blob: cc8bbc1e6bc9b6fe5d980ffe3837d7b4c8843d84 [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/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 * struct array_cache
173 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700188 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800192 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193};
194
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700195struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198};
199
Andrew Mortona737b3e2006-03-22 00:08:11 -0800200/*
Christoph Lametere498be72005-09-09 13:03:32 -0700201 * Need this for bootstrapping a per node allocator.
202 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700205#define CACHE_CACHE 0
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700206#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207
Christoph Lametered11d9e2006-06-30 01:55:45 -0700208static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000209 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700211 int node, struct list_head *list);
212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000214static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700215
Joonsoo Kim76b342b2016-05-19 17:10:26 -0700216static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
217 void **list);
218static inline void fixup_slab_list(struct kmem_cache *cachep,
219 struct kmem_cache_node *n, struct page *page,
220 void **list);
Ingo Molnare0a42722006-06-23 02:03:46 -0700221static int slab_early_init = 1;
222
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000223#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700224
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000225static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700226{
227 INIT_LIST_HEAD(&parent->slabs_full);
228 INIT_LIST_HEAD(&parent->slabs_partial);
229 INIT_LIST_HEAD(&parent->slabs_free);
230 parent->shared = NULL;
231 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800232 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700233 spin_lock_init(&parent->list_lock);
234 parent->free_objects = 0;
235 parent->free_touched = 0;
236}
237
Andrew Mortona737b3e2006-03-22 00:08:11 -0800238#define MAKE_LIST(cachep, listp, slab, nodeid) \
239 do { \
240 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700241 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700242 } while (0)
243
Andrew Mortona737b3e2006-03-22 00:08:11 -0800244#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
245 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
249 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700251#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252#define CFLGS_OFF_SLAB (0x80000000UL)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700253#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
255
256#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800257/*
258 * Optimization question: fewer reaps means less probability for unnessary
259 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100261 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 * which could lock up otherwise freeable slabs.
263 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800264#define REAPTIMEOUT_AC (2*HZ)
265#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267#if STATS
268#define STATS_INC_ACTIVE(x) ((x)->num_active++)
269#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
270#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
271#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700272#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273#define STATS_SET_HIGH(x) \
274 do { \
275 if ((x)->num_active > (x)->high_mark) \
276 (x)->high_mark = (x)->num_active; \
277 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278#define STATS_INC_ERR(x) ((x)->errors++)
279#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700280#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700281#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800282#define STATS_SET_FREEABLE(x, i) \
283 do { \
284 if ((x)->max_freeable < i) \
285 (x)->max_freeable = i; \
286 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
288#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
289#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
290#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
291#else
292#define STATS_INC_ACTIVE(x) do { } while (0)
293#define STATS_DEC_ACTIVE(x) do { } while (0)
294#define STATS_INC_ALLOCED(x) do { } while (0)
295#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700296#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297#define STATS_SET_HIGH(x) do { } while (0)
298#define STATS_INC_ERR(x) do { } while (0)
299#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700300#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700301#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800302#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303#define STATS_INC_ALLOCHIT(x) do { } while (0)
304#define STATS_INC_ALLOCMISS(x) do { } while (0)
305#define STATS_INC_FREEHIT(x) do { } while (0)
306#define STATS_INC_FREEMISS(x) do { } while (0)
307#endif
308
309#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Andrew Mortona737b3e2006-03-22 00:08:11 -0800311/*
312 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 * the end of an object is aligned with the end of the real
316 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800319 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
321 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800324static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800326 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327}
328
David Woodhouseb46b8f12007-05-08 00:22:59 -0700329static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330{
331 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700332 return (unsigned long long*) (objp + obj_offset(cachep) -
333 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334}
335
David Woodhouseb46b8f12007-05-08 00:22:59 -0700336static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337{
338 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
339 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500340 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700341 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400342 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500343 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700344 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345}
346
Pekka Enberg343e0d72006-02-01 03:05:50 -0800347static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
349 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500350 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351}
352
353#else
354
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800355#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700356#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
357#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
359
360#endif
361
Joonsoo Kim03787302014-06-23 13:22:06 -0700362#ifdef CONFIG_DEBUG_SLAB_LEAK
363
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700364static inline bool is_store_user_clean(struct kmem_cache *cachep)
Joonsoo Kim03787302014-06-23 13:22:06 -0700365{
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700366 return atomic_read(&cachep->store_user_clean) == 1;
367}
Joonsoo Kim03787302014-06-23 13:22:06 -0700368
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700369static inline void set_store_user_clean(struct kmem_cache *cachep)
370{
371 atomic_set(&cachep->store_user_clean, 1);
372}
Joonsoo Kim03787302014-06-23 13:22:06 -0700373
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700374static inline void set_store_user_dirty(struct kmem_cache *cachep)
375{
376 if (is_store_user_clean(cachep))
377 atomic_set(&cachep->store_user_clean, 0);
Joonsoo Kim03787302014-06-23 13:22:06 -0700378}
379
380#else
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700381static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
Joonsoo Kim03787302014-06-23 13:22:06 -0700382
383#endif
384
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700386 * Do not go above this order unless 0 objects fit into the slab or
387 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 */
David Rientjes543585c2011-10-18 22:09:24 -0700389#define SLAB_MAX_ORDER_HI 1
390#define SLAB_MAX_ORDER_LO 0
391static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700392static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800394static inline struct kmem_cache *virt_to_cache(const void *obj)
395{
Christoph Lameterb49af682007-05-06 14:49:41 -0700396 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500397 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800398}
399
Joonsoo Kim8456a642013-10-24 10:07:49 +0900400static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800401 unsigned int idx)
402{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900403 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800404}
405
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800406/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500407 * We want to avoid an expensive divide : (offset / cache->size)
408 * Using the fact that size is a constant for a particular cache,
409 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800410 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
411 */
412static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900413 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800414{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900415 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800416 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800417}
418
Joonsoo Kim6fb92432016-03-15 14:54:09 -0700419#define BOOT_CPUCACHE_ENTRIES 1
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000421static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800422 .batchcount = 1,
423 .limit = BOOT_CPUCACHE_ENTRIES,
424 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500425 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800426 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427};
428
Tejun Heo1871e522009-10-29 22:34:13 +0900429static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
Pekka Enberg343e0d72006-02-01 03:05:50 -0800431static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700433 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
Andrew Mortona737b3e2006-03-22 00:08:11 -0800436/*
437 * Calculate the number of objects and left-over bytes for a given buffer size.
438 */
Joonsoo Kim70f75062016-03-15 14:54:53 -0700439static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
440 unsigned long flags, size_t *left_over)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800441{
Joonsoo Kim70f75062016-03-15 14:54:53 -0700442 unsigned int num;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800443 size_t slab_size = PAGE_SIZE << gfporder;
444
445 /*
446 * The slab management structure can be either off the slab or
447 * on it. For the latter case, the memory allocated for a
448 * slab is used for:
449 *
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800450 * - @buffer_size bytes for each object
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700451 * - One freelist_idx_t for each object
452 *
453 * We don't need to consider alignment of freelist because
454 * freelist will be at the end of slab page. The objects will be
455 * at the correct alignment.
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800456 *
457 * If the slab management structure is off the slab, then the
458 * alignment will already be calculated into the size. Because
459 * the slabs are all pages aligned, the objects will be at the
460 * correct alignment when allocated.
461 */
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700462 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700463 num = slab_size / buffer_size;
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700464 *left_over = slab_size % buffer_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800465 } else {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700466 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700467 *left_over = slab_size %
468 (buffer_size + sizeof(freelist_idx_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 }
Joonsoo Kim70f75062016-03-15 14:54:53 -0700470
471 return num;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472}
473
Christoph Lameterf28510d2012-09-11 19:49:38 +0000474#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700475#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Andrew Mortona737b3e2006-03-22 00:08:11 -0800477static void __slab_error(const char *function, struct kmem_cache *cachep,
478 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Joe Perches11705322016-03-17 14:19:50 -0700480 pr_err("slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800481 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030483 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000485#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Paul Menage3395ee02006-12-06 20:32:16 -0800487/*
488 * By default on NUMA we use alien caches to stage the freeing of
489 * objects allocated from other nodes. This causes massive memory
490 * inefficiencies when using fake NUMA setup to split memory into a
491 * large number of small nodes, so it can be disabled on the command
492 * line
493 */
494
495static int use_alien_caches __read_mostly = 1;
496static int __init noaliencache_setup(char *s)
497{
498 use_alien_caches = 0;
499 return 1;
500}
501__setup("noaliencache", noaliencache_setup);
502
David Rientjes3df1ccc2011-10-18 22:09:28 -0700503static int __init slab_max_order_setup(char *str)
504{
505 get_option(&str, &slab_max_order);
506 slab_max_order = slab_max_order < 0 ? 0 :
507 min(slab_max_order, MAX_ORDER - 1);
508 slab_max_order_set = true;
509
510 return 1;
511}
512__setup("slab_max_order=", slab_max_order_setup);
513
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800514#ifdef CONFIG_NUMA
515/*
516 * Special reaping functions for NUMA systems called from cache_reap().
517 * These take care of doing round robin flushing of alien caches (containing
518 * objects freed on different nodes from which they were allocated) and the
519 * flushing of remote pcps by calling drain_node_pages.
520 */
Tejun Heo1871e522009-10-29 22:34:13 +0900521static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800522
523static void init_reap_node(int cpu)
524{
Andrew Morton0edaf862016-05-19 17:10:58 -0700525 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
526 node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800527}
528
529static void next_reap_node(void)
530{
Christoph Lameter909ea962010-12-08 16:22:55 +0100531 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800532
Andrew Morton0edaf862016-05-19 17:10:58 -0700533 node = next_node_in(node, node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100534 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800535}
536
537#else
538#define init_reap_node(cpu) do { } while (0)
539#define next_reap_node(void) do { } while (0)
540#endif
541
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542/*
543 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
544 * via the workqueue/eventd.
545 * Add the CPU number into the expiration time to minimize the possibility of
546 * the CPUs getting into lockstep and contending for the global cache chain
547 * lock.
548 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400549static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550{
Tejun Heo1871e522009-10-29 22:34:13 +0900551 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552
553 /*
554 * When this gets called from do_initcalls via cpucache_init(),
555 * init_workqueues() has already run, so keventd will be setup
556 * at that time.
557 */
David Howells52bad642006-11-22 14:54:01 +0000558 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800559 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700560 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800561 schedule_delayed_work_on(cpu, reap_work,
562 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 }
564}
565
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700566static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100568 /*
569 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300570 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100571 * cache the pointers are not cleared and they could be counted as
572 * valid references during a kmemleak scan. Therefore, kmemleak must
573 * not scan such objects.
574 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700575 kmemleak_no_scan(ac);
576 if (ac) {
577 ac->avail = 0;
578 ac->limit = limit;
579 ac->batchcount = batch;
580 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700582}
583
584static struct array_cache *alloc_arraycache(int node, int entries,
585 int batchcount, gfp_t gfp)
586{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700587 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700588 struct array_cache *ac = NULL;
589
590 ac = kmalloc_node(memsize, gfp, node);
591 init_arraycache(ac, entries, batchcount);
592 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700595static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
596 struct page *page, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700597{
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700598 struct kmem_cache_node *n;
599 int page_node;
600 LIST_HEAD(list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700601
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700602 page_node = page_to_nid(page);
603 n = get_node(cachep, page_node);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700604
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700605 spin_lock(&n->list_lock);
606 free_block(cachep, &objp, 1, page_node, &list);
607 spin_unlock(&n->list_lock);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700608
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700609 slabs_destroy(cachep, &list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700610}
611
Christoph Lameter3ded1752006-03-25 03:06:44 -0800612/*
613 * Transfer objects in one arraycache to another.
614 * Locking must be handled by the caller.
615 *
616 * Return the number of entries transferred.
617 */
618static int transfer_objects(struct array_cache *to,
619 struct array_cache *from, unsigned int max)
620{
621 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700622 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800623
624 if (!nr)
625 return 0;
626
627 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
628 sizeof(void *) *nr);
629
630 from->avail -= nr;
631 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800632 return nr;
633}
634
Christoph Lameter765c4502006-09-27 01:50:08 -0700635#ifndef CONFIG_NUMA
636
637#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000638#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700639
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700640static inline struct alien_cache **alloc_alien_cache(int node,
641 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700642{
Joonsoo Kim88881772016-05-19 17:10:05 -0700643 return NULL;
Christoph Lameter765c4502006-09-27 01:50:08 -0700644}
645
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700646static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700647{
648}
649
650static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
651{
652 return 0;
653}
654
655static inline void *alternate_node_alloc(struct kmem_cache *cachep,
656 gfp_t flags)
657{
658 return NULL;
659}
660
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800661static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700662 gfp_t flags, int nodeid)
663{
664 return NULL;
665}
666
David Rientjes4167e9b2015-04-14 15:46:55 -0700667static inline gfp_t gfp_exact_node(gfp_t flags)
668{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700669 return flags & ~__GFP_NOFAIL;
David Rientjes4167e9b2015-04-14 15:46:55 -0700670}
671
Christoph Lameter765c4502006-09-27 01:50:08 -0700672#else /* CONFIG_NUMA */
673
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800674static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800675static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800676
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700677static struct alien_cache *__alloc_alien_cache(int node, int entries,
678 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700679{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700680 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700681 struct alien_cache *alc = NULL;
682
683 alc = kmalloc_node(memsize, gfp, node);
684 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700685 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700686 return alc;
687}
688
689static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
690{
691 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700692 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700693 int i;
694
695 if (limit > 1)
696 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700697 alc_ptr = kzalloc_node(memsize, gfp, node);
698 if (!alc_ptr)
699 return NULL;
700
701 for_each_node(i) {
702 if (i == node || !node_online(i))
703 continue;
704 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
705 if (!alc_ptr[i]) {
706 for (i--; i >= 0; i--)
707 kfree(alc_ptr[i]);
708 kfree(alc_ptr);
709 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700710 }
711 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700712 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700713}
714
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700715static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700716{
717 int i;
718
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700719 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700720 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700721 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700722 kfree(alc_ptr[i]);
723 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700724}
725
Pekka Enberg343e0d72006-02-01 03:05:50 -0800726static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700727 struct array_cache *ac, int node,
728 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700729{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700730 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700731
732 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000733 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800734 /*
735 * Stuff objects into the remote nodes shared array first.
