blob: 545a170ebf9f66cf0e3716c9cd6f4cb7eef0eda6 [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
Zhi Yong Wu721ae222013-11-08 20:47:37 +0800158 * sort the partial list by the number of objects in use.
Christoph Lameter2086d262007-05-06 14:49:46 -0700159 */
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 */
Roman Bobnievd56791b2013-10-08 15:58:57 -0700936static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
937{
938 kmemleak_alloc(ptr, size, 1, flags);
939}
940
941static inline void kfree_hook(const void *x)
942{
943 kmemleak_free(x);
944}
945
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500946static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
947{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500948 flags &= gfp_allowed_mask;
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500949 lockdep_trace_alloc(flags);
950 might_sleep_if(flags & __GFP_WAIT);
951
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500952 return should_failslab(s->object_size, flags, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500953}
954
Chen Gangd0e0ac92013-07-15 09:05:29 +0800955static inline void slab_post_alloc_hook(struct kmem_cache *s,
956 gfp_t flags, void *object)
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500957{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500958 flags &= gfp_allowed_mask;
Eric Dumazetb3d41882011-02-14 18:35:22 +0100959 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500960 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500961}
962
963static inline void slab_free_hook(struct kmem_cache *s, void *x)
964{
965 kmemleak_free_recursive(x, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500966
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600967 /*
Xie XiuQid1756172013-10-18 09:12:43 +0800968 * Trouble is that we may no longer disable interrupts in the fast path
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600969 * So in order to make the debug calls that expect irqs to be
970 * disabled we need to disable interrupts temporarily.
971 */
972#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
973 {
974 unsigned long flags;
975
976 local_irq_save(flags);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500977 kmemcheck_slab_free(s, x, s->object_size);
978 debug_check_no_locks_freed(x, s->object_size);
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600979 local_irq_restore(flags);
980 }
981#endif
Thomas Gleixnerf9b615d2011-03-24 21:26:46 +0200982 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500983 debug_check_no_obj_freed(x, s->object_size);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500984}
985
986/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700987 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500988 *
989 * list_lock must be held.
Christoph Lameter643b1132007-05-06 14:49:42 -0700990 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500991static void add_full(struct kmem_cache *s,
992 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700993{
Christoph Lameter643b1132007-05-06 14:49:42 -0700994 if (!(s->flags & SLAB_STORE_USER))
995 return;
996
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500997 list_add(&page->lru, &n->full);
998}
Christoph Lameter643b1132007-05-06 14:49:42 -0700999
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001000/*
1001 * list_lock must be held.
1002 */
1003static void remove_full(struct kmem_cache *s, struct page *page)
1004{
1005 if (!(s->flags & SLAB_STORE_USER))
1006 return;
1007
Christoph Lameter643b1132007-05-06 14:49:42 -07001008 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -07001009}
1010
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001011/* Tracking of the number of slabs for debugging purposes */
1012static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1013{
1014 struct kmem_cache_node *n = get_node(s, node);
1015
1016 return atomic_long_read(&n->nr_slabs);
1017}
1018
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001019static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1020{
1021 return atomic_long_read(&n->nr_slabs);
1022}
1023
Christoph Lameter205ab992008-04-14 19:11:40 +03001024static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001025{
1026 struct kmem_cache_node *n = get_node(s, node);
1027
1028 /*
1029 * May be called early in order to allocate a slab for the
1030 * kmem_cache_node structure. Solve the chicken-egg
1031 * dilemma by deferring the increment of the count during
1032 * bootstrap (see early_kmem_cache_node_alloc).
1033 */
Joonsoo Kim338b2642013-01-21 17:01:27 +09001034 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001035 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001036 atomic_long_add(objects, &n->total_objects);
1037 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001038}
Christoph Lameter205ab992008-04-14 19:11:40 +03001039static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001040{
1041 struct kmem_cache_node *n = get_node(s, node);
1042
1043 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001044 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001045}
1046
1047/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001048static void setup_object_debug(struct kmem_cache *s, struct page *page,
1049 void *object)
1050{
1051 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1052 return;
1053
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001054 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001055 init_tracking(s, object);
1056}
1057
Chen Gangd0e0ac92013-07-15 09:05:29 +08001058static noinline int alloc_debug_processing(struct kmem_cache *s,
1059 struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001060 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001061{
1062 if (!check_slab(s, page))
1063 goto bad;
1064
Christoph Lameter81819f02007-05-06 14:49:36 -07001065 if (!check_valid_pointer(s, page, object)) {
1066 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -07001067 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001068 }
1069
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001070 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Christoph Lameter81819f02007-05-06 14:49:36 -07001071 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001072
Christoph Lameter3ec09742007-05-16 22:11:00 -07001073 /* Success perform special debug activities for allocs */
1074 if (s->flags & SLAB_STORE_USER)
1075 set_track(s, object, TRACK_ALLOC, addr);
1076 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001077 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001078 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001079
Christoph Lameter81819f02007-05-06 14:49:36 -07001080bad:
1081 if (PageSlab(page)) {
1082 /*
1083 * If this is a slab page then lets do the best we can
1084 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001085 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001086 */
Christoph Lameter24922682007-07-17 04:03:18 -07001087 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001088 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001089 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 }
1091 return 0;
1092}
1093
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001094static noinline struct kmem_cache_node *free_debug_processing(
1095 struct kmem_cache *s, struct page *page, void *object,
1096 unsigned long addr, unsigned long *flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001097{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001098 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001099
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001100 spin_lock_irqsave(&n->list_lock, *flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001101 slab_lock(page);
1102
Christoph Lameter81819f02007-05-06 14:49:36 -07001103 if (!check_slab(s, page))
1104 goto fail;
1105
1106 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001107 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001108 goto fail;
1109 }
1110
1111 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -07001112 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -07001113 goto fail;
1114 }
1115
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001116 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001117 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001118
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001119 if (unlikely(s != page->slab_cache)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001120 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001121 slab_err(s, page, "Attempt to free object(0x%p) "
1122 "outside of slab", object);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001123 } else if (!page->slab_cache) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001124 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -07001125 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -07001126 object);
Christoph Lameter70d71222007-05-06 14:49:47 -07001127 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -08001128 } else
Christoph Lameter24922682007-07-17 04:03:18 -07001129 object_err(s, page, object,
1130 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -07001131 goto fail;
1132 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001133
Christoph Lameter3ec09742007-05-16 22:11:00 -07001134 if (s->flags & SLAB_STORE_USER)
1135 set_track(s, object, TRACK_FREE, addr);
1136 trace(s, page, object, 0);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001137 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001138out:
Christoph Lameter881db7f2011-06-01 12:25:53 -05001139 slab_unlock(page);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001140 /*
1141 * Keep node_lock to preserve integrity
1142 * until the object is actually freed
1143 */
1144 return n;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001145
Christoph Lameter81819f02007-05-06 14:49:36 -07001146fail:
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001147 slab_unlock(page);
1148 spin_unlock_irqrestore(&n->list_lock, *flags);
Christoph Lameter24922682007-07-17 04:03:18 -07001149 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001150 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001151}
1152
Christoph Lameter41ecc552007-05-09 02:32:44 -07001153static int __init setup_slub_debug(char *str)
1154{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001155 slub_debug = DEBUG_DEFAULT_FLAGS;
1156 if (*str++ != '=' || !*str)
1157 /*
1158 * No options specified. Switch on full debugging.
1159 */
1160 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001161
1162 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001163 /*
1164 * No options but restriction on slabs. This means full
1165 * debugging for slabs matching a pattern.
1166 */
1167 goto check_slabs;
1168
David Rientjesfa5ec8a2009-07-07 00:14:14 -07001169 if (tolower(*str) == 'o') {
1170 /*
1171 * Avoid enabling debugging on caches if its minimum order
1172 * would increase as a result.
1173 */
1174 disable_higher_order_debug = 1;
1175 goto out;
1176 }
1177
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001178 slub_debug = 0;
1179 if (*str == '-')
1180 /*
1181 * Switch off all debugging measures.
1182 */
1183 goto out;
1184
1185 /*
1186 * Determine which debug features should be switched on
1187 */
Pekka Enberg06428782008-01-07 23:20:27 -08001188 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001189 switch (tolower(*str)) {
1190 case 'f':
1191 slub_debug |= SLAB_DEBUG_FREE;
1192 break;
1193 case 'z':
1194 slub_debug |= SLAB_RED_ZONE;
1195 break;
1196 case 'p':
1197 slub_debug |= SLAB_POISON;
1198 break;
1199 case 'u':
1200 slub_debug |= SLAB_STORE_USER;
1201 break;
1202 case 't':
1203 slub_debug |= SLAB_TRACE;
1204 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001205 case 'a':
1206 slub_debug |= SLAB_FAILSLAB;
1207 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001208 default:
1209 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001210 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001211 }
1212 }
1213
1214check_slabs:
1215 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001216 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001217out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001218 return 1;
1219}
1220
1221__setup("slub_debug", setup_slub_debug);
1222
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001223static unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001224 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001225 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001226{
1227 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001228 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001229 */
Christoph Lameterc6f58d92013-11-07 16:29:15 +00001230 if (slub_debug && (!slub_debug_slabs || (name &&
1231 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
David Rientjes3de47212009-07-27 18:30:35 -07001232 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001233
1234 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001235}
1236#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001237static inline void setup_object_debug(struct kmem_cache *s,
1238 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001239
Christoph Lameter3ec09742007-05-16 22:11:00 -07001240static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001241 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001242
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001243static inline struct kmem_cache_node *free_debug_processing(
1244 struct kmem_cache *s, struct page *page, void *object,
1245 unsigned long addr, unsigned long *flags) { return NULL; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001246
Christoph Lameter41ecc552007-05-09 02:32:44 -07001247static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1248 { return 1; }
1249static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001250 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001251static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1252 struct page *page) {}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001253static inline void remove_full(struct kmem_cache *s, struct page *page) {}
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001254static inline unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001255 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001256 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001257{
1258 return flags;
1259}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001260#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001261
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001262#define disable_higher_order_debug 0
1263
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001264static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1265 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001266static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1267 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001268static inline void inc_slabs_node(struct kmem_cache *s, int node,
1269 int objects) {}
1270static inline void dec_slabs_node(struct kmem_cache *s, int node,
1271 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001272
Roman Bobnievd56791b2013-10-08 15:58:57 -07001273static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1274{
1275 kmemleak_alloc(ptr, size, 1, flags);
1276}
1277
1278static inline void kfree_hook(const void *x)
1279{
1280 kmemleak_free(x);
1281}
1282
Christoph Lameter7d550c52010-08-25 14:07:16 -05001283static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1284 { return 0; }
1285
1286static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
Roman Bobnievd56791b2013-10-08 15:58:57 -07001287 void *object)
1288{
1289 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1290 flags & gfp_allowed_mask);
1291}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001292
Roman Bobnievd56791b2013-10-08 15:58:57 -07001293static inline void slab_free_hook(struct kmem_cache *s, void *x)
1294{
1295 kmemleak_free_recursive(x, s->flags);
1296}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001297
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05001298#endif /* CONFIG_SLUB_DEBUG */
Christoph Lameter205ab992008-04-14 19:11:40 +03001299
Christoph Lameter81819f02007-05-06 14:49:36 -07001300/*
1301 * Slab allocation and freeing
1302 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001303static inline struct page *alloc_slab_page(gfp_t flags, int node,
1304 struct kmem_cache_order_objects oo)
1305{
1306 int order = oo_order(oo);
1307
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001308 flags |= __GFP_NOTRACK;
1309
Christoph Lameter2154a332010-07-09 14:07:10 -05001310 if (node == NUMA_NO_NODE)
Christoph Lameter65c33762008-04-14 19:11:40 +03001311 return alloc_pages(flags, order);
1312 else
Minchan Kim6b65aaf2010-04-14 23:58:36 +09001313 return alloc_pages_exact_node(node, flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001314}
1315
Christoph Lameter81819f02007-05-06 14:49:36 -07001316static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1317{
Pekka Enberg06428782008-01-07 23:20:27 -08001318 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001319 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001320 gfp_t alloc_gfp;
Christoph Lameter81819f02007-05-06 14:49:36 -07001321
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001322 flags &= gfp_allowed_mask;
1323
1324 if (flags & __GFP_WAIT)
1325 local_irq_enable();
1326
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001327 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001328
Pekka Enbergba522702009-06-24 21:59:51 +03001329 /*
1330 * Let the initial higher-order allocation fail under memory pressure
1331 * so we fall-back to the minimum order allocation.
1332 */
1333 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1334
1335 page = alloc_slab_page(alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001336 if (unlikely(!page)) {
1337 oo = s->min;
1338 /*
1339 * Allocation may have failed due to fragmentation.
1340 * Try a lower order alloc if possible
1341 */
1342 page = alloc_slab_page(flags, node, oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001343
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001344 if (page)
1345 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001346 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001347
David Rientjes737b7192012-07-09 14:00:38 -07001348 if (kmemcheck_enabled && page
Amerigo Wang5086c389c2009-08-19 21:44:13 +03001349 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001350 int pages = 1 << oo_order(oo);
1351
1352 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1353
1354 /*
1355 * Objects from caches that have a constructor don't get
1356 * cleared when they're allocated, so we need to do it here.
1357 */
1358 if (s->ctor)
1359 kmemcheck_mark_uninitialized_pages(page, pages);
1360 else
1361 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001362 }
1363
David Rientjes737b7192012-07-09 14:00:38 -07001364 if (flags & __GFP_WAIT)
1365 local_irq_disable();
1366 if (!page)
1367 return NULL;
1368
Christoph Lameter834f3d12008-04-14 19:11:31 +03001369 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001370 mod_zone_page_state(page_zone(page),
1371 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1372 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001373 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001374
1375 return page;
1376}
1377
1378static void setup_object(struct kmem_cache *s, struct page *page,
1379 void *object)
1380{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001381 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001382 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001383 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001384}
1385
1386static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1387{
1388 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001389 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001390 void *last;
1391 void *p;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001392 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001393
Christoph Lameter6cb06222007-10-16 01:25:41 -07001394 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001395
Christoph Lameter6cb06222007-10-16 01:25:41 -07001396 page = allocate_slab(s,
1397 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001398 if (!page)
1399 goto out;
1400
Glauber Costa1f458cb2012-12-18 14:22:50 -08001401 order = compound_order(page);
Christoph Lameter205ab992008-04-14 19:11:40 +03001402 inc_slabs_node(s, page_to_nid(page), page->objects);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001403 memcg_bind_pages(s, order);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001404 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001405 __SetPageSlab(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001406 if (page->pfmemalloc)
1407 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001408
1409 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001410
1411 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001412 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001413
1414 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001415 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001416 setup_object(s, page, last);
1417 set_freepointer(s, last, p);
1418 last = p;
1419 }
1420 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001421 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001422
1423 page->freelist = start;
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001424 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001425 page->frozen = 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001426out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001427 return page;
1428}
1429
1430static void __free_slab(struct kmem_cache *s, struct page *page)
1431{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001432 int order = compound_order(page);
1433 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001434
Christoph Lameteraf537b02010-07-09 14:07:14 -05001435 if (kmem_cache_debug(s)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001436 void *p;
1437
1438 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001439 for_each_object(p, s, page_address(page),
1440 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001441 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001442 }
1443
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001444 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001445
Christoph Lameter81819f02007-05-06 14:49:36 -07001446 mod_zone_page_state(page_zone(page),
1447 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1448 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001449 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001450
Mel Gorman072bb0a2012-07-31 16:43:58 -07001451 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001452 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001453
1454 memcg_release_pages(s, order);
Mel Gorman22b751c2013-02-22 16:34:59 -08001455 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001456 if (current->reclaim_state)
1457 current->reclaim_state->reclaimed_slab += pages;
Glauber Costad79923f2012-12-18 14:22:48 -08001458 __free_memcg_kmem_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001459}
1460
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001461#define need_reserve_slab_rcu \
1462 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1463
Christoph Lameter81819f02007-05-06 14:49:36 -07001464static void rcu_free_slab(struct rcu_head *h)
1465{
1466 struct page *page;
1467
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001468 if (need_reserve_slab_rcu)
1469 page = virt_to_head_page(h);
1470 else
1471 page = container_of((struct list_head *)h, struct page, lru);
1472
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001473 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001474}
1475
1476static void free_slab(struct kmem_cache *s, struct page *page)
1477{
1478 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001479 struct rcu_head *head;
1480
1481 if (need_reserve_slab_rcu) {
1482 int order = compound_order(page);
1483 int offset = (PAGE_SIZE << order) - s->reserved;
1484
1485 VM_BUG_ON(s->reserved != sizeof(*head));
1486 head = page_address(page) + offset;
1487 } else {
1488 /*
1489 * RCU free overloads the RCU head over the LRU
1490 */
1491 head = (void *)&page->lru;
1492 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001493
1494 call_rcu(head, rcu_free_slab);
1495 } else
1496 __free_slab(s, page);
1497}
1498
1499static void discard_slab(struct kmem_cache *s, struct page *page)
1500{
Christoph Lameter205ab992008-04-14 19:11:40 +03001501 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001502 free_slab(s, page);
1503}
1504
1505/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001506 * Management of partially allocated slabs.
