blob: e3ba1f2cf60cc0caa5dd7ed059dbe583b477e073 [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
Christoph Lameter881db7f2011-06-01 12:25:53 -05005 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07007 *
Christoph Lametercde53532008-07-04 09:59:22 -07008 * (C) 2007 SGI, Christoph Lameter
Christoph Lameter881db7f2011-06-01 12:25:53 -05009 * (C) 2011 Linux Foundation, Christoph Lameter
Christoph Lameter81819f02007-05-06 14:49:36 -070010 */
11
12#include <linux/mm.h>
Nick Piggin1eb5ac62009-05-05 19:13:44 +100013#include <linux/swap.h> /* struct reclaim_state */
Christoph Lameter81819f02007-05-06 14:49:36 -070014#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
Christoph Lameter97d06602012-07-06 15:25:11 -050019#include "slab.h"
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +040020#include <linux/proc_fs.h>
Andrew Morton3ac38fa2013-04-29 15:08:06 -070021#include <linux/notifier.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070022#include <linux/seq_file.h>
Vegard Nossum5a896d92008-04-04 00:54:48 +020023#include <linux/kmemcheck.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070024#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070028#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070029#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070030#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070031#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090032#include <linux/fault-inject.h>
Pekka Enbergbfa71452011-07-07 22:47:01 +030033#include <linux/stacktrace.h>
Christoph Lameter4de900b2012-01-30 15:53:51 -060034#include <linux/prefetch.h>
Glauber Costa2633d7a2012-12-18 14:22:34 -080035#include <linux/memcontrol.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070036
Richard Kennedy4a923792010-10-21 10:29:19 +010037#include <trace/events/kmem.h>
38
Mel Gorman072bb0a2012-07-31 16:43:58 -070039#include "internal.h"
40
Christoph Lameter81819f02007-05-06 14:49:36 -070041/*
42 * Lock order:
Christoph Lameter18004c52012-07-06 15:25:12 -050043 * 1. slab_mutex (Global Mutex)
Christoph Lameter881db7f2011-06-01 12:25:53 -050044 * 2. node->list_lock
45 * 3. slab_lock(page) (Only on some arches and for debugging)
Christoph Lameter81819f02007-05-06 14:49:36 -070046 *
Christoph Lameter18004c52012-07-06 15:25:12 -050047 * slab_mutex
Christoph Lameter881db7f2011-06-01 12:25:53 -050048 *
Christoph Lameter18004c52012-07-06 15:25:12 -050049 * The role of the slab_mutex is to protect the list of all the slabs
Christoph Lameter881db7f2011-06-01 12:25:53 -050050 * and to synchronize major metadata changes to slab cache structures.
51 *
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
58 *
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -070064 *
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
70 *
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
75 * the list lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070076 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
80 *
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
83 *
Christoph Lameter672bba32007-05-09 02:32:39 -070084 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070086 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070087 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070089 *
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
93 *
94 * Overloading of page flags that are otherwise used for LRU management.
95 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070096 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700108 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -0700109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700111 *
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700114 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700115 */
116
Christoph Lameteraf537b02010-07-09 14:07:14 -0500117static inline int kmem_cache_debug(struct kmem_cache *s)
118{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700119#ifdef CONFIG_SLUB_DEBUG
Christoph Lameteraf537b02010-07-09 14:07:14 -0500120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
Christoph Lameter5577bd82007-05-16 22:10:56 -0700121#else
Christoph Lameteraf537b02010-07-09 14:07:14 -0500122 return 0;
Christoph Lameter5577bd82007-05-16 22:10:56 -0700123#endif
Christoph Lameteraf537b02010-07-09 14:07:14 -0500124}
Christoph Lameter5577bd82007-05-16 22:10:56 -0700125
Joonsoo Kim345c9052013-06-19 14:05:52 +0900126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127{
128#ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
130#else
131 return false;
132#endif
133}
134
Christoph Lameter81819f02007-05-06 14:49:36 -0700135/*
136 * Issues still to be resolved:
137 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700140 * - Variable sizing of the per node arrays
141 */
142
143/* Enable to test recovery from slab corruption on boot */
144#undef SLUB_RESILIENCY_TEST
145
Christoph Lameterb789ef52011-06-01 12:25:49 -0500146/* Enable to log cmpxchg failures */
147#undef SLUB_DEBUG_CMPXCHG
148
Christoph Lameter81819f02007-05-06 14:49:36 -0700149/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800153#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700154
Christoph Lameter2086d262007-05-06 14:49:46 -0700155/*
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in the.
159 */
160#define MAX_PARTIAL 10
161
Christoph Lameter81819f02007-05-06 14:49:36 -0700162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700164
Christoph Lameter81819f02007-05-06 14:49:36 -0700165/*
David Rientjes3de47212009-07-27 18:30:35 -0700166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
168 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700169 */
David Rientjes3de47212009-07-27 18:30:35 -0700170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700171
172/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700173 * Set of flags that will prevent slab merging
174 */
175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +0300176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177 SLAB_FAILSLAB)
Christoph Lameter81819f02007-05-06 14:49:36 -0700178
179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
Vegard Nossum5a896d92008-04-04 00:54:48 +0200180 SLAB_CACHE_DMA | SLAB_NOTRACK)
Christoph Lameter81819f02007-05-06 14:49:36 -0700181
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400182#define OO_SHIFT 16
183#define OO_MASK ((1 << OO_SHIFT) - 1)
Christoph Lameter50d5c412011-06-01 12:25:45 -0500184#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400185
Christoph Lameter81819f02007-05-06 14:49:36 -0700186/* Internal SLUB flags */
Christoph Lameterf90ec392010-07-09 14:07:11 -0500187#define __OBJECT_POISON 0x80000000UL /* Poison object */
Christoph Lameterb789ef52011-06-01 12:25:49 -0500188#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
Christoph Lameter81819f02007-05-06 14:49:36 -0700189
Christoph Lameter81819f02007-05-06 14:49:36 -0700190#ifdef CONFIG_SMP
191static struct notifier_block slab_notifier;
192#endif
193
Christoph Lameter02cbc872007-05-09 02:32:43 -0700194/*
195 * Tracking user of a slab.
196 */
Ben Greeard6543e32011-07-07 11:36:36 -0700197#define TRACK_ADDRS_COUNT 16
Christoph Lameter02cbc872007-05-09 02:32:43 -0700198struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300199 unsigned long addr; /* Called from address */
Ben Greeard6543e32011-07-07 11:36:36 -0700200#ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202#endif
Christoph Lameter02cbc872007-05-09 02:32:43 -0700203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
206};
207
208enum track_item { TRACK_ALLOC, TRACK_FREE };
209
Christoph Lameterab4d5ed2010-10-05 13:57:26 -0500210#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -0700211static int sysfs_slab_add(struct kmem_cache *);
212static int sysfs_slab_alias(struct kmem_cache *, const char *);
213static void sysfs_slab_remove(struct kmem_cache *);
Glauber Costa107dab52012-12-18 14:23:05 -0800214static void memcg_propagate_slab_attrs(struct kmem_cache *s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700215#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700216static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
217static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 { return 0; }
Christoph Lameterdb265ec2012-09-04 23:18:33 +0000219static inline void sysfs_slab_remove(struct kmem_cache *s) { }
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800220
Glauber Costa107dab52012-12-18 14:23:05 -0800221static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
Christoph Lameter81819f02007-05-06 14:49:36 -0700222#endif
223
Christoph Lameter4fdccdf2011-03-22 13:35:00 -0500224static inline void stat(const struct kmem_cache *s, enum stat_item si)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800225{
226#ifdef CONFIG_SLUB_STATS
Christoph Lameter84e554e62009-12-18 16:26:23 -0600227 __this_cpu_inc(s->cpu_slab->stat[si]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800228#endif
229}
230
Christoph Lameter81819f02007-05-06 14:49:36 -0700231/********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
234
Christoph Lameter81819f02007-05-06 14:49:36 -0700235static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
236{
Christoph Lameter81819f02007-05-06 14:49:36 -0700237 return s->node[node];
Christoph Lameter81819f02007-05-06 14:49:36 -0700238}
239
Christoph Lameter6446faa2008-02-15 23:45:26 -0800240/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700241static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
243{
244 void *base;
245
Christoph Lametera973e9d2008-03-01 13:40:44 -0800246 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700247 return 1;
248
Christoph Lametera973e9d2008-03-01 13:40:44 -0800249 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300250 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700251 (object - base) % s->size) {
252 return 0;
253 }
254
255 return 1;
256}
257
Christoph Lameter7656c722007-05-09 02:32:40 -0700258static inline void *get_freepointer(struct kmem_cache *s, void *object)
259{
260 return *(void **)(object + s->offset);
261}
262
Eric Dumazet0ad95002011-12-16 16:25:34 +0100263static void prefetch_freepointer(const struct kmem_cache *s, void *object)
264{
265 prefetch(object + s->offset);
266}
267
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500268static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
269{
270 void *p;
271
272#ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
274#else
275 p = get_freepointer(s, object);