736 * That way we could avoid the overhead of putting the objects
737 * into the free lists and getting them back later.
738 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000739 if (n->shared)
740 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800741
Joonsoo Kim833b7062014-08-06 16:04:33 -0700742 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700743 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000744 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700745 }
746}
747
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800748/*
749 * Called from cache_reap() to regularly drain alien caches round robin.
750 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000751static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800752{
Christoph Lameter909ea962010-12-08 16:22:55 +0100753 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800754
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000755 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700756 struct alien_cache *alc = n->alien[node];
757 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800758
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700759 if (alc) {
760 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700761 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700762 LIST_HEAD(list);
763
764 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700765 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700766 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700767 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800768 }
769 }
770}
771
Andrew Mortona737b3e2006-03-22 00:08:11 -0800772static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700773 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700774{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800775 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700776 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700777 struct array_cache *ac;
778 unsigned long flags;
779
780 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700781 alc = alien[i];
782 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700783 LIST_HEAD(list);
784
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700785 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700786 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700787 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700788 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700789 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700790 }
791 }
792}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700793
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700794static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
795 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700796{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000797 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700798 struct alien_cache *alien = NULL;
799 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700800 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700801
Christoph Lameter18bf8542014-08-06 16:04:11 -0700802 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700803 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700804 if (n->alien && n->alien[page_node]) {
805 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700806 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700807 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700808 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700809 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700810 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700811 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700812 ac->entry[ac->avail++] = objp;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700813 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700814 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700815 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700816 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700817 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700818 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700819 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -0700820 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700821 }
822 return 1;
823}
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700824
825static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
826{
827 int page_node = page_to_nid(virt_to_page(objp));
828 int node = numa_mem_id();
829 /*
830 * Make sure we are not freeing a object from another node to the array
831 * cache on this cpu.
832 */
833 if (likely(node == page_node))
834 return 0;
835
836 return __cache_free_alien(cachep, objp, node, page_node);
837}
David Rientjes4167e9b2015-04-14 15:46:55 -0700838
839/*
Mel Gorman444eb2a42016-03-17 14:19:23 -0700840 * Construct gfp mask to allocate from a specific node but do not reclaim or
841 * warn about failures.
David Rientjes4167e9b2015-04-14 15:46:55 -0700842 */
843static inline gfp_t gfp_exact_node(gfp_t flags)
844{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700845 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
David Rientjes4167e9b2015-04-14 15:46:55 -0700846}
Christoph Lametere498be72005-09-09 13:03:32 -0700847#endif
848
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700849static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
850{
851 struct kmem_cache_node *n;
852
853 /*
854 * Set up the kmem_cache_node for cpu before we can
855 * begin anything. Make sure some other cpu on this
856 * node has not already allocated this
857 */
858 n = get_node(cachep, node);
859 if (n) {
860 spin_lock_irq(&n->list_lock);
861 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
862 cachep->num;
863 spin_unlock_irq(&n->list_lock);
864
865 return 0;
866 }
867
868 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
869 if (!n)
870 return -ENOMEM;
871
872 kmem_cache_node_init(n);
873 n->next_reap = jiffies + REAPTIMEOUT_NODE +
874 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
875
876 n->free_limit =
877 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
878
879 /*
880 * The kmem_cache_nodes don't come and go as CPUs
881 * come and go. slab_mutex is sufficient
882 * protection here.
883 */
884 cachep->node[node] = n;
885
886 return 0;
887}
888
David Rientjes8f9f8d92010-03-27 19:40:47 -0700889/*
Christoph Lameter6a673682013-01-10 19:14:19 +0000890 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000891 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -0700892 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +0000893 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -0700894 * already in use.
895 *
Christoph Lameter18004c52012-07-06 15:25:12 -0500896 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -0700897 */
Christoph Lameter6a673682013-01-10 19:14:19 +0000898static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700899{
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700900 int ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700901 struct kmem_cache *cachep;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700902
Christoph Lameter18004c52012-07-06 15:25:12 -0500903 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700904 ret = init_cache_node(cachep, node, GFP_KERNEL);
905 if (ret)
906 return ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700907 }
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700908
David Rientjes8f9f8d92010-03-27 19:40:47 -0700909 return 0;
910}
911
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700912static int setup_kmem_cache_node(struct kmem_cache *cachep,
913 int node, gfp_t gfp, bool force_change)
914{
915 int ret = -ENOMEM;
916 struct kmem_cache_node *n;
917 struct array_cache *old_shared = NULL;
918 struct array_cache *new_shared = NULL;
919 struct alien_cache **new_alien = NULL;
920 LIST_HEAD(list);
921
922 if (use_alien_caches) {
923 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
924 if (!new_alien)
925 goto fail;
926 }
927
928 if (cachep->shared) {
929 new_shared = alloc_arraycache(node,
930 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
931 if (!new_shared)
932 goto fail;
933 }
934
935 ret = init_cache_node(cachep, node, gfp);
936 if (ret)
937 goto fail;
938
939 n = get_node(cachep, node);
940 spin_lock_irq(&n->list_lock);
941 if (n->shared && force_change) {
942 free_block(cachep, n->shared->entry,
943 n->shared->avail, node, &list);
944 n->shared->avail = 0;
945 }
946
947 if (!n->shared || force_change) {
948 old_shared = n->shared;
949 n->shared = new_shared;
950 new_shared = NULL;
951 }
952
953 if (!n->alien) {
954 n->alien = new_alien;
955 new_alien = NULL;
956 }
957
958 spin_unlock_irq(&n->list_lock);
959 slabs_destroy(cachep, &list);
960
Joonsoo Kim801faf02016-05-19 17:10:31 -0700961 /*
962 * To protect lockless access to n->shared during irq disabled context.
963 * If n->shared isn't NULL in irq disabled context, accessing to it is
964 * guaranteed to be valid until irq is re-enabled, because it will be
965 * freed after synchronize_sched().
966 */
967 if (force_change)
968 synchronize_sched();
969
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700970fail:
971 kfree(old_shared);
972 kfree(new_shared);
973 free_alien_cache(new_alien);
974
975 return ret;
976}
977
Paul Gortmaker0db06282013-06-19 14:53:51 -0400978static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979{
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700980 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000981 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700982 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +1030983 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700984
Christoph Lameter18004c52012-07-06 15:25:12 -0500985 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700986 struct array_cache *nc;
987 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700988 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700989 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700990
Christoph Lameter18bf8542014-08-06 16:04:11 -0700991 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000992 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700993 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700994
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000995 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700996
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000997 /* Free limit for this kmem_cache_node */
998 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700999
1000 /* cpu is dead; no one can alloc from it. */
1001 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1002 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001003 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001004 nc->avail = 0;
1005 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001006
Rusty Russell58463c12009-12-17 11:43:12 -06001007 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001008 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001009 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001010 }
1011
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001012 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001013 if (shared) {
1014 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001015 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001016 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001017 }
1018
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001019 alien = n->alien;
1020 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001021
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001022 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001023
1024 kfree(shared);
1025 if (alien) {
1026 drain_alien_cache(cachep, alien);
1027 free_alien_cache(alien);
1028 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001029
1030free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001031 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001032 }
1033 /*
1034 * In the previous loop, all the objects were freed to
1035 * the respective cache's slabs, now we can go ahead and
1036 * shrink each nodelist to its limit.
1037 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001038 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001039 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001040 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001041 continue;
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001042 drain_freelist(cachep, n, INT_MAX);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001043 }
1044}
1045
Paul Gortmaker0db06282013-06-19 14:53:51 -04001046static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001047{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001048 struct kmem_cache *cachep;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001049 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001050 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001052 /*
1053 * We need to do this right in the beginning since
1054 * alloc_arraycache's are going to use this list.
1055 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001056 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001057 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001058 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001059 if (err < 0)
1060 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001061
1062 /*
1063 * Now we can go ahead with allocating the shared arrays and
1064 * array caches
1065 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001066 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07001067 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1068 if (err)
1069 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001070 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001071
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001072 return 0;
1073bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001074 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001075 return -ENOMEM;
1076}
1077
Paul Gortmaker0db06282013-06-19 14:53:51 -04001078static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001079 unsigned long action, void *hcpu)
1080{
1081 long cpu = (long)hcpu;
1082 int err = 0;
1083
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001085 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001086 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001087 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001088 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001089 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 break;
1091 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001092 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 start_cpu_timer(cpu);
1094 break;
1095#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001096 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001097 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001098 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001099 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001100 * held so that if cache_reap() is invoked it cannot do
1101 * anything expensive but will only modify reap_work
1102 * and reschedule the timer.
1103 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001104 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001105 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001106 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001107 break;
1108 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001109 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001110 start_cpu_timer(cpu);
1111 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001113 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001114 /*
1115 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001116 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001117 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001118 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001119 * structure is usually allocated from kmem_cache_create() and
1120 * gets destroyed at kmem_cache_destroy().
1121 */
Simon Arlott183ff222007-10-20 01:27:18 +02001122 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001123#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001125 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001126 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001127 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001128 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001131 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132}
1133
Paul Gortmaker0db06282013-06-19 14:53:51 -04001134static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001135 &cpuup_callback, NULL, 0
1136};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
David Rientjes8f9f8d92010-03-27 19:40:47 -07001138#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1139/*
1140 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1141 * Returns -EBUSY if all objects cannot be drained so that the node is not
1142 * removed.
1143 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001144 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001145 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001146static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001147{
1148 struct kmem_cache *cachep;
1149 int ret = 0;
1150
Christoph Lameter18004c52012-07-06 15:25:12 -05001151 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001152 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001153
Christoph Lameter18bf8542014-08-06 16:04:11 -07001154 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001155 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001156 continue;
1157
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001158 drain_freelist(cachep, n, INT_MAX);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001159
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001160 if (!list_empty(&n->slabs_full) ||
1161 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001162 ret = -EBUSY;
1163 break;
1164 }
1165 }
1166 return ret;
1167}
1168
1169static int __meminit slab_memory_callback(struct notifier_block *self,
1170 unsigned long action, void *arg)
1171{
1172 struct memory_notify *mnb = arg;
1173 int ret = 0;
1174 int nid;
1175
1176 nid = mnb->status_change_nid;
1177 if (nid < 0)
1178 goto out;
1179
1180 switch (action) {
1181 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001182 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001183 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001184 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001185 break;
1186 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001187 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001188 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001189 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001190 break;
1191 case MEM_ONLINE:
1192 case MEM_OFFLINE:
1193 case MEM_CANCEL_ONLINE:
1194 case MEM_CANCEL_OFFLINE:
1195 break;
1196 }
1197out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001198 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001199}
1200#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1201
Christoph Lametere498be72005-09-09 13:03:32 -07001202/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001203 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001204 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001205static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001206 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001207{
Christoph Lameter6744f082013-01-10 19:12:17 +00001208 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001209
Christoph Lameter6744f082013-01-10 19:12:17 +00001210 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001211 BUG_ON(!ptr);
1212
Christoph Lameter6744f082013-01-10 19:12:17 +00001213 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001214 /*
1215 * Do not assume that spinlocks can be initialized via memcpy:
1216 */
1217 spin_lock_init(&ptr->list_lock);
1218
Christoph Lametere498be72005-09-09 13:03:32 -07001219 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001220 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001221}
1222
Andrew Mortona737b3e2006-03-22 00:08:11 -08001223/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001224 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1225 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001226 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001227static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001228{
1229 int node;
1230
1231 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001232 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001233 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001234 REAPTIMEOUT_NODE +
1235 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001236 }
1237}
1238
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07001239#ifdef CONFIG_SLAB_FREELIST_RANDOM
1240static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
1241 size_t count)
1242{
1243 size_t i;
1244 unsigned int rand;
1245
1246 for (i = 0; i < count; i++)
1247 list[i] = i;
1248
1249 /* Fisher-Yates shuffle */
1250 for (i = count - 1; i > 0; i--) {
1251 rand = prandom_u32_state(state);
1252 rand %= (i + 1);
1253 swap(list[i], list[rand]);
1254 }
1255}
1256
1257/* Create a random sequence per cache */
1258static int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1259{
1260 unsigned int seed, count = cachep->num;
1261 struct rnd_state state;
1262
1263 if (count < 2)
1264 return 0;
1265
1266 /* If it fails, we will just use the global lists */
1267 cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), gfp);
1268 if (!cachep->random_seq)
1269 return -ENOMEM;
1270
1271 /* Get best entropy at this stage */
1272 get_random_bytes_arch(&seed, sizeof(seed));
1273 prandom_seed_state(&state, seed);
1274
1275 freelist_randomize(&state, cachep->random_seq, count);
1276 return 0;
1277}
1278
1279/* Destroy the per-cache random freelist sequence */
1280static void cache_random_seq_destroy(struct kmem_cache *cachep)
1281{
1282 kfree(cachep->random_seq);
1283 cachep->random_seq = NULL;
1284}
1285#else
1286static inline int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1287{
1288 return 0;
1289}
1290static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }
1291#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1292
1293
Pekka Enberg556a1692008-01-25 08:20:51 +02001294/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001295 * Initialisation. Called after the page allocator have been initialised and
1296 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 */
1298void __init kmem_cache_init(void)
1299{
Christoph Lametere498be72005-09-09 13:03:32 -07001300 int i;
1301
Joonsoo Kim68126702013-10-24 10:07:42 +09001302 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1303 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001304 kmem_cache = &kmem_cache_boot;
1305
Joonsoo Kim88881772016-05-19 17:10:05 -07001306 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001307 use_alien_caches = 0;
1308
Christoph Lameter3c583462012-11-28 16:23:01 +00001309 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001310 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001311
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 /*
1313 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001314 * page orders on machines with more than 32MB of memory if
1315 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001317 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001318 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 /* Bootstrap is tricky, because several objects are allocated
1321 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001322 * 1) initialize the kmem_cache cache: it contains the struct
1323 * kmem_cache structures of all caches, except kmem_cache itself:
1324 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001325 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001326 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001327 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001329 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001330 * An __init data area is used for the head array.
