blob: 7eb38dd1cefa2f988be6297c1b75b12369e05827 [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
David Rientjes4167e9b2015-04-14 15:46:55 -0700860static inline gfp_t gfp_exact_node(gfp_t flags)
861{
862 return flags;
863}
864
Christoph Lameter765c4502006-09-27 01:50:08 -0700865#else /* CONFIG_NUMA */
866
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800867static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800868static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800869
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700870static struct alien_cache *__alloc_alien_cache(int node, int entries,
871 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700872{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700873 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700874 struct alien_cache *alc = NULL;
875
876 alc = kmalloc_node(memsize, gfp, node);
877 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700878 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700879 return alc;
880}
881
882static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
883{
884 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700885 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700886 int i;
887
888 if (limit > 1)
889 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700890 alc_ptr = kzalloc_node(memsize, gfp, node);
891 if (!alc_ptr)
892 return NULL;
893
894 for_each_node(i) {
895 if (i == node || !node_online(i))
896 continue;
897 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
898 if (!alc_ptr[i]) {
899 for (i--; i >= 0; i--)
900 kfree(alc_ptr[i]);
901 kfree(alc_ptr);
902 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700903 }
904 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700905 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700906}
907
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700908static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700909{
910 int i;
911
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700912 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700913 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700914 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700915 kfree(alc_ptr[i]);
916 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700917}
918
Pekka Enberg343e0d72006-02-01 03:05:50 -0800919static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700920 struct array_cache *ac, int node,
921 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700922{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700923 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700924
925 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000926 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800927 /*
928 * Stuff objects into the remote nodes shared array first.
929 * That way we could avoid the overhead of putting the objects
930 * into the free lists and getting them back later.
931 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000932 if (n->shared)
933 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800934
Joonsoo Kim833b7062014-08-06 16:04:33 -0700935 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700936 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000937 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700938 }
939}
940
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800941/*
942 * Called from cache_reap() to regularly drain alien caches round robin.
943 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000944static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800945{
Christoph Lameter909ea962010-12-08 16:22:55 +0100946 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800947
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000948 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700949 struct alien_cache *alc = n->alien[node];
950 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800951
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700952 if (alc) {
953 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700954 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700955 LIST_HEAD(list);
956
957 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700958 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700959 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700960 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800961 }
962 }
963}
964
Andrew Mortona737b3e2006-03-22 00:08:11 -0800965static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700966 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700967{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800968 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700969 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700970 struct array_cache *ac;
971 unsigned long flags;
972
973 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700974 alc = alien[i];
975 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700976 LIST_HEAD(list);
977
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700978 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700979 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700980 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700981 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700982 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700983 }
984 }
985}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700986
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700987static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
988 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700989{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000990 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700991 struct alien_cache *alien = NULL;
992 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700993 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700994
Christoph Lameter18bf8542014-08-06 16:04:11 -0700995 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700996 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700997 if (n->alien && n->alien[page_node]) {
998 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700999 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001000 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001001 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001002 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001003 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001004 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001005 ac_put_obj(cachep, ac, objp);
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001006 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001007 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001008 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001009 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001010 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001011 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001012 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001013 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001014 }
1015 return 1;
1016}
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001017
1018static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1019{
1020 int page_node = page_to_nid(virt_to_page(objp));
1021 int node = numa_mem_id();
1022 /*
1023 * Make sure we are not freeing a object from another node to the array
1024 * cache on this cpu.
1025 */
1026 if (likely(node == page_node))
1027 return 0;
1028
1029 return __cache_free_alien(cachep, objp, node, page_node);
1030}
David Rientjes4167e9b2015-04-14 15:46:55 -07001031
1032/*
1033 * Construct gfp mask to allocate from a specific node but do not invoke reclaim
1034 * or warn about failures.
1035 */
1036static inline gfp_t gfp_exact_node(gfp_t flags)
1037{
1038 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_WAIT;
1039}
Christoph Lametere498be72005-09-09 13:03:32 -07001040#endif
1041
David Rientjes8f9f8d92010-03-27 19:40:47 -07001042/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001043 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001044 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001045 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001046 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001047 * already in use.
1048 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001049 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001050 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001051static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001052{
1053 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001054 struct kmem_cache_node *n;
Joonsoo Kim5e804782014-08-06 16:04:40 -07001055 const size_t memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001056
Christoph Lameter18004c52012-07-06 15:25:12 -05001057 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001058 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001059 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001060 * begin anything. Make sure some other cpu on this
1061 * node has not already allocated this
1062 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07001063 n = get_node(cachep, node);
1064 if (!n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001065 n = kmalloc_node(memsize, GFP_KERNEL, node);
1066 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001067 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001068 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001069 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1070 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001071
1072 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001073 * The kmem_cache_nodes don't come and go as CPUs
1074 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001075 * protection here.
1076 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001077 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001078 }
1079
Christoph Lameter18bf8542014-08-06 16:04:11 -07001080 spin_lock_irq(&n->list_lock);
1081 n->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001082 (1 + nr_cpus_node(node)) *
1083 cachep->batchcount + cachep->num;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001084 spin_unlock_irq(&n->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001085 }
1086 return 0;
1087}
1088
Wanpeng Li0fa81032013-07-04 08:33:22 +08001089static inline int slabs_tofree(struct kmem_cache *cachep,
1090 struct kmem_cache_node *n)
1091{
1092 return (n->free_objects + cachep->num - 1) / cachep->num;
1093}
1094
Paul Gortmaker0db06282013-06-19 14:53:51 -04001095static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001097 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001098 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001099 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301100 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001101
Christoph Lameter18004c52012-07-06 15:25:12 -05001102 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001103 struct array_cache *nc;
1104 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001105 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -07001106 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001107
Christoph Lameter18bf8542014-08-06 16:04:11 -07001108 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001109 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001110 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001111
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001112 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001113
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001114 /* Free limit for this kmem_cache_node */
1115 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001116
1117 /* cpu is dead; no one can alloc from it. */
1118 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1119 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001120 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001121 nc->avail = 0;
1122 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001123
Rusty Russell58463c12009-12-17 11:43:12 -06001124 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001126 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001127 }
1128
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001129 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001130 if (shared) {
1131 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001132 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001133 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134 }
1135
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001136 alien = n->alien;
1137 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001138
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001139 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001140
1141 kfree(shared);
1142 if (alien) {
1143 drain_alien_cache(cachep, alien);
1144 free_alien_cache(alien);
1145 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001146
1147free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001148 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001149 }
1150 /*
1151 * In the previous loop, all the objects were freed to
1152 * the respective cache's slabs, now we can go ahead and
1153 * shrink each nodelist to its limit.
1154 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001155 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001156 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001157 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001159 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 }
1161}
1162
Paul Gortmaker0db06282013-06-19 14:53:51 -04001163static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001164{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001165 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001166 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001167 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001168 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001170 /*
1171 * We need to do this right in the beginning since
1172 * alloc_arraycache's are going to use this list.
1173 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001174 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001175 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001176 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001177 if (err < 0)
1178 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001179
1180 /*
1181 * Now we can go ahead with allocating the shared arrays and
1182 * array caches
1183 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001184 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 struct array_cache *shared = NULL;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001186 struct alien_cache **alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001188 if (cachep->shared) {
1189 shared = alloc_arraycache(node,
1190 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001191 0xbaadf00d, GFP_KERNEL);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001192 if (!shared)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193 goto bad;
1194 }
1195 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001196 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001197 if (!alien) {
1198 kfree(shared);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001200 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001201 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07001202 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001203 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001204
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001205 spin_lock_irq(&n->list_lock);
1206 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 /*
1208 * We are serialised from CPU_DEAD or
1209 * CPU_UP_CANCELLED by the cpucontrol lock
1210 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001211 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001212 shared = NULL;
1213 }
1214#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001215 if (!n->alien) {
1216 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001217 alien = NULL;
1218 }
1219#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001220 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001221 kfree(shared);
1222 free_alien_cache(alien);
1223 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001224
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001225 return 0;
1226bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001227 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001228 return -ENOMEM;
1229}
1230
Paul Gortmaker0db06282013-06-19 14:53:51 -04001231static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001232 unsigned long action, void *hcpu)
1233{
1234 long cpu = (long)hcpu;
1235 int err = 0;
1236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001238 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001239 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001240 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001241 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001242 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 break;
1244 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001245 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 start_cpu_timer(cpu);
1247 break;
1248#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001249 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001250 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001251 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001252 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001253 * held so that if cache_reap() is invoked it cannot do
1254 * anything expensive but will only modify reap_work
1255 * and reschedule the timer.
1256 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001257 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001258 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001259 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001260 break;
1261 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001262 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001263 start_cpu_timer(cpu);
1264 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001266 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001267 /*
1268 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001269 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001270 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001271 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001272 * structure is usually allocated from kmem_cache_create() and
1273 * gets destroyed at kmem_cache_destroy().
1274 */
Simon Arlott183ff222007-10-20 01:27:18 +02001275 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001276#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001279 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001281 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001284 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Paul Gortmaker0db06282013-06-19 14:53:51 -04001287static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001288 &cpuup_callback, NULL, 0
1289};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
David Rientjes8f9f8d92010-03-27 19:40:47 -07001291#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1292/*
1293 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1294 * Returns -EBUSY if all objects cannot be drained so that the node is not
1295 * removed.
