blob: 65b5dcb6f67107d6e46515196cfdf5cfbfd8a70b [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
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163#if FREELIST_BYTE_INDEX
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
David Miller30321c72014-05-05 16:20:04 -0400169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900170
Mel Gorman072bb0a2012-07-31 16:43:58 -0700171/*
172 * true if a page was allocated from pfmemalloc reserves for network-based
173 * swap
174 */
175static bool pfmemalloc_active __read_mostly;
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 * struct array_cache
179 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 * Purpose:
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
184 *
185 * The limit is stored in the per-cpu structure to reduce the data cache
186 * footprint.
187 *
188 */
189struct array_cache {
190 unsigned int avail;
191 unsigned int limit;
192 unsigned int batchcount;
193 unsigned int touched;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700194 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
197 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700198 *
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800202 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203};
204
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700205struct alien_cache {
206 spinlock_t lock;
207 struct array_cache ac;
208};
209
Mel Gorman072bb0a2012-07-31 16:43:58 -0700210#define SLAB_OBJ_PFMEMALLOC 1
211static inline bool is_obj_pfmemalloc(void *objp)
212{
213 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
214}
215
216static inline void set_obj_pfmemalloc(void **objp)
217{
218 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
219 return;
220}
221
222static inline void clear_obj_pfmemalloc(void **objp)
223{
224 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
225}
226
Andrew Mortona737b3e2006-03-22 00:08:11 -0800227/*
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 */
231#define BOOT_CPUCACHE_ENTRIES 1
232struct arraycache_init {
233 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800234 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235};
236
237/*
Christoph Lametere498be72005-09-09 13:03:32 -0700238 * Need this for bootstrapping a per node allocator.
239 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700240#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000241static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700242#define CACHE_CACHE 0
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700243#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Christoph Lametered11d9e2006-06-30 01:55:45 -0700245static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000246 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700247static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700248 int node, struct list_head *list);
249static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300250static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000251static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700252
Ingo Molnare0a42722006-06-23 02:03:46 -0700253static int slab_early_init = 1;
254
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000255#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700256
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000257static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700258{
259 INIT_LIST_HEAD(&parent->slabs_full);
260 INIT_LIST_HEAD(&parent->slabs_partial);
261 INIT_LIST_HEAD(&parent->slabs_free);
262 parent->shared = NULL;
263 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800264 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700265 spin_lock_init(&parent->list_lock);
266 parent->free_objects = 0;
267 parent->free_touched = 0;
268}
269
Andrew Mortona737b3e2006-03-22 00:08:11 -0800270#define MAKE_LIST(cachep, listp, slab, nodeid) \
271 do { \
272 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700273 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700274 } while (0)
275
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
277 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700278 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
279 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
280 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
281 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283#define CFLGS_OFF_SLAB (0x80000000UL)
284#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
285
286#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800287/*
288 * Optimization question: fewer reaps means less probability for unnessary
289 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100291 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 * which could lock up otherwise freeable slabs.
293 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800294#define REAPTIMEOUT_AC (2*HZ)
295#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297#if STATS
298#define STATS_INC_ACTIVE(x) ((x)->num_active++)
299#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
300#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
301#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700302#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800303#define STATS_SET_HIGH(x) \
304 do { \
305 if ((x)->num_active > (x)->high_mark) \
306 (x)->high_mark = (x)->num_active; \
307 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308#define STATS_INC_ERR(x) ((x)->errors++)
309#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700310#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700311#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800312#define STATS_SET_FREEABLE(x, i) \
313 do { \
314 if ((x)->max_freeable < i) \
315 (x)->max_freeable = i; \
316 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
318#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
319#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
320#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
321#else
322#define STATS_INC_ACTIVE(x) do { } while (0)
323#define STATS_DEC_ACTIVE(x) do { } while (0)
324#define STATS_INC_ALLOCED(x) do { } while (0)
325#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700326#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327#define STATS_SET_HIGH(x) do { } while (0)
328#define STATS_INC_ERR(x) do { } while (0)
329#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700330#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700331#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800332#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333#define STATS_INC_ALLOCHIT(x) do { } while (0)
334#define STATS_INC_ALLOCMISS(x) do { } while (0)
335#define STATS_INC_FREEHIT(x) do { } while (0)
336#define STATS_INC_FREEMISS(x) do { } while (0)
337#endif
338
339#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Andrew Mortona737b3e2006-03-22 00:08:11 -0800341/*
342 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800344 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 * the end of an object is aligned with the end of the real
346 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800347 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800349 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500350 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
351 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800352 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800354static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800356 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357}
358
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360{
361 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700362 return (unsigned long long*) (objp + obj_offset(cachep) -
363 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364}
365
David Woodhouseb46b8f12007-05-08 00:22:59 -0700366static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367{
368 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
369 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500370 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700371 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400372 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500373 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700374 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
Pekka Enberg343e0d72006-02-01 03:05:50 -0800377static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378{
379 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500380 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381}
382
383#else
384
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800385#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700386#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
387#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
389
390#endif
391
Joonsoo Kim03787302014-06-23 13:22:06 -0700392#define OBJECT_FREE (0)
393#define OBJECT_ACTIVE (1)
394
395#ifdef CONFIG_DEBUG_SLAB_LEAK
396
397static void set_obj_status(struct page *page, int idx, int val)
398{
399 int freelist_size;
400 char *status;
401 struct kmem_cache *cachep = page->slab_cache;
402
403 freelist_size = cachep->num * sizeof(freelist_idx_t);
404 status = (char *)page->freelist + freelist_size;
405 status[idx] = val;
406}
407
408static inline unsigned int get_obj_status(struct page *page, int idx)
409{
410 int freelist_size;
411 char *status;
412 struct kmem_cache *cachep = page->slab_cache;
413
414 freelist_size = cachep->num * sizeof(freelist_idx_t);
415 status = (char *)page->freelist + freelist_size;
416
417 return status[idx];
418}
419
420#else
421static inline void set_obj_status(struct page *page, int idx, int val) {}
422
423#endif
424
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700426 * Do not go above this order unless 0 objects fit into the slab or
427 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 */
David Rientjes543585c2011-10-18 22:09:24 -0700429#define SLAB_MAX_ORDER_HI 1
430#define SLAB_MAX_ORDER_LO 0
431static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700432static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800434static inline struct kmem_cache *virt_to_cache(const void *obj)
435{
Christoph Lameterb49af682007-05-06 14:49:41 -0700436 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500437 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800438}
439
Joonsoo Kim8456a642013-10-24 10:07:49 +0900440static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800441 unsigned int idx)
442{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900443 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800444}
445
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800446/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500447 * We want to avoid an expensive divide : (offset / cache->size)
448 * Using the fact that size is a constant for a particular cache,
449 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800450 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
451 */
452static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900453 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800454{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900455 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800456 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800457}
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000460static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800461 .batchcount = 1,
462 .limit = BOOT_CPUCACHE_ENTRIES,
463 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500464 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800465 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466};
467
Joonsoo Kimedcad252014-08-08 14:19:15 -0700468#define BAD_ALIEN_MAGIC 0x01020304ul
469
Tejun Heo1871e522009-10-29 22:34:13 +0900470static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Pekka Enberg343e0d72006-02-01 03:05:50 -0800472static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700474 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Joonsoo Kim03787302014-06-23 13:22:06 -0700477static size_t calculate_freelist_size(int nr_objs, size_t align)
478{
479 size_t freelist_size;
480
481 freelist_size = nr_objs * sizeof(freelist_idx_t);
482 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
483 freelist_size += nr_objs * sizeof(char);
484
485 if (align)
486 freelist_size = ALIGN(freelist_size, align);
487
488 return freelist_size;
489}
490
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900491static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
492 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900494 int nr_objs;
Joonsoo Kim03787302014-06-23 13:22:06 -0700495 size_t remained_size;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900496 size_t freelist_size;
Joonsoo Kim03787302014-06-23 13:22:06 -0700497 int extra_space = 0;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900498
Joonsoo Kim03787302014-06-23 13:22:06 -0700499 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
500 extra_space = sizeof(char);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900501 /*
502 * Ignore padding for the initial guess. The padding
503 * is at most @align-1 bytes, and @buffer_size is at
504 * least @align. In the worst case, this result will
505 * be one greater than the number of objects that fit
506 * into the memory allocation when taking the padding
507 * into account.
508 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700509 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900510
511 /*
512 * This calculated number will be either the right
513 * amount, or one greater than what we want.
514 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700515 remained_size = slab_size - nr_objs * buffer_size;
516 freelist_size = calculate_freelist_size(nr_objs, align);
517 if (remained_size < freelist_size)
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900518 nr_objs--;
519
520 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800521}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
Andrew Mortona737b3e2006-03-22 00:08:11 -0800523/*
524 * Calculate the number of objects and left-over bytes for a given buffer size.
525 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800526static void cache_estimate(unsigned long gfporder, size_t buffer_size,
527 size_t align, int flags, size_t *left_over,
528 unsigned int *num)
529{
530 int nr_objs;
531 size_t mgmt_size;
532 size_t slab_size = PAGE_SIZE << gfporder;
533
534 /*
535 * The slab management structure can be either off the slab or
536 * on it. For the latter case, the memory allocated for a
537 * slab is used for:
538 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900539 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800540 * - Padding to respect alignment of @align
541 * - @buffer_size bytes for each object
542 *
543 * If the slab management structure is off the slab, then the
544 * alignment will already be calculated into the size. Because
545 * the slabs are all pages aligned, the objects will be at the
546 * correct alignment when allocated.
547 */
548 if (flags & CFLGS_OFF_SLAB) {
549 mgmt_size = 0;
550 nr_objs = slab_size / buffer_size;
551
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800552 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900553 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900554 sizeof(freelist_idx_t), align);
Joonsoo Kim03787302014-06-23 13:22:06 -0700555 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800557 *num = nr_objs;
558 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
Christoph Lameterf28510d2012-09-11 19:49:38 +0000561#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700562#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
Andrew Mortona737b3e2006-03-22 00:08:11 -0800564static void __slab_error(const char *function, struct kmem_cache *cachep,
565 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566{
567 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800568 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030570 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000572#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Paul Menage3395ee02006-12-06 20:32:16 -0800574/*
575 * By default on NUMA we use alien caches to stage the freeing of
576 * objects allocated from other nodes. This causes massive memory
577 * inefficiencies when using fake NUMA setup to split memory into a
578 * large number of small nodes, so it can be disabled on the command
579 * line
580 */
581
582static int use_alien_caches __read_mostly = 1;
583static int __init noaliencache_setup(char *s)
584{
585 use_alien_caches = 0;
586 return 1;
587}
588__setup("noaliencache", noaliencache_setup);
589
David Rientjes3df1ccc2011-10-18 22:09:28 -0700590static int __init slab_max_order_setup(char *str)
591{
592 get_option(&str, &slab_max_order);
593 slab_max_order = slab_max_order < 0 ? 0 :
594 min(slab_max_order, MAX_ORDER - 1);
595 slab_max_order_set = true;
596
597 return 1;
598}
599__setup("slab_max_order=", slab_max_order_setup);
600
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800601#ifdef CONFIG_NUMA
602/*
603 * Special reaping functions for NUMA systems called from cache_reap().
604 * These take care of doing round robin flushing of alien caches (containing
605 * objects freed on different nodes from which they were allocated) and the
606 * flushing of remote pcps by calling drain_node_pages.
607 */
Tejun Heo1871e522009-10-29 22:34:13 +0900608static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800609
610static void init_reap_node(int cpu)
611{
612 int node;
613
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700614 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800615 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800616 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800617
Tejun Heo1871e522009-10-29 22:34:13 +0900618 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800619}
620
621static void next_reap_node(void)
622{
Christoph Lameter909ea962010-12-08 16:22:55 +0100623 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800624
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800625 node = next_node(node, node_online_map);
626 if (unlikely(node >= MAX_NUMNODES))
627 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100628 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800629}
630
631#else
632#define init_reap_node(cpu) do { } while (0)
633#define next_reap_node(void) do { } while (0)
634#endif
635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636/*
637 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
638 * via the workqueue/eventd.
