blob: 0b0550ca85b40c9108b4085873d67babc200885d [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;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -0700236 parent->num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700237}
238
Andrew Mortona737b3e2006-03-22 00:08:11 -0800239#define MAKE_LIST(cachep, listp, slab, nodeid) \
240 do { \
241 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700242 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700243 } while (0)
244
Andrew Mortona737b3e2006-03-22 00:08:11 -0800245#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
246 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700247 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
249 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
250 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700252#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253#define CFLGS_OFF_SLAB (0x80000000UL)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700254#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
256
257#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800258/*
259 * Optimization question: fewer reaps means less probability for unnessary
260 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100262 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 * which could lock up otherwise freeable slabs.
264 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800265#define REAPTIMEOUT_AC (2*HZ)
266#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
268#if STATS
269#define STATS_INC_ACTIVE(x) ((x)->num_active++)
270#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
271#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
272#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700273#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274#define STATS_SET_HIGH(x) \
275 do { \
276 if ((x)->num_active > (x)->high_mark) \
277 (x)->high_mark = (x)->num_active; \
278 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279#define STATS_INC_ERR(x) ((x)->errors++)
280#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700281#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700282#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800283#define STATS_SET_FREEABLE(x, i) \
284 do { \
285 if ((x)->max_freeable < i) \
286 (x)->max_freeable = i; \
287 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
289#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
290#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
291#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
292#else
293#define STATS_INC_ACTIVE(x) do { } while (0)
294#define STATS_DEC_ACTIVE(x) do { } while (0)
295#define STATS_INC_ALLOCED(x) do { } while (0)
296#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700297#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298#define STATS_SET_HIGH(x) do { } while (0)
299#define STATS_INC_ERR(x) do { } while (0)
300#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700301#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700302#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800303#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304#define STATS_INC_ALLOCHIT(x) do { } while (0)
305#define STATS_INC_ALLOCMISS(x) do { } while (0)
306#define STATS_INC_FREEHIT(x) do { } while (0)
307#define STATS_INC_FREEMISS(x) do { } while (0)
308#endif
309
310#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
Andrew Mortona737b3e2006-03-22 00:08:11 -0800312/*
313 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800315 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 * the end of an object is aligned with the end of the real
317 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800318 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800320 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500321 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
322 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800323 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800325static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800327 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328}
329
David Woodhouseb46b8f12007-05-08 00:22:59 -0700330static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
332 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700333 return (unsigned long long*) (objp + obj_offset(cachep) -
334 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335}
336
David Woodhouseb46b8f12007-05-08 00:22:59 -0700337static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338{
339 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
340 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500341 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700342 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400343 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500344 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700345 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346}
347
Pekka Enberg343e0d72006-02-01 03:05:50 -0800348static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349{
350 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500351 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352}
353
354#else
355
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800356#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700357#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
360
361#endif
362
Joonsoo Kim03787302014-06-23 13:22:06 -0700363#ifdef CONFIG_DEBUG_SLAB_LEAK
364
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700365static inline bool is_store_user_clean(struct kmem_cache *cachep)
Joonsoo Kim03787302014-06-23 13:22:06 -0700366{
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700367 return atomic_read(&cachep->store_user_clean) == 1;
368}
Joonsoo Kim03787302014-06-23 13:22:06 -0700369
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700370static inline void set_store_user_clean(struct kmem_cache *cachep)
371{
372 atomic_set(&cachep->store_user_clean, 1);
373}
Joonsoo Kim03787302014-06-23 13:22:06 -0700374
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700375static inline void set_store_user_dirty(struct kmem_cache *cachep)
376{
377 if (is_store_user_clean(cachep))
378 atomic_set(&cachep->store_user_clean, 0);
Joonsoo Kim03787302014-06-23 13:22:06 -0700379}
380
381#else
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700382static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
Joonsoo Kim03787302014-06-23 13:22:06 -0700383
384#endif
385
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700387 * Do not go above this order unless 0 objects fit into the slab or
388 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 */
David Rientjes543585c2011-10-18 22:09:24 -0700390#define SLAB_MAX_ORDER_HI 1
391#define SLAB_MAX_ORDER_LO 0
392static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700393static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800395static inline struct kmem_cache *virt_to_cache(const void *obj)
396{
Christoph Lameterb49af682007-05-06 14:49:41 -0700397 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500398 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800399}
400
Joonsoo Kim8456a642013-10-24 10:07:49 +0900401static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800402 unsigned int idx)
403{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900404 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800405}
406
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800407/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500408 * We want to avoid an expensive divide : (offset / cache->size)
409 * Using the fact that size is a constant for a particular cache,
410 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800411 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
412 */
413static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900414 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800415{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900416 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800417 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800418}
419
Joonsoo Kim6fb92432016-03-15 14:54:09 -0700420#define BOOT_CPUCACHE_ENTRIES 1
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000422static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800423 .batchcount = 1,
424 .limit = BOOT_CPUCACHE_ENTRIES,
425 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500426 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800427 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428};
429
Tejun Heo1871e522009-10-29 22:34:13 +0900430static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Pekka Enberg343e0d72006-02-01 03:05:50 -0800432static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700434 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435}
436
Andrew Mortona737b3e2006-03-22 00:08:11 -0800437/*
438 * Calculate the number of objects and left-over bytes for a given buffer size.
439 */
Joonsoo Kim70f75062016-03-15 14:54:53 -0700440static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
441 unsigned long flags, size_t *left_over)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800442{
Joonsoo Kim70f75062016-03-15 14:54:53 -0700443 unsigned int num;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800444 size_t slab_size = PAGE_SIZE << gfporder;
445
446 /*
447 * The slab management structure can be either off the slab or
448 * on it. For the latter case, the memory allocated for a
449 * slab is used for:
450 *
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800451 * - @buffer_size bytes for each object
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700452 * - One freelist_idx_t for each object
453 *
454 * We don't need to consider alignment of freelist because
455 * freelist will be at the end of slab page. The objects will be
456 * at the correct alignment.
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800457 *
458 * If the slab management structure is off the slab, then the
459 * alignment will already be calculated into the size. Because
460 * the slabs are all pages aligned, the objects will be at the
461 * correct alignment when allocated.
462 */
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700463 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700464 num = slab_size / buffer_size;
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700465 *left_over = slab_size % buffer_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800466 } else {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700467 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700468 *left_over = slab_size %
469 (buffer_size + sizeof(freelist_idx_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 }
Joonsoo Kim70f75062016-03-15 14:54:53 -0700471
472 return num;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
Christoph Lameterf28510d2012-09-11 19:49:38 +0000475#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700476#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477
Andrew Mortona737b3e2006-03-22 00:08:11 -0800478static void __slab_error(const char *function, struct kmem_cache *cachep,
479 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480{
Joe Perches11705322016-03-17 14:19:50 -0700481 pr_err("slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800482 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030484 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000486#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487
Paul Menage3395ee02006-12-06 20:32:16 -0800488/*
489 * By default on NUMA we use alien caches to stage the freeing of
490 * objects allocated from other nodes. This causes massive memory
491 * inefficiencies when using fake NUMA setup to split memory into a
492 * large number of small nodes, so it can be disabled on the command
493 * line
494 */
495
496static int use_alien_caches __read_mostly = 1;
497static int __init noaliencache_setup(char *s)
498{
499 use_alien_caches = 0;
500 return 1;
501}
502__setup("noaliencache", noaliencache_setup);
503
David Rientjes3df1ccc2011-10-18 22:09:28 -0700504static int __init slab_max_order_setup(char *str)
505{
506 get_option(&str, &slab_max_order);
507 slab_max_order = slab_max_order < 0 ? 0 :
508 min(slab_max_order, MAX_ORDER - 1);
509 slab_max_order_set = true;
510
511 return 1;
512}
513__setup("slab_max_order=", slab_max_order_setup);
514
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800515#ifdef CONFIG_NUMA
516/*
517 * Special reaping functions for NUMA systems called from cache_reap().
518 * These take care of doing round robin flushing of alien caches (containing
519 * objects freed on different nodes from which they were allocated) and the
520 * flushing of remote pcps by calling drain_node_pages.
521 */
Tejun Heo1871e522009-10-29 22:34:13 +0900522static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800523
524static void init_reap_node(int cpu)
525{
Andrew Morton0edaf862016-05-19 17:10:58 -0700526 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
527 node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800528}
529
530static void next_reap_node(void)
531{
Christoph Lameter909ea962010-12-08 16:22:55 +0100532 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800533
Andrew Morton0edaf862016-05-19 17:10:58 -0700534 node = next_node_in(node, node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100535 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800536}
537
538#else
539#define init_reap_node(cpu) do { } while (0)
540#define next_reap_node(void) do { } while (0)
541#endif
542
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543/*
544 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
545 * via the workqueue/eventd.
546 * Add the CPU number into the expiration time to minimize the possibility of
547 * the CPUs getting into lockstep and contending for the global cache chain
548 * lock.
549 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400550static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551{
Tejun Heo1871e522009-10-29 22:34:13 +0900552 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553
554 /*
555 * When this gets called from do_initcalls via cpucache_init(),
556 * init_workqueues() has already run, so keventd will be setup
557 * at that time.
558 */
David Howells52bad642006-11-22 14:54:01 +0000559 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800560 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700561 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800562 schedule_delayed_work_on(cpu, reap_work,
563 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 }
565}
566
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700567static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100569 /*
570 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300571 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100572 * cache the pointers are not cleared and they could be counted as
573 * valid references during a kmemleak scan. Therefore, kmemleak must
574 * not scan such objects.
575 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700576 kmemleak_no_scan(ac);
577 if (ac) {
578 ac->avail = 0;
579 ac->limit = limit;
580 ac->batchcount = batch;
581 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700583}
584
585static struct array_cache *alloc_arraycache(int node, int entries,
586 int batchcount, gfp_t gfp)
587{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700588 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700589 struct array_cache *ac = NULL;
590
591 ac = kmalloc_node(memsize, gfp, node);
592 init_arraycache(ac, entries, batchcount);
593 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594}
595
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700596static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
597 struct page *page, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700598{
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700599 struct kmem_cache_node *n;
600 int page_node;
601 LIST_HEAD(list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700602
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700603 page_node = page_to_nid(page);
604 n = get_node(cachep, page_node);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700605
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700606 spin_lock(&n->list_lock);
607 free_block(cachep, &objp, 1, page_node, &list);
608 spin_unlock(&n->list_lock);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700609
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700610 slabs_destroy(cachep, &list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700611}
612
Christoph Lameter3ded1752006-03-25 03:06:44 -0800613/*
614 * Transfer objects in one arraycache to another.
615 * Locking must be handled by the caller.
616 *
617 * Return the number of entries transferred.
618 */
619static int transfer_objects(struct array_cache *to,
620 struct array_cache *from, unsigned int max)
621{
622 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700623 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800624
625 if (!nr)
626 return 0;
627
628 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
629 sizeof(void *) *nr);
630
631 from->avail -= nr;
632 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800633 return nr;
634}
635
Christoph Lameter765c4502006-09-27 01:50:08 -0700636#ifndef CONFIG_NUMA
637
638#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000639#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700640
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700641static inline struct alien_cache **alloc_alien_cache(int node,
642 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700643{
Joonsoo Kim88881772016-05-19 17:10:05 -0700644 return NULL;
Christoph Lameter765c4502006-09-27 01:50:08 -0700645}
646
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700647static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700648{
649}
650
651static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
652{
653 return 0;
654}
655
656static inline void *alternate_node_alloc(struct kmem_cache *cachep,
657 gfp_t flags)
658{
659 return NULL;
660}
661
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800662static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700663 gfp_t flags, int nodeid)
664{
665 return NULL;
666}
667
David Rientjes4167e9b2015-04-14 15:46:55 -0700668static inline gfp_t gfp_exact_node(gfp_t flags)
669{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700670 return flags & ~__GFP_NOFAIL;
David Rientjes4167e9b2015-04-14 15:46:55 -0700671}
672
Christoph Lameter765c4502006-09-27 01:50:08 -0700673#else /* CONFIG_NUMA */
674
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800675static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800676static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800677
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700678static struct alien_cache *__alloc_alien_cache(int node, int entries,
679 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700680{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700681 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700682 struct alien_cache *alc = NULL;
683
684 alc = kmalloc_node(memsize, gfp, node);
685 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700686 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700687 return alc;
688}
689
690static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
691{
692 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700693 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700694 int i;
695
696 if (limit > 1)
697 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700698 alc_ptr = kzalloc_node(memsize, gfp, node);
699 if (!alc_ptr)
700 return NULL;
701
702 for_each_node(i) {
703 if (i == node || !node_online(i))
704 continue;
705 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
706 if (!alc_ptr[i]) {
707 for (i--; i >= 0; i--)
708 kfree(alc_ptr[i]);
709 kfree(alc_ptr);
710 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700711 }
712 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700713 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700714}
715
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700716static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700717{
718 int i;
719
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700720 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700721 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700722 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700723 kfree(alc_ptr[i]);
724 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700725}
726
Pekka Enberg343e0d72006-02-01 03:05:50 -0800727static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700728 struct array_cache *ac, int node,
729 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700730{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700731 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700732
733 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000734 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800735 /*
736 * Stuff objects into the remote nodes shared array first.
737 * That way we could avoid the overhead of putting the objects
738 * into the free lists and getting them back later.
739 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000740 if (n->shared)
741 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800742
Joonsoo Kim833b7062014-08-06 16:04:33 -0700743 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700744 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000745 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700746 }
747}
748
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800749/*
750 * Called from cache_reap() to regularly drain alien caches round robin.