1507 *
1508 * list_lock must be held.
Christoph Lameter81819f02007-05-06 14:49:36 -07001509 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001510static inline void add_partial(struct kmem_cache_node *n,
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001511 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001512{
Christoph Lametere95eed52007-05-06 14:49:44 -07001513 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001514 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001515 list_add_tail(&page->lru, &n->partial);
1516 else
1517 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001518}
1519
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001520/*
1521 * list_lock must be held.
1522 */
1523static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001524 struct page *page)
1525{
1526 list_del(&page->lru);
1527 n->nr_partial--;
1528}
1529
Christoph Lameter81819f02007-05-06 14:49:36 -07001530/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001531 * Remove slab from the partial list, freeze it and
1532 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001533 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001534 * Returns a list of objects or NULL if it fails.
1535 *
Christoph Lameter7ced3712012-05-09 10:09:53 -05001536 * Must hold list_lock since we modify the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001537 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001538static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001539 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001540 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001541{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001542 void *freelist;
1543 unsigned long counters;
1544 struct page new;
1545
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001546 /*
1547 * Zap the freelist and set the frozen bit.
1548 * The old freelist is the list of objects for the
1549 * per cpu allocation list.
1550 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001551 freelist = page->freelist;
1552 counters = page->counters;
1553 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001554 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001555 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001556 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001557 new.freelist = NULL;
1558 } else {
1559 new.freelist = freelist;
1560 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001561
Christoph Lameter7ced3712012-05-09 10:09:53 -05001562 VM_BUG_ON(new.frozen);
1563 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001564
Christoph Lameter7ced3712012-05-09 10:09:53 -05001565 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001566 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001567 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001568 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001569 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001570
1571 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001572 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001573 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001574}
1575
Joonsoo Kim633b0762013-01-21 17:01:25 +09001576static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001577static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001578
Christoph Lameter81819f02007-05-06 14:49:36 -07001579/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001580 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001581 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001582static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1583 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001584{
Christoph Lameter49e22582011-08-09 16:12:27 -05001585 struct page *page, *page2;
1586 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001587 int available = 0;
1588 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001589
1590 /*
1591 * Racy check. If we mistakenly see no partial slabs then we
1592 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001593 * partial slab and there is none available then get_partials()
1594 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001595 */
1596 if (!n || !n->nr_partial)
1597 return NULL;
1598
1599 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001600 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001601 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001602
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001603 if (!pfmemalloc_match(page, flags))
1604 continue;
1605
Joonsoo Kim633b0762013-01-21 17:01:25 +09001606 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001607 if (!t)
1608 break;
1609
Joonsoo Kim633b0762013-01-21 17:01:25 +09001610 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001611 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001612 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001613 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001614 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001615 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001616 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001617 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001618 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001619 if (!kmem_cache_has_cpu_partial(s)
1620 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001621 break;
1622
Christoph Lameter497b66f2011-08-09 16:12:26 -05001623 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001624 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001625 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001626}
1627
1628/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001629 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001630 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001631static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001632 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001633{
1634#ifdef CONFIG_NUMA
1635 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001636 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001637 struct zone *zone;
1638 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001639 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001640 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001641
1642 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001643 * The defrag ratio allows a configuration of the tradeoffs between
1644 * inter node defragmentation and node local allocations. A lower
1645 * defrag_ratio increases the tendency to do local allocations
1646 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001647 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001648 * If the defrag_ratio is set to 0 then kmalloc() always
1649 * returns node local objects. If the ratio is higher then kmalloc()
1650 * may return off node objects because partial slabs are obtained
1651 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001652 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001653 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001654 * defrag_ratio = 1000) then every (well almost) allocation will
1655 * first attempt to defrag slab caches on other nodes. This means
1656 * scanning over all nodes to look for partial slabs which may be
1657 * expensive if we do it every time we are trying to find a slab
1658 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001659 */
Christoph Lameter98246012008-01-07 23:20:26 -08001660 if (!s->remote_node_defrag_ratio ||
1661 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001662 return NULL;
1663
Mel Gormancc9a6c82012-03-21 16:34:11 -07001664 do {
1665 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07001666 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001667 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1668 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001669
Mel Gormancc9a6c82012-03-21 16:34:11 -07001670 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001671
Mel Gormancc9a6c82012-03-21 16:34:11 -07001672 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1673 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001674 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001675 if (object) {
1676 /*
1677 * Return the object even if
1678 * put_mems_allowed indicated that
1679 * the cpuset mems_allowed was
1680 * updated in parallel. It's a
1681 * harmless race between the alloc
1682 * and the cpuset update.
1683 */
1684 put_mems_allowed(cpuset_mems_cookie);
1685 return object;
1686 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001687 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001688 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001689 } while (!put_mems_allowed(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001690#endif
1691 return NULL;
1692}
1693
1694/*
1695 * Get a partial page, lock it and return it.
1696 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001697static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001698 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001699{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001700 void *object;
Christoph Lameter2154a332010-07-09 14:07:10 -05001701 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
Christoph Lameter81819f02007-05-06 14:49:36 -07001702
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001703 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001704 if (object || node != NUMA_NO_NODE)
1705 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001706
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001707 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001708}
1709
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001710#ifdef CONFIG_PREEMPT
1711/*
1712 * Calculate the next globally unique transaction for disambiguiation
1713 * during cmpxchg. The transactions start with the cpu number and are then
1714 * incremented by CONFIG_NR_CPUS.
1715 */
1716#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1717#else
1718/*
1719 * No preemption supported therefore also no need to check for
1720 * different cpus.
1721 */
1722#define TID_STEP 1
1723#endif
1724
1725static inline unsigned long next_tid(unsigned long tid)
1726{
1727 return tid + TID_STEP;
1728}
1729
1730static inline unsigned int tid_to_cpu(unsigned long tid)
1731{
1732 return tid % TID_STEP;
1733}
1734
1735static inline unsigned long tid_to_event(unsigned long tid)
1736{
1737 return tid / TID_STEP;
1738}
1739
1740static inline unsigned int init_tid(int cpu)
1741{
1742 return cpu;
1743}
1744
1745static inline void note_cmpxchg_failure(const char *n,
1746 const struct kmem_cache *s, unsigned long tid)
1747{
1748#ifdef SLUB_DEBUG_CMPXCHG
1749 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1750
1751 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1752
1753#ifdef CONFIG_PREEMPT
1754 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1755 printk("due to cpu change %d -> %d\n",
1756 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1757 else
1758#endif
1759 if (tid_to_event(tid) != tid_to_event(actual_tid))
1760 printk("due to cpu running other code. Event %ld->%ld\n",
1761 tid_to_event(tid), tid_to_event(actual_tid));
1762 else
1763 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1764 actual_tid, tid, next_tid(tid));
1765#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001766 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001767}
1768
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001769static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001770{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001771 int cpu;
1772
1773 for_each_possible_cpu(cpu)
1774 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001775}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001776
1777/*
1778 * Remove the cpu slab
1779 */
Chen Gangd0e0ac92013-07-15 09:05:29 +08001780static void deactivate_slab(struct kmem_cache *s, struct page *page,
1781 void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001782{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001783 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001784 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1785 int lock = 0;
1786 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001787 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001788 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001789 struct page new;
1790 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001791
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001792 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001793 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001794 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001795 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001796
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001797 /*
1798 * Stage one: Free all available per cpu objects back
1799 * to the page freelist while it is still frozen. Leave the
1800 * last one.
1801 *
1802 * There is no need to take the list->lock because the page
1803 * is still frozen.
1804 */
1805 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1806 void *prior;
1807 unsigned long counters;
1808
1809 do {
1810 prior = page->freelist;
1811 counters = page->counters;
1812 set_freepointer(s, freelist, prior);
1813 new.counters = counters;
1814 new.inuse--;
1815 VM_BUG_ON(!new.frozen);
1816
Christoph Lameter1d071712011-07-14 12:49:12 -05001817 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001818 prior, counters,
1819 freelist, new.counters,
1820 "drain percpu freelist"));
1821
1822 freelist = nextfree;
1823 }
1824
1825 /*
1826 * Stage two: Ensure that the page is unfrozen while the
1827 * list presence reflects the actual number of objects
1828 * during unfreeze.
1829 *
1830 * We setup the list membership and then perform a cmpxchg
1831 * with the count. If there is a mismatch then the page
1832 * is not unfrozen but the page is on the wrong list.
1833 *
1834 * Then we restart the process which may have to remove
1835 * the page from the list that we just put it on again
1836 * because the number of objects in the slab may have
1837 * changed.
1838 */
1839redo:
1840
1841 old.freelist = page->freelist;
1842 old.counters = page->counters;
1843 VM_BUG_ON(!old.frozen);
1844
1845 /* Determine target state of the slab */
1846 new.counters = old.counters;
1847 if (freelist) {
1848 new.inuse--;
1849 set_freepointer(s, freelist, old.freelist);
1850 new.freelist = freelist;
1851 } else
1852 new.freelist = old.freelist;
1853
1854 new.frozen = 0;
1855
Christoph Lameter81107182011-08-09 13:01:32 -05001856 if (!new.inuse && n->nr_partial > s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001857 m = M_FREE;
1858 else if (new.freelist) {
1859 m = M_PARTIAL;
1860 if (!lock) {
1861 lock = 1;
1862 /*
1863 * Taking the spinlock removes the possiblity
1864 * that acquire_slab() will see a slab page that
1865 * is frozen
1866 */
1867 spin_lock(&n->list_lock);
1868 }
1869 } else {
1870 m = M_FULL;
1871 if (kmem_cache_debug(s) && !lock) {
1872 lock = 1;
1873 /*
1874 * This also ensures that the scanning of full
1875 * slabs from diagnostic functions will not see
1876 * any frozen slabs.
1877 */
1878 spin_lock(&n->list_lock);
1879 }
1880 }
1881
1882 if (l != m) {
1883
1884 if (l == M_PARTIAL)
1885
1886 remove_partial(n, page);
1887
1888 else if (l == M_FULL)
1889
1890 remove_full(s, page);
1891
1892 if (m == M_PARTIAL) {
1893
1894 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001895 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001896
1897 } else if (m == M_FULL) {
1898
1899 stat(s, DEACTIVATE_FULL);
1900 add_full(s, n, page);
1901
1902 }
1903 }
1904
1905 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001906 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001907 old.freelist, old.counters,
1908 new.freelist, new.counters,
1909 "unfreezing slab"))
1910 goto redo;
1911
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001912 if (lock)
1913 spin_unlock(&n->list_lock);
1914
1915 if (m == M_FREE) {
1916 stat(s, DEACTIVATE_EMPTY);
1917 discard_slab(s, page);
1918 stat(s, FREE_SLAB);
1919 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001920}
1921
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001922/*
1923 * Unfreeze all the cpu partial slabs.
1924 *
Christoph Lameter59a09912012-11-28 16:23:00 +00001925 * This function must be called with interrupts disabled
1926 * for the cpu using c (or some other guarantee must be there
1927 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001928 */
Christoph Lameter59a09912012-11-28 16:23:00 +00001929static void unfreeze_partials(struct kmem_cache *s,
1930 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05001931{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001932#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09001933 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08001934 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001935
1936 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001937 struct page new;
1938 struct page old;
1939
1940 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001941
1942 n2 = get_node(s, page_to_nid(page));
1943 if (n != n2) {
1944 if (n)
1945 spin_unlock(&n->list_lock);
1946
1947 n = n2;
1948 spin_lock(&n->list_lock);
1949 }
Christoph Lameter49e22582011-08-09 16:12:27 -05001950
1951 do {
1952
1953 old.freelist = page->freelist;
1954 old.counters = page->counters;
1955 VM_BUG_ON(!old.frozen);
1956
1957 new.counters = old.counters;
1958 new.freelist = old.freelist;
1959
1960 new.frozen = 0;
1961
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001962 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05001963 old.freelist, old.counters,
1964 new.freelist, new.counters,
1965 "unfreezing slab"));
1966
Joonsoo Kim43d77862012-06-09 02:23:16 +09001967 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08001968 page->next = discard_page;
1969 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001970 } else {
1971 add_partial(n, page, DEACTIVATE_TO_TAIL);
1972 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001973 }
1974 }
1975
1976 if (n)
1977 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08001978
1979 while (discard_page) {
1980 page = discard_page;
1981 discard_page = discard_page->next;
1982
1983 stat(s, DEACTIVATE_EMPTY);
1984 discard_slab(s, page);
1985 stat(s, FREE_SLAB);
1986 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001987#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05001988}
1989
1990/*
1991 * Put a page that was just frozen (in __slab_free) into a partial page
1992 * slot if available. This is done without interrupts disabled and without
1993 * preemption disabled. The cmpxchg is racy and may put the partial page
1994 * onto a random cpus partial slot.
1995 *
1996 * If we did not find a slot then simply move all the partials to the
1997 * per node partial list.
1998 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09001999static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05002000{
Joonsoo Kim345c9052013-06-19 14:05:52 +09002001#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05002002 struct page *oldpage;
2003 int pages;
2004 int pobjects;
2005
2006 do {
2007 pages = 0;
2008 pobjects = 0;
2009 oldpage = this_cpu_read(s->cpu_slab->partial);
2010
2011 if (oldpage) {
2012 pobjects = oldpage->pobjects;
2013 pages = oldpage->pages;
2014 if (drain && pobjects > s->cpu_partial) {
2015 unsigned long flags;
2016 /*
2017 * partial array is full. Move the existing
2018 * set to the per node partial list.
