blob: 7e8bd8d828bc0e5c7e96919c6e24186d8d5382df [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
Christoph Lameter881db7f2011-06-01 12:25:53 -05005 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07007 *
Christoph Lametercde53532008-07-04 09:59:22 -07008 * (C) 2007 SGI, Christoph Lameter
Christoph Lameter881db7f2011-06-01 12:25:53 -05009 * (C) 2011 Linux Foundation, Christoph Lameter
Christoph Lameter81819f02007-05-06 14:49:36 -070010 */
11
12#include <linux/mm.h>
Nick Piggin1eb5ac62009-05-05 19:13:44 +100013#include <linux/swap.h> /* struct reclaim_state */
Christoph Lameter81819f02007-05-06 14:49:36 -070014#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
Christoph Lameter97d06602012-07-06 15:25:11 -050019#include "slab.h"
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040020#include <linux/proc_fs.h>
Andrew Morton3ac38fa2013-04-29 15:08:06 -070021#include <linux/notifier.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070022#include <linux/seq_file.h>
Vegard Nossum5a896d92008-04-04 00:54:48 +020023#include <linux/kmemcheck.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070028#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070029#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070030#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070031#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090032#include <linux/fault-inject.h>
Pekka Enbergbfa71452011-07-07 22:47:01 +030033#include <linux/stacktrace.h>
Christoph Lameter4de900b2012-01-30 15:53:51 -060034#include <linux/prefetch.h>
Glauber Costa2633d7a2012-12-18 14:22:34 -080035#include <linux/memcontrol.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070036
Richard Kennedy4a923792010-10-21 10:29:19 +010037#include <trace/events/kmem.h>
38
Mel Gorman072bb0a2012-07-31 16:43:58 -070039#include "internal.h"
40
Christoph Lameter81819f02007-05-06 14:49:36 -070041/*
42 * Lock order:
Christoph Lameter18004c52012-07-06 15:25:12 -050043 * 1. slab_mutex (Global Mutex)
Christoph Lameter881db7f2011-06-01 12:25:53 -050044 * 2. node->list_lock
45 * 3. slab_lock(page) (Only on some arches and for debugging)
Christoph Lameter81819f02007-05-06 14:49:36 -070046 *
Christoph Lameter18004c52012-07-06 15:25:12 -050047 * slab_mutex
Christoph Lameter881db7f2011-06-01 12:25:53 -050048 *
Christoph Lameter18004c52012-07-06 15:25:12 -050049 * The role of the slab_mutex is to protect the list of all the slabs
Christoph Lameter881db7f2011-06-01 12:25:53 -050050 * and to synchronize major metadata changes to slab cache structures.
51 *
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
58 *
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -070064 *
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
70 *
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
75 * the list lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070076 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
80 *
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
83 *
Christoph Lameter672bba32007-05-09 02:32:39 -070084 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070086 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070087 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070089 *
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
93 *
94 * Overloading of page flags that are otherwise used for LRU management.
95 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070096 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700108 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -0700109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700111 *
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700114 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700115 */
116
Christoph Lameteraf537b02010-07-09 14:07:14 -0500117static inline int kmem_cache_debug(struct kmem_cache *s)
118{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700119#ifdef CONFIG_SLUB_DEBUG
Christoph Lameteraf537b02010-07-09 14:07:14 -0500120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
Christoph Lameter5577bd82007-05-16 22:10:56 -0700121#else
Christoph Lameteraf537b02010-07-09 14:07:14 -0500122 return 0;
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123#endif
Christoph Lameteraf537b02010-07-09 14:07:14 -0500124}
Christoph Lameter5577bd82007-05-16 22:10:56 -0700125
Joonsoo Kim345c9052013-06-19 14:05:52 +0900126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127{
128#ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
130#else
131 return false;
132#endif
133}
134
Christoph Lameter81819f02007-05-06 14:49:36 -0700135/*
136 * Issues still to be resolved:
137 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700140 * - Variable sizing of the per node arrays
141 */
142
143/* Enable to test recovery from slab corruption on boot */
144#undef SLUB_RESILIENCY_TEST
145
Christoph Lameterb789ef52011-06-01 12:25:49 -0500146/* Enable to log cmpxchg failures */
147#undef SLUB_DEBUG_CMPXCHG
148
Christoph Lameter81819f02007-05-06 14:49:36 -0700149/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800153#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700154
Christoph Lameter2086d262007-05-06 14:49:46 -0700155/*
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in the.
159 */
160#define MAX_PARTIAL 10
161
Christoph Lameter81819f02007-05-06 14:49:36 -0700162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700164
Christoph Lameter81819f02007-05-06 14:49:36 -0700165/*
David Rientjes3de47212009-07-27 18:30:35 -0700166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
168 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700169 */
David Rientjes3de47212009-07-27 18:30:35 -0700170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700171
172/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700173 * Set of flags that will prevent slab merging
174 */
175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +0300176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177 SLAB_FAILSLAB)
Christoph Lameter81819f02007-05-06 14:49:36 -0700178
179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
Vegard Nossum5a896d92008-04-04 00:54:48 +0200180 SLAB_CACHE_DMA | SLAB_NOTRACK)
Christoph Lameter81819f02007-05-06 14:49:36 -0700181
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400182#define OO_SHIFT 16
183#define OO_MASK ((1 << OO_SHIFT) - 1)
Christoph Lameter50d5c412011-06-01 12:25:45 -0500184#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400185
Christoph Lameter81819f02007-05-06 14:49:36 -0700186/* Internal SLUB flags */
Christoph Lameterf90ec392010-07-09 14:07:11 -0500187#define __OBJECT_POISON 0x80000000UL /* Poison object */
Christoph Lameterb789ef52011-06-01 12:25:49 -0500188#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
Christoph Lameter81819f02007-05-06 14:49:36 -0700189
Christoph Lameter81819f02007-05-06 14:49:36 -0700190#ifdef CONFIG_SMP
191static struct notifier_block slab_notifier;
192#endif
193
Christoph Lameter02cbc872007-05-09 02:32:43 -0700194/*
195 * Tracking user of a slab.
196 */
Ben Greeard6543e32011-07-07 11:36:36 -0700197#define TRACK_ADDRS_COUNT 16
Christoph Lameter02cbc872007-05-09 02:32:43 -0700198struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300199 unsigned long addr; /* Called from address */
Ben Greeard6543e32011-07-07 11:36:36 -0700200#ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202#endif
Christoph Lameter02cbc872007-05-09 02:32:43 -0700203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
206};
207
208enum track_item { TRACK_ALLOC, TRACK_FREE };
209
Christoph Lameterab4d5ed2010-10-05 13:57:26 -0500210#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -0700211static int sysfs_slab_add(struct kmem_cache *);
212static int sysfs_slab_alias(struct kmem_cache *, const char *);
213static void sysfs_slab_remove(struct kmem_cache *);
Glauber Costa107dab52012-12-18 14:23:05 -0800214static void memcg_propagate_slab_attrs(struct kmem_cache *s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700215#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700216static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
217static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 { return 0; }
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000219static inline void sysfs_slab_remove(struct kmem_cache *s) { }
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800220
Glauber Costa107dab52012-12-18 14:23:05 -0800221static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
Christoph Lameter81819f02007-05-06 14:49:36 -0700222#endif
223
Christoph Lameter4fdccdf2011-03-22 13:35:00 -0500224static inline void stat(const struct kmem_cache *s, enum stat_item si)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800225{
226#ifdef CONFIG_SLUB_STATS
Christoph Lameter84e554e62009-12-18 16:26:23 -0600227 __this_cpu_inc(s->cpu_slab->stat[si]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800228#endif
229}
230
Christoph Lameter81819f02007-05-06 14:49:36 -0700231/********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
234
Christoph Lameter81819f02007-05-06 14:49:36 -0700235static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
236{
Christoph Lameter81819f02007-05-06 14:49:36 -0700237 return s->node[node];
Christoph Lameter81819f02007-05-06 14:49:36 -0700238}
239
Christoph Lameter6446faa2008-02-15 23:45:26 -0800240/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700241static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
243{
244 void *base;
245
Christoph Lametera973e9d2008-03-01 13:40:44 -0800246 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700247 return 1;
248
Christoph Lametera973e9d2008-03-01 13:40:44 -0800249 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300250 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700251 (object - base) % s->size) {
252 return 0;
253 }
254
255 return 1;
256}
257
Christoph Lameter7656c722007-05-09 02:32:40 -0700258static inline void *get_freepointer(struct kmem_cache *s, void *object)
259{
260 return *(void **)(object + s->offset);
261}
262
Eric Dumazet0ad95002011-12-16 16:25:34 +0100263static void prefetch_freepointer(const struct kmem_cache *s, void *object)
264{
265 prefetch(object + s->offset);
266}
267
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500268static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
269{
270 void *p;
271
272#ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
274#else
275 p = get_freepointer(s, object);
276#endif
277 return p;
278}
279
Christoph Lameter7656c722007-05-09 02:32:40 -0700280static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
281{
282 *(void **)(object + s->offset) = fp;
283}
284
285/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300286#define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700288 __p += (__s)->size)
289
Christoph Lameter7656c722007-05-09 02:32:40 -0700290/* Determine object index from a given position */
291static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292{
293 return (p - addr) / s->size;
294}
295
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100296static inline size_t slab_ksize(const struct kmem_cache *s)
297{
298#ifdef CONFIG_SLUB_DEBUG
299 /*
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
302 */
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500304 return s->object_size;
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100305
306#endif
307 /*
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
311 */
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
313 return s->inuse;
314 /*
315 * Else we can use all the padding etc for the allocation
316 */
317 return s->size;
318}
319
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800320static inline int order_objects(int order, unsigned long size, int reserved)
321{
322 return ((PAGE_SIZE << order) - reserved) / size;
323}
324
Christoph Lameter834f3d12008-04-14 19:11:31 +0300325static inline struct kmem_cache_order_objects oo_make(int order,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800326 unsigned long size, int reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300327{
328 struct kmem_cache_order_objects x = {
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800329 (order << OO_SHIFT) + order_objects(order, size, reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300330 };
331
332 return x;
333}
334
335static inline int oo_order(struct kmem_cache_order_objects x)
336{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400337 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300338}
339
340static inline int oo_objects(struct kmem_cache_order_objects x)
341{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400342 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300343}
344
Christoph Lameter881db7f2011-06-01 12:25:53 -0500345/*
346 * Per slab locking using the pagelock
347 */
348static __always_inline void slab_lock(struct page *page)
349{
350 bit_spin_lock(PG_locked, &page->flags);
351}
352
353static __always_inline void slab_unlock(struct page *page)
354{
355 __bit_spin_unlock(PG_locked, &page->flags);
356}
357
Christoph Lameter1d071712011-07-14 12:49:12 -0500358/* Interrupts must be disabled (for the fallback code to work right) */
359static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
362 const char *n)
363{
Christoph Lameter1d071712011-07-14 12:49:12 -0500364 VM_BUG_ON(!irqs_disabled());
Heiko Carstens25654092012-01-12 17:17:33 -0800365#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -0500367 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000368 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500369 freelist_old, counters_old,
370 freelist_new, counters_new))
371 return 1;
372 } else
373#endif
374 {
Christoph Lameter881db7f2011-06-01 12:25:53 -0500375 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800376 if (page->freelist == freelist_old &&
377 page->counters == counters_old) {
Christoph Lameterb789ef52011-06-01 12:25:49 -0500378 page->freelist = freelist_new;
379 page->counters = counters_new;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500380 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500381 return 1;
382 }
Christoph Lameter881db7f2011-06-01 12:25:53 -0500383 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500384 }
385
386 cpu_relax();
387 stat(s, CMPXCHG_DOUBLE_FAIL);
388
389#ifdef SLUB_DEBUG_CMPXCHG
390 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
391#endif
392
393 return 0;
394}
395
Christoph Lameter1d071712011-07-14 12:49:12 -0500396static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
397 void *freelist_old, unsigned long counters_old,
398 void *freelist_new, unsigned long counters_new,
399 const char *n)
400{
Heiko Carstens25654092012-01-12 17:17:33 -0800401#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameter1d071712011-07-14 12:49:12 -0500403 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000404 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameter1d071712011-07-14 12:49:12 -0500405 freelist_old, counters_old,
406 freelist_new, counters_new))
407 return 1;
408 } else
409#endif
410 {
411 unsigned long flags;
412
413 local_irq_save(flags);
414 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800415 if (page->freelist == freelist_old &&
416 page->counters == counters_old) {
Christoph Lameter1d071712011-07-14 12:49:12 -0500417 page->freelist = freelist_new;
418 page->counters = counters_new;
419 slab_unlock(page);
420 local_irq_restore(flags);
421 return 1;
422 }
423 slab_unlock(page);
424 local_irq_restore(flags);
425 }
426
427 cpu_relax();
428 stat(s, CMPXCHG_DOUBLE_FAIL);
429
430#ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432#endif
433
434 return 0;
435}
436
Christoph Lameter41ecc552007-05-09 02:32:44 -0700437#ifdef CONFIG_SLUB_DEBUG
438/*
Christoph Lameter5f80b132011-04-15 14:48:13 -0500439 * Determine a map of object in use on a page.
440 *
Christoph Lameter881db7f2011-06-01 12:25:53 -0500441 * Node listlock must be held to guarantee that the page does
Christoph Lameter5f80b132011-04-15 14:48:13 -0500442 * not vanish from under us.
443 */
444static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
445{
446 void *p;
447 void *addr = page_address(page);
448
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
451}
452
Christoph Lameter41ecc552007-05-09 02:32:44 -0700453/*
454 * Debug settings:
455 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700456#ifdef CONFIG_SLUB_DEBUG_ON
457static int slub_debug = DEBUG_DEFAULT_FLAGS;
458#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700459static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700460#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700461
462static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700463static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700464
Christoph Lameter7656c722007-05-09 02:32:40 -0700465/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700466 * Object debugging
467 */
468static void print_section(char *text, u8 *addr, unsigned int length)
469{
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200470 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
471 length, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700472}
473
Christoph Lameter81819f02007-05-06 14:49:36 -0700474static struct track *get_track(struct kmem_cache *s, void *object,
475 enum track_item alloc)
476{
477 struct track *p;
478
479 if (s->offset)
480 p = object + s->offset + sizeof(void *);
481 else
482 p = object + s->inuse;
483
484 return p + alloc;
485}
486
487static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300488 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700489{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900490 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700491
Christoph Lameter81819f02007-05-06 14:49:36 -0700492 if (addr) {
Ben Greeard6543e32011-07-07 11:36:36 -0700493#ifdef CONFIG_STACKTRACE
494 struct stack_trace trace;
495 int i;
496
497 trace.nr_entries = 0;
498 trace.max_entries = TRACK_ADDRS_COUNT;
499 trace.entries = p->addrs;
500 trace.skip = 3;
501 save_stack_trace(&trace);
502
503 /* See rant in lockdep.c */
504 if (trace.nr_entries != 0 &&
505 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
506 trace.nr_entries--;
507
508 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
509 p->addrs[i] = 0;
510#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700511 p->addr = addr;
512 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400513 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700514 p->when = jiffies;
515 } else
516 memset(p, 0, sizeof(struct track));
517}
518
Christoph Lameter81819f02007-05-06 14:49:36 -0700519static void init_tracking(struct kmem_cache *s, void *object)
520{
Christoph Lameter24922682007-07-17 04:03:18 -0700521 if (!(s->flags & SLAB_STORE_USER))
522 return;
523
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300524 set_track(s, object, TRACK_FREE, 0UL);
525 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700526}
527
528static void print_track(const char *s, struct track *t)
529{
530 if (!t->addr)
531 return;
532
Linus Torvalds7daf7052008-07-14 12:12:53 -0700533 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300534 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Ben Greeard6543e32011-07-07 11:36:36 -0700535#ifdef CONFIG_STACKTRACE
536 {
537 int i;
538 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
539 if (t->addrs[i])
540 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
541 else
542 break;
543 }
544#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700545}
546
Christoph Lameter24922682007-07-17 04:03:18 -0700547static void print_tracking(struct kmem_cache *s, void *object)
548{
549 if (!(s->flags & SLAB_STORE_USER))
550 return;
551
552 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
553 print_track("Freed", get_track(s, object, TRACK_FREE));
554}
555
556static void print_page_info(struct page *page)
557{
Chen Gangd0e0ac92013-07-15 09:05:29 +0800558 printk(KERN_ERR
559 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
560 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700561
562}
563
564static void slab_bug(struct kmem_cache *s, char *fmt, ...)
565{
566 va_list args;
567 char buf[100];
568
569 va_start(args, fmt);
570 vsnprintf(buf, sizeof(buf), fmt, args);
571 va_end(args);
572 printk(KERN_ERR "========================================"
573 "=====================================\n");
Dave Jones265d47e2011-11-15 15:04:00 -0800574 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700575 printk(KERN_ERR "----------------------------------------"
576 "-------------------------------------\n\n");
Dave Jones645df232012-09-18 15:54:12 -0400577
Rusty Russell373d4d02013-01-21 17:17:39 +1030578 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Christoph Lameter24922682007-07-17 04:03:18 -0700579}
580
581static void slab_fix(struct kmem_cache *s, char *fmt, ...)