276#endif
277 return p;
278}
279
Christoph Lameter7656c722007-05-09 02:32:40 -0700280static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
281{
282 *(void **)(object + s->offset) = fp;
283}
284
285/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300286#define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700288 __p += (__s)->size)
289
Christoph Lameter7656c722007-05-09 02:32:40 -0700290/* Determine object index from a given position */
291static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292{
293 return (p - addr) / s->size;
294}
295
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100296static inline size_t slab_ksize(const struct kmem_cache *s)
297{
298#ifdef CONFIG_SLUB_DEBUG
299 /*
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
302 */
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500304 return s->object_size;
Mariusz Kozlowskid71f6062011-02-26 20:10:26 +0100305
306#endif
307 /*
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
311 */
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
313 return s->inuse;
314 /*
315 * Else we can use all the padding etc for the allocation
316 */
317 return s->size;
318}
319
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800320static inline int order_objects(int order, unsigned long size, int reserved)
321{
322 return ((PAGE_SIZE << order) - reserved) / size;
323}
324
Christoph Lameter834f3d12008-04-14 19:11:31 +0300325static inline struct kmem_cache_order_objects oo_make(int order,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800326 unsigned long size, int reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300327{
328 struct kmem_cache_order_objects x = {
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800329 (order << OO_SHIFT) + order_objects(order, size, reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300330 };
331
332 return x;
333}
334
335static inline int oo_order(struct kmem_cache_order_objects x)
336{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400337 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300338}
339
340static inline int oo_objects(struct kmem_cache_order_objects x)
341{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400342 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300343}
344
Christoph Lameter881db7f2011-06-01 12:25:53 -0500345/*
346 * Per slab locking using the pagelock
347 */
348static __always_inline void slab_lock(struct page *page)
349{
350 bit_spin_lock(PG_locked, &page->flags);
351}
352
353static __always_inline void slab_unlock(struct page *page)
354{
355 __bit_spin_unlock(PG_locked, &page->flags);
356}
357
Christoph Lameter1d071712011-07-14 12:49:12 -0500358/* Interrupts must be disabled (for the fallback code to work right) */
359static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
362 const char *n)
363{
Christoph Lameter1d071712011-07-14 12:49:12 -0500364 VM_BUG_ON(!irqs_disabled());
Heiko Carstens25654092012-01-12 17:17:33 -0800365#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -0500367 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000368 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500369 freelist_old, counters_old,
370 freelist_new, counters_new))
371 return 1;
372 } else
373#endif
374 {
Christoph Lameter881db7f2011-06-01 12:25:53 -0500375 slab_lock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500376 if (page->freelist == freelist_old && page->counters == counters_old) {
377 page->freelist = freelist_new;
378 page->counters = counters_new;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500379 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500380 return 1;
381 }
Christoph Lameter881db7f2011-06-01 12:25:53 -0500382 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500383 }
384
385 cpu_relax();
386 stat(s, CMPXCHG_DOUBLE_FAIL);
387
388#ifdef SLUB_DEBUG_CMPXCHG
389 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
390#endif
391
392 return 0;
393}
394
Christoph Lameter1d071712011-07-14 12:49:12 -0500395static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
396 void *freelist_old, unsigned long counters_old,
397 void *freelist_new, unsigned long counters_new,
398 const char *n)
399{
Heiko Carstens25654092012-01-12 17:17:33 -0800400#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
401 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameter1d071712011-07-14 12:49:12 -0500402 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000403 if (cmpxchg_double(&page->freelist, &page->counters,
Christoph Lameter1d071712011-07-14 12:49:12 -0500404 freelist_old, counters_old,
405 freelist_new, counters_new))
406 return 1;
407 } else
408#endif
409 {
410 unsigned long flags;
411
412 local_irq_save(flags);
413 slab_lock(page);
414 if (page->freelist == freelist_old && page->counters == counters_old) {
415 page->freelist = freelist_new;
416 page->counters = counters_new;
417 slab_unlock(page);
418 local_irq_restore(flags);
419 return 1;
420 }
421 slab_unlock(page);
422 local_irq_restore(flags);
423 }
424
425 cpu_relax();
426 stat(s, CMPXCHG_DOUBLE_FAIL);
427
428#ifdef SLUB_DEBUG_CMPXCHG
429 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
430#endif
431
432 return 0;
433}
434
Christoph Lameter41ecc552007-05-09 02:32:44 -0700435#ifdef CONFIG_SLUB_DEBUG
436/*
Christoph Lameter5f80b132011-04-15 14:48:13 -0500437 * Determine a map of object in use on a page.
438 *
Christoph Lameter881db7f2011-06-01 12:25:53 -0500439 * Node listlock must be held to guarantee that the page does
Christoph Lameter5f80b132011-04-15 14:48:13 -0500440 * not vanish from under us.
441 */
442static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
443{
444 void *p;
445 void *addr = page_address(page);
446
447 for (p = page->freelist; p; p = get_freepointer(s, p))
448 set_bit(slab_index(p, s, addr), map);
449}
450
Christoph Lameter41ecc552007-05-09 02:32:44 -0700451/*
452 * Debug settings:
453 */
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700454#ifdef CONFIG_SLUB_DEBUG_ON
455static int slub_debug = DEBUG_DEFAULT_FLAGS;
456#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700457static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700458#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700459
460static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700461static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700462
Christoph Lameter7656c722007-05-09 02:32:40 -0700463/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700464 * Object debugging
465 */
466static void print_section(char *text, u8 *addr, unsigned int length)
467{
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200468 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
469 length, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -0700470}
471
Christoph Lameter81819f02007-05-06 14:49:36 -0700472static struct track *get_track(struct kmem_cache *s, void *object,
473 enum track_item alloc)
474{
475 struct track *p;
476
477 if (s->offset)
478 p = object + s->offset + sizeof(void *);
479 else
480 p = object + s->inuse;
481
482 return p + alloc;
483}
484
485static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300486 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700487{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900488 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700489
Christoph Lameter81819f02007-05-06 14:49:36 -0700490 if (addr) {
Ben Greeard6543e32011-07-07 11:36:36 -0700491#ifdef CONFIG_STACKTRACE
492 struct stack_trace trace;
493 int i;
494
495 trace.nr_entries = 0;
496 trace.max_entries = TRACK_ADDRS_COUNT;
497 trace.entries = p->addrs;
498 trace.skip = 3;
499 save_stack_trace(&trace);
500
501 /* See rant in lockdep.c */
502 if (trace.nr_entries != 0 &&
503 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
504 trace.nr_entries--;
505
506 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
507 p->addrs[i] = 0;
508#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700509 p->addr = addr;
510 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400511 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700512 p->when = jiffies;
513 } else
514 memset(p, 0, sizeof(struct track));
515}
516
Christoph Lameter81819f02007-05-06 14:49:36 -0700517static void init_tracking(struct kmem_cache *s, void *object)
518{
Christoph Lameter24922682007-07-17 04:03:18 -0700519 if (!(s->flags & SLAB_STORE_USER))
520 return;
521
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300522 set_track(s, object, TRACK_FREE, 0UL);
523 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700524}
525
526static void print_track(const char *s, struct track *t)
527{
528 if (!t->addr)
529 return;
530
Linus Torvalds7daf7052008-07-14 12:12:53 -0700531 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300532 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Ben Greeard6543e32011-07-07 11:36:36 -0700533#ifdef CONFIG_STACKTRACE
534 {
535 int i;
536 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
537 if (t->addrs[i])
538 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
539 else
540 break;
541 }
542#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700543}
544
Christoph Lameter24922682007-07-17 04:03:18 -0700545static void print_tracking(struct kmem_cache *s, void *object)
546{
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
549
550 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
551 print_track("Freed", get_track(s, object, TRACK_FREE));
552}
553
554static void print_page_info(struct page *page)
555{
Christoph Lameter39b26462008-04-14 19:11:30 +0300556 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
557 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700558
559}
560
561static void slab_bug(struct kmem_cache *s, char *fmt, ...)
562{
563 va_list args;
564 char buf[100];
565
566 va_start(args, fmt);
567 vsnprintf(buf, sizeof(buf), fmt, args);
568 va_end(args);
569 printk(KERN_ERR "========================================"
570 "=====================================\n");
Dave Jones265d47e2011-11-15 15:04:00 -0800571 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700572 printk(KERN_ERR "----------------------------------------"
573 "-------------------------------------\n\n");
Dave Jones645df232012-09-18 15:54:12 -0400574
Rusty Russell373d4d02013-01-21 17:17:39 +1030575 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Christoph Lameter24922682007-07-17 04:03:18 -0700576}
577
578static void slab_fix(struct kmem_cache *s, char *fmt, ...)