639 * Add the CPU number into the expiration time to minimize the possibility of
640 * the CPUs getting into lockstep and contending for the global cache chain
641 * lock.
642 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400643static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644{
Tejun Heo1871e522009-10-29 22:34:13 +0900645 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
647 /*
648 * When this gets called from do_initcalls via cpucache_init(),
649 * init_workqueues() has already run, so keventd will be setup
650 * at that time.
651 */
David Howells52bad642006-11-22 14:54:01 +0000652 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800653 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700654 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800655 schedule_delayed_work_on(cpu, reap_work,
656 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 }
658}
659
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700660static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100662 /*
663 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300664 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100665 * cache the pointers are not cleared and they could be counted as
666 * valid references during a kmemleak scan. Therefore, kmemleak must
667 * not scan such objects.
668 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700669 kmemleak_no_scan(ac);
670 if (ac) {
671 ac->avail = 0;
672 ac->limit = limit;
673 ac->batchcount = batch;
674 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700676}
677
678static struct array_cache *alloc_arraycache(int node, int entries,
679 int batchcount, gfp_t gfp)
680{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700681 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700682 struct array_cache *ac = NULL;
683
684 ac = kmalloc_node(memsize, gfp, node);
685 init_arraycache(ac, entries, batchcount);
686 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687}
688
Joonsoo Kim8456a642013-10-24 10:07:49 +0900689static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700690{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700691 return PageSlabPfmemalloc(page);
692}
693
694/* Clears pfmemalloc_active if no slabs have pfmalloc set */
695static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
696 struct array_cache *ac)
697{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700698 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
Joonsoo Kim8456a642013-10-24 10:07:49 +0900699 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700700 unsigned long flags;
701
702 if (!pfmemalloc_active)
703 return;
704
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000705 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900706 list_for_each_entry(page, &n->slabs_full, lru)
707 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700708 goto out;
709
Joonsoo Kim8456a642013-10-24 10:07:49 +0900710 list_for_each_entry(page, &n->slabs_partial, lru)
711 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700712 goto out;
713
Joonsoo Kim8456a642013-10-24 10:07:49 +0900714 list_for_each_entry(page, &n->slabs_free, lru)
715 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700716 goto out;
717
718 pfmemalloc_active = false;
719out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000720 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700721}
722
Mel Gorman381760e2012-07-31 16:44:30 -0700723static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700724 gfp_t flags, bool force_refill)
725{
726 int i;
727 void *objp = ac->entry[--ac->avail];
728
729 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
730 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000731 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700732
733 if (gfp_pfmemalloc_allowed(flags)) {
734 clear_obj_pfmemalloc(&objp);
735 return objp;
736 }
737
738 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700739 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700740 /* If a !PFMEMALLOC object is found, swap them */
741 if (!is_obj_pfmemalloc(ac->entry[i])) {
742 objp = ac->entry[i];
743 ac->entry[i] = ac->entry[ac->avail];
744 ac->entry[ac->avail] = objp;
745 return objp;
746 }
747 }
748
749 /*
750 * If there are empty slabs on the slabs_free list and we are
751 * being forced to refill the cache, mark this one !pfmemalloc.
752 */
Christoph Lameter18bf8542014-08-06 16:04:11 -0700753 n = get_node(cachep, numa_mem_id());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000754 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900755 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900756 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700757 clear_obj_pfmemalloc(&objp);
758 recheck_pfmemalloc_active(cachep, ac);
759 return objp;
760 }
761
762 /* No !PFMEMALLOC objects available */
763 ac->avail++;
764 objp = NULL;
765 }
766
767 return objp;
768}
769
Mel Gorman381760e2012-07-31 16:44:30 -0700770static inline void *ac_get_obj(struct kmem_cache *cachep,
771 struct array_cache *ac, gfp_t flags, bool force_refill)
772{
773 void *objp;
774
775 if (unlikely(sk_memalloc_socks()))
776 objp = __ac_get_obj(cachep, ac, flags, force_refill);
777 else
778 objp = ac->entry[--ac->avail];
779
780 return objp;
781}
782
Joonsoo Kimd3aec342014-10-09 15:26:06 -0700783static noinline void *__ac_put_obj(struct kmem_cache *cachep,
784 struct array_cache *ac, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700785{
786 if (unlikely(pfmemalloc_active)) {
787 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700788 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700789 if (PageSlabPfmemalloc(page))
790 set_obj_pfmemalloc(&objp);
791 }
792
Mel Gorman381760e2012-07-31 16:44:30 -0700793 return objp;
794}
795
796static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
797 void *objp)
798{
799 if (unlikely(sk_memalloc_socks()))
800 objp = __ac_put_obj(cachep, ac, objp);
801
Mel Gorman072bb0a2012-07-31 16:43:58 -0700802 ac->entry[ac->avail++] = objp;
803}
804
Christoph Lameter3ded1752006-03-25 03:06:44 -0800805/*
806 * Transfer objects in one arraycache to another.
807 * Locking must be handled by the caller.
808 *
809 * Return the number of entries transferred.
810 */
811static int transfer_objects(struct array_cache *to,
812 struct array_cache *from, unsigned int max)
813{
814 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700815 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800816
817 if (!nr)
818 return 0;
819
820 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
821 sizeof(void *) *nr);
822
823 from->avail -= nr;
824 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800825 return nr;
826}
827
Christoph Lameter765c4502006-09-27 01:50:08 -0700828#ifndef CONFIG_NUMA
829
830#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000831#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700832
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700833static inline struct alien_cache **alloc_alien_cache(int node,
834 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700835{
Joonsoo Kimedcad252014-08-08 14:19:15 -0700836 return (struct alien_cache **)BAD_ALIEN_MAGIC;
Christoph Lameter765c4502006-09-27 01:50:08 -0700837}
838
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700839static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700840{
841}
842
843static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
844{
845 return 0;
846}
847
848static inline void *alternate_node_alloc(struct kmem_cache *cachep,
849 gfp_t flags)
850{
851 return NULL;
852}
853
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800854static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700855 gfp_t flags, int nodeid)
856{
857 return NULL;
858}
859
860#else /* CONFIG_NUMA */
861
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800862static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800863static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800864
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700865static struct alien_cache *__alloc_alien_cache(int node, int entries,
866 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700867{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700868 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700869 struct alien_cache *alc = NULL;
870
871 alc = kmalloc_node(memsize, gfp, node);
872 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700873 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700874 return alc;
875}
876
877static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
878{
879 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700880 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700881 int i;
882
883 if (limit > 1)
884 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700885 alc_ptr = kzalloc_node(memsize, gfp, node);
886 if (!alc_ptr)
887 return NULL;
888
889 for_each_node(i) {
890 if (i == node || !node_online(i))
891 continue;
892 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
893 if (!alc_ptr[i]) {
894 for (i--; i >= 0; i--)
895 kfree(alc_ptr[i]);
896 kfree(alc_ptr);
897 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700898 }
899 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700900 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700901}
902
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700903static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700904{
905 int i;
906
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700907 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700908 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700909 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700910 kfree(alc_ptr[i]);
911 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700912}
913
Pekka Enberg343e0d72006-02-01 03:05:50 -0800914static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700915 struct array_cache *ac, int node,
916 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700917{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700918 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700919
920 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000921 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800922 /*
923 * Stuff objects into the remote nodes shared array first.
924 * That way we could avoid the overhead of putting the objects
925 * into the free lists and getting them back later.
926 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000927 if (n->shared)
928 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800929
Joonsoo Kim833b7062014-08-06 16:04:33 -0700930 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700931 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000932 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700933 }
934}
935
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800936/*
937 * Called from cache_reap() to regularly drain alien caches round robin.
938 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000939static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800940{
Christoph Lameter909ea962010-12-08 16:22:55 +0100941 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800942
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000943 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700944 struct alien_cache *alc = n->alien[node];
945 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800946
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700947 if (alc) {
948 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700949 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700950 LIST_HEAD(list);
951
952 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700953 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700954 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700955 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800956 }
957 }
958}
959
Andrew Mortona737b3e2006-03-22 00:08:11 -0800960static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700961 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700962{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800963 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700964 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700965 struct array_cache *ac;
966 unsigned long flags;
967
968 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700969 alc = alien[i];
970 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700971 LIST_HEAD(list);
972
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700973 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700974 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700975 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700976 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700977 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700978 }
979 }
980}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700981
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700982static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
983 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700984{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000985 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700986 struct alien_cache *alien = NULL;
987 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700988 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700989
Christoph Lameter18bf8542014-08-06 16:04:11 -0700990 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700991 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700992 if (n->alien && n->alien[page_node]) {
993 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700994 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700995 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700996 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700997 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700998 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700999 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001000 ac_put_obj(cachep, ac, objp);
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001001 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001002 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001003 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001004 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001005 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001006 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001007 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001008 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001009 }
1010 return 1;
1011}
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001012
1013static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1014{
1015 int page_node = page_to_nid(virt_to_page(objp));
1016 int node = numa_mem_id();
1017 /*
1018 * Make sure we are not freeing a object from another node to the array
1019 * cache on this cpu.
1020 */
1021 if (likely(node == page_node))
1022 return 0;
1023
1024 return __cache_free_alien(cachep, objp, node, page_node);
1025}
Christoph Lametere498be72005-09-09 13:03:32 -07001026#endif
1027
David Rientjes8f9f8d92010-03-27 19:40:47 -07001028/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001029 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001030 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001031 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001032 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001033 * already in use.
1034 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001035 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001036 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001037static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001038{
1039 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001040 struct kmem_cache_node *n;
Joonsoo Kim5e804782014-08-06 16:04:40 -07001041 const size_t memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001042
Christoph Lameter18004c52012-07-06 15:25:12 -05001043 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001044 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001045 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001046 * begin anything. Make sure some other cpu on this
1047 * node has not already allocated this
1048 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07001049 n = get_node(cachep, node);
1050 if (!n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001051 n = kmalloc_node(memsize, GFP_KERNEL, node);
1052 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001053 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001054 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001055 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1056 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001057
1058 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001059 * The kmem_cache_nodes don't come and go as CPUs
1060 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001061 * protection here.
1062 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001063 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001064 }
1065
Christoph Lameter18bf8542014-08-06 16:04:11 -07001066 spin_lock_irq(&n->list_lock);
1067 n->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001068 (1 + nr_cpus_node(node)) *
1069 cachep->batchcount + cachep->num;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001070 spin_unlock_irq(&n->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001071 }
1072 return 0;
1073}
1074
Wanpeng Li0fa81032013-07-04 08:33:22 +08001075static inline int slabs_tofree(struct kmem_cache *cachep,
1076 struct kmem_cache_node *n)
1077{
1078 return (n->free_objects + cachep->num - 1) / cachep->num;
1079}
1080
Paul Gortmaker0db06282013-06-19 14:53:51 -04001081static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001083 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001084 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001085 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301086 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001087
Christoph Lameter18004c52012-07-06 15:25:12 -05001088 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001089 struct array_cache *nc;
1090 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001091 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -07001092 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001093
Christoph Lameter18bf8542014-08-06 16:04:11 -07001094 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001095 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001096 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001097
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001098 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001099
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001100 /* Free limit for this kmem_cache_node */
1101 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001102
1103 /* cpu is dead; no one can alloc from it. */
1104 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1105 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001106 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001107 nc->avail = 0;
1108 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001109
Rusty Russell58463c12009-12-17 11:43:12 -06001110 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001111 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001112 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001113 }
1114
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001115 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001116 if (shared) {
1117 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001118 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001119 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001120 }
1121
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001122 alien = n->alien;
1123 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001124
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001126
1127 kfree(shared);
1128 if (alien) {
1129 drain_alien_cache(cachep, alien);
1130 free_alien_cache(alien);
1131 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001132
1133free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001134 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001135 }
1136 /*
1137 * In the previous loop, all the objects were freed to
1138 * the respective cache's slabs, now we can go ahead and
1139 * shrink each nodelist to its limit.