582{
583 va_list args;
584 char buf[100];
585
586 va_start(args, fmt);
587 vsnprintf(buf, sizeof(buf), fmt, args);
588 va_end(args);
589 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
590}
591
592static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700593{
594 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800595 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700596
597 print_tracking(s, p);
598
599 print_page_info(page);
600
601 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
602 p, p - addr, get_freepointer(s, p));
603
604 if (p > addr + 16)
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200605 print_section("Bytes b4 ", p - 16, 16);
Christoph Lameter24922682007-07-17 04:03:18 -0700606
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500607 print_section("Object ", p, min_t(unsigned long, s->object_size,
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200608 PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700609 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500610 print_section("Redzone ", p + s->object_size,
611 s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700612
Christoph Lameter81819f02007-05-06 14:49:36 -0700613 if (s->offset)
614 off = s->offset + sizeof(void *);
615 else
616 off = s->inuse;
617
Christoph Lameter24922682007-07-17 04:03:18 -0700618 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700619 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700620
621 if (off != s->size)
622 /* Beginning of the filler is the free pointer */
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200623 print_section("Padding ", p + off, s->size - off);
Christoph Lameter24922682007-07-17 04:03:18 -0700624
625 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700626}
627
628static void object_err(struct kmem_cache *s, struct page *page,
629 u8 *object, char *reason)
630{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700631 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700632 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700633}
634
Chen Gangd0e0ac92013-07-15 09:05:29 +0800635static void slab_err(struct kmem_cache *s, struct page *page,
636 const char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700637{
638 va_list args;
639 char buf[100];
640
Christoph Lameter24922682007-07-17 04:03:18 -0700641 va_start(args, fmt);
642 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700643 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700644 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700645 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700646 dump_stack();
647}
648
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500649static void init_object(struct kmem_cache *s, void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700650{
651 u8 *p = object;
652
653 if (s->flags & __OBJECT_POISON) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500654 memset(p, POISON_FREE, s->object_size - 1);
655 p[s->object_size - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700656 }
657
658 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500659 memset(p + s->object_size, val, s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700660}
661
Christoph Lameter24922682007-07-17 04:03:18 -0700662static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
663 void *from, void *to)
664{
665 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
666 memset(from, data, to - from);
667}
668
669static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
670 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800671 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700672{
673 u8 *fault;
674 u8 *end;
675
Akinobu Mita798248202011-10-31 17:08:07 -0700676 fault = memchr_inv(start, value, bytes);
Christoph Lameter24922682007-07-17 04:03:18 -0700677 if (!fault)
678 return 1;
679
680 end = start + bytes;
681 while (end > fault && end[-1] == value)
682 end--;
683
684 slab_bug(s, "%s overwritten", what);
685 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
686 fault, end - 1, fault[0], value);
687 print_trailer(s, page, object);
688
689 restore_bytes(s, what, value, fault, end);
690 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700691}
692
Christoph Lameter81819f02007-05-06 14:49:36 -0700693/*
694 * Object layout:
695 *
696 * object address
697 * Bytes of the object to be managed.
698 * If the freepointer may overlay the object then the free
699 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700700 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700701 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
702 * 0xa5 (POISON_END)
703 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500704 * object + s->object_size
Christoph Lameter81819f02007-05-06 14:49:36 -0700705 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700706 * Padding is extended by another word if Redzoning is enabled and
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500707 * object_size == inuse.
Christoph Lameter672bba32007-05-09 02:32:39 -0700708 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700709 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
710 * 0xcc (RED_ACTIVE) for objects in use.
711 *
712 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700713 * Meta data starts here.
714 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700715 * A. Free pointer (if we cannot overwrite object on free)
716 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700717 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800718 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700719 * before the word boundary.
720 *
721 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700722 *
723 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700724 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700725 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500726 * If slabcaches are merged then the object_size and inuse boundaries are mostly
Christoph Lameter672bba32007-05-09 02:32:39 -0700727 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700728 * may be used with merged slabcaches.
729 */
730
Christoph Lameter81819f02007-05-06 14:49:36 -0700731static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
732{
733 unsigned long off = s->inuse; /* The end of info */
734
735 if (s->offset)
736 /* Freepointer is placed after the object. */
737 off += sizeof(void *);
738
739 if (s->flags & SLAB_STORE_USER)
740 /* We also have user information there */
741 off += 2 * sizeof(struct track);
742
743 if (s->size == off)
744 return 1;
745
Christoph Lameter24922682007-07-17 04:03:18 -0700746 return check_bytes_and_report(s, page, p, "Object padding",
747 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700748}
749
Christoph Lameter39b26462008-04-14 19:11:30 +0300750/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700751static int slab_pad_check(struct kmem_cache *s, struct page *page)
752{
Christoph Lameter24922682007-07-17 04:03:18 -0700753 u8 *start;
754 u8 *fault;
755 u8 *end;
756 int length;
757 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700758
759 if (!(s->flags & SLAB_POISON))
760 return 1;
761
Christoph Lametera973e9d2008-03-01 13:40:44 -0800762 start = page_address(page);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800763 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
Christoph Lameter39b26462008-04-14 19:11:30 +0300764 end = start + length;
765 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700766 if (!remainder)
767 return 1;
768
Akinobu Mita798248202011-10-31 17:08:07 -0700769 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700770 if (!fault)
771 return 1;
772 while (end > fault && end[-1] == POISON_INUSE)
773 end--;
774
775 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200776 print_section("Padding ", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700777
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200778 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700779 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700780}
781
782static int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500783 void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700784{
785 u8 *p = object;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500786 u8 *endobject = object + s->object_size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700787
788 if (s->flags & SLAB_RED_ZONE) {
Christoph Lameter24922682007-07-17 04:03:18 -0700789 if (!check_bytes_and_report(s, page, object, "Redzone",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500790 endobject, val, s->inuse - s->object_size))
Christoph Lameter81819f02007-05-06 14:49:36 -0700791 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700792 } else {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500793 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800794 check_bytes_and_report(s, page, p, "Alignment padding",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800795 endobject, POISON_INUSE,
796 s->inuse - s->object_size);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800797 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700798 }
799
800 if (s->flags & SLAB_POISON) {
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500801 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700802 (!check_bytes_and_report(s, page, p, "Poison", p,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500803 POISON_FREE, s->object_size - 1) ||
Christoph Lameter24922682007-07-17 04:03:18 -0700804 !check_bytes_and_report(s, page, p, "Poison",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500805 p + s->object_size - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700806 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700807 /*
808 * check_pad_bytes cleans up on its own.
809 */
810 check_pad_bytes(s, page, p);
811 }
812
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500813 if (!s->offset && val == SLUB_RED_ACTIVE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700814 /*
815 * Object and freepointer overlap. Cannot check
816 * freepointer while object is allocated.
817 */
818 return 1;
819
820 /* Check free pointer validity */
821 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
822 object_err(s, page, p, "Freepointer corrupt");
823 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100824 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700825 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700826 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700827 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800828 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700829 return 0;
830 }
831 return 1;
832}
833
834static int check_slab(struct kmem_cache *s, struct page *page)
835{
Christoph Lameter39b26462008-04-14 19:11:30 +0300836 int maxobj;
837
Christoph Lameter81819f02007-05-06 14:49:36 -0700838 VM_BUG_ON(!irqs_disabled());
839
840 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700841 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700842 return 0;
843 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300844
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800845 maxobj = order_objects(compound_order(page), s->size, s->reserved);
Christoph Lameter39b26462008-04-14 19:11:30 +0300846 if (page->objects > maxobj) {
847 slab_err(s, page, "objects %u > max %u",
848 s->name, page->objects, maxobj);
849 return 0;
850 }
851 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700852 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300853 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700854 return 0;
855 }
856 /* Slab_pad_check fixes things up after itself */
857 slab_pad_check(s, page);
858 return 1;
859}
860
861/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700862 * Determine if a certain object on a page is on the freelist. Must hold the
863 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700864 */
865static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
866{
867 int nr = 0;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500868 void *fp;
Christoph Lameter81819f02007-05-06 14:49:36 -0700869 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300870 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700871
Christoph Lameter881db7f2011-06-01 12:25:53 -0500872 fp = page->freelist;
Christoph Lameter39b26462008-04-14 19:11:30 +0300873 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700874 if (fp == search)
875 return 1;
876 if (!check_valid_pointer(s, page, fp)) {
877 if (object) {
878 object_err(s, page, object,
879 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800880 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700881 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700882 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800883 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300884 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700885 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700886 return 0;
887 }
888 break;
889 }
890 object = fp;
891 fp = get_freepointer(s, object);
892 nr++;
893 }
894
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800895 max_objects = order_objects(compound_order(page), s->size, s->reserved);
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400896 if (max_objects > MAX_OBJS_PER_PAGE)
897 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300898
899 if (page->objects != max_objects) {
900 slab_err(s, page, "Wrong number of objects. Found %d but "
901 "should be %d", page->objects, max_objects);
902 page->objects = max_objects;
903 slab_fix(s, "Number of objects adjusted.");
904 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300905 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700906 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300907 "counted were %d", page->inuse, page->objects - nr);
908 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700909 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700910 }
911 return search == NULL;
912}
913
Christoph Lameter0121c6192008-04-29 16:11:12 -0700914static void trace(struct kmem_cache *s, struct page *page, void *object,
915 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700916{
917 if (s->flags & SLAB_TRACE) {
918 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
919 s->name,
920 alloc ? "alloc" : "free",
921 object, page->inuse,
922 page->freelist);
923
924 if (!alloc)
Chen Gangd0e0ac92013-07-15 09:05:29 +0800925 print_section("Object ", (void *)object,
926 s->object_size);
Christoph Lameter3ec09742007-05-16 22:11:00 -0700927
928 dump_stack();
929 }
930}
931
Christoph Lameter643b1132007-05-06 14:49:42 -0700932/*
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500933 * Hooks for other subsystems that check memory allocations. In a typical
934 * production configuration these hooks all should produce no code at all.
935 */
936static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
937{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500938 flags &= gfp_allowed_mask;
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500939 lockdep_trace_alloc(flags);
940 might_sleep_if(flags & __GFP_WAIT);
941
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500942 return should_failslab(s->object_size, flags, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500943}
944
Chen Gangd0e0ac92013-07-15 09:05:29 +0800945static inline void slab_post_alloc_hook(struct kmem_cache *s,
946 gfp_t flags, void *object)
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500947{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500948 flags &= gfp_allowed_mask;
Eric Dumazetb3d41882011-02-14 18:35:22 +0100949 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500950 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500951}
952
953static inline void slab_free_hook(struct kmem_cache *s, void *x)
954{
955 kmemleak_free_recursive(x, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500956
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600957 /*
Xie XiuQid1756172013-10-18 09:12:43 +0800958 * Trouble is that we may no longer disable interrupts in the fast path
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600959 * So in order to make the debug calls that expect irqs to be
960 * disabled we need to disable interrupts temporarily.
961 */
962#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
963 {
964 unsigned long flags;
965
966 local_irq_save(flags);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500967 kmemcheck_slab_free(s, x, s->object_size);
968 debug_check_no_locks_freed(x, s->object_size);
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600969 local_irq_restore(flags);
970 }
971#endif
Thomas Gleixnerf9b615d2011-03-24 21:26:46 +0200972 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500973 debug_check_no_obj_freed(x, s->object_size);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500974}
975
976/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700977 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500978 *
979 * list_lock must be held.
Christoph Lameter643b1132007-05-06 14:49:42 -0700980 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500981static void add_full(struct kmem_cache *s,
982 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700983{
Christoph Lameter643b1132007-05-06 14:49:42 -0700984 if (!(s->flags & SLAB_STORE_USER))
985 return;
986
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500987 list_add(&page->lru, &n->full);
988}
Christoph Lameter643b1132007-05-06 14:49:42 -0700989
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500990/*
991 * list_lock must be held.
992 */
993static void remove_full(struct kmem_cache *s, struct page *page)
994{
995 if (!(s->flags & SLAB_STORE_USER))
996 return;
997
Christoph Lameter643b1132007-05-06 14:49:42 -0700998 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -0700999}
1000
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001001/* Tracking of the number of slabs for debugging purposes */
1002static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1003{
1004 struct kmem_cache_node *n = get_node(s, node);
1005
1006 return atomic_long_read(&n->nr_slabs);
1007}
1008
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001009static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1010{
1011 return atomic_long_read(&n->nr_slabs);
1012}
1013
Christoph Lameter205ab992008-04-14 19:11:40 +03001014static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001015{
1016 struct kmem_cache_node *n = get_node(s, node);
1017
1018 /*
1019 * May be called early in order to allocate a slab for the
1020 * kmem_cache_node structure. Solve the chicken-egg
1021 * dilemma by deferring the increment of the count during
1022 * bootstrap (see early_kmem_cache_node_alloc).
1023 */
Joonsoo Kim338b2642013-01-21 17:01:27 +09001024 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001025 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001026 atomic_long_add(objects, &n->total_objects);
1027 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001028}
Christoph Lameter205ab992008-04-14 19:11:40 +03001029static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001030{
1031 struct kmem_cache_node *n = get_node(s, node);
1032
1033 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001034 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001035}
1036
1037/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001038static void setup_object_debug(struct kmem_cache *s, struct page *page,
1039 void *object)
1040{
1041 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1042 return;
1043
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001044 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001045 init_tracking(s, object);
1046}
1047
Chen Gangd0e0ac92013-07-15 09:05:29 +08001048static noinline int alloc_debug_processing(struct kmem_cache *s,
1049 struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001050 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001051{
1052 if (!check_slab(s, page))
1053 goto bad;
1054
Christoph Lameter81819f02007-05-06 14:49:36 -07001055 if (!check_valid_pointer(s, page, object)) {
1056 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -07001057 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001058 }
1059
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001060 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Christoph Lameter81819f02007-05-06 14:49:36 -07001061 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001062
Christoph Lameter3ec09742007-05-16 22:11:00 -07001063 /* Success perform special debug activities for allocs */
1064 if (s->flags & SLAB_STORE_USER)
1065 set_track(s, object, TRACK_ALLOC, addr);
1066 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001067 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001068 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001069
Christoph Lameter81819f02007-05-06 14:49:36 -07001070bad:
1071 if (PageSlab(page)) {
1072 /*
1073 * If this is a slab page then lets do the best we can
1074 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001075 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001076 */
Christoph Lameter24922682007-07-17 04:03:18 -07001077 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001078 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001079 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001080 }
1081 return 0;
1082}
1083
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001084static noinline struct kmem_cache_node *free_debug_processing(
1085 struct kmem_cache *s, struct page *page, void *object,
1086 unsigned long addr, unsigned long *flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001087{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001088 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001089
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001090 spin_lock_irqsave(&n->list_lock, *flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001091 slab_lock(page);
1092
Christoph Lameter81819f02007-05-06 14:49:36 -07001093 if (!check_slab(s, page))
1094 goto fail;
1095
1096 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001097 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001098 goto fail;
1099 }
1100
1101 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -07001102 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -07001103 goto fail;
1104 }
1105
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001106 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001107 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001108
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001109 if (unlikely(s != page->slab_cache)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001110 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001111 slab_err(s, page, "Attempt to free object(0x%p) "
1112 "outside of slab", object);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001113 } else if (!page->slab_cache) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001114 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -07001115 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -07001116 object);
Christoph Lameter70d71222007-05-06 14:49:47 -07001117 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -08001118 } else
Christoph Lameter24922682007-07-17 04:03:18 -07001119 object_err(s, page, object,
1120 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -07001121 goto fail;
1122 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001123
Christoph Lameter3ec09742007-05-16 22:11:00 -07001124 if (s->flags & SLAB_STORE_USER)
1125 set_track(s, object, TRACK_FREE, addr);
1126 trace(s, page, object, 0);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001127 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001128out:
Christoph Lameter881db7f2011-06-01 12:25:53 -05001129 slab_unlock(page);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001130 /*
1131 * Keep node_lock to preserve integrity
1132 * until the object is actually freed
1133 */
1134 return n;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001135
Christoph Lameter81819f02007-05-06 14:49:36 -07001136fail:
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001137 slab_unlock(page);
1138 spin_unlock_irqrestore(&n->list_lock, *flags);
Christoph Lameter24922682007-07-17 04:03:18 -07001139 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001140 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001141}
1142
Christoph Lameter41ecc552007-05-09 02:32:44 -07001143static int __init setup_slub_debug(char *str)
1144{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001145 slub_debug = DEBUG_DEFAULT_FLAGS;
1146 if (*str++ != '=' || !*str)
1147 /*
1148 * No options specified. Switch on full debugging.
1149 */
1150 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001151
1152 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001153 /*
1154 * No options but restriction on slabs. This means full
1155 * debugging for slabs matching a pattern.
1156 */
1157 goto check_slabs;
1158
David Rientjesfa5ec8a2009-07-07 00:14:14 -07001159 if (tolower(*str) == 'o') {
1160 /*
1161 * Avoid enabling debugging on caches if its minimum order
1162 * would increase as a result.
1163 */
1164 disable_higher_order_debug = 1;
1165 goto out;
1166 }
1167
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001168 slub_debug = 0;
1169 if (*str == '-')
1170 /*
1171 * Switch off all debugging measures.
1172 */
1173 goto out;
1174
1175 /*
1176 * Determine which debug features should be switched on
1177 */
Pekka Enberg06428782008-01-07 23:20:27 -08001178 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001179 switch (tolower(*str)) {
1180 case 'f':
1181 slub_debug |= SLAB_DEBUG_FREE;
1182 break;
1183 case 'z':
1184 slub_debug |= SLAB_RED_ZONE;
1185 break;
1186 case 'p':
1187 slub_debug |= SLAB_POISON;
1188 break;
1189 case 'u':
1190 slub_debug |= SLAB_STORE_USER;
1191 break;
1192 case 't':
1193 slub_debug |= SLAB_TRACE;
1194 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001195 case 'a':
1196 slub_debug |= SLAB_FAILSLAB;
1197 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001198 default:
1199 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001200 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001201 }
1202 }
1203
1204check_slabs:
1205 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001206 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001207out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001208 return 1;
1209}
1210
1211__setup("slub_debug", setup_slub_debug);
1212
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001213static unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001214 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001215 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001216{
1217 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001218 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001219 */
Christoph Lametere1533622008-02-15 23:45:24 -08001220 if (slub_debug && (!slub_debug_slabs ||
David Rientjes3de47212009-07-27 18:30:35 -07001221 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1222 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001223
1224 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001225}
1226#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001227static inline void setup_object_debug(struct kmem_cache *s,
1228 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001229
Christoph Lameter3ec09742007-05-16 22:11:00 -07001230static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001231 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001232
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001233static inline struct kmem_cache_node *free_debug_processing(
1234 struct kmem_cache *s, struct page *page, void *object,
1235 unsigned long addr, unsigned long *flags) { return NULL; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001236
Christoph Lameter41ecc552007-05-09 02:32:44 -07001237static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1238 { return 1; }
1239static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001240 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001241static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1242 struct page *page) {}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001243static inline void remove_full(struct kmem_cache *s, struct page *page) {}
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001244static inline unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001245 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001246 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001247{
1248 return flags;
1249}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001250#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001251
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001252#define disable_higher_order_debug 0
1253
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001254static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1255 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001256static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1257 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001258static inline void inc_slabs_node(struct kmem_cache *s, int node,
1259 int objects) {}
1260static inline void dec_slabs_node(struct kmem_cache *s, int node,
1261 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001262
1263static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1264 { return 0; }
1265
1266static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1267 void *object) {}
1268
1269static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1270
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05001271#endif /* CONFIG_SLUB_DEBUG */
Christoph Lameter205ab992008-04-14 19:11:40 +03001272
Christoph Lameter81819f02007-05-06 14:49:36 -07001273/*
1274 * Slab allocation and freeing
1275 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001276static inline struct page *alloc_slab_page(gfp_t flags, int node,
1277 struct kmem_cache_order_objects oo)
1278{
1279 int order = oo_order(oo);
1280
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001281 flags |= __GFP_NOTRACK;
1282
Christoph Lameter2154a332010-07-09 14:07:10 -05001283 if (node == NUMA_NO_NODE)
Christoph Lameter65c33762008-04-14 19:11:40 +03001284 return alloc_pages(flags, order);
1285 else
Minchan Kim6b65aaf2010-04-14 23:58:36 +09001286 return alloc_pages_exact_node(node, flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001287}
1288
Christoph Lameter81819f02007-05-06 14:49:36 -07001289static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1290{
Pekka Enberg06428782008-01-07 23:20:27 -08001291 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001292 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001293 gfp_t alloc_gfp;
Christoph Lameter81819f02007-05-06 14:49:36 -07001294
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001295 flags &= gfp_allowed_mask;
1296
1297 if (flags & __GFP_WAIT)
1298 local_irq_enable();
1299
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001300 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001301
Pekka Enbergba522702009-06-24 21:59:51 +03001302 /*
1303 * Let the initial higher-order allocation fail under memory pressure
1304 * so we fall-back to the minimum order allocation.