1331 * 3) Create the remaining kmalloc caches, with minimally sized
1332 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001333 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001335 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001336 * the other cache's with kmalloc allocated memory.
1337 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 */
1339
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001340 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341
Eric Dumazet8da34302007-05-06 14:49:29 -07001342 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001343 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001344 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001345 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001346 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001347 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001348 SLAB_HWCACHE_ALIGN);
1349 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001350 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351
Andrew Mortona737b3e2006-03-22 00:08:11 -08001352 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001353 * Initialize the caches that provide memory for the kmem_cache_node
1354 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001355 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001356 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001357 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001358 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001359 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001360
Ingo Molnare0a42722006-06-23 02:03:46 -07001361 slab_early_init = 0;
1362
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001363 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001364 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001365 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
Mel Gorman9c09a952008-01-24 05:49:54 -08001367 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001368 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001369
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001370 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001371 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001372 }
1373 }
1374
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001375 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001376}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001377
Pekka Enberg8429db52009-06-12 15:58:59 +03001378void __init kmem_cache_init_late(void)
1379{
1380 struct kmem_cache *cachep;
1381
Christoph Lameter97d06602012-07-06 15:25:11 -05001382 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001383
Pekka Enberg8429db52009-06-12 15:58:59 +03001384 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001385 mutex_lock(&slab_mutex);
1386 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001387 if (enable_cpucache(cachep, GFP_NOWAIT))
1388 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001389 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001390
Christoph Lameter97d06602012-07-06 15:25:11 -05001391 /* Done! */
1392 slab_state = FULL;
1393
Andrew Mortona737b3e2006-03-22 00:08:11 -08001394 /*
1395 * Register a cpu startup notifier callback that initializes
1396 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 */
1398 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399
David Rientjes8f9f8d92010-03-27 19:40:47 -07001400#ifdef CONFIG_NUMA
1401 /*
1402 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001403 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001404 */
1405 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1406#endif
1407
Andrew Mortona737b3e2006-03-22 00:08:11 -08001408 /*
1409 * The reap timers are started later, with a module init call: That part
1410 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 */
1412}
1413
1414static int __init cpucache_init(void)
1415{
1416 int cpu;
1417
Andrew Mortona737b3e2006-03-22 00:08:11 -08001418 /*
1419 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 */
Christoph Lametere498be72005-09-09 13:03:32 -07001421 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001422 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001423
1424 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001425 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 return 0;
1427}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428__initcall(cpucache_init);
1429
Rafael Aquini8bdec192012-03-09 17:27:27 -03001430static noinline void
1431slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1432{
David Rientjes9a02d692014-06-04 16:06:36 -07001433#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001434 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001435 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001436 unsigned long flags;
1437 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001438 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1439 DEFAULT_RATELIMIT_BURST);
1440
1441 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1442 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001443
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001444 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1445 nodeid, gfpflags, &gfpflags);
1446 pr_warn(" cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001447 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001448
Christoph Lameter18bf8542014-08-06 16:04:11 -07001449 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001450 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1451 unsigned long active_slabs = 0, num_slabs = 0;
1452
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001453 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001454 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001455 active_objs += cachep->num;
1456 active_slabs++;
1457 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001458 list_for_each_entry(page, &n->slabs_partial, lru) {
1459 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001460 active_slabs++;
1461 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001462 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001463 num_slabs++;
1464
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001465 free_objects += n->free_objects;
1466 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001467
1468 num_slabs += active_slabs;
1469 num_objs = num_slabs * cachep->num;
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001470 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
Rafael Aquini8bdec192012-03-09 17:27:27 -03001471 node, active_slabs, num_slabs, active_objs, num_objs,
1472 free_objects);
1473 }
David Rientjes9a02d692014-06-04 16:06:36 -07001474#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001475}
1476
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001478 * Interface to system's page allocator. No need to hold the
1479 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 *
1481 * If we requested dmaable memory, we will get it. Even if we
1482 * did not request dmaable memory, we might get it, but that
1483 * would be relatively rare and ignorable.
1484 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001485static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1486 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487{
1488 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001489 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001490
Glauber Costaa618e892012-06-14 16:17:21 +04001491 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001492 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1493 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001494
Vlastimil Babka96db8002015-09-08 15:03:50 -07001495 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001496 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001497 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001499 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001501 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1502 __free_pages(page, cachep->gfporder);
1503 return NULL;
1504 }
1505
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001506 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001508 add_zone_page_state(page_zone(page),
1509 NR_SLAB_RECLAIMABLE, nr_pages);
1510 else
1511 add_zone_page_state(page_zone(page),
1512 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001513
Joonsoo Kima57a4982013-10-24 10:07:44 +09001514 __SetPageSlab(page);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001515 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1516 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001517 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001518
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001519 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1520 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1521
1522 if (cachep->ctor)
1523 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1524 else
1525 kmemcheck_mark_unallocated_pages(page, nr_pages);
1526 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001527
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001528 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529}
1530
1531/*
1532 * Interface to system's page release.
1533 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001534static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535{
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001536 int order = cachep->gfporder;
1537 unsigned long nr_freed = (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001539 kmemcheck_free_shadow(page, order);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001540
Christoph Lameter972d1a72006-09-25 23:31:51 -07001541 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1542 sub_zone_page_state(page_zone(page),
1543 NR_SLAB_RECLAIMABLE, nr_freed);
1544 else
1545 sub_zone_page_state(page_zone(page),
1546 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001547
Joonsoo Kima57a4982013-10-24 10:07:44 +09001548 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001549 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001550 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001551 page_mapcount_reset(page);
1552 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001553
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 if (current->reclaim_state)
1555 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001556 memcg_uncharge_slab(page, order, cachep);
1557 __free_pages(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558}
1559
1560static void kmem_rcu_free(struct rcu_head *head)
1561{
Joonsoo Kim68126702013-10-24 10:07:42 +09001562 struct kmem_cache *cachep;
1563 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
Joonsoo Kim68126702013-10-24 10:07:42 +09001565 page = container_of(head, struct page, rcu_head);
1566 cachep = page->slab_cache;
1567
1568 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569}
1570
1571#if DEBUG
Joonsoo Kim40b44132016-03-15 14:54:21 -07001572static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1573{
1574 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1575 (cachep->size % PAGE_SIZE) == 0)
1576 return true;
1577
1578 return false;
1579}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
1581#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001582static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001583 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001585 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001587 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 return;
1591
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001592 *addr++ = 0x12345678;
1593 *addr++ = caller;
1594 *addr++ = smp_processor_id();
1595 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 {
1597 unsigned long *sptr = &caller;
1598 unsigned long svalue;
1599
1600 while (!kstack_end(sptr)) {
1601 svalue = *sptr++;
1602 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001603 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 size -= sizeof(unsigned long);
1605 if (size <= sizeof(unsigned long))
1606 break;
1607 }
1608 }
1609
1610 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612}
Joonsoo Kim40b44132016-03-15 14:54:21 -07001613
1614static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1615 int map, unsigned long caller)
1616{
1617 if (!is_debug_pagealloc_cache(cachep))
1618 return;
1619
1620 if (caller)
1621 store_stackinfo(cachep, objp, caller);
1622
1623 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1624}
1625
1626#else
1627static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1628 int map, unsigned long caller) {}
1629
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630#endif
1631
Pekka Enberg343e0d72006-02-01 03:05:50 -08001632static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001634 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001635 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636
1637 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001638 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639}
1640
1641static void dump_line(char *data, int offset, int limit)
1642{
1643 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001644 unsigned char error = 0;
1645 int bad_count = 0;
1646
Joe Perches11705322016-03-17 14:19:50 -07001647 pr_err("%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001648 for (i = 0; i < limit; i++) {
1649 if (data[offset + i] != POISON_FREE) {
1650 error = data[offset + i];
1651 bad_count++;
1652 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001653 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001654 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1655 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001656
1657 if (bad_count == 1) {
1658 error ^= POISON_FREE;
1659 if (!(error & (error - 1))) {
Joe Perches11705322016-03-17 14:19:50 -07001660 pr_err("Single bit error detected. Probably bad RAM.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001661#ifdef CONFIG_X86
Joe Perches11705322016-03-17 14:19:50 -07001662 pr_err("Run memtest86+ or a similar memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001663#else
Joe Perches11705322016-03-17 14:19:50 -07001664 pr_err("Run a memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001665#endif
1666 }
1667 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668}
1669#endif
1670
1671#if DEBUG
1672
Pekka Enberg343e0d72006-02-01 03:05:50 -08001673static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674{
1675 int i, size;
1676 char *realobj;
1677
1678 if (cachep->flags & SLAB_RED_ZONE) {
Joe Perches11705322016-03-17 14:19:50 -07001679 pr_err("Redzone: 0x%llx/0x%llx\n",
1680 *dbg_redzone1(cachep, objp),
1681 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 }
1683
1684 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches11705322016-03-17 14:19:50 -07001685 pr_err("Last user: [<%p>](%pSR)\n",
Joe Perches071361d2012-12-12 10:19:12 -08001686 *dbg_userword(cachep, objp),
1687 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001689 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001690 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001691 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 int limit;
1693 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001694 if (i + limit > size)
1695 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 dump_line(realobj, i, limit);
1697 }
1698}
1699
Pekka Enberg343e0d72006-02-01 03:05:50 -08001700static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701{
1702 char *realobj;
1703 int size, i;
1704 int lines = 0;
1705
Joonsoo Kim40b44132016-03-15 14:54:21 -07001706 if (is_debug_pagealloc_cache(cachep))
1707 return;
1708
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001709 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001710 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001712 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001714 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 exp = POISON_END;
1716 if (realobj[i] != exp) {
1717 int limit;
1718 /* Mismatch ! */
1719 /* Print header */
1720 if (lines == 0) {
Joe Perches11705322016-03-17 14:19:50 -07001721 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1722 print_tainted(), cachep->name,
1723 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 print_objinfo(cachep, objp, 0);
1725 }
1726 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001727 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001729 if (i + limit > size)
1730 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 dump_line(realobj, i, limit);
1732 i += 16;
1733 lines++;
1734 /* Limit to 5 lines */
1735 if (lines > 5)
1736 break;
1737 }
1738 }
1739 if (lines != 0) {
1740 /* Print some data about the neighboring objects, if they
1741 * exist:
1742 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001743 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001744 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
Joonsoo Kim8456a642013-10-24 10:07:49 +09001746 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001748 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001749 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001750 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 print_objinfo(cachep, objp, 2);
1752 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001753 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001754 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001755 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001756 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 print_objinfo(cachep, objp, 2);
1758 }
1759 }
1760}
1761#endif
1762
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001764static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1765 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001766{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767 int i;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001768
1769 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1770 poison_obj(cachep, page->freelist - obj_offset(cachep),
1771 POISON_FREE);
1772 }
1773
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001775 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776
1777 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07001779 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 }
1781 if (cachep->flags & SLAB_RED_ZONE) {
1782 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001783 slab_error(cachep, "start of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001785 slab_error(cachep, "end of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001788}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001790static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1791 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001792{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001793}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794#endif
1795
Randy Dunlap911851e2006-03-22 00:08:14 -08001796/**
1797 * slab_destroy - destroy and release all objects in a slab
1798 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001799 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001800 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001801 * Destroy all the objs in a slab page, and release the mem back to the system.
1802 * Before calling the slab page must have been unlinked from the cache. The
1803 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001804 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001805static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001806{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001807 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001808
Joonsoo Kim8456a642013-10-24 10:07:49 +09001809 freelist = page->freelist;
1810 slab_destroy_debugcheck(cachep, page);
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001811 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1812 call_rcu(&page->rcu_head, kmem_rcu_free);
1813 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001814 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001815
1816 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001817 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001818 * although actual page can be freed in rcu context
1819 */
1820 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001821 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822}
1823
Joonsoo Kim97654df2014-08-06 16:04:25 -07001824static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1825{
1826 struct page *page, *n;
1827
1828 list_for_each_entry_safe(page, n, list, lru) {
1829 list_del(&page->lru);
1830 slab_destroy(cachep, page);
1831 }
1832}
1833
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001835 * calculate_slab_order - calculate size (page order) of slabs
1836 * @cachep: pointer to the cache that is being created
1837 * @size: size of objects to be created in this cache.
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001838 * @flags: slab allocation flags
1839 *
1840 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001841 *
1842 * This could be made much more intelligent. For now, try to avoid using
1843 * high order pages for slabs. When the gfp() functions are more friendly
1844 * towards high-order requests, this should be changed.
