003925 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926 return;
3927 local_irq_save(flags);
3928 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003929 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003930 debug_check_no_locks_freed(objp, c->object_size);
3931
3932 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003933 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 local_irq_restore(flags);
3935}
3936EXPORT_SYMBOL(kfree);
3937
Christoph Lametere498be72005-09-09 13:03:32 -07003938/*
Simon Arlott183ff222007-10-20 01:27:18 +02003939 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003940 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003941static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003942{
3943 int node;
3944 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003945 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003946 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003947
Mel Gorman9c09a952008-01-24 05:49:54 -08003948 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003949
Paul Menage3395ee02006-12-06 20:32:16 -08003950 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003951 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003952 if (!new_alien)
3953 goto fail;
3954 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003955
Eric Dumazet63109842007-05-06 14:49:28 -07003956 new_shared = NULL;
3957 if (cachep->shared) {
3958 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003959 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003960 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003961 if (!new_shared) {
3962 free_alien_cache(new_alien);
3963 goto fail;
3964 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003965 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003966
Andrew Mortona737b3e2006-03-22 00:08:11 -08003967 l3 = cachep->nodelists[node];
3968 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003969 struct array_cache *shared = l3->shared;
3970
Christoph Lametere498be72005-09-09 13:03:32 -07003971 spin_lock_irq(&l3->list_lock);
3972
Christoph Lametercafeb022006-03-25 03:06:46 -08003973 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003974 free_block(cachep, shared->entry,
3975 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003976
Christoph Lametercafeb022006-03-25 03:06:46 -08003977 l3->shared = new_shared;
3978 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003979 l3->alien = new_alien;
3980 new_alien = NULL;
3981 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003982 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003983 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003984 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003985 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003986 free_alien_cache(new_alien);
3987 continue;
3988 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003989 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003990 if (!l3) {
3991 free_alien_cache(new_alien);
3992 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003993 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003994 }
Christoph Lametere498be72005-09-09 13:03:32 -07003995
3996 kmem_list3_init(l3);
3997 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003998 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003999 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07004000 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004001 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004002 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004003 cachep->nodelists[node] = l3;
4004 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004005 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004006
Andrew Mortona737b3e2006-03-22 00:08:11 -08004007fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004008 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08004009 /* Cache is not active yet. Roll back what we did */
4010 node--;
4011 while (node >= 0) {
4012 if (cachep->nodelists[node]) {
4013 l3 = cachep->nodelists[node];
4014
4015 kfree(l3->shared);
4016 free_alien_cache(l3->alien);
4017 kfree(l3);
4018 cachep->nodelists[node] = NULL;
4019 }
4020 node--;
4021 }
4022 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004023 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07004024}
4025
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08004027 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004028 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029};
4030
4031static void do_ccupdate_local(void *info)
4032{
Andrew Mortona737b3e2006-03-22 00:08:11 -08004033 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034 struct array_cache *old;
4035
4036 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08004037 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07004038
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4040 new->new[smp_processor_id()] = old;
4041}
4042
Christoph Lameter18004c52012-07-06 15:25:12 -05004043/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08004044static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004045 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004047 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004048 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004050 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4051 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004052 if (!new)
4053 return -ENOMEM;
4054
Christoph Lametere498be72005-09-09 13:03:32 -07004055 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004056 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004057 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004058 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004059 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004060 kfree(new->new[i]);
4061 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004062 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 }
4064 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004065 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004067 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004068
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070 cachep->batchcount = batchcount;
4071 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004072 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
Christoph Lametere498be72005-09-09 13:03:32 -07004074 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004075 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 if (!ccold)
4077 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004078 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4079 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4080 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 kfree(ccold);
4082 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004083 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004084 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085}
4086
Glauber Costa943a4512012-12-18 14:23:03 -08004087static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4088 int batchcount, int shared, gfp_t gfp)
4089{
4090 int ret;
4091 struct kmem_cache *c = NULL;
4092 int i = 0;
4093
4094 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4095
4096 if (slab_state < FULL)
4097 return ret;
4098
4099 if ((ret < 0) || !is_root_cache(cachep))
4100 return ret;
4101
Glauber Costaebe945c2012-12-18 14:23:10 -08004102 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08004103 for_each_memcg_cache_index(i) {
4104 c = cache_from_memcg(cachep, i);
4105 if (c)
4106 /* return value determined by the parent cache only */
4107 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4108 }
4109
4110 return ret;
4111}
4112
Christoph Lameter18004c52012-07-06 15:25:12 -05004113/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004114static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115{
4116 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08004117 int limit = 0;
4118 int shared = 0;
4119 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
Glauber Costa943a4512012-12-18 14:23:03 -08004121 if (!is_root_cache(cachep)) {
4122 struct kmem_cache *root = memcg_root_cache(cachep);
4123 limit = root->limit;
4124 shared = root->shared;
4125 batchcount = root->batchcount;
4126 }
4127
4128 if (limit && shared && batchcount)
4129 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004130 /*
4131 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 * - create a LIFO ordering, i.e. return objects that are cache-warm
4133 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004134 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 * bufctl chains: array operations are cheaper.