1305 */
1306 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1307
1308 page = alloc_slab_page(alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001309 if (unlikely(!page)) {
1310 oo = s->min;
1311 /*
1312 * Allocation may have failed due to fragmentation.
1313 * Try a lower order alloc if possible
1314 */
1315 page = alloc_slab_page(flags, node, oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001316
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001317 if (page)
1318 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001319 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001320
David Rientjes737b7192012-07-09 14:00:38 -07001321 if (kmemcheck_enabled && page
Amerigo Wang5086c389c2009-08-19 21:44:13 +03001322 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001323 int pages = 1 << oo_order(oo);
1324
1325 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1326
1327 /*
1328 * Objects from caches that have a constructor don't get
1329 * cleared when they're allocated, so we need to do it here.
1330 */
1331 if (s->ctor)
1332 kmemcheck_mark_uninitialized_pages(page, pages);
1333 else
1334 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001335 }
1336
David Rientjes737b7192012-07-09 14:00:38 -07001337 if (flags & __GFP_WAIT)
1338 local_irq_disable();
1339 if (!page)
1340 return NULL;
1341
Christoph Lameter834f3d12008-04-14 19:11:31 +03001342 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001343 mod_zone_page_state(page_zone(page),
1344 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1345 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001346 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001347
1348 return page;
1349}
1350
1351static void setup_object(struct kmem_cache *s, struct page *page,
1352 void *object)
1353{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001354 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001355 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001356 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001357}
1358
1359static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1360{
1361 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001362 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001363 void *last;
1364 void *p;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001365 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001366
Christoph Lameter6cb06222007-10-16 01:25:41 -07001367 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001368
Christoph Lameter6cb06222007-10-16 01:25:41 -07001369 page = allocate_slab(s,
1370 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001371 if (!page)
1372 goto out;
1373
Glauber Costa1f458cb2012-12-18 14:22:50 -08001374 order = compound_order(page);
Christoph Lameter205ab992008-04-14 19:11:40 +03001375 inc_slabs_node(s, page_to_nid(page), page->objects);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001376 memcg_bind_pages(s, order);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001377 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001378 __SetPageSlab(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001379 if (page->pfmemalloc)
1380 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001381
1382 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001383
1384 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001385 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001386
1387 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001388 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001389 setup_object(s, page, last);
1390 set_freepointer(s, last, p);
1391 last = p;
1392 }
1393 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001394 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001395
1396 page->freelist = start;
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001397 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001398 page->frozen = 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001399out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001400 return page;
1401}
1402
1403static void __free_slab(struct kmem_cache *s, struct page *page)
1404{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001405 int order = compound_order(page);
1406 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001407
Christoph Lameteraf537b02010-07-09 14:07:14 -05001408 if (kmem_cache_debug(s)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001409 void *p;
1410
1411 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001412 for_each_object(p, s, page_address(page),
1413 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001414 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001415 }
1416
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001417 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001418
Christoph Lameter81819f02007-05-06 14:49:36 -07001419 mod_zone_page_state(page_zone(page),
1420 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1421 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001422 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001423
Mel Gorman072bb0a2012-07-31 16:43:58 -07001424 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001425 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001426
1427 memcg_release_pages(s, order);
Mel Gorman22b751c2013-02-22 16:34:59 -08001428 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001429 if (current->reclaim_state)
1430 current->reclaim_state->reclaimed_slab += pages;
Glauber Costad79923f2012-12-18 14:22:48 -08001431 __free_memcg_kmem_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001432}
1433
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001434#define need_reserve_slab_rcu \
1435 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1436
Christoph Lameter81819f02007-05-06 14:49:36 -07001437static void rcu_free_slab(struct rcu_head *h)
1438{
1439 struct page *page;
1440
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001441 if (need_reserve_slab_rcu)
1442 page = virt_to_head_page(h);
1443 else
1444 page = container_of((struct list_head *)h, struct page, lru);
1445
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001446 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001447}
1448
1449static void free_slab(struct kmem_cache *s, struct page *page)
1450{
1451 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001452 struct rcu_head *head;
1453
1454 if (need_reserve_slab_rcu) {
1455 int order = compound_order(page);
1456 int offset = (PAGE_SIZE << order) - s->reserved;
1457
1458 VM_BUG_ON(s->reserved != sizeof(*head));
1459 head = page_address(page) + offset;
1460 } else {
1461 /*
1462 * RCU free overloads the RCU head over the LRU
1463 */
1464 head = (void *)&page->lru;
1465 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001466
1467 call_rcu(head, rcu_free_slab);
1468 } else
1469 __free_slab(s, page);
1470}
1471
1472static void discard_slab(struct kmem_cache *s, struct page *page)
1473{
Christoph Lameter205ab992008-04-14 19:11:40 +03001474 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001475 free_slab(s, page);
1476}
1477
1478/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001479 * Management of partially allocated slabs.
1480 *
1481 * list_lock must be held.
Christoph Lameter81819f02007-05-06 14:49:36 -07001482 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001483static inline void add_partial(struct kmem_cache_node *n,
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001484 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001485{
Christoph Lametere95eed52007-05-06 14:49:44 -07001486 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001487 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001488 list_add_tail(&page->lru, &n->partial);
1489 else
1490 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001491}
1492
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001493/*
1494 * list_lock must be held.
1495 */
1496static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001497 struct page *page)
1498{
1499 list_del(&page->lru);
1500 n->nr_partial--;
1501}
1502
Christoph Lameter81819f02007-05-06 14:49:36 -07001503/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001504 * Remove slab from the partial list, freeze it and
1505 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001506 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001507 * Returns a list of objects or NULL if it fails.
1508 *
Christoph Lameter7ced3712012-05-09 10:09:53 -05001509 * Must hold list_lock since we modify the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001510 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001511static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001512 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001513 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001514{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001515 void *freelist;
1516 unsigned long counters;
1517 struct page new;
1518
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001519 /*
1520 * Zap the freelist and set the frozen bit.
1521 * The old freelist is the list of objects for the
1522 * per cpu allocation list.
1523 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001524 freelist = page->freelist;
1525 counters = page->counters;
1526 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001527 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001528 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001529 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001530 new.freelist = NULL;
1531 } else {
1532 new.freelist = freelist;
1533 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001534
Christoph Lameter7ced3712012-05-09 10:09:53 -05001535 VM_BUG_ON(new.frozen);
1536 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001537
Christoph Lameter7ced3712012-05-09 10:09:53 -05001538 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001539 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001540 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001541 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001542 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001543
1544 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001545 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001546 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001547}
1548
Joonsoo Kim633b0762013-01-21 17:01:25 +09001549static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001550static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001551
Christoph Lameter81819f02007-05-06 14:49:36 -07001552/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001553 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001554 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001555static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1556 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001557{
Christoph Lameter49e22582011-08-09 16:12:27 -05001558 struct page *page, *page2;
1559 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001560 int available = 0;
1561 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001562
1563 /*
1564 * Racy check. If we mistakenly see no partial slabs then we
1565 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001566 * partial slab and there is none available then get_partials()
1567 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001568 */
1569 if (!n || !n->nr_partial)
1570 return NULL;
1571
1572 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001573 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001574 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001575
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001576 if (!pfmemalloc_match(page, flags))
1577 continue;
1578
Joonsoo Kim633b0762013-01-21 17:01:25 +09001579 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001580 if (!t)
1581 break;
1582
Joonsoo Kim633b0762013-01-21 17:01:25 +09001583 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001584 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001585 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001586 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001587 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001588 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001589 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001590 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001591 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001592 if (!kmem_cache_has_cpu_partial(s)
1593 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001594 break;
1595
Christoph Lameter497b66f2011-08-09 16:12:26 -05001596 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001597 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001598 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001599}
1600
1601/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001602 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001603 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001604static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001605 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001606{
1607#ifdef CONFIG_NUMA
1608 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001609 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001610 struct zone *zone;
1611 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001612 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001613 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001614
1615 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001616 * The defrag ratio allows a configuration of the tradeoffs between
1617 * inter node defragmentation and node local allocations. A lower
1618 * defrag_ratio increases the tendency to do local allocations
1619 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001620 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001621 * If the defrag_ratio is set to 0 then kmalloc() always
1622 * returns node local objects. If the ratio is higher then kmalloc()
1623 * may return off node objects because partial slabs are obtained
1624 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001625 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001626 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001627 * defrag_ratio = 1000) then every (well almost) allocation will
1628 * first attempt to defrag slab caches on other nodes. This means
1629 * scanning over all nodes to look for partial slabs which may be
1630 * expensive if we do it every time we are trying to find a slab
1631 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001632 */
Christoph Lameter98246012008-01-07 23:20:26 -08001633 if (!s->remote_node_defrag_ratio ||
1634 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001635 return NULL;
1636
Mel Gormancc9a6c82012-03-21 16:34:11 -07001637 do {
1638 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07001639 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001640 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1641 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001642
Mel Gormancc9a6c82012-03-21 16:34:11 -07001643 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001644
Mel Gormancc9a6c82012-03-21 16:34:11 -07001645 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1646 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001647 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001648 if (object) {
1649 /*
1650 * Return the object even if
1651 * put_mems_allowed indicated that
1652 * the cpuset mems_allowed was
1653 * updated in parallel. It's a
1654 * harmless race between the alloc
1655 * and the cpuset update.
1656 */
1657 put_mems_allowed(cpuset_mems_cookie);
1658 return object;
1659 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001660 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001661 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001662 } while (!put_mems_allowed(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001663#endif
1664 return NULL;
1665}
1666
1667/*
1668 * Get a partial page, lock it and return it.
1669 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001670static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001671 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001672{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001673 void *object;
Christoph Lameter2154a332010-07-09 14:07:10 -05001674 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
Christoph Lameter81819f02007-05-06 14:49:36 -07001675
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001676 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001677 if (object || node != NUMA_NO_NODE)
1678 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001679
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001680 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001681}
1682
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001683#ifdef CONFIG_PREEMPT
1684/*
1685 * Calculate the next globally unique transaction for disambiguiation
1686 * during cmpxchg. The transactions start with the cpu number and are then
1687 * incremented by CONFIG_NR_CPUS.
1688 */
1689#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1690#else
1691/*
1692 * No preemption supported therefore also no need to check for
1693 * different cpus.
1694 */
1695#define TID_STEP 1
1696#endif
1697
1698static inline unsigned long next_tid(unsigned long tid)
1699{
1700 return tid + TID_STEP;
1701}
1702
1703static inline unsigned int tid_to_cpu(unsigned long tid)
1704{
1705 return tid % TID_STEP;
1706}
1707
1708static inline unsigned long tid_to_event(unsigned long tid)
1709{
1710 return tid / TID_STEP;
1711}
1712
1713static inline unsigned int init_tid(int cpu)
1714{
1715 return cpu;
1716}
1717
1718static inline void note_cmpxchg_failure(const char *n,
1719 const struct kmem_cache *s, unsigned long tid)
1720{
1721#ifdef SLUB_DEBUG_CMPXCHG
1722 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1723
1724 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1725
1726#ifdef CONFIG_PREEMPT
1727 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1728 printk("due to cpu change %d -> %d\n",
1729 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1730 else
1731#endif
1732 if (tid_to_event(tid) != tid_to_event(actual_tid))
1733 printk("due to cpu running other code. Event %ld->%ld\n",
1734 tid_to_event(tid), tid_to_event(actual_tid));
1735 else
1736 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1737 actual_tid, tid, next_tid(tid));
1738#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001739 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001740}
1741
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001742static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001743{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001744 int cpu;
1745
1746 for_each_possible_cpu(cpu)
1747 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001748}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001749
1750/*
1751 * Remove the cpu slab
1752 */
Chen Gangd0e0ac92013-07-15 09:05:29 +08001753static void deactivate_slab(struct kmem_cache *s, struct page *page,
1754 void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001755{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001756 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001757 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1758 int lock = 0;
1759 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001760 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001761 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001762 struct page new;
1763 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001764
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001765 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001766 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001767 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001768 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001769
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001770 /*
1771 * Stage one: Free all available per cpu objects back
1772 * to the page freelist while it is still frozen. Leave the
1773 * last one.
1774 *
1775 * There is no need to take the list->lock because the page
1776 * is still frozen.
1777 */
1778 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1779 void *prior;
1780 unsigned long counters;
1781
1782 do {
1783 prior = page->freelist;
1784 counters = page->counters;
1785 set_freepointer(s, freelist, prior);
1786 new.counters = counters;
1787 new.inuse--;
1788 VM_BUG_ON(!new.frozen);
1789
Christoph Lameter1d071712011-07-14 12:49:12 -05001790 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001791 prior, counters,
1792 freelist, new.counters,
1793 "drain percpu freelist"));
1794
1795 freelist = nextfree;
1796 }
1797
1798 /*
1799 * Stage two: Ensure that the page is unfrozen while the
1800 * list presence reflects the actual number of objects
1801 * during unfreeze.
1802 *
1803 * We setup the list membership and then perform a cmpxchg
1804 * with the count. If there is a mismatch then the page
1805 * is not unfrozen but the page is on the wrong list.
1806 *
1807 * Then we restart the process which may have to remove
1808 * the page from the list that we just put it on again
1809 * because the number of objects in the slab may have
1810 * changed.
1811 */
1812redo:
1813
1814 old.freelist = page->freelist;
1815 old.counters = page->counters;
1816 VM_BUG_ON(!old.frozen);
1817
1818 /* Determine target state of the slab */
1819 new.counters = old.counters;
1820 if (freelist) {
1821 new.inuse--;
1822 set_freepointer(s, freelist, old.freelist);
1823 new.freelist = freelist;
1824 } else
1825 new.freelist = old.freelist;
1826
1827 new.frozen = 0;
1828
Christoph Lameter81107182011-08-09 13:01:32 -05001829 if (!new.inuse && n->nr_partial > s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001830 m = M_FREE;
1831 else if (new.freelist) {
1832 m = M_PARTIAL;
1833 if (!lock) {
1834 lock = 1;
1835 /*
1836 * Taking the spinlock removes the possiblity
1837 * that acquire_slab() will see a slab page that
1838 * is frozen
1839 */
1840 spin_lock(&n->list_lock);
1841 }
1842 } else {
1843 m = M_FULL;
1844 if (kmem_cache_debug(s) && !lock) {
1845 lock = 1;
1846 /*
1847 * This also ensures that the scanning of full
1848 * slabs from diagnostic functions will not see
1849 * any frozen slabs.
1850 */
1851 spin_lock(&n->list_lock);
1852 }
1853 }
1854
1855 if (l != m) {
1856
1857 if (l == M_PARTIAL)
1858
1859 remove_partial(n, page);
1860
1861 else if (l == M_FULL)
1862
1863 remove_full(s, page);
1864
1865 if (m == M_PARTIAL) {
1866
1867 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001868 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001869
1870 } else if (m == M_FULL) {
1871
1872 stat(s, DEACTIVATE_FULL);
1873 add_full(s, n, page);
1874
1875 }
1876 }
1877
1878 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001879 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001880 old.freelist, old.counters,
1881 new.freelist, new.counters,
1882 "unfreezing slab"))
1883 goto redo;
1884
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001885 if (lock)
1886 spin_unlock(&n->list_lock);
1887
1888 if (m == M_FREE) {
1889 stat(s, DEACTIVATE_EMPTY);
1890 discard_slab(s, page);
1891 stat(s, FREE_SLAB);
1892 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001893}
1894
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001895/*
1896 * Unfreeze all the cpu partial slabs.
1897 *
Christoph Lameter59a09912012-11-28 16:23:00 +00001898 * This function must be called with interrupts disabled
1899 * for the cpu using c (or some other guarantee must be there
1900 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001901 */
Christoph Lameter59a09912012-11-28 16:23:00 +00001902static void unfreeze_partials(struct kmem_cache *s,
1903 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05001904{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001905#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09001906 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08001907 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001908
1909 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001910 struct page new;
1911 struct page old;
1912
1913 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001914
1915 n2 = get_node(s, page_to_nid(page));
1916 if (n != n2) {
1917 if (n)
1918 spin_unlock(&n->list_lock);
1919
1920 n = n2;
1921 spin_lock(&n->list_lock);
1922 }
Christoph Lameter49e22582011-08-09 16:12:27 -05001923
1924 do {
1925
1926 old.freelist = page->freelist;
1927 old.counters = page->counters;
1928 VM_BUG_ON(!old.frozen);
1929
1930 new.counters = old.counters;
1931 new.freelist = old.freelist;
1932
1933 new.frozen = 0;
1934
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001935 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05001936 old.freelist, old.counters,
1937 new.freelist, new.counters,
1938 "unfreezing slab"));
1939
Joonsoo Kim43d77862012-06-09 02:23:16 +09001940 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08001941 page->next = discard_page;
1942 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001943 } else {
1944 add_partial(n, page, DEACTIVATE_TO_TAIL);
1945 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001946 }
1947 }
1948
1949 if (n)
1950 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08001951
1952 while (discard_page) {
1953 page = discard_page;
1954 discard_page = discard_page->next;
1955
1956 stat(s, DEACTIVATE_EMPTY);
1957 discard_slab(s, page);
1958 stat(s, FREE_SLAB);
1959 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001960#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05001961}
1962
1963/*
1964 * Put a page that was just frozen (in __slab_free) into a partial page
1965 * slot if available. This is done without interrupts disabled and without
1966 * preemption disabled. The cmpxchg is racy and may put the partial page
1967 * onto a random cpus partial slot.