1296 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001297 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001298 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001299static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001300{
1301 struct kmem_cache *cachep;
1302 int ret = 0;
1303
Christoph Lameter18004c52012-07-06 15:25:12 -05001304 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001306
Christoph Lameter18bf8542014-08-06 16:04:11 -07001307 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001309 continue;
1310
Wanpeng Li0fa81032013-07-04 08:33:22 +08001311 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001312
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001313 if (!list_empty(&n->slabs_full) ||
1314 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001315 ret = -EBUSY;
1316 break;
1317 }
1318 }
1319 return ret;
1320}
1321
1322static int __meminit slab_memory_callback(struct notifier_block *self,
1323 unsigned long action, void *arg)
1324{
1325 struct memory_notify *mnb = arg;
1326 int ret = 0;
1327 int nid;
1328
1329 nid = mnb->status_change_nid;
1330 if (nid < 0)
1331 goto out;
1332
1333 switch (action) {
1334 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001335 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001336 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001337 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001338 break;
1339 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001340 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001341 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001342 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001343 break;
1344 case MEM_ONLINE:
1345 case MEM_OFFLINE:
1346 case MEM_CANCEL_ONLINE:
1347 case MEM_CANCEL_OFFLINE:
1348 break;
1349 }
1350out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001351 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001352}
1353#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1354
Christoph Lametere498be72005-09-09 13:03:32 -07001355/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001356 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001357 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001358static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001359 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001360{
Christoph Lameter6744f082013-01-10 19:12:17 +00001361 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001362
Christoph Lameter6744f082013-01-10 19:12:17 +00001363 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001364 BUG_ON(!ptr);
1365
Christoph Lameter6744f082013-01-10 19:12:17 +00001366 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001367 /*
1368 * Do not assume that spinlocks can be initialized via memcpy:
1369 */
1370 spin_lock_init(&ptr->list_lock);
1371
Christoph Lametere498be72005-09-09 13:03:32 -07001372 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001373 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001374}
1375
Andrew Mortona737b3e2006-03-22 00:08:11 -08001376/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001377 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1378 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001379 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001380static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001381{
1382 int node;
1383
1384 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001385 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001386 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001387 REAPTIMEOUT_NODE +
1388 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001389 }
1390}
1391
1392/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001393 * Initialisation. Called after the page allocator have been initialised and
1394 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 */
1396void __init kmem_cache_init(void)
1397{
Christoph Lametere498be72005-09-09 13:03:32 -07001398 int i;
1399
Joonsoo Kim68126702013-10-24 10:07:42 +09001400 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1401 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001402 kmem_cache = &kmem_cache_boot;
1403
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001404 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001405 use_alien_caches = 0;
1406
Christoph Lameter3c583462012-11-28 16:23:01 +00001407 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001408 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001409
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 /*
1411 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001412 * page orders on machines with more than 32MB of memory if
1413 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001415 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001416 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001420 * 1) initialize the kmem_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except kmem_cache itself:
1422 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001423 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001424 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001425 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001427 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1430 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001431 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001433 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
1437
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001438 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
Eric Dumazet8da34302007-05-06 14:49:29 -07001440 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001441 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001442 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001443 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001444 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001445 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001446 SLAB_HWCACHE_ALIGN);
1447 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001448 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
Andrew Mortona737b3e2006-03-22 00:08:11 -08001450 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001451 * Initialize the caches that provide memory for the kmem_cache_node
1452 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001453 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001454 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001455 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001456 slab_state = PARTIAL_NODE;
Christoph Lametere498be72005-09-09 13:03:32 -07001457
Ingo Molnare0a42722006-06-23 02:03:46 -07001458 slab_early_init = 0;
1459
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001460 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001461 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001462 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463
Mel Gorman9c09a952008-01-24 05:49:54 -08001464 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001465 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001466
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001467 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001468 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001469 }
1470 }
1471
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001472 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001473}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001474
Pekka Enberg8429db52009-06-12 15:58:59 +03001475void __init kmem_cache_init_late(void)
1476{
1477 struct kmem_cache *cachep;
1478
Christoph Lameter97d06602012-07-06 15:25:11 -05001479 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001480
Pekka Enberg8429db52009-06-12 15:58:59 +03001481 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001482 mutex_lock(&slab_mutex);
1483 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001484 if (enable_cpucache(cachep, GFP_NOWAIT))
1485 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001486 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001487
Christoph Lameter97d06602012-07-06 15:25:11 -05001488 /* Done! */
1489 slab_state = FULL;
1490
Andrew Mortona737b3e2006-03-22 00:08:11 -08001491 /*
1492 * Register a cpu startup notifier callback that initializes
1493 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 */
1495 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
David Rientjes8f9f8d92010-03-27 19:40:47 -07001497#ifdef CONFIG_NUMA
1498 /*
1499 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001500 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001501 */
1502 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1503#endif
1504
Andrew Mortona737b3e2006-03-22 00:08:11 -08001505 /*
1506 * The reap timers are started later, with a module init call: That part
1507 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508 */
1509}
1510
1511static int __init cpucache_init(void)
1512{
1513 int cpu;
1514
Andrew Mortona737b3e2006-03-22 00:08:11 -08001515 /*
1516 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 */
Christoph Lametere498be72005-09-09 13:03:32 -07001518 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001519 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001520
1521 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001522 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 return 0;
1524}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525__initcall(cpucache_init);
1526
Rafael Aquini8bdec192012-03-09 17:27:27 -03001527static noinline void
1528slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1529{
David Rientjes9a02d692014-06-04 16:06:36 -07001530#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001531 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001532 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001533 unsigned long flags;
1534 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001535 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1536 DEFAULT_RATELIMIT_BURST);
1537
1538 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1539 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001540
1541 printk(KERN_WARNING
1542 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1543 nodeid, gfpflags);
1544 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001545 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001546
Christoph Lameter18bf8542014-08-06 16:04:11 -07001547 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001548 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1549 unsigned long active_slabs = 0, num_slabs = 0;
1550
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001551 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001552 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001553 active_objs += cachep->num;
1554 active_slabs++;
1555 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001556 list_for_each_entry(page, &n->slabs_partial, lru) {
1557 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001558 active_slabs++;
1559 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001560 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001561 num_slabs++;
1562
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001563 free_objects += n->free_objects;
1564 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001565
1566 num_slabs += active_slabs;
1567 num_objs = num_slabs * cachep->num;
1568 printk(KERN_WARNING
1569 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1570 node, active_slabs, num_slabs, active_objs, num_objs,
1571 free_objects);
1572 }
David Rientjes9a02d692014-06-04 16:06:36 -07001573#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001574}
1575
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001577 * Interface to system's page allocator. No need to hold the
1578 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 *
1580 * If we requested dmaable memory, we will get it. Even if we
1581 * did not request dmaable memory, we might get it, but that
1582 * would be relatively rare and ignorable.
1583 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001584static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1585 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586{
1587 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001588 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001589
Glauber Costaa618e892012-06-14 16:17:21 +04001590 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001591 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1592 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001593
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001594 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1595 return NULL;
1596
Linus Torvalds517d0862009-06-16 19:50:13 -07001597 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001598 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001599 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001600 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001602 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001604 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001605 if (unlikely(page->pfmemalloc))
1606 pfmemalloc_active = true;
1607
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001608 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001610 add_zone_page_state(page_zone(page),
1611 NR_SLAB_RECLAIMABLE, nr_pages);
1612 else
1613 add_zone_page_state(page_zone(page),
1614 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001615 __SetPageSlab(page);
1616 if (page->pfmemalloc)
1617 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001618
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001619 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1620 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1621
1622 if (cachep->ctor)
1623 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1624 else
1625 kmemcheck_mark_unallocated_pages(page, nr_pages);
1626 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001627
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001628 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629}
1630
1631/*
1632 * Interface to system's page release.