4136 * The numbers are guessed, we should auto-tune as described by
4137 * Bonwick.
4138 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004139 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004141 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004143 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004145 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 limit = 54;
4147 else
4148 limit = 120;
4149
Andrew Mortona737b3e2006-03-22 00:08:11 -08004150 /*
4151 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 * allocation behaviour: Most allocs on one cpu, most free operations
4153 * on another cpu. For these cases, an efficient object passing between
4154 * cpus is necessary. This is provided by a shared array. The array
4155 * replaces Bonwick's magazine layer.
4156 * On uniprocessor, it's functionally equivalent (but less efficient)
4157 * to a larger limit. Thus disabled by default.
4158 */
4159 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004160 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162
4163#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004164 /*
4165 * With debugging enabled, large batchcount lead to excessively long
4166 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 */
4168 if (limit > 32)
4169 limit = 32;
4170#endif
Glauber Costa943a4512012-12-18 14:23:03 -08004171 batchcount = (limit + 1) / 2;
4172skip_setup:
4173 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174 if (err)
4175 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004176 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004177 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178}
4179
Christoph Lameter1b552532006-03-22 00:09:07 -08004180/*
4181 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004182 * necessary. Note that the l3 listlock also protects the array_cache
4183 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004184 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004185static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004186 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187{
4188 int tofree;
4189
Christoph Lameter1b552532006-03-22 00:09:07 -08004190 if (!ac || !ac->avail)
4191 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 if (ac->touched && !force) {
4193 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004194 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004195 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004196 if (ac->avail) {
4197 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4198 if (tofree > ac->avail)
4199 tofree = (ac->avail + 1) / 2;
4200 free_block(cachep, ac->entry, tofree, node);
4201 ac->avail -= tofree;
4202 memmove(ac->entry, &(ac->entry[tofree]),
4203 sizeof(void *) * ac->avail);
4204 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004205 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 }
4207}
4208
4209/**
4210 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004211 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212 *
4213 * Called from workqueue/eventd every few seconds.
4214 * Purpose:
4215 * - clear the per-cpu caches for this CPU.
4216 * - return freeable pages to the main free memory pool.
4217 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004218 * If we cannot acquire the cache chain mutex then just give up - we'll try
4219 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004221static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004223 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004224 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004225 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004226 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227
Christoph Lameter18004c52012-07-06 15:25:12 -05004228 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004230 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231
Christoph Lameter18004c52012-07-06 15:25:12 -05004232 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 check_irq_on();
4234
Christoph Lameter35386e32006-03-22 00:09:05 -08004235 /*
4236 * We only take the l3 lock if absolutely necessary and we
4237 * have established with reasonable certainty that
4238 * we can do some work if the lock was obtained.
4239 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004240 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004241
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004242 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243
Christoph Lameteraab22072006-03-22 00:09:06 -08004244 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245
Christoph Lameter35386e32006-03-22 00:09:05 -08004246 /*
4247 * These are racy checks but it does not matter
4248 * if we skip one check or scan twice.