1968 *
1969 * If we did not find a slot then simply move all the partials to the
1970 * per node partial list.
1971 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09001972static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05001973{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001974#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05001975 struct page *oldpage;
1976 int pages;
1977 int pobjects;
1978
1979 do {
1980 pages = 0;
1981 pobjects = 0;
1982 oldpage = this_cpu_read(s->cpu_slab->partial);
1983
1984 if (oldpage) {
1985 pobjects = oldpage->pobjects;
1986 pages = oldpage->pages;
1987 if (drain && pobjects > s->cpu_partial) {
1988 unsigned long flags;
1989 /*
1990 * partial array is full. Move the existing
1991 * set to the per node partial list.
1992 */
1993 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00001994 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05001995 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09001996 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001997 pobjects = 0;
1998 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08001999 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05002000 }
2001 }
2002
2003 pages++;
2004 pobjects += page->objects - page->inuse;
2005
2006 page->pages = pages;
2007 page->pobjects = pobjects;
2008 page->next = oldpage;
2009
Chen Gangd0e0ac92013-07-15 09:05:29 +08002010 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2011 != oldpage);
Joonsoo Kim345c9052013-06-19 14:05:52 +09002012#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002013}
2014
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002015static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002016{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002017 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002018 deactivate_slab(s, c->page, c->freelist);
2019
2020 c->tid = next_tid(c->tid);
2021 c->page = NULL;
2022 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002023}
2024
2025/*
2026 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002027 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002028 * Called from IPI handler with interrupts disabled.
2029 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002030static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002031{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002032 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002033
Christoph Lameter49e22582011-08-09 16:12:27 -05002034 if (likely(c)) {
2035 if (c->page)
2036 flush_slab(s, c);
2037
Christoph Lameter59a09912012-11-28 16:23:00 +00002038 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002039 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002040}
2041
2042static void flush_cpu_slab(void *d)
2043{
2044 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002045
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002046 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002047}
2048
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002049static bool has_cpu_slab(int cpu, void *info)
2050{
2051 struct kmem_cache *s = info;
2052 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2053
majianpeng02e1a9c2012-05-17 17:03:26 -07002054 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002055}
2056
Christoph Lameter81819f02007-05-06 14:49:36 -07002057static void flush_all(struct kmem_cache *s)
2058{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002059 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002060}
2061
2062/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002063 * Check if the objects in a per cpu structure fit numa
2064 * locality expectations.
2065 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002066static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002067{
2068#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002069 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002070 return 0;
2071#endif
2072 return 1;
2073}
2074
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002075static int count_free(struct page *page)
2076{
2077 return page->objects - page->inuse;
2078}
2079
2080static unsigned long count_partial(struct kmem_cache_node *n,
2081 int (*get_count)(struct page *))
2082{
2083 unsigned long flags;
2084 unsigned long x = 0;
2085 struct page *page;
2086
2087 spin_lock_irqsave(&n->list_lock, flags);
2088 list_for_each_entry(page, &n->partial, lru)
2089 x += get_count(page);
2090 spin_unlock_irqrestore(&n->list_lock, flags);
2091 return x;
2092}
2093
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002094static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2095{
2096#ifdef CONFIG_SLUB_DEBUG
2097 return atomic_long_read(&n->total_objects);
2098#else
2099 return 0;
2100#endif
2101}
2102
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002103static noinline void
2104slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2105{
2106 int node;
2107
2108 printk(KERN_WARNING
2109 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2110 nid, gfpflags);
2111 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002112 "default order: %d, min order: %d\n", s->name, s->object_size,
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002113 s->size, oo_order(s->oo), oo_order(s->min));
2114
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002115 if (oo_order(s->min) > get_order(s->object_size))
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002116 printk(KERN_WARNING " %s debugging increased min order, use "
2117 "slub_debug=O to disable.\n", s->name);
2118
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002119 for_each_online_node(node) {
2120 struct kmem_cache_node *n = get_node(s, node);
2121 unsigned long nr_slabs;
2122 unsigned long nr_objs;
2123 unsigned long nr_free;
2124
2125 if (!n)
2126 continue;
2127
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002128 nr_free = count_partial(n, count_free);
2129 nr_slabs = node_nr_slabs(n);
2130 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002131
2132 printk(KERN_WARNING
2133 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2134 node, nr_slabs, nr_objs, nr_free);
2135 }
2136}
2137
Christoph Lameter497b66f2011-08-09 16:12:26 -05002138static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2139 int node, struct kmem_cache_cpu **pc)
2140{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002141 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002142 struct kmem_cache_cpu *c = *pc;
2143 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002144
Christoph Lameter188fd062012-05-09 10:09:55 -05002145 freelist = get_partial(s, flags, node, c);
2146
2147 if (freelist)
2148 return freelist;
2149
2150 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002151 if (page) {
2152 c = __this_cpu_ptr(s->cpu_slab);
2153 if (c->page)
2154 flush_slab(s, c);
2155
2156 /*
2157 * No other reference to the page yet so we can
2158 * muck around with it freely without cmpxchg
2159 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002160 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002161 page->freelist = NULL;
2162
2163 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002164 c->page = page;
2165 *pc = c;
2166 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002167 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002168
Christoph Lameter6faa6832012-05-09 10:09:51 -05002169 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002170}
2171
Mel Gorman072bb0a2012-07-31 16:43:58 -07002172static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2173{
2174 if (unlikely(PageSlabPfmemalloc(page)))
2175 return gfp_pfmemalloc_allowed(gfpflags);
2176
2177 return true;
2178}
2179
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002180/*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002181 * Check the page->freelist of a page and either transfer the freelist to the
2182 * per cpu freelist or deactivate the page.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002183 *
2184 * The page is still frozen if the return value is not NULL.
2185 *
2186 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002187 *
2188 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002189 */
2190static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2191{
2192 struct page new;
2193 unsigned long counters;
2194 void *freelist;
2195
2196 do {
2197 freelist = page->freelist;
2198 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002199
Christoph Lameter213eeb92011-11-11 14:07:14 -06002200 new.counters = counters;
2201 VM_BUG_ON(!new.frozen);
2202
2203 new.inuse = page->objects;
2204 new.frozen = freelist != NULL;
2205
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002206 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002207 freelist, counters,
2208 NULL, new.counters,
2209 "get_freelist"));
2210
2211 return freelist;
2212}
2213
2214/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002215 * Slow path. The lockless freelist is empty or we need to perform
2216 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002217 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002218 * Processing is still very fast if new objects have been freed to the
2219 * regular freelist. In that case we simply take over the regular freelist
2220 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002221 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002222 * If that is not working then we fall back to the partial lists. We take the
2223 * first element of the freelist as the object to allocate now and move the
2224 * rest of the freelist to the lockless freelist.
2225 *
2226 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002227 * we need to allocate a new slab. This is the slowest path since it involves
2228 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07002229 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002230static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2231 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002232{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002233 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002234 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002235 unsigned long flags;
2236
2237 local_irq_save(flags);
2238#ifdef CONFIG_PREEMPT
2239 /*
2240 * We may have been preempted and rescheduled on a different
2241 * cpu before disabling interrupts. Need to reload cpu area
2242 * pointer.
2243 */
2244 c = this_cpu_ptr(s->cpu_slab);
2245#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002246
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002247 page = c->page;
2248 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002249 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002250redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002251
Christoph Lameter57d437d2012-05-09 10:09:59 -05002252 if (unlikely(!node_match(page, node))) {
Christoph Lametere36a2652011-06-01 12:25:57 -05002253 stat(s, ALLOC_NODE_MISMATCH);
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002254 deactivate_slab(s, page, c->freelist);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002255 c->page = NULL;
2256 c->freelist = NULL;
Christoph Lameterfc59c052011-06-01 12:25:56 -05002257 goto new_slab;
2258 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002259
Mel Gorman072bb0a2012-07-31 16:43:58 -07002260 /*
2261 * By rights, we should be searching for a slab page that was
2262 * PFMEMALLOC but right now, we are losing the pfmemalloc
2263 * information when the page leaves the per-cpu allocator
2264 */
2265 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2266 deactivate_slab(s, page, c->freelist);
2267 c->page = NULL;
2268 c->freelist = NULL;
2269 goto new_slab;
2270 }
2271
Eric Dumazet73736e02011-12-13 04:57:06 +01002272 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002273 freelist = c->freelist;
2274 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002275 goto load_freelist;
2276
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002277 stat(s, ALLOC_SLOWPATH);
2278
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002279 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002280
Christoph Lameter6faa6832012-05-09 10:09:51 -05002281 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002282 c->page = NULL;
2283 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002284 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002285 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002286
Christoph Lameter81819f02007-05-06 14:49:36 -07002287 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002288
Christoph Lameter894b8782007-05-10 03:15:16 -07002289load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002290 /*
2291 * freelist is pointing to the list of objects to be used.
2292 * page is pointing to the page from which the objects are obtained.
2293 * That page must be frozen for per cpu allocations to work.
2294 */
2295 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002296 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002297 c->tid = next_tid(c->tid);
2298 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002299 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002300
Christoph Lameter81819f02007-05-06 14:49:36 -07002301new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002302
Christoph Lameter49e22582011-08-09 16:12:27 -05002303 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002304 page = c->page = c->partial;
2305 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002306 stat(s, CPU_PARTIAL_ALLOC);
2307 c->freelist = NULL;
2308 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002309 }
2310
Christoph Lameter188fd062012-05-09 10:09:55 -05002311 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002312
Christoph Lameterf46974362012-05-09 10:09:54 -05002313 if (unlikely(!freelist)) {
2314 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2315 slab_out_of_memory(s, gfpflags, node);
Christoph Lameter01ad8a72011-04-15 14:48:14 -05002316
Christoph Lameterf46974362012-05-09 10:09:54 -05002317 local_irq_restore(flags);
2318 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002319 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002320
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002321 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002322 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002323 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002324
Christoph Lameter497b66f2011-08-09 16:12:26 -05002325 /* Only entered in the debug case */
Chen Gangd0e0ac92013-07-15 09:05:29 +08002326 if (kmem_cache_debug(s) &&
2327 !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002328 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002329
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002330 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002331 c->page = NULL;
2332 c->freelist = NULL;
Christoph Lametera71ae472011-05-25 09:47:43 -05002333 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002334 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002335}
2336
2337/*
2338 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2339 * have the fastpath folded into their functions. So no function call
2340 * overhead for requests that can be satisfied on the fastpath.
2341 *
2342 * The fastpath works by first checking if the lockless freelist can be used.
2343 * If not then __slab_alloc is called for slow processing.
2344 *
2345 * Otherwise we can simply pick the next object from the lockless free list.
2346 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002347static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002348 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002349{
Christoph Lameter894b8782007-05-10 03:15:16 -07002350 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002351 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002352 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002353 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002354
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002355 if (slab_pre_alloc_hook(s, gfpflags))
Akinobu Mita773ff602008-12-23 19:37:01 +09002356 return NULL;
2357
Glauber Costad79923f2012-12-18 14:22:48 -08002358 s = memcg_kmem_get_cache(s, gfpflags);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002359redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002360 /*
2361 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2362 * enabled. We may switch back and forth between cpus while
2363 * reading from one cpu area. That does not matter as long
2364 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002365 *
2366 * Preemption is disabled for the retrieval of the tid because that
2367 * must occur from the current processor. We cannot allow rescheduling
2368 * on a different processor between the determination of the pointer
2369 * and the retrieval of the tid.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002370 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002371 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002372 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002373
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002374 /*
2375 * The transaction ids are globally unique per cpu and per operation on
2376 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2377 * occurs on the right processor and that there was no operation on the
2378 * linked list in between.
2379 */
2380 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002381 preempt_enable();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002382
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002383 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002384 page = c->page;
Libinac6434e2013-07-18 15:39:51 +08002385 if (unlikely(!object || !node_match(page, node)))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002386 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07002387
2388 else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002389 void *next_object = get_freepointer_safe(s, object);
2390
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002391 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002392 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002393 * operation and if we are on the right processor.
2394 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002395 * The cmpxchg does the following atomically (without lock
2396 * semantics!)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002397 * 1. Relocate first pointer to the current per cpu area.
2398 * 2. Verify that tid and freelist have not been changed
2399 * 3. If they were not changed replace tid and freelist
2400 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002401 * Since this is without lock semantics the protection is only
2402 * against code executing on this cpu *not* from access by
2403 * other cpus.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002404 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002405 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002406 s->cpu_slab->freelist, s->cpu_slab->tid,
2407 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002408 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002409
2410 note_cmpxchg_failure("slab_alloc", s, tid);
2411 goto redo;
2412 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002413 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002414 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002415 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002416
Pekka Enberg74e21342009-11-25 20:14:48 +02002417 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002418 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002419
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002420 slab_post_alloc_hook(s, gfpflags, object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002421
Christoph Lameter894b8782007-05-10 03:15:16 -07002422 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002423}
2424
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002425static __always_inline void *slab_alloc(struct kmem_cache *s,
2426 gfp_t gfpflags, unsigned long addr)
2427{
2428 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2429}
2430
Christoph Lameter81819f02007-05-06 14:49:36 -07002431void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2432{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002433 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002434
Chen Gangd0e0ac92013-07-15 09:05:29 +08002435 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2436 s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002437
2438 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002439}
2440EXPORT_SYMBOL(kmem_cache_alloc);
2441
Li Zefan0f24f122009-12-11 15:45:30 +08002442#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002443void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002444{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002445 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002446 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2447 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002448}
Richard Kennedy4a923792010-10-21 10:29:19 +01002449EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002450#endif
2451
Christoph Lameter81819f02007-05-06 14:49:36 -07002452#ifdef CONFIG_NUMA
2453void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2454{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002455 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002456
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002457 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002458 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002459
2460 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002461}
2462EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002463
Li Zefan0f24f122009-12-11 15:45:30 +08002464#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002465void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002466 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002467 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002468{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002469 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002470
2471 trace_kmalloc_node(_RET_IP_, ret,
2472 size, s->size, gfpflags, node);
2473 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002474}
Richard Kennedy4a923792010-10-21 10:29:19 +01002475EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002476#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002477#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002478
Christoph Lameter81819f02007-05-06 14:49:36 -07002479/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002480 * Slow patch handling. This may still be called frequently since objects
2481 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002482 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002483 * So we still attempt to reduce cache line usage. Just take the slab
2484 * lock and free the item. If there is no additional partial page
2485 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002486 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002487static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterff120592009-12-18 16:26:22 -06002488 void *x, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07002489{
2490 void *prior;
2491 void **object = (void *)x;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002492 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002493 struct page new;
2494 unsigned long counters;
2495 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002496 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002497
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002498 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002499
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002500 if (kmem_cache_debug(s) &&
2501 !(n = free_debug_processing(s, page, x, addr, &flags)))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002502 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002503
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002504 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002505 if (unlikely(n)) {
2506 spin_unlock_irqrestore(&n->list_lock, flags);
2507 n = NULL;
2508 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002509 prior = page->freelist;
2510 counters = page->counters;
2511 set_freepointer(s, object, prior);
2512 new.counters = counters;
2513 was_frozen = new.frozen;
2514 new.inuse--;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002515 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002516
Joonsoo Kim345c9052013-06-19 14:05:52 +09002517 if (kmem_cache_has_cpu_partial(s) && !prior)
Christoph Lameter49e22582011-08-09 16:12:27 -05002518
2519 /*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002520 * Slab was on no list before and will be
2521 * partially empty
2522 * We can defer the list move and instead
2523 * freeze it.
Christoph Lameter49e22582011-08-09 16:12:27 -05002524 */
2525 new.frozen = 1;
2526
2527 else { /* Needs to be taken off a list */
2528
2529 n = get_node(s, page_to_nid(page));
2530 /*
2531 * Speculatively acquire the list_lock.
2532 * If the cmpxchg does not succeed then we may
2533 * drop the list_lock without any processing.
2534 *
2535 * Otherwise the list_lock will synchronize with
2536 * other processors updating the list of slabs.
2537 */
2538 spin_lock_irqsave(&n->list_lock, flags);
2539
2540 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002541 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002542
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002543 } while (!cmpxchg_double_slab(s, page,
2544 prior, counters,
2545 object, new.counters,
2546 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002547
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002548 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002549
2550 /*
2551 * If we just froze the page then put it onto the
2552 * per cpu partial list.
2553 */
Alex Shi8028dce2012-02-03 23:34:56 +08002554 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002555 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002556 stat(s, CPU_PARTIAL_FREE);
2557 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002558 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002559 * The list lock was not taken therefore no list
2560 * activity can be necessary.
2561 */
2562 if (was_frozen)
2563 stat(s, FREE_FROZEN);
Christoph Lameter80f08c12011-06-01 12:25:55 -05002564 return;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002565 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002566
Joonsoo Kim837d6782012-08-16 00:02:40 +09002567 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2568 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002569
Joonsoo Kim837d6782012-08-16 00:02:40 +09002570 /*
2571 * Objects left in the slab. If it was not on the partial list before
2572 * then add it.
2573 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002574 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2575 if (kmem_cache_debug(s))
2576 remove_full(s, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002577 add_partial(n, page, DEACTIVATE_TO_TAIL);
2578 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002579 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002580 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002581 return;
2582
2583slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002584 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002585 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002586 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002587 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002588 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002589 stat(s, FREE_REMOVE_PARTIAL);
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002590 } else
2591 /* Slab must be on the full list */
2592 remove_full(s, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002593
Christoph Lameter80f08c12011-06-01 12:25:55 -05002594 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002595 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002596 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002597}
2598
Christoph Lameter894b8782007-05-10 03:15:16 -07002599/*
2600 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2601 * can perform fastpath freeing without additional function calls.
2602 *
2603 * The fastpath is only possible if we are freeing to the current cpu slab
2604 * of this processor. This typically the case if we have just allocated
2605 * the item before.
2606 *
2607 * If fastpath is not possible then fall back to __slab_free where we deal
2608 * with all sorts of special processing.