579{
580 va_list args;
581 char buf[100];
582
583 va_start(args, fmt);
584 vsnprintf(buf, sizeof(buf), fmt, args);
585 va_end(args);
586 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
587}
588
589static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700590{
591 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800592 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700593
594 print_tracking(s, p);
595
596 print_page_info(page);
597
598 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
599 p, p - addr, get_freepointer(s, p));
600
601 if (p > addr + 16)
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200602 print_section("Bytes b4 ", p - 16, 16);
Christoph Lameter24922682007-07-17 04:03:18 -0700603
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500604 print_section("Object ", p, min_t(unsigned long, s->object_size,
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200605 PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700606 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500607 print_section("Redzone ", p + s->object_size,
608 s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700609
Christoph Lameter81819f02007-05-06 14:49:36 -0700610 if (s->offset)
611 off = s->offset + sizeof(void *);
612 else
613 off = s->inuse;
614
Christoph Lameter24922682007-07-17 04:03:18 -0700615 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700616 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700617
618 if (off != s->size)
619 /* Beginning of the filler is the free pointer */
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200620 print_section("Padding ", p + off, s->size - off);
Christoph Lameter24922682007-07-17 04:03:18 -0700621
622 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700623}
624
625static void object_err(struct kmem_cache *s, struct page *page,
626 u8 *object, char *reason)
627{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700628 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700629 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700630}
631
Christoph Lameter945cf2b2012-09-04 23:18:33 +0000632static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700633{
634 va_list args;
635 char buf[100];
636
Christoph Lameter24922682007-07-17 04:03:18 -0700637 va_start(args, fmt);
638 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700639 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700640 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700641 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700642 dump_stack();
643}
644
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500645static void init_object(struct kmem_cache *s, void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700646{
647 u8 *p = object;
648
649 if (s->flags & __OBJECT_POISON) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500650 memset(p, POISON_FREE, s->object_size - 1);
651 p[s->object_size - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700652 }
653
654 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500655 memset(p + s->object_size, val, s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700656}
657
Christoph Lameter24922682007-07-17 04:03:18 -0700658static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
659 void *from, void *to)
660{
661 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
662 memset(from, data, to - from);
663}
664
665static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
666 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800667 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700668{
669 u8 *fault;
670 u8 *end;
671
Akinobu Mita798248202011-10-31 17:08:07 -0700672 fault = memchr_inv(start, value, bytes);
Christoph Lameter24922682007-07-17 04:03:18 -0700673 if (!fault)
674 return 1;
675
676 end = start + bytes;
677 while (end > fault && end[-1] == value)
678 end--;
679
680 slab_bug(s, "%s overwritten", what);
681 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
682 fault, end - 1, fault[0], value);
683 print_trailer(s, page, object);
684
685 restore_bytes(s, what, value, fault, end);
686 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700687}
688
Christoph Lameter81819f02007-05-06 14:49:36 -0700689/*
690 * Object layout:
691 *
692 * object address
693 * Bytes of the object to be managed.
694 * If the freepointer may overlay the object then the free
695 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700696 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700697 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
698 * 0xa5 (POISON_END)
699 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500700 * object + s->object_size
Christoph Lameter81819f02007-05-06 14:49:36 -0700701 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700702 * Padding is extended by another word if Redzoning is enabled and
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500703 * object_size == inuse.
Christoph Lameter672bba32007-05-09 02:32:39 -0700704 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700705 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
706 * 0xcc (RED_ACTIVE) for objects in use.
707 *
708 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700709 * Meta data starts here.
710 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700711 * A. Free pointer (if we cannot overwrite object on free)
712 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700713 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800714 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700715 * before the word boundary.
716 *
717 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700718 *
719 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700720 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700721 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500722 * If slabcaches are merged then the object_size and inuse boundaries are mostly
Christoph Lameter672bba32007-05-09 02:32:39 -0700723 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700724 * may be used with merged slabcaches.
725 */
726
Christoph Lameter81819f02007-05-06 14:49:36 -0700727static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
728{
729 unsigned long off = s->inuse; /* The end of info */
730
731 if (s->offset)
732 /* Freepointer is placed after the object. */
733 off += sizeof(void *);
734
735 if (s->flags & SLAB_STORE_USER)
736 /* We also have user information there */
737 off += 2 * sizeof(struct track);
738
739 if (s->size == off)
740 return 1;
741
Christoph Lameter24922682007-07-17 04:03:18 -0700742 return check_bytes_and_report(s, page, p, "Object padding",
743 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700744}
745
Christoph Lameter39b26462008-04-14 19:11:30 +0300746/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700747static int slab_pad_check(struct kmem_cache *s, struct page *page)
748{
Christoph Lameter24922682007-07-17 04:03:18 -0700749 u8 *start;
750 u8 *fault;
751 u8 *end;
752 int length;
753 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700754
755 if (!(s->flags & SLAB_POISON))
756 return 1;
757
Christoph Lametera973e9d2008-03-01 13:40:44 -0800758 start = page_address(page);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800759 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
Christoph Lameter39b26462008-04-14 19:11:30 +0300760 end = start + length;
761 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700762 if (!remainder)
763 return 1;
764
Akinobu Mita798248202011-10-31 17:08:07 -0700765 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700766 if (!fault)
767 return 1;
768 while (end > fault && end[-1] == POISON_INUSE)
769 end--;
770
771 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200772 print_section("Padding ", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700773
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200774 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700775 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700776}
777
778static int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500779 void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700780{
781 u8 *p = object;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500782 u8 *endobject = object + s->object_size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700783
784 if (s->flags & SLAB_RED_ZONE) {
Christoph Lameter24922682007-07-17 04:03:18 -0700785 if (!check_bytes_and_report(s, page, object, "Redzone",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500786 endobject, val, s->inuse - s->object_size))
Christoph Lameter81819f02007-05-06 14:49:36 -0700787 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700788 } else {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500789 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800790 check_bytes_and_report(s, page, p, "Alignment padding",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500791 endobject, POISON_INUSE, s->inuse - s->object_size);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800792 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700793 }
794
795 if (s->flags & SLAB_POISON) {
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500796 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700797 (!check_bytes_and_report(s, page, p, "Poison", p,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500798 POISON_FREE, s->object_size - 1) ||
Christoph Lameter24922682007-07-17 04:03:18 -0700799 !check_bytes_and_report(s, page, p, "Poison",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500800 p + s->object_size - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700801 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700802 /*
803 * check_pad_bytes cleans up on its own.
804 */
805 check_pad_bytes(s, page, p);
806 }
807
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500808 if (!s->offset && val == SLUB_RED_ACTIVE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700809 /*
810 * Object and freepointer overlap. Cannot check
811 * freepointer while object is allocated.
812 */
813 return 1;
814
815 /* Check free pointer validity */
816 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
817 object_err(s, page, p, "Freepointer corrupt");
818 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100819 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700820 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700821 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700822 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800823 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700824 return 0;
825 }
826 return 1;
827}
828
829static int check_slab(struct kmem_cache *s, struct page *page)
830{
Christoph Lameter39b26462008-04-14 19:11:30 +0300831 int maxobj;
832
Christoph Lameter81819f02007-05-06 14:49:36 -0700833 VM_BUG_ON(!irqs_disabled());
834
835 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700836 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700837 return 0;
838 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300839
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800840 maxobj = order_objects(compound_order(page), s->size, s->reserved);
Christoph Lameter39b26462008-04-14 19:11:30 +0300841 if (page->objects > maxobj) {
842 slab_err(s, page, "objects %u > max %u",
843 s->name, page->objects, maxobj);
844 return 0;
845 }
846 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700847 slab_err(s, page, "inuse %u > max %u",
Christoph Lameter39b26462008-04-14 19:11:30 +0300848 s->name, page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700849 return 0;
850 }
851 /* Slab_pad_check fixes things up after itself */
852 slab_pad_check(s, page);
853 return 1;
854}
855
856/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700857 * Determine if a certain object on a page is on the freelist. Must hold the
858 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700859 */
860static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
861{
862 int nr = 0;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500863 void *fp;
Christoph Lameter81819f02007-05-06 14:49:36 -0700864 void *object = NULL;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300865 unsigned long max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700866
Christoph Lameter881db7f2011-06-01 12:25:53 -0500867 fp = page->freelist;
Christoph Lameter39b26462008-04-14 19:11:30 +0300868 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700869 if (fp == search)
870 return 1;
871 if (!check_valid_pointer(s, page, fp)) {
872 if (object) {
873 object_err(s, page, object,
874 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800875 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700876 break;
877 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700878 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800879 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300880 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700881 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 return 0;
883 }
884 break;
885 }
886 object = fp;
887 fp = get_freepointer(s, object);
888 nr++;
889 }
890
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800891 max_objects = order_objects(compound_order(page), s->size, s->reserved);
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400892 if (max_objects > MAX_OBJS_PER_PAGE)
893 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300894
895 if (page->objects != max_objects) {
896 slab_err(s, page, "Wrong number of objects. Found %d but "
897 "should be %d", page->objects, max_objects);
898 page->objects = max_objects;
899 slab_fix(s, "Number of objects adjusted.");
900 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300901 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700902 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300903 "counted were %d", page->inuse, page->objects - nr);
904 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700905 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700906 }
907 return search == NULL;
908}
909
Christoph Lameter0121c6192008-04-29 16:11:12 -0700910static void trace(struct kmem_cache *s, struct page *page, void *object,
911 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700912{
913 if (s->flags & SLAB_TRACE) {
914 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
915 s->name,
916 alloc ? "alloc" : "free",
917 object, page->inuse,
918 page->freelist);
919
920 if (!alloc)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500921 print_section("Object ", (void *)object, s->object_size);
Christoph Lameter3ec09742007-05-16 22:11:00 -0700922
923 dump_stack();
924 }
925}
926
Christoph Lameter643b1132007-05-06 14:49:42 -0700927/*
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500928 * Hooks for other subsystems that check memory allocations. In a typical
929 * production configuration these hooks all should produce no code at all.