1845 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001846static size_t calculate_slab_order(struct kmem_cache *cachep,
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07001847 size_t size, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001848{
1849 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001850 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001851
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001852 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001853 unsigned int num;
1854 size_t remainder;
1855
Joonsoo Kim70f75062016-03-15 14:54:53 -07001856 num = cache_estimate(gfporder, size, flags, &remainder);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001857 if (!num)
1858 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001859
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001860 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1861 if (num > SLAB_OBJ_MAX_NUM)
1862 break;
1863
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001864 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001865 struct kmem_cache *freelist_cache;
1866 size_t freelist_size;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001867
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001868 freelist_size = num * sizeof(freelist_idx_t);
1869 freelist_cache = kmalloc_slab(freelist_size, 0u);
1870 if (!freelist_cache)
1871 continue;
1872
1873 /*
1874 * Needed to avoid possible looping condition
Joonsoo Kim76b342b2016-05-19 17:10:26 -07001875 * in cache_grow_begin()
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001876 */
1877 if (OFF_SLAB(freelist_cache))
1878 continue;
1879
1880 /* check if off slab has enough benefit */
1881 if (freelist_cache->size > cachep->size / 2)
1882 continue;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001883 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001884
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001885 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001886 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001887 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001888 left_over = remainder;
1889
1890 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001891 * A VFS-reclaimable slab tends to have most allocations
1892 * as GFP_NOFS and we really don't want to have to be allocating
1893 * higher-order pages when we are unable to shrink dcache.
1894 */
1895 if (flags & SLAB_RECLAIM_ACCOUNT)
1896 break;
1897
1898 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001899 * Large number of objects is good, but very large slabs are
1900 * currently bad for the gfp()s.
1901 */
David Rientjes543585c2011-10-18 22:09:24 -07001902 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001903 break;
1904
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001905 /*
1906 * Acceptable internal fragmentation?
1907 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001908 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001909 break;
1910 }
1911 return left_over;
1912}
1913
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001914static struct array_cache __percpu *alloc_kmem_cache_cpus(
1915 struct kmem_cache *cachep, int entries, int batchcount)
1916{
1917 int cpu;
1918 size_t size;
1919 struct array_cache __percpu *cpu_cache;
1920
1921 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001922 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001923
1924 if (!cpu_cache)
1925 return NULL;
1926
1927 for_each_possible_cpu(cpu) {
1928 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1929 entries, batchcount);
1930 }
1931
1932 return cpu_cache;
1933}
1934
Pekka Enberg83b519e2009-06-10 19:40:04 +03001935static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001936{
Christoph Lameter97d06602012-07-06 15:25:11 -05001937 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001938 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001939
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001940 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1941 if (!cachep->cpu_cache)
1942 return 1;
1943
Christoph Lameter97d06602012-07-06 15:25:11 -05001944 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001945 /* Creation of first cache (kmem_cache). */
1946 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001947 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001948 /* For kmem_cache_node */
1949 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001950 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001951 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001952
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001953 for_each_online_node(node) {
1954 cachep->node[node] = kmalloc_node(
1955 sizeof(struct kmem_cache_node), gfp, node);
1956 BUG_ON(!cachep->node[node]);
1957 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001958 }
1959 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001960
Christoph Lameter6a673682013-01-10 19:14:19 +00001961 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001962 jiffies + REAPTIMEOUT_NODE +
1963 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001964
1965 cpu_cache_get(cachep)->avail = 0;
1966 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1967 cpu_cache_get(cachep)->batchcount = 1;
1968 cpu_cache_get(cachep)->touched = 0;
1969 cachep->batchcount = 1;
1970 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001971 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001972}
1973
Joonsoo Kim12220de2014-10-09 15:26:24 -07001974unsigned long kmem_cache_flags(unsigned long object_size,
1975 unsigned long flags, const char *name,
1976 void (*ctor)(void *))
1977{
1978 return flags;
1979}
1980
1981struct kmem_cache *
1982__kmem_cache_alias(const char *name, size_t size, size_t align,
1983 unsigned long flags, void (*ctor)(void *))
1984{
1985 struct kmem_cache *cachep;
1986
1987 cachep = find_mergeable(size, align, flags, name, ctor);
1988 if (cachep) {
1989 cachep->refcount++;
1990
1991 /*
1992 * Adjust the object sizes so that we clear
1993 * the complete object on kzalloc.
1994 */
1995 cachep->object_size = max_t(int, cachep->object_size, size);
1996 }
1997 return cachep;
1998}
1999
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002000static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
2001 size_t size, unsigned long flags)
2002{
2003 size_t left;
2004
2005 cachep->num = 0;
2006
2007 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
2008 return false;
2009
2010 left = calculate_slab_order(cachep, size,
2011 flags | CFLGS_OBJFREELIST_SLAB);
2012 if (!cachep->num)
2013 return false;
2014
2015 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
2016 return false;
2017
2018 cachep->colour = left / cachep->colour_off;
2019
2020 return true;
2021}
2022
Joonsoo Kim158e3192016-03-15 14:54:35 -07002023static bool set_off_slab_cache(struct kmem_cache *cachep,
2024 size_t size, unsigned long flags)
2025{
2026 size_t left;
2027
2028 cachep->num = 0;
2029
2030 /*
Joonsoo Kim3217fd92016-03-15 14:54:41 -07002031 * Always use on-slab management when SLAB_NOLEAKTRACE
2032 * to avoid recursive calls into kmemleak.
Joonsoo Kim158e3192016-03-15 14:54:35 -07002033 */
Joonsoo Kim158e3192016-03-15 14:54:35 -07002034 if (flags & SLAB_NOLEAKTRACE)
2035 return false;
2036
2037 /*
2038 * Size is large, assume best to place the slab management obj
2039 * off-slab (should allow better packing of objs).
2040 */
2041 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2042 if (!cachep->num)
2043 return false;
2044
2045 /*
2046 * If the slab has been placed off-slab, and we have enough space then
2047 * move it on-slab. This is at the expense of any extra colouring.
2048 */
2049 if (left >= cachep->num * sizeof(freelist_idx_t))
2050 return false;
2051
2052 cachep->colour = left / cachep->colour_off;
2053
2054 return true;
2055}
2056
2057static bool set_on_slab_cache(struct kmem_cache *cachep,
2058 size_t size, unsigned long flags)
2059{
2060 size_t left;
2061
2062 cachep->num = 0;
2063
2064 left = calculate_slab_order(cachep, size, flags);
2065 if (!cachep->num)
2066 return false;
2067
2068 cachep->colour = left / cachep->colour_off;
2069
2070 return true;
2071}
2072
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002073/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002074 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002075 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 *
2078 * Returns a ptr to the cache on success, NULL on failure.
2079 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002080 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 * The flags are
2083 *
2084 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2085 * to catch references to uninitialised memory.
2086 *
2087 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2088 * for buffer overruns.
2089 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2091 * cacheline. This can be beneficial if you're counting cycles as closely
2092 * as davem.
2093 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002094int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002095__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002097 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002098 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002099 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002100 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103#if FORCED_DEBUG
2104 /*
2105 * Enable redzoning and last user accounting, except for caches with
2106 * large objects, if the increased size would increase the object size
2107 * above the next power of two: caches with object sizes just above a
2108 * power of two have a significant amount of internal fragmentation.
2109 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002110 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2111 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002112 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 if (!(flags & SLAB_DESTROY_BY_RCU))
2114 flags |= SLAB_POISON;
2115#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Andrew Mortona737b3e2006-03-22 00:08:11 -08002118 /*
2119 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 * unaligned accesses for some archs when redzoning is used, and makes
2121 * sure any on-slab bufctl's are also correctly aligned.
2122 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002123 if (size & (BYTES_PER_WORD - 1)) {
2124 size += (BYTES_PER_WORD - 1);
2125 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 }
2127
David Woodhouse87a927c2007-07-04 21:26:44 -04002128 if (flags & SLAB_RED_ZONE) {
2129 ralign = REDZONE_ALIGN;
2130 /* If redzoning, ensure that the second redzone is suitably
2131 * aligned, by adjusting the object size accordingly. */
2132 size += REDZONE_ALIGN - 1;
2133 size &= ~(REDZONE_ALIGN - 1);
2134 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002135
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002136 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002137 if (ralign < cachep->align) {
2138 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002140 /* disable debug if necessary */
2141 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002142 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002143 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002144 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002146 cachep->align = ralign;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002147 cachep->colour_off = cache_line_size();
2148 /* Offset must be a multiple of the alignment. */
2149 if (cachep->colour_off < cachep->align)
2150 cachep->colour_off = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151
Pekka Enberg83b519e2009-06-10 19:40:04 +03002152 if (slab_is_available())
2153 gfp = GFP_KERNEL;
2154 else
2155 gfp = GFP_NOWAIT;
2156
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158
Pekka Enbergca5f9702006-09-25 23:31:25 -07002159 /*
2160 * Both debugging options require word-alignment which is calculated
2161 * into align above.
2162 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002165 cachep->obj_offset += sizeof(unsigned long long);
2166 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 }
2168 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002169 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002170 * the real object. But if the second red zone needs to be
2171 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002173 if (flags & SLAB_RED_ZONE)
2174 size += REDZONE_ALIGN;
2175 else
2176 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 }
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002178#endif
2179
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002180 kasan_cache_create(cachep, &size, &flags);
2181
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002182 size = ALIGN(size, cachep->align);
2183 /*
2184 * We should restrict the number of objects in a slab to implement
2185 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2186 */
2187 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2188 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2189
2190#if DEBUG
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002191 /*
2192 * To activate debug pagealloc, off-slab management is necessary
2193 * requirement. In early phase of initialization, small sized slab
2194 * doesn't get initialized so it would not be possible. So, we need
2195 * to check size >= 256. It guarantees that all necessary small
2196 * sized slab is initialized in current slab initialization sequence.
2197 */
Joonsoo Kim40323272016-03-15 14:54:18 -07002198 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
Joonsoo Kimf3a3c322016-03-15 14:54:38 -07002199 size >= 256 && cachep->object_size > cache_line_size()) {
2200 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2201 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2202
2203 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2204 flags |= CFLGS_OFF_SLAB;
2205 cachep->obj_offset += tmp_size - size;
2206 size = tmp_size;
2207 goto done;
2208 }
2209 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 }
2211#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002213 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2214 flags |= CFLGS_OBJFREELIST_SLAB;
2215 goto done;
2216 }
2217
Joonsoo Kim158e3192016-03-15 14:54:35 -07002218 if (set_off_slab_cache(cachep, size, flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 flags |= CFLGS_OFF_SLAB;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002220 goto done;
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002221 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Joonsoo Kim158e3192016-03-15 14:54:35 -07002223 if (set_on_slab_cache(cachep, size, flags))
2224 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225
Joonsoo Kim158e3192016-03-15 14:54:35 -07002226 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002227
Joonsoo Kim158e3192016-03-15 14:54:35 -07002228done:
2229 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002231 cachep->allocflags = __GFP_COMP;
Yang Shia3187e42016-05-19 17:10:41 -07002232 if (flags & SLAB_CACHE_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002233 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002234 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002235 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236
Joonsoo Kim40b44132016-03-15 14:54:21 -07002237#if DEBUG
2238 /*
2239 * If we're going to use the generic kernel_map_pages()
2240 * poisoning, then it's going to smash the contents of
2241 * the redzone and userword anyhow, so switch them off.
2242 */
2243 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2244 (cachep->flags & SLAB_POISON) &&
2245 is_debug_pagealloc_cache(cachep))
2246 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2247#endif
2248
2249 if (OFF_SLAB(cachep)) {
Joonsoo Kim158e3192016-03-15 14:54:35 -07002250 cachep->freelist_cache =
2251 kmalloc_slab(cachep->freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002252 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002254 err = setup_cpu_cache(cachep, gfp);
2255 if (err) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002256 __kmem_cache_release(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002257 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002258 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002260 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
2263#if DEBUG
2264static void check_irq_off(void)
2265{
2266 BUG_ON(!irqs_disabled());
2267}
2268
2269static void check_irq_on(void)
2270{
2271 BUG_ON(irqs_disabled());
2272}
2273
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002274static void check_mutex_acquired(void)
2275{
2276 BUG_ON(!mutex_is_locked(&slab_mutex));
2277}
2278
Pekka Enberg343e0d72006-02-01 03:05:50 -08002279static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280{
2281#ifdef CONFIG_SMP
2282 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002283 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284#endif
2285}
Christoph Lametere498be72005-09-09 13:03:32 -07002286
Pekka Enberg343e0d72006-02-01 03:05:50 -08002287static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002288{
2289#ifdef CONFIG_SMP
2290 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002291 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002292#endif
2293}
2294
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295#else
2296#define check_irq_off() do { } while(0)
2297#define check_irq_on() do { } while(0)
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002298#define check_mutex_acquired() do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002300#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301#endif
2302
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002303static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2304 int node, bool free_all, struct list_head *list)
2305{
2306 int tofree;
2307
2308 if (!ac || !ac->avail)
2309 return;
2310
2311 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2312 if (tofree > ac->avail)
2313 tofree = (ac->avail + 1) / 2;
2314
2315 free_block(cachep, ac->entry, tofree, node, list);
2316 ac->avail -= tofree;
2317 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2318}
Christoph Lameteraab22072006-03-22 00:09:06 -08002319
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320static void do_drain(void *arg)
2321{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002322 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002324 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002325 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002326 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327
2328 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002329 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002330 n = get_node(cachep, node);
2331 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002332 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002333 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002334 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 ac->avail = 0;
2336}
2337
Pekka Enberg343e0d72006-02-01 03:05:50 -08002338static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002340 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002341 int node;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002342 LIST_HEAD(list);
Christoph Lametere498be72005-09-09 13:03:32 -07002343
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002344 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002346 for_each_kmem_cache_node(cachep, node, n)
2347 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002348 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002349
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002350 for_each_kmem_cache_node(cachep, node, n) {
2351 spin_lock_irq(&n->list_lock);
2352 drain_array_locked(cachep, n->shared, node, true, &list);
2353 spin_unlock_irq(&n->list_lock);
2354
2355 slabs_destroy(cachep, &list);
2356 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357}
2358
Christoph Lametered11d9e2006-06-30 01:55:45 -07002359/*
2360 * Remove slabs from the list of free slabs.