4249 */
Christoph Lametere498be72005-09-09 13:03:32 -07004250 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004251 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252
Christoph Lametere498be72005-09-09 13:03:32 -07004253 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254
Christoph Lameteraab22072006-03-22 00:09:06 -08004255 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256
Christoph Lametered11d9e2006-06-30 01:55:45 -07004257 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004258 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004259 else {
4260 int freed;
4261
4262 freed = drain_freelist(searchp, l3, (l3->free_limit +
4263 5 * searchp->num - 1) / (5 * searchp->num));
4264 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004266next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267 cond_resched();
4268 }
4269 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004270 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004271 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004272out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004273 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004274 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275}
4276
Linus Torvalds158a9622008-01-02 13:04:48 -08004277#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004278void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004280 struct slab *slabp;
4281 unsigned long active_objs;
4282 unsigned long num_objs;
4283 unsigned long active_slabs = 0;
4284 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004285 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004287 int node;
4288 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290 active_objs = 0;
4291 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004292 for_each_online_node(node) {
4293 l3 = cachep->nodelists[node];
4294 if (!l3)
4295 continue;
4296
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004297 check_irq_on();
4298 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004299
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004300 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004301 if (slabp->inuse != cachep->num && !error)
4302 error = "slabs_full accounting error";
4303 active_objs += cachep->num;
4304 active_slabs++;
4305 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004306 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004307 if (slabp->inuse == cachep->num && !error)
4308 error = "slabs_partial inuse accounting error";
4309 if (!slabp->inuse && !error)
4310 error = "slabs_partial/inuse accounting error";
4311 active_objs += slabp->inuse;
4312 active_slabs++;
4313 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004314 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004315 if (slabp->inuse && !error)
4316 error = "slabs_free/inuse accounting error";
4317 num_slabs++;
4318 }
4319 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004320 if (l3->shared)
4321 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004322
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004323 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004325 num_slabs += active_slabs;
4326 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004327 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004328 error = "free_objects accounting error";
4329
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004330 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331 if (error)
4332 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4333
Glauber Costa0d7561c2012-10-19 18:20:27 +04004334 sinfo->active_objs = active_objs;
4335 sinfo->num_objs = num_objs;
4336 sinfo->active_slabs = active_slabs;
4337 sinfo->num_slabs = num_slabs;
4338 sinfo->shared_avail = shared_avail;
4339 sinfo->limit = cachep->limit;
4340 sinfo->batchcount = cachep->batchcount;
4341 sinfo->shared = cachep->shared;
4342 sinfo->objects_per_slab = cachep->num;
4343 sinfo->cache_order = cachep->gfporder;
4344}
4345
4346void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4347{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004349 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 unsigned long high = cachep->high_mark;
4351 unsigned long allocs = cachep->num_allocations;
4352 unsigned long grown = cachep->grown;
4353 unsigned long reaped = cachep->reaped;
4354 unsigned long errors = cachep->errors;
4355 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004357 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004358 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004359
Joe Perchese92dd4f2010-03-26 19:27:58 -07004360 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4361 "%4lu %4lu %4lu %4lu %4lu",
4362 allocs, high, grown,
4363 reaped, errors, max_freeable, node_allocs,
4364 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004365 }
4366 /* cpu stats */
4367 {
4368 unsigned long allochit = atomic_read(&cachep->allochit);
4369 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4370 unsigned long freehit = atomic_read(&cachep->freehit);
4371 unsigned long freemiss = atomic_read(&cachep->freemiss);
4372
4373 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004374 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375 }
4376#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004377}
4378
Linus Torvalds1da177e2005-04-16 15:20:36 -07004379#define MAX_SLABINFO_WRITE 128
4380/**
4381 * slabinfo_write - Tuning for the slab allocator
4382 * @file: unused
4383 * @buffer: user buffer
4384 * @count: data length
4385 * @ppos: unused
4386 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004387ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004388 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004389{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004390 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004391 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004392 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004393
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394 if (count > MAX_SLABINFO_WRITE)
4395 return -EINVAL;
4396 if (copy_from_user(&kbuf, buffer, count))
4397 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004398 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004399
4400 tmp = strchr(kbuf, ' ');
4401 if (!tmp)
4402 return -EINVAL;
4403 *tmp = '\0';
4404 tmp++;
4405 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4406 return -EINVAL;
4407
4408 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004409 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004410 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004411 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004412 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004413 if (limit < 1 || batchcount < 1 ||
4414 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004415 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004417 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004418 batchcount, shared,