1633 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001634static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001636 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001638 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001639
Christoph Lameter972d1a72006-09-25 23:31:51 -07001640 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1641 sub_zone_page_state(page_zone(page),
1642 NR_SLAB_RECLAIMABLE, nr_freed);
1643 else
1644 sub_zone_page_state(page_zone(page),
1645 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001646
Joonsoo Kima57a4982013-10-24 10:07:44 +09001647 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001648 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001649 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001650 page_mapcount_reset(page);
1651 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001652
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 if (current->reclaim_state)
1654 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001655 __free_pages(page, cachep->gfporder);
1656 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657}
1658
1659static void kmem_rcu_free(struct rcu_head *head)
1660{
Joonsoo Kim68126702013-10-24 10:07:42 +09001661 struct kmem_cache *cachep;
1662 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663
Joonsoo Kim68126702013-10-24 10:07:42 +09001664 page = container_of(head, struct page, rcu_head);
1665 cachep = page->slab_cache;
1666
1667 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668}
1669
1670#if DEBUG
1671
1672#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001673static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001674 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001676 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001678 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001680 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 return;
1682
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001683 *addr++ = 0x12345678;
1684 *addr++ = caller;
1685 *addr++ = smp_processor_id();
1686 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 {
1688 unsigned long *sptr = &caller;
1689 unsigned long svalue;
1690
1691 while (!kstack_end(sptr)) {
1692 svalue = *sptr++;
1693 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001694 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 size -= sizeof(unsigned long);
1696 if (size <= sizeof(unsigned long))
1697 break;
1698 }
1699 }
1700
1701 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001702 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703}
1704#endif
1705
Pekka Enberg343e0d72006-02-01 03:05:50 -08001706static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001708 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001709 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710
1711 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001712 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713}
1714
1715static void dump_line(char *data, int offset, int limit)
1716{
1717 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001718 unsigned char error = 0;
1719 int bad_count = 0;
1720
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001721 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001722 for (i = 0; i < limit; i++) {
1723 if (data[offset + i] != POISON_FREE) {
1724 error = data[offset + i];
1725 bad_count++;
1726 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001727 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001728 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1729 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001730
1731 if (bad_count == 1) {
1732 error ^= POISON_FREE;
1733 if (!(error & (error - 1))) {
1734 printk(KERN_ERR "Single bit error detected. Probably "
1735 "bad RAM.\n");
1736#ifdef CONFIG_X86
1737 printk(KERN_ERR "Run memtest86+ or a similar memory "
1738 "test tool.\n");
1739#else
1740 printk(KERN_ERR "Run a memory test tool.\n");
1741#endif
1742 }
1743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744}
1745#endif
1746
1747#if DEBUG
1748
Pekka Enberg343e0d72006-02-01 03:05:50 -08001749static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750{
1751 int i, size;
1752 char *realobj;
1753
1754 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001755 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001756 *dbg_redzone1(cachep, objp),
1757 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 }
1759
1760 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001761 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1762 *dbg_userword(cachep, objp),
1763 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001765 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001766 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 int limit;
1769 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001770 if (i + limit > size)
1771 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 dump_line(realobj, i, limit);
1773 }
1774}
1775
Pekka Enberg343e0d72006-02-01 03:05:50 -08001776static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777{
1778 char *realobj;
1779 int size, i;
1780 int lines = 0;
1781
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001782 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001783 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001785 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001787 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 exp = POISON_END;
1789 if (realobj[i] != exp) {
1790 int limit;
1791 /* Mismatch ! */
1792 /* Print header */
1793 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001795 "Slab corruption (%s): %s start=%p, len=%d\n",
1796 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 print_objinfo(cachep, objp, 0);
1798 }
1799 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001802 if (i + limit > size)
1803 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 dump_line(realobj, i, limit);
1805 i += 16;
1806 lines++;
1807 /* Limit to 5 lines */
1808 if (lines > 5)
1809 break;
1810 }
1811 }
1812 if (lines != 0) {
1813 /* Print some data about the neighboring objects, if they
1814 * exist:
1815 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001816 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001817 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818
Joonsoo Kim8456a642013-10-24 10:07:49 +09001819 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001821 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001822 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001824 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 print_objinfo(cachep, objp, 2);
1826 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001828 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001829 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001831 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 print_objinfo(cachep, objp, 2);
1833 }
1834 }
1835}
1836#endif
1837
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001839static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1840 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001841{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 int i;
1843 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001844 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845
1846 if (cachep->flags & SLAB_POISON) {
1847#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001848 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001849 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001850 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001851 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 else
1853 check_poison_obj(cachep, objp);
1854#else
1855 check_poison_obj(cachep, objp);
1856#endif
1857 }
1858 if (cachep->flags & SLAB_RED_ZONE) {
1859 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1860 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1863 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001864 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001867}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001869static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1870 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001871{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001872}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873#endif
1874
Randy Dunlap911851e2006-03-22 00:08:14 -08001875/**
1876 * slab_destroy - destroy and release all objects in a slab
1877 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001878 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001879 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001880 * Destroy all the objs in a slab page, and release the mem back to the system.
1881 * Before calling the slab page must have been unlinked from the cache. The
1882 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001883 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001884static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001885{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001886 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001887
Joonsoo Kim8456a642013-10-24 10:07:49 +09001888 freelist = page->freelist;
1889 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001891 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892
Joonsoo Kim68126702013-10-24 10:07:42 +09001893 /*
1894 * RCU free overloads the RCU head over the LRU.
1895 * slab_page has been overloeaded over the LRU,
1896 * however it is not used from now on so that
1897 * we can use it safely.
1898 */
1899 head = (void *)&page->rcu_head;
1900 call_rcu(head, kmem_rcu_free);
1901
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001903 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001905
1906 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001907 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001908 * although actual page can be freed in rcu context
1909 */
1910 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001911 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912}
1913
Joonsoo Kim97654df2014-08-06 16:04:25 -07001914static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1915{
1916 struct page *page, *n;
1917
1918 list_for_each_entry_safe(page, n, list, lru) {
1919 list_del(&page->lru);
1920 slab_destroy(cachep, page);
1921 }
1922}
1923
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001925 * calculate_slab_order - calculate size (page order) of slabs
1926 * @cachep: pointer to the cache that is being created
1927 * @size: size of objects to be created in this cache.
1928 * @align: required alignment for the objects.
1929 * @flags: slab allocation flags
1930 *
1931 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001932 *
1933 * This could be made much more intelligent. For now, try to avoid using
1934 * high order pages for slabs. When the gfp() functions are more friendly
1935 * towards high-order requests, this should be changed.
1936 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001937static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001938 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001939{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001940 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001941 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001942 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001943
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001944 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001945 unsigned int num;
1946 size_t remainder;
1947
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001948 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001949 if (!num)
1950 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001951
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001952 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1953 if (num > SLAB_OBJ_MAX_NUM)
1954 break;
1955
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001956 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07001957 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001958 /*
1959 * Max number of objs-per-slab for caches which
1960 * use off-slab slabs. Needed to avoid a possible
1961 * looping condition in cache_grow().
1962 */
Joonsoo Kim03787302014-06-23 13:22:06 -07001963 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1964 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001965 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07001966 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001967
1968 if (num > offslab_limit)
1969 break;
1970 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001971
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001972 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001973 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001974 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001975 left_over = remainder;
1976
1977 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001978 * A VFS-reclaimable slab tends to have most allocations
1979 * as GFP_NOFS and we really don't want to have to be allocating
1980 * higher-order pages when we are unable to shrink dcache.
1981 */
1982 if (flags & SLAB_RECLAIM_ACCOUNT)
1983 break;
1984
1985 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001986 * Large number of objects is good, but very large slabs are
1987 * currently bad for the gfp()s.
1988 */
David Rientjes543585c2011-10-18 22:09:24 -07001989 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001990 break;
1991
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001992 /*
1993 * Acceptable internal fragmentation?
1994 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001995 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001996 break;
1997 }
1998 return left_over;
1999}
2000
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002001static struct array_cache __percpu *alloc_kmem_cache_cpus(
2002 struct kmem_cache *cachep, int entries, int batchcount)
2003{
2004 int cpu;
2005 size_t size;
2006 struct array_cache __percpu *cpu_cache;
2007
2008 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07002009 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002010
2011 if (!cpu_cache)
2012 return NULL;
2013
2014 for_each_possible_cpu(cpu) {
2015 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2016 entries, batchcount);
2017 }
2018
2019 return cpu_cache;
2020}
2021
Pekka Enberg83b519e2009-06-10 19:40:04 +03002022static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002023{
Christoph Lameter97d06602012-07-06 15:25:11 -05002024 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002025 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002026
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002027 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2028 if (!cachep->cpu_cache)
2029 return 1;
2030
Christoph Lameter97d06602012-07-06 15:25:11 -05002031 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002032 /* Creation of first cache (kmem_cache). */
2033 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002034 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002035 /* For kmem_cache_node */
2036 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002037 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002038 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002039
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002040 for_each_online_node(node) {
2041 cachep->node[node] = kmalloc_node(
2042 sizeof(struct kmem_cache_node), gfp, node);
2043 BUG_ON(!cachep->node[node]);
2044 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002045 }
2046 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002047
Christoph Lameter6a673682013-01-10 19:14:19 +00002048 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002049 jiffies + REAPTIMEOUT_NODE +
2050 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002051
2052 cpu_cache_get(cachep)->avail = 0;
2053 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2054 cpu_cache_get(cachep)->batchcount = 1;
2055 cpu_cache_get(cachep)->touched = 0;
2056 cachep->batchcount = 1;
2057 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002058 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002059}
2060
Joonsoo Kim12220de2014-10-09 15:26:24 -07002061unsigned long kmem_cache_flags(unsigned long object_size,
2062 unsigned long flags, const char *name,
2063 void (*ctor)(void *))
2064{
2065 return flags;
2066}
2067
2068struct kmem_cache *
2069__kmem_cache_alias(const char *name, size_t size, size_t align,
2070 unsigned long flags, void (*ctor)(void *))
2071{
2072 struct kmem_cache *cachep;
2073
2074 cachep = find_mergeable(size, align, flags, name, ctor);
2075 if (cachep) {
2076 cachep->refcount++;
2077
2078 /*
2079 * Adjust the object sizes so that we clear
2080 * the complete object on kzalloc.
2081 */
2082 cachep->object_size = max_t(int, cachep->object_size, size);
2083 }
2084 return cachep;
2085}
2086
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002087/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002088 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002089 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 *
2092 * Returns a ptr to the cache on success, NULL on failure.
2093 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002094 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 * The flags are
2097 *
2098 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2099 * to catch references to uninitialised memory.
2100 *
2101 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2102 * for buffer overruns.
2103 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2105 * cacheline. This can be beneficial if you're counting cycles as closely
2106 * as davem.
2107 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002108int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002109__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002111 size_t left_over, freelist_size;
2112 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002113 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002114 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002115 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118#if FORCED_DEBUG
2119 /*
2120 * Enable redzoning and last user accounting, except for caches with
2121 * large objects, if the increased size would increase the object size
2122 * above the next power of two: caches with object sizes just above a
2123 * power of two have a significant amount of internal fragmentation.
2124 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002125 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2126 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002127 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 if (!(flags & SLAB_DESTROY_BY_RCU))
2129 flags |= SLAB_POISON;
2130#endif
2131 if (flags & SLAB_DESTROY_BY_RCU)
2132 BUG_ON(flags & SLAB_POISON);
2133#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134
Andrew Mortona737b3e2006-03-22 00:08:11 -08002135 /*
2136 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 * unaligned accesses for some archs when redzoning is used, and makes
2138 * sure any on-slab bufctl's are also correctly aligned.