2361 * Specify the number of slabs to drain in tofree.
2362 *
2363 * Returns the actual number of slabs released.
2364 */
2365static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002366 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002368 struct list_head *p;
2369 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002370 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371
Christoph Lametered11d9e2006-06-30 01:55:45 -07002372 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002373 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002375 spin_lock_irq(&n->list_lock);
2376 p = n->slabs_free.prev;
2377 if (p == &n->slabs_free) {
2378 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002379 goto out;
2380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Joonsoo Kim8456a642013-10-24 10:07:49 +09002382 page = list_entry(p, struct page, lru);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002383 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002384 /*
2385 * Safe to drop the lock. The slab is no longer linked
2386 * to the cache.
2387 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002388 n->free_objects -= cache->num;
2389 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002390 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002391 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002393out:
2394 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395}
2396
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002397int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002398{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002399 int ret = 0;
2400 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002401 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002402
2403 drain_cpu_caches(cachep);
2404
2405 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002406 for_each_kmem_cache_node(cachep, node, n) {
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07002407 drain_freelist(cachep, n, INT_MAX);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002408
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002409 ret += !list_empty(&n->slabs_full) ||
2410 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002411 }
2412 return (ret ? 1 : 0);
2413}
2414
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002415int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416{
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002417 return __kmem_cache_shrink(cachep, false);
2418}
2419
2420void __kmem_cache_release(struct kmem_cache *cachep)
2421{
Christoph Lameter12c36672012-09-04 23:38:33 +00002422 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002423 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002424
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002425 cache_random_seq_destroy(cachep);
2426
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002427 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002428
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002429 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002430 for_each_kmem_cache_node(cachep, i, n) {
2431 kfree(n->shared);
2432 free_alien_cache(n->alien);
2433 kfree(n);
2434 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 }
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
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002459 page->s_mem = addr + colour_off;
2460 page->active = 0;
2461
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002462 if (OBJFREELIST_SLAB(cachep))
2463 freelist = NULL;
2464 else if (OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002466 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002467 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002468 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 return NULL;
2470 } else {
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002471 /* We will use last bytes at the slab for freelist */
2472 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2473 cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 }
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002475
Joonsoo Kim8456a642013-10-24 10:07:49 +09002476 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477}
2478
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002479static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002481 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002482}
2483
2484static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002485 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002486{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002487 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488}
2489
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002490static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491{
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002492#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 int i;
2494
2495 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002496 void *objp = index_to_obj(cachep, page, i);
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002497
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 if (cachep->flags & SLAB_STORE_USER)
2499 *dbg_userword(cachep, objp) = NULL;
2500
2501 if (cachep->flags & SLAB_RED_ZONE) {
2502 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2503 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2504 }
2505 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002506 * Constructors are not allowed to allocate memory from the same
2507 * cache which they are a constructor for. Otherwise, deadlock.
2508 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002510 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2511 kasan_unpoison_object_data(cachep,
2512 objp + obj_offset(cachep));
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002513 cachep->ctor(objp + obj_offset(cachep));
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002514 kasan_poison_object_data(
2515 cachep, objp + obj_offset(cachep));
2516 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517
2518 if (cachep->flags & SLAB_RED_ZONE) {
2519 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002520 slab_error(cachep, "constructor overwrote the end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002522 slab_error(cachep, "constructor overwrote the start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 }
Joonsoo Kim40b44132016-03-15 14:54:21 -07002524 /* need to poison the objs? */
2525 if (cachep->flags & SLAB_POISON) {
2526 poison_obj(cachep, objp, POISON_FREE);
2527 slab_kernel_map(cachep, objp, 0, 0);
2528 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530#endif
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002531}
2532
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002533#ifdef CONFIG_SLAB_FREELIST_RANDOM
2534/* Hold information during a freelist initialization */
2535union freelist_init_state {
2536 struct {
2537 unsigned int pos;
2538 freelist_idx_t *list;
2539 unsigned int count;
2540 unsigned int rand;
2541 };
2542 struct rnd_state rnd_state;
2543};
2544
2545/*
2546 * Initialize the state based on the randomization methode available.
2547 * return true if the pre-computed list is available, false otherwize.
2548 */
2549static bool freelist_state_initialize(union freelist_init_state *state,
2550 struct kmem_cache *cachep,
2551 unsigned int count)
2552{
2553 bool ret;
2554 unsigned int rand;
2555
2556 /* Use best entropy available to define a random shift */
2557 get_random_bytes_arch(&rand, sizeof(rand));
2558
2559 /* Use a random state if the pre-computed list is not available */
2560 if (!cachep->random_seq) {
2561 prandom_seed_state(&state->rnd_state, rand);
2562 ret = false;
2563 } else {
2564 state->list = cachep->random_seq;
2565 state->count = count;
2566 state->pos = 0;
2567 state->rand = rand;
2568 ret = true;
2569 }
2570 return ret;
2571}
2572
2573/* Get the next entry on the list and randomize it using a random shift */
2574static freelist_idx_t next_random_slot(union freelist_init_state *state)
2575{
2576 return (state->list[state->pos++] + state->rand) % state->count;
2577}
2578
2579/*
2580 * Shuffle the freelist initialization state based on pre-computed lists.
2581 * return true if the list was successfully shuffled, false otherwise.
2582 */
2583static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2584{
2585 unsigned int objfreelist = 0, i, count = cachep->num;
2586 union freelist_init_state state;
2587 bool precomputed;
2588
2589 if (count < 2)
2590 return false;
2591
2592 precomputed = freelist_state_initialize(&state, cachep, count);
2593
2594 /* Take a random entry as the objfreelist */
2595 if (OBJFREELIST_SLAB(cachep)) {
2596 if (!precomputed)
2597 objfreelist = count - 1;
2598 else
2599 objfreelist = next_random_slot(&state);
2600 page->freelist = index_to_obj(cachep, page, objfreelist) +
2601 obj_offset(cachep);
2602 count--;
2603 }
2604
2605 /*
2606 * On early boot, generate the list dynamically.
2607 * Later use a pre-computed list for speed.
2608 */
2609 if (!precomputed) {
2610 freelist_randomize(&state.rnd_state, page->freelist, count);
2611 } else {
2612 for (i = 0; i < count; i++)
2613 set_free_obj(page, i, next_random_slot(&state));
2614 }
2615
2616 if (OBJFREELIST_SLAB(cachep))
2617 set_free_obj(page, cachep->num - 1, objfreelist);
2618
2619 return true;
2620}
2621#else
2622static inline bool shuffle_freelist(struct kmem_cache *cachep,
2623 struct page *page)
2624{
2625 return false;
2626}
2627#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2628
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002629static void cache_init_objs(struct kmem_cache *cachep,
2630 struct page *page)
2631{
2632 int i;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002633 void *objp;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002634 bool shuffled;
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002635
2636 cache_init_objs_debug(cachep, page);
2637
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002638 /* Try to randomize the freelist if enabled */
2639 shuffled = shuffle_freelist(cachep, page);
2640
2641 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002642 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2643 obj_offset(cachep);
2644 }
2645
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002646 for (i = 0; i < cachep->num; i++) {
2647 /* constructor could break poison info */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002648 if (DEBUG == 0 && cachep->ctor) {
2649 objp = index_to_obj(cachep, page, i);
2650 kasan_unpoison_object_data(cachep, objp);
2651 cachep->ctor(objp);
2652 kasan_poison_object_data(cachep, objp);
2653 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002654
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002655 if (!shuffled)
2656 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658}
2659
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002660static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002661{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002662 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002663
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002664 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002665 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002666
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002667#if DEBUG
2668 if (cachep->flags & SLAB_STORE_USER)
2669 set_store_user_dirty(cachep);
2670#endif
2671
Matthew Dobson78d382d2006-02-01 03:05:47 -08002672 return objp;
2673}
2674
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002675static void slab_put_obj(struct kmem_cache *cachep,
2676 struct page *page, void *objp)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002677{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002678 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002679#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002680 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002681
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002682 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002683 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002684 if (get_free_obj(page, i) == objnr) {
Joe Perches11705322016-03-17 14:19:50 -07002685 pr_err("slab: double free detected in cache '%s', objp %p\n",
Joe Perches756a025f02016-03-17 14:19:47 -07002686 cachep->name, objp);
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002687 BUG();
2688 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002689 }
2690#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002691 page->active--;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002692 if (!page->freelist)
2693 page->freelist = objp + obj_offset(cachep);
2694
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002695 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002696}
2697
Pekka Enberg47768742006-06-23 02:03:07 -07002698/*
2699 * Map pages beginning at addr to the given cache and slab. This is required
2700 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002701 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002702 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002703static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002704 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002706 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002707 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708}
2709
2710/*
2711 * Grow (by 1) the number of slabs within a cache. This is called by
2712 * kmem_cache_alloc() when there are no active objs left in a cache.
2713 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002714static struct page *cache_grow_begin(struct kmem_cache *cachep,
2715 gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002717 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002718 size_t offset;
2719 gfp_t local_flags;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002720 int page_node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002721 struct kmem_cache_node *n;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002722 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
Andrew Mortona737b3e2006-03-22 00:08:11 -08002724 /*
2725 * Be lazy and only check for valid flags here, keeping it out of the
2726 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002728 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2729 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2730 BUG();
2731 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002732 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 check_irq_off();
Mel Gormand0164ad2015-11-06 16:28:21 -08002735 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 local_irq_enable();
2737
2738 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002739 * Get mem for the objs. Attempt to allocate a physical page from
2740 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002741 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002742 page = kmem_getpages(cachep, local_flags, nodeid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002743 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 goto failed;
2745
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002746 page_node = page_to_nid(page);
2747 n = get_node(cachep, page_node);
Joonsoo Kim03d1d432016-05-19 17:10:20 -07002748
2749 /* Get colour for the slab, and cal the next value. */
2750 n->colour_next++;
2751 if (n->colour_next >= cachep->colour)
2752 n->colour_next = 0;
2753
2754 offset = n->colour_next;
2755 if (offset >= cachep->colour)
2756 offset = 0;
2757
2758 offset *= cachep->colour_off;
2759
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002761 freelist = alloc_slabmgmt(cachep, page, offset,
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002762 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002763 if (OFF_SLAB(cachep) && !freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 goto opps1;
2765
Joonsoo Kim8456a642013-10-24 10:07:49 +09002766 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002768 kasan_poison_slab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002769 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Mel Gormand0164ad2015-11-06 16:28:21 -08002771 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 local_irq_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002774 return page;
2775
Andrew Mortona737b3e2006-03-22 00:08:11 -08002776opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002777 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002778failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002779 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 local_irq_disable();
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002781 return NULL;
2782}
2783
2784static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2785{
2786 struct kmem_cache_node *n;
2787 void *list = NULL;
2788
2789 check_irq_off();
2790
2791 if (!page)
2792 return;
2793
2794 INIT_LIST_HEAD(&page->lru);
2795 n = get_node(cachep, page_to_nid(page));
2796
2797 spin_lock(&n->list_lock);
2798 if (!page->active)
2799 list_add_tail(&page->lru, &(n->slabs_free));
2800 else
2801 fixup_slab_list(cachep, n, page, &list);
2802 STATS_INC_GROWN(cachep);
2803 n->free_objects += cachep->num - page->active;
2804 spin_unlock(&n->list_lock);
2805
2806 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807}
2808
2809#if DEBUG
2810
2811/*
2812 * Perform extra freeing checks:
2813 * - detect bad pointers.
2814 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 */
2816static void kfree_debugcheck(const void *objp)
2817{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 if (!virt_addr_valid(objp)) {
Joe Perches11705322016-03-17 14:19:50 -07002819 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002820 (unsigned long)objp);
2821 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823}
2824
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002825static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2826{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002827 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002828
2829 redzone1 = *dbg_redzone1(cache, obj);
2830 redzone2 = *dbg_redzone2(cache, obj);
2831
2832 /*
2833 * Redzone is ok.
