blob: 04dec48c3ed7a12af3fef4958e928d0ddc31b64e [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>
Ingo Molnar3f8c2452017-02-05 14:31:22 +0100119#include <linux/sched/task_stack.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
Mel Gorman381760e2012-07-31 16:44:30 -0700121#include <net/sock.h>
122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#include <asm/cacheflush.h>
124#include <asm/tlbflush.h>
125#include <asm/page.h>
126
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500127#include <trace/events/kmem.h>
128
Mel Gorman072bb0a2012-07-31 16:43:58 -0700129#include "internal.h"
130
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800131#include "slab.h"
132
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700134 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 * 0 for faster, smaller code (especially in the critical paths).
136 *
137 * STATS - 1 to collect stats for /proc/slabinfo.
138 * 0 for faster, smaller code (especially in the critical paths).
139 *
140 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 */
142
143#ifdef CONFIG_DEBUG_SLAB
144#define DEBUG 1
145#define STATS 1
146#define FORCED_DEBUG 1
147#else
148#define DEBUG 0
149#define STATS 0
150#define FORCED_DEBUG 0
151#endif
152
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153/* Shouldn't this be in a header file somewhere? */
154#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400155#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#ifndef ARCH_KMALLOC_FLAGS
158#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
159#endif
160
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900161#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
162 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163
164#if FREELIST_BYTE_INDEX
165typedef unsigned char freelist_idx_t;
166#else
167typedef unsigned short freelist_idx_t;
168#endif
169
David Miller30321c72014-05-05 16:20:04 -0400170#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900171
Mel Gorman072bb0a2012-07-31 16:43:58 -0700172/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 * struct array_cache
174 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 * Purpose:
176 * - LIFO ordering, to hand out cache-warm objects from _alloc
177 * - reduce the number of linked list operations
178 * - reduce spinlock operations
179 *
180 * The limit is stored in the per-cpu structure to reduce the data cache
181 * footprint.
182 *
183 */
184struct array_cache {
185 unsigned int avail;
186 unsigned int limit;
187 unsigned int batchcount;
188 unsigned int touched;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700189 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800190 * Must have this definition in here for the proper
191 * alignment of array_cache. Also simplifies accessing
192 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800193 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194};
195
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700196struct alien_cache {
197 spinlock_t lock;
198 struct array_cache ac;
199};
200
Andrew Mortona737b3e2006-03-22 00:08:11 -0800201/*
Christoph Lametere498be72005-09-09 13:03:32 -0700202 * Need this for bootstrapping a per node allocator.
203 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700204#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000205static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700206#define CACHE_CACHE 0
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700207#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
Christoph Lametered11d9e2006-06-30 01:55:45 -0700209static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000210 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700211static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700212 int node, struct list_head *list);
213static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300214static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000215static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700216
Joonsoo Kim76b342b2016-05-19 17:10:26 -0700217static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
218 void **list);
219static inline void fixup_slab_list(struct kmem_cache *cachep,
220 struct kmem_cache_node *n, struct page *page,
221 void **list);
Ingo Molnare0a42722006-06-23 02:03:46 -0700222static int slab_early_init = 1;
223
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000224#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700225
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000226static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700227{
228 INIT_LIST_HEAD(&parent->slabs_full);
229 INIT_LIST_HEAD(&parent->slabs_partial);
230 INIT_LIST_HEAD(&parent->slabs_free);
David Rientjesbf00bd32016-12-12 16:41:44 -0800231 parent->total_slabs = 0;
Greg Thelenf728b0a2016-12-12 16:41:41 -0800232 parent->free_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700233 parent->shared = NULL;
234 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800235 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700236 spin_lock_init(&parent->list_lock);
237 parent->free_objects = 0;
238 parent->free_touched = 0;
239}
240
Andrew Mortona737b3e2006-03-22 00:08:11 -0800241#define MAKE_LIST(cachep, listp, slab, nodeid) \
242 do { \
243 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700244 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700245 } while (0)
246
Andrew Mortona737b3e2006-03-22 00:08:11 -0800247#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
248 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700249 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
250 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
252 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700254#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255#define CFLGS_OFF_SLAB (0x80000000UL)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700256#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
258
259#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800260/*
261 * Optimization question: fewer reaps means less probability for unnessary
262 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100264 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 * which could lock up otherwise freeable slabs.
266 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800267#define REAPTIMEOUT_AC (2*HZ)
268#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270#if STATS
271#define STATS_INC_ACTIVE(x) ((x)->num_active++)
272#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
273#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
274#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700275#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276#define STATS_SET_HIGH(x) \
277 do { \
278 if ((x)->num_active > (x)->high_mark) \
279 (x)->high_mark = (x)->num_active; \
280 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281#define STATS_INC_ERR(x) ((x)->errors++)
282#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700283#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700284#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800285#define STATS_SET_FREEABLE(x, i) \
286 do { \
287 if ((x)->max_freeable < i) \
288 (x)->max_freeable = i; \
289 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
291#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
292#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
293#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
294#else
295#define STATS_INC_ACTIVE(x) do { } while (0)
296#define STATS_DEC_ACTIVE(x) do { } while (0)
297#define STATS_INC_ALLOCED(x) do { } while (0)
298#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700299#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300#define STATS_SET_HIGH(x) do { } while (0)
301#define STATS_INC_ERR(x) do { } while (0)
302#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700303#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700304#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800305#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306#define STATS_INC_ALLOCHIT(x) do { } while (0)
307#define STATS_INC_ALLOCMISS(x) do { } while (0)
308#define STATS_INC_FREEHIT(x) do { } while (0)
309#define STATS_INC_FREEMISS(x) do { } while (0)
310#endif
311
312#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Andrew Mortona737b3e2006-03-22 00:08:11 -0800314/*
315 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 * the end of an object is aligned with the end of the real
319 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800322 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
324 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800325 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800327static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800329 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330}
331
David Woodhouseb46b8f12007-05-08 00:22:59 -0700332static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333{
334 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700335 return (unsigned long long*) (objp + obj_offset(cachep) -
336 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337}
338
David Woodhouseb46b8f12007-05-08 00:22:59 -0700339static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340{
341 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
342 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500343 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700344 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400345 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500346 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700347 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348}
349
Pekka Enberg343e0d72006-02-01 03:05:50 -0800350static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351{
352 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500353 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
356#else
357
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800358#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
360#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
362
363#endif
364
Joonsoo Kim03787302014-06-23 13:22:06 -0700365#ifdef CONFIG_DEBUG_SLAB_LEAK
366
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700367static inline bool is_store_user_clean(struct kmem_cache *cachep)
Joonsoo Kim03787302014-06-23 13:22:06 -0700368{
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700369 return atomic_read(&cachep->store_user_clean) == 1;
370}
Joonsoo Kim03787302014-06-23 13:22:06 -0700371
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700372static inline void set_store_user_clean(struct kmem_cache *cachep)
373{
374 atomic_set(&cachep->store_user_clean, 1);
375}
Joonsoo Kim03787302014-06-23 13:22:06 -0700376
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700377static inline void set_store_user_dirty(struct kmem_cache *cachep)
378{
379 if (is_store_user_clean(cachep))
380 atomic_set(&cachep->store_user_clean, 0);
Joonsoo Kim03787302014-06-23 13:22:06 -0700381}
382
383#else
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700384static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
Joonsoo Kim03787302014-06-23 13:22:06 -0700385
386#endif
387
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700389 * Do not go above this order unless 0 objects fit into the slab or
390 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 */
David Rientjes543585c2011-10-18 22:09:24 -0700392#define SLAB_MAX_ORDER_HI 1
393#define SLAB_MAX_ORDER_LO 0
394static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700395static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800397static inline struct kmem_cache *virt_to_cache(const void *obj)
398{
Christoph Lameterb49af682007-05-06 14:49:41 -0700399 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500400 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800401}
402
Joonsoo Kim8456a642013-10-24 10:07:49 +0900403static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800404 unsigned int idx)
405{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900406 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800407}
408
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800409/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500410 * We want to avoid an expensive divide : (offset / cache->size)
411 * Using the fact that size is a constant for a particular cache,
412 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800413 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
414 */
415static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900416 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800417{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900418 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800419 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800420}
421
Joonsoo Kim6fb92432016-03-15 14:54:09 -0700422#define BOOT_CPUCACHE_ENTRIES 1
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000424static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800425 .batchcount = 1,
426 .limit = BOOT_CPUCACHE_ENTRIES,
427 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500428 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430};
431
Tejun Heo1871e522009-10-29 22:34:13 +0900432static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Pekka Enberg343e0d72006-02-01 03:05:50 -0800434static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700436 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
Andrew Mortona737b3e2006-03-22 00:08:11 -0800439/*
440 * Calculate the number of objects and left-over bytes for a given buffer size.
441 */
Joonsoo Kim70f75062016-03-15 14:54:53 -0700442static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
443 unsigned long flags, size_t *left_over)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800444{
Joonsoo Kim70f75062016-03-15 14:54:53 -0700445 unsigned int num;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800446 size_t slab_size = PAGE_SIZE << gfporder;
447
448 /*
449 * The slab management structure can be either off the slab or
450 * on it. For the latter case, the memory allocated for a
451 * slab is used for:
452 *
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800453 * - @buffer_size bytes for each object
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700454 * - One freelist_idx_t for each object
455 *
456 * We don't need to consider alignment of freelist because
457 * freelist will be at the end of slab page. The objects will be
458 * at the correct alignment.
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800459 *
460 * If the slab management structure is off the slab, then the
461 * alignment will already be calculated into the size. Because
462 * the slabs are all pages aligned, the objects will be at the
463 * correct alignment when allocated.
464 */
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700465 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700466 num = slab_size / buffer_size;
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700467 *left_over = slab_size % buffer_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800468 } else {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700469 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700470 *left_over = slab_size %
471 (buffer_size + sizeof(freelist_idx_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 }
Joonsoo Kim70f75062016-03-15 14:54:53 -0700473
474 return num;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Christoph Lameterf28510d2012-09-11 19:49:38 +0000477#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700478#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
Andrew Mortona737b3e2006-03-22 00:08:11 -0800480static void __slab_error(const char *function, struct kmem_cache *cachep,
481 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482{
Joe Perches11705322016-03-17 14:19:50 -0700483 pr_err("slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800484 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030486 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000488#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
Paul Menage3395ee02006-12-06 20:32:16 -0800490/*
491 * By default on NUMA we use alien caches to stage the freeing of
492 * objects allocated from other nodes. This causes massive memory
493 * inefficiencies when using fake NUMA setup to split memory into a
494 * large number of small nodes, so it can be disabled on the command
495 * line
496 */
497
498static int use_alien_caches __read_mostly = 1;
499static int __init noaliencache_setup(char *s)
500{
501 use_alien_caches = 0;
502 return 1;
503}
504__setup("noaliencache", noaliencache_setup);
505
David Rientjes3df1ccc2011-10-18 22:09:28 -0700506static int __init slab_max_order_setup(char *str)
507{
508 get_option(&str, &slab_max_order);
509 slab_max_order = slab_max_order < 0 ? 0 :
510 min(slab_max_order, MAX_ORDER - 1);
511 slab_max_order_set = true;
512
513 return 1;
514}
515__setup("slab_max_order=", slab_max_order_setup);
516
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800517#ifdef CONFIG_NUMA
518/*
519 * Special reaping functions for NUMA systems called from cache_reap().
520 * These take care of doing round robin flushing of alien caches (containing
521 * objects freed on different nodes from which they were allocated) and the
522 * flushing of remote pcps by calling drain_node_pages.
523 */
Tejun Heo1871e522009-10-29 22:34:13 +0900524static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800525
526static void init_reap_node(int cpu)
527{
Andrew Morton0edaf862016-05-19 17:10:58 -0700528 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
529 node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800530}
531
532static void next_reap_node(void)
533{
Christoph Lameter909ea962010-12-08 16:22:55 +0100534 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800535
Andrew Morton0edaf862016-05-19 17:10:58 -0700536 node = next_node_in(node, node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100537 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800538}
539
540#else
541#define init_reap_node(cpu) do { } while (0)
542#define next_reap_node(void) do { } while (0)
543#endif
544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545/*
546 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
547 * via the workqueue/eventd.
548 * Add the CPU number into the expiration time to minimize the possibility of
549 * the CPUs getting into lockstep and contending for the global cache chain
550 * lock.
551 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400552static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
Tejun Heo1871e522009-10-29 22:34:13 +0900554 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
Tejun Heoeac03372016-09-16 15:49:34 -0400556 if (reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800557 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700558 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800559 schedule_delayed_work_on(cpu, reap_work,
560 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 }
562}
563
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700564static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100566 /*
567 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300568 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100569 * cache the pointers are not cleared and they could be counted as
570 * valid references during a kmemleak scan. Therefore, kmemleak must
571 * not scan such objects.
572 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700573 kmemleak_no_scan(ac);
574 if (ac) {
575 ac->avail = 0;
576 ac->limit = limit;
577 ac->batchcount = batch;
578 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700580}
581
582static struct array_cache *alloc_arraycache(int node, int entries,
583 int batchcount, gfp_t gfp)
584{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700585 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700586 struct array_cache *ac = NULL;
587
588 ac = kmalloc_node(memsize, gfp, node);
589 init_arraycache(ac, entries, batchcount);
590 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591}
592
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700593static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
594 struct page *page, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700595{
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700596 struct kmem_cache_node *n;
597 int page_node;
598 LIST_HEAD(list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700599
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700600 page_node = page_to_nid(page);
601 n = get_node(cachep, page_node);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700602
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700603 spin_lock(&n->list_lock);
604 free_block(cachep, &objp, 1, page_node, &list);
605 spin_unlock(&n->list_lock);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700606
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700607 slabs_destroy(cachep, &list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700608}
609
Christoph Lameter3ded1752006-03-25 03:06:44 -0800610/*
611 * Transfer objects in one arraycache to another.
612 * Locking must be handled by the caller.
613 *
614 * Return the number of entries transferred.