2609 */
Pekka Enberg06428782008-01-07 23:20:27 -08002610static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002611 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002612{
2613 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002614 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002615 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002616
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002617 slab_free_hook(s, x);
2618
Christoph Lametera24c5a02011-03-15 12:45:21 -05002619redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002620 /*
2621 * Determine the currently cpus per cpu slab.
2622 * The cpu may change afterward. However that does not matter since
2623 * data is retrieved via this pointer. If we are on the same cpu
2624 * during the cmpxchg then the free will succedd.
2625 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002626 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002627 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002628
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002629 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002630 preempt_enable();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002631
Christoph Lameter442b06b2011-05-17 16:29:31 -05002632 if (likely(page == c->page)) {
Christoph Lameterff120592009-12-18 16:26:22 -06002633 set_freepointer(s, object, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002634
Christoph Lameter933393f2011-12-22 11:58:51 -06002635 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002636 s->cpu_slab->freelist, s->cpu_slab->tid,
2637 c->freelist, tid,
2638 object, next_tid(tid)))) {
2639
2640 note_cmpxchg_failure("slab_free", s, tid);
2641 goto redo;
2642 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002643 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002644 } else
Christoph Lameterff120592009-12-18 16:26:22 -06002645 __slab_free(s, page, x, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002646
Christoph Lameter894b8782007-05-10 03:15:16 -07002647}
2648
Christoph Lameter81819f02007-05-06 14:49:36 -07002649void kmem_cache_free(struct kmem_cache *s, void *x)
2650{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002651 s = cache_from_obj(s, x);
2652 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002653 return;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002654 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002655 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002656}
2657EXPORT_SYMBOL(kmem_cache_free);
2658
Christoph Lameter81819f02007-05-06 14:49:36 -07002659/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002660 * Object placement in a slab is made very easy because we always start at
2661 * offset 0. If we tune the size of the object to the alignment then we can
2662 * get the required alignment by putting one properly sized object after
2663 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002664 *
2665 * Notice that the allocation order determines the sizes of the per cpu
2666 * caches. Each processor has always one slab available for allocations.
2667 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07002668 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07002669 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07002670 */
2671
2672/*
2673 * Mininum / Maximum order of slab pages. This influences locking overhead
2674 * and slab fragmentation. A higher order reduces the number of partial slabs
2675 * and increases the number of allocations possible without having to
2676 * take the list_lock.
2677 */
2678static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03002679static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002680static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002681
2682/*
2683 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07002684 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07002685 */
2686static int slub_nomerge;
2687
2688/*
Christoph Lameter81819f02007-05-06 14:49:36 -07002689 * Calculate the order of allocation given an slab object size.
2690 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002691 * The order of allocation has significant impact on performance and other
2692 * system components. Generally order 0 allocations should be preferred since
2693 * order 0 does not cause fragmentation in the page allocator. Larger objects
2694 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002695 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07002696 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07002697 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002698 * In order to reach satisfactory performance we must ensure that a minimum
2699 * number of objects is in one slab. Otherwise we may generate too much
2700 * activity on the partial lists which requires taking the list_lock. This is
2701 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07002702 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002703 * slub_max_order specifies the order where we begin to stop considering the
2704 * number of objects in a slab as critical. If we reach slub_max_order then
2705 * we try to keep the page order as low as possible. So we accept more waste
2706 * of space in favor of a small page order.
2707 *
2708 * Higher order allocations also allow the placement of more objects in a
2709 * slab and thereby reduce object handling overhead. If the user has
2710 * requested a higher mininum order then we start with that one instead of
2711 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07002712 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002713static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002714 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002715{
2716 int order;
2717 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07002718 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002719
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002720 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04002721 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03002722
Christoph Lameter6300ea72007-07-17 04:03:20 -07002723 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002724 fls(min_objects * size - 1) - PAGE_SHIFT);
2725 order <= max_order; order++) {
2726
Christoph Lameter81819f02007-05-06 14:49:36 -07002727 unsigned long slab_size = PAGE_SIZE << order;
2728
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002729 if (slab_size < min_objects * size + reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002730 continue;
2731
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002732 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002733
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002734 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07002735 break;
2736
2737 }
Christoph Lameter672bba32007-05-09 02:32:39 -07002738
Christoph Lameter81819f02007-05-06 14:49:36 -07002739 return order;
2740}
2741
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002742static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002743{
2744 int order;
2745 int min_objects;
2746 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002747 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002748
2749 /*
2750 * Attempt to find best configuration for a slab. This
2751 * works by first attempting to generate a layout with
2752 * the best configuration and backing off gradually.
2753 *
2754 * First we reduce the acceptable waste in a slab. Then
2755 * we reduce the minimum objects required in a slab.
2756 */
2757 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002758 if (!min_objects)
2759 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002760 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002761 min_objects = min(min_objects, max_objects);
2762
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002763 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002764 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002765 while (fraction >= 4) {
2766 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002767 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002768 if (order <= slub_max_order)
2769 return order;
2770 fraction /= 2;
2771 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03002772 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002773 }
2774
2775 /*
2776 * We were unable to place multiple objects in a slab. Now
2777 * lets see if we can place a single object there.
2778 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002779 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002780 if (order <= slub_max_order)
2781 return order;
2782
2783 /*
2784 * Doh this slab cannot be placed using slub_max_order.
2785 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002786 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03002787 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002788 return order;
2789 return -ENOSYS;
2790}
2791
Pekka Enberg5595cff2008-08-05 09:28:47 +03002792static void
Joonsoo Kim40534972012-05-11 00:50:47 +09002793init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002794{
2795 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002796 spin_lock_init(&n->list_lock);
2797 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002798#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002799 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07002800 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07002801 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002802#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002803}
2804
Christoph Lameter55136592010-08-20 12:37:13 -05002805static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002806{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05002807 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00002808 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002809
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002810 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002811 * Must align to double word boundary for the double cmpxchg
2812 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002813 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002814 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2815 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002816
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002817 if (!s->cpu_slab)
2818 return 0;
2819
2820 init_kmem_cache_cpus(s);
2821
2822 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002823}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002824
Christoph Lameter51df1142010-08-20 12:37:15 -05002825static struct kmem_cache *kmem_cache_node;
2826
Christoph Lameter81819f02007-05-06 14:49:36 -07002827/*
2828 * No kmalloc_node yet so do it by hand. We know that this is the first
2829 * slab on the node for this slabcache. There are no concurrent accesses
2830 * possible.
2831 *
2832 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002833 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2834 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002835 */
Christoph Lameter55136592010-08-20 12:37:13 -05002836static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002837{
2838 struct page *page;
2839 struct kmem_cache_node *n;
2840
Christoph Lameter51df1142010-08-20 12:37:15 -05002841 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07002842
Christoph Lameter51df1142010-08-20 12:37:15 -05002843 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002844
2845 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002846 if (page_to_nid(page) != node) {
2847 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2848 "node %d\n", node);
2849 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2850 "in order to be able to continue\n");
2851 }
2852
Christoph Lameter81819f02007-05-06 14:49:36 -07002853 n = page->freelist;
2854 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002855 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05002856 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05002857 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05002858 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002859#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05002860 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05002861 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002862#endif
Joonsoo Kim40534972012-05-11 00:50:47 +09002863 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002864 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002865
Shaohua Li136333d2011-08-24 08:57:52 +08002866 add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07002867}
2868
2869static void free_kmem_cache_nodes(struct kmem_cache *s)
2870{
2871 int node;
2872
Christoph Lameterf64dc582007-10-16 01:25:33 -07002873 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002874 struct kmem_cache_node *n = s->node[node];
Christoph Lameter51df1142010-08-20 12:37:15 -05002875
Alexander Duyck73367bd2010-05-21 14:41:35 -07002876 if (n)
Christoph Lameter51df1142010-08-20 12:37:15 -05002877 kmem_cache_free(kmem_cache_node, n);
2878
Christoph Lameter81819f02007-05-06 14:49:36 -07002879 s->node[node] = NULL;
2880 }
2881}
2882
Christoph Lameter55136592010-08-20 12:37:13 -05002883static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002884{
2885 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07002886
Christoph Lameterf64dc582007-10-16 01:25:33 -07002887 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002888 struct kmem_cache_node *n;
2889
Alexander Duyck73367bd2010-05-21 14:41:35 -07002890 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05002891 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002892 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07002893 }
Christoph Lameter51df1142010-08-20 12:37:15 -05002894 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05002895 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002896
2897 if (!n) {
2898 free_kmem_cache_nodes(s);
2899 return 0;
2900 }
2901
Christoph Lameter81819f02007-05-06 14:49:36 -07002902 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09002903 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07002904 }
2905 return 1;
2906}
Christoph Lameter81819f02007-05-06 14:49:36 -07002907
David Rientjesc0bdb232009-02-25 09:16:35 +02002908static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002909{
2910 if (min < MIN_PARTIAL)
2911 min = MIN_PARTIAL;
2912 else if (min > MAX_PARTIAL)
2913 min = MAX_PARTIAL;
2914 s->min_partial = min;
2915}
2916
Christoph Lameter81819f02007-05-06 14:49:36 -07002917/*
2918 * calculate_sizes() determines the order and the distribution of data within
2919 * a slab object.
2920 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002921static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002922{
2923 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002924 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002925 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002926
2927 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002928 * Round up object size to the next word boundary. We can only
2929 * place the free pointer at word boundaries and this determines
2930 * the possible location of the free pointer.
2931 */
2932 size = ALIGN(size, sizeof(void *));
2933
2934#ifdef CONFIG_SLUB_DEBUG
2935 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002936 * Determine if we can poison the object itself. If the user of
2937 * the slab may touch the object after free or before allocation
2938 * then we should never poison the object itself.
2939 */
2940 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002941 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002942 s->flags |= __OBJECT_POISON;
2943 else
2944 s->flags &= ~__OBJECT_POISON;
2945
Christoph Lameter81819f02007-05-06 14:49:36 -07002946
2947 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002948 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002949 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002950 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002951 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002952 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07002953 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002954#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002955
2956 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002957 * With that we have determined the number of bytes in actual use
2958 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002959 */
2960 s->inuse = size;
2961
2962 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002963 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002964 /*
2965 * Relocate free pointer after the object if it is not
2966 * permitted to overwrite the first word of the object on
2967 * kmem_cache_free.
2968 *
2969 * This is the case if we do RCU, have a constructor or
2970 * destructor or are poisoning the objects.
2971 */
2972 s->offset = size;
2973 size += sizeof(void *);
2974 }
2975
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002976#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002977 if (flags & SLAB_STORE_USER)
2978 /*
2979 * Need to store information about allocs and frees after
2980 * the object.
2981 */
2982 size += 2 * sizeof(struct track);
2983
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002984 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002985 /*
2986 * Add some empty padding so that we can catch
2987 * overwrites from earlier objects rather than let
2988 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01002989 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07002990 * of the object.
2991 */
2992 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002993#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002994
Christoph Lameter81819f02007-05-06 14:49:36 -07002995 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002996 * SLUB stores one object immediately after another beginning from
2997 * offset 0. In order to align the objects we have to simply size
2998 * each object to conform to the alignment.
2999 */
Christoph Lameter45906852012-11-28 16:23:16 +00003000 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07003001 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003002 if (forced_order >= 0)
3003 order = forced_order;
3004 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003005 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07003006
Christoph Lameter834f3d12008-04-14 19:11:31 +03003007 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003008 return 0;
3009
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003010 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003011 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003012 s->allocflags |= __GFP_COMP;
3013
3014 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003015 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003016
3017 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3018 s->allocflags |= __GFP_RECLAIMABLE;
3019
Christoph Lameter81819f02007-05-06 14:49:36 -07003020 /*
3021 * Determine the number of objects per slab
3022 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003023 s->oo = oo_make(order, size, s->reserved);
3024 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003025 if (oo_objects(s->oo) > oo_objects(s->max))
3026 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003027
Christoph Lameter834f3d12008-04-14 19:11:31 +03003028 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003029}
3030
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003031static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003032{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003033 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003034 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003035
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003036 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3037 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003038
Christoph Lameter06b285d2008-04-14 19:11:41 +03003039 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003040 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003041 if (disable_higher_order_debug) {
3042 /*
3043 * Disable debugging flags that store metadata if the min slab
3044 * order increased.
3045 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003046 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003047 s->flags &= ~DEBUG_METADATA_FLAGS;
3048 s->offset = 0;
3049 if (!calculate_sizes(s, -1))
3050 goto error;
3051 }
3052 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003053
Heiko Carstens25654092012-01-12 17:17:33 -08003054#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3055 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003056 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3057 /* Enable fast mode */
3058 s->flags |= __CMPXCHG_DOUBLE;
3059#endif
3060
David Rientjes3b89d7d2009-02-22 17:40:07 -08003061 /*
3062 * The larger the object size is, the more pages we want on the partial
3063 * list to avoid pounding the page allocator excessively.
3064 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003065 set_min_partial(s, ilog2(s->size) / 2);
3066
3067 /*
3068 * cpu_partial determined the maximum number of objects kept in the
3069 * per cpu partial lists of a processor.
3070 *
3071 * Per cpu partial lists mainly contain slabs that just have one
3072 * object freed. If they are used for allocation then they can be
3073 * filled up again with minimal effort. The slab will never hit the
3074 * per node partial lists and therefore no locking will be required.
3075 *
3076 * This setting also determines
3077 *
3078 * A) The number of objects from per cpu partial slabs dumped to the
3079 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003080 * B) The number of objects in cpu partial slabs to extract from the
Chen Gangd0e0ac92013-07-15 09:05:29 +08003081 * per node list when we run out of per cpu objects. We only fetch
3082 * 50% to keep some capacity around for frees.
Christoph Lameter49e22582011-08-09 16:12:27 -05003083 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003084 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003085 s->cpu_partial = 0;
3086 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003087 s->cpu_partial = 2;
3088 else if (s->size >= 1024)
3089 s->cpu_partial = 6;
3090 else if (s->size >= 256)
3091 s->cpu_partial = 13;
3092 else
3093 s->cpu_partial = 30;
3094
Christoph Lameter81819f02007-05-06 14:49:36 -07003095#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003096 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003097#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003098 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003099 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003100
Christoph Lameter55136592010-08-20 12:37:13 -05003101 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003102 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003103
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003104 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003105error:
3106 if (flags & SLAB_PANIC)
3107 panic("Cannot create slab %s size=%lu realsize=%u "
3108 "order=%u offset=%u flags=%lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +08003109 s->name, (unsigned long)s->size, s->size,
3110 oo_order(s->oo), s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003111 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003112}
Christoph Lameter81819f02007-05-06 14:49:36 -07003113
Christoph Lameter33b12c32008-04-25 12:22:43 -07003114static void list_slab_objects(struct kmem_cache *s, struct page *page,
3115 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003116{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003117#ifdef CONFIG_SLUB_DEBUG
3118 void *addr = page_address(page);
3119 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003120 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3121 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003122 if (!map)
3123 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003124 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003125 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003126
Christoph Lameter5f80b132011-04-15 14:48:13 -05003127 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003128 for_each_object(p, s, addr, page->objects) {
3129
3130 if (!test_bit(slab_index(p, s, addr), map)) {
3131 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3132 p, p - addr);
3133 print_tracking(s, p);
3134 }
3135 }
3136 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003137 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003138#endif
3139}
3140
Christoph Lameter81819f02007-05-06 14:49:36 -07003141/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003142 * Attempt to free all partial slabs on a node.
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003143 * This is called from kmem_cache_close(). We must be the last thread
3144 * using the cache and therefore we do not need to lock anymore.
Christoph Lameter81819f02007-05-06 14:49:36 -07003145 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003146static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003147{
Christoph Lameter81819f02007-05-06 14:49:36 -07003148 struct page *page, *h;
3149
Christoph Lameter33b12c32008-04-25 12:22:43 -07003150 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003151 if (!page->inuse) {
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05003152 remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003153 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003154 } else {
3155 list_slab_objects(s, page,
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003156 "Objects remaining in %s on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003157 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003158 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003159}
3160
3161/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003162 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003163 */
Christoph Lameter0c710012007-07-17 04:03:24 -07003164static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003165{
3166 int node;
3167
3168 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003169 /* Attempt to free all objects */
Christoph Lameterf64dc582007-10-16 01:25:33 -07003170 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003171 struct kmem_cache_node *n = get_node(s, node);
3172
Christoph Lameter599870b2008-04-23 12:36:52 -07003173 free_partial(s, n);
3174 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003175 return 1;
3176 }
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003177 free_percpu(s->cpu_slab);
Christoph Lameter81819f02007-05-06 14:49:36 -07003178 free_kmem_cache_nodes(s);
3179 return 0;
3180}
3181
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003182int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003183{
Christoph Lameter12c36672012-09-04 23:38:33 +00003184 int rc = kmem_cache_close(s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003185
Glauber Costa5413dfb2012-12-18 14:23:13 -08003186 if (!rc) {
3187 /*
3188 * We do the same lock strategy around sysfs_slab_add, see
3189 * __kmem_cache_create. Because this is pretty much the last
3190 * operation we do and the lock will be released shortly after
3191 * that in slab_common.c, we could just move sysfs_slab_remove
3192 * to a later point in common code. We should do that when we
3193 * have a common sysfs framework for all allocators.