2019 */
2020 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00002021 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05002022 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09002023 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05002024 pobjects = 0;
2025 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08002026 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05002027 }
2028 }
2029
2030 pages++;
2031 pobjects += page->objects - page->inuse;
2032
2033 page->pages = pages;
2034 page->pobjects = pobjects;
2035 page->next = oldpage;
2036
Chen Gangd0e0ac92013-07-15 09:05:29 +08002037 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2038 != oldpage);
Joonsoo Kim345c9052013-06-19 14:05:52 +09002039#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002040}
2041
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002042static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002043{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002044 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002045 deactivate_slab(s, c->page, c->freelist);
2046
2047 c->tid = next_tid(c->tid);
2048 c->page = NULL;
2049 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002050}
2051
2052/*
2053 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002054 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002055 * Called from IPI handler with interrupts disabled.
2056 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002057static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002058{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002059 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002060
Christoph Lameter49e22582011-08-09 16:12:27 -05002061 if (likely(c)) {
2062 if (c->page)
2063 flush_slab(s, c);
2064
Christoph Lameter59a09912012-11-28 16:23:00 +00002065 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002066 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002067}
2068
2069static void flush_cpu_slab(void *d)
2070{
2071 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002072
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002073 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002074}
2075
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002076static bool has_cpu_slab(int cpu, void *info)
2077{
2078 struct kmem_cache *s = info;
2079 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2080
majianpeng02e1a9c2012-05-17 17:03:26 -07002081 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002082}
2083
Christoph Lameter81819f02007-05-06 14:49:36 -07002084static void flush_all(struct kmem_cache *s)
2085{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002086 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002087}
2088
2089/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002090 * Check if the objects in a per cpu structure fit numa
2091 * locality expectations.
2092 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002093static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002094{
2095#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002096 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002097 return 0;
2098#endif
2099 return 1;
2100}
2101
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002102static int count_free(struct page *page)
2103{
2104 return page->objects - page->inuse;
2105}
2106
2107static unsigned long count_partial(struct kmem_cache_node *n,
2108 int (*get_count)(struct page *))
2109{
2110 unsigned long flags;
2111 unsigned long x = 0;
2112 struct page *page;
2113
2114 spin_lock_irqsave(&n->list_lock, flags);
2115 list_for_each_entry(page, &n->partial, lru)
2116 x += get_count(page);
2117 spin_unlock_irqrestore(&n->list_lock, flags);
2118 return x;
2119}
2120
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002121static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2122{
2123#ifdef CONFIG_SLUB_DEBUG
2124 return atomic_long_read(&n->total_objects);
2125#else
2126 return 0;
2127#endif
2128}
2129
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002130static noinline void
2131slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2132{
2133 int node;
2134
2135 printk(KERN_WARNING
2136 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2137 nid, gfpflags);
2138 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002139 "default order: %d, min order: %d\n", s->name, s->object_size,
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002140 s->size, oo_order(s->oo), oo_order(s->min));
2141
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002142 if (oo_order(s->min) > get_order(s->object_size))
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002143 printk(KERN_WARNING " %s debugging increased min order, use "
2144 "slub_debug=O to disable.\n", s->name);
2145
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002146 for_each_online_node(node) {
2147 struct kmem_cache_node *n = get_node(s, node);
2148 unsigned long nr_slabs;
2149 unsigned long nr_objs;
2150 unsigned long nr_free;
2151
2152 if (!n)
2153 continue;
2154
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002155 nr_free = count_partial(n, count_free);
2156 nr_slabs = node_nr_slabs(n);
2157 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002158
2159 printk(KERN_WARNING
2160 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2161 node, nr_slabs, nr_objs, nr_free);
2162 }
2163}
2164
Christoph Lameter497b66f2011-08-09 16:12:26 -05002165static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2166 int node, struct kmem_cache_cpu **pc)
2167{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002168 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002169 struct kmem_cache_cpu *c = *pc;
2170 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002171
Christoph Lameter188fd062012-05-09 10:09:55 -05002172 freelist = get_partial(s, flags, node, c);
2173
2174 if (freelist)
2175 return freelist;
2176
2177 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002178 if (page) {
2179 c = __this_cpu_ptr(s->cpu_slab);
2180 if (c->page)
2181 flush_slab(s, c);
2182
2183 /*
2184 * No other reference to the page yet so we can
2185 * muck around with it freely without cmpxchg
2186 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002187 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002188 page->freelist = NULL;
2189
2190 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002191 c->page = page;
2192 *pc = c;
2193 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002194 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002195
Christoph Lameter6faa6832012-05-09 10:09:51 -05002196 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002197}
2198
Mel Gorman072bb0a2012-07-31 16:43:58 -07002199static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2200{
2201 if (unlikely(PageSlabPfmemalloc(page)))
2202 return gfp_pfmemalloc_allowed(gfpflags);
2203
2204 return true;
2205}
2206
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002207/*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002208 * Check the page->freelist of a page and either transfer the freelist to the
2209 * per cpu freelist or deactivate the page.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002210 *
2211 * The page is still frozen if the return value is not NULL.
2212 *
2213 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002214 *
2215 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002216 */
2217static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2218{
2219 struct page new;
2220 unsigned long counters;
2221 void *freelist;
2222
2223 do {
2224 freelist = page->freelist;
2225 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002226
Christoph Lameter213eeb92011-11-11 14:07:14 -06002227 new.counters = counters;
2228 VM_BUG_ON(!new.frozen);
2229
2230 new.inuse = page->objects;
2231 new.frozen = freelist != NULL;
2232
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002233 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002234 freelist, counters,
2235 NULL, new.counters,
2236 "get_freelist"));
2237
2238 return freelist;
2239}
2240
2241/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002242 * Slow path. The lockless freelist is empty or we need to perform
2243 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002244 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002245 * Processing is still very fast if new objects have been freed to the
2246 * regular freelist. In that case we simply take over the regular freelist
2247 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002248 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002249 * If that is not working then we fall back to the partial lists. We take the
2250 * first element of the freelist as the object to allocate now and move the
2251 * rest of the freelist to the lockless freelist.
2252 *
2253 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002254 * we need to allocate a new slab. This is the slowest path since it involves
2255 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07002256 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002257static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2258 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002259{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002260 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002261 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002262 unsigned long flags;
2263
2264 local_irq_save(flags);
2265#ifdef CONFIG_PREEMPT
2266 /*
2267 * We may have been preempted and rescheduled on a different
2268 * cpu before disabling interrupts. Need to reload cpu area
2269 * pointer.
2270 */
2271 c = this_cpu_ptr(s->cpu_slab);
2272#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002273
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002274 page = c->page;
2275 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002276 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002277redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002278
Christoph Lameter57d437d2012-05-09 10:09:59 -05002279 if (unlikely(!node_match(page, node))) {
Christoph Lametere36a2652011-06-01 12:25:57 -05002280 stat(s, ALLOC_NODE_MISMATCH);
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002281 deactivate_slab(s, page, c->freelist);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002282 c->page = NULL;
2283 c->freelist = NULL;
Christoph Lameterfc59c052011-06-01 12:25:56 -05002284 goto new_slab;
2285 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002286
Mel Gorman072bb0a2012-07-31 16:43:58 -07002287 /*
2288 * By rights, we should be searching for a slab page that was
2289 * PFMEMALLOC but right now, we are losing the pfmemalloc
2290 * information when the page leaves the per-cpu allocator
2291 */
2292 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2293 deactivate_slab(s, page, c->freelist);
2294 c->page = NULL;
2295 c->freelist = NULL;
2296 goto new_slab;
2297 }
2298
Eric Dumazet73736e02011-12-13 04:57:06 +01002299 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002300 freelist = c->freelist;
2301 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002302 goto load_freelist;
2303
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002304 stat(s, ALLOC_SLOWPATH);
2305
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002306 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002307
Christoph Lameter6faa6832012-05-09 10:09:51 -05002308 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002309 c->page = NULL;
2310 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002311 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002312 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002313
Christoph Lameter81819f02007-05-06 14:49:36 -07002314 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002315
Christoph Lameter894b8782007-05-10 03:15:16 -07002316load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002317 /*
2318 * freelist is pointing to the list of objects to be used.
2319 * page is pointing to the page from which the objects are obtained.
2320 * That page must be frozen for per cpu allocations to work.
2321 */
2322 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002323 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002324 c->tid = next_tid(c->tid);
2325 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002326 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002327
Christoph Lameter81819f02007-05-06 14:49:36 -07002328new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002329
Christoph Lameter49e22582011-08-09 16:12:27 -05002330 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002331 page = c->page = c->partial;
2332 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002333 stat(s, CPU_PARTIAL_ALLOC);
2334 c->freelist = NULL;
2335 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002336 }
2337
Christoph Lameter188fd062012-05-09 10:09:55 -05002338 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002339
Christoph Lameterf46974362012-05-09 10:09:54 -05002340 if (unlikely(!freelist)) {
2341 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2342 slab_out_of_memory(s, gfpflags, node);
Christoph Lameter01ad8a72011-04-15 14:48:14 -05002343
Christoph Lameterf46974362012-05-09 10:09:54 -05002344 local_irq_restore(flags);
2345 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002346 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002347
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002348 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002349 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002350 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002351
Christoph Lameter497b66f2011-08-09 16:12:26 -05002352 /* Only entered in the debug case */
Chen Gangd0e0ac92013-07-15 09:05:29 +08002353 if (kmem_cache_debug(s) &&
2354 !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002355 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002356
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002357 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002358 c->page = NULL;
2359 c->freelist = NULL;
Christoph Lametera71ae472011-05-25 09:47:43 -05002360 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002361 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002362}
2363
2364/*
2365 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2366 * have the fastpath folded into their functions. So no function call
2367 * overhead for requests that can be satisfied on the fastpath.
2368 *
2369 * The fastpath works by first checking if the lockless freelist can be used.
2370 * If not then __slab_alloc is called for slow processing.
2371 *
2372 * Otherwise we can simply pick the next object from the lockless free list.
2373 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002374static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002375 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002376{
Christoph Lameter894b8782007-05-10 03:15:16 -07002377 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002378 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002379 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002380 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002381
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002382 if (slab_pre_alloc_hook(s, gfpflags))
Akinobu Mita773ff602008-12-23 19:37:01 +09002383 return NULL;
2384
Glauber Costad79923f2012-12-18 14:22:48 -08002385 s = memcg_kmem_get_cache(s, gfpflags);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002386redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002387 /*
2388 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2389 * enabled. We may switch back and forth between cpus while
2390 * reading from one cpu area. That does not matter as long
2391 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002392 *
2393 * Preemption is disabled for the retrieval of the tid because that
2394 * must occur from the current processor. We cannot allow rescheduling
2395 * on a different processor between the determination of the pointer
2396 * and the retrieval of the tid.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002397 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002398 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002399 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002400
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002401 /*
2402 * The transaction ids are globally unique per cpu and per operation on
2403 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2404 * occurs on the right processor and that there was no operation on the
2405 * linked list in between.
2406 */
2407 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002408 preempt_enable();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002409
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002410 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002411 page = c->page;
Libinac6434e2013-07-18 15:39:51 +08002412 if (unlikely(!object || !node_match(page, node)))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002413 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07002414
2415 else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002416 void *next_object = get_freepointer_safe(s, object);
2417
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002418 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002419 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002420 * operation and if we are on the right processor.
2421 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002422 * The cmpxchg does the following atomically (without lock
2423 * semantics!)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002424 * 1. Relocate first pointer to the current per cpu area.
2425 * 2. Verify that tid and freelist have not been changed
2426 * 3. If they were not changed replace tid and freelist
2427 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002428 * Since this is without lock semantics the protection is only
2429 * against code executing on this cpu *not* from access by
2430 * other cpus.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002431 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002432 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002433 s->cpu_slab->freelist, s->cpu_slab->tid,
2434 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002435 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002436
2437 note_cmpxchg_failure("slab_alloc", s, tid);
2438 goto redo;
2439 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002440 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002441 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002442 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002443
Pekka Enberg74e21342009-11-25 20:14:48 +02002444 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002445 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002446
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002447 slab_post_alloc_hook(s, gfpflags, object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002448
Christoph Lameter894b8782007-05-10 03:15:16 -07002449 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002450}
2451
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002452static __always_inline void *slab_alloc(struct kmem_cache *s,
2453 gfp_t gfpflags, unsigned long addr)
2454{
2455 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2456}
2457
Christoph Lameter81819f02007-05-06 14:49:36 -07002458void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2459{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002460 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002461
Chen Gangd0e0ac92013-07-15 09:05:29 +08002462 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2463 s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002464
2465 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002466}
2467EXPORT_SYMBOL(kmem_cache_alloc);
2468
Li Zefan0f24f122009-12-11 15:45:30 +08002469#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002470void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002471{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002472 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002473 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2474 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002475}
Richard Kennedy4a923792010-10-21 10:29:19 +01002476EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002477#endif
2478
Christoph Lameter81819f02007-05-06 14:49:36 -07002479#ifdef CONFIG_NUMA
2480void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2481{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002482 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002483
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002484 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002485 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002486
2487 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002488}
2489EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002490
Li Zefan0f24f122009-12-11 15:45:30 +08002491#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002492void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002493 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002494 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002495{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002496 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002497
2498 trace_kmalloc_node(_RET_IP_, ret,
2499 size, s->size, gfpflags, node);
2500 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002501}
Richard Kennedy4a923792010-10-21 10:29:19 +01002502EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002503#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002504#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002505
Christoph Lameter81819f02007-05-06 14:49:36 -07002506/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002507 * Slow patch handling. This may still be called frequently since objects
2508 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002509 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002510 * So we still attempt to reduce cache line usage. Just take the slab
2511 * lock and free the item. If there is no additional partial page
2512 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002513 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002514static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterff120592009-12-18 16:26:22 -06002515 void *x, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07002516{
2517 void *prior;
2518 void **object = (void *)x;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002519 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002520 struct page new;
2521 unsigned long counters;
2522 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002523 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002524
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002525 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002526
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002527 if (kmem_cache_debug(s) &&
2528 !(n = free_debug_processing(s, page, x, addr, &flags)))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002529 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002530
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002531 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002532 if (unlikely(n)) {
2533 spin_unlock_irqrestore(&n->list_lock, flags);
2534 n = NULL;
2535 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002536 prior = page->freelist;
2537 counters = page->counters;
2538 set_freepointer(s, object, prior);
2539 new.counters = counters;
2540 was_frozen = new.frozen;
2541 new.inuse--;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002542 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002543
Joonsoo Kim345c9052013-06-19 14:05:52 +09002544 if (kmem_cache_has_cpu_partial(s) && !prior)
Christoph Lameter49e22582011-08-09 16:12:27 -05002545
2546 /*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002547 * Slab was on no list before and will be
2548 * partially empty
2549 * We can defer the list move and instead
2550 * freeze it.
Christoph Lameter49e22582011-08-09 16:12:27 -05002551 */
2552 new.frozen = 1;
2553
2554 else { /* Needs to be taken off a list */
2555
2556 n = get_node(s, page_to_nid(page));
2557 /*
2558 * Speculatively acquire the list_lock.
2559 * If the cmpxchg does not succeed then we may
2560 * drop the list_lock without any processing.