615 */
616static int transfer_objects(struct array_cache *to,
617 struct array_cache *from, unsigned int max)
618{
619 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700620 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800621
622 if (!nr)
623 return 0;
624
625 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
626 sizeof(void *) *nr);
627
628 from->avail -= nr;
629 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800630 return nr;
631}
632
Christoph Lameter765c4502006-09-27 01:50:08 -0700633#ifndef CONFIG_NUMA
634
635#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000636#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700637
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700638static inline struct alien_cache **alloc_alien_cache(int node,
639 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700640{
Joonsoo Kim88881772016-05-19 17:10:05 -0700641 return NULL;
Christoph Lameter765c4502006-09-27 01:50:08 -0700642}
643
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700644static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700645{
646}
647
648static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
649{
650 return 0;
651}
652
653static inline void *alternate_node_alloc(struct kmem_cache *cachep,
654 gfp_t flags)
655{
656 return NULL;
657}
658
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800659static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700660 gfp_t flags, int nodeid)
661{
662 return NULL;
663}
664
David Rientjes4167e9b2015-04-14 15:46:55 -0700665static inline gfp_t gfp_exact_node(gfp_t flags)
666{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700667 return flags & ~__GFP_NOFAIL;
David Rientjes4167e9b2015-04-14 15:46:55 -0700668}
669
Christoph Lameter765c4502006-09-27 01:50:08 -0700670#else /* CONFIG_NUMA */
671
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800672static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800673static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800674
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700675static struct alien_cache *__alloc_alien_cache(int node, int entries,
676 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700677{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700678 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700679 struct alien_cache *alc = NULL;
680
681 alc = kmalloc_node(memsize, gfp, node);
682 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700683 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700684 return alc;
685}
686
687static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
688{
689 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700690 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700691 int i;
692
693 if (limit > 1)
694 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700695 alc_ptr = kzalloc_node(memsize, gfp, node);
696 if (!alc_ptr)
697 return NULL;
698
699 for_each_node(i) {
700 if (i == node || !node_online(i))
701 continue;
702 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
703 if (!alc_ptr[i]) {
704 for (i--; i >= 0; i--)
705 kfree(alc_ptr[i]);
706 kfree(alc_ptr);
707 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700708 }
709 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700710 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700711}
712
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700713static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700714{
715 int i;
716
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700717 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700718 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700719 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700720 kfree(alc_ptr[i]);
721 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700722}
723
Pekka Enberg343e0d72006-02-01 03:05:50 -0800724static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700725 struct array_cache *ac, int node,
726 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700727{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700728 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700729
730 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000731 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800732 /*
733 * Stuff objects into the remote nodes shared array first.
734 * That way we could avoid the overhead of putting the objects
735 * into the free lists and getting them back later.
736 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000737 if (n->shared)
738 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800739
Joonsoo Kim833b7062014-08-06 16:04:33 -0700740 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700741 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000742 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700743 }
744}
745
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800746/*
747 * Called from cache_reap() to regularly drain alien caches round robin.
748 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000749static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800750{
Christoph Lameter909ea962010-12-08 16:22:55 +0100751 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800752
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000753 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700754 struct alien_cache *alc = n->alien[node];
755 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800756
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700757 if (alc) {
758 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700759 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700760 LIST_HEAD(list);
761
762 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700763 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700764 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700765 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800766 }
767 }
768}
769
Andrew Mortona737b3e2006-03-22 00:08:11 -0800770static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700771 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700772{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800773 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700774 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700775 struct array_cache *ac;
776 unsigned long flags;
777
778 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700779 alc = alien[i];
780 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700781 LIST_HEAD(list);
782
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700783 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700784 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700785 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700786 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700787 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700788 }
789 }
790}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700791
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700792static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
793 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700794{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000795 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700796 struct alien_cache *alien = NULL;
797 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700798 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700799
Christoph Lameter18bf8542014-08-06 16:04:11 -0700800 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700801 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700802 if (n->alien && n->alien[page_node]) {
803 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700804 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700805 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700806 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700807 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700808 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700809 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700810 ac->entry[ac->avail++] = objp;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700811 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700812 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700813 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700814 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700815 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700816 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700817 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -0700818 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700819 }
820 return 1;
821}
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700822
823static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
824{
825 int page_node = page_to_nid(virt_to_page(objp));
826 int node = numa_mem_id();
827 /*
828 * Make sure we are not freeing a object from another node to the array
829 * cache on this cpu.
830 */
831 if (likely(node == page_node))
832 return 0;
833
834 return __cache_free_alien(cachep, objp, node, page_node);
835}
David Rientjes4167e9b2015-04-14 15:46:55 -0700836
837/*
Mel Gorman444eb2a42016-03-17 14:19:23 -0700838 * Construct gfp mask to allocate from a specific node but do not reclaim or
839 * warn about failures.
David Rientjes4167e9b2015-04-14 15:46:55 -0700840 */
841static inline gfp_t gfp_exact_node(gfp_t flags)
842{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700843 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
David Rientjes4167e9b2015-04-14 15:46:55 -0700844}
Christoph Lametere498be72005-09-09 13:03:32 -0700845#endif
846
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700847static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
848{
849 struct kmem_cache_node *n;
850
851 /*
852 * Set up the kmem_cache_node for cpu before we can
853 * begin anything. Make sure some other cpu on this
854 * node has not already allocated this
855 */
856 n = get_node(cachep, node);
857 if (n) {
858 spin_lock_irq(&n->list_lock);
859 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
860 cachep->num;
861 spin_unlock_irq(&n->list_lock);
862
863 return 0;
864 }
865
866 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
867 if (!n)
868 return -ENOMEM;
869
870 kmem_cache_node_init(n);
871 n->next_reap = jiffies + REAPTIMEOUT_NODE +
872 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
873
874 n->free_limit =
875 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
876
877 /*
878 * The kmem_cache_nodes don't come and go as CPUs
879 * come and go. slab_mutex is sufficient
880 * protection here.
881 */
882 cachep->node[node] = n;
883
884 return 0;
885}
886
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200887#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700888/*
Christoph Lameter6a673682013-01-10 19:14:19 +0000889 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000890 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -0700891 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +0000892 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -0700893 * already in use.
894 *
Christoph Lameter18004c52012-07-06 15:25:12 -0500895 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -0700896 */
Christoph Lameter6a673682013-01-10 19:14:19 +0000897static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700898{
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700899 int ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700900 struct kmem_cache *cachep;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700901
Christoph Lameter18004c52012-07-06 15:25:12 -0500902 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700903 ret = init_cache_node(cachep, node, GFP_KERNEL);
904 if (ret)
905 return ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700906 }
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700907
David Rientjes8f9f8d92010-03-27 19:40:47 -0700908 return 0;
909}
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200910#endif
David Rientjes8f9f8d92010-03-27 19:40:47 -0700911
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700912static int setup_kmem_cache_node(struct kmem_cache *cachep,
913 int node, gfp_t gfp, bool force_change)
914{
915 int ret = -ENOMEM;
916 struct kmem_cache_node *n;
917 struct array_cache *old_shared = NULL;
918 struct array_cache *new_shared = NULL;
919 struct alien_cache **new_alien = NULL;
920 LIST_HEAD(list);
921
922 if (use_alien_caches) {
923 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
924 if (!new_alien)
925 goto fail;
926 }
927
928 if (cachep->shared) {
929 new_shared = alloc_arraycache(node,
930 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
931 if (!new_shared)
932 goto fail;
933 }
934
935 ret = init_cache_node(cachep, node, gfp);
936 if (ret)
937 goto fail;
938
939 n = get_node(cachep, node);
940 spin_lock_irq(&n->list_lock);
941 if (n->shared && force_change) {
942 free_block(cachep, n->shared->entry,
943 n->shared->avail, node, &list);
944 n->shared->avail = 0;
945 }
946
947 if (!n->shared || force_change) {
948 old_shared = n->shared;
949 n->shared = new_shared;
950 new_shared = NULL;
951 }
952
953 if (!n->alien) {
954 n->alien = new_alien;
955 new_alien = NULL;
956 }
957
958 spin_unlock_irq(&n->list_lock);
959 slabs_destroy(cachep, &list);
960
Joonsoo Kim801faf02016-05-19 17:10:31 -0700961 /*
962 * To protect lockless access to n->shared during irq disabled context.
963 * If n->shared isn't NULL in irq disabled context, accessing to it is
964 * guaranteed to be valid until irq is re-enabled, because it will be
965 * freed after synchronize_sched().
966 */
Joonsoo Kim86d9f482016-10-27 17:46:18 -0700967 if (old_shared && force_change)
Joonsoo Kim801faf02016-05-19 17:10:31 -0700968 synchronize_sched();
969
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700970fail:
971 kfree(old_shared);
972 kfree(new_shared);
973 free_alien_cache(new_alien);
974
975 return ret;
976}
977
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200978#ifdef CONFIG_SMP
979
Paul Gortmaker0db06282013-06-19 14:53:51 -0400980static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981{
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700982 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000983 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700984 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +1030985 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700986
Christoph Lameter18004c52012-07-06 15:25:12 -0500987 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700988 struct array_cache *nc;
989 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700990 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700991 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700992
Christoph Lameter18bf8542014-08-06 16:04:11 -0700993 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000994 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700995 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700996
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000997 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700998
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000999 /* Free limit for this kmem_cache_node */
1000 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001001
1002 /* cpu is dead; no one can alloc from it. */
1003 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1004 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001005 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001006 nc->avail = 0;
1007 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001008
Rusty Russell58463c12009-12-17 11:43:12 -06001009 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001010 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001011 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001012 }
1013
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001014 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001015 if (shared) {
1016 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001017 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001018 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001019 }
1020
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001021 alien = n->alien;
1022 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001023
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001024 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001025
1026 kfree(shared);
1027 if (alien) {
1028 drain_alien_cache(cachep, alien);
1029 free_alien_cache(alien);
1030 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001031
1032free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001033 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001034 }
1035 /*
1036 * In the previous loop, all the objects were freed to
1037 * the respective cache's slabs, now we can go ahead and
1038 * shrink each nodelist to its limit.
1039 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001040 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001041 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001042 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001043 continue;
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001044 drain_freelist(cachep, n, INT_MAX);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001045 }
1046}
1047
Paul Gortmaker0db06282013-06-19 14:53:51 -04001048static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001049{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001050 struct kmem_cache *cachep;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001051 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001052 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001054 /*
1055 * We need to do this right in the beginning since
1056 * alloc_arraycache's are going to use this list.
1057 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001058 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001059 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001060 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001061 if (err < 0)
1062 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001063
1064 /*
1065 * Now we can go ahead with allocating the shared arrays and
1066 * array caches
1067 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001068 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07001069 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1070 if (err)
1071 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001072 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001073
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001074 return 0;
1075bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001076 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001077 return -ENOMEM;
1078}
1079
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001080int slab_prepare_cpu(unsigned int cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001081{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001082 int err;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001083
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001084 mutex_lock(&slab_mutex);
1085 err = cpuup_prepare(cpu);
1086 mutex_unlock(&slab_mutex);
1087 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088}
1089
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001090/*
1091 * This is called for a failed online attempt and for a successful
1092 * offline.
1093 *
1094 * Even if all the cpus of a node are down, we don't free the
1095 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1096 * a kmalloc allocation from another cpu for memory from the node of
1097 * the cpu going down. The list3 structure is usually allocated from
1098 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1099 */
1100int slab_dead_cpu(unsigned int cpu)
1101{
1102 mutex_lock(&slab_mutex);
1103 cpuup_canceled(cpu);
1104 mutex_unlock(&slab_mutex);
1105 return 0;
1106}
1107#endif
1108
1109static int slab_online_cpu(unsigned int cpu)
1110{
1111 start_cpu_timer(cpu);
1112 return 0;
1113}
1114
1115static int slab_offline_cpu(unsigned int cpu)
1116{
1117 /*
1118 * Shutdown cache reaper. Note that the slab_mutex is held so
1119 * that if cache_reap() is invoked it cannot do anything
1120 * expensive but will only modify reap_work and reschedule the
1121 * timer.
1122 */
1123 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1124 /* Now the cache_reaper is guaranteed to be not running. */
1125 per_cpu(slab_reap_work, cpu).work.func = NULL;
1126 return 0;
1127}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
David Rientjes8f9f8d92010-03-27 19:40:47 -07001129#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1130/*
1131 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1132 * Returns -EBUSY if all objects cannot be drained so that the node is not
1133 * removed.
1134 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001135 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001136 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001137static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001138{
1139 struct kmem_cache *cachep;
1140 int ret = 0;
1141
Christoph Lameter18004c52012-07-06 15:25:12 -05001142 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001143 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001144
Christoph Lameter18bf8542014-08-06 16:04:11 -07001145 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001146 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001147 continue;
1148
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001149 drain_freelist(cachep, n, INT_MAX);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001150
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001151 if (!list_empty(&n->slabs_full) ||
1152 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001153 ret = -EBUSY;
1154 break;
1155 }
1156 }
1157 return ret;
1158}
1159
1160static int __meminit slab_memory_callback(struct notifier_block *self,
1161 unsigned long action, void *arg)
1162{
1163 struct memory_notify *mnb = arg;
1164 int ret = 0;
1165 int nid;
1166
1167 nid = mnb->status_change_nid;
1168 if (nid < 0)
1169 goto out;
1170
1171 switch (action) {
1172 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001173 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001174 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001175 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001176 break;
1177 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001178 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001179 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001180 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001181 break;
1182 case MEM_ONLINE:
1183 case MEM_OFFLINE:
1184 case MEM_CANCEL_ONLINE:
1185 case MEM_CANCEL_OFFLINE:
1186 break;
1187 }
1188out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001189 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001190}
1191#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1192
Christoph Lametere498be72005-09-09 13:03:32 -07001193/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001194 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001195 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001196static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001197 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001198{
Christoph Lameter6744f082013-01-10 19:12:17 +00001199 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001200
Christoph Lameter6744f082013-01-10 19:12:17 +00001201 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001202 BUG_ON(!ptr);
1203
Christoph Lameter6744f082013-01-10 19:12:17 +00001204 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001205 /*
1206 * Do not assume that spinlocks can be initialized via memcpy:
1207 */
1208 spin_lock_init(&ptr->list_lock);
1209
Christoph Lametere498be72005-09-09 13:03:32 -07001210 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001211 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001212}
1213
Andrew Mortona737b3e2006-03-22 00:08:11 -08001214/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001215 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1216 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001217 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001218static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001219{
1220 int node;
1221
1222 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001223 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001224 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001225 REAPTIMEOUT_NODE +
1226 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001227 }
1228}
1229
1230/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001231 * Initialisation. Called after the page allocator have been initialised and
1232 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 */
1234void __init kmem_cache_init(void)
1235{
Christoph Lametere498be72005-09-09 13:03:32 -07001236 int i;
1237
Joonsoo Kim68126702013-10-24 10:07:42 +09001238 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1239 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001240 kmem_cache = &kmem_cache_boot;
1241
Joonsoo Kim88881772016-05-19 17:10:05 -07001242 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001243 use_alien_caches = 0;
1244
Christoph Lameter3c583462012-11-28 16:23:01 +00001245 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001246 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001247
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 /*
1249 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001250 * page orders on machines with more than 32MB of memory if
1251 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001253 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001254 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 /* Bootstrap is tricky, because several objects are allocated
1257 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001258 * 1) initialize the kmem_cache cache: it contains the struct
1259 * kmem_cache structures of all caches, except kmem_cache itself:
1260 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001261 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001262 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001263 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001265 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001266 * An __init data area is used for the head array.