930 */
931static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
932{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500933 flags &= gfp_allowed_mask;
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500934 lockdep_trace_alloc(flags);
935 might_sleep_if(flags & __GFP_WAIT);
936
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500937 return should_failslab(s->object_size, flags, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500938}
939
940static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
941{
Christoph Lameterc1d50832010-08-20 12:37:17 -0500942 flags &= gfp_allowed_mask;
Eric Dumazetb3d41882011-02-14 18:35:22 +0100943 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500944 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500945}
946
947static inline void slab_free_hook(struct kmem_cache *s, void *x)
948{
949 kmemleak_free_recursive(x, s->flags);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500950
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600951 /*
952 * Trouble is that we may no longer disable interupts in the fast path
953 * So in order to make the debug calls that expect irqs to be
954 * disabled we need to disable interrupts temporarily.
955 */
956#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
957 {
958 unsigned long flags;
959
960 local_irq_save(flags);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500961 kmemcheck_slab_free(s, x, s->object_size);
962 debug_check_no_locks_freed(x, s->object_size);
Christoph Lameterd3f661d2011-02-25 11:38:52 -0600963 local_irq_restore(flags);
964 }
965#endif
Thomas Gleixnerf9b615d2011-03-24 21:26:46 +0200966 if (!(s->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500967 debug_check_no_obj_freed(x, s->object_size);
Christoph Lameterc016b0b2010-08-20 12:37:16 -0500968}
969
970/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700971 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500972 *
973 * list_lock must be held.
Christoph Lameter643b1132007-05-06 14:49:42 -0700974 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500975static void add_full(struct kmem_cache *s,
976 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700977{
Christoph Lameter643b1132007-05-06 14:49:42 -0700978 if (!(s->flags & SLAB_STORE_USER))
979 return;
980
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500981 list_add(&page->lru, &n->full);
982}
Christoph Lameter643b1132007-05-06 14:49:42 -0700983
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500984/*
985 * list_lock must be held.
986 */
987static void remove_full(struct kmem_cache *s, struct page *page)
988{
989 if (!(s->flags & SLAB_STORE_USER))
990 return;
991
Christoph Lameter643b1132007-05-06 14:49:42 -0700992 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -0700993}
994
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300995/* Tracking of the number of slabs for debugging purposes */
996static inline unsigned long slabs_node(struct kmem_cache *s, int node)
997{
998 struct kmem_cache_node *n = get_node(s, node);
999
1000 return atomic_long_read(&n->nr_slabs);
1001}
1002
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001003static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1004{
1005 return atomic_long_read(&n->nr_slabs);
1006}
1007
Christoph Lameter205ab992008-04-14 19:11:40 +03001008static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001009{
1010 struct kmem_cache_node *n = get_node(s, node);
1011
1012 /*
1013 * May be called early in order to allocate a slab for the
1014 * kmem_cache_node structure. Solve the chicken-egg
1015 * dilemma by deferring the increment of the count during
1016 * bootstrap (see early_kmem_cache_node_alloc).
1017 */
Joonsoo Kim338b2642013-01-21 17:01:27 +09001018 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001019 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001020 atomic_long_add(objects, &n->total_objects);
1021 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001022}
Christoph Lameter205ab992008-04-14 19:11:40 +03001023static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001024{
1025 struct kmem_cache_node *n = get_node(s, node);
1026
1027 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001028 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001029}
1030
1031/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001032static void setup_object_debug(struct kmem_cache *s, struct page *page,
1033 void *object)
1034{
1035 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1036 return;
1037
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001038 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001039 init_tracking(s, object);
1040}
1041
Christoph Lameter15370662010-08-20 12:37:12 -05001042static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001043 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001044{
1045 if (!check_slab(s, page))
1046 goto bad;
1047
Christoph Lameter81819f02007-05-06 14:49:36 -07001048 if (!check_valid_pointer(s, page, object)) {
1049 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -07001050 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001051 }
1052
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001053 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Christoph Lameter81819f02007-05-06 14:49:36 -07001054 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -07001055
Christoph Lameter3ec09742007-05-16 22:11:00 -07001056 /* Success perform special debug activities for allocs */
1057 if (s->flags & SLAB_STORE_USER)
1058 set_track(s, object, TRACK_ALLOC, addr);
1059 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001060 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001061 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001062
Christoph Lameter81819f02007-05-06 14:49:36 -07001063bad:
1064 if (PageSlab(page)) {
1065 /*
1066 * If this is a slab page then lets do the best we can
1067 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001068 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001069 */
Christoph Lameter24922682007-07-17 04:03:18 -07001070 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001071 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001072 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001073 }
1074 return 0;
1075}
1076
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001077static noinline struct kmem_cache_node *free_debug_processing(
1078 struct kmem_cache *s, struct page *page, void *object,
1079 unsigned long addr, unsigned long *flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001080{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001081 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001082
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001083 spin_lock_irqsave(&n->list_lock, *flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001084 slab_lock(page);
1085
Christoph Lameter81819f02007-05-06 14:49:36 -07001086 if (!check_slab(s, page))
1087 goto fail;
1088
1089 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001090 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001091 goto fail;
1092 }
1093
1094 if (on_freelist(s, page, object)) {
Christoph Lameter24922682007-07-17 04:03:18 -07001095 object_err(s, page, object, "Object already free");
Christoph Lameter81819f02007-05-06 14:49:36 -07001096 goto fail;
1097 }
1098
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001099 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001100 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001101
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001102 if (unlikely(s != page->slab_cache)) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08001103 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -07001104 slab_err(s, page, "Attempt to free object(0x%p) "
1105 "outside of slab", object);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001106 } else if (!page->slab_cache) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001107 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -07001108 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -07001109 object);
Christoph Lameter70d71222007-05-06 14:49:47 -07001110 dump_stack();
Pekka Enberg06428782008-01-07 23:20:27 -08001111 } else
Christoph Lameter24922682007-07-17 04:03:18 -07001112 object_err(s, page, object,
1113 "page slab pointer corrupt.");
Christoph Lameter81819f02007-05-06 14:49:36 -07001114 goto fail;
1115 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001116
Christoph Lameter3ec09742007-05-16 22:11:00 -07001117 if (s->flags & SLAB_STORE_USER)
1118 set_track(s, object, TRACK_FREE, addr);
1119 trace(s, page, object, 0);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001120 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001121out:
Christoph Lameter881db7f2011-06-01 12:25:53 -05001122 slab_unlock(page);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001123 /*
1124 * Keep node_lock to preserve integrity
1125 * until the object is actually freed
1126 */
1127 return n;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001128
Christoph Lameter81819f02007-05-06 14:49:36 -07001129fail:
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001130 slab_unlock(page);
1131 spin_unlock_irqrestore(&n->list_lock, *flags);
Christoph Lameter24922682007-07-17 04:03:18 -07001132 slab_fix(s, "Object at 0x%p not freed", object);
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001133 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001134}
1135
Christoph Lameter41ecc552007-05-09 02:32:44 -07001136static int __init setup_slub_debug(char *str)
1137{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001138 slub_debug = DEBUG_DEFAULT_FLAGS;
1139 if (*str++ != '=' || !*str)
1140 /*
1141 * No options specified. Switch on full debugging.
1142 */
1143 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001144
1145 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001146 /*
1147 * No options but restriction on slabs. This means full
1148 * debugging for slabs matching a pattern.
1149 */
1150 goto check_slabs;
1151
David Rientjesfa5ec8a2009-07-07 00:14:14 -07001152 if (tolower(*str) == 'o') {
1153 /*
1154 * Avoid enabling debugging on caches if its minimum order
1155 * would increase as a result.
1156 */
1157 disable_higher_order_debug = 1;
1158 goto out;
1159 }
1160
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001161 slub_debug = 0;
1162 if (*str == '-')
1163 /*
1164 * Switch off all debugging measures.