2139 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002140 if (size & (BYTES_PER_WORD - 1)) {
2141 size += (BYTES_PER_WORD - 1);
2142 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 }
2144
David Woodhouse87a927c2007-07-04 21:26:44 -04002145 if (flags & SLAB_RED_ZONE) {
2146 ralign = REDZONE_ALIGN;
2147 /* If redzoning, ensure that the second redzone is suitably
2148 * aligned, by adjusting the object size accordingly. */
2149 size += REDZONE_ALIGN - 1;
2150 size &= ~(REDZONE_ALIGN - 1);
2151 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002152
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002153 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002154 if (ralign < cachep->align) {
2155 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002157 /* disable debug if necessary */
2158 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002159 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002160 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002161 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002163 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164
Pekka Enberg83b519e2009-06-10 19:40:04 +03002165 if (slab_is_available())
2166 gfp = GFP_KERNEL;
2167 else
2168 gfp = GFP_NOWAIT;
2169
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171
Pekka Enbergca5f9702006-09-25 23:31:25 -07002172 /*
2173 * Both debugging options require word-alignment which is calculated
2174 * into align above.
2175 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002178 cachep->obj_offset += sizeof(unsigned long long);
2179 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 }
2181 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002182 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002183 * the real object. But if the second red zone needs to be
2184 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002186 if (flags & SLAB_RED_ZONE)
2187 size += REDZONE_ALIGN;
2188 else
2189 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 }
2191#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002192 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002193 && cachep->object_size > cache_line_size()
2194 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2195 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 size = PAGE_SIZE;
2197 }
2198#endif
2199#endif
2200
Ingo Molnare0a42722006-06-23 02:03:46 -07002201 /*
2202 * Determine if the slab management is 'on' or 'off' slab.
2203 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002204 * it too early on. Always use on-slab management when
2205 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002206 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002207 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002208 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 /*
2210 * Size is large, assume best to place the slab management obj
2211 * off-slab (should allow better packing of objs).
2212 */
2213 flags |= CFLGS_OFF_SLAB;
2214
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002215 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002216 /*
2217 * We should restrict the number of objects in a slab to implement
2218 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2219 */
2220 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2221 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002223 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002225 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002226 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002227
Joonsoo Kim03787302014-06-23 13:22:06 -07002228 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229
2230 /*
2231 * If the slab has been placed off-slab, and we have enough space then
2232 * move it on-slab. This is at the expense of any extra colouring.
2233 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002234 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002236 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 }
2238
2239 if (flags & CFLGS_OFF_SLAB) {
2240 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002241 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302242
2243#ifdef CONFIG_PAGE_POISONING
2244 /* If we're going to use the generic kernel_map_pages()
2245 * poisoning, then it's going to smash the contents of
2246 * the redzone and userword anyhow, so switch them off.
2247 */
2248 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2249 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2250#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 }
2252
2253 cachep->colour_off = cache_line_size();
2254 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002255 if (cachep->colour_off < cachep->align)
2256 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002257 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002258 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002260 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002261 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002262 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002263 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002264 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002266 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002267 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002268 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002269 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002270 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002271 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2272 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002273 * But leave a BUG_ON for some lucky dude.
2274 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002275 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002276 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002278 err = setup_cpu_cache(cachep, gfp);
2279 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002280 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002281 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002282 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002284 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
2287#if DEBUG
2288static void check_irq_off(void)
2289{
2290 BUG_ON(!irqs_disabled());
2291}
2292
2293static void check_irq_on(void)
2294{
2295 BUG_ON(irqs_disabled());
2296}
2297
Pekka Enberg343e0d72006-02-01 03:05:50 -08002298static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299{
2300#ifdef CONFIG_SMP
2301 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002302 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303#endif
2304}
Christoph Lametere498be72005-09-09 13:03:32 -07002305
Pekka Enberg343e0d72006-02-01 03:05:50 -08002306static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002307{
2308#ifdef CONFIG_SMP
2309 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002310 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002311#endif
2312}
2313
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314#else
2315#define check_irq_off() do { } while(0)
2316#define check_irq_on() do { } while(0)
2317#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002318#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319#endif
2320
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002321static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002322 struct array_cache *ac,
2323 int force, int node);
2324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325static void do_drain(void *arg)
2326{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002327 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002329 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002330 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002331 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332
2333 check_irq_off();
Pekka Enberg9a2dba4b2006-02-01 03:05:49 -08002334 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002335 n = get_node(cachep, node);
2336 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002337 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002338 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002339 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 ac->avail = 0;
2341}
2342
Pekka Enberg343e0d72006-02-01 03:05:50 -08002343static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002345 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002346 int node;
2347
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002348 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002350 for_each_kmem_cache_node(cachep, node, n)
2351 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002352 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002353
Christoph Lameter18bf8542014-08-06 16:04:11 -07002354 for_each_kmem_cache_node(cachep, node, n)
2355 drain_array(cachep, n, n->shared, 1, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356}
2357
Christoph Lametered11d9e2006-06-30 01:55:45 -07002358/*
2359 * Remove slabs from the list of free slabs.
2360 * Specify the number of slabs to drain in tofree.
2361 *
2362 * Returns the actual number of slabs released.
2363 */
2364static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002365 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002367 struct list_head *p;
2368 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002369 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370
Christoph Lametered11d9e2006-06-30 01:55:45 -07002371 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002372 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002374 spin_lock_irq(&n->list_lock);
2375 p = n->slabs_free.prev;
2376 if (p == &n->slabs_free) {
2377 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002378 goto out;
2379 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380
Joonsoo Kim8456a642013-10-24 10:07:49 +09002381 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002383 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002385 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002386 /*
2387 * Safe to drop the lock. The slab is no longer linked
2388 * to the cache.
2389 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002390 n->free_objects -= cache->num;
2391 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002392 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002393 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002395out:
2396 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397}
2398
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002399int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002400{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002401 int ret = 0;
2402 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002403 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002404
2405 drain_cpu_caches(cachep);
2406
2407 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002408 for_each_kmem_cache_node(cachep, node, n) {
Wanpeng Li0fa81032013-07-04 08:33:22 +08002409 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002410
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002411 ret += !list_empty(&n->slabs_full) ||
2412 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002413 }
2414 return (ret ? 1 : 0);
2415}
2416
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002417int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418{
Christoph Lameter12c36672012-09-04 23:38:33 +00002419 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002420 struct kmem_cache_node *n;
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002421 int rc = __kmem_cache_shrink(cachep, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
Christoph Lameter12c36672012-09-04 23:38:33 +00002423 if (rc)
2424 return rc;
2425
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002426 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002427
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002428 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002429 for_each_kmem_cache_node(cachep, i, n) {
2430 kfree(n->shared);
2431 free_alien_cache(n->alien);
2432 kfree(n);
2433 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002435 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002438/*
2439 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002440 *
2441 * For a slab cache when the slab descriptor is off-slab, the
2442 * slab descriptor can't come from the same cache which is being created,
2443 * Because if it is the case, that means we defer the creation of
2444 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2445 * And we eventually call down to __kmem_cache_create(), which
2446 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2447 * This is a "chicken-and-egg" problem.
2448 *
2449 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2450 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002451 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002452static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002453 struct page *page, int colour_off,
2454 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002456 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002457 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002458
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 if (OFF_SLAB(cachep)) {
2460 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002461 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002462 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002463 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 return NULL;
2465 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002466 freelist = addr + colour_off;
2467 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002469 page->active = 0;
2470 page->s_mem = addr + colour_off;
2471 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472}
2473
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002474static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002476 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002477}
2478
2479static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002480 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002481{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002482 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483}
2484
Pekka Enberg343e0d72006-02-01 03:05:50 -08002485static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002486 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487{
2488 int i;
2489
2490 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002491 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492#if DEBUG
2493 /* need to poison the objs? */
2494 if (cachep->flags & SLAB_POISON)
2495 poison_obj(cachep, objp, POISON_FREE);
2496 if (cachep->flags & SLAB_STORE_USER)
2497 *dbg_userword(cachep, objp) = NULL;
2498
2499 if (cachep->flags & SLAB_RED_ZONE) {
2500 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2501 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2502 }
2503 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002504 * Constructors are not allowed to allocate memory from the same
2505 * cache which they are a constructor for. Otherwise, deadlock.
2506 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 */
2508 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002509 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510
2511 if (cachep->flags & SLAB_RED_ZONE) {
2512 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2513 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002514 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2516 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002517 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002519 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002520 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002521 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002522 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523#else
2524 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002525 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002527 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002528 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530}
2531
Pekka Enberg343e0d72006-02-01 03:05:50 -08002532static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002534 if (CONFIG_ZONE_DMA_FLAG) {
2535 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002536 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002537 else
Glauber Costaa618e892012-06-14 16:17:21 +04002538 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002539 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540}
2541
Joonsoo Kim8456a642013-10-24 10:07:49 +09002542static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002543 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002544{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002545 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002546
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002547 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002548 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002549#if DEBUG
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#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002552
2553 return objp;
2554}
2555
Joonsoo Kim8456a642013-10-24 10:07:49 +09002556static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002557 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002558{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002559 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002560#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002561 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002562
Matthew Dobson78d382d2006-02-01 03:05:47 -08002563 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002564 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002565
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002566 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002567 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002568 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002569 printk(KERN_ERR "slab: double free detected in cache "
2570 "'%s', objp %p\n", cachep->name, objp);
2571 BUG();
2572 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002573 }
2574#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002575 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002576 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002577}
2578
Pekka Enberg47768742006-06-23 02:03:07 -07002579/*
2580 * Map pages beginning at addr to the given cache and slab. This is required
2581 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002582 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002583 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002584static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002585 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002587 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002588 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589}
2590
2591/*
2592 * Grow (by 1) the number of slabs within a cache. This is called by
2593 * kmem_cache_alloc() when there are no active objs left in a cache.