1267 * 3) Create the remaining kmalloc caches, with minimally sized
1268 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001269 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001271 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001272 * the other cache's with kmalloc allocated memory.
1273 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 */
1275
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001276 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Eric Dumazet8da34302007-05-06 14:49:29 -07001278 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001279 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001280 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001281 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001282 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001283 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001284 SLAB_HWCACHE_ALIGN);
1285 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001286 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
Andrew Mortona737b3e2006-03-22 00:08:11 -08001288 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001289 * Initialize the caches that provide memory for the kmem_cache_node
1290 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001291 */
Vlastimil Babkaaf3b5f82017-02-22 15:41:05 -08001292 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache(
1293 kmalloc_info[INDEX_NODE].name,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001294 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001295 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001296 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001297
Ingo Molnare0a42722006-06-23 02:03:46 -07001298 slab_early_init = 0;
1299
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001300 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001301 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001302 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Mel Gorman9c09a952008-01-24 05:49:54 -08001304 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001306
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001307 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001309 }
1310 }
1311
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001312 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001313}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001314
Pekka Enberg8429db52009-06-12 15:58:59 +03001315void __init kmem_cache_init_late(void)
1316{
1317 struct kmem_cache *cachep;
1318
Christoph Lameter97d06602012-07-06 15:25:11 -05001319 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001320
Pekka Enberg8429db52009-06-12 15:58:59 +03001321 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001322 mutex_lock(&slab_mutex);
1323 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001324 if (enable_cpucache(cachep, GFP_NOWAIT))
1325 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001326 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001327
Christoph Lameter97d06602012-07-06 15:25:11 -05001328 /* Done! */
1329 slab_state = FULL;
1330
David Rientjes8f9f8d92010-03-27 19:40:47 -07001331#ifdef CONFIG_NUMA
1332 /*
1333 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001334 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001335 */
1336 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1337#endif
1338
Andrew Mortona737b3e2006-03-22 00:08:11 -08001339 /*
1340 * The reap timers are started later, with a module init call: That part
1341 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 */
1343}
1344
1345static int __init cpucache_init(void)
1346{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001347 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Andrew Mortona737b3e2006-03-22 00:08:11 -08001349 /*
1350 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 */
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001352 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1353 slab_online_cpu, slab_offline_cpu);
1354 WARN_ON(ret < 0);
Glauber Costaa164f8962012-06-21 00:59:18 +04001355
1356 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001357 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 return 0;
1359}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360__initcall(cpucache_init);
1361
Rafael Aquini8bdec192012-03-09 17:27:27 -03001362static noinline void
1363slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1364{
David Rientjes9a02d692014-06-04 16:06:36 -07001365#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001366 struct kmem_cache_node *n;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001367 unsigned long flags;
1368 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001369 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1370 DEFAULT_RATELIMIT_BURST);
1371
1372 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1373 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001374
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001375 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1376 nodeid, gfpflags, &gfpflags);
1377 pr_warn(" cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001378 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001379
Christoph Lameter18bf8542014-08-06 16:04:11 -07001380 for_each_kmem_cache_node(cachep, node, n) {
David Rientjesbf00bd32016-12-12 16:41:44 -08001381 unsigned long total_slabs, free_slabs, free_objs;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001382
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001383 spin_lock_irqsave(&n->list_lock, flags);
David Rientjesbf00bd32016-12-12 16:41:44 -08001384 total_slabs = n->total_slabs;
1385 free_slabs = n->free_slabs;
1386 free_objs = n->free_objects;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001387 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001388
David Rientjesbf00bd32016-12-12 16:41:44 -08001389 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1390 node, total_slabs - free_slabs, total_slabs,
1391 (total_slabs * cachep->num) - free_objs,
1392 total_slabs * cachep->num);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001393 }
David Rientjes9a02d692014-06-04 16:06:36 -07001394#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001395}
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001398 * Interface to system's page allocator. No need to hold the
1399 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 *
1401 * If we requested dmaable memory, we will get it. Even if we
1402 * did not request dmaable memory, we might get it, but that
1403 * would be relatively rare and ignorable.
1404 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001405static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1406 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407{
1408 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001409 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001410
Glauber Costaa618e892012-06-14 16:17:21 +04001411 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001412 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1413 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001414
Vlastimil Babka96db8002015-09-08 15:03:50 -07001415 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001416 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001417 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001419 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001421 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1422 __free_pages(page, cachep->gfporder);
1423 return NULL;
1424 }
1425
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001426 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Johannes Weiner7779f212017-07-06 15:40:55 -07001428 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages);
Christoph Lameter972d1a72006-09-25 23:31:51 -07001429 else
Johannes Weiner7779f212017-07-06 15:40:55 -07001430 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001431
Joonsoo Kima57a4982013-10-24 10:07:44 +09001432 __SetPageSlab(page);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001433 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1434 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001435 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001436
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001437 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1438 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1439
1440 if (cachep->ctor)
1441 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1442 else
1443 kmemcheck_mark_unallocated_pages(page, nr_pages);
1444 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001445
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001446 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447}
1448
1449/*
1450 * Interface to system's page release.
1451 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001452static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453{
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001454 int order = cachep->gfporder;
1455 unsigned long nr_freed = (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001457 kmemcheck_free_shadow(page, order);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001458
Christoph Lameter972d1a72006-09-25 23:31:51 -07001459 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Johannes Weiner7779f212017-07-06 15:40:55 -07001460 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, -nr_freed);
Christoph Lameter972d1a72006-09-25 23:31:51 -07001461 else
Johannes Weiner7779f212017-07-06 15:40:55 -07001462 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, -nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001463
Joonsoo Kima57a4982013-10-24 10:07:44 +09001464 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001465 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001466 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001467 page_mapcount_reset(page);
1468 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001469
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 if (current->reclaim_state)
1471 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001472 memcg_uncharge_slab(page, order, cachep);
1473 __free_pages(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474}
1475
1476static void kmem_rcu_free(struct rcu_head *head)
1477{
Joonsoo Kim68126702013-10-24 10:07:42 +09001478 struct kmem_cache *cachep;
1479 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480
Joonsoo Kim68126702013-10-24 10:07:42 +09001481 page = container_of(head, struct page, rcu_head);
1482 cachep = page->slab_cache;
1483
1484 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485}
1486
1487#if DEBUG
Joonsoo Kim40b44132016-03-15 14:54:21 -07001488static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1489{
1490 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1491 (cachep->size % PAGE_SIZE) == 0)
1492 return true;
1493
1494 return false;
1495}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
1497#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001498static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001499 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001501 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001503 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001505 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 return;
1507
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 *addr++ = 0x12345678;
1509 *addr++ = caller;
1510 *addr++ = smp_processor_id();
1511 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 {
1513 unsigned long *sptr = &caller;
1514 unsigned long svalue;
1515
1516 while (!kstack_end(sptr)) {
1517 svalue = *sptr++;
1518 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001519 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 size -= sizeof(unsigned long);
1521 if (size <= sizeof(unsigned long))
1522 break;
1523 }
1524 }
1525
1526 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001527 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528}
Joonsoo Kim40b44132016-03-15 14:54:21 -07001529
1530static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1531 int map, unsigned long caller)
1532{
1533 if (!is_debug_pagealloc_cache(cachep))
1534 return;
1535
1536 if (caller)
1537 store_stackinfo(cachep, objp, caller);
1538
1539 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1540}
1541
1542#else
1543static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1544 int map, unsigned long caller) {}
1545
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546#endif
1547
Pekka Enberg343e0d72006-02-01 03:05:50 -08001548static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001550 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001551 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
1553 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001554 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555}
1556
1557static void dump_line(char *data, int offset, int limit)
1558{
1559 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001560 unsigned char error = 0;
1561 int bad_count = 0;
1562
Joe Perches11705322016-03-17 14:19:50 -07001563 pr_err("%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001564 for (i = 0; i < limit; i++) {
1565 if (data[offset + i] != POISON_FREE) {
1566 error = data[offset + i];
1567 bad_count++;
1568 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001569 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001570 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1571 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001572
1573 if (bad_count == 1) {
1574 error ^= POISON_FREE;
1575 if (!(error & (error - 1))) {
Joe Perches11705322016-03-17 14:19:50 -07001576 pr_err("Single bit error detected. Probably bad RAM.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001577#ifdef CONFIG_X86
Joe Perches11705322016-03-17 14:19:50 -07001578 pr_err("Run memtest86+ or a similar memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001579#else
Joe Perches11705322016-03-17 14:19:50 -07001580 pr_err("Run a memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001581#endif
1582 }
1583 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584}
1585#endif
1586
1587#if DEBUG
1588
Pekka Enberg343e0d72006-02-01 03:05:50 -08001589static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590{
1591 int i, size;
1592 char *realobj;
1593
1594 if (cachep->flags & SLAB_RED_ZONE) {
Joe Perches11705322016-03-17 14:19:50 -07001595 pr_err("Redzone: 0x%llx/0x%llx\n",
1596 *dbg_redzone1(cachep, objp),
1597 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 }
1599
1600 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches11705322016-03-17 14:19:50 -07001601 pr_err("Last user: [<%p>](%pSR)\n",
Joe Perches071361d2012-12-12 10:19:12 -08001602 *dbg_userword(cachep, objp),
1603 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001605 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001606 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001607 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 int limit;
1609 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001610 if (i + limit > size)
1611 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 dump_line(realobj, i, limit);
1613 }
1614}
1615
Pekka Enberg343e0d72006-02-01 03:05:50 -08001616static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617{
1618 char *realobj;
1619 int size, i;
1620 int lines = 0;
1621
Joonsoo Kim40b44132016-03-15 14:54:21 -07001622 if (is_debug_pagealloc_cache(cachep))
1623 return;
1624
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001625 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001626 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001628 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001630 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 exp = POISON_END;
1632 if (realobj[i] != exp) {
1633 int limit;
1634 /* Mismatch ! */
1635 /* Print header */
1636 if (lines == 0) {
Joe Perches11705322016-03-17 14:19:50 -07001637 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1638 print_tainted(), cachep->name,
1639 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 print_objinfo(cachep, objp, 0);
1641 }
1642 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001643 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001645 if (i + limit > size)
1646 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 dump_line(realobj, i, limit);
1648 i += 16;
1649 lines++;
1650 /* Limit to 5 lines */
1651 if (lines > 5)
1652 break;
1653 }
1654 }
1655 if (lines != 0) {
1656 /* Print some data about the neighboring objects, if they
1657 * exist:
1658 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001659 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001660 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
Joonsoo Kim8456a642013-10-24 10:07:49 +09001662 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001664 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001665 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001666 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 print_objinfo(cachep, objp, 2);
1668 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001669 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001670 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001671 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001672 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 print_objinfo(cachep, objp, 2);
1674 }
1675 }
1676}
1677#endif
1678
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001680static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1681 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001682{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 int i;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001684
1685 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1686 poison_obj(cachep, page->freelist - obj_offset(cachep),
1687 POISON_FREE);
1688 }
1689
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001691 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692
1693 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07001695 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 }
1697 if (cachep->flags & SLAB_RED_ZONE) {
1698 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07001699 slab_error(cachep, "start of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07001701 slab_error(cachep, "end of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001704}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001706static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1707 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001708{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001709}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710#endif
1711
Randy Dunlap911851e2006-03-22 00:08:14 -08001712/**
1713 * slab_destroy - destroy and release all objects in a slab
1714 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001715 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001716 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001717 * Destroy all the objs in a slab page, and release the mem back to the system.
1718 * Before calling the slab page must have been unlinked from the cache. The
1719 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001720 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001721static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001722{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001723 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001724
Joonsoo Kim8456a642013-10-24 10:07:49 +09001725 freelist = page->freelist;
1726 slab_destroy_debugcheck(cachep, page);
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08001727 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001728 call_rcu(&page->rcu_head, kmem_rcu_free);
1729 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001730 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001731
1732 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001733 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001734 * although actual page can be freed in rcu context
1735 */
1736 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001737 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738}
1739
Joonsoo Kim97654df2014-08-06 16:04:25 -07001740static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1741{
1742 struct page *page, *n;
1743
1744 list_for_each_entry_safe(page, n, list, lru) {
1745 list_del(&page->lru);
1746 slab_destroy(cachep, page);
1747 }
1748}
1749
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001751 * calculate_slab_order - calculate size (page order) of slabs
1752 * @cachep: pointer to the cache that is being created
1753 * @size: size of objects to be created in this cache.
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001754 * @flags: slab allocation flags
1755 *
1756 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001757 *
1758 * This could be made much more intelligent. For now, try to avoid using
1759 * high order pages for slabs. When the gfp() functions are more friendly
1760 * towards high-order requests, this should be changed.
1761 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001762static size_t calculate_slab_order(struct kmem_cache *cachep,
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07001763 size_t size, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001764{
1765 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001766 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001767
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001768 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001769 unsigned int num;
1770 size_t remainder;
1771
Joonsoo Kim70f75062016-03-15 14:54:53 -07001772 num = cache_estimate(gfporder, size, flags, &remainder);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001773 if (!num)
1774 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001775
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001776 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1777 if (num > SLAB_OBJ_MAX_NUM)
1778 break;
1779
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001780 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001781 struct kmem_cache *freelist_cache;
1782 size_t freelist_size;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001783
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001784 freelist_size = num * sizeof(freelist_idx_t);
1785 freelist_cache = kmalloc_slab(freelist_size, 0u);
1786 if (!freelist_cache)
1787 continue;
1788
1789 /*
1790 * Needed to avoid possible looping condition
Joonsoo Kim76b342b2016-05-19 17:10:26 -07001791 * in cache_grow_begin()
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001792 */
1793 if (OFF_SLAB(freelist_cache))
1794 continue;
1795
1796 /* check if off slab has enough benefit */
1797 if (freelist_cache->size > cachep->size / 2)
1798 continue;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001799 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001800
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001801 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001802 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001803 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001804 left_over = remainder;
1805
1806 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001807 * A VFS-reclaimable slab tends to have most allocations
1808 * as GFP_NOFS and we really don't want to have to be allocating
1809 * higher-order pages when we are unable to shrink dcache.