2561 *
2562 * Otherwise the list_lock will synchronize with
2563 * other processors updating the list of slabs.
2564 */
2565 spin_lock_irqsave(&n->list_lock, flags);
2566
2567 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002568 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002569
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002570 } while (!cmpxchg_double_slab(s, page,
2571 prior, counters,
2572 object, new.counters,
2573 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002574
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002575 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002576
2577 /*
2578 * If we just froze the page then put it onto the
2579 * per cpu partial list.
2580 */
Alex Shi8028dce2012-02-03 23:34:56 +08002581 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002582 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002583 stat(s, CPU_PARTIAL_FREE);
2584 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002585 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002586 * The list lock was not taken therefore no list
2587 * activity can be necessary.
2588 */
2589 if (was_frozen)
2590 stat(s, FREE_FROZEN);
Christoph Lameter80f08c12011-06-01 12:25:55 -05002591 return;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002592 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002593
Joonsoo Kim837d6782012-08-16 00:02:40 +09002594 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2595 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002596
Joonsoo Kim837d6782012-08-16 00:02:40 +09002597 /*
2598 * Objects left in the slab. If it was not on the partial list before
2599 * then add it.
2600 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002601 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2602 if (kmem_cache_debug(s))
2603 remove_full(s, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002604 add_partial(n, page, DEACTIVATE_TO_TAIL);
2605 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002606 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002607 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002608 return;
2609
2610slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002611 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002612 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002613 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002614 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002615 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002616 stat(s, FREE_REMOVE_PARTIAL);
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002617 } else
2618 /* Slab must be on the full list */
2619 remove_full(s, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002620
Christoph Lameter80f08c12011-06-01 12:25:55 -05002621 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002622 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002623 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002624}
2625
Christoph Lameter894b8782007-05-10 03:15:16 -07002626/*
2627 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2628 * can perform fastpath freeing without additional function calls.
2629 *
2630 * The fastpath is only possible if we are freeing to the current cpu slab
2631 * of this processor. This typically the case if we have just allocated
2632 * the item before.
2633 *
2634 * If fastpath is not possible then fall back to __slab_free where we deal
2635 * with all sorts of special processing.
2636 */
Pekka Enberg06428782008-01-07 23:20:27 -08002637static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002638 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002639{
2640 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002641 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002642 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002643
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002644 slab_free_hook(s, x);
2645
Christoph Lametera24c5a02011-03-15 12:45:21 -05002646redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002647 /*
2648 * Determine the currently cpus per cpu slab.
2649 * The cpu may change afterward. However that does not matter since
2650 * data is retrieved via this pointer. If we are on the same cpu
2651 * during the cmpxchg then the free will succedd.
2652 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002653 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002654 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002655
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002656 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002657 preempt_enable();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002658
Christoph Lameter442b06b2011-05-17 16:29:31 -05002659 if (likely(page == c->page)) {
Christoph Lameterff120592009-12-18 16:26:22 -06002660 set_freepointer(s, object, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002661
Christoph Lameter933393f2011-12-22 11:58:51 -06002662 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002663 s->cpu_slab->freelist, s->cpu_slab->tid,
2664 c->freelist, tid,
2665 object, next_tid(tid)))) {
2666
2667 note_cmpxchg_failure("slab_free", s, tid);
2668 goto redo;
2669 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002670 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002671 } else
Christoph Lameterff120592009-12-18 16:26:22 -06002672 __slab_free(s, page, x, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002673
Christoph Lameter894b8782007-05-10 03:15:16 -07002674}
2675
Christoph Lameter81819f02007-05-06 14:49:36 -07002676void kmem_cache_free(struct kmem_cache *s, void *x)
2677{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002678 s = cache_from_obj(s, x);
2679 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002680 return;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002681 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002682 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002683}
2684EXPORT_SYMBOL(kmem_cache_free);
2685
Christoph Lameter81819f02007-05-06 14:49:36 -07002686/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002687 * Object placement in a slab is made very easy because we always start at
2688 * offset 0. If we tune the size of the object to the alignment then we can
2689 * get the required alignment by putting one properly sized object after
2690 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002691 *
2692 * Notice that the allocation order determines the sizes of the per cpu
2693 * caches. Each processor has always one slab available for allocations.
2694 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07002695 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07002696 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07002697 */
2698
2699/*
2700 * Mininum / Maximum order of slab pages. This influences locking overhead
2701 * and slab fragmentation. A higher order reduces the number of partial slabs
2702 * and increases the number of allocations possible without having to
2703 * take the list_lock.
2704 */
2705static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03002706static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002707static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002708
2709/*
2710 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07002711 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07002712 */
2713static int slub_nomerge;
2714
2715/*
Christoph Lameter81819f02007-05-06 14:49:36 -07002716 * Calculate the order of allocation given an slab object size.
2717 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002718 * The order of allocation has significant impact on performance and other
2719 * system components. Generally order 0 allocations should be preferred since
2720 * order 0 does not cause fragmentation in the page allocator. Larger objects
2721 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002722 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07002723 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07002724 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002725 * In order to reach satisfactory performance we must ensure that a minimum
2726 * number of objects is in one slab. Otherwise we may generate too much
2727 * activity on the partial lists which requires taking the list_lock. This is
2728 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07002729 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002730 * slub_max_order specifies the order where we begin to stop considering the
2731 * number of objects in a slab as critical. If we reach slub_max_order then
2732 * we try to keep the page order as low as possible. So we accept more waste
2733 * of space in favor of a small page order.
2734 *
2735 * Higher order allocations also allow the placement of more objects in a
2736 * slab and thereby reduce object handling overhead. If the user has
2737 * requested a higher mininum order then we start with that one instead of
2738 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07002739 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002740static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002741 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002742{
2743 int order;
2744 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07002745 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002746
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002747 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04002748 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03002749
Christoph Lameter6300ea72007-07-17 04:03:20 -07002750 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002751 fls(min_objects * size - 1) - PAGE_SHIFT);
2752 order <= max_order; order++) {
2753
Christoph Lameter81819f02007-05-06 14:49:36 -07002754 unsigned long slab_size = PAGE_SIZE << order;
2755
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002756 if (slab_size < min_objects * size + reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002757 continue;
2758
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002759 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002760
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002761 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07002762 break;
2763
2764 }
Christoph Lameter672bba32007-05-09 02:32:39 -07002765
Christoph Lameter81819f02007-05-06 14:49:36 -07002766 return order;
2767}
2768
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002769static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002770{
2771 int order;
2772 int min_objects;
2773 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002774 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002775
2776 /*
2777 * Attempt to find best configuration for a slab. This
2778 * works by first attempting to generate a layout with
2779 * the best configuration and backing off gradually.
2780 *
2781 * First we reduce the acceptable waste in a slab. Then
2782 * we reduce the minimum objects required in a slab.
2783 */
2784 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002785 if (!min_objects)
2786 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002787 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002788 min_objects = min(min_objects, max_objects);
2789
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002790 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002791 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002792 while (fraction >= 4) {
2793 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002794 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002795 if (order <= slub_max_order)
2796 return order;
2797 fraction /= 2;
2798 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03002799 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002800 }
2801
2802 /*
2803 * We were unable to place multiple objects in a slab. Now
2804 * lets see if we can place a single object there.
2805 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002806 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002807 if (order <= slub_max_order)
2808 return order;
2809
2810 /*
2811 * Doh this slab cannot be placed using slub_max_order.
2812 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002813 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03002814 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002815 return order;
2816 return -ENOSYS;
2817}
2818
Pekka Enberg5595cff2008-08-05 09:28:47 +03002819static void
Joonsoo Kim40534972012-05-11 00:50:47 +09002820init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002821{
2822 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002823 spin_lock_init(&n->list_lock);
2824 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002825#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002826 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07002827 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07002828 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002829#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002830}
2831
Christoph Lameter55136592010-08-20 12:37:13 -05002832static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002833{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05002834 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00002835 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002836
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002837 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002838 * Must align to double word boundary for the double cmpxchg
2839 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002840 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002841 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2842 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002843
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002844 if (!s->cpu_slab)
2845 return 0;
2846
2847 init_kmem_cache_cpus(s);
2848
2849 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002850}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002851
Christoph Lameter51df1142010-08-20 12:37:15 -05002852static struct kmem_cache *kmem_cache_node;
2853
Christoph Lameter81819f02007-05-06 14:49:36 -07002854/*
2855 * No kmalloc_node yet so do it by hand. We know that this is the first
2856 * slab on the node for this slabcache. There are no concurrent accesses
2857 * possible.
2858 *
Zhi Yong Wu721ae222013-11-08 20:47:37 +08002859 * Note that this function only works on the kmem_cache_node
2860 * when allocating for the kmem_cache_node. This is used for bootstrapping
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002861 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002862 */
Christoph Lameter55136592010-08-20 12:37:13 -05002863static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002864{
2865 struct page *page;
2866 struct kmem_cache_node *n;
2867
Christoph Lameter51df1142010-08-20 12:37:15 -05002868 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07002869
Christoph Lameter51df1142010-08-20 12:37:15 -05002870 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002871
2872 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002873 if (page_to_nid(page) != node) {
2874 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2875 "node %d\n", node);
2876 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2877 "in order to be able to continue\n");
2878 }
2879
Christoph Lameter81819f02007-05-06 14:49:36 -07002880 n = page->freelist;
2881 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002882 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05002883 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05002884 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05002885 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002886#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05002887 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05002888 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002889#endif
Joonsoo Kim40534972012-05-11 00:50:47 +09002890 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002891 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002892
Shaohua Li136333d2011-08-24 08:57:52 +08002893 add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07002894}
2895
2896static void free_kmem_cache_nodes(struct kmem_cache *s)
2897{
2898 int node;
2899
Christoph Lameterf64dc582007-10-16 01:25:33 -07002900 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002901 struct kmem_cache_node *n = s->node[node];
Christoph Lameter51df1142010-08-20 12:37:15 -05002902
Alexander Duyck73367bd2010-05-21 14:41:35 -07002903 if (n)
Christoph Lameter51df1142010-08-20 12:37:15 -05002904 kmem_cache_free(kmem_cache_node, n);
2905
Christoph Lameter81819f02007-05-06 14:49:36 -07002906 s->node[node] = NULL;
2907 }
2908}
2909
Christoph Lameter55136592010-08-20 12:37:13 -05002910static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002911{
2912 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07002913
Christoph Lameterf64dc582007-10-16 01:25:33 -07002914 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002915 struct kmem_cache_node *n;
2916
Alexander Duyck73367bd2010-05-21 14:41:35 -07002917 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05002918 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002919 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07002920 }
Christoph Lameter51df1142010-08-20 12:37:15 -05002921 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05002922 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002923
2924 if (!n) {
2925 free_kmem_cache_nodes(s);
2926 return 0;
2927 }
2928
Christoph Lameter81819f02007-05-06 14:49:36 -07002929 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09002930 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07002931 }
2932 return 1;
2933}
Christoph Lameter81819f02007-05-06 14:49:36 -07002934
David Rientjesc0bdb232009-02-25 09:16:35 +02002935static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002936{
2937 if (min < MIN_PARTIAL)
2938 min = MIN_PARTIAL;
2939 else if (min > MAX_PARTIAL)
2940 min = MAX_PARTIAL;
2941 s->min_partial = min;
2942}
2943
Christoph Lameter81819f02007-05-06 14:49:36 -07002944/*
2945 * calculate_sizes() determines the order and the distribution of data within
2946 * a slab object.
2947 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002948static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002949{
2950 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002951 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002952 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002953
2954 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002955 * Round up object size to the next word boundary. We can only
2956 * place the free pointer at word boundaries and this determines
2957 * the possible location of the free pointer.
2958 */
2959 size = ALIGN(size, sizeof(void *));
2960
2961#ifdef CONFIG_SLUB_DEBUG
2962 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002963 * Determine if we can poison the object itself. If the user of
2964 * the slab may touch the object after free or before allocation
2965 * then we should never poison the object itself.
2966 */
2967 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002968 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002969 s->flags |= __OBJECT_POISON;
2970 else
2971 s->flags &= ~__OBJECT_POISON;
2972
Christoph Lameter81819f02007-05-06 14:49:36 -07002973
2974 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002975 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002976 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002977 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002978 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002979 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07002980 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002981#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002982
2983 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002984 * With that we have determined the number of bytes in actual use
2985 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002986 */
2987 s->inuse = size;
2988
2989 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002990 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002991 /*
2992 * Relocate free pointer after the object if it is not
2993 * permitted to overwrite the first word of the object on
2994 * kmem_cache_free.
2995 *
2996 * This is the case if we do RCU, have a constructor or
2997 * destructor or are poisoning the objects.
2998 */
2999 s->offset = size;
3000 size += sizeof(void *);
3001 }
3002
Christoph Lameterc12b3c62007-05-23 13:57:31 -07003003#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07003004 if (flags & SLAB_STORE_USER)
3005 /*
3006 * Need to store information about allocs and frees after
3007 * the object.
3008 */
3009 size += 2 * sizeof(struct track);
3010
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07003011 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07003012 /*
3013 * Add some empty padding so that we can catch
3014 * overwrites from earlier objects rather than let
3015 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01003016 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07003017 * of the object.
3018 */
3019 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003020#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07003021
Christoph Lameter81819f02007-05-06 14:49:36 -07003022 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003023 * SLUB stores one object immediately after another beginning from
3024 * offset 0. In order to align the objects we have to simply size
3025 * each object to conform to the alignment.
3026 */
Christoph Lameter45906852012-11-28 16:23:16 +00003027 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07003028 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003029 if (forced_order >= 0)
3030 order = forced_order;
3031 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003032 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07003033
Christoph Lameter834f3d12008-04-14 19:11:31 +03003034 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003035 return 0;
3036
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003037 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003038 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003039 s->allocflags |= __GFP_COMP;
3040
3041 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003042 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003043
3044 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3045 s->allocflags |= __GFP_RECLAIMABLE;
3046
Christoph Lameter81819f02007-05-06 14:49:36 -07003047 /*
3048 * Determine the number of objects per slab
3049 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003050 s->oo = oo_make(order, size, s->reserved);
3051 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003052 if (oo_objects(s->oo) > oo_objects(s->max))
3053 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003054
Christoph Lameter834f3d12008-04-14 19:11:31 +03003055 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003056}
3057
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003058static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003059{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003060 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003061 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003062
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003063 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3064 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003065
Christoph Lameter06b285d2008-04-14 19:11:41 +03003066 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003067 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003068 if (disable_higher_order_debug) {
3069 /*
3070 * Disable debugging flags that store metadata if the min slab
3071 * order increased.
3072 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003073 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003074 s->flags &= ~DEBUG_METADATA_FLAGS;
3075 s->offset = 0;
3076 if (!calculate_sizes(s, -1))
3077 goto error;
3078 }
3079 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003080
Heiko Carstens25654092012-01-12 17:17:33 -08003081#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3082 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003083 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3084 /* Enable fast mode */
3085 s->flags |= __CMPXCHG_DOUBLE;
3086#endif
3087
David Rientjes3b89d7d2009-02-22 17:40:07 -08003088 /*
3089 * The larger the object size is, the more pages we want on the partial
3090 * list to avoid pounding the page allocator excessively.
3091 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003092 set_min_partial(s, ilog2(s->size) / 2);
3093
3094 /*
3095 * cpu_partial determined the maximum number of objects kept in the
3096 * per cpu partial lists of a processor.