751 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000752static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800753{
Christoph Lameter909ea962010-12-08 16:22:55 +0100754 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800755
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000756 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700757 struct alien_cache *alc = n->alien[node];
758 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800759
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700760 if (alc) {
761 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700762 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700763 LIST_HEAD(list);
764
765 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700766 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700767 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700768 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800769 }
770 }
771}
772
Andrew Mortona737b3e2006-03-22 00:08:11 -0800773static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700774 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700775{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800776 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700777 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700778 struct array_cache *ac;
779 unsigned long flags;
780
781 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700782 alc = alien[i];
783 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700784 LIST_HEAD(list);
785
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700786 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700787 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700788 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700789 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700790 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700791 }
792 }
793}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700794
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700795static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
796 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700797{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000798 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700799 struct alien_cache *alien = NULL;
800 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700801 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700802
Christoph Lameter18bf8542014-08-06 16:04:11 -0700803 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700804 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700805 if (n->alien && n->alien[page_node]) {
806 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700807 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700808 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700809 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700810 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700811 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700812 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700813 ac->entry[ac->avail++] = objp;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700814 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700815 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700816 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700817 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700818 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700819 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700820 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -0700821 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700822 }
823 return 1;
824}
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700825
826static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
827{
828 int page_node = page_to_nid(virt_to_page(objp));
829 int node = numa_mem_id();
830 /*
831 * Make sure we are not freeing a object from another node to the array
832 * cache on this cpu.
833 */
834 if (likely(node == page_node))
835 return 0;
836
837 return __cache_free_alien(cachep, objp, node, page_node);
838}
David Rientjes4167e9b2015-04-14 15:46:55 -0700839
840/*
Mel Gorman444eb2a42016-03-17 14:19:23 -0700841 * Construct gfp mask to allocate from a specific node but do not reclaim or
842 * warn about failures.
David Rientjes4167e9b2015-04-14 15:46:55 -0700843 */
844static inline gfp_t gfp_exact_node(gfp_t flags)
845{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700846 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
David Rientjes4167e9b2015-04-14 15:46:55 -0700847}
Christoph Lametere498be72005-09-09 13:03:32 -0700848#endif
849
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700850static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
851{
852 struct kmem_cache_node *n;
853
854 /*
855 * Set up the kmem_cache_node for cpu before we can
856 * begin anything. Make sure some other cpu on this
857 * node has not already allocated this
858 */
859 n = get_node(cachep, node);
860 if (n) {
861 spin_lock_irq(&n->list_lock);
862 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
863 cachep->num;
864 spin_unlock_irq(&n->list_lock);
865
866 return 0;
867 }
868
869 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
870 if (!n)
871 return -ENOMEM;
872
873 kmem_cache_node_init(n);
874 n->next_reap = jiffies + REAPTIMEOUT_NODE +
875 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
876
877 n->free_limit =
878 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
879
880 /*
881 * The kmem_cache_nodes don't come and go as CPUs
882 * come and go. slab_mutex is sufficient
883 * protection here.
884 */
885 cachep->node[node] = n;
886
887 return 0;
888}
889
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200890#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700891/*
Christoph Lameter6a673682013-01-10 19:14:19 +0000892 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000893 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -0700894 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +0000895 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -0700896 * already in use.
897 *
Christoph Lameter18004c52012-07-06 15:25:12 -0500898 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -0700899 */
Christoph Lameter6a673682013-01-10 19:14:19 +0000900static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700901{
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700902 int ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700903 struct kmem_cache *cachep;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700904
Christoph Lameter18004c52012-07-06 15:25:12 -0500905 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700906 ret = init_cache_node(cachep, node, GFP_KERNEL);
907 if (ret)
908 return ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700909 }
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700910
David Rientjes8f9f8d92010-03-27 19:40:47 -0700911 return 0;
912}
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200913#endif
David Rientjes8f9f8d92010-03-27 19:40:47 -0700914
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700915static int setup_kmem_cache_node(struct kmem_cache *cachep,
916 int node, gfp_t gfp, bool force_change)
917{
918 int ret = -ENOMEM;
919 struct kmem_cache_node *n;
920 struct array_cache *old_shared = NULL;
921 struct array_cache *new_shared = NULL;
922 struct alien_cache **new_alien = NULL;
923 LIST_HEAD(list);
924
925 if (use_alien_caches) {
926 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
927 if (!new_alien)
928 goto fail;
929 }
930
931 if (cachep->shared) {
932 new_shared = alloc_arraycache(node,
933 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
934 if (!new_shared)
935 goto fail;
936 }
937
938 ret = init_cache_node(cachep, node, gfp);
939 if (ret)
940 goto fail;
941
942 n = get_node(cachep, node);
943 spin_lock_irq(&n->list_lock);
944 if (n->shared && force_change) {
945 free_block(cachep, n->shared->entry,
946 n->shared->avail, node, &list);
947 n->shared->avail = 0;
948 }
949
950 if (!n->shared || force_change) {
951 old_shared = n->shared;
952 n->shared = new_shared;
953 new_shared = NULL;
954 }
955
956 if (!n->alien) {
957 n->alien = new_alien;
958 new_alien = NULL;
959 }
960
961 spin_unlock_irq(&n->list_lock);
962 slabs_destroy(cachep, &list);
963
Joonsoo Kim801faf02016-05-19 17:10:31 -0700964 /*
965 * To protect lockless access to n->shared during irq disabled context.
966 * If n->shared isn't NULL in irq disabled context, accessing to it is
967 * guaranteed to be valid until irq is re-enabled, because it will be
968 * freed after synchronize_sched().
969 */
Joonsoo Kim86d9f482016-10-27 17:46:18 -0700970 if (old_shared && force_change)
Joonsoo Kim801faf02016-05-19 17:10:31 -0700971 synchronize_sched();
972
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700973fail:
974 kfree(old_shared);
975 kfree(new_shared);
976 free_alien_cache(new_alien);
977
978 return ret;
979}
980
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200981#ifdef CONFIG_SMP
982
Paul Gortmaker0db06282013-06-19 14:53:51 -0400983static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984{
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700985 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000986 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700987 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +1030988 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700989
Christoph Lameter18004c52012-07-06 15:25:12 -0500990 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700991 struct array_cache *nc;
992 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700993 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700994 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700995
Christoph Lameter18bf8542014-08-06 16:04:11 -0700996 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000997 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700998 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700999
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001000 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001001
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001002 /* Free limit for this kmem_cache_node */
1003 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001004
1005 /* cpu is dead; no one can alloc from it. */
1006 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1007 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001008 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001009 nc->avail = 0;
1010 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001011
Rusty Russell58463c12009-12-17 11:43:12 -06001012 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001013 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001014 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001015 }
1016
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001017 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001018 if (shared) {
1019 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001020 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001021 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001022 }
1023
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001024 alien = n->alien;
1025 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001026
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001027 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001028
1029 kfree(shared);
1030 if (alien) {
1031 drain_alien_cache(cachep, alien);
1032 free_alien_cache(alien);
1033 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001034
1035free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001036 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001037 }
1038 /*
1039 * In the previous loop, all the objects were freed to
1040 * the respective cache's slabs, now we can go ahead and
1041 * shrink each nodelist to its limit.
1042 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001043 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001044 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001045 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001046 continue;
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001047 drain_freelist(cachep, n, INT_MAX);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001048 }
1049}
1050
Paul Gortmaker0db06282013-06-19 14:53:51 -04001051static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001052{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001053 struct kmem_cache *cachep;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001054 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001055 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001057 /*
1058 * We need to do this right in the beginning since
1059 * alloc_arraycache's are going to use this list.
1060 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001061 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001062 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001063 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001064 if (err < 0)
1065 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001066
1067 /*
1068 * Now we can go ahead with allocating the shared arrays and
1069 * array caches
1070 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001071 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07001072 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1073 if (err)
1074 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001075 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001076
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001077 return 0;
1078bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001079 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001080 return -ENOMEM;
1081}
1082
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001083int slab_prepare_cpu(unsigned int cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001084{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001085 int err;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001086
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001087 mutex_lock(&slab_mutex);
1088 err = cpuup_prepare(cpu);
1089 mutex_unlock(&slab_mutex);
1090 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091}
1092
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001093/*
1094 * This is called for a failed online attempt and for a successful
1095 * offline.
1096 *
1097 * Even if all the cpus of a node are down, we don't free the
1098 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1099 * a kmalloc allocation from another cpu for memory from the node of
1100 * the cpu going down. The list3 structure is usually allocated from
1101 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1102 */
1103int slab_dead_cpu(unsigned int cpu)
1104{
1105 mutex_lock(&slab_mutex);
1106 cpuup_canceled(cpu);
1107 mutex_unlock(&slab_mutex);
1108 return 0;
1109}
1110#endif
1111
1112static int slab_online_cpu(unsigned int cpu)
1113{
1114 start_cpu_timer(cpu);
1115 return 0;
1116}
1117
1118static int slab_offline_cpu(unsigned int cpu)
1119{
1120 /*
1121 * Shutdown cache reaper. Note that the slab_mutex is held so
1122 * that if cache_reap() is invoked it cannot do anything
1123 * expensive but will only modify reap_work and reschedule the
1124 * timer.
1125 */
1126 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1127 /* Now the cache_reaper is guaranteed to be not running. */
1128 per_cpu(slab_reap_work, cpu).work.func = NULL;
1129 return 0;
1130}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
David Rientjes8f9f8d92010-03-27 19:40:47 -07001132#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1133/*
1134 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1135 * Returns -EBUSY if all objects cannot be drained so that the node is not
1136 * removed.
1137 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001138 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001139 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001140static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001141{
1142 struct kmem_cache *cachep;
1143 int ret = 0;
1144
Christoph Lameter18004c52012-07-06 15:25:12 -05001145 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001146 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001147
Christoph Lameter18bf8542014-08-06 16:04:11 -07001148 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001149 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001150 continue;
1151
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001152 drain_freelist(cachep, n, INT_MAX);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001153
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001154 if (!list_empty(&n->slabs_full) ||
1155 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001156 ret = -EBUSY;
1157 break;
1158 }
1159 }
1160 return ret;
1161}
1162
1163static int __meminit slab_memory_callback(struct notifier_block *self,
1164 unsigned long action, void *arg)
1165{
1166 struct memory_notify *mnb = arg;
1167 int ret = 0;
1168 int nid;
1169
1170 nid = mnb->status_change_nid;
1171 if (nid < 0)
1172 goto out;
1173
1174 switch (action) {
1175 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001176 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001177 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001178 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001179 break;
1180 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001181 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001182 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001183 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001184 break;
1185 case MEM_ONLINE:
1186 case MEM_OFFLINE:
1187 case MEM_CANCEL_ONLINE:
1188 case MEM_CANCEL_OFFLINE:
1189 break;
1190 }
1191out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001192 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001193}
1194#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1195
Christoph Lametere498be72005-09-09 13:03:32 -07001196/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001197 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001198 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001199static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001200 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001201{
Christoph Lameter6744f082013-01-10 19:12:17 +00001202 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001203
Christoph Lameter6744f082013-01-10 19:12:17 +00001204 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001205 BUG_ON(!ptr);
1206
Christoph Lameter6744f082013-01-10 19:12:17 +00001207 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001208 /*
1209 * Do not assume that spinlocks can be initialized via memcpy:
1210 */
1211 spin_lock_init(&ptr->list_lock);
1212
Christoph Lametere498be72005-09-09 13:03:32 -07001213 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001214 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001215}
1216
Andrew Mortona737b3e2006-03-22 00:08:11 -08001217/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001218 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1219 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001220 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001221static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001222{
1223 int node;
1224
1225 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001226 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001227 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001228 REAPTIMEOUT_NODE +
1229 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001230 }
1231}
1232
1233/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001234 * Initialisation. Called after the page allocator have been initialised and
1235 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 */
1237void __init kmem_cache_init(void)
1238{
Christoph Lametere498be72005-09-09 13:03:32 -07001239 int i;
1240
Joonsoo Kim68126702013-10-24 10:07:42 +09001241 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1242 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001243 kmem_cache = &kmem_cache_boot;
1244
Joonsoo Kim88881772016-05-19 17:10:05 -07001245 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001246 use_alien_caches = 0;
1247
Christoph Lameter3c583462012-11-28 16:23:01 +00001248 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001249 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001250
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 /*
1252 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001253 * page orders on machines with more than 32MB of memory if
1254 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001256 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001257 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 /* Bootstrap is tricky, because several objects are allocated
1260 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001261 * 1) initialize the kmem_cache cache: it contains the struct
1262 * kmem_cache structures of all caches, except kmem_cache itself:
1263 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001264 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001265 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001266 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001268 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001269 * An __init data area is used for the head array.
1270 * 3) Create the remaining kmalloc caches, with minimally sized
1271 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001272 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001274 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001275 * the other cache's with kmalloc allocated memory.