1140 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001141 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001142 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001143 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001144 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001145 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001146 }
1147}
1148
Paul Gortmaker0db06282013-06-19 14:53:51 -04001149static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001150{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001151 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001152 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001153 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001154 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001156 /*
1157 * We need to do this right in the beginning since
1158 * alloc_arraycache's are going to use this list.
1159 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001160 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001161 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001162 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001163 if (err < 0)
1164 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001165
1166 /*
1167 * Now we can go ahead with allocating the shared arrays and
1168 * array caches
1169 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001170 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001171 struct array_cache *shared = NULL;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001172 struct alien_cache **alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001173
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001174 if (cachep->shared) {
1175 shared = alloc_arraycache(node,
1176 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001177 0xbaadf00d, GFP_KERNEL);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001178 if (!shared)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001179 goto bad;
1180 }
1181 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001182 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001183 if (!alien) {
1184 kfree(shared);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001186 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07001188 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001189 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001190
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001191 spin_lock_irq(&n->list_lock);
1192 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193 /*
1194 * We are serialised from CPU_DEAD or
1195 * CPU_UP_CANCELLED by the cpucontrol lock
1196 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001197 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001198 shared = NULL;
1199 }
1200#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001201 if (!n->alien) {
1202 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001203 alien = NULL;
1204 }
1205#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001206 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 kfree(shared);
1208 free_alien_cache(alien);
1209 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001210
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001211 return 0;
1212bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001213 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001214 return -ENOMEM;
1215}
1216
Paul Gortmaker0db06282013-06-19 14:53:51 -04001217static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001218 unsigned long action, void *hcpu)
1219{
1220 long cpu = (long)hcpu;
1221 int err = 0;
1222
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001224 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001225 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001226 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001227 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001228 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 break;
1230 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001231 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 start_cpu_timer(cpu);
1233 break;
1234#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001235 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001236 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001237 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001238 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001239 * held so that if cache_reap() is invoked it cannot do
1240 * anything expensive but will only modify reap_work
1241 * and reschedule the timer.
1242 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001243 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001244 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001245 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001246 break;
1247 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001248 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001249 start_cpu_timer(cpu);
1250 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001252 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001253 /*
1254 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001255 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001256 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001257 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001258 * structure is usually allocated from kmem_cache_create() and
1259 * gets destroyed at kmem_cache_destroy().
1260 */
Simon Arlott183ff222007-10-20 01:27:18 +02001261 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001262#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001264 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001265 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001266 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001267 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001270 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271}
1272
Paul Gortmaker0db06282013-06-19 14:53:51 -04001273static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001274 &cpuup_callback, NULL, 0
1275};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276
David Rientjes8f9f8d92010-03-27 19:40:47 -07001277#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1278/*
1279 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1280 * Returns -EBUSY if all objects cannot be drained so that the node is not
1281 * removed.
1282 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001283 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001284 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001285static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001286{
1287 struct kmem_cache *cachep;
1288 int ret = 0;
1289
Christoph Lameter18004c52012-07-06 15:25:12 -05001290 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001291 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001292
Christoph Lameter18bf8542014-08-06 16:04:11 -07001293 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001294 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001295 continue;
1296
Wanpeng Li0fa81032013-07-04 08:33:22 +08001297 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001298
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001299 if (!list_empty(&n->slabs_full) ||
1300 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001301 ret = -EBUSY;
1302 break;
1303 }
1304 }
1305 return ret;
1306}
1307
1308static int __meminit slab_memory_callback(struct notifier_block *self,
1309 unsigned long action, void *arg)
1310{
1311 struct memory_notify *mnb = arg;
1312 int ret = 0;
1313 int nid;
1314
1315 nid = mnb->status_change_nid;
1316 if (nid < 0)
1317 goto out;
1318
1319 switch (action) {
1320 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001321 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001322 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001323 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001324 break;
1325 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001326 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001327 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001328 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001329 break;
1330 case MEM_ONLINE:
1331 case MEM_OFFLINE:
1332 case MEM_CANCEL_ONLINE:
1333 case MEM_CANCEL_OFFLINE:
1334 break;
1335 }
1336out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001337 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001338}
1339#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1340
Christoph Lametere498be72005-09-09 13:03:32 -07001341/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001342 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001343 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001344static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001345 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001346{
Christoph Lameter6744f082013-01-10 19:12:17 +00001347 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001348
Christoph Lameter6744f082013-01-10 19:12:17 +00001349 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001350 BUG_ON(!ptr);
1351
Christoph Lameter6744f082013-01-10 19:12:17 +00001352 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001353 /*
1354 * Do not assume that spinlocks can be initialized via memcpy:
1355 */
1356 spin_lock_init(&ptr->list_lock);
1357
Christoph Lametere498be72005-09-09 13:03:32 -07001358 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001359 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001360}
1361
Andrew Mortona737b3e2006-03-22 00:08:11 -08001362/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001363 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1364 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001365 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001366static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001367{
1368 int node;
1369
1370 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001371 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001372 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001373 REAPTIMEOUT_NODE +
1374 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001375 }
1376}
1377
1378/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001379 * Initialisation. Called after the page allocator have been initialised and
1380 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 */
1382void __init kmem_cache_init(void)
1383{
Christoph Lametere498be72005-09-09 13:03:32 -07001384 int i;
1385
Joonsoo Kim68126702013-10-24 10:07:42 +09001386 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1387 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001388 kmem_cache = &kmem_cache_boot;
1389
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001390 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001391 use_alien_caches = 0;
1392
Christoph Lameter3c583462012-11-28 16:23:01 +00001393 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001394 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001395
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 /*
1397 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001398 * page orders on machines with more than 32MB of memory if
1399 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001401 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001402 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 /* Bootstrap is tricky, because several objects are allocated
1405 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001406 * 1) initialize the kmem_cache cache: it contains the struct
1407 * kmem_cache structures of all caches, except kmem_cache itself:
1408 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001409 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001410 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001411 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001413 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001414 * An __init data area is used for the head array.
1415 * 3) Create the remaining kmalloc caches, with minimally sized
1416 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001417 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001419 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001420 * the other cache's with kmalloc allocated memory.
1421 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 */
1423
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001424 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Eric Dumazet8da34302007-05-06 14:49:29 -07001426 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001427 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001428 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001429 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001430 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001431 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001432 SLAB_HWCACHE_ALIGN);
1433 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001434 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435
Andrew Mortona737b3e2006-03-22 00:08:11 -08001436 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001437 * Initialize the caches that provide memory for the kmem_cache_node
1438 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001439 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001440 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001441 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001442 slab_state = PARTIAL_NODE;
Christoph Lametere498be72005-09-09 13:03:32 -07001443
Ingo Molnare0a42722006-06-23 02:03:46 -07001444 slab_early_init = 0;
1445
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001446 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001447 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001448 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
Mel Gorman9c09a952008-01-24 05:49:54 -08001450 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001451 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001452
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001453 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001454 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001455 }
1456 }
1457
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001458 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001459}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001460
Pekka Enberg8429db52009-06-12 15:58:59 +03001461void __init kmem_cache_init_late(void)
1462{
1463 struct kmem_cache *cachep;
1464
Christoph Lameter97d06602012-07-06 15:25:11 -05001465 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001466
Pekka Enberg8429db52009-06-12 15:58:59 +03001467 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001468 mutex_lock(&slab_mutex);
1469 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001470 if (enable_cpucache(cachep, GFP_NOWAIT))
1471 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001472 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001473
Christoph Lameter97d06602012-07-06 15:25:11 -05001474 /* Done! */
1475 slab_state = FULL;
1476
Andrew Mortona737b3e2006-03-22 00:08:11 -08001477 /*
1478 * Register a cpu startup notifier callback that initializes
1479 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 */
1481 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
David Rientjes8f9f8d92010-03-27 19:40:47 -07001483#ifdef CONFIG_NUMA
1484 /*
1485 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001486 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001487 */
1488 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1489#endif
1490
Andrew Mortona737b3e2006-03-22 00:08:11 -08001491 /*
1492 * The reap timers are started later, with a module init call: That part
1493 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 */
1495}
1496
1497static int __init cpucache_init(void)
1498{
1499 int cpu;
1500
Andrew Mortona737b3e2006-03-22 00:08:11 -08001501 /*
1502 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 */
Christoph Lametere498be72005-09-09 13:03:32 -07001504 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001505 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001506
1507 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001508 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 return 0;
1510}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511__initcall(cpucache_init);
1512
Rafael Aquini8bdec192012-03-09 17:27:27 -03001513static noinline void
1514slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1515{
David Rientjes9a02d692014-06-04 16:06:36 -07001516#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001517 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001518 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001519 unsigned long flags;
1520 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001521 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1522 DEFAULT_RATELIMIT_BURST);
1523
1524 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1525 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001526
1527 printk(KERN_WARNING
1528 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1529 nodeid, gfpflags);
1530 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001531 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001532
Christoph Lameter18bf8542014-08-06 16:04:11 -07001533 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001534 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1535 unsigned long active_slabs = 0, num_slabs = 0;
1536
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001537 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001538 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001539 active_objs += cachep->num;
1540 active_slabs++;
1541 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001542 list_for_each_entry(page, &n->slabs_partial, lru) {
1543 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001544 active_slabs++;
1545 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001546 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001547 num_slabs++;
1548
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001549 free_objects += n->free_objects;
1550 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001551
1552 num_slabs += active_slabs;
1553 num_objs = num_slabs * cachep->num;
1554 printk(KERN_WARNING
1555 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1556 node, active_slabs, num_slabs, active_objs, num_objs,
1557 free_objects);
1558 }
David Rientjes9a02d692014-06-04 16:06:36 -07001559#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001560}
1561
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001563 * Interface to system's page allocator. No need to hold the
1564 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 *
1566 * If we requested dmaable memory, we will get it. Even if we
1567 * did not request dmaable memory, we might get it, but that
1568 * would be relatively rare and ignorable.
1569 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001570static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1571 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572{
1573 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001574 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001575
Glauber Costaa618e892012-06-14 16:17:21 +04001576 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001577 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1578 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001579
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001580 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1581 return NULL;
1582
Linus Torvalds517d0862009-06-16 19:50:13 -07001583 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001584 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001585 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001586 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001588 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001590 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001591 if (unlikely(page->pfmemalloc))
1592 pfmemalloc_active = true;
1593
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001594 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001596 add_zone_page_state(page_zone(page),
1597 NR_SLAB_RECLAIMABLE, nr_pages);
1598 else
1599 add_zone_page_state(page_zone(page),
1600 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001601 __SetPageSlab(page);
1602 if (page->pfmemalloc)
1603 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001604
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001605 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1606 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1607
1608 if (cachep->ctor)
1609 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1610 else
1611 kmemcheck_mark_unallocated_pages(page, nr_pages);
1612 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001613
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001614 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615}
1616
1617/*
1618 * Interface to system's page release.