3097 *
3098 * Per cpu partial lists mainly contain slabs that just have one
3099 * object freed. If they are used for allocation then they can be
3100 * filled up again with minimal effort. The slab will never hit the
3101 * per node partial lists and therefore no locking will be required.
3102 *
3103 * This setting also determines
3104 *
3105 * A) The number of objects from per cpu partial slabs dumped to the
3106 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003107 * B) The number of objects in cpu partial slabs to extract from the
Chen Gangd0e0ac92013-07-15 09:05:29 +08003108 * per node list when we run out of per cpu objects. We only fetch
3109 * 50% to keep some capacity around for frees.
Christoph Lameter49e22582011-08-09 16:12:27 -05003110 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003111 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003112 s->cpu_partial = 0;
3113 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003114 s->cpu_partial = 2;
3115 else if (s->size >= 1024)
3116 s->cpu_partial = 6;
3117 else if (s->size >= 256)
3118 s->cpu_partial = 13;
3119 else
3120 s->cpu_partial = 30;
3121
Christoph Lameter81819f02007-05-06 14:49:36 -07003122#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003123 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003124#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003125 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003126 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003127
Christoph Lameter55136592010-08-20 12:37:13 -05003128 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003129 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003130
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003131 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003132error:
3133 if (flags & SLAB_PANIC)
3134 panic("Cannot create slab %s size=%lu realsize=%u "
3135 "order=%u offset=%u flags=%lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +08003136 s->name, (unsigned long)s->size, s->size,
3137 oo_order(s->oo), s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003138 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003139}
Christoph Lameter81819f02007-05-06 14:49:36 -07003140
Christoph Lameter33b12c32008-04-25 12:22:43 -07003141static void list_slab_objects(struct kmem_cache *s, struct page *page,
3142 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003143{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003144#ifdef CONFIG_SLUB_DEBUG
3145 void *addr = page_address(page);
3146 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003147 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3148 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003149 if (!map)
3150 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003151 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003152 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003153
Christoph Lameter5f80b132011-04-15 14:48:13 -05003154 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003155 for_each_object(p, s, addr, page->objects) {
3156
3157 if (!test_bit(slab_index(p, s, addr), map)) {
3158 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3159 p, p - addr);
3160 print_tracking(s, p);
3161 }
3162 }
3163 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003164 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003165#endif
3166}
3167
Christoph Lameter81819f02007-05-06 14:49:36 -07003168/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003169 * Attempt to free all partial slabs on a node.
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003170 * This is called from kmem_cache_close(). We must be the last thread
3171 * using the cache and therefore we do not need to lock anymore.
Christoph Lameter81819f02007-05-06 14:49:36 -07003172 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003173static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003174{
Christoph Lameter81819f02007-05-06 14:49:36 -07003175 struct page *page, *h;
3176
Christoph Lameter33b12c32008-04-25 12:22:43 -07003177 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003178 if (!page->inuse) {
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05003179 remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003180 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003181 } else {
3182 list_slab_objects(s, page,
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003183 "Objects remaining in %s on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003184 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003185 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003186}
3187
3188/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003189 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003190 */
Christoph Lameter0c710012007-07-17 04:03:24 -07003191static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003192{
3193 int node;
3194
3195 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003196 /* Attempt to free all objects */
Christoph Lameterf64dc582007-10-16 01:25:33 -07003197 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003198 struct kmem_cache_node *n = get_node(s, node);
3199
Christoph Lameter599870b2008-04-23 12:36:52 -07003200 free_partial(s, n);
3201 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003202 return 1;
3203 }
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003204 free_percpu(s->cpu_slab);
Christoph Lameter81819f02007-05-06 14:49:36 -07003205 free_kmem_cache_nodes(s);
3206 return 0;
3207}
3208
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003209int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003210{
Christoph Lameter12c36672012-09-04 23:38:33 +00003211 int rc = kmem_cache_close(s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003212
Glauber Costa5413dfb2012-12-18 14:23:13 -08003213 if (!rc) {
3214 /*
3215 * We do the same lock strategy around sysfs_slab_add, see
3216 * __kmem_cache_create. Because this is pretty much the last
3217 * operation we do and the lock will be released shortly after
3218 * that in slab_common.c, we could just move sysfs_slab_remove
3219 * to a later point in common code. We should do that when we
3220 * have a common sysfs framework for all allocators.
3221 */
3222 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003223 sysfs_slab_remove(s);
Glauber Costa5413dfb2012-12-18 14:23:13 -08003224 mutex_lock(&slab_mutex);
3225 }
Christoph Lameter12c36672012-09-04 23:38:33 +00003226
3227 return rc;
Christoph Lameter81819f02007-05-06 14:49:36 -07003228}
Christoph Lameter81819f02007-05-06 14:49:36 -07003229
3230/********************************************************************
3231 * Kmalloc subsystem
3232 *******************************************************************/
3233
Christoph Lameter81819f02007-05-06 14:49:36 -07003234static int __init setup_slub_min_order(char *str)
3235{
Pekka Enberg06428782008-01-07 23:20:27 -08003236 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003237
3238 return 1;
3239}
3240
3241__setup("slub_min_order=", setup_slub_min_order);
3242
3243static int __init setup_slub_max_order(char *str)
3244{
Pekka Enberg06428782008-01-07 23:20:27 -08003245 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003246 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003247
3248 return 1;
3249}
3250
3251__setup("slub_max_order=", setup_slub_max_order);
3252
3253static int __init setup_slub_min_objects(char *str)
3254{
Pekka Enberg06428782008-01-07 23:20:27 -08003255 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003256
3257 return 1;
3258}
3259
3260__setup("slub_min_objects=", setup_slub_min_objects);
3261
3262static int __init setup_slub_nomerge(char *str)
3263{
3264 slub_nomerge = 1;
3265 return 1;
3266}
3267
3268__setup("slub_nomerge", setup_slub_nomerge);
3269
Christoph Lameter81819f02007-05-06 14:49:36 -07003270void *__kmalloc(size_t size, gfp_t flags)
3271{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003272 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003273 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003274
Christoph Lameter95a05b42013-01-10 19:14:19 +00003275 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003276 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003277
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003278 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003279
3280 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003281 return s;
3282
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003283 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003284
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003285 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003286
3287 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003288}
3289EXPORT_SYMBOL(__kmalloc);
3290
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003291#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003292static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3293{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003294 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003295 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003296
Glauber Costad79923f2012-12-18 14:22:48 -08003297 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003298 page = alloc_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003299 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003300 ptr = page_address(page);
3301
Roman Bobnievd56791b2013-10-08 15:58:57 -07003302 kmalloc_large_node_hook(ptr, size, flags);
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003303 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003304}
3305
Christoph Lameter81819f02007-05-06 14:49:36 -07003306void *__kmalloc_node(size_t size, gfp_t flags, int node)
3307{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003308 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003309 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003310
Christoph Lameter95a05b42013-01-10 19:14:19 +00003311 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003312 ret = kmalloc_large_node(size, flags, node);
3313
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003314 trace_kmalloc_node(_RET_IP_, ret,
3315 size, PAGE_SIZE << get_order(size),
3316 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003317
3318 return ret;
3319 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003320
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003321 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003322
3323 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003324 return s;
3325
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003326 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003327
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003328 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003329
3330 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003331}
3332EXPORT_SYMBOL(__kmalloc_node);
3333#endif
3334
3335size_t ksize(const void *object)
3336{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003337 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003338
Christoph Lameteref8b4522007-10-16 01:24:46 -07003339 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003340 return 0;
3341
Vegard Nossum294a80a2007-12-04 23:45:30 -08003342 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003343
Pekka Enberg76994412008-05-22 19:22:25 +03003344 if (unlikely(!PageSlab(page))) {
3345 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003346 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003347 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003348
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003349 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003350}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003351EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003352
3353void kfree(const void *x)
3354{
Christoph Lameter81819f02007-05-06 14:49:36 -07003355 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003356 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003357
Pekka Enberg2121db72009-03-25 11:05:57 +02003358 trace_kfree(_RET_IP_, x);
3359
Satyam Sharma2408c552007-10-16 01:24:44 -07003360 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003361 return;
3362
Christoph Lameterb49af682007-05-06 14:49:41 -07003363 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003364 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003365 BUG_ON(!PageCompound(page));
Roman Bobnievd56791b2013-10-08 15:58:57 -07003366 kfree_hook(x);
Glauber Costad79923f2012-12-18 14:22:48 -08003367 __free_memcg_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003368 return;
3369 }
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003370 slab_free(page->slab_cache, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003371}
3372EXPORT_SYMBOL(kfree);
3373
Christoph Lameter2086d262007-05-06 14:49:46 -07003374/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003375 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3376 * the remaining slabs by the number of items in use. The slabs with the
3377 * most items in use come first. New allocations will then fill those up
3378 * and thus they can be removed from the partial lists.
3379 *
3380 * The slabs with the least items are placed last. This results in them
3381 * being allocated from last increasing the chance that the last objects
3382 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003383 */
3384int kmem_cache_shrink(struct kmem_cache *s)
3385{
3386 int node;
3387 int i;
3388 struct kmem_cache_node *n;
3389 struct page *page;
3390 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03003391 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07003392 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03003393 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07003394 unsigned long flags;
3395
3396 if (!slabs_by_inuse)
3397 return -ENOMEM;
3398
3399 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003400 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07003401 n = get_node(s, node);
3402
3403 if (!n->nr_partial)
3404 continue;
3405
Christoph Lameter834f3d12008-04-14 19:11:31 +03003406 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07003407 INIT_LIST_HEAD(slabs_by_inuse + i);
3408
3409 spin_lock_irqsave(&n->list_lock, flags);
3410
3411 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003412 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07003413 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003414 * Note that concurrent frees may occur while we hold the
3415 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003416 */
3417 list_for_each_entry_safe(page, t, &n->partial, lru) {
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003418 list_move(&page->lru, slabs_by_inuse + page->inuse);
3419 if (!page->inuse)
3420 n->nr_partial--;
Christoph Lameter2086d262007-05-06 14:49:46 -07003421 }
3422
Christoph Lameter2086d262007-05-06 14:49:46 -07003423 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003424 * Rebuild the partial list with the slabs filled up most
3425 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07003426 */
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003427 for (i = objects - 1; i > 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07003428 list_splice(slabs_by_inuse + i, n->partial.prev);
3429
Christoph Lameter2086d262007-05-06 14:49:46 -07003430 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003431
3432 /* Release empty slabs */
3433 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3434 discard_slab(s, page);
Christoph Lameter2086d262007-05-06 14:49:46 -07003435 }
3436
3437 kfree(slabs_by_inuse);
3438 return 0;
3439}
3440EXPORT_SYMBOL(kmem_cache_shrink);
3441
Yasunori Gotob9049e22007-10-21 16:41:37 -07003442static int slab_mem_going_offline_callback(void *arg)
3443{
3444 struct kmem_cache *s;
3445
Christoph Lameter18004c52012-07-06 15:25:12 -05003446 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003447 list_for_each_entry(s, &slab_caches, list)
3448 kmem_cache_shrink(s);
Christoph Lameter18004c52012-07-06 15:25:12 -05003449 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003450
3451 return 0;
3452}
3453
3454static void slab_mem_offline_callback(void *arg)
3455{
3456 struct kmem_cache_node *n;
3457 struct kmem_cache *s;
3458 struct memory_notify *marg = arg;
3459 int offline_node;
3460
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003461 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003462
3463 /*
3464 * If the node still has available memory. we need kmem_cache_node
3465 * for it yet.
3466 */
3467 if (offline_node < 0)
3468 return;
3469
Christoph Lameter18004c52012-07-06 15:25:12 -05003470 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003471 list_for_each_entry(s, &slab_caches, list) {
3472 n = get_node(s, offline_node);
3473 if (n) {
3474 /*
3475 * if n->nr_slabs > 0, slabs still exist on the node
3476 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003477 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003478 * callback. So, we must fail.
3479 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003480 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003481
3482 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003483 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003484 }
3485 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003486 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003487}
3488
3489static int slab_mem_going_online_callback(void *arg)
3490{
3491 struct kmem_cache_node *n;
3492 struct kmem_cache *s;
3493 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003494 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003495 int ret = 0;
3496
3497 /*
3498 * If the node's memory is already available, then kmem_cache_node is
3499 * already created. Nothing to do.
3500 */
3501 if (nid < 0)
3502 return 0;
3503
3504 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003505 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003506 * allocate a kmem_cache_node structure in order to bring the node
3507 * online.
3508 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003509 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003510 list_for_each_entry(s, &slab_caches, list) {
3511 /*
3512 * XXX: kmem_cache_alloc_node will fallback to other nodes
3513 * since memory is not yet available from the node that
3514 * is brought up.
3515 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003516 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003517 if (!n) {
3518 ret = -ENOMEM;
3519 goto out;
3520 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003521 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003522 s->node[nid] = n;
3523 }
3524out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003525 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003526 return ret;
3527}
3528
3529static int slab_memory_callback(struct notifier_block *self,
3530 unsigned long action, void *arg)
3531{
3532 int ret = 0;
3533
3534 switch (action) {
3535 case MEM_GOING_ONLINE:
3536 ret = slab_mem_going_online_callback(arg);
3537 break;
3538 case MEM_GOING_OFFLINE:
3539 ret = slab_mem_going_offline_callback(arg);
3540 break;
3541 case MEM_OFFLINE:
3542 case MEM_CANCEL_ONLINE:
3543 slab_mem_offline_callback(arg);
3544 break;
3545 case MEM_ONLINE:
3546 case MEM_CANCEL_OFFLINE:
3547 break;
3548 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003549 if (ret)
3550 ret = notifier_from_errno(ret);
3551 else
3552 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003553 return ret;
3554}
3555
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003556static struct notifier_block slab_memory_callback_nb = {
3557 .notifier_call = slab_memory_callback,
3558 .priority = SLAB_CALLBACK_PRI,
3559};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003560
Christoph Lameter81819f02007-05-06 14:49:36 -07003561/********************************************************************
3562 * Basic setup of slabs
3563 *******************************************************************/
3564
Christoph Lameter51df1142010-08-20 12:37:15 -05003565/*
3566 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003567 * the page allocator. Allocate them properly then fix up the pointers
3568 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003569 */
3570
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003571static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003572{
3573 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003574 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameter51df1142010-08-20 12:37:15 -05003575
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003576 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003577
Glauber Costa7d557b32013-02-22 20:20:00 +04003578 /*
3579 * This runs very early, and only the boot processor is supposed to be
3580 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3581 * IPIs around.