2594 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002595static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002596 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002598 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002599 size_t offset;
2600 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002601 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602
Andrew Mortona737b3e2006-03-22 00:08:11 -08002603 /*
2604 * Be lazy and only check for valid flags here, keeping it out of the
2605 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002607 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2608 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2609 BUG();
2610 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002611 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002613 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002615 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002616 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617
2618 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002619 offset = n->colour_next;
2620 n->colour_next++;
2621 if (n->colour_next >= cachep->colour)
2622 n->colour_next = 0;
2623 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002625 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626
2627 if (local_flags & __GFP_WAIT)
2628 local_irq_enable();
2629
2630 /*
2631 * The test for missing atomic flag is performed here, rather than
2632 * the more obvious place, simply to reduce the critical path length
2633 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2634 * will eventually be caught here (where it matters).
2635 */
2636 kmem_flagcheck(cachep, flags);
2637
Andrew Mortona737b3e2006-03-22 00:08:11 -08002638 /*
2639 * Get mem for the objs. Attempt to allocate a physical page from
2640 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002641 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002642 if (!page)
2643 page = kmem_getpages(cachep, local_flags, nodeid);
2644 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 goto failed;
2646
2647 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002648 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002649 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002650 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 goto opps1;
2652
Joonsoo Kim8456a642013-10-24 10:07:49 +09002653 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654
Joonsoo Kim8456a642013-10-24 10:07:49 +09002655 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656
2657 if (local_flags & __GFP_WAIT)
2658 local_irq_disable();
2659 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002660 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661
2662 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002663 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002665 n->free_objects += cachep->num;
2666 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002668opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002669 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002670failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 if (local_flags & __GFP_WAIT)
2672 local_irq_disable();
2673 return 0;
2674}
2675
2676#if DEBUG
2677
2678/*
2679 * Perform extra freeing checks:
2680 * - detect bad pointers.
2681 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682 */
2683static void kfree_debugcheck(const void *objp)
2684{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 if (!virt_addr_valid(objp)) {
2686 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002687 (unsigned long)objp);
2688 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690}
2691
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002692static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2693{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002694 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002695
2696 redzone1 = *dbg_redzone1(cache, obj);
2697 redzone2 = *dbg_redzone2(cache, obj);
2698
2699 /*
2700 * Redzone is ok.
2701 */
2702 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2703 return;
2704
2705 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2706 slab_error(cache, "double free detected");
2707 else
2708 slab_error(cache, "memory outside object was overwritten");
2709
David Woodhouseb46b8f12007-05-08 00:22:59 -07002710 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002711 obj, redzone1, redzone2);
2712}
2713
Pekka Enberg343e0d72006-02-01 03:05:50 -08002714static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002715 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002718 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002720 BUG_ON(virt_to_cache(objp) != cachep);
2721
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002722 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002724 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002727 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2729 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2730 }
2731 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002732 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
Joonsoo Kim8456a642013-10-24 10:07:49 +09002734 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735
2736 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002737 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738
Joonsoo Kim03787302014-06-23 13:22:06 -07002739 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 if (cachep->flags & SLAB_POISON) {
2741#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002742 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002743 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002744 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002745 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 } else {
2747 poison_obj(cachep, objp, POISON_FREE);
2748 }
2749#else
2750 poison_obj(cachep, objp, POISON_FREE);
2751#endif
2752 }
2753 return objp;
2754}
2755
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756#else
2757#define kfree_debugcheck(x) do { } while(0)
2758#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759#endif
2760
Mel Gorman072bb0a2012-07-31 16:43:58 -07002761static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2762 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
2764 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002765 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002767 int node;
2768
Joe Korty6d2144d2008-03-05 15:04:59 -08002769 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002770 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002771 if (unlikely(force_refill))
2772 goto force_grow;
2773retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002774 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 batchcount = ac->batchcount;
2776 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002777 /*
2778 * If there was little recent activity on this cache, then
2779 * perform only a partial refill. Otherwise we could generate
2780 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 */
2782 batchcount = BATCHREFILL_LIMIT;
2783 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07002784 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002786 BUG_ON(ac->avail > 0 || !n);
2787 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002788
Christoph Lameter3ded1752006-03-25 03:06:44 -08002789 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002790 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2791 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002792 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002793 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002794
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 while (batchcount > 0) {
2796 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002797 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002799 entry = n->slabs_partial.next;
2800 if (entry == &n->slabs_partial) {
2801 n->free_touched = 1;
2802 entry = n->slabs_free.next;
2803 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 goto must_grow;
2805 }
2806
Joonsoo Kim8456a642013-10-24 10:07:49 +09002807 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002809
2810 /*
2811 * The slab was either on partial or free list so
2812 * there must be at least one object available for
2813 * allocation.
2814 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002815 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002816
Joonsoo Kim8456a642013-10-24 10:07:49 +09002817 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 STATS_INC_ALLOCED(cachep);
2819 STATS_INC_ACTIVE(cachep);
2820 STATS_SET_HIGH(cachep);
2821
Joonsoo Kim8456a642013-10-24 10:07:49 +09002822 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002823 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825
2826 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002827 list_del(&page->lru);
2828 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002829 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002831 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 }
2833
Andrew Mortona737b3e2006-03-22 00:08:11 -08002834must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002835 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002836alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002837 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838
2839 if (unlikely(!ac->avail)) {
2840 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002841force_grow:
David Rientjes4167e9b2015-04-14 15:46:55 -07002842 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002843
Andrew Mortona737b3e2006-03-22 00:08:11 -08002844 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba4b2006-02-01 03:05:49 -08002845 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002846 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002847
2848 /* no objects in sight? abort */
2849 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 return NULL;
2851
Andrew Mortona737b3e2006-03-22 00:08:11 -08002852 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 goto retry;
2854 }
2855 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002856
2857 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858}
2859
Andrew Mortona737b3e2006-03-22 00:08:11 -08002860static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2861 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862{
2863 might_sleep_if(flags & __GFP_WAIT);
2864#if DEBUG
2865 kmem_flagcheck(cachep, flags);
2866#endif
2867}
2868
2869#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002870static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002871 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872{
Joonsoo Kim03787302014-06-23 13:22:06 -07002873 struct page *page;
2874
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002875 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002877 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002879 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002880 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002881 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 else
2883 check_poison_obj(cachep, objp);
2884#else
2885 check_poison_obj(cachep, objp);
2886#endif
2887 poison_obj(cachep, objp, POISON_INUSE);
2888 }
2889 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002890 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891
2892 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002893 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2894 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2895 slab_error(cachep, "double free, or memory outside"
2896 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002897 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002898 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002899 objp, *dbg_redzone1(cachep, objp),
2900 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 }
2902 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2903 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2904 }
Joonsoo Kim03787302014-06-23 13:22:06 -07002905
2906 page = virt_to_head_page(objp);
2907 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002908 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002909 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002910 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002911 if (ARCH_SLAB_MINALIGN &&
2912 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002913 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002914 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 return objp;
2917}
2918#else
2919#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2920#endif
2921
Akinobu Mita773ff602008-12-23 19:37:01 +09002922static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002923{
Joonsoo Kim8a9c61d2014-08-06 16:04:20 -07002924 if (unlikely(cachep == kmem_cache))
Akinobu Mita773ff602008-12-23 19:37:01 +09002925 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002926
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002927 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002928}
2929
Pekka Enberg343e0d72006-02-01 03:05:50 -08002930static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002934 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935
Alok N Kataria5c382302005-09-27 21:45:46 -07002936 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002937
Pekka Enberg9a2dba4b2006-02-01 03:05:49 -08002938 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002941 objp = ac_get_obj(cachep, ac, flags, false);
2942
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002943 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002944 * Allow for the possibility all avail objects are not allowed
2945 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002946 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07002947 if (objp) {
2948 STATS_INC_ALLOCHIT(cachep);
2949 goto out;
2950 }
2951 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07002953
2954 STATS_INC_ALLOCMISS(cachep);
2955 objp = cache_alloc_refill(cachep, flags, force_refill);
2956 /*
2957 * the 'ac' may be updated by cache_alloc_refill(),
2958 * and kmemleak_erase() requires its correct value.
2959 */
2960 ac = cpu_cache_get(cachep);
2961
2962out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01002963 /*
2964 * To avoid a false negative, if an object that is in one of the
2965 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2966 * treat the array pointers as a reference to the object.
2967 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09002968 if (objp)
2969 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07002970 return objp;
2971}
2972
Christoph Lametere498be72005-09-09 13:03:32 -07002973#ifdef CONFIG_NUMA
2974/*
Zefan Li2ad654b2014-09-25 09:41:02 +08002975 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002976 *
2977 * If we are in_interrupt, then process context, including cpusets and
2978 * mempolicy, may not apply and should not be used for allocation policy.
2979 */
2980static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2981{
2982 int nid_alloc, nid_here;
2983
Christoph Lameter765c4502006-09-27 01:50:08 -07002984 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08002985 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002986 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002987 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07002988 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002989 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07002990 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002991 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08002992 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08002993 return NULL;
2994}
2995
2996/*
Christoph Lameter765c4502006-09-27 01:50:08 -07002997 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08002998 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00002999 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003000 * perform an allocation without specifying a node. This allows the page
3001 * allocator to do its reclaim / fallback magic. We then insert the
3002 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003003 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003004static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003005{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003006 struct zonelist *zonelist;
3007 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003008 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003009 struct zone *zone;
3010 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003011 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003012 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003013 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003014
3015 if (flags & __GFP_THISNODE)
3016 return NULL;
3017
Christoph Lameter6cb06222007-10-16 01:25:41 -07003018 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003019
Mel Gormancc9a6c82012-03-21 16:34:11 -07003020retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003021 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003022 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003023
Christoph Lameter3c517a62006-12-06 20:33:29 -08003024retry:
3025 /*
3026 * Look through allowed nodes for objects available
3027 * from existing per node queues.