1619 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001620static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001622 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001624 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001625
Christoph Lameter972d1a72006-09-25 23:31:51 -07001626 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1627 sub_zone_page_state(page_zone(page),
1628 NR_SLAB_RECLAIMABLE, nr_freed);
1629 else
1630 sub_zone_page_state(page_zone(page),
1631 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001632
Joonsoo Kima57a4982013-10-24 10:07:44 +09001633 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001634 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001635 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001636 page_mapcount_reset(page);
1637 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001638
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 if (current->reclaim_state)
1640 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001641 __free_pages(page, cachep->gfporder);
1642 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643}
1644
1645static void kmem_rcu_free(struct rcu_head *head)
1646{
Joonsoo Kim68126702013-10-24 10:07:42 +09001647 struct kmem_cache *cachep;
1648 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Joonsoo Kim68126702013-10-24 10:07:42 +09001650 page = container_of(head, struct page, rcu_head);
1651 cachep = page->slab_cache;
1652
1653 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654}
1655
1656#if DEBUG
1657
1658#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001659static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001660 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001662 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001664 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001666 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 return;
1668
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001669 *addr++ = 0x12345678;
1670 *addr++ = caller;
1671 *addr++ = smp_processor_id();
1672 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 {
1674 unsigned long *sptr = &caller;
1675 unsigned long svalue;
1676
1677 while (!kstack_end(sptr)) {
1678 svalue = *sptr++;
1679 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001680 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 size -= sizeof(unsigned long);
1682 if (size <= sizeof(unsigned long))
1683 break;
1684 }
1685 }
1686
1687 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001688 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689}
1690#endif
1691
Pekka Enberg343e0d72006-02-01 03:05:50 -08001692static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001694 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001695 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696
1697 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001698 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699}
1700
1701static void dump_line(char *data, int offset, int limit)
1702{
1703 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001704 unsigned char error = 0;
1705 int bad_count = 0;
1706
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001707 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001708 for (i = 0; i < limit; i++) {
1709 if (data[offset + i] != POISON_FREE) {
1710 error = data[offset + i];
1711 bad_count++;
1712 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001713 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001714 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1715 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001716
1717 if (bad_count == 1) {
1718 error ^= POISON_FREE;
1719 if (!(error & (error - 1))) {
1720 printk(KERN_ERR "Single bit error detected. Probably "
1721 "bad RAM.\n");
1722#ifdef CONFIG_X86
1723 printk(KERN_ERR "Run memtest86+ or a similar memory "
1724 "test tool.\n");
1725#else
1726 printk(KERN_ERR "Run a memory test tool.\n");
1727#endif
1728 }
1729 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730}
1731#endif
1732
1733#if DEBUG
1734
Pekka Enberg343e0d72006-02-01 03:05:50 -08001735static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736{
1737 int i, size;
1738 char *realobj;
1739
1740 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001741 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001742 *dbg_redzone1(cachep, objp),
1743 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 }
1745
1746 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001747 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1748 *dbg_userword(cachep, objp),
1749 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001751 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001752 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001753 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 int limit;
1755 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001756 if (i + limit > size)
1757 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 dump_line(realobj, i, limit);
1759 }
1760}
1761
Pekka Enberg343e0d72006-02-01 03:05:50 -08001762static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763{
1764 char *realobj;
1765 int size, i;
1766 int lines = 0;
1767
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001768 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001769 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 exp = POISON_END;
1775 if (realobj[i] != exp) {
1776 int limit;
1777 /* Mismatch ! */
1778 /* Print header */
1779 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001780 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001781 "Slab corruption (%s): %s start=%p, len=%d\n",
1782 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 print_objinfo(cachep, objp, 0);
1784 }
1785 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001786 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001788 if (i + limit > size)
1789 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 dump_line(realobj, i, limit);
1791 i += 16;
1792 lines++;
1793 /* Limit to 5 lines */
1794 if (lines > 5)
1795 break;
1796 }
1797 }
1798 if (lines != 0) {
1799 /* Print some data about the neighboring objects, if they
1800 * exist:
1801 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001802 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001803 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804
Joonsoo Kim8456a642013-10-24 10:07:49 +09001805 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001807 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001808 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001810 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 print_objinfo(cachep, objp, 2);
1812 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001813 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001814 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001815 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001817 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 print_objinfo(cachep, objp, 2);
1819 }
1820 }
1821}
1822#endif
1823
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001825static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1826 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001827{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 int i;
1829 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001830 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831
1832 if (cachep->flags & SLAB_POISON) {
1833#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001834 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001835 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001836 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001837 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 else
1839 check_poison_obj(cachep, objp);
1840#else
1841 check_poison_obj(cachep, objp);
1842#endif
1843 }
1844 if (cachep->flags & SLAB_RED_ZONE) {
1845 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1846 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001847 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1849 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001850 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001853}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001855static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1856 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001857{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001858}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859#endif
1860
Randy Dunlap911851e2006-03-22 00:08:14 -08001861/**
1862 * slab_destroy - destroy and release all objects in a slab
1863 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001864 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001865 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001866 * Destroy all the objs in a slab page, and release the mem back to the system.
1867 * Before calling the slab page must have been unlinked from the cache. The
1868 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001869 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001870static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001871{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001872 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873
Joonsoo Kim8456a642013-10-24 10:07:49 +09001874 freelist = page->freelist;
1875 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001877 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878
Joonsoo Kim68126702013-10-24 10:07:42 +09001879 /*
1880 * RCU free overloads the RCU head over the LRU.
1881 * slab_page has been overloeaded over the LRU,
1882 * however it is not used from now on so that
1883 * we can use it safely.
1884 */
1885 head = (void *)&page->rcu_head;
1886 call_rcu(head, kmem_rcu_free);
1887
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001889 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001891
1892 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001893 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001894 * although actual page can be freed in rcu context
1895 */
1896 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001897 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898}
1899
Joonsoo Kim97654df2014-08-06 16:04:25 -07001900static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1901{
1902 struct page *page, *n;
1903
1904 list_for_each_entry_safe(page, n, list, lru) {
1905 list_del(&page->lru);
1906 slab_destroy(cachep, page);
1907 }
1908}
1909
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001911 * calculate_slab_order - calculate size (page order) of slabs
1912 * @cachep: pointer to the cache that is being created
1913 * @size: size of objects to be created in this cache.
1914 * @align: required alignment for the objects.
1915 * @flags: slab allocation flags
1916 *
1917 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001918 *
1919 * This could be made much more intelligent. For now, try to avoid using
1920 * high order pages for slabs. When the gfp() functions are more friendly
1921 * towards high-order requests, this should be changed.
1922 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001923static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001924 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001925{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001926 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001927 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001928 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001929
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001930 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001931 unsigned int num;
1932 size_t remainder;
1933
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001934 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001935 if (!num)
1936 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001937
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001938 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1939 if (num > SLAB_OBJ_MAX_NUM)
1940 break;
1941
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001942 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07001943 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001944 /*
1945 * Max number of objs-per-slab for caches which
1946 * use off-slab slabs. Needed to avoid a possible
1947 * looping condition in cache_grow().
1948 */
Joonsoo Kim03787302014-06-23 13:22:06 -07001949 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1950 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001951 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07001952 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001953
1954 if (num > offslab_limit)
1955 break;
1956 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001957
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001958 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001959 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001960 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001961 left_over = remainder;
1962
1963 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001964 * A VFS-reclaimable slab tends to have most allocations
1965 * as GFP_NOFS and we really don't want to have to be allocating
1966 * higher-order pages when we are unable to shrink dcache.
1967 */
1968 if (flags & SLAB_RECLAIM_ACCOUNT)
1969 break;
1970
1971 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001972 * Large number of objects is good, but very large slabs are
1973 * currently bad for the gfp()s.
1974 */
David Rientjes543585c2011-10-18 22:09:24 -07001975 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001976 break;
1977
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001978 /*
1979 * Acceptable internal fragmentation?
1980 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001981 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001982 break;
1983 }
1984 return left_over;
1985}
1986
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001987static struct array_cache __percpu *alloc_kmem_cache_cpus(
1988 struct kmem_cache *cachep, int entries, int batchcount)
1989{
1990 int cpu;
1991 size_t size;
1992 struct array_cache __percpu *cpu_cache;
1993
1994 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001995 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001996
1997 if (!cpu_cache)
1998 return NULL;
1999
2000 for_each_possible_cpu(cpu) {
2001 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2002 entries, batchcount);
2003 }
2004
2005 return cpu_cache;
2006}
2007
Pekka Enberg83b519e2009-06-10 19:40:04 +03002008static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002009{
Christoph Lameter97d06602012-07-06 15:25:11 -05002010 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002011 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002012
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002013 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2014 if (!cachep->cpu_cache)
2015 return 1;
2016
Christoph Lameter97d06602012-07-06 15:25:11 -05002017 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002018 /* Creation of first cache (kmem_cache). */
2019 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002020 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002021 /* For kmem_cache_node */
2022 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002023 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002024 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002025
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002026 for_each_online_node(node) {
2027 cachep->node[node] = kmalloc_node(
2028 sizeof(struct kmem_cache_node), gfp, node);
2029 BUG_ON(!cachep->node[node]);
2030 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002031 }
2032 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002033
Christoph Lameter6a673682013-01-10 19:14:19 +00002034 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002035 jiffies + REAPTIMEOUT_NODE +
2036 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002037
2038 cpu_cache_get(cachep)->avail = 0;
2039 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2040 cpu_cache_get(cachep)->batchcount = 1;
2041 cpu_cache_get(cachep)->touched = 0;
2042 cachep->batchcount = 1;
2043 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002044 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002045}
2046
Joonsoo Kim12220de2014-10-09 15:26:24 -07002047unsigned long kmem_cache_flags(unsigned long object_size,
2048 unsigned long flags, const char *name,
2049 void (*ctor)(void *))
2050{
2051 return flags;
2052}
2053
2054struct kmem_cache *
2055__kmem_cache_alias(const char *name, size_t size, size_t align,
2056 unsigned long flags, void (*ctor)(void *))
2057{
2058 struct kmem_cache *cachep;
2059
2060 cachep = find_mergeable(size, align, flags, name, ctor);
2061 if (cachep) {
2062 cachep->refcount++;
2063
2064 /*
2065 * Adjust the object sizes so that we clear
2066 * the complete object on kzalloc.
2067 */
2068 cachep->object_size = max_t(int, cachep->object_size, size);
2069 }
2070 return cachep;
2071}
2072
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002073/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002074 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002075 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 *
2078 * Returns a ptr to the cache on success, NULL on failure.
2079 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002080 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 * The flags are
2083 *
2084 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2085 * to catch references to uninitialised memory.
2086 *
2087 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2088 * for buffer overruns.
2089 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2091 * cacheline. This can be beneficial if you're counting cycles as closely
2092 * as davem.
2093 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002094int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002095__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002097 size_t left_over, freelist_size;
2098 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002099 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002100 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002101 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104#if FORCED_DEBUG
2105 /*
2106 * Enable redzoning and last user accounting, except for caches with
2107 * large objects, if the increased size would increase the object size
2108 * above the next power of two: caches with object sizes just above a
2109 * power of two have a significant amount of internal fragmentation.
2110 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002111 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2112 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002113 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 if (!(flags & SLAB_DESTROY_BY_RCU))
2115 flags |= SLAB_POISON;
2116#endif
2117 if (flags & SLAB_DESTROY_BY_RCU)
2118 BUG_ON(flags & SLAB_POISON);
2119#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Andrew Mortona737b3e2006-03-22 00:08:11 -08002121 /*
2122 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 * unaligned accesses for some archs when redzoning is used, and makes
2124 * sure any on-slab bufctl's are also correctly aligned.
2125 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002126 if (size & (BYTES_PER_WORD - 1)) {
2127 size += (BYTES_PER_WORD - 1);
2128 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 }
2130
David Woodhouse87a927c2007-07-04 21:26:44 -04002131 if (flags & SLAB_RED_ZONE) {
2132 ralign = REDZONE_ALIGN;
2133 /* If redzoning, ensure that the second redzone is suitably
2134 * aligned, by adjusting the object size accordingly. */
2135 size += REDZONE_ALIGN - 1;
2136 size &= ~(REDZONE_ALIGN - 1);
2137 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002138
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002139 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002140 if (ralign < cachep->align) {
2141 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002143 /* disable debug if necessary */
2144 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002145 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002146 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002147 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002149 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150
Pekka Enberg83b519e2009-06-10 19:40:04 +03002151 if (slab_is_available())
2152 gfp = GFP_KERNEL;
2153 else
2154 gfp = GFP_NOWAIT;
2155
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157
Pekka Enbergca5f9702006-09-25 23:31:25 -07002158 /*
2159 * Both debugging options require word-alignment which is calculated
2160 * into align above.