3582 */
3583 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter51df1142010-08-20 12:37:15 -05003584 for_each_node_state(node, N_NORMAL_MEMORY) {
3585 struct kmem_cache_node *n = get_node(s, node);
3586 struct page *p;
3587
3588 if (n) {
3589 list_for_each_entry(p, &n->partial, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003590 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003591
Li Zefan607bf322011-04-12 15:22:26 +08003592#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter51df1142010-08-20 12:37:15 -05003593 list_for_each_entry(p, &n->full, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003594 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003595#endif
3596 }
3597 }
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003598 list_add(&s->list, &slab_caches);
3599 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003600}
3601
Christoph Lameter81819f02007-05-06 14:49:36 -07003602void __init kmem_cache_init(void)
3603{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003604 static __initdata struct kmem_cache boot_kmem_cache,
3605 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003606
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003607 if (debug_guardpage_minorder())
3608 slub_max_order = 0;
3609
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003610 kmem_cache_node = &boot_kmem_cache_node;
3611 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003612
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003613 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3614 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003615
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003616 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003617
3618 /* Able to allocate the per node structures */
3619 slab_state = PARTIAL;
3620
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003621 create_boot_cache(kmem_cache, "kmem_cache",
3622 offsetof(struct kmem_cache, node) +
3623 nr_node_ids * sizeof(struct kmem_cache_node *),
3624 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003625
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003626 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003627
Christoph Lameter51df1142010-08-20 12:37:15 -05003628 /*
3629 * Allocate kmem_cache_node properly from the kmem_cache slab.
3630 * kmem_cache_node is separately allocated so no need to
3631 * update any list pointers.
3632 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003633 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003634
3635 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003636 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003637
3638#ifdef CONFIG_SMP
3639 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003640#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003641
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003642 printk(KERN_INFO
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003643 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003644 " CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003645 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003646 slub_min_order, slub_max_order, slub_min_objects,
3647 nr_cpu_ids, nr_node_ids);
3648}
3649
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003650void __init kmem_cache_init_late(void)
3651{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003652}
3653
Christoph Lameter81819f02007-05-06 14:49:36 -07003654/*
3655 * Find a mergeable slab cache
3656 */
3657static int slab_unmergeable(struct kmem_cache *s)
3658{
3659 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3660 return 1;
3661
Christoph Lameterc59def92007-05-16 22:10:50 -07003662 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003663 return 1;
3664
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003665 /*
3666 * We may have set a slab to be unmergeable during bootstrap.
3667 */
3668 if (s->refcount < 0)
3669 return 1;
3670
Christoph Lameter81819f02007-05-06 14:49:36 -07003671 return 0;
3672}
3673
Glauber Costa2633d7a2012-12-18 14:22:34 -08003674static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003675 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003676 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003677{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003678 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003679
3680 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3681 return NULL;
3682
Christoph Lameterc59def92007-05-16 22:10:50 -07003683 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003684 return NULL;
3685
3686 size = ALIGN(size, sizeof(void *));
3687 align = calculate_alignment(flags, align, size);
3688 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003689 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003690
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003691 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003692 if (slab_unmergeable(s))
3693 continue;
3694
3695 if (size > s->size)
3696 continue;
3697
Christoph Lameterba0268a2007-09-11 15:24:11 -07003698 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003699 continue;
3700 /*
3701 * Check if alignment is compatible.
3702 * Courtesy of Adrian Drzewiecki
3703 */
Pekka Enberg06428782008-01-07 23:20:27 -08003704 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003705 continue;
3706
3707 if (s->size - size >= sizeof(void *))
3708 continue;
3709
Glauber Costa2633d7a2012-12-18 14:22:34 -08003710 if (!cache_match_memcg(s, memcg))
3711 continue;
3712
Christoph Lameter81819f02007-05-06 14:49:36 -07003713 return s;
3714 }
3715 return NULL;
3716}
3717
Glauber Costa2633d7a2012-12-18 14:22:34 -08003718struct kmem_cache *
3719__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3720 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003721{
3722 struct kmem_cache *s;
3723
Glauber Costa2633d7a2012-12-18 14:22:34 -08003724 s = find_mergeable(memcg, size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003725 if (s) {
3726 s->refcount++;
3727 /*
3728 * Adjust the object sizes so that we clear
3729 * the complete object on kzalloc.
3730 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003731 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003732 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003733
David Rientjes7b8f3b62008-12-17 22:09:46 -08003734 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003735 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003736 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003737 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003738 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003739
Christoph Lametercbb79692012-09-05 00:18:32 +00003740 return s;
3741}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003742
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003743int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003744{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003745 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003746
Pekka Enbergaac3a162012-09-05 12:07:44 +03003747 err = kmem_cache_open(s, flags);
3748 if (err)
3749 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003750
Christoph Lameter45530c42012-11-28 16:23:07 +00003751 /* Mutex is not taken during early boot */
3752 if (slab_state <= UP)
3753 return 0;
3754
Glauber Costa107dab52012-12-18 14:23:05 -08003755 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003756 mutex_unlock(&slab_mutex);
3757 err = sysfs_slab_add(s);
3758 mutex_lock(&slab_mutex);
Christoph Lameter20cea962012-07-06 15:25:13 -05003759
Pekka Enbergaac3a162012-09-05 12:07:44 +03003760 if (err)
3761 kmem_cache_close(s);
3762
3763 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07003764}
Christoph Lameter81819f02007-05-06 14:49:36 -07003765
Christoph Lameter81819f02007-05-06 14:49:36 -07003766#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003767/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003768 * Use the cpu notifier to insure that the cpu slabs are flushed when
3769 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003770 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04003771static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07003772 unsigned long action, void *hcpu)
3773{
3774 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003775 struct kmem_cache *s;
3776 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003777
3778 switch (action) {
3779 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003780 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003781 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003782 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05003783 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003784 list_for_each_entry(s, &slab_caches, list) {
3785 local_irq_save(flags);
3786 __flush_cpu_slab(s, cpu);
3787 local_irq_restore(flags);
3788 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003789 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003790 break;
3791 default:
3792 break;
3793 }
3794 return NOTIFY_OK;
3795}
3796
Paul Gortmaker0db06282013-06-19 14:53:51 -04003797static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003798 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003799};
Christoph Lameter81819f02007-05-06 14:49:36 -07003800
3801#endif
3802
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003803void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003804{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003805 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003806 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003807
Christoph Lameter95a05b42013-01-10 19:14:19 +00003808 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003809 return kmalloc_large(size, gfpflags);
3810
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003811 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003812
Satyam Sharma2408c552007-10-16 01:24:44 -07003813 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003814 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003815
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003816 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003817
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003818 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003819 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003820
3821 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003822}
3823
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003824#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07003825void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003826 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003827{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003828 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003829 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003830
Christoph Lameter95a05b42013-01-10 19:14:19 +00003831 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08003832 ret = kmalloc_large_node(size, gfpflags, node);
3833
3834 trace_kmalloc_node(caller, ret,
3835 size, PAGE_SIZE << get_order(size),
3836 gfpflags, node);
3837
3838 return ret;
3839 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02003840
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003841 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003842
Satyam Sharma2408c552007-10-16 01:24:44 -07003843 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003844 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003845
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003846 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003847
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003848 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003849 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003850
3851 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003852}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003853#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003854
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003855#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03003856static int count_inuse(struct page *page)
3857{
3858 return page->inuse;
3859}
3860
3861static int count_total(struct page *page)
3862{
3863 return page->objects;
3864}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003865#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03003866
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003867#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07003868static int validate_slab(struct kmem_cache *s, struct page *page,
3869 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003870{
3871 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003872 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003873
3874 if (!check_slab(s, page) ||
3875 !on_freelist(s, page, NULL))
3876 return 0;
3877
3878 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003879 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003880
Christoph Lameter5f80b132011-04-15 14:48:13 -05003881 get_map(s, page, map);
3882 for_each_object(p, s, addr, page->objects) {
3883 if (test_bit(slab_index(p, s, addr), map))
3884 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3885 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003886 }
3887
Christoph Lameter224a88b2008-04-14 19:11:31 +03003888 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003889 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02003890 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003891 return 0;
3892 return 1;
3893}
3894
Christoph Lameter434e2452007-07-17 04:03:30 -07003895static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3896 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003897{
Christoph Lameter881db7f2011-06-01 12:25:53 -05003898 slab_lock(page);
3899 validate_slab(s, page, map);
3900 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003901}
3902
Christoph Lameter434e2452007-07-17 04:03:30 -07003903static int validate_slab_node(struct kmem_cache *s,
3904 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003905{
3906 unsigned long count = 0;
3907 struct page *page;
3908 unsigned long flags;
3909
3910 spin_lock_irqsave(&n->list_lock, flags);
3911
3912 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003913 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003914 count++;
3915 }
3916 if (count != n->nr_partial)
3917 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3918 "counter=%ld\n", s->name, count, n->nr_partial);
3919
3920 if (!(s->flags & SLAB_STORE_USER))
3921 goto out;
3922
3923 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003924 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003925 count++;
3926 }
3927 if (count != atomic_long_read(&n->nr_slabs))
3928 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3929 "counter=%ld\n", s->name, count,
3930 atomic_long_read(&n->nr_slabs));
3931
3932out:
3933 spin_unlock_irqrestore(&n->list_lock, flags);
3934 return count;
3935}
3936
Christoph Lameter434e2452007-07-17 04:03:30 -07003937static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003938{
3939 int node;
3940 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003941 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003942 sizeof(unsigned long), GFP_KERNEL);
3943
3944 if (!map)
3945 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003946
3947 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003948 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003949 struct kmem_cache_node *n = get_node(s, node);
3950
Christoph Lameter434e2452007-07-17 04:03:30 -07003951 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003952 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003953 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003954 return count;
3955}
Christoph Lameter88a420e2007-05-06 14:49:45 -07003956/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003957 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003958 * and freed.
3959 */
3960
3961struct location {
3962 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003963 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003964 long long sum_time;
3965 long min_time;
3966 long max_time;
3967 long min_pid;
3968 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303969 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003970 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003971};
3972
3973struct loc_track {
3974 unsigned long max;
3975 unsigned long count;
3976 struct location *loc;
3977};
3978
3979static void free_loc_track(struct loc_track *t)
3980{
3981 if (t->max)
3982 free_pages((unsigned long)t->loc,
3983 get_order(sizeof(struct location) * t->max));
3984}
3985
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003986static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003987{
3988 struct location *l;
3989 int order;
3990
Christoph Lameter88a420e2007-05-06 14:49:45 -07003991 order = get_order(sizeof(struct location) * max);
3992
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003993 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003994 if (!l)
3995 return 0;
3996
3997 if (t->count) {
3998 memcpy(l, t->loc, sizeof(struct location) * t->count);
3999 free_loc_track(t);
4000 }
4001 t->max = max;
4002 t->loc = l;
4003 return 1;
4004}
4005
4006static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004007 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004008{
4009 long start, end, pos;
4010 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004011 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004012 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004013
4014 start = -1;
4015 end = t->count;
4016
4017 for ( ; ; ) {
4018 pos = start + (end - start + 1) / 2;
4019
4020 /*
4021 * There is nothing at "end". If we end up there
4022 * we need to add something to before end.
4023 */
4024 if (pos == end)
4025 break;
4026
4027 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004028 if (track->addr == caddr) {
4029
4030 l = &t->loc[pos];
4031 l->count++;
4032 if (track->when) {
4033 l->sum_time += age;
4034 if (age < l->min_time)
4035 l->min_time = age;
4036 if (age > l->max_time)
4037 l->max_time = age;
4038
4039 if (track->pid < l->min_pid)
4040 l->min_pid = track->pid;
4041 if (track->pid > l->max_pid)
4042 l->max_pid = track->pid;
4043
Rusty Russell174596a2009-01-01 10:12:29 +10304044 cpumask_set_cpu(track->cpu,
4045 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004046 }
4047 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004048 return 1;
4049 }
4050
Christoph Lameter45edfa52007-05-09 02:32:45 -07004051 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004052 end = pos;
4053 else
4054 start = pos;
4055 }
4056
4057 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004058 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004059 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004060 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004061 return 0;
4062
4063 l = t->loc + pos;
4064 if (pos < t->count)
4065 memmove(l + 1, l,
4066 (t->count - pos) * sizeof(struct location));
4067 t->count++;
4068 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004069 l->addr = track->addr;
4070 l->sum_time = age;
4071 l->min_time = age;
4072 l->max_time = age;
4073 l->min_pid = track->pid;
4074 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304075 cpumask_clear(to_cpumask(l->cpus));
4076 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004077 nodes_clear(l->nodes);
4078 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004079 return 1;
4080}
4081
4082static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004083 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004084 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004085{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004086 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004087 void *p;
4088
Christoph Lameter39b26462008-04-14 19:11:30 +03004089 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004090 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004091
Christoph Lameter224a88b2008-04-14 19:11:31 +03004092 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004093 if (!test_bit(slab_index(p, s, addr), map))
4094 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004095}
4096
4097static int list_locations(struct kmem_cache *s, char *buf,
4098 enum track_item alloc)
4099{
Harvey Harrisone374d482008-01-31 15:20:50 -08004100 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004101 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004102 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004103 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004104 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4105 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004106
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004107 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4108 GFP_TEMPORARY)) {
4109 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004110 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004111 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004112 /* Push back cpu slabs */
4113 flush_all(s);
4114
Christoph Lameterf64dc582007-10-16 01:25:33 -07004115 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004116 struct kmem_cache_node *n = get_node(s, node);
4117 unsigned long flags;
4118 struct page *page;
4119
Christoph Lameter9e869432007-08-22 14:01:56 -07004120 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004121 continue;
4122
4123 spin_lock_irqsave(&n->list_lock, flags);