1810 */
1811 if (flags & SLAB_RECLAIM_ACCOUNT)
1812 break;
1813
1814 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001815 * Large number of objects is good, but very large slabs are
1816 * currently bad for the gfp()s.
1817 */
David Rientjes543585c2011-10-18 22:09:24 -07001818 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001819 break;
1820
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001821 /*
1822 * Acceptable internal fragmentation?
1823 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001824 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001825 break;
1826 }
1827 return left_over;
1828}
1829
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001830static struct array_cache __percpu *alloc_kmem_cache_cpus(
1831 struct kmem_cache *cachep, int entries, int batchcount)
1832{
1833 int cpu;
1834 size_t size;
1835 struct array_cache __percpu *cpu_cache;
1836
1837 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001838 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001839
1840 if (!cpu_cache)
1841 return NULL;
1842
1843 for_each_possible_cpu(cpu) {
1844 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1845 entries, batchcount);
1846 }
1847
1848 return cpu_cache;
1849}
1850
Fabian Frederickbd721ea2016-08-02 14:03:33 -07001851static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001852{
Christoph Lameter97d06602012-07-06 15:25:11 -05001853 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001854 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001855
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001856 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1857 if (!cachep->cpu_cache)
1858 return 1;
1859
Christoph Lameter97d06602012-07-06 15:25:11 -05001860 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001861 /* Creation of first cache (kmem_cache). */
1862 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001863 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001864 /* For kmem_cache_node */
1865 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001866 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001867 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001868
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001869 for_each_online_node(node) {
1870 cachep->node[node] = kmalloc_node(
1871 sizeof(struct kmem_cache_node), gfp, node);
1872 BUG_ON(!cachep->node[node]);
1873 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001874 }
1875 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001876
Christoph Lameter6a673682013-01-10 19:14:19 +00001877 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001878 jiffies + REAPTIMEOUT_NODE +
1879 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001880
1881 cpu_cache_get(cachep)->avail = 0;
1882 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1883 cpu_cache_get(cachep)->batchcount = 1;
1884 cpu_cache_get(cachep)->touched = 0;
1885 cachep->batchcount = 1;
1886 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001887 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001888}
1889
Joonsoo Kim12220de2014-10-09 15:26:24 -07001890unsigned long kmem_cache_flags(unsigned long object_size,
1891 unsigned long flags, const char *name,
1892 void (*ctor)(void *))
1893{
1894 return flags;
1895}
1896
1897struct kmem_cache *
1898__kmem_cache_alias(const char *name, size_t size, size_t align,
1899 unsigned long flags, void (*ctor)(void *))
1900{
1901 struct kmem_cache *cachep;
1902
1903 cachep = find_mergeable(size, align, flags, name, ctor);
1904 if (cachep) {
1905 cachep->refcount++;
1906
1907 /*
1908 * Adjust the object sizes so that we clear
1909 * the complete object on kzalloc.
1910 */
1911 cachep->object_size = max_t(int, cachep->object_size, size);
1912 }
1913 return cachep;
1914}
1915
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001916static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1917 size_t size, unsigned long flags)
1918{
1919 size_t left;
1920
1921 cachep->num = 0;
1922
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08001923 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001924 return false;
1925
1926 left = calculate_slab_order(cachep, size,
1927 flags | CFLGS_OBJFREELIST_SLAB);
1928 if (!cachep->num)
1929 return false;
1930
1931 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1932 return false;
1933
1934 cachep->colour = left / cachep->colour_off;
1935
1936 return true;
1937}
1938
Joonsoo Kim158e3192016-03-15 14:54:35 -07001939static bool set_off_slab_cache(struct kmem_cache *cachep,
1940 size_t size, unsigned long flags)
1941{
1942 size_t left;
1943
1944 cachep->num = 0;
1945
1946 /*
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001947 * Always use on-slab management when SLAB_NOLEAKTRACE
1948 * to avoid recursive calls into kmemleak.
Joonsoo Kim158e3192016-03-15 14:54:35 -07001949 */
Joonsoo Kim158e3192016-03-15 14:54:35 -07001950 if (flags & SLAB_NOLEAKTRACE)
1951 return false;
1952
1953 /*
1954 * Size is large, assume best to place the slab management obj
1955 * off-slab (should allow better packing of objs).
1956 */
1957 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1958 if (!cachep->num)
1959 return false;
1960
1961 /*
1962 * If the slab has been placed off-slab, and we have enough space then
1963 * move it on-slab. This is at the expense of any extra colouring.
1964 */
1965 if (left >= cachep->num * sizeof(freelist_idx_t))
1966 return false;
1967
1968 cachep->colour = left / cachep->colour_off;
1969
1970 return true;
1971}
1972
1973static bool set_on_slab_cache(struct kmem_cache *cachep,
1974 size_t size, unsigned long flags)
1975{
1976 size_t left;
1977
1978 cachep->num = 0;
1979
1980 left = calculate_slab_order(cachep, size, flags);
1981 if (!cachep->num)
1982 return false;
1983
1984 cachep->colour = left / cachep->colour_off;
1985
1986 return true;
1987}
1988
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001989/**
Christoph Lameter039363f2012-07-06 15:25:10 -05001990 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08001991 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 *
1994 * Returns a ptr to the cache on success, NULL on failure.
1995 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09001996 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 * The flags are
1999 *
2000 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2001 * to catch references to uninitialised memory.
2002 *
2003 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2004 * for buffer overruns.
2005 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2007 * cacheline. This can be beneficial if you're counting cycles as closely
2008 * as davem.
2009 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002010int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002011__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002013 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002014 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002015 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002016 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019#if FORCED_DEBUG
2020 /*
2021 * Enable redzoning and last user accounting, except for caches with
2022 * large objects, if the increased size would increase the object size
2023 * above the next power of two: caches with object sizes just above a
2024 * power of two have a significant amount of internal fragmentation.
2025 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002026 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2027 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002028 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08002029 if (!(flags & SLAB_TYPESAFE_BY_RCU))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 flags |= SLAB_POISON;
2031#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033
Andrew Mortona737b3e2006-03-22 00:08:11 -08002034 /*
2035 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 * unaligned accesses for some archs when redzoning is used, and makes
2037 * sure any on-slab bufctl's are also correctly aligned.
2038 */
Canjiang Lue0771952017-07-06 15:36:37 -07002039 size = ALIGN(size, BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040
David Woodhouse87a927c2007-07-04 21:26:44 -04002041 if (flags & SLAB_RED_ZONE) {
2042 ralign = REDZONE_ALIGN;
2043 /* If redzoning, ensure that the second redzone is suitably
2044 * aligned, by adjusting the object size accordingly. */
Canjiang Lue0771952017-07-06 15:36:37 -07002045 size = ALIGN(size, REDZONE_ALIGN);
David Woodhouse87a927c2007-07-04 21:26:44 -04002046 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002047
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002048 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002049 if (ralign < cachep->align) {
2050 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002052 /* disable debug if necessary */
2053 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002054 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002055 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002056 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002058 cachep->align = ralign;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002059 cachep->colour_off = cache_line_size();
2060 /* Offset must be a multiple of the alignment. */
2061 if (cachep->colour_off < cachep->align)
2062 cachep->colour_off = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063
Pekka Enberg83b519e2009-06-10 19:40:04 +03002064 if (slab_is_available())
2065 gfp = GFP_KERNEL;
2066 else
2067 gfp = GFP_NOWAIT;
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070
Pekka Enbergca5f9702006-09-25 23:31:25 -07002071 /*
2072 * Both debugging options require word-alignment which is calculated
2073 * into align above.
2074 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002077 cachep->obj_offset += sizeof(unsigned long long);
2078 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 }
2080 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002081 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002082 * the real object. But if the second red zone needs to be
2083 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002085 if (flags & SLAB_RED_ZONE)
2086 size += REDZONE_ALIGN;
2087 else
2088 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 }
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002090#endif
2091
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002092 kasan_cache_create(cachep, &size, &flags);
2093
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002094 size = ALIGN(size, cachep->align);
2095 /*
2096 * We should restrict the number of objects in a slab to implement
2097 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2098 */
2099 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2100 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2101
2102#if DEBUG
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002103 /*
2104 * To activate debug pagealloc, off-slab management is necessary
2105 * requirement. In early phase of initialization, small sized slab
2106 * doesn't get initialized so it would not be possible. So, we need
2107 * to check size >= 256. It guarantees that all necessary small
2108 * sized slab is initialized in current slab initialization sequence.
2109 */
Joonsoo Kim40323272016-03-15 14:54:18 -07002110 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
Joonsoo Kimf3a3c322016-03-15 14:54:38 -07002111 size >= 256 && cachep->object_size > cache_line_size()) {
2112 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2113 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2114
2115 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2116 flags |= CFLGS_OFF_SLAB;
2117 cachep->obj_offset += tmp_size - size;
2118 size = tmp_size;
2119 goto done;
2120 }
2121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 }
2123#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002125 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2126 flags |= CFLGS_OBJFREELIST_SLAB;
2127 goto done;
2128 }
2129
Joonsoo Kim158e3192016-03-15 14:54:35 -07002130 if (set_off_slab_cache(cachep, size, flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 flags |= CFLGS_OFF_SLAB;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002132 goto done;
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002133 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134
Joonsoo Kim158e3192016-03-15 14:54:35 -07002135 if (set_on_slab_cache(cachep, size, flags))
2136 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137
Joonsoo Kim158e3192016-03-15 14:54:35 -07002138 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002139
Joonsoo Kim158e3192016-03-15 14:54:35 -07002140done:
2141 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002143 cachep->allocflags = __GFP_COMP;
Yang Shia3187e42016-05-19 17:10:41 -07002144 if (flags & SLAB_CACHE_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002145 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002146 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002147 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148
Joonsoo Kim40b44132016-03-15 14:54:21 -07002149#if DEBUG
2150 /*
2151 * If we're going to use the generic kernel_map_pages()
2152 * poisoning, then it's going to smash the contents of
2153 * the redzone and userword anyhow, so switch them off.
2154 */
2155 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2156 (cachep->flags & SLAB_POISON) &&
2157 is_debug_pagealloc_cache(cachep))
2158 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2159#endif
2160
2161 if (OFF_SLAB(cachep)) {
Joonsoo Kim158e3192016-03-15 14:54:35 -07002162 cachep->freelist_cache =
2163 kmalloc_slab(cachep->freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002164 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002166 err = setup_cpu_cache(cachep, gfp);
2167 if (err) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002168 __kmem_cache_release(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002169 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002172 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
2175#if DEBUG
2176static void check_irq_off(void)
2177{
2178 BUG_ON(!irqs_disabled());
2179}
2180
2181static void check_irq_on(void)
2182{
2183 BUG_ON(irqs_disabled());
2184}
2185
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002186static void check_mutex_acquired(void)
2187{
2188 BUG_ON(!mutex_is_locked(&slab_mutex));
2189}
2190
Pekka Enberg343e0d72006-02-01 03:05:50 -08002191static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192{
2193#ifdef CONFIG_SMP
2194 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002195 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196#endif
2197}
Christoph Lametere498be72005-09-09 13:03:32 -07002198
Pekka Enberg343e0d72006-02-01 03:05:50 -08002199static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002200{
2201#ifdef CONFIG_SMP
2202 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002203 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002204#endif
2205}
2206
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207#else
2208#define check_irq_off() do { } while(0)
2209#define check_irq_on() do { } while(0)
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002210#define check_mutex_acquired() do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002212#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213#endif
2214
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002215static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2216 int node, bool free_all, struct list_head *list)
2217{
2218 int tofree;
2219
2220 if (!ac || !ac->avail)
2221 return;
2222
2223 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2224 if (tofree > ac->avail)
2225 tofree = (ac->avail + 1) / 2;
2226
2227 free_block(cachep, ac->entry, tofree, node, list);
2228 ac->avail -= tofree;
2229 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2230}
Christoph Lameteraab22072006-03-22 00:09:06 -08002231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232static void do_drain(void *arg)
2233{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002234 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002236 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002237 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002238 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239
2240 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002241 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002242 n = get_node(cachep, node);
2243 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002244 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002245 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002246 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 ac->avail = 0;
2248}
2249
Pekka Enberg343e0d72006-02-01 03:05:50 -08002250static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002252 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002253 int node;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002254 LIST_HEAD(list);
Christoph Lametere498be72005-09-09 13:03:32 -07002255
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002256 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002258 for_each_kmem_cache_node(cachep, node, n)
2259 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002260 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002261
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002262 for_each_kmem_cache_node(cachep, node, n) {
2263 spin_lock_irq(&n->list_lock);
2264 drain_array_locked(cachep, n->shared, node, true, &list);
2265 spin_unlock_irq(&n->list_lock);
2266
2267 slabs_destroy(cachep, &list);
2268 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269}
2270
Christoph Lametered11d9e2006-06-30 01:55:45 -07002271/*
2272 * Remove slabs from the list of free slabs.
2273 * Specify the number of slabs to drain in tofree.
2274 *
2275 * Returns the actual number of slabs released.
2276 */
2277static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002278 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002280 struct list_head *p;
2281 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002282 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Christoph Lametered11d9e2006-06-30 01:55:45 -07002284 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002285 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002287 spin_lock_irq(&n->list_lock);
2288 p = n->slabs_free.prev;
2289 if (p == &n->slabs_free) {
2290 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002291 goto out;
2292 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
Joonsoo Kim8456a642013-10-24 10:07:49 +09002294 page = list_entry(p, struct page, lru);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002295 list_del(&page->lru);
Greg Thelenf728b0a2016-12-12 16:41:41 -08002296 n->free_slabs--;
David Rientjesbf00bd32016-12-12 16:41:44 -08002297 n->total_slabs--;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002298 /*
2299 * Safe to drop the lock. The slab is no longer linked
2300 * to the cache.