2161 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002164 cachep->obj_offset += sizeof(unsigned long long);
2165 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 }
2167 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002168 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002169 * the real object. But if the second red zone needs to be
2170 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002172 if (flags & SLAB_RED_ZONE)
2173 size += REDZONE_ALIGN;
2174 else
2175 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 }
2177#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002178 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002179 && cachep->object_size > cache_line_size()
2180 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2181 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 size = PAGE_SIZE;
2183 }
2184#endif
2185#endif
2186
Ingo Molnare0a42722006-06-23 02:03:46 -07002187 /*
2188 * Determine if the slab management is 'on' or 'off' slab.
2189 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002190 * it too early on. Always use on-slab management when
2191 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002192 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002193 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002194 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 /*
2196 * Size is large, assume best to place the slab management obj
2197 * off-slab (should allow better packing of objs).
2198 */
2199 flags |= CFLGS_OFF_SLAB;
2200
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002201 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002202 /*
2203 * We should restrict the number of objects in a slab to implement
2204 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2205 */
2206 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2207 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002209 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002211 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002212 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002213
Joonsoo Kim03787302014-06-23 13:22:06 -07002214 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215
2216 /*
2217 * If the slab has been placed off-slab, and we have enough space then
2218 * move it on-slab. This is at the expense of any extra colouring.
2219 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002220 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002222 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 }
2224
2225 if (flags & CFLGS_OFF_SLAB) {
2226 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002227 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302228
2229#ifdef CONFIG_PAGE_POISONING
2230 /* If we're going to use the generic kernel_map_pages()
2231 * poisoning, then it's going to smash the contents of
2232 * the redzone and userword anyhow, so switch them off.
2233 */
2234 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2235 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2236#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 }
2238
2239 cachep->colour_off = cache_line_size();
2240 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002241 if (cachep->colour_off < cachep->align)
2242 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002243 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002244 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002246 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002247 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002248 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002249 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002250 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002252 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002253 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002254 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002255 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002256 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002257 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2258 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002259 * But leave a BUG_ON for some lucky dude.
2260 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002261 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002262 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002264 err = setup_cpu_cache(cachep, gfp);
2265 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002266 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002267 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002268 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002270 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
2273#if DEBUG
2274static void check_irq_off(void)
2275{
2276 BUG_ON(!irqs_disabled());
2277}
2278
2279static void check_irq_on(void)
2280{
2281 BUG_ON(irqs_disabled());
2282}
2283
Pekka Enberg343e0d72006-02-01 03:05:50 -08002284static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285{
2286#ifdef CONFIG_SMP
2287 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002288 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289#endif
2290}
Christoph Lametere498be72005-09-09 13:03:32 -07002291
Pekka Enberg343e0d72006-02-01 03:05:50 -08002292static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002293{
2294#ifdef CONFIG_SMP
2295 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002296 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002297#endif
2298}
2299
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300#else
2301#define check_irq_off() do { } while(0)
2302#define check_irq_on() do { } while(0)
2303#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002304#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305#endif
2306
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002307static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002308 struct array_cache *ac,
2309 int force, int node);
2310
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311static void do_drain(void *arg)
2312{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002313 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002315 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002316 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002317 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
2319 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002320 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002321 n = get_node(cachep, node);
2322 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002323 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002324 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002325 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 ac->avail = 0;
2327}
2328
Pekka Enberg343e0d72006-02-01 03:05:50 -08002329static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002331 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002332 int node;
2333
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002334 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002336 for_each_kmem_cache_node(cachep, node, n)
2337 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002338 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002339
Christoph Lameter18bf8542014-08-06 16:04:11 -07002340 for_each_kmem_cache_node(cachep, node, n)
2341 drain_array(cachep, n, n->shared, 1, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342}
2343
Christoph Lametered11d9e2006-06-30 01:55:45 -07002344/*
2345 * Remove slabs from the list of free slabs.
2346 * Specify the number of slabs to drain in tofree.
2347 *
2348 * Returns the actual number of slabs released.
2349 */
2350static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002351 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002353 struct list_head *p;
2354 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002355 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356
Christoph Lametered11d9e2006-06-30 01:55:45 -07002357 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002358 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002360 spin_lock_irq(&n->list_lock);
2361 p = n->slabs_free.prev;
2362 if (p == &n->slabs_free) {
2363 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002364 goto out;
2365 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366
Joonsoo Kim8456a642013-10-24 10:07:49 +09002367 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002369 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002371 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002372 /*
2373 * Safe to drop the lock. The slab is no longer linked
2374 * to the cache.
2375 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002376 n->free_objects -= cache->num;
2377 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002378 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002379 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002381out:
2382 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383}
2384
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002385int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002386{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002387 int ret = 0;
2388 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002389 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002390
2391 drain_cpu_caches(cachep);
2392
2393 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002394 for_each_kmem_cache_node(cachep, node, n) {
Wanpeng Li0fa81032013-07-04 08:33:22 +08002395 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002396
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002397 ret += !list_empty(&n->slabs_full) ||
2398 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002399 }
2400 return (ret ? 1 : 0);
2401}
2402
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002403int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404{
Christoph Lameter12c36672012-09-04 23:38:33 +00002405 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002406 struct kmem_cache_node *n;
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002407 int rc = __kmem_cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408
Christoph Lameter12c36672012-09-04 23:38:33 +00002409 if (rc)
2410 return rc;
2411
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002412 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002413
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002414 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002415 for_each_kmem_cache_node(cachep, i, n) {
2416 kfree(n->shared);
2417 free_alien_cache(n->alien);
2418 kfree(n);
2419 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002421 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002424/*
2425 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002426 *
2427 * For a slab cache when the slab descriptor is off-slab, the
2428 * slab descriptor can't come from the same cache which is being created,
2429 * Because if it is the case, that means we defer the creation of
2430 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2431 * And we eventually call down to __kmem_cache_create(), which
2432 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2433 * This is a "chicken-and-egg" problem.
2434 *
2435 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2436 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002437 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002438static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002439 struct page *page, int colour_off,
2440 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002442 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002443 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 if (OFF_SLAB(cachep)) {
2446 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002447 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002448 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002449 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 return NULL;
2451 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002452 freelist = addr + colour_off;
2453 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002455 page->active = 0;
2456 page->s_mem = addr + colour_off;
2457 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458}
2459
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002460static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002462 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002463}
2464
2465static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002466 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002467{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002468 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
Pekka Enberg343e0d72006-02-01 03:05:50 -08002471static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002472 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473{
2474 int i;
2475
2476 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002477 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478#if DEBUG
2479 /* need to poison the objs? */
2480 if (cachep->flags & SLAB_POISON)
2481 poison_obj(cachep, objp, POISON_FREE);
2482 if (cachep->flags & SLAB_STORE_USER)
2483 *dbg_userword(cachep, objp) = NULL;
2484
2485 if (cachep->flags & SLAB_RED_ZONE) {
2486 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2487 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2488 }
2489 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002490 * Constructors are not allowed to allocate memory from the same
2491 * cache which they are a constructor for. Otherwise, deadlock.
2492 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 */
2494 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002495 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496
2497 if (cachep->flags & SLAB_RED_ZONE) {
2498 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2499 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002500 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2502 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002503 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002505 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002506 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002507 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002508 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509#else
2510 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002511 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002513 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002514 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516}
2517
Pekka Enberg343e0d72006-02-01 03:05:50 -08002518static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002520 if (CONFIG_ZONE_DMA_FLAG) {
2521 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002522 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002523 else
Glauber Costaa618e892012-06-14 16:17:21 +04002524 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002525 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526}
2527
Joonsoo Kim8456a642013-10-24 10:07:49 +09002528static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002529 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002530{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002531 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002532
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002533 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002534 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002535#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002536 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002537#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002538
2539 return objp;
2540}
2541
Joonsoo Kim8456a642013-10-24 10:07:49 +09002542static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002543 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002544{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002545 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002546#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002547 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002548
Matthew Dobson78d382d2006-02-01 03:05:47 -08002549 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002550 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002551
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002552 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002553 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002554 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002555 printk(KERN_ERR "slab: double free detected in cache "
2556 "'%s', objp %p\n", cachep->name, objp);
2557 BUG();
2558 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002559 }
2560#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002561 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002562 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002563}
2564
Pekka Enberg47768742006-06-23 02:03:07 -07002565/*
2566 * Map pages beginning at addr to the given cache and slab. This is required
2567 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002568 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002569 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002570static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002571 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002573 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002574 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575}
2576
2577/*
2578 * Grow (by 1) the number of slabs within a cache. This is called by
2579 * kmem_cache_alloc() when there are no active objs left in a cache.
2580 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002581static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002582 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002584 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002585 size_t offset;
2586 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002587 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Andrew Mortona737b3e2006-03-22 00:08:11 -08002589 /*
2590 * Be lazy and only check for valid flags here, keeping it out of the
2591 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002593 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2594 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2595 BUG();
2596 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002597 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002599 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002601 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002602 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603
2604 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002605 offset = n->colour_next;
2606 n->colour_next++;
2607 if (n->colour_next >= cachep->colour)
2608 n->colour_next = 0;
2609 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002611 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612
2613 if (local_flags & __GFP_WAIT)
2614 local_irq_enable();
2615
2616 /*
2617 * The test for missing atomic flag is performed here, rather than
2618 * the more obvious place, simply to reduce the critical path length
2619 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2620 * will eventually be caught here (where it matters).
2621 */
2622 kmem_flagcheck(cachep, flags);
2623
Andrew Mortona737b3e2006-03-22 00:08:11 -08002624 /*
2625 * Get mem for the objs. Attempt to allocate a physical page from
2626 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002627 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002628 if (!page)
2629 page = kmem_getpages(cachep, local_flags, nodeid);
2630 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 goto failed;
2632
2633 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002634 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002635 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002636 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 goto opps1;
2638
Joonsoo Kim8456a642013-10-24 10:07:49 +09002639 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640
Joonsoo Kim8456a642013-10-24 10:07:49 +09002641 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642
2643 if (local_flags & __GFP_WAIT)
2644 local_irq_disable();
2645 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002646 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647
2648 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002649 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002651 n->free_objects += cachep->num;
2652 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002654opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002655 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002656failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 if (local_flags & __GFP_WAIT)
2658 local_irq_disable();
2659 return 0;
2660}
2661
2662#if DEBUG
2663
2664/*
2665 * Perform extra freeing checks:
2666 * - detect bad pointers.
2667 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 */
2669static void kfree_debugcheck(const void *objp)
2670{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 if (!virt_addr_valid(objp)) {
2672 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 (unsigned long)objp);
2674 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676}
2677
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002678static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2679{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002680 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002681
2682 redzone1 = *dbg_redzone1(cache, obj);
2683 redzone2 = *dbg_redzone2(cache, obj);
2684
2685 /*
2686 * Redzone is ok.