2834 */
2835 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2836 return;
2837
2838 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2839 slab_error(cache, "double free detected");
2840 else
2841 slab_error(cache, "memory outside object was overwritten");
2842
Joe Perches11705322016-03-17 14:19:50 -07002843 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2844 obj, redzone1, redzone2);
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002845}
2846
Pekka Enberg343e0d72006-02-01 03:05:50 -08002847static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002848 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002851 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002853 BUG_ON(virt_to_cache(objp) != cachep);
2854
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002855 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002857 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002860 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2862 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2863 }
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002864 if (cachep->flags & SLAB_STORE_USER) {
2865 set_store_user_dirty(cachep);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002866 *dbg_userword(cachep, objp) = (void *)caller;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002867 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868
Joonsoo Kim8456a642013-10-24 10:07:49 +09002869 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870
2871 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002872 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 poison_obj(cachep, objp, POISON_FREE);
Joonsoo Kim40b44132016-03-15 14:54:21 -07002876 slab_kernel_map(cachep, objp, 0, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 }
2878 return objp;
2879}
2880
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881#else
2882#define kfree_debugcheck(x) do { } while(0)
2883#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884#endif
2885
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002886static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2887 void **list)
2888{
2889#if DEBUG
2890 void *next = *list;
2891 void *objp;
2892
2893 while (next) {
2894 objp = next - obj_offset(cachep);
2895 next = *(void **)next;
2896 poison_obj(cachep, objp, POISON_FREE);
2897 }
2898#endif
2899}
2900
Joonsoo Kimd8410232016-03-15 14:54:44 -07002901static inline void fixup_slab_list(struct kmem_cache *cachep,
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002902 struct kmem_cache_node *n, struct page *page,
2903 void **list)
Joonsoo Kimd8410232016-03-15 14:54:44 -07002904{
2905 /* move slabp to correct slabp list: */
2906 list_del(&page->lru);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002907 if (page->active == cachep->num) {
Joonsoo Kimd8410232016-03-15 14:54:44 -07002908 list_add(&page->lru, &n->slabs_full);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002909 if (OBJFREELIST_SLAB(cachep)) {
2910#if DEBUG
2911 /* Poisoning will be done without holding the lock */
2912 if (cachep->flags & SLAB_POISON) {
2913 void **objp = page->freelist;
2914
2915 *objp = *list;
2916 *list = objp;
2917 }
2918#endif
2919 page->freelist = NULL;
2920 }
2921 } else
Joonsoo Kimd8410232016-03-15 14:54:44 -07002922 list_add(&page->lru, &n->slabs_partial);
2923}
2924
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002925/* Try to find non-pfmemalloc slab if needed */
2926static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2927 struct page *page, bool pfmemalloc)
2928{
2929 if (!page)
2930 return NULL;
2931
2932 if (pfmemalloc)
2933 return page;
2934
2935 if (!PageSlabPfmemalloc(page))
2936 return page;
2937
2938 /* No need to keep pfmemalloc slab if we have enough free objects */
2939 if (n->free_objects > n->free_limit) {
2940 ClearPageSlabPfmemalloc(page);
2941 return page;
2942 }
2943
2944 /* Move pfmemalloc slab to the end of list to speed up next search */
2945 list_del(&page->lru);
2946 if (!page->active)
2947 list_add_tail(&page->lru, &n->slabs_free);
2948 else
2949 list_add_tail(&page->lru, &n->slabs_partial);
2950
2951 list_for_each_entry(page, &n->slabs_partial, lru) {
2952 if (!PageSlabPfmemalloc(page))
2953 return page;
2954 }
2955
2956 list_for_each_entry(page, &n->slabs_free, lru) {
2957 if (!PageSlabPfmemalloc(page))
2958 return page;
2959 }
2960
2961 return NULL;
2962}
2963
2964static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
Geliang Tang7aa0d222016-01-14 15:18:02 -08002965{
2966 struct page *page;
2967
2968 page = list_first_entry_or_null(&n->slabs_partial,
2969 struct page, lru);
2970 if (!page) {
2971 n->free_touched = 1;
2972 page = list_first_entry_or_null(&n->slabs_free,
2973 struct page, lru);
2974 }
2975
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002976 if (sk_memalloc_socks())
2977 return get_valid_first_slab(n, page, pfmemalloc);
2978
Geliang Tang7aa0d222016-01-14 15:18:02 -08002979 return page;
2980}
2981
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002982static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2983 struct kmem_cache_node *n, gfp_t flags)
2984{
2985 struct page *page;
2986 void *obj;
2987 void *list = NULL;
2988
2989 if (!gfp_pfmemalloc_allowed(flags))
2990 return NULL;
2991
2992 spin_lock(&n->list_lock);
2993 page = get_first_slab(n, true);
2994 if (!page) {
2995 spin_unlock(&n->list_lock);
2996 return NULL;
2997 }
2998
2999 obj = slab_get_obj(cachep, page);
3000 n->free_objects--;
3001
3002 fixup_slab_list(cachep, n, page, &list);
3003
3004 spin_unlock(&n->list_lock);
3005 fixup_objfreelist_debug(cachep, &list);
3006
3007 return obj;
3008}
3009
Joonsoo Kim213b4692016-05-19 17:10:29 -07003010/*
3011 * Slab list should be fixed up by fixup_slab_list() for existing slab
3012 * or cache_grow_end() for new slab
3013 */
3014static __always_inline int alloc_block(struct kmem_cache *cachep,
3015 struct array_cache *ac, struct page *page, int batchcount)
3016{
3017 /*
3018 * There must be at least one object available for
3019 * allocation.
3020 */
3021 BUG_ON(page->active >= cachep->num);
3022
3023 while (page->active < cachep->num && batchcount--) {
3024 STATS_INC_ALLOCED(cachep);
3025 STATS_INC_ACTIVE(cachep);
3026 STATS_SET_HIGH(cachep);
3027
3028 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
3029 }
3030
3031 return batchcount;
3032}
3033
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003034static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035{
3036 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003037 struct kmem_cache_node *n;
Joonsoo Kim801faf02016-05-19 17:10:31 -07003038 struct array_cache *ac, *shared;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003039 int node;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003040 void *list = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003041 struct page *page;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003042
Joe Korty6d2144d2008-03-05 15:04:59 -08003043 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003044 node = numa_mem_id();
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003045
Joe Korty6d2144d2008-03-05 15:04:59 -08003046 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 batchcount = ac->batchcount;
3048 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003049 /*
3050 * If there was little recent activity on this cache, then
3051 * perform only a partial refill. Otherwise we could generate
3052 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 */
3054 batchcount = BATCHREFILL_LIMIT;
3055 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07003056 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003058 BUG_ON(ac->avail > 0 || !n);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003059 shared = READ_ONCE(n->shared);
3060 if (!n->free_objects && (!shared || !shared->avail))
3061 goto direct_grow;
3062
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003063 spin_lock(&n->list_lock);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003064 shared = READ_ONCE(n->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003065
Christoph Lameter3ded1752006-03-25 03:06:44 -08003066 /* See if we can refill from the shared array */
Joonsoo Kim801faf02016-05-19 17:10:31 -07003067 if (shared && transfer_objects(ac, shared, batchcount)) {
3068 shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003069 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003070 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003071
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 while (batchcount > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 /* Get slab alloc is to come from. */
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003074 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003075 if (!page)
3076 goto must_grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003079
Joonsoo Kim213b4692016-05-19 17:10:29 -07003080 batchcount = alloc_block(cachep, ac, page, batchcount);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003081 fixup_slab_list(cachep, n, page, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 }
3083
Andrew Mortona737b3e2006-03-22 00:08:11 -08003084must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003085 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003087 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003088 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089
Joonsoo Kim801faf02016-05-19 17:10:31 -07003090direct_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 if (unlikely(!ac->avail)) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003092 /* Check if we can use obj in pfmemalloc slab */
3093 if (sk_memalloc_socks()) {
3094 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3095
3096 if (obj)
3097 return obj;
3098 }
3099
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003100 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
Christoph Lametere498be72005-09-09 13:03:32 -07003101
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003102 /*
3103 * cache_grow_begin() can reenable interrupts,
3104 * then ac could change.
3105 */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003106 ac = cpu_cache_get(cachep);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003107 if (!ac->avail && page)
3108 alloc_block(cachep, ac, page, batchcount);
3109 cache_grow_end(cachep, page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003110
Joonsoo Kim213b4692016-05-19 17:10:29 -07003111 if (!ac->avail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 }
3114 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003115
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003116 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117}
3118
Andrew Mortona737b3e2006-03-22 00:08:11 -08003119static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3120 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121{
Mel Gormand0164ad2015-11-06 16:28:21 -08003122 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123}
3124
3125#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003126static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003127 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003129 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003131 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07003133 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 poison_obj(cachep, objp, POISON_INUSE);
3135 }
3136 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003137 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
3139 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003140 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3141 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
Joe Perches756a025f02016-03-17 14:19:47 -07003142 slab_error(cachep, "double free, or memory outside object was overwritten");
Joe Perches11705322016-03-17 14:19:50 -07003143 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3144 objp, *dbg_redzone1(cachep, objp),
3145 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 }
3147 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3148 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3149 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003150
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003151 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003152 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003153 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003154 if (ARCH_SLAB_MINALIGN &&
3155 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Joe Perches11705322016-03-17 14:19:50 -07003156 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003157 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003158 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 return objp;
3160}
3161#else
3162#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3163#endif
3164
Pekka Enberg343e0d72006-02-01 03:05:50 -08003165static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003167 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168 struct array_cache *ac;
3169
Alok N Kataria5c382302005-09-27 21:45:46 -07003170 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003171
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003172 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 ac->touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003175 objp = ac->entry[--ac->avail];
Mel Gorman072bb0a2012-07-31 16:43:58 -07003176
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003177 STATS_INC_ALLOCHIT(cachep);
3178 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003180
3181 STATS_INC_ALLOCMISS(cachep);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003182 objp = cache_alloc_refill(cachep, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003183 /*
3184 * the 'ac' may be updated by cache_alloc_refill(),
3185 * and kmemleak_erase() requires its correct value.
3186 */
3187 ac = cpu_cache_get(cachep);
3188
3189out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003190 /*
3191 * To avoid a false negative, if an object that is in one of the
3192 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3193 * treat the array pointers as a reference to the object.
3194 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003195 if (objp)
3196 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003197 return objp;
3198}
3199
Christoph Lametere498be72005-09-09 13:03:32 -07003200#ifdef CONFIG_NUMA
3201/*
Zefan Li2ad654b2014-09-25 09:41:02 +08003202 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003203 *
3204 * If we are in_interrupt, then process context, including cpusets and
3205 * mempolicy, may not apply and should not be used for allocation policy.
3206 */
3207static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3208{
3209 int nid_alloc, nid_here;
3210
Christoph Lameter765c4502006-09-27 01:50:08 -07003211 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003212 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003213 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003214 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003215 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003216 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003217 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003218 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003219 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003220 return NULL;
3221}
3222
3223/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003224 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003225 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003226 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003227 * perform an allocation without specifying a node. This allows the page
3228 * allocator to do its reclaim / fallback magic. We then insert the
3229 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003230 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003231static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003232{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003233 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07003234 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003235 struct zone *zone;
3236 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003237 void *obj = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003238 struct page *page;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003239 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003240 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003241
3242 if (flags & __GFP_THISNODE)
3243 return NULL;
3244
Mel Gormancc9a6c82012-03-21 16:34:11 -07003245retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003246 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003247 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003248
Christoph Lameter3c517a62006-12-06 20:33:29 -08003249retry:
3250 /*
3251 * Look through allowed nodes for objects available
3252 * from existing per node queues.
3253 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003254 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3255 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003256
Vladimir Davydov061d7072014-12-12 16:58:25 -08003257 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003258 get_node(cache, nid) &&
3259 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003260 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003261 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003262 if (obj)
3263 break;
3264 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003265 }
3266
Christoph Lametercfce6602007-05-06 14:50:17 -07003267 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003268 /*
3269 * This allocation will be performed within the constraints
3270 * of the current cpuset / memory policy requirements.
3271 * We may trigger various forms of reclaim on the allowed
3272 * set and go into memory reserves if necessary.
3273 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003274 page = cache_grow_begin(cache, flags, numa_mem_id());
3275 cache_grow_end(cache, page);
3276 if (page) {
3277 nid = page_to_nid(page);
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003278 obj = ____cache_alloc_node(cache,
3279 gfp_exact_node(flags), nid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003280
Christoph Lameter3c517a62006-12-06 20:33:29 -08003281 /*
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003282 * Another processor may allocate the objects in
3283 * the slab since we are not holding any locks.
Christoph Lameter3c517a62006-12-06 20:33:29 -08003284 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003285 if (!obj)
3286 goto retry;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003287 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003288 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003289
Mel Gormand26914d2014-04-03 14:47:24 -07003290 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003291 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003292 return obj;
3293}
3294
3295/*
Christoph Lametere498be72005-09-09 13:03:32 -07003296 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003298static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003299 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003300{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003301 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003302 struct kmem_cache_node *n;
Joonsoo Kim213b4692016-05-19 17:10:29 -07003303 void *obj = NULL;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003304 void *list = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003306 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003307 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003308 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003309
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003310 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003311 spin_lock(&n->list_lock);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003312 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003313 if (!page)
3314 goto must_grow;
Christoph Lametere498be72005-09-09 13:03:32 -07003315
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003316 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003317
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 STATS_INC_NODEALLOCS(cachep);
3319 STATS_INC_ACTIVE(cachep);
3320 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003321
Joonsoo Kim8456a642013-10-24 10:07:49 +09003322 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003323
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003324 obj = slab_get_obj(cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003325 n->free_objects--;
Christoph Lametere498be72005-09-09 13:03:32 -07003326
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003327 fixup_slab_list(cachep, n, page, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003328
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003329 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003330 fixup_objfreelist_debug(cachep, &list);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003331 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003332
Andrew Mortona737b3e2006-03-22 00:08:11 -08003333must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003334 spin_unlock(&n->list_lock);
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003335 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003336 if (page) {
3337 /* This slab isn't counted yet so don't update free_objects */
3338 obj = slab_get_obj(cachep, page);
3339 }
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003340 cache_grow_end(cachep, page);
Christoph Lametere498be72005-09-09 13:03:32 -07003341
Joonsoo Kim213b4692016-05-19 17:10:29 -07003342 return obj ? obj : fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003343}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003344
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003345static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003346slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003347 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003348{
3349 unsigned long save_flags;
3350 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003351 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003352
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003353 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003354 cachep = slab_pre_alloc_hook(cachep, flags);
3355 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003356 return NULL;
3357
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003358 cache_alloc_debugcheck_before(cachep, flags);
3359 local_irq_save(save_flags);
3360
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003361 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003362 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003363
Christoph Lameter18bf8542014-08-06 16:04:11 -07003364 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003365 /* Node not bootstrapped yet */
3366 ptr = fallback_alloc(cachep, flags);
3367 goto out;
3368 }
3369
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003370 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003371 /*
3372 * Use the locally cached objects if possible.
3373 * However ____cache_alloc does not allow fallback
3374 * to other nodes. It may fail while we still have
3375 * objects on other nodes available.