3194 */
3195 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003196 sysfs_slab_remove(s);
Glauber Costa5413dfb2012-12-18 14:23:13 -08003197 mutex_lock(&slab_mutex);
3198 }
Christoph Lameter12c36672012-09-04 23:38:33 +00003199
3200 return rc;
Christoph Lameter81819f02007-05-06 14:49:36 -07003201}
Christoph Lameter81819f02007-05-06 14:49:36 -07003202
3203/********************************************************************
3204 * Kmalloc subsystem
3205 *******************************************************************/
3206
Christoph Lameter81819f02007-05-06 14:49:36 -07003207static int __init setup_slub_min_order(char *str)
3208{
Pekka Enberg06428782008-01-07 23:20:27 -08003209 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003210
3211 return 1;
3212}
3213
3214__setup("slub_min_order=", setup_slub_min_order);
3215
3216static int __init setup_slub_max_order(char *str)
3217{
Pekka Enberg06428782008-01-07 23:20:27 -08003218 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003219 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003220
3221 return 1;
3222}
3223
3224__setup("slub_max_order=", setup_slub_max_order);
3225
3226static int __init setup_slub_min_objects(char *str)
3227{
Pekka Enberg06428782008-01-07 23:20:27 -08003228 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003229
3230 return 1;
3231}
3232
3233__setup("slub_min_objects=", setup_slub_min_objects);
3234
3235static int __init setup_slub_nomerge(char *str)
3236{
3237 slub_nomerge = 1;
3238 return 1;
3239}
3240
3241__setup("slub_nomerge", setup_slub_nomerge);
3242
Christoph Lameter81819f02007-05-06 14:49:36 -07003243void *__kmalloc(size_t size, gfp_t flags)
3244{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003245 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003246 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003247
Christoph Lameter95a05b42013-01-10 19:14:19 +00003248 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003249 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003250
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003251 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003252
3253 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003254 return s;
3255
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003256 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003257
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003258 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003259
3260 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003261}
3262EXPORT_SYMBOL(__kmalloc);
3263
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003264#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003265static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3266{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003267 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003268 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003269
Glauber Costad79923f2012-12-18 14:22:48 -08003270 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003271 page = alloc_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003272 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003273 ptr = page_address(page);
3274
3275 kmemleak_alloc(ptr, size, 1, flags);
3276 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003277}
3278
Christoph Lameter81819f02007-05-06 14:49:36 -07003279void *__kmalloc_node(size_t size, gfp_t flags, int node)
3280{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003281 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003282 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003283
Christoph Lameter95a05b42013-01-10 19:14:19 +00003284 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003285 ret = kmalloc_large_node(size, flags, node);
3286
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003287 trace_kmalloc_node(_RET_IP_, ret,
3288 size, PAGE_SIZE << get_order(size),
3289 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003290
3291 return ret;
3292 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003293
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003294 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003295
3296 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003297 return s;
3298
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003299 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003300
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003301 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003302
3303 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003304}
3305EXPORT_SYMBOL(__kmalloc_node);
3306#endif
3307
3308size_t ksize(const void *object)
3309{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003310 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003311
Christoph Lameteref8b4522007-10-16 01:24:46 -07003312 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003313 return 0;
3314
Vegard Nossum294a80a2007-12-04 23:45:30 -08003315 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003316
Pekka Enberg76994412008-05-22 19:22:25 +03003317 if (unlikely(!PageSlab(page))) {
3318 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003319 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003320 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003321
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003322 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003323}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003324EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003325
3326void kfree(const void *x)
3327{
Christoph Lameter81819f02007-05-06 14:49:36 -07003328 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003329 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003330
Pekka Enberg2121db72009-03-25 11:05:57 +02003331 trace_kfree(_RET_IP_, x);
3332
Satyam Sharma2408c552007-10-16 01:24:44 -07003333 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003334 return;
3335
Christoph Lameterb49af682007-05-06 14:49:41 -07003336 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003337 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003338 BUG_ON(!PageCompound(page));
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003339 kmemleak_free(x);
Glauber Costad79923f2012-12-18 14:22:48 -08003340 __free_memcg_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003341 return;
3342 }
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003343 slab_free(page->slab_cache, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003344}
3345EXPORT_SYMBOL(kfree);
3346
Christoph Lameter2086d262007-05-06 14:49:46 -07003347/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003348 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3349 * the remaining slabs by the number of items in use. The slabs with the
3350 * most items in use come first. New allocations will then fill those up
3351 * and thus they can be removed from the partial lists.
3352 *
3353 * The slabs with the least items are placed last. This results in them
3354 * being allocated from last increasing the chance that the last objects
3355 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003356 */
3357int kmem_cache_shrink(struct kmem_cache *s)
3358{
3359 int node;
3360 int i;
3361 struct kmem_cache_node *n;
3362 struct page *page;
3363 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03003364 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07003365 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03003366 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07003367 unsigned long flags;
3368
3369 if (!slabs_by_inuse)
3370 return -ENOMEM;
3371
3372 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003373 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07003374 n = get_node(s, node);
3375
3376 if (!n->nr_partial)
3377 continue;
3378
Christoph Lameter834f3d12008-04-14 19:11:31 +03003379 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07003380 INIT_LIST_HEAD(slabs_by_inuse + i);
3381
3382 spin_lock_irqsave(&n->list_lock, flags);
3383
3384 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003385 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07003386 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003387 * Note that concurrent frees may occur while we hold the
3388 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003389 */
3390 list_for_each_entry_safe(page, t, &n->partial, lru) {
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003391 list_move(&page->lru, slabs_by_inuse + page->inuse);
3392 if (!page->inuse)
3393 n->nr_partial--;
Christoph Lameter2086d262007-05-06 14:49:46 -07003394 }
3395
Christoph Lameter2086d262007-05-06 14:49:46 -07003396 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003397 * Rebuild the partial list with the slabs filled up most
3398 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07003399 */
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003400 for (i = objects - 1; i > 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07003401 list_splice(slabs_by_inuse + i, n->partial.prev);
3402
Christoph Lameter2086d262007-05-06 14:49:46 -07003403 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003404
3405 /* Release empty slabs */
3406 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3407 discard_slab(s, page);
Christoph Lameter2086d262007-05-06 14:49:46 -07003408 }
3409
3410 kfree(slabs_by_inuse);
3411 return 0;
3412}
3413EXPORT_SYMBOL(kmem_cache_shrink);
3414
Yasunori Gotob9049e22007-10-21 16:41:37 -07003415static int slab_mem_going_offline_callback(void *arg)
3416{
3417 struct kmem_cache *s;
3418
Christoph Lameter18004c52012-07-06 15:25:12 -05003419 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003420 list_for_each_entry(s, &slab_caches, list)
3421 kmem_cache_shrink(s);
Christoph Lameter18004c52012-07-06 15:25:12 -05003422 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003423
3424 return 0;
3425}
3426
3427static void slab_mem_offline_callback(void *arg)
3428{
3429 struct kmem_cache_node *n;
3430 struct kmem_cache *s;
3431 struct memory_notify *marg = arg;
3432 int offline_node;
3433
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003434 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003435
3436 /*
3437 * If the node still has available memory. we need kmem_cache_node
3438 * for it yet.
3439 */
3440 if (offline_node < 0)
3441 return;
3442
Christoph Lameter18004c52012-07-06 15:25:12 -05003443 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003444 list_for_each_entry(s, &slab_caches, list) {
3445 n = get_node(s, offline_node);
3446 if (n) {
3447 /*
3448 * if n->nr_slabs > 0, slabs still exist on the node
3449 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003450 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003451 * callback. So, we must fail.
3452 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003453 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003454
3455 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003456 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003457 }
3458 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003459 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003460}
3461
3462static int slab_mem_going_online_callback(void *arg)
3463{
3464 struct kmem_cache_node *n;
3465 struct kmem_cache *s;
3466 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003467 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003468 int ret = 0;
3469
3470 /*
3471 * If the node's memory is already available, then kmem_cache_node is
3472 * already created. Nothing to do.
3473 */
3474 if (nid < 0)
3475 return 0;
3476
3477 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003478 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003479 * allocate a kmem_cache_node structure in order to bring the node
3480 * online.
3481 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003482 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003483 list_for_each_entry(s, &slab_caches, list) {
3484 /*
3485 * XXX: kmem_cache_alloc_node will fallback to other nodes
3486 * since memory is not yet available from the node that
3487 * is brought up.
3488 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003489 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003490 if (!n) {
3491 ret = -ENOMEM;
3492 goto out;
3493 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003494 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003495 s->node[nid] = n;
3496 }
3497out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003498 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003499 return ret;
3500}
3501
3502static int slab_memory_callback(struct notifier_block *self,
3503 unsigned long action, void *arg)
3504{
3505 int ret = 0;
3506
3507 switch (action) {
3508 case MEM_GOING_ONLINE:
3509 ret = slab_mem_going_online_callback(arg);
3510 break;
3511 case MEM_GOING_OFFLINE:
3512 ret = slab_mem_going_offline_callback(arg);
3513 break;
3514 case MEM_OFFLINE:
3515 case MEM_CANCEL_ONLINE:
3516 slab_mem_offline_callback(arg);
3517 break;
3518 case MEM_ONLINE:
3519 case MEM_CANCEL_OFFLINE:
3520 break;
3521 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003522 if (ret)
3523 ret = notifier_from_errno(ret);
3524 else
3525 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003526 return ret;
3527}
3528
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003529static struct notifier_block slab_memory_callback_nb = {
3530 .notifier_call = slab_memory_callback,
3531 .priority = SLAB_CALLBACK_PRI,
3532};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003533
Christoph Lameter81819f02007-05-06 14:49:36 -07003534/********************************************************************
3535 * Basic setup of slabs
3536 *******************************************************************/
3537
Christoph Lameter51df1142010-08-20 12:37:15 -05003538/*
3539 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003540 * the page allocator. Allocate them properly then fix up the pointers
3541 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003542 */
3543
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003544static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003545{
3546 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003547 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameter51df1142010-08-20 12:37:15 -05003548
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003549 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003550
Glauber Costa7d557b32013-02-22 20:20:00 +04003551 /*
3552 * This runs very early, and only the boot processor is supposed to be
3553 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3554 * IPIs around.
3555 */
3556 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter51df1142010-08-20 12:37:15 -05003557 for_each_node_state(node, N_NORMAL_MEMORY) {
3558 struct kmem_cache_node *n = get_node(s, node);
3559 struct page *p;
3560
3561 if (n) {
3562 list_for_each_entry(p, &n->partial, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003563 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003564
Li Zefan607bf322011-04-12 15:22:26 +08003565#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter51df1142010-08-20 12:37:15 -05003566 list_for_each_entry(p, &n->full, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003567 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003568#endif
3569 }
3570 }
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003571 list_add(&s->list, &slab_caches);
3572 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003573}
3574
Christoph Lameter81819f02007-05-06 14:49:36 -07003575void __init kmem_cache_init(void)
3576{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003577 static __initdata struct kmem_cache boot_kmem_cache,
3578 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003579
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003580 if (debug_guardpage_minorder())
3581 slub_max_order = 0;
3582
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003583 kmem_cache_node = &boot_kmem_cache_node;
3584 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003585
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003586 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3587 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003588
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003589 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003590
3591 /* Able to allocate the per node structures */
3592 slab_state = PARTIAL;
3593
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003594 create_boot_cache(kmem_cache, "kmem_cache",
3595 offsetof(struct kmem_cache, node) +
3596 nr_node_ids * sizeof(struct kmem_cache_node *),
3597 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003598
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003599 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003600
Christoph Lameter51df1142010-08-20 12:37:15 -05003601 /*
3602 * Allocate kmem_cache_node properly from the kmem_cache slab.
3603 * kmem_cache_node is separately allocated so no need to
3604 * update any list pointers.
3605 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003606 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003607
3608 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003609 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003610
3611#ifdef CONFIG_SMP
3612 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003613#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003614
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003615 printk(KERN_INFO
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003616 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003617 " CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003618 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003619 slub_min_order, slub_max_order, slub_min_objects,
3620 nr_cpu_ids, nr_node_ids);
3621}
3622
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003623void __init kmem_cache_init_late(void)
3624{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003625}
3626
Christoph Lameter81819f02007-05-06 14:49:36 -07003627/*
3628 * Find a mergeable slab cache
3629 */
3630static int slab_unmergeable(struct kmem_cache *s)
3631{
3632 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3633 return 1;
3634
Christoph Lameterc59def92007-05-16 22:10:50 -07003635 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003636 return 1;
3637
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003638 /*
3639 * We may have set a slab to be unmergeable during bootstrap.
3640 */
3641 if (s->refcount < 0)
3642 return 1;
3643
Christoph Lameter81819f02007-05-06 14:49:36 -07003644 return 0;
3645}
3646
Glauber Costa2633d7a2012-12-18 14:22:34 -08003647static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003648 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003649 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003650{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003651 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003652
3653 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3654 return NULL;
3655
Christoph Lameterc59def92007-05-16 22:10:50 -07003656 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003657 return NULL;
3658
3659 size = ALIGN(size, sizeof(void *));
3660 align = calculate_alignment(flags, align, size);
3661 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003662 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003663
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003664 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003665 if (slab_unmergeable(s))
3666 continue;
3667
3668 if (size > s->size)
3669 continue;
3670
Christoph Lameterba0268a2007-09-11 15:24:11 -07003671 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003672 continue;
3673 /*
3674 * Check if alignment is compatible.
3675 * Courtesy of Adrian Drzewiecki
3676 */
Pekka Enberg06428782008-01-07 23:20:27 -08003677 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003678 continue;
3679
3680 if (s->size - size >= sizeof(void *))
3681 continue;
3682
Glauber Costa2633d7a2012-12-18 14:22:34 -08003683 if (!cache_match_memcg(s, memcg))
3684 continue;
3685
Christoph Lameter81819f02007-05-06 14:49:36 -07003686 return s;
3687 }
3688 return NULL;
3689}
3690
Glauber Costa2633d7a2012-12-18 14:22:34 -08003691struct kmem_cache *
3692__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3693 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003694{
3695 struct kmem_cache *s;
3696
Glauber Costa2633d7a2012-12-18 14:22:34 -08003697 s = find_mergeable(memcg, size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003698 if (s) {
3699 s->refcount++;
3700 /*
3701 * Adjust the object sizes so that we clear
3702 * the complete object on kzalloc.
3703 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003704 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003705 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003706
David Rientjes7b8f3b62008-12-17 22:09:46 -08003707 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003708 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003709 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003710 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003711 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003712
Christoph Lametercbb79692012-09-05 00:18:32 +00003713 return s;
3714}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003715
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003716int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003717{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003718 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003719
Pekka Enbergaac3a162012-09-05 12:07:44 +03003720 err = kmem_cache_open(s, flags);
3721 if (err)
3722 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003723
Christoph Lameter45530c42012-11-28 16:23:07 +00003724 /* Mutex is not taken during early boot */
3725 if (slab_state <= UP)
3726 return 0;
3727
Glauber Costa107dab52012-12-18 14:23:05 -08003728 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003729 mutex_unlock(&slab_mutex);
3730 err = sysfs_slab_add(s);
3731 mutex_lock(&slab_mutex);
Christoph Lameter20cea962012-07-06 15:25:13 -05003732
Pekka Enbergaac3a162012-09-05 12:07:44 +03003733 if (err)
3734 kmem_cache_close(s);
3735
3736 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07003737}
Christoph Lameter81819f02007-05-06 14:49:36 -07003738
Christoph Lameter81819f02007-05-06 14:49:36 -07003739#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003740/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003741 * Use the cpu notifier to insure that the cpu slabs are flushed when
3742 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003743 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04003744static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07003745 unsigned long action, void *hcpu)
3746{
3747 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003748 struct kmem_cache *s;
3749 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003750
3751 switch (action) {
3752 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003753 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003754 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003755 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05003756 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003757 list_for_each_entry(s, &slab_caches, list) {
3758 local_irq_save(flags);
3759 __flush_cpu_slab(s, cpu);
3760 local_irq_restore(flags);
3761 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003762 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003763 break;
3764 default:
3765 break;
3766 }
3767 return NOTIFY_OK;
3768}
3769
Paul Gortmaker0db06282013-06-19 14:53:51 -04003770static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003771 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003772};
Christoph Lameter81819f02007-05-06 14:49:36 -07003773
3774#endif
3775
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003776void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003777{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003778 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003779 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003780
Christoph Lameter95a05b42013-01-10 19:14:19 +00003781 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003782 return kmalloc_large(size, gfpflags);
3783
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003784 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003785
Satyam Sharma2408c552007-10-16 01:24:44 -07003786 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003787 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003788
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003789 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003790
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003791 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003792 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003793
3794 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003795}
3796
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003797#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07003798void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003799 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003800{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003801 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003802 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003803
Christoph Lameter95a05b42013-01-10 19:14:19 +00003804 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08003805 ret = kmalloc_large_node(size, gfpflags, node);
3806
3807 trace_kmalloc_node(caller, ret,
3808 size, PAGE_SIZE << get_order(size),
3809 gfpflags, node);
3810
3811 return ret;
3812 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02003813
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003814 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003815
Satyam Sharma2408c552007-10-16 01:24:44 -07003816 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003817 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003818
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003819 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003820
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003821 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003822 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003823
3824 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003825}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003826#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003827
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003828#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03003829static int count_inuse(struct page *page)
3830{
3831 return page->inuse;
3832}
3833
3834static int count_total(struct page *page)
3835{
3836 return page->objects;
3837}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003838#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03003839
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003840#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07003841static int validate_slab(struct kmem_cache *s, struct page *page,
3842 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003843{
3844 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003845 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003846
3847 if (!check_slab(s, page) ||
3848 !on_freelist(s, page, NULL))
3849 return 0;
3850
3851 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003852 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003853
Christoph Lameter5f80b132011-04-15 14:48:13 -05003854 get_map(s, page, map);
3855 for_each_object(p, s, addr, page->objects) {
3856 if (test_bit(slab_index(p, s, addr), map))
3857 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3858 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003859 }
3860
Christoph Lameter224a88b2008-04-14 19:11:31 +03003861 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003862 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02003863 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003864 return 0;
3865 return 1;
3866}
3867
Christoph Lameter434e2452007-07-17 04:03:30 -07003868static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3869 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003870{
Christoph Lameter881db7f2011-06-01 12:25:53 -05003871 slab_lock(page);
3872 validate_slab(s, page, map);
3873 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003874}
3875
Christoph Lameter434e2452007-07-17 04:03:30 -07003876static int validate_slab_node(struct kmem_cache *s,
3877 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003878{
3879 unsigned long count = 0;
3880 struct page *page;
3881 unsigned long flags;
3882
3883 spin_lock_irqsave(&n->list_lock, flags);
3884
3885 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003886 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003887 count++;
3888 }
3889 if (count != n->nr_partial)
3890 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3891 "counter=%ld\n", s->name, count, n->nr_partial);
3892
3893 if (!(s->flags & SLAB_STORE_USER))
3894 goto out;
3895
3896 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003897 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003898 count++;
3899 }
3900 if (count != atomic_long_read(&n->nr_slabs))
3901 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3902 "counter=%ld\n", s->name, count,
3903 atomic_long_read(&n->nr_slabs));
3904
3905out:
3906 spin_unlock_irqrestore(&n->list_lock, flags);
3907 return count;
3908}
3909
Christoph Lameter434e2452007-07-17 04:03:30 -07003910static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003911{
3912 int node;
3913 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003914 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003915 sizeof(unsigned long), GFP_KERNEL);
3916
3917 if (!map)
3918 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003919
3920 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003921 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003922 struct kmem_cache_node *n = get_node(s, node);
3923
Christoph Lameter434e2452007-07-17 04:03:30 -07003924 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003925 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003926 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003927 return count;
3928}
Christoph Lameter88a420e2007-05-06 14:49:45 -07003929/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003930 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003931 * and freed.