1165 */
1166 goto out;
1167
1168 /*
1169 * Determine which debug features should be switched on
1170 */
Pekka Enberg06428782008-01-07 23:20:27 -08001171 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001172 switch (tolower(*str)) {
1173 case 'f':
1174 slub_debug |= SLAB_DEBUG_FREE;
1175 break;
1176 case 'z':
1177 slub_debug |= SLAB_RED_ZONE;
1178 break;
1179 case 'p':
1180 slub_debug |= SLAB_POISON;
1181 break;
1182 case 'u':
1183 slub_debug |= SLAB_STORE_USER;
1184 break;
1185 case 't':
1186 slub_debug |= SLAB_TRACE;
1187 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001188 case 'a':
1189 slub_debug |= SLAB_FAILSLAB;
1190 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001191 default:
1192 printk(KERN_ERR "slub_debug option '%c' "
Pekka Enberg06428782008-01-07 23:20:27 -08001193 "unknown. skipped\n", *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001194 }
1195 }
1196
1197check_slabs:
1198 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001199 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001200out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001201 return 1;
1202}
1203
1204__setup("slub_debug", setup_slub_debug);
1205
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001206static unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001207 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001208 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001209{
1210 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001211 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001212 */
Christoph Lametere1533622008-02-15 23:45:24 -08001213 if (slub_debug && (!slub_debug_slabs ||
David Rientjes3de47212009-07-27 18:30:35 -07001214 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1215 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001216
1217 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001218}
1219#else
Christoph Lameter3ec09742007-05-16 22:11:00 -07001220static inline void setup_object_debug(struct kmem_cache *s,
1221 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001222
Christoph Lameter3ec09742007-05-16 22:11:00 -07001223static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001224 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001225
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001226static inline struct kmem_cache_node *free_debug_processing(
1227 struct kmem_cache *s, struct page *page, void *object,
1228 unsigned long addr, unsigned long *flags) { return NULL; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001229
Christoph Lameter41ecc552007-05-09 02:32:44 -07001230static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1231 { return 1; }
1232static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001233 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001234static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1235 struct page *page) {}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001236static inline void remove_full(struct kmem_cache *s, struct page *page) {}
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001237static inline unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001238 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001239 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001240{
1241 return flags;
1242}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001243#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001244
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001245#define disable_higher_order_debug 0
1246
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001247static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1248 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001249static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1250 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001251static inline void inc_slabs_node(struct kmem_cache *s, int node,
1252 int objects) {}
1253static inline void dec_slabs_node(struct kmem_cache *s, int node,
1254 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001255
1256static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1257 { return 0; }
1258
1259static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1260 void *object) {}
1261
1262static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1263
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05001264#endif /* CONFIG_SLUB_DEBUG */
Christoph Lameter205ab992008-04-14 19:11:40 +03001265
Christoph Lameter81819f02007-05-06 14:49:36 -07001266/*
1267 * Slab allocation and freeing
1268 */
Christoph Lameter65c33762008-04-14 19:11:40 +03001269static inline struct page *alloc_slab_page(gfp_t flags, int node,
1270 struct kmem_cache_order_objects oo)
1271{
1272 int order = oo_order(oo);
1273
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001274 flags |= __GFP_NOTRACK;
1275
Christoph Lameter2154a332010-07-09 14:07:10 -05001276 if (node == NUMA_NO_NODE)
Christoph Lameter65c33762008-04-14 19:11:40 +03001277 return alloc_pages(flags, order);
1278 else
Minchan Kim6b65aaf2010-04-14 23:58:36 +09001279 return alloc_pages_exact_node(node, flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001280}
1281
Christoph Lameter81819f02007-05-06 14:49:36 -07001282static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1283{
Pekka Enberg06428782008-01-07 23:20:27 -08001284 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001285 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001286 gfp_t alloc_gfp;
Christoph Lameter81819f02007-05-06 14:49:36 -07001287
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001288 flags &= gfp_allowed_mask;
1289
1290 if (flags & __GFP_WAIT)
1291 local_irq_enable();
1292
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001293 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001294
Pekka Enbergba522702009-06-24 21:59:51 +03001295 /*
1296 * Let the initial higher-order allocation fail under memory pressure
1297 * so we fall-back to the minimum order allocation.
1298 */
1299 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1300
1301 page = alloc_slab_page(alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001302 if (unlikely(!page)) {
1303 oo = s->min;
1304 /*
1305 * Allocation may have failed due to fragmentation.
1306 * Try a lower order alloc if possible
1307 */
1308 page = alloc_slab_page(flags, node, oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001309
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001310 if (page)
1311 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001312 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001313
David Rientjes737b7192012-07-09 14:00:38 -07001314 if (kmemcheck_enabled && page
Amerigo Wang5086c389c2009-08-19 21:44:13 +03001315 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001316 int pages = 1 << oo_order(oo);
1317
1318 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1319
1320 /*
1321 * Objects from caches that have a constructor don't get
1322 * cleared when they're allocated, so we need to do it here.
1323 */
1324 if (s->ctor)
1325 kmemcheck_mark_uninitialized_pages(page, pages);
1326 else
1327 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001328 }
1329
David Rientjes737b7192012-07-09 14:00:38 -07001330 if (flags & __GFP_WAIT)
1331 local_irq_disable();
1332 if (!page)
1333 return NULL;
1334
Christoph Lameter834f3d12008-04-14 19:11:31 +03001335 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001336 mod_zone_page_state(page_zone(page),
1337 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1338 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Christoph Lameter65c33762008-04-14 19:11:40 +03001339 1 << oo_order(oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07001340
1341 return page;
1342}
1343
1344static void setup_object(struct kmem_cache *s, struct page *page,
1345 void *object)
1346{
Christoph Lameter3ec09742007-05-16 22:11:00 -07001347 setup_object_debug(s, page, object);
Christoph Lameter4f104932007-05-06 14:50:17 -07001348 if (unlikely(s->ctor))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001349 s->ctor(object);
Christoph Lameter81819f02007-05-06 14:49:36 -07001350}
1351
1352static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1353{
1354 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001355 void *start;
Christoph Lameter81819f02007-05-06 14:49:36 -07001356 void *last;
1357 void *p;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001358 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001359
Christoph Lameter6cb06222007-10-16 01:25:41 -07001360 BUG_ON(flags & GFP_SLAB_BUG_MASK);
Christoph Lameter81819f02007-05-06 14:49:36 -07001361
Christoph Lameter6cb06222007-10-16 01:25:41 -07001362 page = allocate_slab(s,
1363 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001364 if (!page)
1365 goto out;
1366
Glauber Costa1f458cb2012-12-18 14:22:50 -08001367 order = compound_order(page);
Christoph Lameter205ab992008-04-14 19:11:40 +03001368 inc_slabs_node(s, page_to_nid(page), page->objects);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001369 memcg_bind_pages(s, order);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001370 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001371 __SetPageSlab(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001372 if (page->pfmemalloc)
1373 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001374
1375 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001376
1377 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001378 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001379
1380 last = start;
Christoph Lameter224a88b2008-04-14 19:11:31 +03001381 for_each_object(p, s, start, page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001382 setup_object(s, page, last);
1383 set_freepointer(s, last, p);
1384 last = p;
1385 }
1386 setup_object(s, page, last);
Christoph Lametera973e9d2008-03-01 13:40:44 -08001387 set_freepointer(s, last, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001388
1389 page->freelist = start;
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001390 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001391 page->frozen = 1;
Christoph Lameter81819f02007-05-06 14:49:36 -07001392out:
Christoph Lameter81819f02007-05-06 14:49:36 -07001393 return page;
1394}
1395
1396static void __free_slab(struct kmem_cache *s, struct page *page)
1397{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001398 int order = compound_order(page);
1399 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001400
Christoph Lameteraf537b02010-07-09 14:07:14 -05001401 if (kmem_cache_debug(s)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001402 void *p;
1403
1404 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001405 for_each_object(p, s, page_address(page),
1406 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001407 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001408 }
1409
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001410 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001411
Christoph Lameter81819f02007-05-06 14:49:36 -07001412 mod_zone_page_state(page_zone(page),
1413 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1414 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001415 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001416
Mel Gorman072bb0a2012-07-31 16:43:58 -07001417 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001418 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001419
1420 memcg_release_pages(s, order);
Mel Gorman22b751c2013-02-22 16:34:59 -08001421 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001422 if (current->reclaim_state)
1423 current->reclaim_state->reclaimed_slab += pages;
Glauber Costad79923f2012-12-18 14:22:48 -08001424 __free_memcg_kmem_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001425}
1426
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001427#define need_reserve_slab_rcu \
1428 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1429
Christoph Lameter81819f02007-05-06 14:49:36 -07001430static void rcu_free_slab(struct rcu_head *h)
1431{
1432 struct page *page;
1433
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001434 if (need_reserve_slab_rcu)
1435 page = virt_to_head_page(h);
1436 else
1437 page = container_of((struct list_head *)h, struct page, lru);
1438
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001439 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001440}
1441
1442static void free_slab(struct kmem_cache *s, struct page *page)
1443{
1444 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001445 struct rcu_head *head;
1446
1447 if (need_reserve_slab_rcu) {
1448 int order = compound_order(page);
1449 int offset = (PAGE_SIZE << order) - s->reserved;
1450
1451 VM_BUG_ON(s->reserved != sizeof(*head));
1452 head = page_address(page) + offset;
1453 } else {
1454 /*
1455 * RCU free overloads the RCU head over the LRU
1456 */
1457 head = (void *)&page->lru;
1458 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001459
1460 call_rcu(head, rcu_free_slab);
1461 } else
1462 __free_slab(s, page);
1463}
1464
1465static void discard_slab(struct kmem_cache *s, struct page *page)
1466{
Christoph Lameter205ab992008-04-14 19:11:40 +03001467 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001468 free_slab(s, page);
1469}
1470
1471/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001472 * Management of partially allocated slabs.