2301 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002302 n->free_objects -= cache->num;
2303 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002304 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002305 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002307out:
2308 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309}
2310
Tejun Heoc9fc5862017-02-22 15:41:27 -08002311int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002312{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002313 int ret = 0;
2314 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002315 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002316
2317 drain_cpu_caches(cachep);
2318
2319 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002320 for_each_kmem_cache_node(cachep, node, n) {
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07002321 drain_freelist(cachep, n, INT_MAX);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002322
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002323 ret += !list_empty(&n->slabs_full) ||
2324 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002325 }
2326 return (ret ? 1 : 0);
2327}
2328
Tejun Heoc9fc5862017-02-22 15:41:27 -08002329#ifdef CONFIG_MEMCG
2330void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2331{
2332 __kmem_cache_shrink(cachep);
2333}
2334#endif
2335
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002336int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337{
Tejun Heoc9fc5862017-02-22 15:41:27 -08002338 return __kmem_cache_shrink(cachep);
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002339}
2340
2341void __kmem_cache_release(struct kmem_cache *cachep)
2342{
Christoph Lameter12c36672012-09-04 23:38:33 +00002343 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002344 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002345
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002346 cache_random_seq_destroy(cachep);
2347
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002348 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002349
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002350 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002351 for_each_kmem_cache_node(cachep, i, n) {
2352 kfree(n->shared);
2353 free_alien_cache(n->alien);
2354 kfree(n);
2355 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002359/*
2360 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002361 *
2362 * For a slab cache when the slab descriptor is off-slab, the
2363 * slab descriptor can't come from the same cache which is being created,
2364 * Because if it is the case, that means we defer the creation of
2365 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2366 * And we eventually call down to __kmem_cache_create(), which
2367 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2368 * This is a "chicken-and-egg" problem.
2369 *
2370 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2371 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002372 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002373static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002374 struct page *page, int colour_off,
2375 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002377 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002378 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002379
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002380 page->s_mem = addr + colour_off;
2381 page->active = 0;
2382
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002383 if (OBJFREELIST_SLAB(cachep))
2384 freelist = NULL;
2385 else if (OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002387 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002388 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002389 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 return NULL;
2391 } else {
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002392 /* We will use last bytes at the slab for freelist */
2393 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2394 cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 }
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002396
Joonsoo Kim8456a642013-10-24 10:07:49 +09002397 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398}
2399
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002400static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002402 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002403}
2404
2405static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002406 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002407{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002408 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409}
2410
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002411static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412{
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002413#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 int i;
2415
2416 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002417 void *objp = index_to_obj(cachep, page, i);
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002418
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 if (cachep->flags & SLAB_STORE_USER)
2420 *dbg_userword(cachep, objp) = NULL;
2421
2422 if (cachep->flags & SLAB_RED_ZONE) {
2423 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2424 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2425 }
2426 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002427 * Constructors are not allowed to allocate memory from the same
2428 * cache which they are a constructor for. Otherwise, deadlock.
2429 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430 */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002431 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2432 kasan_unpoison_object_data(cachep,
2433 objp + obj_offset(cachep));
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002434 cachep->ctor(objp + obj_offset(cachep));
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002435 kasan_poison_object_data(
2436 cachep, objp + obj_offset(cachep));
2437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438
2439 if (cachep->flags & SLAB_RED_ZONE) {
2440 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07002441 slab_error(cachep, "constructor overwrote the end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07002443 slab_error(cachep, "constructor overwrote the start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 }
Joonsoo Kim40b44132016-03-15 14:54:21 -07002445 /* need to poison the objs? */
2446 if (cachep->flags & SLAB_POISON) {
2447 poison_obj(cachep, objp, POISON_FREE);
2448 slab_kernel_map(cachep, objp, 0, 0);
2449 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002450 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451#endif
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002452}
2453
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002454#ifdef CONFIG_SLAB_FREELIST_RANDOM
2455/* Hold information during a freelist initialization */
2456union freelist_init_state {
2457 struct {
2458 unsigned int pos;
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002459 unsigned int *list;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002460 unsigned int count;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002461 };
2462 struct rnd_state rnd_state;
2463};
2464
2465/*
2466 * Initialize the state based on the randomization methode available.
2467 * return true if the pre-computed list is available, false otherwize.
2468 */
2469static bool freelist_state_initialize(union freelist_init_state *state,
2470 struct kmem_cache *cachep,
2471 unsigned int count)
2472{
2473 bool ret;
2474 unsigned int rand;
2475
2476 /* Use best entropy available to define a random shift */
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002477 rand = get_random_int();
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002478
2479 /* Use a random state if the pre-computed list is not available */
2480 if (!cachep->random_seq) {
2481 prandom_seed_state(&state->rnd_state, rand);
2482 ret = false;
2483 } else {
2484 state->list = cachep->random_seq;
2485 state->count = count;
John Sperbeckc4e490c2017-01-10 16:58:24 -08002486 state->pos = rand % count;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002487 ret = true;
2488 }
2489 return ret;
2490}
2491
2492/* Get the next entry on the list and randomize it using a random shift */
2493static freelist_idx_t next_random_slot(union freelist_init_state *state)
2494{
John Sperbeckc4e490c2017-01-10 16:58:24 -08002495 if (state->pos >= state->count)
2496 state->pos = 0;
2497 return state->list[state->pos++];
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002498}
2499
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002500/* Swap two freelist entries */
2501static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2502{
2503 swap(((freelist_idx_t *)page->freelist)[a],
2504 ((freelist_idx_t *)page->freelist)[b]);
2505}
2506
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002507/*
2508 * Shuffle the freelist initialization state based on pre-computed lists.
2509 * return true if the list was successfully shuffled, false otherwise.
2510 */
2511static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2512{
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002513 unsigned int objfreelist = 0, i, rand, count = cachep->num;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002514 union freelist_init_state state;
2515 bool precomputed;
2516
2517 if (count < 2)
2518 return false;
2519
2520 precomputed = freelist_state_initialize(&state, cachep, count);
2521
2522 /* Take a random entry as the objfreelist */
2523 if (OBJFREELIST_SLAB(cachep)) {
2524 if (!precomputed)
2525 objfreelist = count - 1;
2526 else
2527 objfreelist = next_random_slot(&state);
2528 page->freelist = index_to_obj(cachep, page, objfreelist) +
2529 obj_offset(cachep);
2530 count--;
2531 }
2532
2533 /*
2534 * On early boot, generate the list dynamically.
2535 * Later use a pre-computed list for speed.
2536 */
2537 if (!precomputed) {
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002538 for (i = 0; i < count; i++)
2539 set_free_obj(page, i, i);
2540
2541 /* Fisher-Yates shuffle */
2542 for (i = count - 1; i > 0; i--) {
2543 rand = prandom_u32_state(&state.rnd_state);
2544 rand %= (i + 1);
2545 swap_free_obj(page, i, rand);
2546 }
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002547 } else {
2548 for (i = 0; i < count; i++)
2549 set_free_obj(page, i, next_random_slot(&state));
2550 }
2551
2552 if (OBJFREELIST_SLAB(cachep))
2553 set_free_obj(page, cachep->num - 1, objfreelist);
2554
2555 return true;
2556}
2557#else
2558static inline bool shuffle_freelist(struct kmem_cache *cachep,
2559 struct page *page)
2560{
2561 return false;
2562}
2563#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2564
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002565static void cache_init_objs(struct kmem_cache *cachep,
2566 struct page *page)
2567{
2568 int i;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002569 void *objp;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002570 bool shuffled;
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002571
2572 cache_init_objs_debug(cachep, page);
2573
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002574 /* Try to randomize the freelist if enabled */
2575 shuffled = shuffle_freelist(cachep, page);
2576
2577 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002578 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2579 obj_offset(cachep);
2580 }
2581
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002582 for (i = 0; i < cachep->num; i++) {
Andrey Ryabininb3cbd9b2016-08-02 14:02:52 -07002583 objp = index_to_obj(cachep, page, i);
2584 kasan_init_slab_obj(cachep, objp);
2585
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002586 /* constructor could break poison info */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002587 if (DEBUG == 0 && cachep->ctor) {
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002588 kasan_unpoison_object_data(cachep, objp);
2589 cachep->ctor(objp);
2590 kasan_poison_object_data(cachep, objp);
2591 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002592
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002593 if (!shuffled)
2594 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596}
2597
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002598static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002599{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002600 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002601
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002602 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002603 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002604
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002605#if DEBUG
2606 if (cachep->flags & SLAB_STORE_USER)
2607 set_store_user_dirty(cachep);
2608#endif
2609
Matthew Dobson78d382d2006-02-01 03:05:47 -08002610 return objp;
2611}
2612
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002613static void slab_put_obj(struct kmem_cache *cachep,
2614 struct page *page, void *objp)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002615{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002616 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002617#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002618 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002619
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002620 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002621 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002622 if (get_free_obj(page, i) == objnr) {
Joe Perches11705322016-03-17 14:19:50 -07002623 pr_err("slab: double free detected in cache '%s', objp %p\n",
Joe Perches756a0252016-03-17 14:19:47 -07002624 cachep->name, objp);
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002625 BUG();
2626 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002627 }
2628#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002629 page->active--;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002630 if (!page->freelist)
2631 page->freelist = objp + obj_offset(cachep);
2632
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002633 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002634}
2635
Pekka Enberg47768742006-06-23 02:03:07 -07002636/*
2637 * Map pages beginning at addr to the given cache and slab. This is required
2638 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002639 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002640 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002641static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002642 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002644 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002645 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646}
2647
2648/*
2649 * Grow (by 1) the number of slabs within a cache. This is called by
2650 * kmem_cache_alloc() when there are no active objs left in a cache.
2651 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002652static struct page *cache_grow_begin(struct kmem_cache *cachep,
2653 gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002655 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002656 size_t offset;
2657 gfp_t local_flags;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002658 int page_node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002659 struct kmem_cache_node *n;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002660 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661
Andrew Mortona737b3e2006-03-22 00:08:11 -08002662 /*
2663 * Be lazy and only check for valid flags here, keeping it out of the
2664 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002666 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
Michal Hockobacdcb32016-07-26 15:22:02 -07002667 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
Michal Hocko72baeef0c2016-07-26 15:22:05 -07002668 flags &= ~GFP_SLAB_BUG_MASK;
2669 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2670 invalid_mask, &invalid_mask, flags, &flags);
2671 dump_stack();
Andrew Mortonc871ac42014-12-10 15:42:25 -08002672 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002673 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 check_irq_off();
Mel Gormand0164ad2015-11-06 16:28:21 -08002676 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 local_irq_enable();
2678
2679 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002680 * Get mem for the objs. Attempt to allocate a physical page from
2681 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002682 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002683 page = kmem_getpages(cachep, local_flags, nodeid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002684 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 goto failed;
2686
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002687 page_node = page_to_nid(page);
2688 n = get_node(cachep, page_node);
Joonsoo Kim03d1d432016-05-19 17:10:20 -07002689
2690 /* Get colour for the slab, and cal the next value. */
2691 n->colour_next++;
2692 if (n->colour_next >= cachep->colour)
2693 n->colour_next = 0;
2694
2695 offset = n->colour_next;
2696 if (offset >= cachep->colour)
2697 offset = 0;
2698
2699 offset *= cachep->colour_off;
2700
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002702 freelist = alloc_slabmgmt(cachep, page, offset,
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002703 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002704 if (OFF_SLAB(cachep) && !freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 goto opps1;
2706
Joonsoo Kim8456a642013-10-24 10:07:49 +09002707 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002709 kasan_poison_slab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002710 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
Mel Gormand0164ad2015-11-06 16:28:21 -08002712 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 local_irq_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002715 return page;
2716
Andrew Mortona737b3e2006-03-22 00:08:11 -08002717opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002718 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002719failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002720 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 local_irq_disable();
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002722 return NULL;
2723}
2724
2725static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2726{
2727 struct kmem_cache_node *n;
2728 void *list = NULL;
2729
2730 check_irq_off();
2731
2732 if (!page)
2733 return;
2734
2735 INIT_LIST_HEAD(&page->lru);
2736 n = get_node(cachep, page_to_nid(page));
2737
2738 spin_lock(&n->list_lock);
David Rientjesbf00bd32016-12-12 16:41:44 -08002739 n->total_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002740 if (!page->active) {
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002741 list_add_tail(&page->lru, &(n->slabs_free));
Greg Thelenf728b0a2016-12-12 16:41:41 -08002742 n->free_slabs++;
David Rientjesbf00bd32016-12-12 16:41:44 -08002743 } else
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002744 fixup_slab_list(cachep, n, page, &list);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07002745
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002746 STATS_INC_GROWN(cachep);
2747 n->free_objects += cachep->num - page->active;
2748 spin_unlock(&n->list_lock);
2749
2750 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751}
2752
2753#if DEBUG
2754
2755/*
2756 * Perform extra freeing checks:
2757 * - detect bad pointers.
2758 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 */
2760static void kfree_debugcheck(const void *objp)
2761{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 if (!virt_addr_valid(objp)) {
Joe Perches11705322016-03-17 14:19:50 -07002763 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002764 (unsigned long)objp);
2765 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767}
2768
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002769static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2770{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002771 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002772
2773 redzone1 = *dbg_redzone1(cache, obj);
2774 redzone2 = *dbg_redzone2(cache, obj);
2775
2776 /*
2777 * Redzone is ok.