3028 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003029 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3030 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003031
Vladimir Davydov061d7072014-12-12 16:58:25 -08003032 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003033 get_node(cache, nid) &&
3034 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003035 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003036 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003037 if (obj)
3038 break;
3039 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003040 }
3041
Christoph Lametercfce6602007-05-06 14:50:17 -07003042 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003043 /*
3044 * This allocation will be performed within the constraints
3045 * of the current cpuset / memory policy requirements.
3046 * We may trigger various forms of reclaim on the allowed
3047 * set and go into memory reserves if necessary.
3048 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003049 struct page *page;
3050
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003051 if (local_flags & __GFP_WAIT)
3052 local_irq_enable();
3053 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003054 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003055 if (local_flags & __GFP_WAIT)
3056 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003057 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003058 /*
3059 * Insert into the appropriate per node queues
3060 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003061 nid = page_to_nid(page);
3062 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003063 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003064 gfp_exact_node(flags), nid);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003065 if (!obj)
3066 /*
3067 * Another processor may allocate the
3068 * objects in the slab since we are
3069 * not holding any locks.
3070 */
3071 goto retry;
3072 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003073 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003074 obj = NULL;
3075 }
3076 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003077 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003078
Mel Gormand26914d2014-04-03 14:47:24 -07003079 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003080 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003081 return obj;
3082}
3083
3084/*
Christoph Lametere498be72005-09-09 13:03:32 -07003085 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003087static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003088 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003089{
3090 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003091 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003092 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003093 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003094 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003096 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003097 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003098 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003099
Andrew Mortona737b3e2006-03-22 00:08:11 -08003100retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003101 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003102 spin_lock(&n->list_lock);
3103 entry = n->slabs_partial.next;
3104 if (entry == &n->slabs_partial) {
3105 n->free_touched = 1;
3106 entry = n->slabs_free.next;
3107 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 goto must_grow;
3109 }
Christoph Lametere498be72005-09-09 13:03:32 -07003110
Joonsoo Kim8456a642013-10-24 10:07:49 +09003111 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003112 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003113
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003114 STATS_INC_NODEALLOCS(cachep);
3115 STATS_INC_ACTIVE(cachep);
3116 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003117
Joonsoo Kim8456a642013-10-24 10:07:49 +09003118 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003119
Joonsoo Kim8456a642013-10-24 10:07:49 +09003120 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003121 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003122 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003123 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003124
Joonsoo Kim8456a642013-10-24 10:07:49 +09003125 if (page->active == cachep->num)
3126 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003127 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003128 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003129
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003130 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003131 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003132
Andrew Mortona737b3e2006-03-22 00:08:11 -08003133must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003134 spin_unlock(&n->list_lock);
David Rientjes4167e9b2015-04-14 15:46:55 -07003135 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003136 if (x)
3137 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003138
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003139 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003140
Andrew Mortona737b3e2006-03-22 00:08:11 -08003141done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003142 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003143}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003144
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003145static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003146slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003147 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003148{
3149 unsigned long save_flags;
3150 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003151 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003152
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003153 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003154
Nick Piggincf40bd12009-01-21 08:12:39 +01003155 lockdep_trace_alloc(flags);
3156
Akinobu Mita773ff602008-12-23 19:37:01 +09003157 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003158 return NULL;
3159
Glauber Costad79923f2012-12-18 14:22:48 -08003160 cachep = memcg_kmem_get_cache(cachep, flags);
3161
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003162 cache_alloc_debugcheck_before(cachep, flags);
3163 local_irq_save(save_flags);
3164
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003165 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003166 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003167
Christoph Lameter18bf8542014-08-06 16:04:11 -07003168 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003169 /* Node not bootstrapped yet */
3170 ptr = fallback_alloc(cachep, flags);
3171 goto out;
3172 }
3173
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003174 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003175 /*
3176 * Use the locally cached objects if possible.
3177 * However ____cache_alloc does not allow fallback
3178 * to other nodes. It may fail while we still have
3179 * objects on other nodes available.
3180 */
3181 ptr = ____cache_alloc(cachep, flags);
3182 if (ptr)
3183 goto out;
3184 }
3185 /* ___cache_alloc_node can fall back to other nodes */
3186 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3187 out:
3188 local_irq_restore(save_flags);
3189 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003190 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003191 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003192
Joe Perches5087c822013-09-10 17:02:51 -07003193 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003194 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003195 if (unlikely(flags & __GFP_ZERO))
3196 memset(ptr, 0, cachep->object_size);
3197 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003198
Vladimir Davydov8135be52014-12-12 16:56:38 -08003199 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003200 return ptr;
3201}
3202
3203static __always_inline void *
3204__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3205{
3206 void *objp;
3207
Zefan Li2ad654b2014-09-25 09:41:02 +08003208 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003209 objp = alternate_node_alloc(cache, flags);
3210 if (objp)
3211 goto out;
3212 }
3213 objp = ____cache_alloc(cache, flags);
3214
3215 /*
3216 * We may just have run out of memory on the local node.
3217 * ____cache_alloc_node() knows how to locate memory on other nodes
3218 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003219 if (!objp)
3220 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003221
3222 out:
3223 return objp;
3224}
3225#else
3226
3227static __always_inline void *
3228__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3229{
3230 return ____cache_alloc(cachep, flags);
3231}
3232
3233#endif /* CONFIG_NUMA */
3234
3235static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003236slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003237{
3238 unsigned long save_flags;
3239 void *objp;
3240
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003241 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003242
Nick Piggincf40bd12009-01-21 08:12:39 +01003243 lockdep_trace_alloc(flags);
3244
Akinobu Mita773ff602008-12-23 19:37:01 +09003245 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003246 return NULL;
3247
Glauber Costad79923f2012-12-18 14:22:48 -08003248 cachep = memcg_kmem_get_cache(cachep, flags);
3249
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003250 cache_alloc_debugcheck_before(cachep, flags);
3251 local_irq_save(save_flags);
3252 objp = __do_cache_alloc(cachep, flags);
3253 local_irq_restore(save_flags);
3254 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003255 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003256 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003257 prefetchw(objp);
3258
Joe Perches5087c822013-09-10 17:02:51 -07003259 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003260 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003261 if (unlikely(flags & __GFP_ZERO))
3262 memset(objp, 0, cachep->object_size);
3263 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003264
Vladimir Davydov8135be52014-12-12 16:56:38 -08003265 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003266 return objp;
3267}
Christoph Lametere498be72005-09-09 13:03:32 -07003268
3269/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003270 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003271 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003272 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003273static void free_block(struct kmem_cache *cachep, void **objpp,
3274 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275{
3276 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003277 struct kmem_cache_node *n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278
3279 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003280 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003281 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282
Mel Gorman072bb0a2012-07-31 16:43:58 -07003283 clear_obj_pfmemalloc(&objpp[i]);
3284 objp = objpp[i];
3285
Joonsoo Kim8456a642013-10-24 10:07:49 +09003286 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003287 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003288 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003289 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003291 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292
3293 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003294 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003295 if (n->free_objects > n->free_limit) {
3296 n->free_objects -= cachep->num;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003297 list_add_tail(&page->lru, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003299 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 }
3301 } else {
3302 /* Unconditionally move a slab to the end of the
3303 * partial list on free - maximum time for the
3304 * other objects to be freed, too.
3305 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003306 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307 }
3308 }
3309}
3310
Pekka Enberg343e0d72006-02-01 03:05:50 -08003311static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312{
3313 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003314 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003315 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003316 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317
3318 batchcount = ac->batchcount;
3319#if DEBUG
3320 BUG_ON(!batchcount || batchcount > ac->avail);
3321#endif
3322 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003323 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003324 spin_lock(&n->list_lock);
3325 if (n->shared) {
3326 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003327 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 if (max) {
3329 if (batchcount > max)
3330 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003331 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003332 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 shared_array->avail += batchcount;
3334 goto free_done;
3335 }
3336 }
3337
Joonsoo Kim97654df2014-08-06 16:04:25 -07003338 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003339free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340#if STATS
3341 {
3342 int i = 0;
3343 struct list_head *p;
3344
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003345 p = n->slabs_free.next;
3346 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003347 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348
Joonsoo Kim8456a642013-10-24 10:07:49 +09003349 page = list_entry(p, struct page, lru);
3350 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351
3352 i++;
3353 p = p->next;
3354 }
3355 STATS_SET_FREEABLE(cachep, i);
3356 }
3357#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003358 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003359 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003361 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362}
3363
3364/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003365 * Release an obj back to its cache. If the obj has a constructed state, it must
3366 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003368static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003369 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370{
Pekka Enberg9a2dba4b2006-02-01 03:05:49 -08003371 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372
3373 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003374 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003375 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003377 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003378
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003379 /*
3380 * Skip calling cache_free_alien() when the platform is not numa.
3381 * This will avoid cache misses that happen while accessing slabp (which
3382 * is per page memory reference) to get nodeid. Instead use a global
3383 * variable to skip the call, which is mostly likely to be present in
3384 * the cache.
3385 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003386 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003387 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003388
Joonsoo Kim3d880192014-10-09 15:26:04 -07003389 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 } else {
3392 STATS_INC_FREEMISS(cachep);
3393 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003395
Mel Gorman072bb0a2012-07-31 16:43:58 -07003396 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397}
3398
3399/**
3400 * kmem_cache_alloc - Allocate an object
3401 * @cachep: The cache to allocate from.
3402 * @flags: See kmalloc().