3932 */
3933
3934struct location {
3935 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003936 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003937 long long sum_time;
3938 long min_time;
3939 long max_time;
3940 long min_pid;
3941 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303942 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003943 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003944};
3945
3946struct loc_track {
3947 unsigned long max;
3948 unsigned long count;
3949 struct location *loc;
3950};
3951
3952static void free_loc_track(struct loc_track *t)
3953{
3954 if (t->max)
3955 free_pages((unsigned long)t->loc,
3956 get_order(sizeof(struct location) * t->max));
3957}
3958
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003959static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003960{
3961 struct location *l;
3962 int order;
3963
Christoph Lameter88a420e2007-05-06 14:49:45 -07003964 order = get_order(sizeof(struct location) * max);
3965
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003966 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003967 if (!l)
3968 return 0;
3969
3970 if (t->count) {
3971 memcpy(l, t->loc, sizeof(struct location) * t->count);
3972 free_loc_track(t);
3973 }
3974 t->max = max;
3975 t->loc = l;
3976 return 1;
3977}
3978
3979static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07003980 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003981{
3982 long start, end, pos;
3983 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003984 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003985 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003986
3987 start = -1;
3988 end = t->count;
3989
3990 for ( ; ; ) {
3991 pos = start + (end - start + 1) / 2;
3992
3993 /*
3994 * There is nothing at "end". If we end up there
3995 * we need to add something to before end.
3996 */
3997 if (pos == end)
3998 break;
3999
4000 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004001 if (track->addr == caddr) {
4002
4003 l = &t->loc[pos];
4004 l->count++;
4005 if (track->when) {
4006 l->sum_time += age;
4007 if (age < l->min_time)
4008 l->min_time = age;
4009 if (age > l->max_time)
4010 l->max_time = age;
4011
4012 if (track->pid < l->min_pid)
4013 l->min_pid = track->pid;
4014 if (track->pid > l->max_pid)
4015 l->max_pid = track->pid;
4016
Rusty Russell174596a2009-01-01 10:12:29 +10304017 cpumask_set_cpu(track->cpu,
4018 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004019 }
4020 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004021 return 1;
4022 }
4023
Christoph Lameter45edfa52007-05-09 02:32:45 -07004024 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004025 end = pos;
4026 else
4027 start = pos;
4028 }
4029
4030 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004031 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004032 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004033 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004034 return 0;
4035
4036 l = t->loc + pos;
4037 if (pos < t->count)
4038 memmove(l + 1, l,
4039 (t->count - pos) * sizeof(struct location));
4040 t->count++;
4041 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004042 l->addr = track->addr;
4043 l->sum_time = age;
4044 l->min_time = age;
4045 l->max_time = age;
4046 l->min_pid = track->pid;
4047 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304048 cpumask_clear(to_cpumask(l->cpus));
4049 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004050 nodes_clear(l->nodes);
4051 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004052 return 1;
4053}
4054
4055static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004056 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004057 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004058{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004059 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004060 void *p;
4061
Christoph Lameter39b26462008-04-14 19:11:30 +03004062 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004063 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004064
Christoph Lameter224a88b2008-04-14 19:11:31 +03004065 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004066 if (!test_bit(slab_index(p, s, addr), map))
4067 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004068}
4069
4070static int list_locations(struct kmem_cache *s, char *buf,
4071 enum track_item alloc)
4072{
Harvey Harrisone374d482008-01-31 15:20:50 -08004073 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004074 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004075 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004076 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004077 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4078 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004079
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004080 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4081 GFP_TEMPORARY)) {
4082 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004083 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004084 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004085 /* Push back cpu slabs */
4086 flush_all(s);
4087
Christoph Lameterf64dc582007-10-16 01:25:33 -07004088 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004089 struct kmem_cache_node *n = get_node(s, node);
4090 unsigned long flags;
4091 struct page *page;
4092
Christoph Lameter9e869432007-08-22 14:01:56 -07004093 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004094 continue;
4095
4096 spin_lock_irqsave(&n->list_lock, flags);
4097 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004098 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004099 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004100 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004101 spin_unlock_irqrestore(&n->list_lock, flags);
4102 }
4103
4104 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004105 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004106
Hugh Dickins9c246242008-12-09 13:14:27 -08004107 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004108 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004109 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004110
4111 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004112 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004113 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004114 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004115
4116 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004117 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004118 l->min_time,
4119 (long)div_u64(l->sum_time, l->count),
4120 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004121 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004122 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004123 l->min_time);
4124
4125 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004126 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004127 l->min_pid, l->max_pid);
4128 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004129 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004130 l->min_pid);
4131
Rusty Russell174596a2009-01-01 10:12:29 +10304132 if (num_online_cpus() > 1 &&
4133 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004134 len < PAGE_SIZE - 60) {
4135 len += sprintf(buf + len, " cpus=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004136 len += cpulist_scnprintf(buf + len,
4137 PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10304138 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004139 }
4140
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004141 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004142 len < PAGE_SIZE - 60) {
4143 len += sprintf(buf + len, " nodes=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004144 len += nodelist_scnprintf(buf + len,
4145 PAGE_SIZE - len - 50,
4146 l->nodes);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004147 }
4148
Harvey Harrisone374d482008-01-31 15:20:50 -08004149 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07004150 }
4151
4152 free_loc_track(&t);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004153 kfree(map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004154 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08004155 len += sprintf(buf, "No data\n");
4156 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004157}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004158#endif
Christoph Lameter88a420e2007-05-06 14:49:45 -07004159
Christoph Lametera5a84752010-10-05 13:57:27 -05004160#ifdef SLUB_RESILIENCY_TEST
4161static void resiliency_test(void)
4162{
4163 u8 *p;
4164
Christoph Lameter95a05b42013-01-10 19:14:19 +00004165 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
Christoph Lametera5a84752010-10-05 13:57:27 -05004166
4167 printk(KERN_ERR "SLUB resiliency testing\n");
4168 printk(KERN_ERR "-----------------------\n");
4169 printk(KERN_ERR "A. Corruption after allocation\n");
4170
4171 p = kzalloc(16, GFP_KERNEL);
4172 p[16] = 0x12;
4173 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4174 " 0x12->0x%p\n\n", p + 16);
4175
4176 validate_slab_cache(kmalloc_caches[4]);
4177
4178 /* Hmmm... The next two are dangerous */
4179 p = kzalloc(32, GFP_KERNEL);
4180 p[32 + sizeof(void *)] = 0x34;
4181 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4182 " 0x34 -> -0x%p\n", p);
4183 printk(KERN_ERR
4184 "If allocated object is overwritten then not detectable\n\n");
4185
4186 validate_slab_cache(kmalloc_caches[5]);
4187 p = kzalloc(64, GFP_KERNEL);
4188 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4189 *p = 0x56;
4190 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4191 p);
4192 printk(KERN_ERR
4193 "If allocated object is overwritten then not detectable\n\n");
4194 validate_slab_cache(kmalloc_caches[6]);
4195
4196 printk(KERN_ERR "\nB. Corruption after free\n");
4197 p = kzalloc(128, GFP_KERNEL);
4198 kfree(p);
4199 *p = 0x78;
4200 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4201 validate_slab_cache(kmalloc_caches[7]);
4202
4203 p = kzalloc(256, GFP_KERNEL);
4204 kfree(p);
4205 p[50] = 0x9a;
4206 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4207 p);
4208 validate_slab_cache(kmalloc_caches[8]);
4209
4210 p = kzalloc(512, GFP_KERNEL);
4211 kfree(p);
4212 p[512] = 0xab;
4213 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4214 validate_slab_cache(kmalloc_caches[9]);
4215}
4216#else
4217#ifdef CONFIG_SYSFS
4218static void resiliency_test(void) {};
4219#endif
4220#endif
4221
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004222#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -07004223enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03004224 SL_ALL, /* All slabs */
4225 SL_PARTIAL, /* Only partially allocated slabs */
4226 SL_CPU, /* Only slabs used for cpu caches */
4227 SL_OBJECTS, /* Determine allocated objects not slabs */
4228 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07004229};
4230
Christoph Lameter205ab992008-04-14 19:11:40 +03004231#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004232#define SO_PARTIAL (1 << SL_PARTIAL)
4233#define SO_CPU (1 << SL_CPU)
4234#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03004235#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004236
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004237static ssize_t show_slab_objects(struct kmem_cache *s,
4238 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07004239{
4240 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07004241 int node;
4242 int x;
4243 unsigned long *nodes;
Christoph Lameter81819f02007-05-06 14:49:36 -07004244
Chen Gange35e1a92013-07-12 08:23:48 +08004245 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004246 if (!nodes)
4247 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004248
Christoph Lameter205ab992008-04-14 19:11:40 +03004249 if (flags & SO_CPU) {
4250 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004251
Christoph Lameter205ab992008-04-14 19:11:40 +03004252 for_each_possible_cpu(cpu) {
Chen Gangd0e0ac92013-07-15 09:05:29 +08004253 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4254 cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004255 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004256 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004257
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004258 page = ACCESS_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004259 if (!page)
4260 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004261
Christoph Lameterec3ab082012-05-09 10:09:56 -05004262 node = page_to_nid(page);
4263 if (flags & SO_TOTAL)
4264 x = page->objects;
4265 else if (flags & SO_OBJECTS)
4266 x = page->inuse;
4267 else
4268 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004269
Christoph Lameterec3ab082012-05-09 10:09:56 -05004270 total += x;
4271 nodes[node] += x;
4272
4273 page = ACCESS_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004274 if (page) {
4275 x = page->pobjects;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004276 total += x;
4277 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004278 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004279 }
4280 }
4281
Christoph Lameter04d94872011-01-10 10:15:15 -06004282 lock_memory_hotplug();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004283#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004284 if (flags & SO_ALL) {
4285 for_each_node_state(node, N_NORMAL_MEMORY) {
4286 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07004287
Chen Gangd0e0ac92013-07-15 09:05:29 +08004288 if (flags & SO_TOTAL)
4289 x = atomic_long_read(&n->total_objects);
4290 else if (flags & SO_OBJECTS)
4291 x = atomic_long_read(&n->total_objects) -
4292 count_partial(n, count_free);
Christoph Lameter205ab992008-04-14 19:11:40 +03004293 else
4294 x = atomic_long_read(&n->nr_slabs);
4295 total += x;
4296 nodes[node] += x;
4297 }
4298
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004299 } else
4300#endif
4301 if (flags & SO_PARTIAL) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004302 for_each_node_state(node, N_NORMAL_MEMORY) {
4303 struct kmem_cache_node *n = get_node(s, node);
4304
4305 if (flags & SO_TOTAL)
4306 x = count_partial(n, count_total);
4307 else if (flags & SO_OBJECTS)
4308 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004309 else
4310 x = n->nr_partial;
4311 total += x;
4312 nodes[node] += x;
4313 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004314 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004315 x = sprintf(buf, "%lu", total);
4316#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07004317 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07004318 if (nodes[node])
4319 x += sprintf(buf + x, " N%d=%lu",
4320 node, nodes[node]);
4321#endif
Christoph Lameter04d94872011-01-10 10:15:15 -06004322 unlock_memory_hotplug();
Christoph Lameter81819f02007-05-06 14:49:36 -07004323 kfree(nodes);
4324 return x + sprintf(buf + x, "\n");
4325}
4326
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004327#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004328static int any_slab_objects(struct kmem_cache *s)
4329{
4330 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004331
4332 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004333 struct kmem_cache_node *n = get_node(s, node);
4334
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004335 if (!n)
4336 continue;
4337
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004338 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004339 return 1;
4340 }
4341 return 0;
4342}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004343#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004344
4345#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004346#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004347
4348struct slab_attribute {
4349 struct attribute attr;
4350 ssize_t (*show)(struct kmem_cache *s, char *buf);
4351 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4352};
4353
4354#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004355 static struct slab_attribute _name##_attr = \
4356 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004357
4358#define SLAB_ATTR(_name) \
4359 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004360 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004361
Christoph Lameter81819f02007-05-06 14:49:36 -07004362static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4363{
4364 return sprintf(buf, "%d\n", s->size);
4365}
4366SLAB_ATTR_RO(slab_size);
4367
4368static ssize_t align_show(struct kmem_cache *s, char *buf)
4369{
4370 return sprintf(buf, "%d\n", s->align);
4371}
4372SLAB_ATTR_RO(align);
4373
4374static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4375{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004376 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004377}
4378SLAB_ATTR_RO(object_size);
4379
4380static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4381{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004382 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004383}
4384SLAB_ATTR_RO(objs_per_slab);
4385
Christoph Lameter06b285d2008-04-14 19:11:41 +03004386static ssize_t order_store(struct kmem_cache *s,
4387 const char *buf, size_t length)
4388{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004389 unsigned long order;
4390 int err;
4391
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004392 err = kstrtoul(buf, 10, &order);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004393 if (err)
4394 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004395
4396 if (order > slub_max_order || order < slub_min_order)
4397 return -EINVAL;
4398
4399 calculate_sizes(s, order);
4400 return length;
4401}
4402
Christoph Lameter81819f02007-05-06 14:49:36 -07004403static ssize_t order_show(struct kmem_cache *s, char *buf)
4404{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004405 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004406}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004407SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004408
David Rientjes73d342b2009-02-22 17:40:09 -08004409static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4410{
4411 return sprintf(buf, "%lu\n", s->min_partial);
4412}
4413
4414static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4415 size_t length)
4416{
4417 unsigned long min;
4418 int err;
4419
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004420 err = kstrtoul(buf, 10, &min);
David Rientjes73d342b2009-02-22 17:40:09 -08004421 if (err)
4422 return err;
4423
David Rientjesc0bdb232009-02-25 09:16:35 +02004424 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004425 return length;
4426}
4427SLAB_ATTR(min_partial);
4428
Christoph Lameter49e22582011-08-09 16:12:27 -05004429static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4430{
4431 return sprintf(buf, "%u\n", s->cpu_partial);
4432}
4433
4434static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4435 size_t length)
4436{
4437 unsigned long objects;
4438 int err;
4439
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004440 err = kstrtoul(buf, 10, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05004441 if (err)
4442 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004443 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004444 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004445
4446 s->cpu_partial = objects;
4447 flush_all(s);
4448 return length;
4449}
4450SLAB_ATTR(cpu_partial);
4451
Christoph Lameter81819f02007-05-06 14:49:36 -07004452static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4453{
Joe Perches62c70bc2011-01-13 15:45:52 -08004454 if (!s->ctor)
4455 return 0;
4456 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004457}
4458SLAB_ATTR_RO(ctor);
4459
Christoph Lameter81819f02007-05-06 14:49:36 -07004460static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4461{
4462 return sprintf(buf, "%d\n", s->refcount - 1);
4463}
4464SLAB_ATTR_RO(aliases);
4465
Christoph Lameter81819f02007-05-06 14:49:36 -07004466static ssize_t partial_show(struct kmem_cache *s, char *buf)
4467{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004468 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004469}
4470SLAB_ATTR_RO(partial);
4471
4472static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4473{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004474 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004475}
4476SLAB_ATTR_RO(cpu_slabs);
4477
4478static ssize_t objects_show(struct kmem_cache *s, char *buf)
4479{
Christoph Lameter205ab992008-04-14 19:11:40 +03004480 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004481}
4482SLAB_ATTR_RO(objects);
4483
Christoph Lameter205ab992008-04-14 19:11:40 +03004484static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4485{
4486 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4487}
4488SLAB_ATTR_RO(objects_partial);
4489
Christoph Lameter49e22582011-08-09 16:12:27 -05004490static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4491{
4492 int objects = 0;
4493 int pages = 0;
4494 int cpu;
4495 int len;
4496
4497 for_each_online_cpu(cpu) {
4498 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4499
4500 if (page) {
4501 pages += page->pages;
4502 objects += page->pobjects;
4503 }
4504 }
4505
4506 len = sprintf(buf, "%d(%d)", objects, pages);
4507
4508#ifdef CONFIG_SMP
4509 for_each_online_cpu(cpu) {
4510 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4511
4512 if (page && len < PAGE_SIZE - 20)
4513 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4514 page->pobjects, page->pages);
4515 }
4516#endif
4517 return len + sprintf(buf + len, "\n");
4518}
4519SLAB_ATTR_RO(slabs_cpu_partial);
4520
Christoph Lameter81819f02007-05-06 14:49:36 -07004521static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4522{
4523 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4524}
4525
4526static ssize_t reclaim_account_store(struct kmem_cache *s,
4527 const char *buf, size_t length)
4528{
4529 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4530 if (buf[0] == '1')
4531 s->flags |= SLAB_RECLAIM_ACCOUNT;
4532 return length;
4533}
4534SLAB_ATTR(reclaim_account);
4535
4536static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4537{
Christoph Lameter5af60832007-05-06 14:49:56 -07004538 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004539}
4540SLAB_ATTR_RO(hwcache_align);
4541
4542#ifdef CONFIG_ZONE_DMA
4543static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4544{
4545 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4546}
4547SLAB_ATTR_RO(cache_dma);
4548#endif
4549
4550static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4551{
4552 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4553}
4554SLAB_ATTR_RO(destroy_by_rcu);
4555
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004556static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4557{
4558 return sprintf(buf, "%d\n", s->reserved);
4559}
4560SLAB_ATTR_RO(reserved);
4561
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004562#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004563static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4564{
4565 return show_slab_objects(s, buf, SO_ALL);
4566}
4567SLAB_ATTR_RO(slabs);
4568
4569static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4570{
4571 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4572}
4573SLAB_ATTR_RO(total_objects);
4574
4575static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4576{
4577 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4578}
4579
4580static ssize_t sanity_checks_store(struct kmem_cache *s,
4581 const char *buf, size_t length)
4582{
4583 s->flags &= ~SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004584 if (buf[0] == '1') {
4585 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004586 s->flags |= SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004587 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004588 return length;
4589}
4590SLAB_ATTR(sanity_checks);
4591
4592static ssize_t trace_show(struct kmem_cache *s, char *buf)
4593{
4594 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4595}
4596
4597static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4598 size_t length)
4599{