2778 */
2779 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2780 return;
2781
2782 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2783 slab_error(cache, "double free detected");
2784 else
2785 slab_error(cache, "memory outside object was overwritten");
2786
Joe Perches11705322016-03-17 14:19:50 -07002787 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2788 obj, redzone1, redzone2);
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002789}
2790
Pekka Enberg343e0d72006-02-01 03:05:50 -08002791static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002792 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002795 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002797 BUG_ON(virt_to_cache(objp) != cachep);
2798
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002799 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002801 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002804 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2806 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2807 }
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002808 if (cachep->flags & SLAB_STORE_USER) {
2809 set_store_user_dirty(cachep);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002810 *dbg_userword(cachep, objp) = (void *)caller;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002811 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812
Joonsoo Kim8456a642013-10-24 10:07:49 +09002813 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002816 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 poison_obj(cachep, objp, POISON_FREE);
Joonsoo Kim40b44132016-03-15 14:54:21 -07002820 slab_kernel_map(cachep, objp, 0, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 }
2822 return objp;
2823}
2824
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825#else
2826#define kfree_debugcheck(x) do { } while(0)
2827#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828#endif
2829
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002830static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2831 void **list)
2832{
2833#if DEBUG
2834 void *next = *list;
2835 void *objp;
2836
2837 while (next) {
2838 objp = next - obj_offset(cachep);
2839 next = *(void **)next;
2840 poison_obj(cachep, objp, POISON_FREE);
2841 }
2842#endif
2843}
2844
Joonsoo Kimd8410232016-03-15 14:54:44 -07002845static inline void fixup_slab_list(struct kmem_cache *cachep,
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002846 struct kmem_cache_node *n, struct page *page,
2847 void **list)
Joonsoo Kimd8410232016-03-15 14:54:44 -07002848{
2849 /* move slabp to correct slabp list: */
2850 list_del(&page->lru);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002851 if (page->active == cachep->num) {
Joonsoo Kimd8410232016-03-15 14:54:44 -07002852 list_add(&page->lru, &n->slabs_full);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002853 if (OBJFREELIST_SLAB(cachep)) {
2854#if DEBUG
2855 /* Poisoning will be done without holding the lock */
2856 if (cachep->flags & SLAB_POISON) {
2857 void **objp = page->freelist;
2858
2859 *objp = *list;
2860 *list = objp;
2861 }
2862#endif
2863 page->freelist = NULL;
2864 }
2865 } else
Joonsoo Kimd8410232016-03-15 14:54:44 -07002866 list_add(&page->lru, &n->slabs_partial);
2867}
2868
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002869/* Try to find non-pfmemalloc slab if needed */
2870static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
David Rientjesbf00bd32016-12-12 16:41:44 -08002871 struct page *page, bool pfmemalloc)
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002872{
2873 if (!page)
2874 return NULL;
2875
2876 if (pfmemalloc)
2877 return page;
2878
2879 if (!PageSlabPfmemalloc(page))
2880 return page;
2881
2882 /* No need to keep pfmemalloc slab if we have enough free objects */
2883 if (n->free_objects > n->free_limit) {
2884 ClearPageSlabPfmemalloc(page);
2885 return page;
2886 }
2887
2888 /* Move pfmemalloc slab to the end of list to speed up next search */
2889 list_del(&page->lru);
David Rientjesbf00bd32016-12-12 16:41:44 -08002890 if (!page->active) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002891 list_add_tail(&page->lru, &n->slabs_free);
David Rientjesbf00bd32016-12-12 16:41:44 -08002892 n->free_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002893 } else
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002894 list_add_tail(&page->lru, &n->slabs_partial);
2895
2896 list_for_each_entry(page, &n->slabs_partial, lru) {
2897 if (!PageSlabPfmemalloc(page))
2898 return page;
2899 }
2900
Greg Thelenf728b0a2016-12-12 16:41:41 -08002901 n->free_touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002902 list_for_each_entry(page, &n->slabs_free, lru) {
Greg Thelenf728b0a2016-12-12 16:41:41 -08002903 if (!PageSlabPfmemalloc(page)) {
David Rientjesbf00bd32016-12-12 16:41:44 -08002904 n->free_slabs--;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002905 return page;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002906 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002907 }
2908
2909 return NULL;
2910}
2911
2912static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
Geliang Tang7aa0d222016-01-14 15:18:02 -08002913{
2914 struct page *page;
2915
Greg Thelenf728b0a2016-12-12 16:41:41 -08002916 assert_spin_locked(&n->list_lock);
David Rientjesbf00bd32016-12-12 16:41:44 -08002917 page = list_first_entry_or_null(&n->slabs_partial, struct page, lru);
Geliang Tang7aa0d222016-01-14 15:18:02 -08002918 if (!page) {
2919 n->free_touched = 1;
David Rientjesbf00bd32016-12-12 16:41:44 -08002920 page = list_first_entry_or_null(&n->slabs_free, struct page,
2921 lru);
Greg Thelenf728b0a2016-12-12 16:41:41 -08002922 if (page)
David Rientjesbf00bd32016-12-12 16:41:44 -08002923 n->free_slabs--;
Geliang Tang7aa0d222016-01-14 15:18:02 -08002924 }
2925
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002926 if (sk_memalloc_socks())
David Rientjesbf00bd32016-12-12 16:41:44 -08002927 page = get_valid_first_slab(n, page, pfmemalloc);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002928
Geliang Tang7aa0d222016-01-14 15:18:02 -08002929 return page;
2930}
2931
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002932static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2933 struct kmem_cache_node *n, gfp_t flags)
2934{
2935 struct page *page;
2936 void *obj;
2937 void *list = NULL;
2938
2939 if (!gfp_pfmemalloc_allowed(flags))
2940 return NULL;
2941
2942 spin_lock(&n->list_lock);
2943 page = get_first_slab(n, true);
2944 if (!page) {
2945 spin_unlock(&n->list_lock);
2946 return NULL;
2947 }
2948
2949 obj = slab_get_obj(cachep, page);
2950 n->free_objects--;
2951
2952 fixup_slab_list(cachep, n, page, &list);
2953
2954 spin_unlock(&n->list_lock);
2955 fixup_objfreelist_debug(cachep, &list);
2956
2957 return obj;
2958}
2959
Joonsoo Kim213b4692016-05-19 17:10:29 -07002960/*
2961 * Slab list should be fixed up by fixup_slab_list() for existing slab
2962 * or cache_grow_end() for new slab
2963 */
2964static __always_inline int alloc_block(struct kmem_cache *cachep,
2965 struct array_cache *ac, struct page *page, int batchcount)
2966{
2967 /*
2968 * There must be at least one object available for
2969 * allocation.
2970 */
2971 BUG_ON(page->active >= cachep->num);
2972
2973 while (page->active < cachep->num && batchcount--) {
2974 STATS_INC_ALLOCED(cachep);
2975 STATS_INC_ACTIVE(cachep);
2976 STATS_SET_HIGH(cachep);
2977
2978 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2979 }
2980
2981 return batchcount;
2982}
2983
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002984static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985{
2986 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002987 struct kmem_cache_node *n;
Joonsoo Kim801faf02016-05-19 17:10:31 -07002988 struct array_cache *ac, *shared;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002989 int node;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002990 void *list = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002991 struct page *page;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002992
Joe Korty6d2144d2008-03-05 15:04:59 -08002993 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002994 node = numa_mem_id();
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002995
Joe Korty6d2144d2008-03-05 15:04:59 -08002996 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997 batchcount = ac->batchcount;
2998 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002999 /*
3000 * If there was little recent activity on this cache, then
3001 * perform only a partial refill. Otherwise we could generate
3002 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 */
3004 batchcount = BATCHREFILL_LIMIT;
3005 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07003006 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003008 BUG_ON(ac->avail > 0 || !n);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003009 shared = READ_ONCE(n->shared);
3010 if (!n->free_objects && (!shared || !shared->avail))
3011 goto direct_grow;
3012
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003013 spin_lock(&n->list_lock);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003014 shared = READ_ONCE(n->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003015
Christoph Lameter3ded1752006-03-25 03:06:44 -08003016 /* See if we can refill from the shared array */
Joonsoo Kim801faf02016-05-19 17:10:31 -07003017 if (shared && transfer_objects(ac, shared, batchcount)) {
3018 shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003019 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003020 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003021
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 while (batchcount > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 /* Get slab alloc is to come from. */
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003024 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003025 if (!page)
3026 goto must_grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003029
Joonsoo Kim213b4692016-05-19 17:10:29 -07003030 batchcount = alloc_block(cachep, ac, page, batchcount);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003031 fixup_slab_list(cachep, n, page, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032 }
3033
Andrew Mortona737b3e2006-03-22 00:08:11 -08003034must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003035 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003036alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003037 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003038 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039
Joonsoo Kim801faf02016-05-19 17:10:31 -07003040direct_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 if (unlikely(!ac->avail)) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003042 /* Check if we can use obj in pfmemalloc slab */
3043 if (sk_memalloc_socks()) {
3044 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3045
3046 if (obj)
3047 return obj;
3048 }
3049
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003050 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
Christoph Lametere498be72005-09-09 13:03:32 -07003051
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003052 /*
3053 * cache_grow_begin() can reenable interrupts,
3054 * then ac could change.
3055 */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003056 ac = cpu_cache_get(cachep);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003057 if (!ac->avail && page)
3058 alloc_block(cachep, ac, page, batchcount);
3059 cache_grow_end(cachep, page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003060
Joonsoo Kim213b4692016-05-19 17:10:29 -07003061 if (!ac->avail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 }
3064 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003065
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003066 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067}
3068
Andrew Mortona737b3e2006-03-22 00:08:11 -08003069static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3070 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071{
Mel Gormand0164ad2015-11-06 16:28:21 -08003072 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073}
3074
3075#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003076static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003077 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003079 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003081 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07003083 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 poison_obj(cachep, objp, POISON_INUSE);
3085 }
3086 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003087 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088
3089 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3091 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
Joe Perches756a0252016-03-17 14:19:47 -07003092 slab_error(cachep, "double free, or memory outside object was overwritten");
Joe Perches11705322016-03-17 14:19:50 -07003093 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3094 objp, *dbg_redzone1(cachep, objp),
3095 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 }
3097 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3098 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3099 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003100
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003101 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003102 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003103 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003104 if (ARCH_SLAB_MINALIGN &&
3105 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Joe Perches11705322016-03-17 14:19:50 -07003106 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003107 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003108 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109 return objp;
3110}
3111#else
3112#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3113#endif
3114
Pekka Enberg343e0d72006-02-01 03:05:50 -08003115static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003117 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 struct array_cache *ac;
3119
Alok N Kataria5c382302005-09-27 21:45:46 -07003120 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003121
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003122 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124 ac->touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003125 objp = ac->entry[--ac->avail];
Mel Gorman072bb0a2012-07-31 16:43:58 -07003126
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003127 STATS_INC_ALLOCHIT(cachep);
3128 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003130
3131 STATS_INC_ALLOCMISS(cachep);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003132 objp = cache_alloc_refill(cachep, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003133 /*
3134 * the 'ac' may be updated by cache_alloc_refill(),
3135 * and kmemleak_erase() requires its correct value.
3136 */
3137 ac = cpu_cache_get(cachep);
3138
3139out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003140 /*
3141 * To avoid a false negative, if an object that is in one of the
3142 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3143 * treat the array pointers as a reference to the object.
3144 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003145 if (objp)
3146 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003147 return objp;
3148}
3149
Christoph Lametere498be72005-09-09 13:03:32 -07003150#ifdef CONFIG_NUMA
3151/*
Zefan Li2ad654b2014-09-25 09:41:02 +08003152 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003153 *
3154 * If we are in_interrupt, then process context, including cpusets and
3155 * mempolicy, may not apply and should not be used for allocation policy.
3156 */
3157static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3158{
3159 int nid_alloc, nid_here;
3160
Christoph Lameter765c4502006-09-27 01:50:08 -07003161 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003162 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003163 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003164 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003165 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003166 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003167 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003168 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003169 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003170 return NULL;
3171}
3172
3173/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003174 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003175 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003176 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003177 * perform an allocation without specifying a node. This allows the page
3178 * allocator to do its reclaim / fallback magic. We then insert the
3179 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003180 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003181static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003182{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003183 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07003184 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003185 struct zone *zone;
3186 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003187 void *obj = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003188 struct page *page;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003189 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003190 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003191
3192 if (flags & __GFP_THISNODE)
3193 return NULL;
3194
Mel Gormancc9a6c82012-03-21 16:34:11 -07003195retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003196 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003197 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003198
Christoph Lameter3c517a62006-12-06 20:33:29 -08003199retry:
3200 /*
3201 * Look through allowed nodes for objects available
3202 * from existing per node queues.
3203 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003204 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3205 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003206
Vladimir Davydov061d7072014-12-12 16:58:25 -08003207 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003208 get_node(cache, nid) &&
3209 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003210 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003211 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003212 if (obj)
3213 break;
3214 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003215 }
3216
Christoph Lametercfce6602007-05-06 14:50:17 -07003217 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 /*
3219 * This allocation will be performed within the constraints
3220 * of the current cpuset / memory policy requirements.
3221 * We may trigger various forms of reclaim on the allowed
3222 * set and go into memory reserves if necessary.
3223 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003224 page = cache_grow_begin(cache, flags, numa_mem_id());
3225 cache_grow_end(cache, page);
3226 if (page) {
3227 nid = page_to_nid(page);
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003228 obj = ____cache_alloc_node(cache,
3229 gfp_exact_node(flags), nid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003230
Christoph Lameter3c517a62006-12-06 20:33:29 -08003231 /*
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003232 * Another processor may allocate the objects in
3233 * the slab since we are not holding any locks.
Christoph Lameter3c517a62006-12-06 20:33:29 -08003234 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003235 if (!obj)
3236 goto retry;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003237 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003238 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003239
Mel Gormand26914d2014-04-03 14:47:24 -07003240 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003241 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003242 return obj;
3243}
3244
3245/*
Christoph Lametere498be72005-09-09 13:03:32 -07003246 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003248static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003249 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003250{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003251 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003252 struct kmem_cache_node *n;
Joonsoo Kim213b4692016-05-19 17:10:29 -07003253 void *obj = NULL;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003254 void *list = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003256 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003257 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003258 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003259
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003260 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003261 spin_lock(&n->list_lock);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003262 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003263 if (!page)
3264 goto must_grow;
Christoph Lametere498be72005-09-09 13:03:32 -07003265
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003266 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003267
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003268 STATS_INC_NODEALLOCS(cachep);
3269 STATS_INC_ACTIVE(cachep);
3270 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003271
Joonsoo Kim8456a642013-10-24 10:07:49 +09003272 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003274 obj = slab_get_obj(cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003275 n->free_objects--;
Christoph Lametere498be72005-09-09 13:03:32 -07003276
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003277 fixup_slab_list(cachep, n, page, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003278
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003279 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003280 fixup_objfreelist_debug(cachep, &list);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003281 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003282
Andrew Mortona737b3e2006-03-22 00:08:11 -08003283must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003284 spin_unlock(&n->list_lock);
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003285 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003286 if (page) {
3287 /* This slab isn't counted yet so don't update free_objects */
3288 obj = slab_get_obj(cachep, page);
3289 }
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003290 cache_grow_end(cachep, page);
Christoph Lametere498be72005-09-09 13:03:32 -07003291
Joonsoo Kim213b4692016-05-19 17:10:29 -07003292 return obj ? obj : fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003293}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003294
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003295static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003296slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003297 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003298{
3299 unsigned long save_flags;
3300 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003301 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003302
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003303 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003304 cachep = slab_pre_alloc_hook(cachep, flags);
3305 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003306 return NULL;
3307
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003308 cache_alloc_debugcheck_before(cachep, flags);
3309 local_irq_save(save_flags);
3310
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003311 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003312 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003313
Christoph Lameter18bf8542014-08-06 16:04:11 -07003314 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003315 /* Node not bootstrapped yet */
3316 ptr = fallback_alloc(cachep, flags);
3317 goto out;
3318 }
3319
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003320 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003321 /*
3322 * Use the locally cached objects if possible.
3323 * However ____cache_alloc does not allow fallback
3324 * to other nodes. It may fail while we still have
3325 * objects on other nodes available.