1276 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 */
1278
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001279 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
Eric Dumazet8da34302007-05-06 14:49:29 -07001281 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001282 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001283 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001284 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001285 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001286 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001287 SLAB_HWCACHE_ALIGN);
1288 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001289 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Andrew Mortona737b3e2006-03-22 00:08:11 -08001291 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001292 * Initialize the caches that provide memory for the kmem_cache_node
1293 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001294 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001295 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001296 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001297 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001298 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001299
Ingo Molnare0a42722006-06-23 02:03:46 -07001300 slab_early_init = 0;
1301
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001302 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001303 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001304 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305
Mel Gorman9c09a952008-01-24 05:49:54 -08001306 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001307 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001308
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001309 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001310 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001311 }
1312 }
1313
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001314 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001315}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001316
Pekka Enberg8429db52009-06-12 15:58:59 +03001317void __init kmem_cache_init_late(void)
1318{
1319 struct kmem_cache *cachep;
1320
Christoph Lameter97d06602012-07-06 15:25:11 -05001321 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001322
Pekka Enberg8429db52009-06-12 15:58:59 +03001323 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001324 mutex_lock(&slab_mutex);
1325 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001326 if (enable_cpucache(cachep, GFP_NOWAIT))
1327 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001328 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001329
Christoph Lameter97d06602012-07-06 15:25:11 -05001330 /* Done! */
1331 slab_state = FULL;
1332
David Rientjes8f9f8d92010-03-27 19:40:47 -07001333#ifdef CONFIG_NUMA
1334 /*
1335 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001336 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001337 */
1338 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1339#endif
1340
Andrew Mortona737b3e2006-03-22 00:08:11 -08001341 /*
1342 * The reap timers are started later, with a module init call: That part
1343 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 */
1345}
1346
1347static int __init cpucache_init(void)
1348{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001349 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350
Andrew Mortona737b3e2006-03-22 00:08:11 -08001351 /*
1352 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 */
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001354 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1355 slab_online_cpu, slab_offline_cpu);
1356 WARN_ON(ret < 0);
Glauber Costaa164f8962012-06-21 00:59:18 +04001357
1358 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001359 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 return 0;
1361}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362__initcall(cpucache_init);
1363
Rafael Aquini8bdec192012-03-09 17:27:27 -03001364static noinline void
1365slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1366{
David Rientjes9a02d692014-06-04 16:06:36 -07001367#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001368 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001369 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001370 unsigned long flags;
1371 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001372 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1373 DEFAULT_RATELIMIT_BURST);
1374
1375 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1376 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001377
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001378 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1379 nodeid, gfpflags, &gfpflags);
1380 pr_warn(" cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001381 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001382
Christoph Lameter18bf8542014-08-06 16:04:11 -07001383 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001384 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1385 unsigned long active_slabs = 0, num_slabs = 0;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07001386 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
1387 unsigned long num_slabs_full;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001388
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001389 spin_lock_irqsave(&n->list_lock, flags);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07001390 num_slabs = n->num_slabs;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001391 list_for_each_entry(page, &n->slabs_partial, lru) {
1392 active_objs += page->active;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07001393 num_slabs_partial++;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001394 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001395 list_for_each_entry(page, &n->slabs_free, lru)
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07001396 num_slabs_free++;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001397
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001398 free_objects += n->free_objects;
1399 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001400
Rafael Aquini8bdec192012-03-09 17:27:27 -03001401 num_objs = num_slabs * cachep->num;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07001402 active_slabs = num_slabs - num_slabs_free;
1403 num_slabs_full = num_slabs -
1404 (num_slabs_partial + num_slabs_free);
1405 active_objs += (num_slabs_full * cachep->num);
1406
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001407 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
Rafael Aquini8bdec192012-03-09 17:27:27 -03001408 node, active_slabs, num_slabs, active_objs, num_objs,
1409 free_objects);
1410 }
David Rientjes9a02d692014-06-04 16:06:36 -07001411#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001412}
1413
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001415 * Interface to system's page allocator. No need to hold the
1416 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417 *
1418 * If we requested dmaable memory, we will get it. Even if we
1419 * did not request dmaable memory, we might get it, but that
1420 * would be relatively rare and ignorable.
1421 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001422static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1423 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424{
1425 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001426 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001427
Glauber Costaa618e892012-06-14 16:17:21 +04001428 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001429 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1430 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001431
Vlastimil Babka96db8002015-09-08 15:03:50 -07001432 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001433 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001434 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001436 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001438 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1439 __free_pages(page, cachep->gfporder);
1440 return NULL;
1441 }
1442
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001443 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001445 add_zone_page_state(page_zone(page),
1446 NR_SLAB_RECLAIMABLE, nr_pages);
1447 else
1448 add_zone_page_state(page_zone(page),
1449 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001450
Joonsoo Kima57a4982013-10-24 10:07:44 +09001451 __SetPageSlab(page);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001452 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1453 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001454 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001455
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001456 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1457 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1458
1459 if (cachep->ctor)
1460 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1461 else
1462 kmemcheck_mark_unallocated_pages(page, nr_pages);
1463 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001464
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001465 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466}
1467
1468/*
1469 * Interface to system's page release.
1470 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001471static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472{
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001473 int order = cachep->gfporder;
1474 unsigned long nr_freed = (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001476 kmemcheck_free_shadow(page, order);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001477
Christoph Lameter972d1a72006-09-25 23:31:51 -07001478 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1479 sub_zone_page_state(page_zone(page),
1480 NR_SLAB_RECLAIMABLE, nr_freed);
1481 else
1482 sub_zone_page_state(page_zone(page),
1483 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001484
Joonsoo Kima57a4982013-10-24 10:07:44 +09001485 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001486 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001487 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001488 page_mapcount_reset(page);
1489 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001490
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 if (current->reclaim_state)
1492 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001493 memcg_uncharge_slab(page, order, cachep);
1494 __free_pages(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495}
1496
1497static void kmem_rcu_free(struct rcu_head *head)
1498{
Joonsoo Kim68126702013-10-24 10:07:42 +09001499 struct kmem_cache *cachep;
1500 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501
Joonsoo Kim68126702013-10-24 10:07:42 +09001502 page = container_of(head, struct page, rcu_head);
1503 cachep = page->slab_cache;
1504
1505 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506}
1507
1508#if DEBUG
Joonsoo Kim40b44132016-03-15 14:54:21 -07001509static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1510{
1511 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1512 (cachep->size % PAGE_SIZE) == 0)
1513 return true;
1514
1515 return false;
1516}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
1518#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001519static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001520 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001522 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001524 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001526 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 return;
1528
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001529 *addr++ = 0x12345678;
1530 *addr++ = caller;
1531 *addr++ = smp_processor_id();
1532 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 {
1534 unsigned long *sptr = &caller;
1535 unsigned long svalue;
1536
1537 while (!kstack_end(sptr)) {
1538 svalue = *sptr++;
1539 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001540 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 size -= sizeof(unsigned long);
1542 if (size <= sizeof(unsigned long))
1543 break;
1544 }
1545 }
1546
1547 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001548 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549}
Joonsoo Kim40b44132016-03-15 14:54:21 -07001550
1551static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1552 int map, unsigned long caller)
1553{
1554 if (!is_debug_pagealloc_cache(cachep))
1555 return;
1556
1557 if (caller)
1558 store_stackinfo(cachep, objp, caller);
1559
1560 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1561}
1562
1563#else
1564static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1565 int map, unsigned long caller) {}
1566
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567#endif
1568
Pekka Enberg343e0d72006-02-01 03:05:50 -08001569static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001571 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001572 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573
1574 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001575 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576}
1577
1578static void dump_line(char *data, int offset, int limit)
1579{
1580 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001581 unsigned char error = 0;
1582 int bad_count = 0;
1583
Joe Perches11705322016-03-17 14:19:50 -07001584 pr_err("%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001585 for (i = 0; i < limit; i++) {
1586 if (data[offset + i] != POISON_FREE) {
1587 error = data[offset + i];
1588 bad_count++;
1589 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001590 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001591 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1592 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001593
1594 if (bad_count == 1) {
1595 error ^= POISON_FREE;
1596 if (!(error & (error - 1))) {
Joe Perches11705322016-03-17 14:19:50 -07001597 pr_err("Single bit error detected. Probably bad RAM.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001598#ifdef CONFIG_X86
Joe Perches11705322016-03-17 14:19:50 -07001599 pr_err("Run memtest86+ or a similar memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001600#else
Joe Perches11705322016-03-17 14:19:50 -07001601 pr_err("Run a memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001602#endif
1603 }
1604 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605}
1606#endif
1607
1608#if DEBUG
1609
Pekka Enberg343e0d72006-02-01 03:05:50 -08001610static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611{
1612 int i, size;
1613 char *realobj;
1614
1615 if (cachep->flags & SLAB_RED_ZONE) {
Joe Perches11705322016-03-17 14:19:50 -07001616 pr_err("Redzone: 0x%llx/0x%llx\n",
1617 *dbg_redzone1(cachep, objp),
1618 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 }
1620
1621 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches11705322016-03-17 14:19:50 -07001622 pr_err("Last user: [<%p>](%pSR)\n",
Joe Perches071361d2012-12-12 10:19:12 -08001623 *dbg_userword(cachep, objp),
1624 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001626 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001627 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001628 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 int limit;
1630 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001631 if (i + limit > size)
1632 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 dump_line(realobj, i, limit);
1634 }
1635}
1636
Pekka Enberg343e0d72006-02-01 03:05:50 -08001637static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638{
1639 char *realobj;
1640 int size, i;
1641 int lines = 0;
1642
Joonsoo Kim40b44132016-03-15 14:54:21 -07001643 if (is_debug_pagealloc_cache(cachep))
1644 return;
1645
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001646 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001647 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001649 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 exp = POISON_END;
1653 if (realobj[i] != exp) {
1654 int limit;
1655 /* Mismatch ! */
1656 /* Print header */
1657 if (lines == 0) {
Joe Perches11705322016-03-17 14:19:50 -07001658 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1659 print_tainted(), cachep->name,
1660 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 print_objinfo(cachep, objp, 0);
1662 }
1663 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001664 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001666 if (i + limit > size)
1667 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 dump_line(realobj, i, limit);
1669 i += 16;
1670 lines++;
1671 /* Limit to 5 lines */
1672 if (lines > 5)
1673 break;
1674 }
1675 }
1676 if (lines != 0) {
1677 /* Print some data about the neighboring objects, if they
1678 * exist:
1679 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001680 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001681 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682
Joonsoo Kim8456a642013-10-24 10:07:49 +09001683 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001685 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001686 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001687 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 print_objinfo(cachep, objp, 2);
1689 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001690 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001691 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001692 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001693 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 print_objinfo(cachep, objp, 2);
1695 }
1696 }
1697}
1698#endif
1699
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001701static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1702 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001703{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 int i;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001705
1706 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1707 poison_obj(cachep, page->freelist - obj_offset(cachep),
1708 POISON_FREE);
1709 }
1710
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001712 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
1714 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07001716 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 }
1718 if (cachep->flags & SLAB_RED_ZONE) {
1719 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001720 slab_error(cachep, "start of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001722 slab_error(cachep, "end of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001725}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001727static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1728 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001729{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001730}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731#endif
1732
Randy Dunlap911851e2006-03-22 00:08:14 -08001733/**
1734 * slab_destroy - destroy and release all objects in a slab
1735 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001736 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001737 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001738 * Destroy all the objs in a slab page, and release the mem back to the system.
1739 * Before calling the slab page must have been unlinked from the cache. The
1740 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001741 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001742static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001743{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001744 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001745
Joonsoo Kim8456a642013-10-24 10:07:49 +09001746 freelist = page->freelist;
1747 slab_destroy_debugcheck(cachep, page);
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001748 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1749 call_rcu(&page->rcu_head, kmem_rcu_free);
1750 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001751 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001752
1753 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001754 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001755 * although actual page can be freed in rcu context
1756 */
1757 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001758 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759}
1760
Joonsoo Kim97654df2014-08-06 16:04:25 -07001761static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1762{
1763 struct page *page, *n;
1764
1765 list_for_each_entry_safe(page, n, list, lru) {
1766 list_del(&page->lru);
1767 slab_destroy(cachep, page);
1768 }
1769}
1770
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001772 * calculate_slab_order - calculate size (page order) of slabs
1773 * @cachep: pointer to the cache that is being created
1774 * @size: size of objects to be created in this cache.
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001775 * @flags: slab allocation flags
1776 *
1777 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001778 *
1779 * This could be made much more intelligent. For now, try to avoid using
1780 * high order pages for slabs. When the gfp() functions are more friendly
1781 * towards high-order requests, this should be changed.
1782 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001783static size_t calculate_slab_order(struct kmem_cache *cachep,
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07001784 size_t size, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001785{
1786 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001787 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001788
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001789 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001790 unsigned int num;
1791 size_t remainder;
1792
Joonsoo Kim70f75062016-03-15 14:54:53 -07001793 num = cache_estimate(gfporder, size, flags, &remainder);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001794 if (!num)
1795 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001796
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001797 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1798 if (num > SLAB_OBJ_MAX_NUM)
1799 break;
1800
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001801 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001802 struct kmem_cache *freelist_cache;
1803 size_t freelist_size;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001804
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001805 freelist_size = num * sizeof(freelist_idx_t);
1806 freelist_cache = kmalloc_slab(freelist_size, 0u);
1807 if (!freelist_cache)
1808 continue;
1809
1810 /*
1811 * Needed to avoid possible looping condition
Joonsoo Kim76b342b2016-05-19 17:10:26 -07001812 * in cache_grow_begin()
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001813 */
1814 if (OFF_SLAB(freelist_cache))
1815 continue;
1816
1817 /* check if off slab has enough benefit */
1818 if (freelist_cache->size > cachep->size / 2)
1819 continue;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001820 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001821
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001822 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001823 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001824 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001825 left_over = remainder;
1826
1827 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001828 * A VFS-reclaimable slab tends to have most allocations
1829 * as GFP_NOFS and we really don't want to have to be allocating
1830 * higher-order pages when we are unable to shrink dcache.
1831 */
1832 if (flags & SLAB_RECLAIM_ACCOUNT)
1833 break;
1834
1835 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001836 * Large number of objects is good, but very large slabs are
1837 * currently bad for the gfp()s.
1838 */
David Rientjes543585c2011-10-18 22:09:24 -07001839 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001840 break;
1841
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001842 /*
1843 * Acceptable internal fragmentation?