4124 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004125 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004126 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004127 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004128 spin_unlock_irqrestore(&n->list_lock, flags);
4129 }
4130
4131 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004132 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004133
Hugh Dickins9c246242008-12-09 13:14:27 -08004134 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004135 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004136 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004137
4138 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004139 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004140 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004141 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004142
4143 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004144 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004145 l->min_time,
4146 (long)div_u64(l->sum_time, l->count),
4147 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004148 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004149 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004150 l->min_time);
4151
4152 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004153 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004154 l->min_pid, l->max_pid);
4155 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004156 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004157 l->min_pid);
4158
Rusty Russell174596a2009-01-01 10:12:29 +10304159 if (num_online_cpus() > 1 &&
4160 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004161 len < PAGE_SIZE - 60) {
4162 len += sprintf(buf + len, " cpus=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004163 len += cpulist_scnprintf(buf + len,
4164 PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10304165 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004166 }
4167
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004168 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004169 len < PAGE_SIZE - 60) {
4170 len += sprintf(buf + len, " nodes=");
Chen Gangd0e0ac92013-07-15 09:05:29 +08004171 len += nodelist_scnprintf(buf + len,
4172 PAGE_SIZE - len - 50,
4173 l->nodes);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004174 }
4175
Harvey Harrisone374d482008-01-31 15:20:50 -08004176 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07004177 }
4178
4179 free_loc_track(&t);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004180 kfree(map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004181 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08004182 len += sprintf(buf, "No data\n");
4183 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004184}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004185#endif
Christoph Lameter88a420e2007-05-06 14:49:45 -07004186
Christoph Lametera5a84752010-10-05 13:57:27 -05004187#ifdef SLUB_RESILIENCY_TEST
4188static void resiliency_test(void)
4189{
4190 u8 *p;
4191
Christoph Lameter95a05b42013-01-10 19:14:19 +00004192 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
Christoph Lametera5a84752010-10-05 13:57:27 -05004193
4194 printk(KERN_ERR "SLUB resiliency testing\n");
4195 printk(KERN_ERR "-----------------------\n");
4196 printk(KERN_ERR "A. Corruption after allocation\n");
4197
4198 p = kzalloc(16, GFP_KERNEL);
4199 p[16] = 0x12;
4200 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4201 " 0x12->0x%p\n\n", p + 16);
4202
4203 validate_slab_cache(kmalloc_caches[4]);
4204
4205 /* Hmmm... The next two are dangerous */
4206 p = kzalloc(32, GFP_KERNEL);
4207 p[32 + sizeof(void *)] = 0x34;
4208 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4209 " 0x34 -> -0x%p\n", p);
4210 printk(KERN_ERR
4211 "If allocated object is overwritten then not detectable\n\n");
4212
4213 validate_slab_cache(kmalloc_caches[5]);
4214 p = kzalloc(64, GFP_KERNEL);
4215 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4216 *p = 0x56;
4217 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4218 p);
4219 printk(KERN_ERR
4220 "If allocated object is overwritten then not detectable\n\n");
4221 validate_slab_cache(kmalloc_caches[6]);
4222
4223 printk(KERN_ERR "\nB. Corruption after free\n");
4224 p = kzalloc(128, GFP_KERNEL);
4225 kfree(p);
4226 *p = 0x78;
4227 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4228 validate_slab_cache(kmalloc_caches[7]);
4229
4230 p = kzalloc(256, GFP_KERNEL);
4231 kfree(p);
4232 p[50] = 0x9a;
4233 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4234 p);
4235 validate_slab_cache(kmalloc_caches[8]);
4236
4237 p = kzalloc(512, GFP_KERNEL);
4238 kfree(p);
4239 p[512] = 0xab;
4240 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4241 validate_slab_cache(kmalloc_caches[9]);
4242}
4243#else
4244#ifdef CONFIG_SYSFS
4245static void resiliency_test(void) {};
4246#endif
4247#endif
4248
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004249#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -07004250enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03004251 SL_ALL, /* All slabs */
4252 SL_PARTIAL, /* Only partially allocated slabs */
4253 SL_CPU, /* Only slabs used for cpu caches */
4254 SL_OBJECTS, /* Determine allocated objects not slabs */
4255 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07004256};
4257
Christoph Lameter205ab992008-04-14 19:11:40 +03004258#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004259#define SO_PARTIAL (1 << SL_PARTIAL)
4260#define SO_CPU (1 << SL_CPU)
4261#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03004262#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004263
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004264static ssize_t show_slab_objects(struct kmem_cache *s,
4265 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07004266{
4267 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07004268 int node;
4269 int x;
4270 unsigned long *nodes;
Christoph Lameter81819f02007-05-06 14:49:36 -07004271
Chen Gange35e1a92013-07-12 08:23:48 +08004272 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004273 if (!nodes)
4274 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004275
Christoph Lameter205ab992008-04-14 19:11:40 +03004276 if (flags & SO_CPU) {
4277 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004278
Christoph Lameter205ab992008-04-14 19:11:40 +03004279 for_each_possible_cpu(cpu) {
Chen Gangd0e0ac92013-07-15 09:05:29 +08004280 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4281 cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004282 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004283 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004284
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004285 page = ACCESS_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004286 if (!page)
4287 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004288
Christoph Lameterec3ab082012-05-09 10:09:56 -05004289 node = page_to_nid(page);
4290 if (flags & SO_TOTAL)
4291 x = page->objects;
4292 else if (flags & SO_OBJECTS)
4293 x = page->inuse;
4294 else
4295 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004296
Christoph Lameterec3ab082012-05-09 10:09:56 -05004297 total += x;
4298 nodes[node] += x;
4299
4300 page = ACCESS_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004301 if (page) {
4302 x = page->pobjects;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004303 total += x;
4304 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004305 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004306 }
4307 }
4308
Christoph Lameter04d94872011-01-10 10:15:15 -06004309 lock_memory_hotplug();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004310#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004311 if (flags & SO_ALL) {
4312 for_each_node_state(node, N_NORMAL_MEMORY) {
4313 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07004314
Chen Gangd0e0ac92013-07-15 09:05:29 +08004315 if (flags & SO_TOTAL)
4316 x = atomic_long_read(&n->total_objects);
4317 else if (flags & SO_OBJECTS)
4318 x = atomic_long_read(&n->total_objects) -
4319 count_partial(n, count_free);
Christoph Lameter205ab992008-04-14 19:11:40 +03004320 else
4321 x = atomic_long_read(&n->nr_slabs);
4322 total += x;
4323 nodes[node] += x;
4324 }
4325
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004326 } else
4327#endif
4328 if (flags & SO_PARTIAL) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004329 for_each_node_state(node, N_NORMAL_MEMORY) {
4330 struct kmem_cache_node *n = get_node(s, node);
4331
4332 if (flags & SO_TOTAL)
4333 x = count_partial(n, count_total);
4334 else if (flags & SO_OBJECTS)
4335 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004336 else
4337 x = n->nr_partial;
4338 total += x;
4339 nodes[node] += x;
4340 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004341 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004342 x = sprintf(buf, "%lu", total);
4343#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07004344 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07004345 if (nodes[node])
4346 x += sprintf(buf + x, " N%d=%lu",
4347 node, nodes[node]);
4348#endif
Christoph Lameter04d94872011-01-10 10:15:15 -06004349 unlock_memory_hotplug();
Christoph Lameter81819f02007-05-06 14:49:36 -07004350 kfree(nodes);
4351 return x + sprintf(buf + x, "\n");
4352}
4353
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004354#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004355static int any_slab_objects(struct kmem_cache *s)
4356{
4357 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004358
4359 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004360 struct kmem_cache_node *n = get_node(s, node);
4361
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004362 if (!n)
4363 continue;
4364
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004365 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004366 return 1;
4367 }
4368 return 0;
4369}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004370#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004371
4372#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004373#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004374
4375struct slab_attribute {
4376 struct attribute attr;
4377 ssize_t (*show)(struct kmem_cache *s, char *buf);
4378 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4379};
4380
4381#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004382 static struct slab_attribute _name##_attr = \
4383 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004384
4385#define SLAB_ATTR(_name) \
4386 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004387 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004388
Christoph Lameter81819f02007-05-06 14:49:36 -07004389static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4390{
4391 return sprintf(buf, "%d\n", s->size);
4392}
4393SLAB_ATTR_RO(slab_size);
4394
4395static ssize_t align_show(struct kmem_cache *s, char *buf)
4396{
4397 return sprintf(buf, "%d\n", s->align);
4398}
4399SLAB_ATTR_RO(align);
4400
4401static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4402{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004403 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004404}
4405SLAB_ATTR_RO(object_size);
4406
4407static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4408{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004409 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004410}
4411SLAB_ATTR_RO(objs_per_slab);
4412
Christoph Lameter06b285d2008-04-14 19:11:41 +03004413static ssize_t order_store(struct kmem_cache *s,
4414 const char *buf, size_t length)
4415{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004416 unsigned long order;
4417 int err;
4418
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004419 err = kstrtoul(buf, 10, &order);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004420 if (err)
4421 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004422
4423 if (order > slub_max_order || order < slub_min_order)
4424 return -EINVAL;
4425
4426 calculate_sizes(s, order);
4427 return length;
4428}
4429
Christoph Lameter81819f02007-05-06 14:49:36 -07004430static ssize_t order_show(struct kmem_cache *s, char *buf)
4431{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004432 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004433}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004434SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004435
David Rientjes73d342b2009-02-22 17:40:09 -08004436static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4437{
4438 return sprintf(buf, "%lu\n", s->min_partial);
4439}
4440
4441static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4442 size_t length)
4443{
4444 unsigned long min;
4445 int err;
4446
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004447 err = kstrtoul(buf, 10, &min);
David Rientjes73d342b2009-02-22 17:40:09 -08004448 if (err)
4449 return err;
4450
David Rientjesc0bdb232009-02-25 09:16:35 +02004451 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004452 return length;
4453}
4454SLAB_ATTR(min_partial);
4455
Christoph Lameter49e22582011-08-09 16:12:27 -05004456static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4457{
4458 return sprintf(buf, "%u\n", s->cpu_partial);
4459}
4460
4461static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4462 size_t length)
4463{
4464 unsigned long objects;
4465 int err;
4466
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004467 err = kstrtoul(buf, 10, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05004468 if (err)
4469 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004470 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004471 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004472
4473 s->cpu_partial = objects;
4474 flush_all(s);
4475 return length;
4476}
4477SLAB_ATTR(cpu_partial);
4478
Christoph Lameter81819f02007-05-06 14:49:36 -07004479static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4480{
Joe Perches62c70bc2011-01-13 15:45:52 -08004481 if (!s->ctor)
4482 return 0;
4483 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004484}
4485SLAB_ATTR_RO(ctor);
4486
Christoph Lameter81819f02007-05-06 14:49:36 -07004487static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4488{
4489 return sprintf(buf, "%d\n", s->refcount - 1);
4490}
4491SLAB_ATTR_RO(aliases);
4492
Christoph Lameter81819f02007-05-06 14:49:36 -07004493static ssize_t partial_show(struct kmem_cache *s, char *buf)
4494{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004495 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004496}
4497SLAB_ATTR_RO(partial);
4498
4499static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4500{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004501 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004502}
4503SLAB_ATTR_RO(cpu_slabs);
4504
4505static ssize_t objects_show(struct kmem_cache *s, char *buf)
4506{
Christoph Lameter205ab992008-04-14 19:11:40 +03004507 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004508}
4509SLAB_ATTR_RO(objects);
4510
Christoph Lameter205ab992008-04-14 19:11:40 +03004511static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4512{
4513 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4514}
4515SLAB_ATTR_RO(objects_partial);
4516
Christoph Lameter49e22582011-08-09 16:12:27 -05004517static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4518{
4519 int objects = 0;
4520 int pages = 0;
4521 int cpu;
4522 int len;
4523
4524 for_each_online_cpu(cpu) {
4525 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4526
4527 if (page) {
4528 pages += page->pages;
4529 objects += page->pobjects;
4530 }
4531 }
4532
4533 len = sprintf(buf, "%d(%d)", objects, pages);
4534
4535#ifdef CONFIG_SMP
4536 for_each_online_cpu(cpu) {
4537 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4538
4539 if (page && len < PAGE_SIZE - 20)
4540 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4541 page->pobjects, page->pages);
4542 }
4543#endif
4544 return len + sprintf(buf + len, "\n");
4545}
4546SLAB_ATTR_RO(slabs_cpu_partial);
4547
Christoph Lameter81819f02007-05-06 14:49:36 -07004548static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4549{
4550 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4551}
4552
4553static ssize_t reclaim_account_store(struct kmem_cache *s,
4554 const char *buf, size_t length)
4555{
4556 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4557 if (buf[0] == '1')
4558 s->flags |= SLAB_RECLAIM_ACCOUNT;
4559 return length;
4560}
4561SLAB_ATTR(reclaim_account);
4562
4563static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4564{
Christoph Lameter5af60832007-05-06 14:49:56 -07004565 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004566}
4567SLAB_ATTR_RO(hwcache_align);
4568
4569#ifdef CONFIG_ZONE_DMA
4570static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4571{
4572 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4573}
4574SLAB_ATTR_RO(cache_dma);
4575#endif
4576
4577static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4578{
4579 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4580}
4581SLAB_ATTR_RO(destroy_by_rcu);
4582
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004583static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4584{
4585 return sprintf(buf, "%d\n", s->reserved);
4586}
4587SLAB_ATTR_RO(reserved);
4588
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004589#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004590static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4591{
4592 return show_slab_objects(s, buf, SO_ALL);
4593}
4594SLAB_ATTR_RO(slabs);
4595
4596static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4597{
4598 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4599}
4600SLAB_ATTR_RO(total_objects);
4601
4602static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4603{
4604 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4605}
4606
4607static ssize_t sanity_checks_store(struct kmem_cache *s,
4608 const char *buf, size_t length)
4609{
4610 s->flags &= ~SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004611 if (buf[0] == '1') {
4612 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004613 s->flags |= SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004614 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004615 return length;
4616}
4617SLAB_ATTR(sanity_checks);
4618
4619static ssize_t trace_show(struct kmem_cache *s, char *buf)
4620{
4621 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4622}
4623
4624static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4625 size_t length)
4626{
4627 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004628 if (buf[0] == '1') {
4629 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004630 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004631 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004632 return length;
4633}
4634SLAB_ATTR(trace);
4635
Christoph Lameter81819f02007-05-06 14:49:36 -07004636static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4637{
4638 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4639}
4640
4641static ssize_t red_zone_store(struct kmem_cache *s,
4642 const char *buf, size_t length)
4643{
4644 if (any_slab_objects(s))
4645 return -EBUSY;
4646
4647 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004648 if (buf[0] == '1') {
4649 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004650 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004651 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004652 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004653 return length;
4654}
4655SLAB_ATTR(red_zone);
4656
4657static ssize_t poison_show(struct kmem_cache *s, char *buf)
4658{
4659 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4660}
4661
4662static ssize_t poison_store(struct kmem_cache *s,
4663 const char *buf, size_t length)
4664{
4665 if (any_slab_objects(s))
4666 return -EBUSY;
4667
4668 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004669 if (buf[0] == '1') {
4670 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004671 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004672 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004673 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004674 return length;
4675}
4676SLAB_ATTR(poison);
4677
4678static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4679{
4680 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4681}
4682
4683static ssize_t store_user_store(struct kmem_cache *s,
4684 const char *buf, size_t length)
4685{
4686 if (any_slab_objects(s))
4687 return -EBUSY;
4688
4689 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004690 if (buf[0] == '1') {
4691 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004692 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004693 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004694 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004695 return length;
4696}
4697SLAB_ATTR(store_user);
4698
Christoph Lameter53e15af2007-05-06 14:49:43 -07004699static ssize_t validate_show(struct kmem_cache *s, char *buf)
4700{
4701 return 0;
4702}
4703
4704static ssize_t validate_store(struct kmem_cache *s,
4705 const char *buf, size_t length)
4706{
Christoph Lameter434e2452007-07-17 04:03:30 -07004707 int ret = -EINVAL;
4708
4709 if (buf[0] == '1') {
4710 ret = validate_slab_cache(s);
4711 if (ret >= 0)
4712 ret = length;
4713 }
4714 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004715}