2687 */
2688 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2689 return;
2690
2691 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2692 slab_error(cache, "double free detected");
2693 else
2694 slab_error(cache, "memory outside object was overwritten");
2695
David Woodhouseb46b8f12007-05-08 00:22:59 -07002696 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002697 obj, redzone1, redzone2);
2698}
2699
Pekka Enberg343e0d72006-02-01 03:05:50 -08002700static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002701 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002704 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002706 BUG_ON(virt_to_cache(objp) != cachep);
2707
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002708 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002710 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002713 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2715 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2716 }
2717 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002718 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719
Joonsoo Kim8456a642013-10-24 10:07:49 +09002720 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
2722 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002723 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724
Joonsoo Kim03787302014-06-23 13:22:06 -07002725 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 if (cachep->flags & SLAB_POISON) {
2727#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002728 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002729 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002730 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002731 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 } else {
2733 poison_obj(cachep, objp, POISON_FREE);
2734 }
2735#else
2736 poison_obj(cachep, objp, POISON_FREE);
2737#endif
2738 }
2739 return objp;
2740}
2741
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742#else
2743#define kfree_debugcheck(x) do { } while(0)
2744#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745#endif
2746
Mel Gorman072bb0a2012-07-31 16:43:58 -07002747static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2748 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749{
2750 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002751 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002753 int node;
2754
Joe Korty6d2144d2008-03-05 15:04:59 -08002755 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002756 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002757 if (unlikely(force_refill))
2758 goto force_grow;
2759retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002760 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 batchcount = ac->batchcount;
2762 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002763 /*
2764 * If there was little recent activity on this cache, then
2765 * perform only a partial refill. Otherwise we could generate
2766 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 */
2768 batchcount = BATCHREFILL_LIMIT;
2769 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07002770 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002772 BUG_ON(ac->avail > 0 || !n);
2773 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002774
Christoph Lameter3ded1752006-03-25 03:06:44 -08002775 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002776 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2777 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002778 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002779 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002780
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 while (batchcount > 0) {
2782 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002783 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002785 entry = n->slabs_partial.next;
2786 if (entry == &n->slabs_partial) {
2787 n->free_touched = 1;
2788 entry = n->slabs_free.next;
2789 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 goto must_grow;
2791 }
2792
Joonsoo Kim8456a642013-10-24 10:07:49 +09002793 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002795
2796 /*
2797 * The slab was either on partial or free list so
2798 * there must be at least one object available for
2799 * allocation.
2800 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002801 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002802
Joonsoo Kim8456a642013-10-24 10:07:49 +09002803 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 STATS_INC_ALLOCED(cachep);
2805 STATS_INC_ACTIVE(cachep);
2806 STATS_SET_HIGH(cachep);
2807
Joonsoo Kim8456a642013-10-24 10:07:49 +09002808 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002809 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811
2812 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002813 list_del(&page->lru);
2814 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002815 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002817 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 }
2819
Andrew Mortona737b3e2006-03-22 00:08:11 -08002820must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002821 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002822alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002823 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
2825 if (unlikely(!ac->avail)) {
2826 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002827force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002828 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002829
Andrew Mortona737b3e2006-03-22 00:08:11 -08002830 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002831 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002832 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002833
2834 /* no objects in sight? abort */
2835 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 return NULL;
2837
Andrew Mortona737b3e2006-03-22 00:08:11 -08002838 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 goto retry;
2840 }
2841 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002842
2843 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844}
2845
Andrew Mortona737b3e2006-03-22 00:08:11 -08002846static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2847 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848{
2849 might_sleep_if(flags & __GFP_WAIT);
2850#if DEBUG
2851 kmem_flagcheck(cachep, flags);
2852#endif
2853}
2854
2855#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002856static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002857 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858{
Joonsoo Kim03787302014-06-23 13:22:06 -07002859 struct page *page;
2860
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002861 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002863 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002865 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002866 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002867 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 else
2869 check_poison_obj(cachep, objp);
2870#else
2871 check_poison_obj(cachep, objp);
2872#endif
2873 poison_obj(cachep, objp, POISON_INUSE);
2874 }
2875 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002876 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
2878 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002879 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2880 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2881 slab_error(cachep, "double free, or memory outside"
2882 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002884 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002885 objp, *dbg_redzone1(cachep, objp),
2886 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 }
2888 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2889 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2890 }
Joonsoo Kim03787302014-06-23 13:22:06 -07002891
2892 page = virt_to_head_page(objp);
2893 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002894 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002895 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002896 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002897 if (ARCH_SLAB_MINALIGN &&
2898 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002899 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002900 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002901 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 return objp;
2903}
2904#else
2905#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2906#endif
2907
Akinobu Mita773ff602008-12-23 19:37:01 +09002908static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002909{
Joonsoo Kim8a9c61d2014-08-06 16:04:20 -07002910 if (unlikely(cachep == kmem_cache))
Akinobu Mita773ff602008-12-23 19:37:01 +09002911 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002912
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002913 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002914}
2915
Pekka Enberg343e0d72006-02-01 03:05:50 -08002916static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002920 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921
Alok N Kataria5c382302005-09-27 21:45:46 -07002922 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002923
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002924 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002927 objp = ac_get_obj(cachep, ac, flags, false);
2928
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002929 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002930 * Allow for the possibility all avail objects are not allowed
2931 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002932 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07002933 if (objp) {
2934 STATS_INC_ALLOCHIT(cachep);
2935 goto out;
2936 }
2937 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07002939
2940 STATS_INC_ALLOCMISS(cachep);
2941 objp = cache_alloc_refill(cachep, flags, force_refill);
2942 /*
2943 * the 'ac' may be updated by cache_alloc_refill(),
2944 * and kmemleak_erase() requires its correct value.
2945 */
2946 ac = cpu_cache_get(cachep);
2947
2948out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01002949 /*
2950 * To avoid a false negative, if an object that is in one of the
2951 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2952 * treat the array pointers as a reference to the object.
2953 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09002954 if (objp)
2955 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07002956 return objp;
2957}
2958
Christoph Lametere498be72005-09-09 13:03:32 -07002959#ifdef CONFIG_NUMA
2960/*
Zefan Li2ad654b2014-09-25 09:41:02 +08002961 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002962 *
2963 * If we are in_interrupt, then process context, including cpusets and
2964 * mempolicy, may not apply and should not be used for allocation policy.
2965 */
2966static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2967{
2968 int nid_alloc, nid_here;
2969
Christoph Lameter765c4502006-09-27 01:50:08 -07002970 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08002971 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002972 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002973 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07002974 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002975 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07002976 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002977 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08002978 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08002979 return NULL;
2980}
2981
2982/*
Christoph Lameter765c4502006-09-27 01:50:08 -07002983 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08002984 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00002985 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08002986 * perform an allocation without specifying a node. This allows the page
2987 * allocator to do its reclaim / fallback magic. We then insert the
2988 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07002989 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08002990static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07002991{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08002992 struct zonelist *zonelist;
2993 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07002994 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002995 struct zone *zone;
2996 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07002997 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08002998 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07002999 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003000
3001 if (flags & __GFP_THISNODE)
3002 return NULL;
3003
Christoph Lameter6cb06222007-10-16 01:25:41 -07003004 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003005
Mel Gormancc9a6c82012-03-21 16:34:11 -07003006retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003007 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003008 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003009
Christoph Lameter3c517a62006-12-06 20:33:29 -08003010retry:
3011 /*
3012 * Look through allowed nodes for objects available
3013 * from existing per node queues.
3014 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003015 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3016 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003017
Vladimir Davydov061d7072014-12-12 16:58:25 -08003018 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003019 get_node(cache, nid) &&
3020 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003021 obj = ____cache_alloc_node(cache,
3022 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003023 if (obj)
3024 break;
3025 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003026 }
3027
Christoph Lametercfce6602007-05-06 14:50:17 -07003028 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003029 /*
3030 * This allocation will be performed within the constraints
3031 * of the current cpuset / memory policy requirements.
3032 * We may trigger various forms of reclaim on the allowed
3033 * set and go into memory reserves if necessary.
3034 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003035 struct page *page;
3036
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003037 if (local_flags & __GFP_WAIT)
3038 local_irq_enable();
3039 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003040 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003041 if (local_flags & __GFP_WAIT)
3042 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003043 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003044 /*
3045 * Insert into the appropriate per node queues
3046 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003047 nid = page_to_nid(page);
3048 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003049 obj = ____cache_alloc_node(cache,
3050 flags | GFP_THISNODE, nid);
3051 if (!obj)
3052 /*
3053 * Another processor may allocate the
3054 * objects in the slab since we are
3055 * not holding any locks.
3056 */
3057 goto retry;
3058 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003059 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003060 obj = NULL;
3061 }
3062 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003063 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003064
Mel Gormand26914d2014-04-03 14:47:24 -07003065 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003066 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003067 return obj;
3068}
3069
3070/*
Christoph Lametere498be72005-09-09 13:03:32 -07003071 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003073static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003074 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003075{
3076 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003077 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003078 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003079 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003080 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003082 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003083 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003084 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003085
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003087 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003088 spin_lock(&n->list_lock);
3089 entry = n->slabs_partial.next;
3090 if (entry == &n->slabs_partial) {
3091 n->free_touched = 1;
3092 entry = n->slabs_free.next;
3093 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003094 goto must_grow;
3095 }
Christoph Lametere498be72005-09-09 13:03:32 -07003096
Joonsoo Kim8456a642013-10-24 10:07:49 +09003097 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003098 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003099
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003100 STATS_INC_NODEALLOCS(cachep);
3101 STATS_INC_ACTIVE(cachep);
3102 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003103
Joonsoo Kim8456a642013-10-24 10:07:49 +09003104 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003105
Joonsoo Kim8456a642013-10-24 10:07:49 +09003106 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003107 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003109 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003110
Joonsoo Kim8456a642013-10-24 10:07:49 +09003111 if (page->active == cachep->num)
3112 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003113 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003114 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003115
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003116 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003117 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003118
Andrew Mortona737b3e2006-03-22 00:08:11 -08003119must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003120 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003121 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003122 if (x)
3123 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003124
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003125 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003126
Andrew Mortona737b3e2006-03-22 00:08:11 -08003127done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003128 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003129}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003130
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003131static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003132slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003133 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003134{
3135 unsigned long save_flags;
3136 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003137 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003138
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003139 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003140
Nick Piggincf40bd12009-01-21 08:12:39 +01003141 lockdep_trace_alloc(flags);
3142
Akinobu Mita773ff602008-12-23 19:37:01 +09003143 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003144 return NULL;
3145
Glauber Costad79923f2012-12-18 14:22:48 -08003146 cachep = memcg_kmem_get_cache(cachep, flags);
3147
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003148 cache_alloc_debugcheck_before(cachep, flags);
3149 local_irq_save(save_flags);
3150
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003151 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003152 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003153
Christoph Lameter18bf8542014-08-06 16:04:11 -07003154 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003155 /* Node not bootstrapped yet */
3156 ptr = fallback_alloc(cachep, flags);
3157 goto out;
3158 }
3159
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003160 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003161 /*
3162 * Use the locally cached objects if possible.
3163 * However ____cache_alloc does not allow fallback
3164 * to other nodes. It may fail while we still have
3165 * objects on other nodes available.
3166 */
3167 ptr = ____cache_alloc(cachep, flags);
3168 if (ptr)
3169 goto out;
3170 }
3171 /* ___cache_alloc_node can fall back to other nodes */
3172 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3173 out:
3174 local_irq_restore(save_flags);
3175 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003176 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003177 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003178
Joe Perches5087c822013-09-10 17:02:51 -07003179 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003180 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003181 if (unlikely(flags & __GFP_ZERO))
3182 memset(ptr, 0, cachep->object_size);
3183 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003184
Vladimir Davydov8135be52014-12-12 16:56:38 -08003185 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003186 return ptr;
3187}
3188
3189static __always_inline void *
3190__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3191{
3192 void *objp;
3193
Zefan Li2ad654b2014-09-25 09:41:02 +08003194 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003195 objp = alternate_node_alloc(cache, flags);
3196 if (objp)
3197 goto out;
3198 }
3199 objp = ____cache_alloc(cache, flags);
3200
3201 /*
3202 * We may just have run out of memory on the local node.