3376 */
3377 ptr = ____cache_alloc(cachep, flags);
3378 if (ptr)
3379 goto out;
3380 }
3381 /* ___cache_alloc_node can fall back to other nodes */
3382 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3383 out:
3384 local_irq_restore(save_flags);
3385 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3386
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003387 if (unlikely(flags & __GFP_ZERO) && ptr)
3388 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003389
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003390 slab_post_alloc_hook(cachep, flags, 1, &ptr);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003391 return ptr;
3392}
3393
3394static __always_inline void *
3395__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3396{
3397 void *objp;
3398
Zefan Li2ad654b2014-09-25 09:41:02 +08003399 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003400 objp = alternate_node_alloc(cache, flags);
3401 if (objp)
3402 goto out;
3403 }
3404 objp = ____cache_alloc(cache, flags);
3405
3406 /*
3407 * We may just have run out of memory on the local node.
3408 * ____cache_alloc_node() knows how to locate memory on other nodes
3409 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003410 if (!objp)
3411 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003412
3413 out:
3414 return objp;
3415}
3416#else
3417
3418static __always_inline void *
3419__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3420{
3421 return ____cache_alloc(cachep, flags);
3422}
3423
3424#endif /* CONFIG_NUMA */
3425
3426static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003427slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003428{
3429 unsigned long save_flags;
3430 void *objp;
3431
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003432 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003433 cachep = slab_pre_alloc_hook(cachep, flags);
3434 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003435 return NULL;
3436
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003437 cache_alloc_debugcheck_before(cachep, flags);
3438 local_irq_save(save_flags);
3439 objp = __do_cache_alloc(cachep, flags);
3440 local_irq_restore(save_flags);
3441 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3442 prefetchw(objp);
3443
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003444 if (unlikely(flags & __GFP_ZERO) && objp)
3445 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003446
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003447 slab_post_alloc_hook(cachep, flags, 1, &objp);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003448 return objp;
3449}
Christoph Lametere498be72005-09-09 13:03:32 -07003450
3451/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003452 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003453 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003454 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003455static void free_block(struct kmem_cache *cachep, void **objpp,
3456 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457{
3458 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003459 struct kmem_cache_node *n = get_node(cachep, node);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003460 struct page *page;
3461
3462 n->free_objects += nr_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
3464 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003465 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003466 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467
Mel Gorman072bb0a2012-07-31 16:43:58 -07003468 objp = objpp[i];
3469
Joonsoo Kim8456a642013-10-24 10:07:49 +09003470 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003471 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003472 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003473 slab_put_obj(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 STATS_DEC_ACTIVE(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475
3476 /* fixup slab chains */
Joonsoo Kim6052b782016-05-19 17:10:17 -07003477 if (page->active == 0)
3478 list_add(&page->lru, &n->slabs_free);
3479 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 /* Unconditionally move a slab to the end of the
3481 * partial list on free - maximum time for the
3482 * other objects to be freed, too.
3483 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003484 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 }
3486 }
Joonsoo Kim6052b782016-05-19 17:10:17 -07003487
3488 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3489 n->free_objects -= cachep->num;
3490
3491 page = list_last_entry(&n->slabs_free, struct page, lru);
3492 list_del(&page->lru);
3493 list_add(&page->lru, list);
3494 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495}
3496
Pekka Enberg343e0d72006-02-01 03:05:50 -08003497static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498{
3499 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003500 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003501 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003502 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
3504 batchcount = ac->batchcount;
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003505
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003507 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003508 spin_lock(&n->list_lock);
3509 if (n->shared) {
3510 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003511 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 if (max) {
3513 if (batchcount > max)
3514 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003515 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003516 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 shared_array->avail += batchcount;
3518 goto free_done;
3519 }
3520 }
3521
Joonsoo Kim97654df2014-08-06 16:04:25 -07003522 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003523free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524#if STATS
3525 {
3526 int i = 0;
Geliang Tang73c02192016-01-14 15:17:59 -08003527 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528
Geliang Tang73c02192016-01-14 15:17:59 -08003529 list_for_each_entry(page, &n->slabs_free, lru) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003530 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531
3532 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 }
3534 STATS_SET_FREEABLE(cachep, i);
3535 }
3536#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003537 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003538 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003540 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541}
3542
3543/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003544 * Release an obj back to its cache. If the obj has a constructed state, it must
3545 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003547static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003548 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549{
Alexander Potapenko55834c52016-05-20 16:59:11 -07003550 /* Put the object into the quarantine, don't touch it for now. */
3551 if (kasan_slab_free(cachep, objp))
3552 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553
Alexander Potapenko55834c52016-05-20 16:59:11 -07003554 ___cache_free(cachep, objp, caller);
3555}
3556
3557void ___cache_free(struct kmem_cache *cachep, void *objp,
3558 unsigned long caller)
3559{
3560 struct array_cache *ac = cpu_cache_get(cachep);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003561
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003563 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003564 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003566 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003567
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003568 /*
3569 * Skip calling cache_free_alien() when the platform is not numa.
3570 * This will avoid cache misses that happen while accessing slabp (which
3571 * is per page memory reference) to get nodeid. Instead use a global
3572 * variable to skip the call, which is mostly likely to be present in
3573 * the cache.
3574 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003575 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003576 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003577
Joonsoo Kim3d880192014-10-09 15:26:04 -07003578 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 } else {
3581 STATS_INC_FREEMISS(cachep);
3582 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003584
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003585 if (sk_memalloc_socks()) {
3586 struct page *page = virt_to_head_page(objp);
3587
3588 if (unlikely(PageSlabPfmemalloc(page))) {
3589 cache_free_pfmemalloc(cachep, page, objp);
3590 return;
3591 }
3592 }
3593
3594 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595}
3596
3597/**
3598 * kmem_cache_alloc - Allocate an object
3599 * @cachep: The cache to allocate from.
3600 * @flags: See kmalloc().
3601 *
3602 * Allocate an object from this cache. The flags are only relevant
3603 * if the cache has no available objects.
3604 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003605void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003607 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003608
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003609 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003610 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003611 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003612
3613 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614}
3615EXPORT_SYMBOL(kmem_cache_alloc);
3616
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003617static __always_inline void
3618cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3619 size_t size, void **p, unsigned long caller)
3620{
3621 size_t i;
3622
3623 for (i = 0; i < size; i++)
3624 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3625}
3626
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003627int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003628 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07003629{
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003630 size_t i;
3631
3632 s = slab_pre_alloc_hook(s, flags);
3633 if (!s)
3634 return 0;
3635
3636 cache_alloc_debugcheck_before(s, flags);
3637
3638 local_irq_disable();
3639 for (i = 0; i < size; i++) {
3640 void *objp = __do_cache_alloc(s, flags);
3641
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003642 if (unlikely(!objp))
3643 goto error;
3644 p[i] = objp;
3645 }
3646 local_irq_enable();
3647
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003648 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3649
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003650 /* Clear memory outside IRQ disabled section */
3651 if (unlikely(flags & __GFP_ZERO))
3652 for (i = 0; i < size; i++)
3653 memset(p[i], 0, s->object_size);
3654
3655 slab_post_alloc_hook(s, flags, size, p);
3656 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3657 return size;
3658error:
3659 local_irq_enable();
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003660 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003661 slab_post_alloc_hook(s, flags, i, p);
3662 __kmem_cache_free_bulk(s, i, p);
3663 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07003664}
3665EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3666
Li Zefan0f24f122009-12-11 15:45:30 +08003667#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003668void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003669kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003670{
Steven Rostedt85beb582010-11-24 16:23:34 -05003671 void *ret;
3672
Ezequiel Garcia48356302012-09-08 17:47:57 -03003673 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003674
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003675 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003676 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003677 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003678 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003679}
Steven Rostedt85beb582010-11-24 16:23:34 -05003680EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003681#endif
3682
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003684/**
3685 * kmem_cache_alloc_node - Allocate an object on the specified node
3686 * @cachep: The cache to allocate from.
3687 * @flags: See kmalloc().
3688 * @nodeid: node number of the target node.
3689 *
3690 * Identical to kmem_cache_alloc but it will allocate memory on the given
3691 * node, which can improve the performance for cpu bound structures.
3692 *
3693 * Fallback to other node is possible if __GFP_THISNODE is not set.
3694 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003695void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3696{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003697 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003698
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003699 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003700 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003701 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003702 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003703
3704 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706EXPORT_SYMBOL(kmem_cache_alloc_node);
3707
Li Zefan0f24f122009-12-11 15:45:30 +08003708#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003709void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003710 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003711 int nodeid,
3712 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003713{
Steven Rostedt85beb582010-11-24 16:23:34 -05003714 void *ret;
3715
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003716 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003717
3718 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003719 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003720 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003721 flags, nodeid);
3722 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003723}
Steven Rostedt85beb582010-11-24 16:23:34 -05003724EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003725#endif
3726
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003727static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003728__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003729{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003730 struct kmem_cache *cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003731 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003732
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003733 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003734 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3735 return cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003736 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003737 kasan_kmalloc(cachep, ret, size, flags);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003738
3739 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003740}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003741
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003742void *__kmalloc_node(size_t size, gfp_t flags, int node)
3743{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003744 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003745}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003746EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003747
3748void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003749 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003750{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003751 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003752}
3753EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003754#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755
3756/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003757 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003759 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003760 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003762static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003763 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003765 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003766 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003768 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003769 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3770 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003771 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003772
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003773 kasan_kmalloc(cachep, ret, size, flags);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003774 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003775 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003776
3777 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003778}
3779
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003780void *__kmalloc(size_t size, gfp_t flags)
3781{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003782 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783}
3784EXPORT_SYMBOL(__kmalloc);
3785
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003786void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003787{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003788 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003789}
3790EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003791
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792/**
3793 * kmem_cache_free - Deallocate an object
3794 * @cachep: The cache the allocation was from.
3795 * @objp: The previously allocated object.
3796 *
3797 * Free an object which was previously allocated from this
3798 * cache.
3799 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003800void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801{
3802 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003803 cachep = cache_from_obj(cachep, objp);
3804 if (!cachep)
3805 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806
3807 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003808 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003809 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003810 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003811 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003812 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003813
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003814 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003815}
3816EXPORT_SYMBOL(kmem_cache_free);
3817
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003818void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3819{
3820 struct kmem_cache *s;
3821 size_t i;
3822
3823 local_irq_disable();
3824 for (i = 0; i < size; i++) {
3825 void *objp = p[i];
3826
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07003827 if (!orig_s) /* called via kfree_bulk */
3828 s = virt_to_cache(objp);
3829 else
3830 s = cache_from_obj(orig_s, objp);
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003831
3832 debug_check_no_locks_freed(objp, s->object_size);
3833 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3834 debug_check_no_obj_freed(objp, s->object_size);
3835
3836 __cache_free(s, objp, _RET_IP_);
3837 }
3838 local_irq_enable();
3839
3840 /* FIXME: add tracing */
3841}
3842EXPORT_SYMBOL(kmem_cache_free_bulk);
3843
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 * kfree - free previously allocated memory
3846 * @objp: pointer returned by kmalloc.
3847 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003848 * If @objp is NULL, no operation is performed.
3849 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 * Don't free memory not originally allocated by kmalloc()
3851 * or you will run into trouble.
3852 */
3853void kfree(const void *objp)
3854{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003855 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 unsigned long flags;
3857
Pekka Enberg2121db72009-03-25 11:05:57 +02003858 trace_kfree(_RET_IP_, objp);
3859
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003860 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 return;
3862 local_irq_save(flags);
3863 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003864 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003865 debug_check_no_locks_freed(objp, c->object_size);
3866
3867 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003868 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 local_irq_restore(flags);
3870}
3871EXPORT_SYMBOL(kfree);
3872
Christoph Lametere498be72005-09-09 13:03:32 -07003873/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003874 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003875 */
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003876static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003877{
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003878 int ret;
Christoph Lametere498be72005-09-09 13:03:32 -07003879 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003880 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07003881
Mel Gorman9c09a952008-01-24 05:49:54 -08003882 for_each_online_node(node) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003883 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3884 if (ret)
Christoph Lametere498be72005-09-09 13:03:32 -07003885 goto fail;
3886
Christoph Lametere498be72005-09-09 13:03:32 -07003887 }
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003888
Christoph Lametercafeb022006-03-25 03:06:46 -08003889 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003890
Andrew Mortona737b3e2006-03-22 00:08:11 -08003891fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003892 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003893 /* Cache is not active yet. Roll back what we did */
3894 node--;
3895 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003896 n = get_node(cachep, node);
3897 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003898 kfree(n->shared);
3899 free_alien_cache(n->alien);
3900 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003901 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003902 }
3903 node--;
3904 }
3905 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003906 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003907}
3908
Christoph Lameter18004c52012-07-06 15:25:12 -05003909/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003910static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003911 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003913 struct array_cache __percpu *cpu_cache, *prev;
3914 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003916 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3917 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003918 return -ENOMEM;
3919
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003920 prev = cachep->cpu_cache;
3921 cachep->cpu_cache = cpu_cache;
3922 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003923
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 cachep->batchcount = batchcount;
3926 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003927 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003929 if (!prev)
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003930 goto setup_node;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003931
3932 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003933 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003934 int node;
3935 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003936 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003937
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003938 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003939 n = get_node(cachep, node);
3940 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003941 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003942 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003943 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003945 free_percpu(prev);
3946
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003947setup_node:
3948 return setup_kmem_cache_nodes(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949}
3950
Glauber Costa943a4512012-12-18 14:23:03 -08003951static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3952 int batchcount, int shared, gfp_t gfp)
3953{
3954 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003955 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003956
3957 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3958
3959 if (slab_state < FULL)
3960 return ret;
3961
3962 if ((ret < 0) || !is_root_cache(cachep))
3963 return ret;
3964
Vladimir Davydov426589f2015-02-12 14:59:23 -08003965 lockdep_assert_held(&slab_mutex);
3966 for_each_memcg_cache(c, cachep) {
3967 /* return value determined by the root cache only */
3968 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003969 }
3970
3971 return ret;
3972}
3973
Christoph Lameter18004c52012-07-06 15:25:12 -05003974/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003975static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976{
3977 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003978 int limit = 0;
3979 int shared = 0;
3980 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003982 err = cache_random_seq_create(cachep, gfp);
3983 if (err)
3984 goto end;
3985
Glauber Costa943a4512012-12-18 14:23:03 -08003986 if (!is_root_cache(cachep)) {
3987 struct kmem_cache *root = memcg_root_cache(cachep);
3988 limit = root->limit;
3989 shared = root->shared;
3990 batchcount = root->batchcount;
3991 }
3992
3993 if (limit && shared && batchcount)
3994 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003995 /*
3996 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 * - create a LIFO ordering, i.e. return objects that are cache-warm
3998 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003999 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 * bufctl chains: array operations are cheaper.