4600 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004601 if (buf[0] == '1') {
4602 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004603 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004604 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004605 return length;
4606}
4607SLAB_ATTR(trace);
4608
Christoph Lameter81819f02007-05-06 14:49:36 -07004609static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4610{
4611 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4612}
4613
4614static ssize_t red_zone_store(struct kmem_cache *s,
4615 const char *buf, size_t length)
4616{
4617 if (any_slab_objects(s))
4618 return -EBUSY;
4619
4620 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004621 if (buf[0] == '1') {
4622 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004623 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004624 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004625 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004626 return length;
4627}
4628SLAB_ATTR(red_zone);
4629
4630static ssize_t poison_show(struct kmem_cache *s, char *buf)
4631{
4632 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4633}
4634
4635static ssize_t poison_store(struct kmem_cache *s,
4636 const char *buf, size_t length)
4637{
4638 if (any_slab_objects(s))
4639 return -EBUSY;
4640
4641 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004642 if (buf[0] == '1') {
4643 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004644 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004645 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004646 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004647 return length;
4648}
4649SLAB_ATTR(poison);
4650
4651static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4652{
4653 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4654}
4655
4656static ssize_t store_user_store(struct kmem_cache *s,
4657 const char *buf, size_t length)
4658{
4659 if (any_slab_objects(s))
4660 return -EBUSY;
4661
4662 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004663 if (buf[0] == '1') {
4664 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004665 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004666 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004667 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004668 return length;
4669}
4670SLAB_ATTR(store_user);
4671
Christoph Lameter53e15af2007-05-06 14:49:43 -07004672static ssize_t validate_show(struct kmem_cache *s, char *buf)
4673{
4674 return 0;
4675}
4676
4677static ssize_t validate_store(struct kmem_cache *s,
4678 const char *buf, size_t length)
4679{
Christoph Lameter434e2452007-07-17 04:03:30 -07004680 int ret = -EINVAL;
4681
4682 if (buf[0] == '1') {
4683 ret = validate_slab_cache(s);
4684 if (ret >= 0)
4685 ret = length;
4686 }
4687 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004688}
4689SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004690
4691static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4692{
4693 if (!(s->flags & SLAB_STORE_USER))
4694 return -ENOSYS;
4695 return list_locations(s, buf, TRACK_ALLOC);
4696}
4697SLAB_ATTR_RO(alloc_calls);
4698
4699static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4700{
4701 if (!(s->flags & SLAB_STORE_USER))
4702 return -ENOSYS;
4703 return list_locations(s, buf, TRACK_FREE);
4704}
4705SLAB_ATTR_RO(free_calls);
4706#endif /* CONFIG_SLUB_DEBUG */
4707
4708#ifdef CONFIG_FAILSLAB
4709static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4710{
4711 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4712}
4713
4714static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4715 size_t length)
4716{
4717 s->flags &= ~SLAB_FAILSLAB;
4718 if (buf[0] == '1')
4719 s->flags |= SLAB_FAILSLAB;
4720 return length;
4721}
4722SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004723#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07004724
Christoph Lameter2086d262007-05-06 14:49:46 -07004725static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4726{
4727 return 0;
4728}
4729
4730static ssize_t shrink_store(struct kmem_cache *s,
4731 const char *buf, size_t length)
4732{
4733 if (buf[0] == '1') {
4734 int rc = kmem_cache_shrink(s);
4735
4736 if (rc)
4737 return rc;
4738 } else
4739 return -EINVAL;
4740 return length;
4741}
4742SLAB_ATTR(shrink);
4743
Christoph Lameter81819f02007-05-06 14:49:36 -07004744#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004745static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004746{
Christoph Lameter98246012008-01-07 23:20:26 -08004747 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004748}
4749
Christoph Lameter98246012008-01-07 23:20:26 -08004750static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004751 const char *buf, size_t length)
4752{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004753 unsigned long ratio;
4754 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004755
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004756 err = kstrtoul(buf, 10, &ratio);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004757 if (err)
4758 return err;
4759
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004760 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004761 s->remote_node_defrag_ratio = ratio * 10;
4762
Christoph Lameter81819f02007-05-06 14:49:36 -07004763 return length;
4764}
Christoph Lameter98246012008-01-07 23:20:26 -08004765SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004766#endif
4767
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004768#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004769static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4770{
4771 unsigned long sum = 0;
4772 int cpu;
4773 int len;
4774 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4775
4776 if (!data)
4777 return -ENOMEM;
4778
4779 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004780 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004781
4782 data[cpu] = x;
4783 sum += x;
4784 }
4785
4786 len = sprintf(buf, "%lu", sum);
4787
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004788#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004789 for_each_online_cpu(cpu) {
4790 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004791 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004792 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004793#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004794 kfree(data);
4795 return len + sprintf(buf + len, "\n");
4796}
4797
David Rientjes78eb00c2009-10-15 02:20:22 -07004798static void clear_stat(struct kmem_cache *s, enum stat_item si)
4799{
4800 int cpu;
4801
4802 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004803 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07004804}
4805
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004806#define STAT_ATTR(si, text) \
4807static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4808{ \
4809 return show_stat(s, buf, si); \
4810} \
David Rientjes78eb00c2009-10-15 02:20:22 -07004811static ssize_t text##_store(struct kmem_cache *s, \
4812 const char *buf, size_t length) \
4813{ \
4814 if (buf[0] != '0') \
4815 return -EINVAL; \
4816 clear_stat(s, si); \
4817 return length; \
4818} \
4819SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004820
4821STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4822STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4823STAT_ATTR(FREE_FASTPATH, free_fastpath);
4824STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4825STAT_ATTR(FREE_FROZEN, free_frozen);
4826STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4827STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4828STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4829STAT_ATTR(ALLOC_SLAB, alloc_slab);
4830STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05004831STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004832STAT_ATTR(FREE_SLAB, free_slab);
4833STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4834STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4835STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4836STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4837STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4838STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05004839STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03004840STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05004841STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4842STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05004843STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4844STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08004845STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4846STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004847#endif
4848
Pekka Enberg06428782008-01-07 23:20:27 -08004849static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004850 &slab_size_attr.attr,
4851 &object_size_attr.attr,
4852 &objs_per_slab_attr.attr,
4853 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004854 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004855 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004856 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004857 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004858 &partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4860 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004861 &aliases_attr.attr,
4862 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05004866 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004867 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004868 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004869#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004870 &total_objects_attr.attr,
4871 &slabs_attr.attr,
4872 &sanity_checks_attr.attr,
4873 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004874 &red_zone_attr.attr,
4875 &poison_attr.attr,
4876 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004877 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004878 &alloc_calls_attr.attr,
4879 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004880#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004881#ifdef CONFIG_ZONE_DMA
4882 &cache_dma_attr.attr,
4883#endif
4884#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004885 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004886#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004887#ifdef CONFIG_SLUB_STATS
4888 &alloc_fastpath_attr.attr,
4889 &alloc_slowpath_attr.attr,
4890 &free_fastpath_attr.attr,
4891 &free_slowpath_attr.attr,
4892 &free_frozen_attr.attr,
4893 &free_add_partial_attr.attr,
4894 &free_remove_partial_attr.attr,
4895 &alloc_from_partial_attr.attr,
4896 &alloc_slab_attr.attr,
4897 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05004898 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004899 &free_slab_attr.attr,
4900 &cpuslab_flush_attr.attr,
4901 &deactivate_full_attr.attr,
4902 &deactivate_empty_attr.attr,
4903 &deactivate_to_head_attr.attr,
4904 &deactivate_to_tail_attr.attr,
4905 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05004906 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004907 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05004908 &cmpxchg_double_fail_attr.attr,
4909 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004910 &cpu_partial_alloc_attr.attr,
4911 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08004912 &cpu_partial_node_attr.attr,
4913 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004914#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03004915#ifdef CONFIG_FAILSLAB
4916 &failslab_attr.attr,
4917#endif
4918
Christoph Lameter81819f02007-05-06 14:49:36 -07004919 NULL
4920};
4921
4922static struct attribute_group slab_attr_group = {
4923 .attrs = slab_attrs,
4924};
4925
4926static ssize_t slab_attr_show(struct kobject *kobj,
4927 struct attribute *attr,
4928 char *buf)
4929{
4930 struct slab_attribute *attribute;
4931 struct kmem_cache *s;
4932 int err;
4933
4934 attribute = to_slab_attr(attr);
4935 s = to_slab(kobj);
4936
4937 if (!attribute->show)
4938 return -EIO;
4939
4940 err = attribute->show(s, buf);
4941
4942 return err;
4943}
4944
4945static ssize_t slab_attr_store(struct kobject *kobj,
4946 struct attribute *attr,
4947 const char *buf, size_t len)
4948{
4949 struct slab_attribute *attribute;
4950 struct kmem_cache *s;
4951 int err;
4952
4953 attribute = to_slab_attr(attr);
4954 s = to_slab(kobj);
4955
4956 if (!attribute->store)
4957 return -EIO;
4958
4959 err = attribute->store(s, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08004960#ifdef CONFIG_MEMCG_KMEM
4961 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4962 int i;
Christoph Lameter81819f02007-05-06 14:49:36 -07004963
Glauber Costa107dab52012-12-18 14:23:05 -08004964 mutex_lock(&slab_mutex);
4965 if (s->max_attr_size < len)
4966 s->max_attr_size = len;
4967
Glauber Costaebe945c2012-12-18 14:23:10 -08004968 /*
4969 * This is a best effort propagation, so this function's return
4970 * value will be determined by the parent cache only. This is
4971 * basically because not all attributes will have a well
4972 * defined semantics for rollbacks - most of the actions will
4973 * have permanent effects.
4974 *
4975 * Returning the error value of any of the children that fail
4976 * is not 100 % defined, in the sense that users seeing the
4977 * error code won't be able to know anything about the state of
4978 * the cache.
4979 *
4980 * Only returning the error code for the parent cache at least
4981 * has well defined semantics. The cache being written to
4982 * directly either failed or succeeded, in which case we loop
4983 * through the descendants with best-effort propagation.
4984 */
Glauber Costa107dab52012-12-18 14:23:05 -08004985 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08004986 struct kmem_cache *c = cache_from_memcg_idx(s, i);
Glauber Costa107dab52012-12-18 14:23:05 -08004987 if (c)
4988 attribute->store(c, buf, len);
4989 }
4990 mutex_unlock(&slab_mutex);
4991 }
4992#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004993 return err;
4994}
4995
Glauber Costa107dab52012-12-18 14:23:05 -08004996static void memcg_propagate_slab_attrs(struct kmem_cache *s)
4997{
4998#ifdef CONFIG_MEMCG_KMEM
4999 int i;
5000 char *buffer = NULL;
5001
5002 if (!is_root_cache(s))
5003 return;
5004
5005 /*
5006 * This mean this cache had no attribute written. Therefore, no point
5007 * in copying default values around
5008 */
5009 if (!s->max_attr_size)
5010 return;
5011
5012 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5013 char mbuf[64];
5014 char *buf;
5015 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5016
5017 if (!attr || !attr->store || !attr->show)
5018 continue;
5019
5020 /*
5021 * It is really bad that we have to allocate here, so we will
5022 * do it only as a fallback. If we actually allocate, though,
5023 * we can just use the allocated buffer until the end.
5024 *
5025 * Most of the slub attributes will tend to be very small in
5026 * size, but sysfs allows buffers up to a page, so they can
5027 * theoretically happen.
5028 */
5029 if (buffer)
5030 buf = buffer;
5031 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5032 buf = mbuf;
5033 else {
5034 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5035 if (WARN_ON(!buffer))
5036 continue;
5037 buf = buffer;
5038 }
5039
5040 attr->show(s->memcg_params->root_cache, buf);
5041 attr->store(s, buf, strlen(buf));
5042 }
5043
5044 if (buffer)
5045 free_page((unsigned long)buffer);
5046#endif
5047}
5048
Emese Revfy52cf25d2010-01-19 02:58:23 +01005049static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005050 .show = slab_attr_show,
5051 .store = slab_attr_store,
5052};
5053
5054static struct kobj_type slab_ktype = {
5055 .sysfs_ops = &slab_sysfs_ops,
5056};
5057
5058static int uevent_filter(struct kset *kset, struct kobject *kobj)
5059{
5060 struct kobj_type *ktype = get_ktype(kobj);
5061
5062 if (ktype == &slab_ktype)
5063 return 1;
5064 return 0;
5065}
5066
Emese Revfy9cd43612009-12-31 14:52:51 +01005067static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005068 .filter = uevent_filter,
5069};
5070
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005071static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005072
5073#define ID_STR_LENGTH 64
5074
5075/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005076 *
5077 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005078 */
5079static char *create_unique_id(struct kmem_cache *s)
5080{
5081 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5082 char *p = name;
5083
5084 BUG_ON(!name);
5085
5086 *p++ = ':';
5087 /*
5088 * First flags affecting slabcache operations. We will only
5089 * get here for aliasable slabs so we do not need to support
5090 * too many flags. The flags here must cover all flags that
5091 * are matched during merging to guarantee that the id is
5092 * unique.
5093 */
5094 if (s->flags & SLAB_CACHE_DMA)
5095 *p++ = 'd';
5096 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5097 *p++ = 'a';
5098 if (s->flags & SLAB_DEBUG_FREE)
5099 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005100 if (!(s->flags & SLAB_NOTRACK))
5101 *p++ = 't';
Christoph Lameter81819f02007-05-06 14:49:36 -07005102 if (p != name + 1)
5103 *p++ = '-';
5104 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005105
5106#ifdef CONFIG_MEMCG_KMEM
5107 if (!is_root_cache(s))
Chen Gangd0e0ac92013-07-15 09:05:29 +08005108 p += sprintf(p, "-%08d",
5109 memcg_cache_id(s->memcg_params->memcg));
Glauber Costa2633d7a2012-12-18 14:22:34 -08005110#endif
5111
Christoph Lameter81819f02007-05-06 14:49:36 -07005112 BUG_ON(p > name + ID_STR_LENGTH - 1);
5113 return name;
5114}
5115
5116static int sysfs_slab_add(struct kmem_cache *s)
5117{
5118 int err;
5119 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005120 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005121
Christoph Lameter81819f02007-05-06 14:49:36 -07005122 if (unmergeable) {
5123 /*
5124 * Slabcache can never be merged so we can use the name proper.
5125 * This is typically the case for debug situations. In that
5126 * case we can catch duplicate names easily.
5127 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005128 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005129 name = s->name;
5130 } else {
5131 /*
5132 * Create a unique name for the slab as a target
5133 * for the symlinks.
5134 */
5135 name = create_unique_id(s);
5136 }
5137
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005138 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005139 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5140 if (err) {
5141 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005142 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005143 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005144
5145 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005146 if (err) {
5147 kobject_del(&s->kobj);
5148 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005149 return err;
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005150 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005151 kobject_uevent(&s->kobj, KOBJ_ADD);
5152 if (!unmergeable) {
5153 /* Setup first alias */
5154 sysfs_slab_alias(s, s->name);
5155 kfree(name);
5156 }
5157 return 0;
5158}
5159
5160static void sysfs_slab_remove(struct kmem_cache *s)
5161{
Christoph Lameter97d06602012-07-06 15:25:11 -05005162 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005163 /*
5164 * Sysfs has not been setup yet so no need to remove the
5165 * cache from sysfs.
5166 */
5167 return;
5168
Christoph Lameter81819f02007-05-06 14:49:36 -07005169 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5170 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005171 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005172}
5173
5174/*
5175 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005176 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005177 */
5178struct saved_alias {
5179 struct kmem_cache *s;
5180 const char *name;
5181 struct saved_alias *next;
5182};
5183
Adrian Bunk5af328a2007-07-17 04:03:27 -07005184static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005185
5186static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5187{
5188 struct saved_alias *al;
5189
Christoph Lameter97d06602012-07-06 15:25:11 -05005190 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005191 /*
5192 * If we have a leftover link then remove it.
5193 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005194 sysfs_remove_link(&slab_kset->kobj, name);
5195 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005196 }
5197
5198 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5199 if (!al)
5200 return -ENOMEM;
5201
5202 al->s = s;
5203 al->name = name;
5204 al->next = alias_list;
5205 alias_list = al;
5206 return 0;
5207}
5208
5209static int __init slab_sysfs_init(void)
5210{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005211 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005212 int err;
5213
Christoph Lameter18004c52012-07-06 15:25:12 -05005214 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005215
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005216 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005217 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005218 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005219 printk(KERN_ERR "Cannot register slab subsystem.\n");
5220 return -ENOSYS;
5221 }
5222
Christoph Lameter97d06602012-07-06 15:25:11 -05005223 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005224
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005225 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005226 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005227 if (err)
5228 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5229 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005230 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005231
5232 while (alias_list) {
5233 struct saved_alias *al = alias_list;
5234
5235 alias_list = alias_list->next;
5236 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005237 if (err)
5238 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
Julia Lawall068ce412012-07-08 13:37:40 +02005239 " %s to sysfs\n", al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005240 kfree(al);
5241 }
5242
Christoph Lameter18004c52012-07-06 15:25:12 -05005243 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005244 resiliency_test();
5245 return 0;
5246}
5247
5248__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005249#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005250
5251/*
5252 * The /proc/slabinfo ABI
5253 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005254#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005255void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005256{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005257 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005258 unsigned long nr_objs = 0;
5259 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005260 int node;
5261
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005262 for_each_online_node(node) {
5263 struct kmem_cache_node *n = get_node(s, node);
5264
5265 if (!n)
5266 continue;
5267
Wanpeng Lic17fd132013-07-04 08:33:26 +08005268 nr_slabs += node_nr_slabs(n);
5269 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005270 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005271 }
5272
Glauber Costa0d7561c2012-10-19 18:20:27 +04005273 sinfo->active_objs = nr_objs - nr_free;
5274 sinfo->num_objs = nr_objs;
5275 sinfo->active_slabs = nr_slabs;
5276 sinfo->num_slabs = nr_slabs;
5277 sinfo->objects_per_slab = oo_objects(s->oo);
5278 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005279}
5280
Glauber Costa0d7561c2012-10-19 18:20:27 +04005281void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005282{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005283}
5284
Glauber Costab7454ad2012-10-19 18:20:25 +04005285ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5286 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005287{
Glauber Costab7454ad2012-10-19 18:20:25 +04005288 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005289}
Linus Torvalds158a9622008-01-02 13:04:48 -08005290#endif /* CONFIG_SLABINFO */