4716SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004717
4718static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4719{
4720 if (!(s->flags & SLAB_STORE_USER))
4721 return -ENOSYS;
4722 return list_locations(s, buf, TRACK_ALLOC);
4723}
4724SLAB_ATTR_RO(alloc_calls);
4725
4726static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4727{
4728 if (!(s->flags & SLAB_STORE_USER))
4729 return -ENOSYS;
4730 return list_locations(s, buf, TRACK_FREE);
4731}
4732SLAB_ATTR_RO(free_calls);
4733#endif /* CONFIG_SLUB_DEBUG */
4734
4735#ifdef CONFIG_FAILSLAB
4736static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4737{
4738 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4739}
4740
4741static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4742 size_t length)
4743{
4744 s->flags &= ~SLAB_FAILSLAB;
4745 if (buf[0] == '1')
4746 s->flags |= SLAB_FAILSLAB;
4747 return length;
4748}
4749SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004750#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07004751
Christoph Lameter2086d262007-05-06 14:49:46 -07004752static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4753{
4754 return 0;
4755}
4756
4757static ssize_t shrink_store(struct kmem_cache *s,
4758 const char *buf, size_t length)
4759{
4760 if (buf[0] == '1') {
4761 int rc = kmem_cache_shrink(s);
4762
4763 if (rc)
4764 return rc;
4765 } else
4766 return -EINVAL;
4767 return length;
4768}
4769SLAB_ATTR(shrink);
4770
Christoph Lameter81819f02007-05-06 14:49:36 -07004771#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004772static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004773{
Christoph Lameter98246012008-01-07 23:20:26 -08004774 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004775}
4776
Christoph Lameter98246012008-01-07 23:20:26 -08004777static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004778 const char *buf, size_t length)
4779{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004780 unsigned long ratio;
4781 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004782
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004783 err = kstrtoul(buf, 10, &ratio);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004784 if (err)
4785 return err;
4786
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004787 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004788 s->remote_node_defrag_ratio = ratio * 10;
4789
Christoph Lameter81819f02007-05-06 14:49:36 -07004790 return length;
4791}
Christoph Lameter98246012008-01-07 23:20:26 -08004792SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004793#endif
4794
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004795#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004796static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4797{
4798 unsigned long sum = 0;
4799 int cpu;
4800 int len;
4801 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4802
4803 if (!data)
4804 return -ENOMEM;
4805
4806 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004807 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004808
4809 data[cpu] = x;
4810 sum += x;
4811 }
4812
4813 len = sprintf(buf, "%lu", sum);
4814
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004815#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004816 for_each_online_cpu(cpu) {
4817 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004818 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004819 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004820#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004821 kfree(data);
4822 return len + sprintf(buf + len, "\n");
4823}
4824
David Rientjes78eb00c2009-10-15 02:20:22 -07004825static void clear_stat(struct kmem_cache *s, enum stat_item si)
4826{
4827 int cpu;
4828
4829 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004830 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07004831}
4832
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004833#define STAT_ATTR(si, text) \
4834static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4835{ \
4836 return show_stat(s, buf, si); \
4837} \
David Rientjes78eb00c2009-10-15 02:20:22 -07004838static ssize_t text##_store(struct kmem_cache *s, \
4839 const char *buf, size_t length) \
4840{ \
4841 if (buf[0] != '0') \
4842 return -EINVAL; \
4843 clear_stat(s, si); \
4844 return length; \
4845} \
4846SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004847
4848STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4849STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4850STAT_ATTR(FREE_FASTPATH, free_fastpath);
4851STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4852STAT_ATTR(FREE_FROZEN, free_frozen);
4853STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4854STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4855STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4856STAT_ATTR(ALLOC_SLAB, alloc_slab);
4857STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05004858STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004859STAT_ATTR(FREE_SLAB, free_slab);
4860STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4861STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4862STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4863STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4864STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4865STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05004866STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03004867STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05004868STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4869STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05004870STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4871STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08004872STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4873STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004874#endif
4875
Pekka Enberg06428782008-01-07 23:20:27 -08004876static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004877 &slab_size_attr.attr,
4878 &object_size_attr.attr,
4879 &objs_per_slab_attr.attr,
4880 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004881 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004882 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004883 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004884 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004885 &partial_attr.attr,
4886 &cpu_slabs_attr.attr,
4887 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004888 &aliases_attr.attr,
4889 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004890 &hwcache_align_attr.attr,
4891 &reclaim_account_attr.attr,
4892 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05004893 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004894 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004895 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004896#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004897 &total_objects_attr.attr,
4898 &slabs_attr.attr,
4899 &sanity_checks_attr.attr,
4900 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004901 &red_zone_attr.attr,
4902 &poison_attr.attr,
4903 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004904 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004905 &alloc_calls_attr.attr,
4906 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004907#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004908#ifdef CONFIG_ZONE_DMA
4909 &cache_dma_attr.attr,
4910#endif
4911#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004912 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004913#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004914#ifdef CONFIG_SLUB_STATS
4915 &alloc_fastpath_attr.attr,
4916 &alloc_slowpath_attr.attr,
4917 &free_fastpath_attr.attr,
4918 &free_slowpath_attr.attr,
4919 &free_frozen_attr.attr,
4920 &free_add_partial_attr.attr,
4921 &free_remove_partial_attr.attr,
4922 &alloc_from_partial_attr.attr,
4923 &alloc_slab_attr.attr,
4924 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05004925 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004926 &free_slab_attr.attr,
4927 &cpuslab_flush_attr.attr,
4928 &deactivate_full_attr.attr,
4929 &deactivate_empty_attr.attr,
4930 &deactivate_to_head_attr.attr,
4931 &deactivate_to_tail_attr.attr,
4932 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05004933 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004934 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05004935 &cmpxchg_double_fail_attr.attr,
4936 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004937 &cpu_partial_alloc_attr.attr,
4938 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08004939 &cpu_partial_node_attr.attr,
4940 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004941#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03004942#ifdef CONFIG_FAILSLAB
4943 &failslab_attr.attr,
4944#endif
4945
Christoph Lameter81819f02007-05-06 14:49:36 -07004946 NULL
4947};
4948
4949static struct attribute_group slab_attr_group = {
4950 .attrs = slab_attrs,
4951};
4952
4953static ssize_t slab_attr_show(struct kobject *kobj,
4954 struct attribute *attr,
4955 char *buf)
4956{
4957 struct slab_attribute *attribute;
4958 struct kmem_cache *s;
4959 int err;
4960
4961 attribute = to_slab_attr(attr);
4962 s = to_slab(kobj);
4963
4964 if (!attribute->show)
4965 return -EIO;
4966
4967 err = attribute->show(s, buf);
4968
4969 return err;
4970}
4971
4972static ssize_t slab_attr_store(struct kobject *kobj,
4973 struct attribute *attr,
4974 const char *buf, size_t len)
4975{
4976 struct slab_attribute *attribute;
4977 struct kmem_cache *s;
4978 int err;
4979
4980 attribute = to_slab_attr(attr);
4981 s = to_slab(kobj);
4982
4983 if (!attribute->store)
4984 return -EIO;
4985
4986 err = attribute->store(s, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08004987#ifdef CONFIG_MEMCG_KMEM
4988 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4989 int i;
Christoph Lameter81819f02007-05-06 14:49:36 -07004990
Glauber Costa107dab52012-12-18 14:23:05 -08004991 mutex_lock(&slab_mutex);
4992 if (s->max_attr_size < len)
4993 s->max_attr_size = len;
4994
Glauber Costaebe945c2012-12-18 14:23:10 -08004995 /*
4996 * This is a best effort propagation, so this function's return
4997 * value will be determined by the parent cache only. This is
4998 * basically because not all attributes will have a well
4999 * defined semantics for rollbacks - most of the actions will
5000 * have permanent effects.
5001 *
5002 * Returning the error value of any of the children that fail
5003 * is not 100 % defined, in the sense that users seeing the
5004 * error code won't be able to know anything about the state of
5005 * the cache.
5006 *
5007 * Only returning the error code for the parent cache at least
5008 * has well defined semantics. The cache being written to
5009 * directly either failed or succeeded, in which case we loop
5010 * through the descendants with best-effort propagation.
5011 */
Glauber Costa107dab52012-12-18 14:23:05 -08005012 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08005013 struct kmem_cache *c = cache_from_memcg_idx(s, i);
Glauber Costa107dab52012-12-18 14:23:05 -08005014 if (c)
5015 attribute->store(c, buf, len);
5016 }
5017 mutex_unlock(&slab_mutex);
5018 }
5019#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005020 return err;
5021}
5022
Glauber Costa107dab52012-12-18 14:23:05 -08005023static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5024{
5025#ifdef CONFIG_MEMCG_KMEM
5026 int i;
5027 char *buffer = NULL;
5028
5029 if (!is_root_cache(s))
5030 return;
5031
5032 /*
5033 * This mean this cache had no attribute written. Therefore, no point
5034 * in copying default values around
5035 */
5036 if (!s->max_attr_size)
5037 return;
5038
5039 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5040 char mbuf[64];
5041 char *buf;
5042 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5043
5044 if (!attr || !attr->store || !attr->show)
5045 continue;
5046
5047 /*
5048 * It is really bad that we have to allocate here, so we will
5049 * do it only as a fallback. If we actually allocate, though,
5050 * we can just use the allocated buffer until the end.
5051 *
5052 * Most of the slub attributes will tend to be very small in
5053 * size, but sysfs allows buffers up to a page, so they can
5054 * theoretically happen.
5055 */
5056 if (buffer)
5057 buf = buffer;
5058 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5059 buf = mbuf;
5060 else {
5061 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5062 if (WARN_ON(!buffer))
5063 continue;
5064 buf = buffer;
5065 }
5066
5067 attr->show(s->memcg_params->root_cache, buf);
5068 attr->store(s, buf, strlen(buf));
5069 }
5070
5071 if (buffer)
5072 free_page((unsigned long)buffer);
5073#endif
5074}
5075
Emese Revfy52cf25d2010-01-19 02:58:23 +01005076static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005077 .show = slab_attr_show,
5078 .store = slab_attr_store,
5079};
5080
5081static struct kobj_type slab_ktype = {
5082 .sysfs_ops = &slab_sysfs_ops,
5083};
5084
5085static int uevent_filter(struct kset *kset, struct kobject *kobj)
5086{
5087 struct kobj_type *ktype = get_ktype(kobj);
5088
5089 if (ktype == &slab_ktype)
5090 return 1;
5091 return 0;
5092}
5093
Emese Revfy9cd43612009-12-31 14:52:51 +01005094static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005095 .filter = uevent_filter,
5096};
5097
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005098static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005099
5100#define ID_STR_LENGTH 64
5101
5102/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005103 *
5104 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005105 */
5106static char *create_unique_id(struct kmem_cache *s)
5107{
5108 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5109 char *p = name;
5110
5111 BUG_ON(!name);
5112
5113 *p++ = ':';
5114 /*
5115 * First flags affecting slabcache operations. We will only
5116 * get here for aliasable slabs so we do not need to support
5117 * too many flags. The flags here must cover all flags that
5118 * are matched during merging to guarantee that the id is
5119 * unique.
5120 */
5121 if (s->flags & SLAB_CACHE_DMA)
5122 *p++ = 'd';
5123 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5124 *p++ = 'a';
5125 if (s->flags & SLAB_DEBUG_FREE)
5126 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005127 if (!(s->flags & SLAB_NOTRACK))
5128 *p++ = 't';
Christoph Lameter81819f02007-05-06 14:49:36 -07005129 if (p != name + 1)
5130 *p++ = '-';
5131 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005132
5133#ifdef CONFIG_MEMCG_KMEM
5134 if (!is_root_cache(s))
Chen Gangd0e0ac92013-07-15 09:05:29 +08005135 p += sprintf(p, "-%08d",
5136 memcg_cache_id(s->memcg_params->memcg));
Glauber Costa2633d7a2012-12-18 14:22:34 -08005137#endif
5138
Christoph Lameter81819f02007-05-06 14:49:36 -07005139 BUG_ON(p > name + ID_STR_LENGTH - 1);
5140 return name;
5141}
5142
5143static int sysfs_slab_add(struct kmem_cache *s)
5144{
5145 int err;
5146 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005147 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005148
Christoph Lameter81819f02007-05-06 14:49:36 -07005149 if (unmergeable) {
5150 /*
5151 * Slabcache can never be merged so we can use the name proper.
5152 * This is typically the case for debug situations. In that
5153 * case we can catch duplicate names easily.
5154 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005155 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005156 name = s->name;
5157 } else {
5158 /*
5159 * Create a unique name for the slab as a target
5160 * for the symlinks.
5161 */
5162 name = create_unique_id(s);
5163 }
5164
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005165 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005166 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5167 if (err) {
5168 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005169 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005170 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005171
5172 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005173 if (err) {
5174 kobject_del(&s->kobj);
5175 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005176 return err;
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005177 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005178 kobject_uevent(&s->kobj, KOBJ_ADD);
5179 if (!unmergeable) {
5180 /* Setup first alias */
5181 sysfs_slab_alias(s, s->name);
5182 kfree(name);
5183 }
5184 return 0;
5185}
5186
5187static void sysfs_slab_remove(struct kmem_cache *s)
5188{
Christoph Lameter97d06602012-07-06 15:25:11 -05005189 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005190 /*
5191 * Sysfs has not been setup yet so no need to remove the
5192 * cache from sysfs.
5193 */
5194 return;
5195
Christoph Lameter81819f02007-05-06 14:49:36 -07005196 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5197 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005198 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005199}
5200
5201/*
5202 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005203 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005204 */
5205struct saved_alias {
5206 struct kmem_cache *s;
5207 const char *name;
5208 struct saved_alias *next;
5209};
5210
Adrian Bunk5af328a2007-07-17 04:03:27 -07005211static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005212
5213static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5214{
5215 struct saved_alias *al;
5216
Christoph Lameter97d06602012-07-06 15:25:11 -05005217 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005218 /*
5219 * If we have a leftover link then remove it.
5220 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005221 sysfs_remove_link(&slab_kset->kobj, name);
5222 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005223 }
5224
5225 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5226 if (!al)
5227 return -ENOMEM;
5228
5229 al->s = s;
5230 al->name = name;
5231 al->next = alias_list;
5232 alias_list = al;
5233 return 0;
5234}
5235
5236static int __init slab_sysfs_init(void)
5237{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005238 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005239 int err;
5240
Christoph Lameter18004c52012-07-06 15:25:12 -05005241 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005242
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005243 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005244 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005245 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005246 printk(KERN_ERR "Cannot register slab subsystem.\n");
5247 return -ENOSYS;
5248 }
5249
Christoph Lameter97d06602012-07-06 15:25:11 -05005250 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005251
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005252 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005253 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005254 if (err)
5255 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5256 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005257 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005258
5259 while (alias_list) {
5260 struct saved_alias *al = alias_list;
5261
5262 alias_list = alias_list->next;
5263 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005264 if (err)
5265 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
Julia Lawall068ce412012-07-08 13:37:40 +02005266 " %s to sysfs\n", al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005267 kfree(al);
5268 }
5269
Christoph Lameter18004c52012-07-06 15:25:12 -05005270 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005271 resiliency_test();
5272 return 0;
5273}
5274
5275__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005276#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005277
5278/*
5279 * The /proc/slabinfo ABI
5280 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005281#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005282void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005283{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005284 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005285 unsigned long nr_objs = 0;
5286 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005287 int node;
5288
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005289 for_each_online_node(node) {
5290 struct kmem_cache_node *n = get_node(s, node);
5291
5292 if (!n)
5293 continue;
5294
Wanpeng Lic17fd132013-07-04 08:33:26 +08005295 nr_slabs += node_nr_slabs(n);
5296 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005297 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005298 }
5299
Glauber Costa0d7561c2012-10-19 18:20:27 +04005300 sinfo->active_objs = nr_objs - nr_free;
5301 sinfo->num_objs = nr_objs;
5302 sinfo->active_slabs = nr_slabs;
5303 sinfo->num_slabs = nr_slabs;
5304 sinfo->objects_per_slab = oo_objects(s->oo);
5305 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005306}
5307
Glauber Costa0d7561c2012-10-19 18:20:27 +04005308void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005309{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005310}
5311
Glauber Costab7454ad2012-10-19 18:20:25 +04005312ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5313 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005314{
Glauber Costab7454ad2012-10-19 18:20:25 +04005315 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005316}
Linus Torvalds158a9622008-01-02 13:04:48 -08005317#endif /* CONFIG_SLABINFO */