3203 * ____cache_alloc_node() knows how to locate memory on other nodes
3204 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003205 if (!objp)
3206 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003207
3208 out:
3209 return objp;
3210}
3211#else
3212
3213static __always_inline void *
3214__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3215{
3216 return ____cache_alloc(cachep, flags);
3217}
3218
3219#endif /* CONFIG_NUMA */
3220
3221static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003222slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003223{
3224 unsigned long save_flags;
3225 void *objp;
3226
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003227 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003228
Nick Piggincf40bd12009-01-21 08:12:39 +01003229 lockdep_trace_alloc(flags);
3230
Akinobu Mita773ff602008-12-23 19:37:01 +09003231 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003232 return NULL;
3233
Glauber Costad79923f2012-12-18 14:22:48 -08003234 cachep = memcg_kmem_get_cache(cachep, flags);
3235
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003236 cache_alloc_debugcheck_before(cachep, flags);
3237 local_irq_save(save_flags);
3238 objp = __do_cache_alloc(cachep, flags);
3239 local_irq_restore(save_flags);
3240 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003241 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003242 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003243 prefetchw(objp);
3244
Joe Perches5087c822013-09-10 17:02:51 -07003245 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003246 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003247 if (unlikely(flags & __GFP_ZERO))
3248 memset(objp, 0, cachep->object_size);
3249 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003250
Vladimir Davydov8135be52014-12-12 16:56:38 -08003251 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003252 return objp;
3253}
Christoph Lametere498be72005-09-09 13:03:32 -07003254
3255/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003256 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003257 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003258 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003259static void free_block(struct kmem_cache *cachep, void **objpp,
3260 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261{
3262 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003263 struct kmem_cache_node *n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264
3265 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003266 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003267 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268
Mel Gorman072bb0a2012-07-31 16:43:58 -07003269 clear_obj_pfmemalloc(&objpp[i]);
3270 objp = objpp[i];
3271
Joonsoo Kim8456a642013-10-24 10:07:49 +09003272 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003273 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003274 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003275 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003277 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278
3279 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003280 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003281 if (n->free_objects > n->free_limit) {
3282 n->free_objects -= cachep->num;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003283 list_add_tail(&page->lru, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003285 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 }
3287 } else {
3288 /* Unconditionally move a slab to the end of the
3289 * partial list on free - maximum time for the
3290 * other objects to be freed, too.
3291 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003292 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 }
3294 }
3295}
3296
Pekka Enberg343e0d72006-02-01 03:05:50 -08003297static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298{
3299 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003300 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003301 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003302 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303
3304 batchcount = ac->batchcount;
3305#if DEBUG
3306 BUG_ON(!batchcount || batchcount > ac->avail);
3307#endif
3308 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003309 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003310 spin_lock(&n->list_lock);
3311 if (n->shared) {
3312 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 if (max) {
3315 if (batchcount > max)
3316 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003317 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 shared_array->avail += batchcount;
3320 goto free_done;
3321 }
3322 }
3323
Joonsoo Kim97654df2014-08-06 16:04:25 -07003324 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003325free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326#if STATS
3327 {
3328 int i = 0;
3329 struct list_head *p;
3330
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003331 p = n->slabs_free.next;
3332 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003333 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
Joonsoo Kim8456a642013-10-24 10:07:49 +09003335 page = list_entry(p, struct page, lru);
3336 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337
3338 i++;
3339 p = p->next;
3340 }
3341 STATS_SET_FREEABLE(cachep, i);
3342 }
3343#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003344 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003345 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003347 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348}
3349
3350/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003351 * Release an obj back to its cache. If the obj has a constructed state, it must
3352 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003354static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003355 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003357 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358
3359 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003360 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003361 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003363 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003364
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003365 /*
3366 * Skip calling cache_free_alien() when the platform is not numa.
3367 * This will avoid cache misses that happen while accessing slabp (which
3368 * is per page memory reference) to get nodeid. Instead use a global
3369 * variable to skip the call, which is mostly likely to be present in
3370 * the cache.
3371 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003372 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003373 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003374
Joonsoo Kim3d880192014-10-09 15:26:04 -07003375 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 } else {
3378 STATS_INC_FREEMISS(cachep);
3379 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003381
Mel Gorman072bb0a2012-07-31 16:43:58 -07003382 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383}
3384
3385/**
3386 * kmem_cache_alloc - Allocate an object
3387 * @cachep: The cache to allocate from.
3388 * @flags: See kmalloc().
3389 *
3390 * Allocate an object from this cache. The flags are only relevant
3391 * if the cache has no available objects.
3392 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003393void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003395 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003396
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003397 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003398 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003399
3400 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401}
3402EXPORT_SYMBOL(kmem_cache_alloc);
3403
Li Zefan0f24f122009-12-11 15:45:30 +08003404#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003405void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003406kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003407{
Steven Rostedt85beb582010-11-24 16:23:34 -05003408 void *ret;
3409
Ezequiel Garcia48356302012-09-08 17:47:57 -03003410 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003411
3412 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003413 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003414 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003415}
Steven Rostedt85beb582010-11-24 16:23:34 -05003416EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003417#endif
3418
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003420/**
3421 * kmem_cache_alloc_node - Allocate an object on the specified node
3422 * @cachep: The cache to allocate from.
3423 * @flags: See kmalloc().
3424 * @nodeid: node number of the target node.
3425 *
3426 * Identical to kmem_cache_alloc but it will allocate memory on the given
3427 * node, which can improve the performance for cpu bound structures.
3428 *
3429 * Fallback to other node is possible if __GFP_THISNODE is not set.
3430 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003431void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3432{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003433 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003434
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003435 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003436 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003437 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003438
3439 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003440}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441EXPORT_SYMBOL(kmem_cache_alloc_node);
3442
Li Zefan0f24f122009-12-11 15:45:30 +08003443#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003444void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003445 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003446 int nodeid,
3447 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003448{
Steven Rostedt85beb582010-11-24 16:23:34 -05003449 void *ret;
3450
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003451 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003452
Steven Rostedt85beb582010-11-24 16:23:34 -05003453 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003454 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003455 flags, nodeid);
3456 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003457}
Steven Rostedt85beb582010-11-24 16:23:34 -05003458EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003459#endif
3460
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003461static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003462__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003463{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003464 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003465
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003466 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003467 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3468 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003469 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003470}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003471
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003472void *__kmalloc_node(size_t size, gfp_t flags, int node)
3473{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003474 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003475}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003476EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003477
3478void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003479 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003480{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003481 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003482}
3483EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003484#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485
3486/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003487 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003489 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003490 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003492static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003493 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003495 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003496 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003498 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003499 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3500 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003501 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003502
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003503 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003504 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003505
3506 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003507}
3508
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003509void *__kmalloc(size_t size, gfp_t flags)
3510{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003511 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512}
3513EXPORT_SYMBOL(__kmalloc);
3514
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003515void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003516{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003517 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003518}
3519EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003520
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521/**
3522 * kmem_cache_free - Deallocate an object
3523 * @cachep: The cache the allocation was from.
3524 * @objp: The previously allocated object.
3525 *
3526 * Free an object which was previously allocated from this
3527 * cache.
3528 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003529void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530{
3531 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003532 cachep = cache_from_obj(cachep, objp);
3533 if (!cachep)
3534 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535
3536 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003537 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003538 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003539 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003540 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003542
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003543 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544}
3545EXPORT_SYMBOL(kmem_cache_free);
3546
3547/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 * kfree - free previously allocated memory
3549 * @objp: pointer returned by kmalloc.
3550 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003551 * If @objp is NULL, no operation is performed.
3552 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 * Don't free memory not originally allocated by kmalloc()
3554 * or you will run into trouble.
3555 */
3556void kfree(const void *objp)
3557{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003558 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 unsigned long flags;
3560
Pekka Enberg2121db72009-03-25 11:05:57 +02003561 trace_kfree(_RET_IP_, objp);
3562
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003563 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 return;
3565 local_irq_save(flags);
3566 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003567 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003568 debug_check_no_locks_freed(objp, c->object_size);
3569
3570 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003571 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 local_irq_restore(flags);
3573}
3574EXPORT_SYMBOL(kfree);
3575
Christoph Lametere498be72005-09-09 13:03:32 -07003576/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003577 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003578 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003579static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003580{
3581 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003582 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003583 struct array_cache *new_shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07003584 struct alien_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003585
Mel Gorman9c09a952008-01-24 05:49:54 -08003586 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003587
LQYMGTb455def2014-12-10 15:42:13 -08003588 if (use_alien_caches) {
3589 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3590 if (!new_alien)
3591 goto fail;
3592 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003593
Eric Dumazet63109842007-05-06 14:49:28 -07003594 new_shared = NULL;
3595 if (cachep->shared) {
3596 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003597 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003598 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003599 if (!new_shared) {
3600 free_alien_cache(new_alien);
3601 goto fail;
3602 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003603 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003604
Christoph Lameter18bf8542014-08-06 16:04:11 -07003605 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003606 if (n) {
3607 struct array_cache *shared = n->shared;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003608 LIST_HEAD(list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003609
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003610 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003611
Christoph Lametercafeb022006-03-25 03:06:46 -08003612 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003613 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07003614 shared->avail, node, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003615
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003616 n->shared = new_shared;
3617 if (!n->alien) {
3618 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003619 new_alien = NULL;
3620 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003621 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003622 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003623 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003624 slabs_destroy(cachep, &list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003625 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003626 free_alien_cache(new_alien);
3627 continue;
3628 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003629 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3630 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003631 free_alien_cache(new_alien);
3632 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003633 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003634 }
Christoph Lametere498be72005-09-09 13:03:32 -07003635
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003636 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003637 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3638 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003639 n->shared = new_shared;
3640 n->alien = new_alien;
3641 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003642 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003643 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003644 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003645 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003646
Andrew Mortona737b3e2006-03-22 00:08:11 -08003647fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003648 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003649 /* Cache is not active yet. Roll back what we did */
3650 node--;
3651 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003652 n = get_node(cachep, node);
3653 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003654 kfree(n->shared);
3655 free_alien_cache(n->alien);
3656 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003657 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003658 }
3659 node--;
3660 }
3661 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003662 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003663}
3664
Christoph Lameter18004c52012-07-06 15:25:12 -05003665/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003666static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003667 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003669 struct array_cache __percpu *cpu_cache, *prev;
3670 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003672 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3673 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003674 return -ENOMEM;
3675
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003676 prev = cachep->cpu_cache;
3677 cachep->cpu_cache = cpu_cache;
3678 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003679
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 cachep->batchcount = batchcount;
3682 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003683 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003685 if (!prev)
3686 goto alloc_node;
3687
3688 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003689 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003690 int node;
3691 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003692 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003693
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003694 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003695 n = get_node(cachep, node);
3696 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003697 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003698 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003699 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003701 free_percpu(prev);
3702
3703alloc_node:
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003704 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705}
3706
Glauber Costa943a4512012-12-18 14:23:03 -08003707static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3708 int batchcount, int shared, gfp_t gfp)
3709{
3710 int ret;
3711 struct kmem_cache *c = NULL;
3712 int i = 0;
3713
3714 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3715
3716 if (slab_state < FULL)
3717 return ret;
3718
3719 if ((ret < 0) || !is_root_cache(cachep))
3720 return ret;
3721
Glauber Costaebe945c2012-12-18 14:23:10 -08003722 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003723 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003724 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003725 if (c)
3726 /* return value determined by the parent cache only */
3727 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3728 }
3729
3730 return ret;
3731}
3732
Christoph Lameter18004c52012-07-06 15:25:12 -05003733/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003734static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735{
3736 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003737 int limit = 0;
3738 int shared = 0;
3739 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740
Glauber Costa943a4512012-12-18 14:23:03 -08003741 if (!is_root_cache(cachep)) {
3742 struct kmem_cache *root = memcg_root_cache(cachep);
3743 limit = root->limit;
3744 shared = root->shared;
3745 batchcount = root->batchcount;
3746 }
3747
3748 if (limit && shared && batchcount)
3749 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003750 /*
3751 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 * - create a LIFO ordering, i.e. return objects that are cache-warm
3753 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003754 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755 * bufctl chains: array operations are cheaper.