1844 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001845 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001846 break;
1847 }
1848 return left_over;
1849}
1850
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001851static struct array_cache __percpu *alloc_kmem_cache_cpus(
1852 struct kmem_cache *cachep, int entries, int batchcount)
1853{
1854 int cpu;
1855 size_t size;
1856 struct array_cache __percpu *cpu_cache;
1857
1858 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001859 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001860
1861 if (!cpu_cache)
1862 return NULL;
1863
1864 for_each_possible_cpu(cpu) {
1865 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1866 entries, batchcount);
1867 }
1868
1869 return cpu_cache;
1870}
1871
Fabian Frederickbd721ea2016-08-02 14:03:33 -07001872static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001873{
Christoph Lameter97d06602012-07-06 15:25:11 -05001874 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001875 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001876
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001877 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1878 if (!cachep->cpu_cache)
1879 return 1;
1880
Christoph Lameter97d06602012-07-06 15:25:11 -05001881 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001882 /* Creation of first cache (kmem_cache). */
1883 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001884 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001885 /* For kmem_cache_node */
1886 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001887 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001888 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001889
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001890 for_each_online_node(node) {
1891 cachep->node[node] = kmalloc_node(
1892 sizeof(struct kmem_cache_node), gfp, node);
1893 BUG_ON(!cachep->node[node]);
1894 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001895 }
1896 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001897
Christoph Lameter6a673682013-01-10 19:14:19 +00001898 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001899 jiffies + REAPTIMEOUT_NODE +
1900 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001901
1902 cpu_cache_get(cachep)->avail = 0;
1903 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1904 cpu_cache_get(cachep)->batchcount = 1;
1905 cpu_cache_get(cachep)->touched = 0;
1906 cachep->batchcount = 1;
1907 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001908 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001909}
1910
Joonsoo Kim12220de2014-10-09 15:26:24 -07001911unsigned long kmem_cache_flags(unsigned long object_size,
1912 unsigned long flags, const char *name,
1913 void (*ctor)(void *))
1914{
1915 return flags;
1916}
1917
1918struct kmem_cache *
1919__kmem_cache_alias(const char *name, size_t size, size_t align,
1920 unsigned long flags, void (*ctor)(void *))
1921{
1922 struct kmem_cache *cachep;
1923
1924 cachep = find_mergeable(size, align, flags, name, ctor);
1925 if (cachep) {
1926 cachep->refcount++;
1927
1928 /*
1929 * Adjust the object sizes so that we clear
1930 * the complete object on kzalloc.
1931 */
1932 cachep->object_size = max_t(int, cachep->object_size, size);
1933 }
1934 return cachep;
1935}
1936
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001937static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1938 size_t size, unsigned long flags)
1939{
1940 size_t left;
1941
1942 cachep->num = 0;
1943
1944 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1945 return false;
1946
1947 left = calculate_slab_order(cachep, size,
1948 flags | CFLGS_OBJFREELIST_SLAB);
1949 if (!cachep->num)
1950 return false;
1951
1952 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1953 return false;
1954
1955 cachep->colour = left / cachep->colour_off;
1956
1957 return true;
1958}
1959
Joonsoo Kim158e3192016-03-15 14:54:35 -07001960static bool set_off_slab_cache(struct kmem_cache *cachep,
1961 size_t size, unsigned long flags)
1962{
1963 size_t left;
1964
1965 cachep->num = 0;
1966
1967 /*
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001968 * Always use on-slab management when SLAB_NOLEAKTRACE
1969 * to avoid recursive calls into kmemleak.
Joonsoo Kim158e3192016-03-15 14:54:35 -07001970 */
Joonsoo Kim158e3192016-03-15 14:54:35 -07001971 if (flags & SLAB_NOLEAKTRACE)
1972 return false;
1973
1974 /*
1975 * Size is large, assume best to place the slab management obj
1976 * off-slab (should allow better packing of objs).
1977 */
1978 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1979 if (!cachep->num)
1980 return false;
1981
1982 /*
1983 * If the slab has been placed off-slab, and we have enough space then
1984 * move it on-slab. This is at the expense of any extra colouring.
1985 */
1986 if (left >= cachep->num * sizeof(freelist_idx_t))
1987 return false;
1988
1989 cachep->colour = left / cachep->colour_off;
1990
1991 return true;
1992}
1993
1994static bool set_on_slab_cache(struct kmem_cache *cachep,
1995 size_t size, unsigned long flags)
1996{
1997 size_t left;
1998
1999 cachep->num = 0;
2000
2001 left = calculate_slab_order(cachep, size, flags);
2002 if (!cachep->num)
2003 return false;
2004
2005 cachep->colour = left / cachep->colour_off;
2006
2007 return true;
2008}
2009
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002011 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002012 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 *
2015 * Returns a ptr to the cache on success, NULL on failure.
2016 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002017 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 * The flags are
2020 *
2021 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2022 * to catch references to uninitialised memory.
2023 *
2024 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2025 * for buffer overruns.
2026 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2028 * cacheline. This can be beneficial if you're counting cycles as closely
2029 * as davem.
2030 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002031int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002032__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002034 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002035 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002036 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002037 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040#if FORCED_DEBUG
2041 /*
2042 * Enable redzoning and last user accounting, except for caches with
2043 * large objects, if the increased size would increase the object size
2044 * above the next power of two: caches with object sizes just above a
2045 * power of two have a significant amount of internal fragmentation.
2046 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002047 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2048 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002049 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 if (!(flags & SLAB_DESTROY_BY_RCU))
2051 flags |= SLAB_POISON;
2052#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054
Andrew Mortona737b3e2006-03-22 00:08:11 -08002055 /*
2056 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 * unaligned accesses for some archs when redzoning is used, and makes
2058 * sure any on-slab bufctl's are also correctly aligned.
2059 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002060 if (size & (BYTES_PER_WORD - 1)) {
2061 size += (BYTES_PER_WORD - 1);
2062 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 }
2064
David Woodhouse87a927c2007-07-04 21:26:44 -04002065 if (flags & SLAB_RED_ZONE) {
2066 ralign = REDZONE_ALIGN;
2067 /* If redzoning, ensure that the second redzone is suitably
2068 * aligned, by adjusting the object size accordingly. */
2069 size += REDZONE_ALIGN - 1;
2070 size &= ~(REDZONE_ALIGN - 1);
2071 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002072
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002073 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002074 if (ralign < cachep->align) {
2075 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002077 /* disable debug if necessary */
2078 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002079 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002080 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002081 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002083 cachep->align = ralign;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002084 cachep->colour_off = cache_line_size();
2085 /* Offset must be a multiple of the alignment. */
2086 if (cachep->colour_off < cachep->align)
2087 cachep->colour_off = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
Pekka Enberg83b519e2009-06-10 19:40:04 +03002089 if (slab_is_available())
2090 gfp = GFP_KERNEL;
2091 else
2092 gfp = GFP_NOWAIT;
2093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095
Pekka Enbergca5f9702006-09-25 23:31:25 -07002096 /*
2097 * Both debugging options require word-alignment which is calculated
2098 * into align above.
2099 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002102 cachep->obj_offset += sizeof(unsigned long long);
2103 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 }
2105 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002106 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002107 * the real object. But if the second red zone needs to be
2108 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002110 if (flags & SLAB_RED_ZONE)
2111 size += REDZONE_ALIGN;
2112 else
2113 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 }
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002115#endif
2116
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002117 kasan_cache_create(cachep, &size, &flags);
2118
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002119 size = ALIGN(size, cachep->align);
2120 /*
2121 * We should restrict the number of objects in a slab to implement
2122 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2123 */
2124 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2125 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2126
2127#if DEBUG
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002128 /*
2129 * To activate debug pagealloc, off-slab management is necessary
2130 * requirement. In early phase of initialization, small sized slab
2131 * doesn't get initialized so it would not be possible. So, we need
2132 * to check size >= 256. It guarantees that all necessary small
2133 * sized slab is initialized in current slab initialization sequence.
2134 */
Joonsoo Kim40323272016-03-15 14:54:18 -07002135 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
Joonsoo Kimf3a3c322016-03-15 14:54:38 -07002136 size >= 256 && cachep->object_size > cache_line_size()) {
2137 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2138 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2139
2140 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2141 flags |= CFLGS_OFF_SLAB;
2142 cachep->obj_offset += tmp_size - size;
2143 size = tmp_size;
2144 goto done;
2145 }
2146 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 }
2148#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002150 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2151 flags |= CFLGS_OBJFREELIST_SLAB;
2152 goto done;
2153 }
2154
Joonsoo Kim158e3192016-03-15 14:54:35 -07002155 if (set_off_slab_cache(cachep, size, flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 flags |= CFLGS_OFF_SLAB;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002157 goto done;
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002158 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
Joonsoo Kim158e3192016-03-15 14:54:35 -07002160 if (set_on_slab_cache(cachep, size, flags))
2161 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
Joonsoo Kim158e3192016-03-15 14:54:35 -07002163 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002164
Joonsoo Kim158e3192016-03-15 14:54:35 -07002165done:
2166 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002168 cachep->allocflags = __GFP_COMP;
Yang Shia3187e42016-05-19 17:10:41 -07002169 if (flags & SLAB_CACHE_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002170 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002171 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002172 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173
Joonsoo Kim40b44132016-03-15 14:54:21 -07002174#if DEBUG
2175 /*
2176 * If we're going to use the generic kernel_map_pages()
2177 * poisoning, then it's going to smash the contents of
2178 * the redzone and userword anyhow, so switch them off.
2179 */
2180 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2181 (cachep->flags & SLAB_POISON) &&
2182 is_debug_pagealloc_cache(cachep))
2183 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2184#endif
2185
2186 if (OFF_SLAB(cachep)) {
Joonsoo Kim158e3192016-03-15 14:54:35 -07002187 cachep->freelist_cache =
2188 kmalloc_slab(cachep->freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002191 err = setup_cpu_cache(cachep, gfp);
2192 if (err) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002193 __kmem_cache_release(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002194 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002195 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002197 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199
2200#if DEBUG
2201static void check_irq_off(void)
2202{
2203 BUG_ON(!irqs_disabled());
2204}
2205
2206static void check_irq_on(void)
2207{
2208 BUG_ON(irqs_disabled());
2209}
2210
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002211static void check_mutex_acquired(void)
2212{
2213 BUG_ON(!mutex_is_locked(&slab_mutex));
2214}
2215
Pekka Enberg343e0d72006-02-01 03:05:50 -08002216static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217{
2218#ifdef CONFIG_SMP
2219 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002220 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221#endif
2222}
Christoph Lametere498be72005-09-09 13:03:32 -07002223
Pekka Enberg343e0d72006-02-01 03:05:50 -08002224static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002225{
2226#ifdef CONFIG_SMP
2227 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002228 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002229#endif
2230}
2231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232#else
2233#define check_irq_off() do { } while(0)
2234#define check_irq_on() do { } while(0)
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002235#define check_mutex_acquired() do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002237#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238#endif
2239
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002240static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2241 int node, bool free_all, struct list_head *list)
2242{
2243 int tofree;
2244
2245 if (!ac || !ac->avail)
2246 return;
2247
2248 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2249 if (tofree > ac->avail)
2250 tofree = (ac->avail + 1) / 2;
2251
2252 free_block(cachep, ac->entry, tofree, node, list);
2253 ac->avail -= tofree;
2254 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2255}
Christoph Lameteraab22072006-03-22 00:09:06 -08002256
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257static void do_drain(void *arg)
2258{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002259 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002261 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002262 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002263 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264
2265 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002266 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002267 n = get_node(cachep, node);
2268 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002269 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002270 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002271 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 ac->avail = 0;
2273}
2274
Pekka Enberg343e0d72006-02-01 03:05:50 -08002275static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002277 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002278 int node;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002279 LIST_HEAD(list);
Christoph Lametere498be72005-09-09 13:03:32 -07002280
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002281 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002283 for_each_kmem_cache_node(cachep, node, n)
2284 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002285 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002286
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002287 for_each_kmem_cache_node(cachep, node, n) {
2288 spin_lock_irq(&n->list_lock);
2289 drain_array_locked(cachep, n->shared, node, true, &list);
2290 spin_unlock_irq(&n->list_lock);
2291
2292 slabs_destroy(cachep, &list);
2293 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294}
2295
Christoph Lametered11d9e2006-06-30 01:55:45 -07002296/*
2297 * Remove slabs from the list of free slabs.
2298 * Specify the number of slabs to drain in tofree.
2299 *
2300 * Returns the actual number of slabs released.
2301 */
2302static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002303 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002305 struct list_head *p;
2306 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002307 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308
Christoph Lametered11d9e2006-06-30 01:55:45 -07002309 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002310 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002312 spin_lock_irq(&n->list_lock);
2313 p = n->slabs_free.prev;
2314 if (p == &n->slabs_free) {
2315 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002316 goto out;
2317 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
Joonsoo Kim8456a642013-10-24 10:07:49 +09002319 page = list_entry(p, struct page, lru);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002320 list_del(&page->lru);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07002321 n->num_slabs--;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002322 /*
2323 * Safe to drop the lock. The slab is no longer linked
2324 * to the cache.
2325 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002326 n->free_objects -= cache->num;
2327 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002328 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002329 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002331out:
2332 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333}
2334
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002335int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002336{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002337 int ret = 0;
2338 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002339 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002340
2341 drain_cpu_caches(cachep);
2342
2343 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002344 for_each_kmem_cache_node(cachep, node, n) {
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07002345 drain_freelist(cachep, n, INT_MAX);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002346
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002347 ret += !list_empty(&n->slabs_full) ||
2348 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002349 }
2350 return (ret ? 1 : 0);
2351}
2352
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002353int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354{
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002355 return __kmem_cache_shrink(cachep, false);
2356}
2357
2358void __kmem_cache_release(struct kmem_cache *cachep)
2359{
Christoph Lameter12c36672012-09-04 23:38:33 +00002360 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002361 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002362
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002363 cache_random_seq_destroy(cachep);
2364
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002365 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002366
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002367 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002368 for_each_kmem_cache_node(cachep, i, n) {
2369 kfree(n->shared);
2370 free_alien_cache(n->alien);
2371 kfree(n);
2372 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002376/*
2377 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002378 *
2379 * For a slab cache when the slab descriptor is off-slab, the
2380 * slab descriptor can't come from the same cache which is being created,
2381 * Because if it is the case, that means we defer the creation of
2382 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2383 * And we eventually call down to __kmem_cache_create(), which
2384 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2385 * This is a "chicken-and-egg" problem.