4419 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420 }
4421 break;
4422 }
4423 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004424 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004425 if (res >= 0)
4426 res = count;
4427 return res;
4428}
Al Viro871751e2006-03-25 03:06:39 -08004429
4430#ifdef CONFIG_DEBUG_SLAB_LEAK
4431
4432static void *leaks_start(struct seq_file *m, loff_t *pos)
4433{
Christoph Lameter18004c52012-07-06 15:25:12 -05004434 mutex_lock(&slab_mutex);
4435 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004436}
4437
4438static inline int add_caller(unsigned long *n, unsigned long v)
4439{
4440 unsigned long *p;
4441 int l;
4442 if (!v)
4443 return 1;
4444 l = n[1];
4445 p = n + 2;
4446 while (l) {
4447 int i = l/2;
4448 unsigned long *q = p + 2 * i;
4449 if (*q == v) {
4450 q[1]++;
4451 return 1;
4452 }
4453 if (*q > v) {
4454 l = i;
4455 } else {
4456 p = q + 2;
4457 l -= i + 1;
4458 }
4459 }
4460 if (++n[1] == n[0])
4461 return 0;
4462 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4463 p[0] = v;
4464 p[1] = 1;
4465 return 1;
4466}
4467
4468static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4469{
4470 void *p;
4471 int i;
4472 if (n[0] == n[1])
4473 return;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004474 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
Al Viro871751e2006-03-25 03:06:39 -08004475 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4476 continue;
4477 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4478 return;
4479 }
4480}
4481
4482static void show_symbol(struct seq_file *m, unsigned long address)
4483{
4484#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004485 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004486 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004487
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004488 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004489 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004490 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004491 seq_printf(m, " [%s]", modname);
4492 return;
4493 }
4494#endif
4495 seq_printf(m, "%p", (void *)address);
4496}
4497
4498static int leaks_show(struct seq_file *m, void *p)
4499{
Thierry Reding0672aa72012-06-22 19:42:49 +02004500 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Al Viro871751e2006-03-25 03:06:39 -08004501 struct slab *slabp;
4502 struct kmem_list3 *l3;
4503 const char *name;
4504 unsigned long *n = m->private;
4505 int node;
4506 int i;
4507
4508 if (!(cachep->flags & SLAB_STORE_USER))
4509 return 0;
4510 if (!(cachep->flags & SLAB_RED_ZONE))
4511 return 0;
4512
4513 /* OK, we can do it */
4514
4515 n[1] = 0;
4516
4517 for_each_online_node(node) {
4518 l3 = cachep->nodelists[node];
4519 if (!l3)
4520 continue;
4521
4522 check_irq_on();
4523 spin_lock_irq(&l3->list_lock);
4524
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004525 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004526 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004527 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004528 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004529 spin_unlock_irq(&l3->list_lock);
4530 }
4531 name = cachep->name;
4532 if (n[0] == n[1]) {
4533 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004534 mutex_unlock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004535 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4536 if (!m->private) {
4537 /* Too bad, we are really out */
4538 m->private = n;
Christoph Lameter18004c52012-07-06 15:25:12 -05004539 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004540 return -ENOMEM;
4541 }
4542 *(unsigned long *)m->private = n[0] * 2;
4543 kfree(n);
Christoph Lameter18004c52012-07-06 15:25:12 -05004544 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004545 /* Now make sure this entry will be retried */
4546 m->count = m->size;
4547 return 0;
4548 }
4549 for (i = 0; i < n[1]; i++) {
4550 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4551 show_symbol(m, n[2*i+2]);
4552 seq_putc(m, '\n');
4553 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004554
Al Viro871751e2006-03-25 03:06:39 -08004555 return 0;
4556}
4557
Glauber Costab7454ad2012-10-19 18:20:25 +04004558static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4559{
4560 return seq_list_next(p, &slab_caches, pos);
4561}
4562
4563static void s_stop(struct seq_file *m, void *p)
4564{
4565 mutex_unlock(&slab_mutex);
4566}
4567
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004568static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004569 .start = leaks_start,
4570 .next = s_next,
4571 .stop = s_stop,
4572 .show = leaks_show,
4573};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004574
4575static int slabstats_open(struct inode *inode, struct file *file)
4576{
4577 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4578 int ret = -ENOMEM;
4579 if (n) {
4580 ret = seq_open(file, &slabstats_op);
4581 if (!ret) {
4582 struct seq_file *m = file->private_data;
4583 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4584 m->private = n;
4585 n = NULL;
4586 }
4587 kfree(n);
4588 }
4589 return ret;
4590}
4591
4592static const struct file_operations proc_slabstats_operations = {
4593 .open = slabstats_open,
4594 .read = seq_read,
4595 .llseek = seq_lseek,
4596 .release = seq_release_private,
4597};
Al Viro871751e2006-03-25 03:06:39 -08004598#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004599
4600static int __init slab_proc_init(void)
4601{
4602#ifdef CONFIG_DEBUG_SLAB_LEAK
4603 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4604#endif
4605 return 0;
4606}
4607module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004608#endif
4609
Manfred Spraul00e145b2005-09-03 15:55:07 -07004610/**
4611 * ksize - get the actual amount of memory allocated for a given object
4612 * @objp: Pointer to the object
4613 *
4614 * kmalloc may internally round up allocations and return more memory
4615 * than requested. ksize() can be used to determine the actual amount of
4616 * memory allocated. The caller may use this additional memory, even though
4617 * a smaller amount of memory was initially specified with the kmalloc call.
4618 * The caller must guarantee that objp points to a valid object previously
4619 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4620 * must not be freed during the duration of the call.
4621 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004622size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004623{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004624 BUG_ON(!objp);
4625 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004626 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004627
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004628 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004629}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004630EXPORT_SYMBOL(ksize);