3756 * The numbers are guessed, we should auto-tune as described by
3757 * Bonwick.
3758 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003759 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003761 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003763 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003765 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 limit = 54;
3767 else
3768 limit = 120;
3769
Andrew Mortona737b3e2006-03-22 00:08:11 -08003770 /*
3771 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 * allocation behaviour: Most allocs on one cpu, most free operations
3773 * on another cpu. For these cases, an efficient object passing between
3774 * cpus is necessary. This is provided by a shared array. The array
3775 * replaces Bonwick's magazine layer.
3776 * On uniprocessor, it's functionally equivalent (but less efficient)
3777 * to a larger limit. Thus disabled by default.
3778 */
3779 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003780 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782
3783#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003784 /*
3785 * With debugging enabled, large batchcount lead to excessively long
3786 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 */
3788 if (limit > 32)
3789 limit = 32;
3790#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003791 batchcount = (limit + 1) / 2;
3792skip_setup:
3793 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 if (err)
3795 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003796 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003797 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798}
3799
Christoph Lameter1b552532006-03-22 00:09:07 -08003800/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003801 * Drain an array if it contains any elements taking the node lock only if
3802 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003803 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003804 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003805static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003806 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807{
Joonsoo Kim97654df2014-08-06 16:04:25 -07003808 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809 int tofree;
3810
Christoph Lameter1b552532006-03-22 00:09:07 -08003811 if (!ac || !ac->avail)
3812 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 if (ac->touched && !force) {
3814 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003815 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003816 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003817 if (ac->avail) {
3818 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3819 if (tofree > ac->avail)
3820 tofree = (ac->avail + 1) / 2;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003821 free_block(cachep, ac->entry, tofree, node, &list);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003822 ac->avail -= tofree;
3823 memmove(ac->entry, &(ac->entry[tofree]),
3824 sizeof(void *) * ac->avail);
3825 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003826 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003827 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828 }
3829}
3830
3831/**
3832 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003833 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834 *
3835 * Called from workqueue/eventd every few seconds.
3836 * Purpose:
3837 * - clear the per-cpu caches for this CPU.
3838 * - return freeable pages to the main free memory pool.
3839 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003840 * If we cannot acquire the cache chain mutex then just give up - we'll try
3841 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003843static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003845 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003846 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003847 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003848 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849
Christoph Lameter18004c52012-07-06 15:25:12 -05003850 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003852 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853
Christoph Lameter18004c52012-07-06 15:25:12 -05003854 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 check_irq_on();
3856
Christoph Lameter35386e32006-03-22 00:09:05 -08003857 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003858 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003859 * have established with reasonable certainty that
3860 * we can do some work if the lock was obtained.
3861 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07003862 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08003863
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003864 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003866 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867
Christoph Lameter35386e32006-03-22 00:09:05 -08003868 /*
3869 * These are racy checks but it does not matter
3870 * if we skip one check or scan twice.
3871 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003872 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003873 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003875 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003877 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003879 if (n->free_touched)
3880 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003881 else {
3882 int freed;
3883
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003884 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003885 5 * searchp->num - 1) / (5 * searchp->num));
3886 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003888next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 cond_resched();
3890 }
3891 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05003892 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003893 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003894out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08003895 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003896 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897}
3898
Linus Torvalds158a9622008-01-02 13:04:48 -08003899#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04003900void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003902 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003903 unsigned long active_objs;
3904 unsigned long num_objs;
3905 unsigned long active_slabs = 0;
3906 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003907 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003909 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003910 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912 active_objs = 0;
3913 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003914 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07003915
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003916 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003917 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003918
Joonsoo Kim8456a642013-10-24 10:07:49 +09003919 list_for_each_entry(page, &n->slabs_full, lru) {
3920 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07003921 error = "slabs_full accounting error";
3922 active_objs += cachep->num;
3923 active_slabs++;
3924 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003925 list_for_each_entry(page, &n->slabs_partial, lru) {
3926 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003927 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003928 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003929 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003930 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07003931 active_slabs++;
3932 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003933 list_for_each_entry(page, &n->slabs_free, lru) {
3934 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003935 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07003936 num_slabs++;
3937 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003938 free_objects += n->free_objects;
3939 if (n->shared)
3940 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003941
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003942 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003944 num_slabs += active_slabs;
3945 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003946 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 error = "free_objects accounting error";
3948
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003949 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 if (error)
3951 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3952
Glauber Costa0d7561c2012-10-19 18:20:27 +04003953 sinfo->active_objs = active_objs;
3954 sinfo->num_objs = num_objs;
3955 sinfo->active_slabs = active_slabs;
3956 sinfo->num_slabs = num_slabs;
3957 sinfo->shared_avail = shared_avail;
3958 sinfo->limit = cachep->limit;
3959 sinfo->batchcount = cachep->batchcount;
3960 sinfo->shared = cachep->shared;
3961 sinfo->objects_per_slab = cachep->num;
3962 sinfo->cache_order = cachep->gfporder;
3963}
3964
3965void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3966{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003968 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 unsigned long high = cachep->high_mark;
3970 unsigned long allocs = cachep->num_allocations;
3971 unsigned long grown = cachep->grown;
3972 unsigned long reaped = cachep->reaped;
3973 unsigned long errors = cachep->errors;
3974 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003976 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003977 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978
Joe Perchese92dd4f2010-03-26 19:27:58 -07003979 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
3980 "%4lu %4lu %4lu %4lu %4lu",
3981 allocs, high, grown,
3982 reaped, errors, max_freeable, node_allocs,
3983 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 }
3985 /* cpu stats */
3986 {
3987 unsigned long allochit = atomic_read(&cachep->allochit);
3988 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3989 unsigned long freehit = atomic_read(&cachep->freehit);
3990 unsigned long freemiss = atomic_read(&cachep->freemiss);
3991
3992 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003993 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 }
3995#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996}
3997
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998#define MAX_SLABINFO_WRITE 128
3999/**
4000 * slabinfo_write - Tuning for the slab allocator
4001 * @file: unused
4002 * @buffer: user buffer
4003 * @count: data length
4004 * @ppos: unused
4005 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004006ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004007 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004009 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004011 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004012
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 if (count > MAX_SLABINFO_WRITE)
4014 return -EINVAL;
4015 if (copy_from_user(&kbuf, buffer, count))
4016 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004017 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018
4019 tmp = strchr(kbuf, ' ');
4020 if (!tmp)
4021 return -EINVAL;
4022 *tmp = '\0';
4023 tmp++;
4024 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4025 return -EINVAL;
4026
4027 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004028 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004030 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004032 if (limit < 1 || batchcount < 1 ||
4033 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004034 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004036 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004037 batchcount, shared,
4038 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 }
4040 break;
4041 }
4042 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004043 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 if (res >= 0)
4045 res = count;
4046 return res;
4047}
Al Viro871751e2006-03-25 03:06:39 -08004048
4049#ifdef CONFIG_DEBUG_SLAB_LEAK
4050
Al Viro871751e2006-03-25 03:06:39 -08004051static inline int add_caller(unsigned long *n, unsigned long v)
4052{
4053 unsigned long *p;
4054 int l;
4055 if (!v)
4056 return 1;
4057 l = n[1];
4058 p = n + 2;
4059 while (l) {
4060 int i = l/2;
4061 unsigned long *q = p + 2 * i;
4062 if (*q == v) {
4063 q[1]++;
4064 return 1;
4065 }
4066 if (*q > v) {
4067 l = i;
4068 } else {
4069 p = q + 2;
4070 l -= i + 1;
4071 }
4072 }
4073 if (++n[1] == n[0])
4074 return 0;
4075 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4076 p[0] = v;
4077 p[1] = 1;
4078 return 1;
4079}
4080
Joonsoo Kim8456a642013-10-24 10:07:49 +09004081static void handle_slab(unsigned long *n, struct kmem_cache *c,
4082 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004083{
4084 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004085 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004086
Al Viro871751e2006-03-25 03:06:39 -08004087 if (n[0] == n[1])
4088 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004089 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004090 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004091 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004092
Al Viro871751e2006-03-25 03:06:39 -08004093 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4094 return;
4095 }
4096}
4097
4098static void show_symbol(struct seq_file *m, unsigned long address)
4099{
4100#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004101 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004102 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004103
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004104 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004105 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004106 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004107 seq_printf(m, " [%s]", modname);
4108 return;
4109 }
4110#endif
4111 seq_printf(m, "%p", (void *)address);
4112}
4113
4114static int leaks_show(struct seq_file *m, void *p)
4115{
Thierry Reding0672aa72012-06-22 19:42:49 +02004116 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004117 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004118 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004119 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004120 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004121 int node;
4122 int i;
4123
4124 if (!(cachep->flags & SLAB_STORE_USER))
4125 return 0;
4126 if (!(cachep->flags & SLAB_RED_ZONE))
4127 return 0;
4128
4129 /* OK, we can do it */
4130
Christoph Lameterdb845062013-02-05 18:45:23 +00004131 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004132
Christoph Lameter18bf8542014-08-06 16:04:11 -07004133 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004134
4135 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004136 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004137
Joonsoo Kim8456a642013-10-24 10:07:49 +09004138 list_for_each_entry(page, &n->slabs_full, lru)
4139 handle_slab(x, cachep, page);
4140 list_for_each_entry(page, &n->slabs_partial, lru)
4141 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004142 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004143 }
4144 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004145 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004146 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004147 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004148 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004149 if (!m->private) {
4150 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004151 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004152 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004153 return -ENOMEM;
4154 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004155 *(unsigned long *)m->private = x[0] * 2;
4156 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004157 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004158 /* Now make sure this entry will be retried */
4159 m->count = m->size;
4160 return 0;
4161 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004162 for (i = 0; i < x[1]; i++) {
4163 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4164 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004165 seq_putc(m, '\n');
4166 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004167
Al Viro871751e2006-03-25 03:06:39 -08004168 return 0;
4169}
4170
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004171static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004172 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004173 .next = slab_next,
4174 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004175 .show = leaks_show,
4176};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004177
4178static int slabstats_open(struct inode *inode, struct file *file)
4179{
Rob Jonesb208ce32014-10-09 15:28:03 -07004180 unsigned long *n;
4181
4182 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4183 if (!n)
4184 return -ENOMEM;
4185
4186 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4187
4188 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004189}
4190
4191static const struct file_operations proc_slabstats_operations = {
4192 .open = slabstats_open,
4193 .read = seq_read,
4194 .llseek = seq_lseek,
4195 .release = seq_release_private,
4196};
Al Viro871751e2006-03-25 03:06:39 -08004197#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004198
4199static int __init slab_proc_init(void)
4200{
4201#ifdef CONFIG_DEBUG_SLAB_LEAK
4202 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4203#endif
4204 return 0;
4205}
4206module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207#endif
4208
Manfred Spraul00e145b2005-09-03 15:55:07 -07004209/**
4210 * ksize - get the actual amount of memory allocated for a given object
4211 * @objp: Pointer to the object
4212 *
4213 * kmalloc may internally round up allocations and return more memory
4214 * than requested. ksize() can be used to determine the actual amount of
4215 * memory allocated. The caller may use this additional memory, even though
4216 * a smaller amount of memory was initially specified with the kmalloc call.
4217 * The caller must guarantee that objp points to a valid object previously
4218 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4219 * must not be freed during the duration of the call.
4220 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004221size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004223 BUG_ON(!objp);
4224 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004225 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004227 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004228}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004229EXPORT_SYMBOL(ksize);