2386 *
2387 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2388 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002389 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002390static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002391 struct page *page, int colour_off,
2392 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002394 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002395 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002397 page->s_mem = addr + colour_off;
2398 page->active = 0;
2399
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002400 if (OBJFREELIST_SLAB(cachep))
2401 freelist = NULL;
2402 else if (OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002404 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002405 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002406 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 return NULL;
2408 } else {
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002409 /* We will use last bytes at the slab for freelist */
2410 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2411 cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 }
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002413
Joonsoo Kim8456a642013-10-24 10:07:49 +09002414 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415}
2416
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002417static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002419 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002420}
2421
2422static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002423 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002424{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002425 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426}
2427
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002428static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429{
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002430#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 int i;
2432
2433 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002434 void *objp = index_to_obj(cachep, page, i);
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002435
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 if (cachep->flags & SLAB_STORE_USER)
2437 *dbg_userword(cachep, objp) = NULL;
2438
2439 if (cachep->flags & SLAB_RED_ZONE) {
2440 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2441 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2442 }
2443 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002444 * Constructors are not allowed to allocate memory from the same
2445 * cache which they are a constructor for. Otherwise, deadlock.
2446 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002448 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2449 kasan_unpoison_object_data(cachep,
2450 objp + obj_offset(cachep));
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002451 cachep->ctor(objp + obj_offset(cachep));
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002452 kasan_poison_object_data(
2453 cachep, objp + obj_offset(cachep));
2454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
2456 if (cachep->flags & SLAB_RED_ZONE) {
2457 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002458 slab_error(cachep, "constructor overwrote the end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002460 slab_error(cachep, "constructor overwrote the start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 }
Joonsoo Kim40b44132016-03-15 14:54:21 -07002462 /* need to poison the objs? */
2463 if (cachep->flags & SLAB_POISON) {
2464 poison_obj(cachep, objp, POISON_FREE);
2465 slab_kernel_map(cachep, objp, 0, 0);
2466 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002467 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468#endif
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002469}
2470
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002471#ifdef CONFIG_SLAB_FREELIST_RANDOM
2472/* Hold information during a freelist initialization */
2473union freelist_init_state {
2474 struct {
2475 unsigned int pos;
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002476 unsigned int *list;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002477 unsigned int count;
2478 unsigned int rand;
2479 };
2480 struct rnd_state rnd_state;
2481};
2482
2483/*
2484 * Initialize the state based on the randomization methode available.
2485 * return true if the pre-computed list is available, false otherwize.
2486 */
2487static bool freelist_state_initialize(union freelist_init_state *state,
2488 struct kmem_cache *cachep,
2489 unsigned int count)
2490{
2491 bool ret;
2492 unsigned int rand;
2493
2494 /* Use best entropy available to define a random shift */
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002495 rand = get_random_int();
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002496
2497 /* Use a random state if the pre-computed list is not available */
2498 if (!cachep->random_seq) {
2499 prandom_seed_state(&state->rnd_state, rand);
2500 ret = false;
2501 } else {
2502 state->list = cachep->random_seq;
2503 state->count = count;
2504 state->pos = 0;
2505 state->rand = rand;
2506 ret = true;
2507 }
2508 return ret;
2509}
2510
2511/* Get the next entry on the list and randomize it using a random shift */
2512static freelist_idx_t next_random_slot(union freelist_init_state *state)
2513{
2514 return (state->list[state->pos++] + state->rand) % state->count;
2515}
2516
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002517/* Swap two freelist entries */
2518static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2519{
2520 swap(((freelist_idx_t *)page->freelist)[a],
2521 ((freelist_idx_t *)page->freelist)[b]);
2522}
2523
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002524/*
2525 * Shuffle the freelist initialization state based on pre-computed lists.
2526 * return true if the list was successfully shuffled, false otherwise.
2527 */
2528static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2529{
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002530 unsigned int objfreelist = 0, i, rand, count = cachep->num;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002531 union freelist_init_state state;
2532 bool precomputed;
2533
2534 if (count < 2)
2535 return false;
2536
2537 precomputed = freelist_state_initialize(&state, cachep, count);
2538
2539 /* Take a random entry as the objfreelist */
2540 if (OBJFREELIST_SLAB(cachep)) {
2541 if (!precomputed)
2542 objfreelist = count - 1;
2543 else
2544 objfreelist = next_random_slot(&state);
2545 page->freelist = index_to_obj(cachep, page, objfreelist) +
2546 obj_offset(cachep);
2547 count--;
2548 }
2549
2550 /*
2551 * On early boot, generate the list dynamically.
2552 * Later use a pre-computed list for speed.
2553 */
2554 if (!precomputed) {
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002555 for (i = 0; i < count; i++)
2556 set_free_obj(page, i, i);
2557
2558 /* Fisher-Yates shuffle */
2559 for (i = count - 1; i > 0; i--) {
2560 rand = prandom_u32_state(&state.rnd_state);
2561 rand %= (i + 1);
2562 swap_free_obj(page, i, rand);
2563 }
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002564 } else {
2565 for (i = 0; i < count; i++)
2566 set_free_obj(page, i, next_random_slot(&state));
2567 }
2568
2569 if (OBJFREELIST_SLAB(cachep))
2570 set_free_obj(page, cachep->num - 1, objfreelist);
2571
2572 return true;
2573}
2574#else
2575static inline bool shuffle_freelist(struct kmem_cache *cachep,
2576 struct page *page)
2577{
2578 return false;
2579}
2580#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2581
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002582static void cache_init_objs(struct kmem_cache *cachep,
2583 struct page *page)
2584{
2585 int i;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002586 void *objp;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002587 bool shuffled;
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002588
2589 cache_init_objs_debug(cachep, page);
2590
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002591 /* Try to randomize the freelist if enabled */
2592 shuffled = shuffle_freelist(cachep, page);
2593
2594 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002595 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2596 obj_offset(cachep);
2597 }
2598
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002599 for (i = 0; i < cachep->num; i++) {
Andrey Ryabininb3cbd9b2016-08-02 14:02:52 -07002600 objp = index_to_obj(cachep, page, i);
2601 kasan_init_slab_obj(cachep, objp);
2602
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002603 /* constructor could break poison info */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002604 if (DEBUG == 0 && cachep->ctor) {
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002605 kasan_unpoison_object_data(cachep, objp);
2606 cachep->ctor(objp);
2607 kasan_poison_object_data(cachep, objp);
2608 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002609
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002610 if (!shuffled)
2611 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613}
2614
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002615static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002616{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002617 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002618
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002619 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002620 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002621
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002622#if DEBUG
2623 if (cachep->flags & SLAB_STORE_USER)
2624 set_store_user_dirty(cachep);
2625#endif
2626
Matthew Dobson78d382d2006-02-01 03:05:47 -08002627 return objp;
2628}
2629
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002630static void slab_put_obj(struct kmem_cache *cachep,
2631 struct page *page, void *objp)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002632{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002633 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002634#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002635 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002636
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002637 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002638 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002639 if (get_free_obj(page, i) == objnr) {
Joe Perches11705322016-03-17 14:19:50 -07002640 pr_err("slab: double free detected in cache '%s', objp %p\n",
Joe Perches756a025f02016-03-17 14:19:47 -07002641 cachep->name, objp);
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002642 BUG();
2643 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002644 }
2645#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002646 page->active--;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002647 if (!page->freelist)
2648 page->freelist = objp + obj_offset(cachep);
2649
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002650 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002651}
2652
Pekka Enberg47768742006-06-23 02:03:07 -07002653/*
2654 * Map pages beginning at addr to the given cache and slab. This is required
2655 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002656 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002657 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002658static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002659 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002661 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002662 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663}
2664
2665/*
2666 * Grow (by 1) the number of slabs within a cache. This is called by
2667 * kmem_cache_alloc() when there are no active objs left in a cache.
2668 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002669static struct page *cache_grow_begin(struct kmem_cache *cachep,
2670 gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002672 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 size_t offset;
2674 gfp_t local_flags;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002675 int page_node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002676 struct kmem_cache_node *n;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002677 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678
Andrew Mortona737b3e2006-03-22 00:08:11 -08002679 /*
2680 * Be lazy and only check for valid flags here, keeping it out of the
2681 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002683 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
Michal Hockobacdcb32016-07-26 15:22:02 -07002684 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
Michal Hocko72baeef0c2016-07-26 15:22:05 -07002685 flags &= ~GFP_SLAB_BUG_MASK;
2686 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2687 invalid_mask, &invalid_mask, flags, &flags);
2688 dump_stack();
Andrew Mortonc871ac42014-12-10 15:42:25 -08002689 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002690 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 check_irq_off();
Mel Gormand0164ad2015-11-06 16:28:21 -08002693 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 local_irq_enable();
2695
2696 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002697 * Get mem for the objs. Attempt to allocate a physical page from
2698 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002699 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002700 page = kmem_getpages(cachep, local_flags, nodeid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002701 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 goto failed;
2703
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002704 page_node = page_to_nid(page);
2705 n = get_node(cachep, page_node);
Joonsoo Kim03d1d432016-05-19 17:10:20 -07002706
2707 /* Get colour for the slab, and cal the next value. */
2708 n->colour_next++;
2709 if (n->colour_next >= cachep->colour)
2710 n->colour_next = 0;
2711
2712 offset = n->colour_next;
2713 if (offset >= cachep->colour)
2714 offset = 0;
2715
2716 offset *= cachep->colour_off;
2717
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002719 freelist = alloc_slabmgmt(cachep, page, offset,
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002720 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002721 if (OFF_SLAB(cachep) && !freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 goto opps1;
2723
Joonsoo Kim8456a642013-10-24 10:07:49 +09002724 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002726 kasan_poison_slab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002727 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728
Mel Gormand0164ad2015-11-06 16:28:21 -08002729 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 local_irq_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002732 return page;
2733
Andrew Mortona737b3e2006-03-22 00:08:11 -08002734opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002735 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002736failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002737 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 local_irq_disable();
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002739 return NULL;
2740}
2741
2742static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2743{
2744 struct kmem_cache_node *n;
2745 void *list = NULL;
2746
2747 check_irq_off();
2748
2749 if (!page)
2750 return;
2751
2752 INIT_LIST_HEAD(&page->lru);
2753 n = get_node(cachep, page_to_nid(page));
2754
2755 spin_lock(&n->list_lock);
2756 if (!page->active)
2757 list_add_tail(&page->lru, &(n->slabs_free));
2758 else
2759 fixup_slab_list(cachep, n, page, &list);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07002760
2761 n->num_slabs++;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002762 STATS_INC_GROWN(cachep);
2763 n->free_objects += cachep->num - page->active;
2764 spin_unlock(&n->list_lock);
2765
2766 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767}
2768
2769#if DEBUG
2770
2771/*
2772 * Perform extra freeing checks:
2773 * - detect bad pointers.
2774 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 */
2776static void kfree_debugcheck(const void *objp)
2777{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 if (!virt_addr_valid(objp)) {
Joe Perches11705322016-03-17 14:19:50 -07002779 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002780 (unsigned long)objp);
2781 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783}
2784
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002785static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2786{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002787 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002788
2789 redzone1 = *dbg_redzone1(cache, obj);
2790 redzone2 = *dbg_redzone2(cache, obj);
2791
2792 /*
2793 * Redzone is ok.
2794 */
2795 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2796 return;
2797
2798 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2799 slab_error(cache, "double free detected");
2800 else
2801 slab_error(cache, "memory outside object was overwritten");
2802
Joe Perches11705322016-03-17 14:19:50 -07002803 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2804 obj, redzone1, redzone2);
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002805}
2806
Pekka Enberg343e0d72006-02-01 03:05:50 -08002807static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002808 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002811 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002813 BUG_ON(virt_to_cache(objp) != cachep);
2814
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002815 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002817 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002820 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2822 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2823 }
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002824 if (cachep->flags & SLAB_STORE_USER) {
2825 set_store_user_dirty(cachep);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002826 *dbg_userword(cachep, objp) = (void *)caller;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002827 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
Joonsoo Kim8456a642013-10-24 10:07:49 +09002829 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
2831 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002832 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 poison_obj(cachep, objp, POISON_FREE);
Joonsoo Kim40b44132016-03-15 14:54:21 -07002836 slab_kernel_map(cachep, objp, 0, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 }
2838 return objp;
2839}
2840
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841#else
2842#define kfree_debugcheck(x) do { } while(0)
2843#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844#endif
2845
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002846static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2847 void **list)
2848{
2849#if DEBUG
2850 void *next = *list;
2851 void *objp;
2852
2853 while (next) {
2854 objp = next - obj_offset(cachep);
2855 next = *(void **)next;
2856 poison_obj(cachep, objp, POISON_FREE);
2857 }
2858#endif
2859}
2860
Joonsoo Kimd8410232016-03-15 14:54:44 -07002861static inline void fixup_slab_list(struct kmem_cache *cachep,
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002862 struct kmem_cache_node *n, struct page *page,
2863 void **list)
Joonsoo Kimd8410232016-03-15 14:54:44 -07002864{
2865 /* move slabp to correct slabp list: */
2866 list_del(&page->lru);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002867 if (page->active == cachep->num) {
Joonsoo Kimd8410232016-03-15 14:54:44 -07002868 list_add(&page->lru, &n->slabs_full);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002869 if (OBJFREELIST_SLAB(cachep)) {
2870#if DEBUG
2871 /* Poisoning will be done without holding the lock */
2872 if (cachep->flags & SLAB_POISON) {
2873 void **objp = page->freelist;
2874
2875 *objp = *list;
2876 *list = objp;
2877 }
2878#endif
2879 page->freelist = NULL;
2880 }
2881 } else
Joonsoo Kimd8410232016-03-15 14:54:44 -07002882 list_add(&page->lru, &n->slabs_partial);
2883}
2884
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002885/* Try to find non-pfmemalloc slab if needed */
2886static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2887 struct page *page, bool pfmemalloc)
2888{
2889 if (!page)
2890 return NULL;
2891
2892 if (pfmemalloc)
2893 return page;
2894
2895 if (!PageSlabPfmemalloc(page))
2896 return page;
2897
2898 /* No need to keep pfmemalloc slab if we have enough free objects */
2899 if (n->free_objects > n->free_limit) {
2900 ClearPageSlabPfmemalloc(page);
2901 return page;
2902 }
2903
2904 /* Move pfmemalloc slab to the end of list to speed up next search */
2905 list_del(&page->lru);
2906 if (!page->active)
2907 list_add_tail(&page->lru, &n->slabs_free);
2908 else
2909 list_add_tail(&page->lru, &n->slabs_partial);
2910
2911 list_for_each_entry(page, &n->slabs_partial, lru) {
2912 if (!PageSlabPfmemalloc(page))
2913 return page;
2914 }
2915
2916 list_for_each_entry(page, &n->slabs_free, lru) {
2917 if (!PageSlabPfmemalloc(page))
2918 return page;
2919 }
2920
2921 return NULL;
2922}
2923
2924static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
Geliang Tang7aa0d222016-01-14 15:18:02 -08002925{
2926 struct page *page;
2927
2928 page = list_first_entry_or_null(&n->slabs_partial,
2929 struct page, lru);
2930 if (!page) {
2931 n->free_touched = 1;
2932 page = list_first_entry_or_null(&n->slabs_free,
2933 struct page, lru);
2934 }
2935
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002936 if (sk_memalloc_socks())
2937 return get_valid_first_slab(n, page, pfmemalloc);
2938
Geliang Tang7aa0d222016-01-14 15:18:02 -08002939 return page;
2940}
2941
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002942static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2943 struct kmem_cache_node *n, gfp_t flags)
2944{
2945 struct page *page;
2946 void *obj;
2947 void *list = NULL;
2948
2949 if (!gfp_pfmemalloc_allowed(flags))
2950 return NULL;
2951
2952 spin_lock(&n->list_lock);
2953 page = get_first_slab(n, true);
2954 if (!page) {
2955 spin_unlock(&n->list_lock);
2956 return NULL;
2957 }
2958
2959 obj = slab_get_obj(cachep, page);
2960 n->free_objects--;
2961
2962 fixup_slab_list(cachep, n, page, &list);
2963
2964 spin_unlock(&n->list_lock);
2965 fixup_objfreelist_debug(cachep, &list);
2966
2967 return obj;
2968}
2969
Joonsoo Kim213b4692016-05-19 17:10:29 -07002970/*
2971 * Slab list should be fixed up by fixup_slab_list() for existing slab
2972 * or cache_grow_end() for new slab
2973 */
2974static __always_inline int alloc_block(struct kmem_cache *cachep,
2975 struct array_cache *ac, struct page *page, int batchcount)
2976{
2977 /*
2978 * There must be at least one object available for
2979 * allocation.