4001 * The numbers are guessed, we should auto-tune as described by
4002 * Bonwick.
4003 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004004 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004006 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004008 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004010 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 limit = 54;
4012 else
4013 limit = 120;
4014
Andrew Mortona737b3e2006-03-22 00:08:11 -08004015 /*
4016 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 * allocation behaviour: Most allocs on one cpu, most free operations
4018 * on another cpu. For these cases, an efficient object passing between
4019 * cpus is necessary. This is provided by a shared array. The array
4020 * replaces Bonwick's magazine layer.
4021 * On uniprocessor, it's functionally equivalent (but less efficient)
4022 * to a larger limit. Thus disabled by default.
4023 */
4024 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004025 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
4028#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004029 /*
4030 * With debugging enabled, large batchcount lead to excessively long
4031 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032 */
4033 if (limit > 32)
4034 limit = 32;
4035#endif
Glauber Costa943a4512012-12-18 14:23:03 -08004036 batchcount = (limit + 1) / 2;
4037skip_setup:
4038 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07004039end:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 if (err)
Joe Perches11705322016-03-17 14:19:50 -07004041 pr_err("enable_cpucache failed for %s, error %d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004042 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004043 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044}
4045
Christoph Lameter1b552532006-03-22 00:09:07 -08004046/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004047 * Drain an array if it contains any elements taking the node lock only if
4048 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004049 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004050 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004051static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004052 struct array_cache *ac, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053{
Joonsoo Kim97654df2014-08-06 16:04:25 -07004054 LIST_HEAD(list);
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004055
4056 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4057 check_mutex_acquired();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058
Christoph Lameter1b552532006-03-22 00:09:07 -08004059 if (!ac || !ac->avail)
4060 return;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004061
4062 if (ac->touched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 ac->touched = 0;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004064 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 }
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004066
4067 spin_lock_irq(&n->list_lock);
4068 drain_array_locked(cachep, ac, node, false, &list);
4069 spin_unlock_irq(&n->list_lock);
4070
4071 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072}
4073
4074/**
4075 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004076 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077 *
4078 * Called from workqueue/eventd every few seconds.
4079 * Purpose:
4080 * - clear the per-cpu caches for this CPU.
4081 * - return freeable pages to the main free memory pool.
4082 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004083 * If we cannot acquire the cache chain mutex then just give up - we'll try
4084 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004086static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004088 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004089 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004090 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004091 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092
Christoph Lameter18004c52012-07-06 15:25:12 -05004093 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004095 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096
Christoph Lameter18004c52012-07-06 15:25:12 -05004097 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 check_irq_on();
4099
Christoph Lameter35386e32006-03-22 00:09:05 -08004100 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004101 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004102 * have established with reasonable certainty that
4103 * we can do some work if the lock was obtained.
4104 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07004105 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08004106
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004107 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004109 drain_array(searchp, n, cpu_cache_get(searchp), node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110
Christoph Lameter35386e32006-03-22 00:09:05 -08004111 /*
4112 * These are racy checks but it does not matter
4113 * if we skip one check or scan twice.
4114 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004115 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004116 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004118 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004120 drain_array(searchp, n, n->shared, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004122 if (n->free_touched)
4123 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004124 else {
4125 int freed;
4126
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004127 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004128 5 * searchp->num - 1) / (5 * searchp->num));
4129 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004131next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 cond_resched();
4133 }
4134 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004135 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004136 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004137out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004138 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004139 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140}
4141
Linus Torvalds158a9622008-01-02 13:04:48 -08004142#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004143void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004145 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004146 unsigned long active_objs;
4147 unsigned long num_objs;
4148 unsigned long active_slabs = 0;
4149 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004150 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004152 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004153 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 active_objs = 0;
4156 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07004157 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07004158
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004159 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004160 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004161
Joonsoo Kim8456a642013-10-24 10:07:49 +09004162 list_for_each_entry(page, &n->slabs_full, lru) {
4163 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004164 error = "slabs_full accounting error";
4165 active_objs += cachep->num;
4166 active_slabs++;
4167 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004168 list_for_each_entry(page, &n->slabs_partial, lru) {
4169 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004170 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004171 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004172 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004173 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004174 active_slabs++;
4175 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004176 list_for_each_entry(page, &n->slabs_free, lru) {
4177 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004178 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004179 num_slabs++;
4180 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004181 free_objects += n->free_objects;
4182 if (n->shared)
4183 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004184
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004185 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004187 num_slabs += active_slabs;
4188 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004189 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 error = "free_objects accounting error";
4191
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004192 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193 if (error)
Joe Perches11705322016-03-17 14:19:50 -07004194 pr_err("slab: cache %s error: %s\n", name, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004195
Glauber Costa0d7561c2012-10-19 18:20:27 +04004196 sinfo->active_objs = active_objs;
4197 sinfo->num_objs = num_objs;
4198 sinfo->active_slabs = active_slabs;
4199 sinfo->num_slabs = num_slabs;
4200 sinfo->shared_avail = shared_avail;
4201 sinfo->limit = cachep->limit;
4202 sinfo->batchcount = cachep->batchcount;
4203 sinfo->shared = cachep->shared;
4204 sinfo->objects_per_slab = cachep->num;
4205 sinfo->cache_order = cachep->gfporder;
4206}
4207
4208void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4209{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004211 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212 unsigned long high = cachep->high_mark;
4213 unsigned long allocs = cachep->num_allocations;
4214 unsigned long grown = cachep->grown;
4215 unsigned long reaped = cachep->reaped;
4216 unsigned long errors = cachep->errors;
4217 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004219 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004220 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221
Joe Perches756a025f02016-03-17 14:19:47 -07004222 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
Joe Perchese92dd4f2010-03-26 19:27:58 -07004223 allocs, high, grown,
4224 reaped, errors, max_freeable, node_allocs,
4225 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 }
4227 /* cpu stats */
4228 {
4229 unsigned long allochit = atomic_read(&cachep->allochit);
4230 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4231 unsigned long freehit = atomic_read(&cachep->freehit);
4232 unsigned long freemiss = atomic_read(&cachep->freemiss);
4233
4234 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004235 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 }
4237#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238}
4239
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240#define MAX_SLABINFO_WRITE 128
4241/**
4242 * slabinfo_write - Tuning for the slab allocator
4243 * @file: unused
4244 * @buffer: user buffer
4245 * @count: data length
4246 * @ppos: unused
4247 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004248ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004249 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004251 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004253 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004254
Linus Torvalds1da177e2005-04-16 15:20:36 -07004255 if (count > MAX_SLABINFO_WRITE)
4256 return -EINVAL;
4257 if (copy_from_user(&kbuf, buffer, count))
4258 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004259 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260
4261 tmp = strchr(kbuf, ' ');
4262 if (!tmp)
4263 return -EINVAL;
4264 *tmp = '\0';
4265 tmp++;
4266 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4267 return -EINVAL;
4268
4269 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004270 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004271 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004272 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004274 if (limit < 1 || batchcount < 1 ||
4275 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004276 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004278 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004279 batchcount, shared,
4280 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281 }
4282 break;
4283 }
4284 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004285 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 if (res >= 0)
4287 res = count;
4288 return res;
4289}
Al Viro871751e2006-03-25 03:06:39 -08004290
4291#ifdef CONFIG_DEBUG_SLAB_LEAK
4292
Al Viro871751e2006-03-25 03:06:39 -08004293static inline int add_caller(unsigned long *n, unsigned long v)
4294{
4295 unsigned long *p;
4296 int l;
4297 if (!v)
4298 return 1;
4299 l = n[1];
4300 p = n + 2;
4301 while (l) {
4302 int i = l/2;
4303 unsigned long *q = p + 2 * i;
4304 if (*q == v) {
4305 q[1]++;
4306 return 1;
4307 }
4308 if (*q > v) {
4309 l = i;
4310 } else {
4311 p = q + 2;
4312 l -= i + 1;
4313 }
4314 }
4315 if (++n[1] == n[0])
4316 return 0;
4317 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4318 p[0] = v;
4319 p[1] = 1;
4320 return 1;
4321}
4322
Joonsoo Kim8456a642013-10-24 10:07:49 +09004323static void handle_slab(unsigned long *n, struct kmem_cache *c,
4324 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004325{
4326 void *p;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004327 int i, j;
4328 unsigned long v;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004329
Al Viro871751e2006-03-25 03:06:39 -08004330 if (n[0] == n[1])
4331 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004332 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004333 bool active = true;
4334
4335 for (j = page->active; j < c->num; j++) {
4336 if (get_free_obj(page, j) == i) {
4337 active = false;
4338 break;
4339 }
4340 }
4341
4342 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004343 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004344
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004345 /*
4346 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4347 * mapping is established when actual object allocation and
4348 * we could mistakenly access the unmapped object in the cpu
4349 * cache.
4350 */
4351 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4352 continue;
4353
4354 if (!add_caller(n, v))
Al Viro871751e2006-03-25 03:06:39 -08004355 return;
4356 }
4357}
4358
4359static void show_symbol(struct seq_file *m, unsigned long address)
4360{
4361#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004362 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004363 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004364
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004365 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004366 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004367 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004368 seq_printf(m, " [%s]", modname);
4369 return;
4370 }
4371#endif
4372 seq_printf(m, "%p", (void *)address);
4373}
4374
4375static int leaks_show(struct seq_file *m, void *p)
4376{
Thierry Reding0672aa72012-06-22 19:42:49 +02004377 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004378 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004379 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004380 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004381 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004382 int node;
4383 int i;
4384
4385 if (!(cachep->flags & SLAB_STORE_USER))
4386 return 0;
4387 if (!(cachep->flags & SLAB_RED_ZONE))
4388 return 0;
4389
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004390 /*
4391 * Set store_user_clean and start to grab stored user information
4392 * for all objects on this cache. If some alloc/free requests comes
4393 * during the processing, information would be wrong so restart
4394 * whole processing.
4395 */
4396 do {
4397 set_store_user_clean(cachep);
4398 drain_cpu_caches(cachep);
Al Viro871751e2006-03-25 03:06:39 -08004399
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004400 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004401
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004402 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004403
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004404 check_irq_on();
4405 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004406
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004407 list_for_each_entry(page, &n->slabs_full, lru)
4408 handle_slab(x, cachep, page);
4409 list_for_each_entry(page, &n->slabs_partial, lru)
4410 handle_slab(x, cachep, page);
4411 spin_unlock_irq(&n->list_lock);
4412 }
4413 } while (!is_store_user_clean(cachep));
4414
Al Viro871751e2006-03-25 03:06:39 -08004415 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004416 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004417 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004418 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004419 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004420 if (!m->private) {
4421 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004422 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004423 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004424 return -ENOMEM;
4425 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004426 *(unsigned long *)m->private = x[0] * 2;
4427 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004428 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004429 /* Now make sure this entry will be retried */
4430 m->count = m->size;
4431 return 0;
4432 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004433 for (i = 0; i < x[1]; i++) {
4434 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4435 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004436 seq_putc(m, '\n');
4437 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004438
Al Viro871751e2006-03-25 03:06:39 -08004439 return 0;
4440}
4441
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004442static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004443 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004444 .next = slab_next,
4445 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004446 .show = leaks_show,
4447};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004448
4449static int slabstats_open(struct inode *inode, struct file *file)
4450{
Rob Jonesb208ce32014-10-09 15:28:03 -07004451 unsigned long *n;
4452
4453 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4454 if (!n)
4455 return -ENOMEM;
4456
4457 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4458
4459 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004460}
4461
4462static const struct file_operations proc_slabstats_operations = {
4463 .open = slabstats_open,
4464 .read = seq_read,
4465 .llseek = seq_lseek,
4466 .release = seq_release_private,
4467};
Al Viro871751e2006-03-25 03:06:39 -08004468#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004469
4470static int __init slab_proc_init(void)
4471{
4472#ifdef CONFIG_DEBUG_SLAB_LEAK
4473 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4474#endif
4475 return 0;
4476}
4477module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004478#endif
4479
Manfred Spraul00e145b2005-09-03 15:55:07 -07004480/**
4481 * ksize - get the actual amount of memory allocated for a given object
4482 * @objp: Pointer to the object
4483 *
4484 * kmalloc may internally round up allocations and return more memory
4485 * than requested. ksize() can be used to determine the actual amount of
4486 * memory allocated. The caller may use this additional memory, even though
4487 * a smaller amount of memory was initially specified with the kmalloc call.
4488 * The caller must guarantee that objp points to a valid object previously
4489 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4490 * must not be freed during the duration of the call.
4491 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004492size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493{
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004494 size_t size;
4495
Christoph Lameteref8b4522007-10-16 01:24:46 -07004496 BUG_ON(!objp);
4497 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004498 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004499
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004500 size = virt_to_cache(objp)->object_size;
4501 /* We assume that ksize callers could use the whole allocated area,
4502 * so we need to unpoison this area.
4503 */
Alexander Potapenko4ebb31a42016-05-20 16:59:14 -07004504 kasan_unpoison_shadow(objp, size);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004505
4506 return size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004507}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004508EXPORT_SYMBOL(ksize);