3326 */
3327 ptr = ____cache_alloc(cachep, flags);
3328 if (ptr)
3329 goto out;
3330 }
3331 /* ___cache_alloc_node can fall back to other nodes */
3332 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3333 out:
3334 local_irq_restore(save_flags);
3335 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3336
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003337 if (unlikely(flags & __GFP_ZERO) && ptr)
3338 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003339
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003340 slab_post_alloc_hook(cachep, flags, 1, &ptr);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003341 return ptr;
3342}
3343
3344static __always_inline void *
3345__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3346{
3347 void *objp;
3348
Zefan Li2ad654b2014-09-25 09:41:02 +08003349 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003350 objp = alternate_node_alloc(cache, flags);
3351 if (objp)
3352 goto out;
3353 }
3354 objp = ____cache_alloc(cache, flags);
3355
3356 /*
3357 * We may just have run out of memory on the local node.
3358 * ____cache_alloc_node() knows how to locate memory on other nodes
3359 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003360 if (!objp)
3361 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003362
3363 out:
3364 return objp;
3365}
3366#else
3367
3368static __always_inline void *
3369__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3370{
3371 return ____cache_alloc(cachep, flags);
3372}
3373
3374#endif /* CONFIG_NUMA */
3375
3376static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003377slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003378{
3379 unsigned long save_flags;
3380 void *objp;
3381
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003382 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003383 cachep = slab_pre_alloc_hook(cachep, flags);
3384 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003385 return NULL;
3386
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003387 cache_alloc_debugcheck_before(cachep, flags);
3388 local_irq_save(save_flags);
3389 objp = __do_cache_alloc(cachep, flags);
3390 local_irq_restore(save_flags);
3391 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3392 prefetchw(objp);
3393
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003394 if (unlikely(flags & __GFP_ZERO) && objp)
3395 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003396
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003397 slab_post_alloc_hook(cachep, flags, 1, &objp);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003398 return objp;
3399}
Christoph Lametere498be72005-09-09 13:03:32 -07003400
3401/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003402 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003403 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003404 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003405static void free_block(struct kmem_cache *cachep, void **objpp,
3406 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407{
3408 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003409 struct kmem_cache_node *n = get_node(cachep, node);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003410 struct page *page;
3411
3412 n->free_objects += nr_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413
3414 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003415 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003416 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417
Mel Gorman072bb0a2012-07-31 16:43:58 -07003418 objp = objpp[i];
3419
Joonsoo Kim8456a642013-10-24 10:07:49 +09003420 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003421 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003422 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003423 slab_put_obj(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 STATS_DEC_ACTIVE(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
3426 /* fixup slab chains */
Greg Thelenf728b0a2016-12-12 16:41:41 -08003427 if (page->active == 0) {
Joonsoo Kim6052b782016-05-19 17:10:17 -07003428 list_add(&page->lru, &n->slabs_free);
Greg Thelenf728b0a2016-12-12 16:41:41 -08003429 n->free_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08003430 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 /* Unconditionally move a slab to the end of the
3432 * partial list on free - maximum time for the
3433 * other objects to be freed, too.
3434 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003435 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436 }
3437 }
Joonsoo Kim6052b782016-05-19 17:10:17 -07003438
3439 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3440 n->free_objects -= cachep->num;
3441
3442 page = list_last_entry(&n->slabs_free, struct page, lru);
Wei Yongjunde24bae2016-07-26 15:22:11 -07003443 list_move(&page->lru, list);
Greg Thelenf728b0a2016-12-12 16:41:41 -08003444 n->free_slabs--;
David Rientjesbf00bd32016-12-12 16:41:44 -08003445 n->total_slabs--;
Joonsoo Kim6052b782016-05-19 17:10:17 -07003446 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447}
3448
Pekka Enberg343e0d72006-02-01 03:05:50 -08003449static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450{
3451 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003452 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003453 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003454 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455
3456 batchcount = ac->batchcount;
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003457
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003459 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003460 spin_lock(&n->list_lock);
3461 if (n->shared) {
3462 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003463 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 if (max) {
3465 if (batchcount > max)
3466 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003467 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003468 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 shared_array->avail += batchcount;
3470 goto free_done;
3471 }
3472 }
3473
Joonsoo Kim97654df2014-08-06 16:04:25 -07003474 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003475free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476#if STATS
3477 {
3478 int i = 0;
Geliang Tang73c02192016-01-14 15:17:59 -08003479 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480
Geliang Tang73c02192016-01-14 15:17:59 -08003481 list_for_each_entry(page, &n->slabs_free, lru) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003482 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
3484 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 }
3486 STATS_SET_FREEABLE(cachep, i);
3487 }
3488#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003489 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003490 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003492 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493}
3494
3495/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003496 * Release an obj back to its cache. If the obj has a constructed state, it must
3497 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003499static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003500 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501{
Alexander Potapenko55834c52016-05-20 16:59:11 -07003502 /* Put the object into the quarantine, don't touch it for now. */
3503 if (kasan_slab_free(cachep, objp))
3504 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505
Alexander Potapenko55834c52016-05-20 16:59:11 -07003506 ___cache_free(cachep, objp, caller);
3507}
3508
3509void ___cache_free(struct kmem_cache *cachep, void *objp,
3510 unsigned long caller)
3511{
3512 struct array_cache *ac = cpu_cache_get(cachep);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003513
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003515 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003516 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003518 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003519
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003520 /*
3521 * Skip calling cache_free_alien() when the platform is not numa.
3522 * This will avoid cache misses that happen while accessing slabp (which
3523 * is per page memory reference) to get nodeid. Instead use a global
3524 * variable to skip the call, which is mostly likely to be present in
3525 * the cache.
3526 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003527 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003528 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003529
Joonsoo Kim3d880192014-10-09 15:26:04 -07003530 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 } else {
3533 STATS_INC_FREEMISS(cachep);
3534 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003536
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003537 if (sk_memalloc_socks()) {
3538 struct page *page = virt_to_head_page(objp);
3539
3540 if (unlikely(PageSlabPfmemalloc(page))) {
3541 cache_free_pfmemalloc(cachep, page, objp);
3542 return;
3543 }
3544 }
3545
3546 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547}
3548
3549/**
3550 * kmem_cache_alloc - Allocate an object
3551 * @cachep: The cache to allocate from.
3552 * @flags: See kmalloc().
3553 *
3554 * Allocate an object from this cache. The flags are only relevant
3555 * if the cache has no available objects.
3556 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003557void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003559 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003560
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003561 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003562 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003563 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003564
3565 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566}
3567EXPORT_SYMBOL(kmem_cache_alloc);
3568
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003569static __always_inline void
3570cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3571 size_t size, void **p, unsigned long caller)
3572{
3573 size_t i;
3574
3575 for (i = 0; i < size; i++)
3576 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3577}
3578
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003579int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003580 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07003581{
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003582 size_t i;
3583
3584 s = slab_pre_alloc_hook(s, flags);
3585 if (!s)
3586 return 0;
3587
3588 cache_alloc_debugcheck_before(s, flags);
3589
3590 local_irq_disable();
3591 for (i = 0; i < size; i++) {
3592 void *objp = __do_cache_alloc(s, flags);
3593
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003594 if (unlikely(!objp))
3595 goto error;
3596 p[i] = objp;
3597 }
3598 local_irq_enable();
3599
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003600 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3601
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003602 /* Clear memory outside IRQ disabled section */
3603 if (unlikely(flags & __GFP_ZERO))
3604 for (i = 0; i < size; i++)
3605 memset(p[i], 0, s->object_size);
3606
3607 slab_post_alloc_hook(s, flags, size, p);
3608 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3609 return size;
3610error:
3611 local_irq_enable();
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003612 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003613 slab_post_alloc_hook(s, flags, i, p);
3614 __kmem_cache_free_bulk(s, i, p);
3615 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07003616}
3617EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3618
Li Zefan0f24f122009-12-11 15:45:30 +08003619#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003620void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003621kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003622{
Steven Rostedt85beb582010-11-24 16:23:34 -05003623 void *ret;
3624
Ezequiel Garcia48356302012-09-08 17:47:57 -03003625 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003626
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003627 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003628 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003629 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003630 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003631}
Steven Rostedt85beb582010-11-24 16:23:34 -05003632EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003633#endif
3634
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003636/**
3637 * kmem_cache_alloc_node - Allocate an object on the specified node
3638 * @cachep: The cache to allocate from.
3639 * @flags: See kmalloc().
3640 * @nodeid: node number of the target node.
3641 *
3642 * Identical to kmem_cache_alloc but it will allocate memory on the given
3643 * node, which can improve the performance for cpu bound structures.
3644 *
3645 * Fallback to other node is possible if __GFP_THISNODE is not set.
3646 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003647void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3648{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003649 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003650
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003651 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003652 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003653 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003654 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003655
3656 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003657}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658EXPORT_SYMBOL(kmem_cache_alloc_node);
3659
Li Zefan0f24f122009-12-11 15:45:30 +08003660#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003661void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003662 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003663 int nodeid,
3664 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003665{
Steven Rostedt85beb582010-11-24 16:23:34 -05003666 void *ret;
3667
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003668 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003669
3670 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003671 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003672 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003673 flags, nodeid);
3674 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003675}
Steven Rostedt85beb582010-11-24 16:23:34 -05003676EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003677#endif
3678
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003679static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003680__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003681{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003682 struct kmem_cache *cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003683 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003684
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003685 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003686 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3687 return cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003688 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003689 kasan_kmalloc(cachep, ret, size, flags);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003690
3691 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003692}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003693
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003694void *__kmalloc_node(size_t size, gfp_t flags, int node)
3695{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003696 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003697}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003698EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003699
3700void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003701 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003702{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003703 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003704}
3705EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003706#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707
3708/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003709 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003711 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003712 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003714static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003715 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003717 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003718 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003720 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003721 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3722 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003723 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003724
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003725 kasan_kmalloc(cachep, ret, size, flags);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003726 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003727 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003728
3729 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003730}
3731
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003732void *__kmalloc(size_t size, gfp_t flags)
3733{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003734 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735}
3736EXPORT_SYMBOL(__kmalloc);
3737
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003738void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003739{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003740 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003741}
3742EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003743
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744/**
3745 * kmem_cache_free - Deallocate an object
3746 * @cachep: The cache the allocation was from.
3747 * @objp: The previously allocated object.
3748 *
3749 * Free an object which was previously allocated from this
3750 * cache.
3751 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003752void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753{
3754 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003755 cachep = cache_from_obj(cachep, objp);
3756 if (!cachep)
3757 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758
3759 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003760 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003761 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003762 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003763 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003765
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003766 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767}
3768EXPORT_SYMBOL(kmem_cache_free);
3769
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003770void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3771{
3772 struct kmem_cache *s;
3773 size_t i;
3774
3775 local_irq_disable();
3776 for (i = 0; i < size; i++) {
3777 void *objp = p[i];
3778
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07003779 if (!orig_s) /* called via kfree_bulk */
3780 s = virt_to_cache(objp);
3781 else
3782 s = cache_from_obj(orig_s, objp);
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003783
3784 debug_check_no_locks_freed(objp, s->object_size);
3785 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3786 debug_check_no_obj_freed(objp, s->object_size);
3787
3788 __cache_free(s, objp, _RET_IP_);
3789 }
3790 local_irq_enable();
3791
3792 /* FIXME: add tracing */
3793}
3794EXPORT_SYMBOL(kmem_cache_free_bulk);
3795
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 * kfree - free previously allocated memory
3798 * @objp: pointer returned by kmalloc.
3799 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003800 * If @objp is NULL, no operation is performed.
3801 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 * Don't free memory not originally allocated by kmalloc()
3803 * or you will run into trouble.
3804 */
3805void kfree(const void *objp)
3806{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003807 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 unsigned long flags;
3809
Pekka Enberg2121db72009-03-25 11:05:57 +02003810 trace_kfree(_RET_IP_, objp);
3811
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003812 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 return;
3814 local_irq_save(flags);
3815 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003816 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003817 debug_check_no_locks_freed(objp, c->object_size);
3818
3819 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003820 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 local_irq_restore(flags);
3822}
3823EXPORT_SYMBOL(kfree);
3824
Christoph Lametere498be72005-09-09 13:03:32 -07003825/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003826 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003827 */
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003828static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003829{
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003830 int ret;
Christoph Lametere498be72005-09-09 13:03:32 -07003831 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003832 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07003833
Mel Gorman9c09a952008-01-24 05:49:54 -08003834 for_each_online_node(node) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003835 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3836 if (ret)
Christoph Lametere498be72005-09-09 13:03:32 -07003837 goto fail;
3838
Christoph Lametere498be72005-09-09 13:03:32 -07003839 }
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003840
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003842
Andrew Mortona737b3e2006-03-22 00:08:11 -08003843fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003844 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003845 /* Cache is not active yet. Roll back what we did */
3846 node--;
3847 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003848 n = get_node(cachep, node);
3849 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003850 kfree(n->shared);
3851 free_alien_cache(n->alien);
3852 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003853 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003854 }
3855 node--;
3856 }
3857 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003858 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003859}
3860
Christoph Lameter18004c52012-07-06 15:25:12 -05003861/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003862static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003863 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003865 struct array_cache __percpu *cpu_cache, *prev;
3866 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003868 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3869 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003870 return -ENOMEM;
3871
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003872 prev = cachep->cpu_cache;
3873 cachep->cpu_cache = cpu_cache;
Greg Thelena87c75f2017-05-03 14:51:47 -07003874 /*
3875 * Without a previous cpu_cache there's no need to synchronize remote
3876 * cpus, so skip the IPIs.
3877 */
3878 if (prev)
3879 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003880
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882 cachep->batchcount = batchcount;
3883 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003884 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003886 if (!prev)
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003887 goto setup_node;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003888
3889 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003890 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003891 int node;
3892 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003893 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003894
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003895 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003896 n = get_node(cachep, node);
3897 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003898 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003899 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003900 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003902 free_percpu(prev);
3903
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003904setup_node:
3905 return setup_kmem_cache_nodes(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906}
3907
Glauber Costa943a4512012-12-18 14:23:03 -08003908static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3909 int batchcount, int shared, gfp_t gfp)
3910{
3911 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003912 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003913
3914 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3915
3916 if (slab_state < FULL)
3917 return ret;
3918
3919 if ((ret < 0) || !is_root_cache(cachep))
3920 return ret;
3921
Vladimir Davydov426589f2015-02-12 14:59:23 -08003922 lockdep_assert_held(&slab_mutex);
3923 for_each_memcg_cache(c, cachep) {
3924 /* return value determined by the root cache only */
3925 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003926 }
3927
3928 return ret;
3929}
3930
Christoph Lameter18004c52012-07-06 15:25:12 -05003931/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003932static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933{
3934 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003935 int limit = 0;
3936 int shared = 0;
3937 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938
Thomas Garnier7c00fce2016-07-26 15:21:56 -07003939 err = cache_random_seq_create(cachep, cachep->num, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003940 if (err)
3941 goto end;
3942
Glauber Costa943a4512012-12-18 14:23:03 -08003943 if (!is_root_cache(cachep)) {
3944 struct kmem_cache *root = memcg_root_cache(cachep);
3945 limit = root->limit;
3946 shared = root->shared;
3947 batchcount = root->batchcount;
3948 }
3949
3950 if (limit && shared && batchcount)
3951 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003952 /*
3953 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 * - create a LIFO ordering, i.e. return objects that are cache-warm
3955 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003956 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 * bufctl chains: array operations are cheaper.