2980 */
2981 BUG_ON(page->active >= cachep->num);
2982
2983 while (page->active < cachep->num && batchcount--) {
2984 STATS_INC_ALLOCED(cachep);
2985 STATS_INC_ACTIVE(cachep);
2986 STATS_SET_HIGH(cachep);
2987
2988 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2989 }
2990
2991 return batchcount;
2992}
2993
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002994static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995{
2996 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002997 struct kmem_cache_node *n;
Joonsoo Kim801faf02016-05-19 17:10:31 -07002998 struct array_cache *ac, *shared;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002999 int node;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003000 void *list = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003001 struct page *page;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003002
Joe Korty6d2144d2008-03-05 15:04:59 -08003003 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003004 node = numa_mem_id();
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003005
Joe Korty6d2144d2008-03-05 15:04:59 -08003006 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 batchcount = ac->batchcount;
3008 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003009 /*
3010 * If there was little recent activity on this cache, then
3011 * perform only a partial refill. Otherwise we could generate
3012 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 */
3014 batchcount = BATCHREFILL_LIMIT;
3015 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07003016 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003018 BUG_ON(ac->avail > 0 || !n);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003019 shared = READ_ONCE(n->shared);
3020 if (!n->free_objects && (!shared || !shared->avail))
3021 goto direct_grow;
3022
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003023 spin_lock(&n->list_lock);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003024 shared = READ_ONCE(n->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003025
Christoph Lameter3ded1752006-03-25 03:06:44 -08003026 /* See if we can refill from the shared array */
Joonsoo Kim801faf02016-05-19 17:10:31 -07003027 if (shared && transfer_objects(ac, shared, batchcount)) {
3028 shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003029 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003030 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003031
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032 while (batchcount > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 /* Get slab alloc is to come from. */
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003034 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003035 if (!page)
3036 goto must_grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003039
Joonsoo Kim213b4692016-05-19 17:10:29 -07003040 batchcount = alloc_block(cachep, ac, page, batchcount);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003041 fixup_slab_list(cachep, n, page, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042 }
3043
Andrew Mortona737b3e2006-03-22 00:08:11 -08003044must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003045 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003046alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003047 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003048 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049
Joonsoo Kim801faf02016-05-19 17:10:31 -07003050direct_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 if (unlikely(!ac->avail)) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003052 /* Check if we can use obj in pfmemalloc slab */
3053 if (sk_memalloc_socks()) {
3054 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3055
3056 if (obj)
3057 return obj;
3058 }
3059
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003060 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
Christoph Lametere498be72005-09-09 13:03:32 -07003061
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003062 /*
3063 * cache_grow_begin() can reenable interrupts,
3064 * then ac could change.
3065 */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003066 ac = cpu_cache_get(cachep);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003067 if (!ac->avail && page)
3068 alloc_block(cachep, ac, page, batchcount);
3069 cache_grow_end(cachep, page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003070
Joonsoo Kim213b4692016-05-19 17:10:29 -07003071 if (!ac->avail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 }
3074 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003075
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003076 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077}
3078
Andrew Mortona737b3e2006-03-22 00:08:11 -08003079static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3080 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081{
Mel Gormand0164ad2015-11-06 16:28:21 -08003082 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083}
3084
3085#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003087 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003089 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07003093 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 poison_obj(cachep, objp, POISON_INUSE);
3095 }
3096 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003097 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098
3099 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003100 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3101 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
Joe Perches756a025f02016-03-17 14:19:47 -07003102 slab_error(cachep, "double free, or memory outside object was overwritten");
Joe Perches11705322016-03-17 14:19:50 -07003103 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3104 objp, *dbg_redzone1(cachep, objp),
3105 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 }
3107 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3108 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3109 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003110
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003111 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003112 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003113 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003114 if (ARCH_SLAB_MINALIGN &&
3115 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Joe Perches11705322016-03-17 14:19:50 -07003116 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003117 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003118 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 return objp;
3120}
3121#else
3122#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3123#endif
3124
Pekka Enberg343e0d72006-02-01 03:05:50 -08003125static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003127 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 struct array_cache *ac;
3129
Alok N Kataria5c382302005-09-27 21:45:46 -07003130 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003131
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003132 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 ac->touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003135 objp = ac->entry[--ac->avail];
Mel Gorman072bb0a2012-07-31 16:43:58 -07003136
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003137 STATS_INC_ALLOCHIT(cachep);
3138 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003140
3141 STATS_INC_ALLOCMISS(cachep);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003142 objp = cache_alloc_refill(cachep, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003143 /*
3144 * the 'ac' may be updated by cache_alloc_refill(),
3145 * and kmemleak_erase() requires its correct value.
3146 */
3147 ac = cpu_cache_get(cachep);
3148
3149out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003150 /*
3151 * To avoid a false negative, if an object that is in one of the
3152 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3153 * treat the array pointers as a reference to the object.
3154 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003155 if (objp)
3156 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003157 return objp;
3158}
3159
Christoph Lametere498be72005-09-09 13:03:32 -07003160#ifdef CONFIG_NUMA
3161/*
Zefan Li2ad654b2014-09-25 09:41:02 +08003162 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003163 *
3164 * If we are in_interrupt, then process context, including cpusets and
3165 * mempolicy, may not apply and should not be used for allocation policy.
3166 */
3167static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3168{
3169 int nid_alloc, nid_here;
3170
Christoph Lameter765c4502006-09-27 01:50:08 -07003171 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003172 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003173 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003174 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003175 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003176 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003177 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003178 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003179 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003180 return NULL;
3181}
3182
3183/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003184 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003186 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003187 * perform an allocation without specifying a node. This allows the page
3188 * allocator to do its reclaim / fallback magic. We then insert the
3189 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003190 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003191static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003192{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003193 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07003194 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003195 struct zone *zone;
3196 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003197 void *obj = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003198 struct page *page;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003199 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003200 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003201
3202 if (flags & __GFP_THISNODE)
3203 return NULL;
3204
Mel Gormancc9a6c82012-03-21 16:34:11 -07003205retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003206 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003207 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003208
Christoph Lameter3c517a62006-12-06 20:33:29 -08003209retry:
3210 /*
3211 * Look through allowed nodes for objects available
3212 * from existing per node queues.
3213 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003214 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3215 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003216
Vladimir Davydov061d7072014-12-12 16:58:25 -08003217 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003218 get_node(cache, nid) &&
3219 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003220 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003221 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003222 if (obj)
3223 break;
3224 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003225 }
3226
Christoph Lametercfce6602007-05-06 14:50:17 -07003227 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003228 /*
3229 * This allocation will be performed within the constraints
3230 * of the current cpuset / memory policy requirements.
3231 * We may trigger various forms of reclaim on the allowed
3232 * set and go into memory reserves if necessary.
3233 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003234 page = cache_grow_begin(cache, flags, numa_mem_id());
3235 cache_grow_end(cache, page);
3236 if (page) {
3237 nid = page_to_nid(page);
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003238 obj = ____cache_alloc_node(cache,
3239 gfp_exact_node(flags), nid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003240
Christoph Lameter3c517a62006-12-06 20:33:29 -08003241 /*
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003242 * Another processor may allocate the objects in
3243 * the slab since we are not holding any locks.
Christoph Lameter3c517a62006-12-06 20:33:29 -08003244 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003245 if (!obj)
3246 goto retry;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003247 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003248 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003249
Mel Gormand26914d2014-04-03 14:47:24 -07003250 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003251 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003252 return obj;
3253}
3254
3255/*
Christoph Lametere498be72005-09-09 13:03:32 -07003256 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003258static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003259 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003260{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003261 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003262 struct kmem_cache_node *n;
Joonsoo Kim213b4692016-05-19 17:10:29 -07003263 void *obj = NULL;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003264 void *list = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003266 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003267 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003268 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003269
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003270 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003271 spin_lock(&n->list_lock);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003272 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003273 if (!page)
3274 goto must_grow;
Christoph Lametere498be72005-09-09 13:03:32 -07003275
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003276 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003277
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003278 STATS_INC_NODEALLOCS(cachep);
3279 STATS_INC_ACTIVE(cachep);
3280 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Joonsoo Kim8456a642013-10-24 10:07:49 +09003282 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003283
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003284 obj = slab_get_obj(cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003285 n->free_objects--;
Christoph Lametere498be72005-09-09 13:03:32 -07003286
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003287 fixup_slab_list(cachep, n, page, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003288
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003289 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003290 fixup_objfreelist_debug(cachep, &list);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003291 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003292
Andrew Mortona737b3e2006-03-22 00:08:11 -08003293must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003294 spin_unlock(&n->list_lock);
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003295 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003296 if (page) {
3297 /* This slab isn't counted yet so don't update free_objects */
3298 obj = slab_get_obj(cachep, page);
3299 }
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003300 cache_grow_end(cachep, page);
Christoph Lametere498be72005-09-09 13:03:32 -07003301
Joonsoo Kim213b4692016-05-19 17:10:29 -07003302 return obj ? obj : fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003303}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003304
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003305static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003306slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003307 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003308{
3309 unsigned long save_flags;
3310 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003311 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003312
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003313 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003314 cachep = slab_pre_alloc_hook(cachep, flags);
3315 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003316 return NULL;
3317
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003318 cache_alloc_debugcheck_before(cachep, flags);
3319 local_irq_save(save_flags);
3320
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003321 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003322 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003323
Christoph Lameter18bf8542014-08-06 16:04:11 -07003324 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003325 /* Node not bootstrapped yet */
3326 ptr = fallback_alloc(cachep, flags);
3327 goto out;
3328 }
3329
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003330 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003331 /*
3332 * Use the locally cached objects if possible.
3333 * However ____cache_alloc does not allow fallback
3334 * to other nodes. It may fail while we still have
3335 * objects on other nodes available.
3336 */
3337 ptr = ____cache_alloc(cachep, flags);
3338 if (ptr)
3339 goto out;
3340 }
3341 /* ___cache_alloc_node can fall back to other nodes */
3342 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3343 out:
3344 local_irq_restore(save_flags);
3345 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3346
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003347 if (unlikely(flags & __GFP_ZERO) && ptr)
3348 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003349
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003350 slab_post_alloc_hook(cachep, flags, 1, &ptr);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003351 return ptr;
3352}
3353
3354static __always_inline void *
3355__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3356{
3357 void *objp;
3358
Zefan Li2ad654b2014-09-25 09:41:02 +08003359 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003360 objp = alternate_node_alloc(cache, flags);
3361 if (objp)
3362 goto out;
3363 }
3364 objp = ____cache_alloc(cache, flags);
3365
3366 /*
3367 * We may just have run out of memory on the local node.