1473 *
1474 * list_lock must be held.
Christoph Lameter81819f02007-05-06 14:49:36 -07001475 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001476static inline void add_partial(struct kmem_cache_node *n,
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001477 struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001478{
Christoph Lametere95eed52007-05-06 14:49:44 -07001479 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001480 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001481 list_add_tail(&page->lru, &n->partial);
1482 else
1483 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001484}
1485
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001486/*
1487 * list_lock must be held.
1488 */
1489static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001490 struct page *page)
1491{
1492 list_del(&page->lru);
1493 n->nr_partial--;
1494}
1495
Christoph Lameter81819f02007-05-06 14:49:36 -07001496/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001497 * Remove slab from the partial list, freeze it and
1498 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001499 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001500 * Returns a list of objects or NULL if it fails.
1501 *
Christoph Lameter7ced3712012-05-09 10:09:53 -05001502 * Must hold list_lock since we modify the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001503 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001504static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001505 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001506 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001507{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001508 void *freelist;
1509 unsigned long counters;
1510 struct page new;
1511
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001512 /*
1513 * Zap the freelist and set the frozen bit.
1514 * The old freelist is the list of objects for the
1515 * per cpu allocation list.
1516 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001517 freelist = page->freelist;
1518 counters = page->counters;
1519 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001520 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001521 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001522 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001523 new.freelist = NULL;
1524 } else {
1525 new.freelist = freelist;
1526 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001527
Christoph Lameter7ced3712012-05-09 10:09:53 -05001528 VM_BUG_ON(new.frozen);
1529 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001530
Christoph Lameter7ced3712012-05-09 10:09:53 -05001531 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001532 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001533 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001534 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001535 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001536
1537 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001538 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001539 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001540}
1541
Joonsoo Kim633b0762013-01-21 17:01:25 +09001542static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001543static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001544
Christoph Lameter81819f02007-05-06 14:49:36 -07001545/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001546 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001547 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001548static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1549 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001550{
Christoph Lameter49e22582011-08-09 16:12:27 -05001551 struct page *page, *page2;
1552 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001553 int available = 0;
1554 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001555
1556 /*
1557 * Racy check. If we mistakenly see no partial slabs then we
1558 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001559 * partial slab and there is none available then get_partials()
1560 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001561 */
1562 if (!n || !n->nr_partial)
1563 return NULL;
1564
1565 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001566 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001567 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001568
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001569 if (!pfmemalloc_match(page, flags))
1570 continue;
1571
Joonsoo Kim633b0762013-01-21 17:01:25 +09001572 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001573 if (!t)
1574 break;
1575
Joonsoo Kim633b0762013-01-21 17:01:25 +09001576 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001577 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001578 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001579 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001580 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001581 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001582 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001583 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001584 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001585 if (!kmem_cache_has_cpu_partial(s)
1586 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001587 break;
1588
Christoph Lameter497b66f2011-08-09 16:12:26 -05001589 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001590 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001591 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001592}
1593
1594/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001595 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001596 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001597static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001598 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001599{
1600#ifdef CONFIG_NUMA
1601 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001602 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001603 struct zone *zone;
1604 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001605 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001606 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001607
1608 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001609 * The defrag ratio allows a configuration of the tradeoffs between
1610 * inter node defragmentation and node local allocations. A lower
1611 * defrag_ratio increases the tendency to do local allocations
1612 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001613 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001614 * If the defrag_ratio is set to 0 then kmalloc() always
1615 * returns node local objects. If the ratio is higher then kmalloc()
1616 * may return off node objects because partial slabs are obtained
1617 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001618 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001619 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001620 * defrag_ratio = 1000) then every (well almost) allocation will
1621 * first attempt to defrag slab caches on other nodes. This means
1622 * scanning over all nodes to look for partial slabs which may be
1623 * expensive if we do it every time we are trying to find a slab
1624 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001625 */
Christoph Lameter98246012008-01-07 23:20:26 -08001626 if (!s->remote_node_defrag_ratio ||
1627 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001628 return NULL;
1629
Mel Gormancc9a6c82012-03-21 16:34:11 -07001630 do {
1631 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07001632 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001633 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1634 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001635
Mel Gormancc9a6c82012-03-21 16:34:11 -07001636 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001637
Mel Gormancc9a6c82012-03-21 16:34:11 -07001638 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1639 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001640 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001641 if (object) {
1642 /*
1643 * Return the object even if
1644 * put_mems_allowed indicated that
1645 * the cpuset mems_allowed was
1646 * updated in parallel. It's a
1647 * harmless race between the alloc
1648 * and the cpuset update.
1649 */
1650 put_mems_allowed(cpuset_mems_cookie);
1651 return object;
1652 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001653 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001654 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001655 } while (!put_mems_allowed(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001656#endif
1657 return NULL;
1658}
1659
1660/*
1661 * Get a partial page, lock it and return it.
1662 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001663static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001664 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001665{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001666 void *object;
Christoph Lameter2154a332010-07-09 14:07:10 -05001667 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
Christoph Lameter81819f02007-05-06 14:49:36 -07001668
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001669 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001670 if (object || node != NUMA_NO_NODE)
1671 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001672
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001673 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001674}
1675
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001676#ifdef CONFIG_PREEMPT
1677/*
1678 * Calculate the next globally unique transaction for disambiguiation
1679 * during cmpxchg. The transactions start with the cpu number and are then
1680 * incremented by CONFIG_NR_CPUS.
1681 */
1682#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1683#else
1684/*
1685 * No preemption supported therefore also no need to check for
1686 * different cpus.
1687 */
1688#define TID_STEP 1
1689#endif
1690
1691static inline unsigned long next_tid(unsigned long tid)
1692{
1693 return tid + TID_STEP;
1694}
1695
1696static inline unsigned int tid_to_cpu(unsigned long tid)
1697{
1698 return tid % TID_STEP;
1699}
1700
1701static inline unsigned long tid_to_event(unsigned long tid)
1702{
1703 return tid / TID_STEP;
1704}
1705
1706static inline unsigned int init_tid(int cpu)
1707{
1708 return cpu;
1709}
1710
1711static inline void note_cmpxchg_failure(const char *n,
1712 const struct kmem_cache *s, unsigned long tid)
1713{
1714#ifdef SLUB_DEBUG_CMPXCHG
1715 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1716
1717 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1718
1719#ifdef CONFIG_PREEMPT
1720 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1721 printk("due to cpu change %d -> %d\n",
1722 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1723 else
1724#endif
1725 if (tid_to_event(tid) != tid_to_event(actual_tid))
1726 printk("due to cpu running other code. Event %ld->%ld\n",
1727 tid_to_event(tid), tid_to_event(actual_tid));
1728 else
1729 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1730 actual_tid, tid, next_tid(tid));
1731#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001732 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001733}
1734
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001735static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001736{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001737 int cpu;
1738
1739 for_each_possible_cpu(cpu)
1740 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001741}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001742
1743/*
1744 * Remove the cpu slab
1745 */
Christoph Lameterc17dda42012-05-09 10:09:57 -05001746static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001747{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001748 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001749 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1750 int lock = 0;
1751 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001752 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001753 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001754 struct page new;
1755 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001756
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001757 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001758 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001759 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001760 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001761
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001762 /*
1763 * Stage one: Free all available per cpu objects back
1764 * to the page freelist while it is still frozen. Leave the
1765 * last one.
1766 *
1767 * There is no need to take the list->lock because the page
1768 * is still frozen.
1769 */
1770 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1771 void *prior;
1772 unsigned long counters;
1773
1774 do {
1775 prior = page->freelist;
1776 counters = page->counters;
1777 set_freepointer(s, freelist, prior);
1778 new.counters = counters;
1779 new.inuse--;
1780 VM_BUG_ON(!new.frozen);
1781
Christoph Lameter1d071712011-07-14 12:49:12 -05001782 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001783 prior, counters,
1784 freelist, new.counters,
1785 "drain percpu freelist"));
1786
1787 freelist = nextfree;
1788 }
1789
1790 /*
1791 * Stage two: Ensure that the page is unfrozen while the
1792 * list presence reflects the actual number of objects
1793 * during unfreeze.