3958 * The numbers are guessed, we should auto-tune as described by
3959 * Bonwick.
3960 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003961 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003963 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003965 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003967 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 limit = 54;
3969 else
3970 limit = 120;
3971
Andrew Mortona737b3e2006-03-22 00:08:11 -08003972 /*
3973 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 * allocation behaviour: Most allocs on one cpu, most free operations
3975 * on another cpu. For these cases, an efficient object passing between
3976 * cpus is necessary. This is provided by a shared array. The array
3977 * replaces Bonwick's magazine layer.
3978 * On uniprocessor, it's functionally equivalent (but less efficient)
3979 * to a larger limit. Thus disabled by default.
3980 */
3981 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003982 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984
3985#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003986 /*
3987 * With debugging enabled, large batchcount lead to excessively long
3988 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 */
3990 if (limit > 32)
3991 limit = 32;
3992#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003993 batchcount = (limit + 1) / 2;
3994skip_setup:
3995 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003996end:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 if (err)
Joe Perches11705322016-03-17 14:19:50 -07003998 pr_err("enable_cpucache failed for %s, error %d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003999 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004000 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001}
4002
Christoph Lameter1b552532006-03-22 00:09:07 -08004003/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004004 * Drain an array if it contains any elements taking the node lock only if
4005 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004006 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004007 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004008static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004009 struct array_cache *ac, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010{
Joonsoo Kim97654df2014-08-06 16:04:25 -07004011 LIST_HEAD(list);
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004012
4013 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4014 check_mutex_acquired();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015
Christoph Lameter1b552532006-03-22 00:09:07 -08004016 if (!ac || !ac->avail)
4017 return;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004018
4019 if (ac->touched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020 ac->touched = 0;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004021 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 }
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004023
4024 spin_lock_irq(&n->list_lock);
4025 drain_array_locked(cachep, ac, node, false, &list);
4026 spin_unlock_irq(&n->list_lock);
4027
4028 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029}
4030
4031/**
4032 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004033 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034 *
4035 * Called from workqueue/eventd every few seconds.
4036 * Purpose:
4037 * - clear the per-cpu caches for this CPU.
4038 * - return freeable pages to the main free memory pool.
4039 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004040 * If we cannot acquire the cache chain mutex then just give up - we'll try
4041 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004043static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004045 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004046 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004047 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004048 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049
Christoph Lameter18004c52012-07-06 15:25:12 -05004050 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004052 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053
Christoph Lameter18004c52012-07-06 15:25:12 -05004054 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 check_irq_on();
4056
Christoph Lameter35386e32006-03-22 00:09:05 -08004057 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004058 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004059 * have established with reasonable certainty that
4060 * we can do some work if the lock was obtained.
4061 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07004062 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08004063
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004064 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004066 drain_array(searchp, n, cpu_cache_get(searchp), node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067
Christoph Lameter35386e32006-03-22 00:09:05 -08004068 /*
4069 * These are racy checks but it does not matter
4070 * if we skip one check or scan twice.
4071 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004072 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004073 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004075 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004077 drain_array(searchp, n, n->shared, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004079 if (n->free_touched)
4080 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004081 else {
4082 int freed;
4083
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004084 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004085 5 * searchp->num - 1) / (5 * searchp->num));
4086 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004088next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 cond_resched();
4090 }
4091 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004092 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004093 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004094out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004095 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004096 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097}
4098
Linus Torvalds158a9622008-01-02 13:04:48 -08004099#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004100void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101{
Greg Thelenf728b0a2016-12-12 16:41:41 -08004102 unsigned long active_objs, num_objs, active_slabs;
David Rientjesbf00bd32016-12-12 16:41:44 -08004103 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4104 unsigned long free_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004105 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004106 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107
Christoph Lameter18bf8542014-08-06 16:04:11 -07004108 for_each_kmem_cache_node(cachep, node, n) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004109 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004110 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004111
David Rientjesbf00bd32016-12-12 16:41:44 -08004112 total_slabs += n->total_slabs;
4113 free_slabs += n->free_slabs;
Greg Thelenf728b0a2016-12-12 16:41:41 -08004114 free_objs += n->free_objects;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004115
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004116 if (n->shared)
4117 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004118
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004119 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120 }
David Rientjesbf00bd32016-12-12 16:41:44 -08004121 num_objs = total_slabs * cachep->num;
4122 active_slabs = total_slabs - free_slabs;
Greg Thelenf728b0a2016-12-12 16:41:41 -08004123 active_objs = num_objs - free_objs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124
Glauber Costa0d7561c2012-10-19 18:20:27 +04004125 sinfo->active_objs = active_objs;
4126 sinfo->num_objs = num_objs;
4127 sinfo->active_slabs = active_slabs;
David Rientjesbf00bd32016-12-12 16:41:44 -08004128 sinfo->num_slabs = total_slabs;
Glauber Costa0d7561c2012-10-19 18:20:27 +04004129 sinfo->shared_avail = shared_avail;
4130 sinfo->limit = cachep->limit;
4131 sinfo->batchcount = cachep->batchcount;
4132 sinfo->shared = cachep->shared;
4133 sinfo->objects_per_slab = cachep->num;
4134 sinfo->cache_order = cachep->gfporder;
4135}
4136
4137void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4138{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004140 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141 unsigned long high = cachep->high_mark;
4142 unsigned long allocs = cachep->num_allocations;
4143 unsigned long grown = cachep->grown;
4144 unsigned long reaped = cachep->reaped;
4145 unsigned long errors = cachep->errors;
4146 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004148 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004149 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150
Joe Perches756a0252016-03-17 14:19:47 -07004151 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
Joe Perchese92dd4f2010-03-26 19:27:58 -07004152 allocs, high, grown,
4153 reaped, errors, max_freeable, node_allocs,
4154 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 }
4156 /* cpu stats */
4157 {
4158 unsigned long allochit = atomic_read(&cachep->allochit);
4159 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4160 unsigned long freehit = atomic_read(&cachep->freehit);
4161 unsigned long freemiss = atomic_read(&cachep->freemiss);
4162
4163 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004164 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 }
4166#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167}
4168
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169#define MAX_SLABINFO_WRITE 128
4170/**
4171 * slabinfo_write - Tuning for the slab allocator
4172 * @file: unused
4173 * @buffer: user buffer
4174 * @count: data length
4175 * @ppos: unused
4176 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004177ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004178 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004180 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004182 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004183
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184 if (count > MAX_SLABINFO_WRITE)
4185 return -EINVAL;
4186 if (copy_from_user(&kbuf, buffer, count))
4187 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004188 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189
4190 tmp = strchr(kbuf, ' ');
4191 if (!tmp)
4192 return -EINVAL;
4193 *tmp = '\0';
4194 tmp++;
4195 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4196 return -EINVAL;
4197
4198 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004199 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004201 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004203 if (limit < 1 || batchcount < 1 ||
4204 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004205 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004207 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004208 batchcount, shared,
4209 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 }
4211 break;
4212 }
4213 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004214 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 if (res >= 0)
4216 res = count;
4217 return res;
4218}
Al Viro871751e2006-03-25 03:06:39 -08004219
4220#ifdef CONFIG_DEBUG_SLAB_LEAK
4221
Al Viro871751e2006-03-25 03:06:39 -08004222static inline int add_caller(unsigned long *n, unsigned long v)
4223{
4224 unsigned long *p;
4225 int l;
4226 if (!v)
4227 return 1;
4228 l = n[1];
4229 p = n + 2;
4230 while (l) {
4231 int i = l/2;
4232 unsigned long *q = p + 2 * i;
4233 if (*q == v) {
4234 q[1]++;
4235 return 1;
4236 }
4237 if (*q > v) {
4238 l = i;
4239 } else {
4240 p = q + 2;
4241 l -= i + 1;
4242 }
4243 }
4244 if (++n[1] == n[0])
4245 return 0;
4246 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4247 p[0] = v;
4248 p[1] = 1;
4249 return 1;
4250}
4251
Joonsoo Kim8456a642013-10-24 10:07:49 +09004252static void handle_slab(unsigned long *n, struct kmem_cache *c,
4253 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004254{
4255 void *p;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004256 int i, j;
4257 unsigned long v;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004258
Al Viro871751e2006-03-25 03:06:39 -08004259 if (n[0] == n[1])
4260 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004261 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004262 bool active = true;
4263
4264 for (j = page->active; j < c->num; j++) {
4265 if (get_free_obj(page, j) == i) {
4266 active = false;
4267 break;
4268 }
4269 }
4270
4271 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004272 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004273
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004274 /*
4275 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4276 * mapping is established when actual object allocation and
4277 * we could mistakenly access the unmapped object in the cpu
4278 * cache.
4279 */
4280 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4281 continue;
4282
4283 if (!add_caller(n, v))
Al Viro871751e2006-03-25 03:06:39 -08004284 return;
4285 }
4286}
4287
4288static void show_symbol(struct seq_file *m, unsigned long address)
4289{
4290#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004291 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004292 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004293
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004294 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004295 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004296 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004297 seq_printf(m, " [%s]", modname);
4298 return;
4299 }
4300#endif
4301 seq_printf(m, "%p", (void *)address);
4302}
4303
4304static int leaks_show(struct seq_file *m, void *p)
4305{
Thierry Reding0672aa72012-06-22 19:42:49 +02004306 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004307 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004308 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004309 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004310 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004311 int node;
4312 int i;
4313
4314 if (!(cachep->flags & SLAB_STORE_USER))
4315 return 0;
4316 if (!(cachep->flags & SLAB_RED_ZONE))
4317 return 0;
4318
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004319 /*
4320 * Set store_user_clean and start to grab stored user information
4321 * for all objects on this cache. If some alloc/free requests comes
4322 * during the processing, information would be wrong so restart
4323 * whole processing.
4324 */
4325 do {
4326 set_store_user_clean(cachep);
4327 drain_cpu_caches(cachep);
Al Viro871751e2006-03-25 03:06:39 -08004328
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004329 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004330
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004331 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004332
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004333 check_irq_on();
4334 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004335
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004336 list_for_each_entry(page, &n->slabs_full, lru)
4337 handle_slab(x, cachep, page);
4338 list_for_each_entry(page, &n->slabs_partial, lru)
4339 handle_slab(x, cachep, page);
4340 spin_unlock_irq(&n->list_lock);
4341 }
4342 } while (!is_store_user_clean(cachep));
4343
Al Viro871751e2006-03-25 03:06:39 -08004344 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004345 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004346 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004347 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004348 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004349 if (!m->private) {
4350 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004351 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004352 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004353 return -ENOMEM;
4354 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004355 *(unsigned long *)m->private = x[0] * 2;
4356 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004357 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004358 /* Now make sure this entry will be retried */
4359 m->count = m->size;
4360 return 0;
4361 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004362 for (i = 0; i < x[1]; i++) {
4363 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4364 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004365 seq_putc(m, '\n');
4366 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004367
Al Viro871751e2006-03-25 03:06:39 -08004368 return 0;
4369}
4370
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004371static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004372 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004373 .next = slab_next,
4374 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004375 .show = leaks_show,
4376};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004377
4378static int slabstats_open(struct inode *inode, struct file *file)
4379{
Rob Jonesb208ce32014-10-09 15:28:03 -07004380 unsigned long *n;
4381
4382 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4383 if (!n)
4384 return -ENOMEM;
4385
4386 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4387
4388 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004389}
4390
4391static const struct file_operations proc_slabstats_operations = {
4392 .open = slabstats_open,
4393 .read = seq_read,
4394 .llseek = seq_lseek,
4395 .release = seq_release_private,
4396};
Al Viro871751e2006-03-25 03:06:39 -08004397#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004398
4399static int __init slab_proc_init(void)
4400{
4401#ifdef CONFIG_DEBUG_SLAB_LEAK
4402 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4403#endif
4404 return 0;
4405}
4406module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004407#endif
4408
Kees Cook04385fc2016-06-23 15:20:59 -07004409#ifdef CONFIG_HARDENED_USERCOPY
4410/*
4411 * Rejects objects that are incorrectly sized.
4412 *
4413 * Returns NULL if check passes, otherwise const char * to name of cache
4414 * to indicate an error.
4415 */
4416const char *__check_heap_object(const void *ptr, unsigned long n,
4417 struct page *page)
4418{
4419 struct kmem_cache *cachep;
4420 unsigned int objnr;
4421 unsigned long offset;
4422
4423 /* Find and validate object. */
4424 cachep = page->slab_cache;
4425 objnr = obj_to_index(cachep, page, (void *)ptr);
4426 BUG_ON(objnr >= cachep->num);
4427
4428 /* Find offset within object. */
4429 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4430
4431 /* Allow address range falling entirely within object size. */
4432 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4433 return NULL;
4434
4435 return cachep->name;
4436}
4437#endif /* CONFIG_HARDENED_USERCOPY */
4438
Manfred Spraul00e145b2005-09-03 15:55:07 -07004439/**
4440 * ksize - get the actual amount of memory allocated for a given object
4441 * @objp: Pointer to the object
4442 *
4443 * kmalloc may internally round up allocations and return more memory
4444 * than requested. ksize() can be used to determine the actual amount of
4445 * memory allocated. The caller may use this additional memory, even though
4446 * a smaller amount of memory was initially specified with the kmalloc call.
4447 * The caller must guarantee that objp points to a valid object previously
4448 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4449 * must not be freed during the duration of the call.
4450 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004451size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004452{
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004453 size_t size;
4454
Christoph Lameteref8b4522007-10-16 01:24:46 -07004455 BUG_ON(!objp);
4456 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004457 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004459 size = virt_to_cache(objp)->object_size;
4460 /* We assume that ksize callers could use the whole allocated area,
4461 * so we need to unpoison this area.
4462 */
Alexander Potapenko4ebb31a42016-05-20 16:59:14 -07004463 kasan_unpoison_shadow(objp, size);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004464
4465 return size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004466}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004467EXPORT_SYMBOL(ksize);