3368 * ____cache_alloc_node() knows how to locate memory on other nodes
3369 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003370 if (!objp)
3371 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003372
3373 out:
3374 return objp;
3375}
3376#else
3377
3378static __always_inline void *
3379__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3380{
3381 return ____cache_alloc(cachep, flags);
3382}
3383
3384#endif /* CONFIG_NUMA */
3385
3386static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003387slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003388{
3389 unsigned long save_flags;
3390 void *objp;
3391
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003392 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003393 cachep = slab_pre_alloc_hook(cachep, flags);
3394 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003395 return NULL;
3396
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003397 cache_alloc_debugcheck_before(cachep, flags);
3398 local_irq_save(save_flags);
3399 objp = __do_cache_alloc(cachep, flags);
3400 local_irq_restore(save_flags);
3401 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3402 prefetchw(objp);
3403
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003404 if (unlikely(flags & __GFP_ZERO) && objp)
3405 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003406
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003407 slab_post_alloc_hook(cachep, flags, 1, &objp);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003408 return objp;
3409}
Christoph Lametere498be72005-09-09 13:03:32 -07003410
3411/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003412 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003413 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003414 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003415static void free_block(struct kmem_cache *cachep, void **objpp,
3416 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417{
3418 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003419 struct kmem_cache_node *n = get_node(cachep, node);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003420 struct page *page;
3421
3422 n->free_objects += nr_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
3424 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003425 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003426 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427
Mel Gorman072bb0a2012-07-31 16:43:58 -07003428 objp = objpp[i];
3429
Joonsoo Kim8456a642013-10-24 10:07:49 +09003430 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003431 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003432 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003433 slab_put_obj(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 STATS_DEC_ACTIVE(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435
3436 /* fixup slab chains */
Joonsoo Kim6052b782016-05-19 17:10:17 -07003437 if (page->active == 0)
3438 list_add(&page->lru, &n->slabs_free);
3439 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 /* Unconditionally move a slab to the end of the
3441 * partial list on free - maximum time for the
3442 * other objects to be freed, too.
3443 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003444 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 }
3446 }
Joonsoo Kim6052b782016-05-19 17:10:17 -07003447
3448 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3449 n->free_objects -= cachep->num;
3450
3451 page = list_last_entry(&n->slabs_free, struct page, lru);
Wei Yongjunde24bae2016-07-26 15:22:11 -07003452 list_move(&page->lru, list);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07003453 n->num_slabs--;
Joonsoo Kim6052b782016-05-19 17:10:17 -07003454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455}
3456
Pekka Enberg343e0d72006-02-01 03:05:50 -08003457static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458{
3459 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003460 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003461 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003462 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
3464 batchcount = ac->batchcount;
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003465
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003467 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003468 spin_lock(&n->list_lock);
3469 if (n->shared) {
3470 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003471 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 if (max) {
3473 if (batchcount > max)
3474 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003475 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003476 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 shared_array->avail += batchcount;
3478 goto free_done;
3479 }
3480 }
3481
Joonsoo Kim97654df2014-08-06 16:04:25 -07003482 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003483free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484#if STATS
3485 {
3486 int i = 0;
Geliang Tang73c02192016-01-14 15:17:59 -08003487 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488
Geliang Tang73c02192016-01-14 15:17:59 -08003489 list_for_each_entry(page, &n->slabs_free, lru) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003490 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491
3492 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 }
3494 STATS_SET_FREEABLE(cachep, i);
3495 }
3496#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003497 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003498 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003500 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501}
3502
3503/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003504 * Release an obj back to its cache. If the obj has a constructed state, it must
3505 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003507static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003508 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509{
Alexander Potapenko55834c52016-05-20 16:59:11 -07003510 /* Put the object into the quarantine, don't touch it for now. */
3511 if (kasan_slab_free(cachep, objp))
3512 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513
Alexander Potapenko55834c52016-05-20 16:59:11 -07003514 ___cache_free(cachep, objp, caller);
3515}
3516
3517void ___cache_free(struct kmem_cache *cachep, void *objp,
3518 unsigned long caller)
3519{
3520 struct array_cache *ac = cpu_cache_get(cachep);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003521
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003523 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003524 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003526 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003527
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003528 /*
3529 * Skip calling cache_free_alien() when the platform is not numa.
3530 * This will avoid cache misses that happen while accessing slabp (which
3531 * is per page memory reference) to get nodeid. Instead use a global
3532 * variable to skip the call, which is mostly likely to be present in
3533 * the cache.
3534 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003535 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003536 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003537
Joonsoo Kim3d880192014-10-09 15:26:04 -07003538 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 } else {
3541 STATS_INC_FREEMISS(cachep);
3542 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003544
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003545 if (sk_memalloc_socks()) {
3546 struct page *page = virt_to_head_page(objp);
3547
3548 if (unlikely(PageSlabPfmemalloc(page))) {
3549 cache_free_pfmemalloc(cachep, page, objp);
3550 return;
3551 }
3552 }
3553
3554 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555}
3556
3557/**
3558 * kmem_cache_alloc - Allocate an object
3559 * @cachep: The cache to allocate from.
3560 * @flags: See kmalloc().
3561 *
3562 * Allocate an object from this cache. The flags are only relevant
3563 * if the cache has no available objects.
3564 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003565void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003567 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003568
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003569 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003570 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003571 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003572
3573 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
3575EXPORT_SYMBOL(kmem_cache_alloc);
3576
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003577static __always_inline void
3578cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3579 size_t size, void **p, unsigned long caller)
3580{
3581 size_t i;
3582
3583 for (i = 0; i < size; i++)
3584 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3585}
3586
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003587int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003588 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07003589{
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003590 size_t i;
3591
3592 s = slab_pre_alloc_hook(s, flags);
3593 if (!s)
3594 return 0;
3595
3596 cache_alloc_debugcheck_before(s, flags);
3597
3598 local_irq_disable();
3599 for (i = 0; i < size; i++) {
3600 void *objp = __do_cache_alloc(s, flags);
3601
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003602 if (unlikely(!objp))
3603 goto error;
3604 p[i] = objp;
3605 }
3606 local_irq_enable();
3607
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003608 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3609
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003610 /* Clear memory outside IRQ disabled section */
3611 if (unlikely(flags & __GFP_ZERO))
3612 for (i = 0; i < size; i++)
3613 memset(p[i], 0, s->object_size);
3614
3615 slab_post_alloc_hook(s, flags, size, p);
3616 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3617 return size;
3618error:
3619 local_irq_enable();
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003620 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003621 slab_post_alloc_hook(s, flags, i, p);
3622 __kmem_cache_free_bulk(s, i, p);
3623 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07003624}
3625EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3626
Li Zefan0f24f122009-12-11 15:45:30 +08003627#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003628void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003629kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003630{
Steven Rostedt85beb582010-11-24 16:23:34 -05003631 void *ret;
3632
Ezequiel Garcia48356302012-09-08 17:47:57 -03003633 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003634
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003635 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003636 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003637 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003638 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003639}
Steven Rostedt85beb582010-11-24 16:23:34 -05003640EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003641#endif
3642
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003644/**
3645 * kmem_cache_alloc_node - Allocate an object on the specified node
3646 * @cachep: The cache to allocate from.
3647 * @flags: See kmalloc().
3648 * @nodeid: node number of the target node.
3649 *
3650 * Identical to kmem_cache_alloc but it will allocate memory on the given
3651 * node, which can improve the performance for cpu bound structures.
3652 *
3653 * Fallback to other node is possible if __GFP_THISNODE is not set.
3654 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003655void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3656{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003657 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003658
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003659 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003660 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003661 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003662 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003663
3664 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003665}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666EXPORT_SYMBOL(kmem_cache_alloc_node);
3667
Li Zefan0f24f122009-12-11 15:45:30 +08003668#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003669void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003670 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003671 int nodeid,
3672 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003673{
Steven Rostedt85beb582010-11-24 16:23:34 -05003674 void *ret;
3675
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003676 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003677
3678 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003679 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003680 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003681 flags, nodeid);
3682 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003683}
Steven Rostedt85beb582010-11-24 16:23:34 -05003684EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003685#endif
3686
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003687static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003688__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003689{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003690 struct kmem_cache *cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003691 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003692
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003693 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003694 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3695 return cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003696 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003697 kasan_kmalloc(cachep, ret, size, flags);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003698
3699 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003700}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003701
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003702void *__kmalloc_node(size_t size, gfp_t flags, int node)
3703{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003704 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003706EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003707
3708void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003709 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003710{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003711 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003712}
3713EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003714#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715
3716/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003717 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003719 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003720 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003722static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003723 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003725 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003726 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003728 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3730 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003731 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003732
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003733 kasan_kmalloc(cachep, ret, size, flags);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003734 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003735 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003736
3737 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003738}
3739
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003740void *__kmalloc(size_t size, gfp_t flags)
3741{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003742 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743}
3744EXPORT_SYMBOL(__kmalloc);
3745
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003746void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003747{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003748 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003749}
3750EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003751
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752/**
3753 * kmem_cache_free - Deallocate an object
3754 * @cachep: The cache the allocation was from.
3755 * @objp: The previously allocated object.
3756 *
3757 * Free an object which was previously allocated from this
3758 * cache.
3759 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003760void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761{
3762 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003763 cachep = cache_from_obj(cachep, objp);
3764 if (!cachep)
3765 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766
3767 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003768 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003769 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003770 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003771 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003773
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003774 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775}
3776EXPORT_SYMBOL(kmem_cache_free);
3777
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003778void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3779{
3780 struct kmem_cache *s;
3781 size_t i;
3782
3783 local_irq_disable();
3784 for (i = 0; i < size; i++) {
3785 void *objp = p[i];
3786
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07003787 if (!orig_s) /* called via kfree_bulk */
3788 s = virt_to_cache(objp);
3789 else
3790 s = cache_from_obj(orig_s, objp);
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003791
3792 debug_check_no_locks_freed(objp, s->object_size);
3793 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3794 debug_check_no_obj_freed(objp, s->object_size);
3795
3796 __cache_free(s, objp, _RET_IP_);
3797 }
3798 local_irq_enable();
3799
3800 /* FIXME: add tracing */
3801}
3802EXPORT_SYMBOL(kmem_cache_free_bulk);
3803
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 * kfree - free previously allocated memory
3806 * @objp: pointer returned by kmalloc.
3807 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003808 * If @objp is NULL, no operation is performed.
3809 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810 * Don't free memory not originally allocated by kmalloc()
3811 * or you will run into trouble.
3812 */
3813void kfree(const void *objp)
3814{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003815 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 unsigned long flags;
3817
Pekka Enberg2121db72009-03-25 11:05:57 +02003818 trace_kfree(_RET_IP_, objp);
3819
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003820 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 return;
3822 local_irq_save(flags);
3823 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003824 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003825 debug_check_no_locks_freed(objp, c->object_size);
3826
3827 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003828 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 local_irq_restore(flags);
3830}
3831EXPORT_SYMBOL(kfree);
3832
Christoph Lametere498be72005-09-09 13:03:32 -07003833/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003834 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003835 */
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003836static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003837{
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003838 int ret;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003840 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07003841
Mel Gorman9c09a952008-01-24 05:49:54 -08003842 for_each_online_node(node) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003843 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3844 if (ret)
Christoph Lametere498be72005-09-09 13:03:32 -07003845 goto fail;
3846
Christoph Lametere498be72005-09-09 13:03:32 -07003847 }
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003848
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003850
Andrew Mortona737b3e2006-03-22 00:08:11 -08003851fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003852 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003853 /* Cache is not active yet. Roll back what we did */
3854 node--;
3855 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003856 n = get_node(cachep, node);
3857 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003858 kfree(n->shared);
3859 free_alien_cache(n->alien);
3860 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003861 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003862 }
3863 node--;
3864 }
3865 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003866 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003867}
3868
Christoph Lameter18004c52012-07-06 15:25:12 -05003869/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003870static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003871 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003873 struct array_cache __percpu *cpu_cache, *prev;
3874 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003876 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3877 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003878 return -ENOMEM;
3879
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003880 prev = cachep->cpu_cache;
3881 cachep->cpu_cache = cpu_cache;
3882 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003883
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 cachep->batchcount = batchcount;
3886 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003887 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003889 if (!prev)
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003890 goto setup_node;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003891
3892 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003893 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003894 int node;
3895 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003896 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003897
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003898 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003899 n = get_node(cachep, node);
3900 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003901 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003902 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003903 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003905 free_percpu(prev);
3906
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003907setup_node:
3908 return setup_kmem_cache_nodes(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909}
3910
Glauber Costa943a4512012-12-18 14:23:03 -08003911static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3912 int batchcount, int shared, gfp_t gfp)
3913{
3914 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003915 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003916
3917 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3918
3919 if (slab_state < FULL)
3920 return ret;
3921
3922 if ((ret < 0) || !is_root_cache(cachep))
3923 return ret;
3924
Vladimir Davydov426589f2015-02-12 14:59:23 -08003925 lockdep_assert_held(&slab_mutex);
3926 for_each_memcg_cache(c, cachep) {
3927 /* return value determined by the root cache only */
3928 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003929 }
3930
3931 return ret;
3932}
3933
Christoph Lameter18004c52012-07-06 15:25:12 -05003934/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003935static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936{
3937 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003938 int limit = 0;
3939 int shared = 0;
3940 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941
Thomas Garnier7c00fce2016-07-26 15:21:56 -07003942 err = cache_random_seq_create(cachep, cachep->num, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003943 if (err)
3944 goto end;
3945
Glauber Costa943a4512012-12-18 14:23:03 -08003946 if (!is_root_cache(cachep)) {
3947 struct kmem_cache *root = memcg_root_cache(cachep);
3948 limit = root->limit;
3949 shared = root->shared;
3950 batchcount = root->batchcount;
3951 }
3952
3953 if (limit && shared && batchcount)
3954 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003955 /*
3956 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 * - create a LIFO ordering, i.e. return objects that are cache-warm
3958 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003959 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960 * bufctl chains: array operations are cheaper.
3961 * The numbers are guessed, we should auto-tune as described by
3962 * Bonwick.
3963 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003964 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003966 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003968 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003970 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 limit = 54;
3972 else
3973 limit = 120;
3974
Andrew Mortona737b3e2006-03-22 00:08:11 -08003975 /*
3976 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 * allocation behaviour: Most allocs on one cpu, most free operations
3978 * on another cpu. For these cases, an efficient object passing between
3979 * cpus is necessary. This is provided by a shared array. The array
3980 * replaces Bonwick's magazine layer.