3403 *
3404 * Allocate an object from this cache. The flags are only relevant
3405 * if the cache has no available objects.
3406 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003407void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003409 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003410
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003411 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003412 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003413
3414 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415}
3416EXPORT_SYMBOL(kmem_cache_alloc);
3417
Li Zefan0f24f122009-12-11 15:45:30 +08003418#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003419void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003420kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003421{
Steven Rostedt85beb582010-11-24 16:23:34 -05003422 void *ret;
3423
Ezequiel Garcia48356302012-09-08 17:47:57 -03003424 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003425
3426 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003427 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003428 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003429}
Steven Rostedt85beb582010-11-24 16:23:34 -05003430EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003431#endif
3432
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003434/**
3435 * kmem_cache_alloc_node - Allocate an object on the specified node
3436 * @cachep: The cache to allocate from.
3437 * @flags: See kmalloc().
3438 * @nodeid: node number of the target node.
3439 *
3440 * Identical to kmem_cache_alloc but it will allocate memory on the given
3441 * node, which can improve the performance for cpu bound structures.
3442 *
3443 * Fallback to other node is possible if __GFP_THISNODE is not set.
3444 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003445void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3446{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003447 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003448
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003449 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003450 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003451 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003452
3453 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003454}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455EXPORT_SYMBOL(kmem_cache_alloc_node);
3456
Li Zefan0f24f122009-12-11 15:45:30 +08003457#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003458void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003459 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003460 int nodeid,
3461 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003462{
Steven Rostedt85beb582010-11-24 16:23:34 -05003463 void *ret;
3464
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003465 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003466
Steven Rostedt85beb582010-11-24 16:23:34 -05003467 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003468 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003469 flags, nodeid);
3470 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003471}
Steven Rostedt85beb582010-11-24 16:23:34 -05003472EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003473#endif
3474
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003475static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003476__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003477{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003478 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003479
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003480 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003481 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3482 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003483 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003484}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003485
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003486void *__kmalloc_node(size_t size, gfp_t flags, int node)
3487{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003488 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003489}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003490EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003491
3492void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003493 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003494{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003495 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003496}
3497EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003498#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499
3500/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003501 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003503 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003504 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003506static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003507 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003509 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003510 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003512 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003513 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3514 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003515 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003516
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003517 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003518 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003519
3520 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003521}
3522
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003523void *__kmalloc(size_t size, gfp_t flags)
3524{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003525 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526}
3527EXPORT_SYMBOL(__kmalloc);
3528
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003529void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003530{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003531 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003532}
3533EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003534
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535/**
3536 * kmem_cache_free - Deallocate an object
3537 * @cachep: The cache the allocation was from.
3538 * @objp: The previously allocated object.
3539 *
3540 * Free an object which was previously allocated from this
3541 * cache.
3542 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003543void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544{
3545 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003546 cachep = cache_from_obj(cachep, objp);
3547 if (!cachep)
3548 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549
3550 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003551 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003552 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003553 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003554 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003556
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003557 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558}
3559EXPORT_SYMBOL(kmem_cache_free);
3560
3561/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 * kfree - free previously allocated memory
3563 * @objp: pointer returned by kmalloc.
3564 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003565 * If @objp is NULL, no operation is performed.
3566 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 * Don't free memory not originally allocated by kmalloc()
3568 * or you will run into trouble.
3569 */
3570void kfree(const void *objp)
3571{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003572 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 unsigned long flags;
3574
Pekka Enberg2121db72009-03-25 11:05:57 +02003575 trace_kfree(_RET_IP_, objp);
3576
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003577 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 return;
3579 local_irq_save(flags);
3580 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003581 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003582 debug_check_no_locks_freed(objp, c->object_size);
3583
3584 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003585 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 local_irq_restore(flags);
3587}
3588EXPORT_SYMBOL(kfree);
3589
Christoph Lametere498be72005-09-09 13:03:32 -07003590/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003591 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003592 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003593static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003594{
3595 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003596 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003597 struct array_cache *new_shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07003598 struct alien_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003599
Mel Gorman9c09a952008-01-24 05:49:54 -08003600 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003601
LQYMGTb455def2014-12-10 15:42:13 -08003602 if (use_alien_caches) {
3603 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3604 if (!new_alien)
3605 goto fail;
3606 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003607
Eric Dumazet63109842007-05-06 14:49:28 -07003608 new_shared = NULL;
3609 if (cachep->shared) {
3610 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003611 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003612 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003613 if (!new_shared) {
3614 free_alien_cache(new_alien);
3615 goto fail;
3616 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003617 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003618
Christoph Lameter18bf8542014-08-06 16:04:11 -07003619 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003620 if (n) {
3621 struct array_cache *shared = n->shared;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003622 LIST_HEAD(list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003623
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003624 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003625
Christoph Lametercafeb022006-03-25 03:06:46 -08003626 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003627 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07003628 shared->avail, node, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003629
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003630 n->shared = new_shared;
3631 if (!n->alien) {
3632 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003633 new_alien = NULL;
3634 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003635 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003636 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003637 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003638 slabs_destroy(cachep, &list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003639 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003640 free_alien_cache(new_alien);
3641 continue;
3642 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003643 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3644 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003645 free_alien_cache(new_alien);
3646 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003647 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003648 }
Christoph Lametere498be72005-09-09 13:03:32 -07003649
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003650 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003651 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3652 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003653 n->shared = new_shared;
3654 n->alien = new_alien;
3655 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003656 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003657 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003658 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003659 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003660
Andrew Mortona737b3e2006-03-22 00:08:11 -08003661fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003662 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003663 /* Cache is not active yet. Roll back what we did */
3664 node--;
3665 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003666 n = get_node(cachep, node);
3667 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003668 kfree(n->shared);
3669 free_alien_cache(n->alien);
3670 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003671 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003672 }
3673 node--;
3674 }
3675 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003676 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003677}
3678
Christoph Lameter18004c52012-07-06 15:25:12 -05003679/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003680static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003681 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003683 struct array_cache __percpu *cpu_cache, *prev;
3684 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003686 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3687 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003688 return -ENOMEM;
3689
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003690 prev = cachep->cpu_cache;
3691 cachep->cpu_cache = cpu_cache;
3692 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003693
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 cachep->batchcount = batchcount;
3696 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003697 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003699 if (!prev)
3700 goto alloc_node;
3701
3702 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003703 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003704 int node;
3705 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003706 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003707
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003708 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003709 n = get_node(cachep, node);
3710 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003711 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003712 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003713 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003715 free_percpu(prev);
3716
3717alloc_node:
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003718 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719}
3720
Glauber Costa943a4512012-12-18 14:23:03 -08003721static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3722 int batchcount, int shared, gfp_t gfp)
3723{
3724 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003725 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003726
3727 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3728
3729 if (slab_state < FULL)
3730 return ret;
3731
3732 if ((ret < 0) || !is_root_cache(cachep))
3733 return ret;
3734
Vladimir Davydov426589f2015-02-12 14:59:23 -08003735 lockdep_assert_held(&slab_mutex);
3736 for_each_memcg_cache(c, cachep) {
3737 /* return value determined by the root cache only */
3738 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003739 }
3740
3741 return ret;
3742}
3743
Christoph Lameter18004c52012-07-06 15:25:12 -05003744/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003745static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746{
3747 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003748 int limit = 0;
3749 int shared = 0;
3750 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751
Glauber Costa943a4512012-12-18 14:23:03 -08003752 if (!is_root_cache(cachep)) {
3753 struct kmem_cache *root = memcg_root_cache(cachep);
3754 limit = root->limit;
3755 shared = root->shared;
3756 batchcount = root->batchcount;
3757 }
3758
3759 if (limit && shared && batchcount)
3760 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003761 /*
3762 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 * - create a LIFO ordering, i.e. return objects that are cache-warm
3764 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003765 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 * bufctl chains: array operations are cheaper.
3767 * The numbers are guessed, we should auto-tune as described by
3768 * Bonwick.
3769 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003770 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003772 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003774 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003776 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 limit = 54;
3778 else
3779 limit = 120;
3780
Andrew Mortona737b3e2006-03-22 00:08:11 -08003781 /*
3782 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 * allocation behaviour: Most allocs on one cpu, most free operations
3784 * on another cpu. For these cases, an efficient object passing between
3785 * cpus is necessary. This is provided by a shared array. The array
3786 * replaces Bonwick's magazine layer.
3787 * On uniprocessor, it's functionally equivalent (but less efficient)
3788 * to a larger limit. Thus disabled by default.
3789 */
3790 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003791 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793
3794#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003795 /*
3796 * With debugging enabled, large batchcount lead to excessively long
3797 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 */
3799 if (limit > 32)
3800 limit = 32;
3801#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003802 batchcount = (limit + 1) / 2;
3803skip_setup:
3804 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 if (err)
3806 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003807 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003808 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809}
3810
Christoph Lameter1b552532006-03-22 00:09:07 -08003811/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003812 * Drain an array if it contains any elements taking the node lock only if
3813 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003814 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003815 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003816static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003817 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818{
Joonsoo Kim97654df2014-08-06 16:04:25 -07003819 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820 int tofree;
3821
Christoph Lameter1b552532006-03-22 00:09:07 -08003822 if (!ac || !ac->avail)
3823 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824 if (ac->touched && !force) {
3825 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003826 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003827 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003828 if (ac->avail) {
3829 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3830 if (tofree > ac->avail)
3831 tofree = (ac->avail + 1) / 2;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003832 free_block(cachep, ac->entry, tofree, node, &list);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003833 ac->avail -= tofree;
3834 memmove(ac->entry, &(ac->entry[tofree]),
3835 sizeof(void *) * ac->avail);
3836 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003837 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003838 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839 }
3840}
3841
3842/**
3843 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003844 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 *
3846 * Called from workqueue/eventd every few seconds.