1794 *
1795 * We setup the list membership and then perform a cmpxchg
1796 * with the count. If there is a mismatch then the page
1797 * is not unfrozen but the page is on the wrong list.
1798 *
1799 * Then we restart the process which may have to remove
1800 * the page from the list that we just put it on again
1801 * because the number of objects in the slab may have
1802 * changed.
1803 */
1804redo:
1805
1806 old.freelist = page->freelist;
1807 old.counters = page->counters;
1808 VM_BUG_ON(!old.frozen);
1809
1810 /* Determine target state of the slab */
1811 new.counters = old.counters;
1812 if (freelist) {
1813 new.inuse--;
1814 set_freepointer(s, freelist, old.freelist);
1815 new.freelist = freelist;
1816 } else
1817 new.freelist = old.freelist;
1818
1819 new.frozen = 0;
1820
Christoph Lameter81107182011-08-09 13:01:32 -05001821 if (!new.inuse && n->nr_partial > s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001822 m = M_FREE;
1823 else if (new.freelist) {
1824 m = M_PARTIAL;
1825 if (!lock) {
1826 lock = 1;
1827 /*
1828 * Taking the spinlock removes the possiblity
1829 * that acquire_slab() will see a slab page that
1830 * is frozen
1831 */
1832 spin_lock(&n->list_lock);
1833 }
1834 } else {
1835 m = M_FULL;
1836 if (kmem_cache_debug(s) && !lock) {
1837 lock = 1;
1838 /*
1839 * This also ensures that the scanning of full
1840 * slabs from diagnostic functions will not see
1841 * any frozen slabs.
1842 */
1843 spin_lock(&n->list_lock);
1844 }
1845 }
1846
1847 if (l != m) {
1848
1849 if (l == M_PARTIAL)
1850
1851 remove_partial(n, page);
1852
1853 else if (l == M_FULL)
1854
1855 remove_full(s, page);
1856
1857 if (m == M_PARTIAL) {
1858
1859 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001860 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001861
1862 } else if (m == M_FULL) {
1863
1864 stat(s, DEACTIVATE_FULL);
1865 add_full(s, n, page);
1866
1867 }
1868 }
1869
1870 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001871 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001872 old.freelist, old.counters,
1873 new.freelist, new.counters,
1874 "unfreezing slab"))
1875 goto redo;
1876
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001877 if (lock)
1878 spin_unlock(&n->list_lock);
1879
1880 if (m == M_FREE) {
1881 stat(s, DEACTIVATE_EMPTY);
1882 discard_slab(s, page);
1883 stat(s, FREE_SLAB);
1884 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001885}
1886
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001887/*
1888 * Unfreeze all the cpu partial slabs.
1889 *
Christoph Lameter59a09912012-11-28 16:23:00 +00001890 * This function must be called with interrupts disabled
1891 * for the cpu using c (or some other guarantee must be there
1892 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001893 */
Christoph Lameter59a09912012-11-28 16:23:00 +00001894static void unfreeze_partials(struct kmem_cache *s,
1895 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05001896{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001897#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09001898 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08001899 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001900
1901 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001902 struct page new;
1903 struct page old;
1904
1905 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001906
1907 n2 = get_node(s, page_to_nid(page));
1908 if (n != n2) {
1909 if (n)
1910 spin_unlock(&n->list_lock);
1911
1912 n = n2;
1913 spin_lock(&n->list_lock);
1914 }
Christoph Lameter49e22582011-08-09 16:12:27 -05001915
1916 do {
1917
1918 old.freelist = page->freelist;
1919 old.counters = page->counters;
1920 VM_BUG_ON(!old.frozen);
1921
1922 new.counters = old.counters;
1923 new.freelist = old.freelist;
1924
1925 new.frozen = 0;
1926
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001927 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05001928 old.freelist, old.counters,
1929 new.freelist, new.counters,
1930 "unfreezing slab"));
1931
Joonsoo Kim43d77862012-06-09 02:23:16 +09001932 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08001933 page->next = discard_page;
1934 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001935 } else {
1936 add_partial(n, page, DEACTIVATE_TO_TAIL);
1937 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001938 }
1939 }
1940
1941 if (n)
1942 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08001943
1944 while (discard_page) {
1945 page = discard_page;
1946 discard_page = discard_page->next;
1947
1948 stat(s, DEACTIVATE_EMPTY);
1949 discard_slab(s, page);
1950 stat(s, FREE_SLAB);
1951 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001952#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05001953}
1954
1955/*
1956 * Put a page that was just frozen (in __slab_free) into a partial page
1957 * slot if available. This is done without interrupts disabled and without
1958 * preemption disabled. The cmpxchg is racy and may put the partial page
1959 * onto a random cpus partial slot.
1960 *
1961 * If we did not find a slot then simply move all the partials to the
1962 * per node partial list.
1963 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09001964static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05001965{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001966#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05001967 struct page *oldpage;
1968 int pages;
1969 int pobjects;
1970
1971 do {
1972 pages = 0;
1973 pobjects = 0;
1974 oldpage = this_cpu_read(s->cpu_slab->partial);
1975
1976 if (oldpage) {
1977 pobjects = oldpage->pobjects;
1978 pages = oldpage->pages;
1979 if (drain && pobjects > s->cpu_partial) {
1980 unsigned long flags;
1981 /*
1982 * partial array is full. Move the existing
1983 * set to the per node partial list.
1984 */
1985 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00001986 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05001987 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09001988 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001989 pobjects = 0;
1990 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08001991 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05001992 }
1993 }
1994
1995 pages++;
1996 pobjects += page->objects - page->inuse;
1997
1998 page->pages = pages;
1999 page->pobjects = pobjects;
2000 page->next = oldpage;
2001
Christoph Lameter933393f2011-12-22 11:58:51 -06002002 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
Joonsoo Kim345c9052013-06-19 14:05:52 +09002003#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002004}
2005
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002006static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002007{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002008 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002009 deactivate_slab(s, c->page, c->freelist);
2010
2011 c->tid = next_tid(c->tid);
2012 c->page = NULL;
2013 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002014}
2015
2016/*
2017 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002018 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002019 * Called from IPI handler with interrupts disabled.
2020 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002021static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002022{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002023 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002024
Christoph Lameter49e22582011-08-09 16:12:27 -05002025 if (likely(c)) {
2026 if (c->page)
2027 flush_slab(s, c);
2028
Christoph Lameter59a09912012-11-28 16:23:00 +00002029 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002030 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002031}
2032
2033static void flush_cpu_slab(void *d)
2034{
2035 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002036
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002037 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002038}
2039
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002040static bool has_cpu_slab(int cpu, void *info)
2041{
2042 struct kmem_cache *s = info;
2043 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2044
majianpeng02e1a9c2012-05-17 17:03:26 -07002045 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002046}
2047
Christoph Lameter81819f02007-05-06 14:49:36 -07002048static void flush_all(struct kmem_cache *s)
2049{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002050 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002051}
2052
2053/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002054 * Check if the objects in a per cpu structure fit numa
2055 * locality expectations.
2056 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002057static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002058{
2059#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002060 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002061 return 0;
2062#endif
2063 return 1;
2064}
2065
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002066static int count_free(struct page *page)
2067{
2068 return page->objects - page->inuse;
2069}
2070
2071static unsigned long count_partial(struct kmem_cache_node *n,
2072 int (*get_count)(struct page *))
2073{
2074 unsigned long flags;
2075 unsigned long x = 0;
2076 struct page *page;
2077
2078 spin_lock_irqsave(&n->list_lock, flags);
2079 list_for_each_entry(page, &n->partial, lru)
2080 x += get_count(page);
2081 spin_unlock_irqrestore(&n->list_lock, flags);
2082 return x;
2083}
2084
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002085static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2086{
2087#ifdef CONFIG_SLUB_DEBUG
2088 return atomic_long_read(&n->total_objects);
2089#else
2090 return 0;
2091#endif
2092}
2093
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002094static noinline void
2095slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2096{
2097 int node;
2098
2099 printk(KERN_WARNING
2100 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2101 nid, gfpflags);
2102 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002103 "default order: %d, min order: %d\n", s->name, s->object_size,
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002104 s->size, oo_order(s->oo), oo_order(s->min));
2105
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002106 if (oo_order(s->min) > get_order(s->object_size))
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002107 printk(KERN_WARNING " %s debugging increased min order, use "
2108 "slub_debug=O to disable.\n", s->name);
2109
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002110 for_each_online_node(node) {
2111 struct kmem_cache_node *n = get_node(s, node);
2112 unsigned long nr_slabs;
2113 unsigned long nr_objs;
2114 unsigned long nr_free;
2115
2116 if (!n)
2117 continue;
2118
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002119 nr_free = count_partial(n, count_free);
2120 nr_slabs = node_nr_slabs(n);
2121 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002122
2123 printk(KERN_WARNING
2124 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2125 node, nr_slabs, nr_objs, nr_free);
2126 }
2127}
2128
Christoph Lameter497b66f2011-08-09 16:12:26 -05002129static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2130 int node, struct kmem_cache_cpu **pc)
2131{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002132 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002133 struct kmem_cache_cpu *c = *pc;
2134 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002135
Christoph Lameter188fd062012-05-09 10:09:55 -05002136 freelist = get_partial(s, flags, node, c);
2137
2138 if (freelist)
2139 return freelist;
2140
2141 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002142 if (page) {
2143 c = __this_cpu_ptr(s->cpu_slab);
2144 if (c->page)
2145 flush_slab(s, c);
2146
2147 /*
2148 * No other reference to the page yet so we can
2149 * muck around with it freely without cmpxchg
2150 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002151 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002152 page->freelist = NULL;
2153
2154 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002155 c->page = page;
2156 *pc = c;
2157 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002158 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002159
Christoph Lameter6faa6832012-05-09 10:09:51 -05002160 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002161}
2162
Mel Gorman072bb0a2012-07-31 16:43:58 -07002163static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2164{
2165 if (unlikely(PageSlabPfmemalloc(page)))
2166 return gfp_pfmemalloc_allowed(gfpflags);
2167
2168 return true;
2169}
2170
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002171/*
Christoph Lameter213eeb92011-11-11 14:07:14 -06002172 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2173 * or deactivate the page.