3981 * On uniprocessor, it's functionally equivalent (but less efficient)
3982 * to a larger limit. Thus disabled by default.
3983 */
3984 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003985 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987
3988#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003989 /*
3990 * With debugging enabled, large batchcount lead to excessively long
3991 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 */
3993 if (limit > 32)
3994 limit = 32;
3995#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003996 batchcount = (limit + 1) / 2;
3997skip_setup:
3998 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003999end:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 if (err)
Joe Perches11705322016-03-17 14:19:50 -07004001 pr_err("enable_cpucache failed for %s, error %d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004002 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004003 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004}
4005
Christoph Lameter1b552532006-03-22 00:09:07 -08004006/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004007 * Drain an array if it contains any elements taking the node lock only if
4008 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004009 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004010 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004011static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004012 struct array_cache *ac, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013{
Joonsoo Kim97654df2014-08-06 16:04:25 -07004014 LIST_HEAD(list);
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004015
4016 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4017 check_mutex_acquired();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018
Christoph Lameter1b552532006-03-22 00:09:07 -08004019 if (!ac || !ac->avail)
4020 return;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004021
4022 if (ac->touched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 ac->touched = 0;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004024 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025 }
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004026
4027 spin_lock_irq(&n->list_lock);
4028 drain_array_locked(cachep, ac, node, false, &list);
4029 spin_unlock_irq(&n->list_lock);
4030
4031 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032}
4033
4034/**
4035 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004036 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 *
4038 * Called from workqueue/eventd every few seconds.
4039 * Purpose:
4040 * - clear the per-cpu caches for this CPU.
4041 * - return freeable pages to the main free memory pool.
4042 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004043 * If we cannot acquire the cache chain mutex then just give up - we'll try
4044 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004046static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004048 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004049 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004050 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004051 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Christoph Lameter18004c52012-07-06 15:25:12 -05004053 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004055 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056
Christoph Lameter18004c52012-07-06 15:25:12 -05004057 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058 check_irq_on();
4059
Christoph Lameter35386e32006-03-22 00:09:05 -08004060 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004061 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004062 * have established with reasonable certainty that
4063 * we can do some work if the lock was obtained.
4064 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07004065 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08004066
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004067 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004069 drain_array(searchp, n, cpu_cache_get(searchp), node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070
Christoph Lameter35386e32006-03-22 00:09:05 -08004071 /*
4072 * These are racy checks but it does not matter
4073 * if we skip one check or scan twice.
4074 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004075 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004076 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004078 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004080 drain_array(searchp, n, n->shared, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004082 if (n->free_touched)
4083 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004084 else {
4085 int freed;
4086
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004087 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004088 5 * searchp->num - 1) / (5 * searchp->num));
4089 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004091next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 cond_resched();
4093 }
4094 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004095 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004096 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004097out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004098 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004099 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100}
4101
Linus Torvalds158a9622008-01-02 13:04:48 -08004102#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004103void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004105 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004106 unsigned long active_objs;
4107 unsigned long num_objs;
4108 unsigned long active_slabs = 0;
4109 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004110 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
4111 unsigned long num_slabs_full = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004112 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004114 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004115 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117 active_objs = 0;
4118 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07004119 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07004120
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004121 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004122 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004123
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004124 num_slabs += n->num_slabs;
4125
Joonsoo Kim8456a642013-10-24 10:07:49 +09004126 list_for_each_entry(page, &n->slabs_partial, lru) {
4127 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004128 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004129 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004130 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004131 active_objs += page->active;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004132 num_slabs_partial++;
Christoph Lametere498be72005-09-09 13:03:32 -07004133 }
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004134
Joonsoo Kim8456a642013-10-24 10:07:49 +09004135 list_for_each_entry(page, &n->slabs_free, lru) {
4136 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004137 error = "slabs_free accounting error";
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004138 num_slabs_free++;
Christoph Lametere498be72005-09-09 13:03:32 -07004139 }
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004140
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004141 free_objects += n->free_objects;
4142 if (n->shared)
4143 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004144
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004145 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004147 num_objs = num_slabs * cachep->num;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004148 active_slabs = num_slabs - num_slabs_free;
4149 num_slabs_full = num_slabs - (num_slabs_partial + num_slabs_free);
4150 active_objs += (num_slabs_full * cachep->num);
4151
Christoph Lametere498be72005-09-09 13:03:32 -07004152 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 error = "free_objects accounting error";
4154
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004155 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 if (error)
Joe Perches11705322016-03-17 14:19:50 -07004157 pr_err("slab: cache %s error: %s\n", name, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158
Glauber Costa0d7561c2012-10-19 18:20:27 +04004159 sinfo->active_objs = active_objs;
4160 sinfo->num_objs = num_objs;
4161 sinfo->active_slabs = active_slabs;
4162 sinfo->num_slabs = num_slabs;
4163 sinfo->shared_avail = shared_avail;
4164 sinfo->limit = cachep->limit;
4165 sinfo->batchcount = cachep->batchcount;
4166 sinfo->shared = cachep->shared;
4167 sinfo->objects_per_slab = cachep->num;
4168 sinfo->cache_order = cachep->gfporder;
4169}
4170
4171void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4172{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004174 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175 unsigned long high = cachep->high_mark;
4176 unsigned long allocs = cachep->num_allocations;
4177 unsigned long grown = cachep->grown;
4178 unsigned long reaped = cachep->reaped;
4179 unsigned long errors = cachep->errors;
4180 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004182 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004183 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184
Joe Perches756a025f02016-03-17 14:19:47 -07004185 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
Joe Perchese92dd4f2010-03-26 19:27:58 -07004186 allocs, high, grown,
4187 reaped, errors, max_freeable, node_allocs,
4188 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189 }
4190 /* cpu stats */
4191 {
4192 unsigned long allochit = atomic_read(&cachep->allochit);
4193 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4194 unsigned long freehit = atomic_read(&cachep->freehit);
4195 unsigned long freemiss = atomic_read(&cachep->freemiss);
4196
4197 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004198 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 }
4200#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201}
4202
Linus Torvalds1da177e2005-04-16 15:20:36 -07004203#define MAX_SLABINFO_WRITE 128
4204/**
4205 * slabinfo_write - Tuning for the slab allocator
4206 * @file: unused
4207 * @buffer: user buffer
4208 * @count: data length
4209 * @ppos: unused
4210 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004211ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004212 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004214 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004216 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004217
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 if (count > MAX_SLABINFO_WRITE)
4219 return -EINVAL;
4220 if (copy_from_user(&kbuf, buffer, count))
4221 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004222 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223
4224 tmp = strchr(kbuf, ' ');
4225 if (!tmp)
4226 return -EINVAL;
4227 *tmp = '\0';
4228 tmp++;
4229 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4230 return -EINVAL;
4231
4232 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004233 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004235 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004237 if (limit < 1 || batchcount < 1 ||
4238 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004239 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004241 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004242 batchcount, shared,
4243 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244 }
4245 break;
4246 }
4247 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004248 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249 if (res >= 0)
4250 res = count;
4251 return res;
4252}
Al Viro871751e2006-03-25 03:06:39 -08004253
4254#ifdef CONFIG_DEBUG_SLAB_LEAK
4255
Al Viro871751e2006-03-25 03:06:39 -08004256static inline int add_caller(unsigned long *n, unsigned long v)
4257{
4258 unsigned long *p;
4259 int l;
4260 if (!v)
4261 return 1;
4262 l = n[1];
4263 p = n + 2;
4264 while (l) {
4265 int i = l/2;
4266 unsigned long *q = p + 2 * i;
4267 if (*q == v) {
4268 q[1]++;
4269 return 1;
4270 }
4271 if (*q > v) {
4272 l = i;
4273 } else {
4274 p = q + 2;
4275 l -= i + 1;
4276 }
4277 }
4278 if (++n[1] == n[0])
4279 return 0;
4280 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4281 p[0] = v;
4282 p[1] = 1;
4283 return 1;
4284}
4285
Joonsoo Kim8456a642013-10-24 10:07:49 +09004286static void handle_slab(unsigned long *n, struct kmem_cache *c,
4287 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004288{
4289 void *p;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004290 int i, j;
4291 unsigned long v;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004292
Al Viro871751e2006-03-25 03:06:39 -08004293 if (n[0] == n[1])
4294 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004295 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004296 bool active = true;
4297
4298 for (j = page->active; j < c->num; j++) {
4299 if (get_free_obj(page, j) == i) {
4300 active = false;
4301 break;
4302 }
4303 }
4304
4305 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004306 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004307
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004308 /*
4309 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4310 * mapping is established when actual object allocation and
4311 * we could mistakenly access the unmapped object in the cpu
4312 * cache.
4313 */
4314 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4315 continue;
4316
4317 if (!add_caller(n, v))
Al Viro871751e2006-03-25 03:06:39 -08004318 return;
4319 }
4320}
4321
4322static void show_symbol(struct seq_file *m, unsigned long address)
4323{
4324#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004325 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004326 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004327
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004328 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004329 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004330 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004331 seq_printf(m, " [%s]", modname);
4332 return;
4333 }
4334#endif
4335 seq_printf(m, "%p", (void *)address);
4336}
4337
4338static int leaks_show(struct seq_file *m, void *p)
4339{
Thierry Reding0672aa72012-06-22 19:42:49 +02004340 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004341 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004342 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004343 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004344 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004345 int node;
4346 int i;
4347
4348 if (!(cachep->flags & SLAB_STORE_USER))
4349 return 0;
4350 if (!(cachep->flags & SLAB_RED_ZONE))
4351 return 0;
4352
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004353 /*
4354 * Set store_user_clean and start to grab stored user information
4355 * for all objects on this cache. If some alloc/free requests comes
4356 * during the processing, information would be wrong so restart
4357 * whole processing.
4358 */
4359 do {
4360 set_store_user_clean(cachep);
4361 drain_cpu_caches(cachep);
Al Viro871751e2006-03-25 03:06:39 -08004362
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004363 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004364
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004365 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004366
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004367 check_irq_on();
4368 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004369
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004370 list_for_each_entry(page, &n->slabs_full, lru)
4371 handle_slab(x, cachep, page);
4372 list_for_each_entry(page, &n->slabs_partial, lru)
4373 handle_slab(x, cachep, page);
4374 spin_unlock_irq(&n->list_lock);
4375 }
4376 } while (!is_store_user_clean(cachep));
4377
Al Viro871751e2006-03-25 03:06:39 -08004378 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004379 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004380 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004381 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004382 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004383 if (!m->private) {
4384 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004385 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004386 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004387 return -ENOMEM;
4388 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004389 *(unsigned long *)m->private = x[0] * 2;
4390 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004391 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004392 /* Now make sure this entry will be retried */
4393 m->count = m->size;
4394 return 0;
4395 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004396 for (i = 0; i < x[1]; i++) {
4397 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4398 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004399 seq_putc(m, '\n');
4400 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004401
Al Viro871751e2006-03-25 03:06:39 -08004402 return 0;
4403}
4404
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004405static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004406 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004407 .next = slab_next,
4408 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004409 .show = leaks_show,
4410};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004411
4412static int slabstats_open(struct inode *inode, struct file *file)
4413{
Rob Jonesb208ce32014-10-09 15:28:03 -07004414 unsigned long *n;
4415
4416 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4417 if (!n)
4418 return -ENOMEM;
4419
4420 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4421
4422 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004423}
4424
4425static const struct file_operations proc_slabstats_operations = {
4426 .open = slabstats_open,
4427 .read = seq_read,
4428 .llseek = seq_lseek,
4429 .release = seq_release_private,
4430};
Al Viro871751e2006-03-25 03:06:39 -08004431#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004432
4433static int __init slab_proc_init(void)
4434{
4435#ifdef CONFIG_DEBUG_SLAB_LEAK
4436 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4437#endif
4438 return 0;
4439}
4440module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441#endif
4442
Kees Cook04385fc2016-06-23 15:20:59 -07004443#ifdef CONFIG_HARDENED_USERCOPY
4444/*
4445 * Rejects objects that are incorrectly sized.
4446 *
4447 * Returns NULL if check passes, otherwise const char * to name of cache
4448 * to indicate an error.
4449 */
4450const char *__check_heap_object(const void *ptr, unsigned long n,
4451 struct page *page)
4452{
4453 struct kmem_cache *cachep;
4454 unsigned int objnr;
4455 unsigned long offset;
4456
4457 /* Find and validate object. */
4458 cachep = page->slab_cache;
4459 objnr = obj_to_index(cachep, page, (void *)ptr);
4460 BUG_ON(objnr >= cachep->num);
4461
4462 /* Find offset within object. */
4463 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4464
4465 /* Allow address range falling entirely within object size. */
4466 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4467 return NULL;
4468
4469 return cachep->name;
4470}
4471#endif /* CONFIG_HARDENED_USERCOPY */
4472
Manfred Spraul00e145b2005-09-03 15:55:07 -07004473/**
4474 * ksize - get the actual amount of memory allocated for a given object
4475 * @objp: Pointer to the object
4476 *
4477 * kmalloc may internally round up allocations and return more memory
4478 * than requested. ksize() can be used to determine the actual amount of
4479 * memory allocated. The caller may use this additional memory, even though
4480 * a smaller amount of memory was initially specified with the kmalloc call.
4481 * The caller must guarantee that objp points to a valid object previously
4482 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4483 * must not be freed during the duration of the call.
4484 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004485size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004486{
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004487 size_t size;
4488
Christoph Lameteref8b4522007-10-16 01:24:46 -07004489 BUG_ON(!objp);
4490 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004491 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004492
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004493 size = virt_to_cache(objp)->object_size;
4494 /* We assume that ksize callers could use the whole allocated area,
4495 * so we need to unpoison this area.
4496 */
Alexander Potapenko4ebb31a42016-05-20 16:59:14 -07004497 kasan_unpoison_shadow(objp, size);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004498
4499 return size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004501EXPORT_SYMBOL(ksize);