3847 * Purpose:
3848 * - clear the per-cpu caches for this CPU.
3849 * - return freeable pages to the main free memory pool.
3850 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003851 * If we cannot acquire the cache chain mutex then just give up - we'll try
3852 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003854static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003856 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003857 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003858 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003859 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860
Christoph Lameter18004c52012-07-06 15:25:12 -05003861 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003863 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864
Christoph Lameter18004c52012-07-06 15:25:12 -05003865 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 check_irq_on();
3867
Christoph Lameter35386e32006-03-22 00:09:05 -08003868 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003869 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003870 * have established with reasonable certainty that
3871 * we can do some work if the lock was obtained.
3872 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07003873 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08003874
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003875 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003877 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878
Christoph Lameter35386e32006-03-22 00:09:05 -08003879 /*
3880 * These are racy checks but it does not matter
3881 * if we skip one check or scan twice.
3882 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003883 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003884 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003886 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003888 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003890 if (n->free_touched)
3891 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003892 else {
3893 int freed;
3894
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003895 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003896 5 * searchp->num - 1) / (5 * searchp->num));
3897 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003899next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900 cond_resched();
3901 }
3902 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05003903 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003904 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003905out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08003906 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003907 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908}
3909
Linus Torvalds158a9622008-01-02 13:04:48 -08003910#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04003911void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003913 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003914 unsigned long active_objs;
3915 unsigned long num_objs;
3916 unsigned long active_slabs = 0;
3917 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003918 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003920 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003921 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923 active_objs = 0;
3924 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003925 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07003926
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003927 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003928 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003929
Joonsoo Kim8456a642013-10-24 10:07:49 +09003930 list_for_each_entry(page, &n->slabs_full, lru) {
3931 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07003932 error = "slabs_full accounting error";
3933 active_objs += cachep->num;
3934 active_slabs++;
3935 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003936 list_for_each_entry(page, &n->slabs_partial, lru) {
3937 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003938 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003939 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003940 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003941 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07003942 active_slabs++;
3943 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003944 list_for_each_entry(page, &n->slabs_free, lru) {
3945 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003946 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07003947 num_slabs++;
3948 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003949 free_objects += n->free_objects;
3950 if (n->shared)
3951 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003952
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003953 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003955 num_slabs += active_slabs;
3956 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003957 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 error = "free_objects accounting error";
3959
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003960 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 if (error)
3962 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3963
Glauber Costa0d7561c2012-10-19 18:20:27 +04003964 sinfo->active_objs = active_objs;
3965 sinfo->num_objs = num_objs;
3966 sinfo->active_slabs = active_slabs;
3967 sinfo->num_slabs = num_slabs;
3968 sinfo->shared_avail = shared_avail;
3969 sinfo->limit = cachep->limit;
3970 sinfo->batchcount = cachep->batchcount;
3971 sinfo->shared = cachep->shared;
3972 sinfo->objects_per_slab = cachep->num;
3973 sinfo->cache_order = cachep->gfporder;
3974}
3975
3976void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3977{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003979 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 unsigned long high = cachep->high_mark;
3981 unsigned long allocs = cachep->num_allocations;
3982 unsigned long grown = cachep->grown;
3983 unsigned long reaped = cachep->reaped;
3984 unsigned long errors = cachep->errors;
3985 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003987 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003988 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989
Joe Perchese92dd4f2010-03-26 19:27:58 -07003990 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
3991 "%4lu %4lu %4lu %4lu %4lu",
3992 allocs, high, grown,
3993 reaped, errors, max_freeable, node_allocs,
3994 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 }
3996 /* cpu stats */
3997 {
3998 unsigned long allochit = atomic_read(&cachep->allochit);
3999 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4000 unsigned long freehit = atomic_read(&cachep->freehit);
4001 unsigned long freemiss = atomic_read(&cachep->freemiss);
4002
4003 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004004 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 }
4006#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007}
4008
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009#define MAX_SLABINFO_WRITE 128
4010/**
4011 * slabinfo_write - Tuning for the slab allocator
4012 * @file: unused
4013 * @buffer: user buffer
4014 * @count: data length
4015 * @ppos: unused
4016 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004017ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004018 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004020 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004022 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004023
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024 if (count > MAX_SLABINFO_WRITE)
4025 return -EINVAL;
4026 if (copy_from_user(&kbuf, buffer, count))
4027 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004028 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029
4030 tmp = strchr(kbuf, ' ');
4031 if (!tmp)
4032 return -EINVAL;
4033 *tmp = '\0';
4034 tmp++;
4035 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4036 return -EINVAL;
4037
4038 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004039 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004041 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004043 if (limit < 1 || batchcount < 1 ||
4044 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004045 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004047 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004048 batchcount, shared,
4049 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 }
4051 break;
4052 }
4053 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004054 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 if (res >= 0)
4056 res = count;
4057 return res;
4058}
Al Viro871751e2006-03-25 03:06:39 -08004059
4060#ifdef CONFIG_DEBUG_SLAB_LEAK
4061
Al Viro871751e2006-03-25 03:06:39 -08004062static inline int add_caller(unsigned long *n, unsigned long v)
4063{
4064 unsigned long *p;
4065 int l;
4066 if (!v)
4067 return 1;
4068 l = n[1];
4069 p = n + 2;
4070 while (l) {
4071 int i = l/2;
4072 unsigned long *q = p + 2 * i;
4073 if (*q == v) {
4074 q[1]++;
4075 return 1;
4076 }
4077 if (*q > v) {
4078 l = i;
4079 } else {
4080 p = q + 2;
4081 l -= i + 1;
4082 }
4083 }
4084 if (++n[1] == n[0])
4085 return 0;
4086 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4087 p[0] = v;
4088 p[1] = 1;
4089 return 1;
4090}
4091
Joonsoo Kim8456a642013-10-24 10:07:49 +09004092static void handle_slab(unsigned long *n, struct kmem_cache *c,
4093 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004094{
4095 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004096 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004097
Al Viro871751e2006-03-25 03:06:39 -08004098 if (n[0] == n[1])
4099 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004100 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004101 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004102 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004103
Al Viro871751e2006-03-25 03:06:39 -08004104 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4105 return;
4106 }
4107}
4108
4109static void show_symbol(struct seq_file *m, unsigned long address)
4110{
4111#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004112 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004113 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004114
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004115 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004116 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004117 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004118 seq_printf(m, " [%s]", modname);
4119 return;
4120 }
4121#endif
4122 seq_printf(m, "%p", (void *)address);
4123}
4124
4125static int leaks_show(struct seq_file *m, void *p)
4126{
Thierry Reding0672aa72012-06-22 19:42:49 +02004127 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004128 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004129 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004130 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004131 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004132 int node;
4133 int i;
4134
4135 if (!(cachep->flags & SLAB_STORE_USER))
4136 return 0;
4137 if (!(cachep->flags & SLAB_RED_ZONE))
4138 return 0;
4139
4140 /* OK, we can do it */
4141
Christoph Lameterdb845062013-02-05 18:45:23 +00004142 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004143
Christoph Lameter18bf8542014-08-06 16:04:11 -07004144 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004145
4146 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004147 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004148
Joonsoo Kim8456a642013-10-24 10:07:49 +09004149 list_for_each_entry(page, &n->slabs_full, lru)
4150 handle_slab(x, cachep, page);
4151 list_for_each_entry(page, &n->slabs_partial, lru)
4152 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004153 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004154 }
4155 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004156 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004157 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004158 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004159 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004160 if (!m->private) {
4161 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004162 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004163 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004164 return -ENOMEM;
4165 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004166 *(unsigned long *)m->private = x[0] * 2;
4167 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004168 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004169 /* Now make sure this entry will be retried */
4170 m->count = m->size;
4171 return 0;
4172 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004173 for (i = 0; i < x[1]; i++) {
4174 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4175 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004176 seq_putc(m, '\n');
4177 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004178
Al Viro871751e2006-03-25 03:06:39 -08004179 return 0;
4180}
4181
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004182static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004183 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004184 .next = slab_next,
4185 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004186 .show = leaks_show,
4187};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004188
4189static int slabstats_open(struct inode *inode, struct file *file)
4190{
Rob Jonesb208ce32014-10-09 15:28:03 -07004191 unsigned long *n;
4192
4193 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4194 if (!n)
4195 return -ENOMEM;
4196
4197 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4198
4199 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004200}
4201
4202static const struct file_operations proc_slabstats_operations = {
4203 .open = slabstats_open,
4204 .read = seq_read,
4205 .llseek = seq_lseek,
4206 .release = seq_release_private,
4207};
Al Viro871751e2006-03-25 03:06:39 -08004208#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004209
4210static int __init slab_proc_init(void)
4211{
4212#ifdef CONFIG_DEBUG_SLAB_LEAK
4213 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4214#endif
4215 return 0;
4216}
4217module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218#endif
4219
Manfred Spraul00e145b2005-09-03 15:55:07 -07004220/**
4221 * ksize - get the actual amount of memory allocated for a given object
4222 * @objp: Pointer to the object
4223 *
4224 * kmalloc may internally round up allocations and return more memory
4225 * than requested. ksize() can be used to determine the actual amount of
4226 * memory allocated. The caller may use this additional memory, even though
4227 * a smaller amount of memory was initially specified with the kmalloc call.
4228 * The caller must guarantee that objp points to a valid object previously
4229 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4230 * must not be freed during the duration of the call.
4231 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004232size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004234 BUG_ON(!objp);
4235 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004236 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004237
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004238 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004240EXPORT_SYMBOL(ksize);