2174 *
2175 * The page is still frozen if the return value is not NULL.
2176 *
2177 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002178 *
2179 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002180 */
2181static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2182{
2183 struct page new;
2184 unsigned long counters;
2185 void *freelist;
2186
2187 do {
2188 freelist = page->freelist;
2189 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002190
Christoph Lameter213eeb92011-11-11 14:07:14 -06002191 new.counters = counters;
2192 VM_BUG_ON(!new.frozen);
2193
2194 new.inuse = page->objects;
2195 new.frozen = freelist != NULL;
2196
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002197 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002198 freelist, counters,
2199 NULL, new.counters,
2200 "get_freelist"));
2201
2202 return freelist;
2203}
2204
2205/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002206 * Slow path. The lockless freelist is empty or we need to perform
2207 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002208 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002209 * Processing is still very fast if new objects have been freed to the
2210 * regular freelist. In that case we simply take over the regular freelist
2211 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002212 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002213 * If that is not working then we fall back to the partial lists. We take the
2214 * first element of the freelist as the object to allocate now and move the
2215 * rest of the freelist to the lockless freelist.
2216 *
2217 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002218 * we need to allocate a new slab. This is the slowest path since it involves
2219 * a call to the page allocator and the setup of a new slab.
Christoph Lameter81819f02007-05-06 14:49:36 -07002220 */
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002221static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2222 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002223{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002224 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002225 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002226 unsigned long flags;
2227
2228 local_irq_save(flags);
2229#ifdef CONFIG_PREEMPT
2230 /*
2231 * We may have been preempted and rescheduled on a different
2232 * cpu before disabling interrupts. Need to reload cpu area
2233 * pointer.
2234 */
2235 c = this_cpu_ptr(s->cpu_slab);
2236#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002237
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002238 page = c->page;
2239 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002240 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002241redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002242
Christoph Lameter57d437d2012-05-09 10:09:59 -05002243 if (unlikely(!node_match(page, node))) {
Christoph Lametere36a2652011-06-01 12:25:57 -05002244 stat(s, ALLOC_NODE_MISMATCH);
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002245 deactivate_slab(s, page, c->freelist);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002246 c->page = NULL;
2247 c->freelist = NULL;
Christoph Lameterfc59c052011-06-01 12:25:56 -05002248 goto new_slab;
2249 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002250
Mel Gorman072bb0a2012-07-31 16:43:58 -07002251 /*
2252 * By rights, we should be searching for a slab page that was
2253 * PFMEMALLOC but right now, we are losing the pfmemalloc
2254 * information when the page leaves the per-cpu allocator
2255 */
2256 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2257 deactivate_slab(s, page, c->freelist);
2258 c->page = NULL;
2259 c->freelist = NULL;
2260 goto new_slab;
2261 }
2262
Eric Dumazet73736e02011-12-13 04:57:06 +01002263 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002264 freelist = c->freelist;
2265 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002266 goto load_freelist;
2267
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002268 stat(s, ALLOC_SLOWPATH);
2269
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002270 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002271
Christoph Lameter6faa6832012-05-09 10:09:51 -05002272 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002273 c->page = NULL;
2274 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002275 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002276 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002277
Christoph Lameter81819f02007-05-06 14:49:36 -07002278 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002279
Christoph Lameter894b8782007-05-10 03:15:16 -07002280load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002281 /*
2282 * freelist is pointing to the list of objects to be used.
2283 * page is pointing to the page from which the objects are obtained.
2284 * That page must be frozen for per cpu allocations to work.
2285 */
2286 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002287 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002288 c->tid = next_tid(c->tid);
2289 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002290 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002291
Christoph Lameter81819f02007-05-06 14:49:36 -07002292new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002293
Christoph Lameter49e22582011-08-09 16:12:27 -05002294 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002295 page = c->page = c->partial;
2296 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002297 stat(s, CPU_PARTIAL_ALLOC);
2298 c->freelist = NULL;
2299 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002300 }
2301
Christoph Lameter188fd062012-05-09 10:09:55 -05002302 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002303
Christoph Lameterf46974362012-05-09 10:09:54 -05002304 if (unlikely(!freelist)) {
2305 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2306 slab_out_of_memory(s, gfpflags, node);
Christoph Lameter01ad8a72011-04-15 14:48:14 -05002307
Christoph Lameterf46974362012-05-09 10:09:54 -05002308 local_irq_restore(flags);
2309 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002310 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002311
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002312 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002313 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002314 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002315
Christoph Lameter497b66f2011-08-09 16:12:26 -05002316 /* Only entered in the debug case */
Christoph Lameter5091b742012-07-31 16:44:00 -07002317 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002318 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002319
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002320 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002321 c->page = NULL;
2322 c->freelist = NULL;
Christoph Lametera71ae472011-05-25 09:47:43 -05002323 local_irq_restore(flags);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002324 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002325}
2326
2327/*
2328 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2329 * have the fastpath folded into their functions. So no function call
2330 * overhead for requests that can be satisfied on the fastpath.
2331 *
2332 * The fastpath works by first checking if the lockless freelist can be used.
2333 * If not then __slab_alloc is called for slow processing.
2334 *
2335 * Otherwise we can simply pick the next object from the lockless free list.
2336 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002337static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002338 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002339{
Christoph Lameter894b8782007-05-10 03:15:16 -07002340 void **object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002341 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002342 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002343 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002344
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002345 if (slab_pre_alloc_hook(s, gfpflags))
Akinobu Mita773ff602008-12-23 19:37:01 +09002346 return NULL;
2347
Glauber Costad79923f2012-12-18 14:22:48 -08002348 s = memcg_kmem_get_cache(s, gfpflags);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002349redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002350 /*
2351 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2352 * enabled. We may switch back and forth between cpus while
2353 * reading from one cpu area. That does not matter as long
2354 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002355 *
2356 * Preemption is disabled for the retrieval of the tid because that
2357 * must occur from the current processor. We cannot allow rescheduling
2358 * on a different processor between the determination of the pointer
2359 * and the retrieval of the tid.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002360 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002361 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002362 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002363
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002364 /*
2365 * The transaction ids are globally unique per cpu and per operation on
2366 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2367 * occurs on the right processor and that there was no operation on the
2368 * linked list in between.
2369 */
2370 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002371 preempt_enable();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002372
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002373 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002374 page = c->page;
Steven Rostedtc25f1952013-01-17 13:10:58 -05002375 if (unlikely(!object || !page || !node_match(page, node)))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002376 object = __slab_alloc(s, gfpflags, node, addr, c);
Christoph Lameter894b8782007-05-10 03:15:16 -07002377
2378 else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002379 void *next_object = get_freepointer_safe(s, object);
2380
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002381 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002382 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002383 * operation and if we are on the right processor.
2384 *
2385 * The cmpxchg does the following atomically (without lock semantics!)
2386 * 1. Relocate first pointer to the current per cpu area.
2387 * 2. Verify that tid and freelist have not been changed
2388 * 3. If they were not changed replace tid and freelist
2389 *
2390 * Since this is without lock semantics the protection is only against
2391 * code executing on this cpu *not* from access by other cpus.
2392 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002393 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002394 s->cpu_slab->freelist, s->cpu_slab->tid,
2395 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002396 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002397
2398 note_cmpxchg_failure("slab_alloc", s, tid);
2399 goto redo;
2400 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002401 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002402 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002403 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002404
Pekka Enberg74e21342009-11-25 20:14:48 +02002405 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002406 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002407
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002408 slab_post_alloc_hook(s, gfpflags, object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002409
Christoph Lameter894b8782007-05-10 03:15:16 -07002410 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002411}
2412
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002413static __always_inline void *slab_alloc(struct kmem_cache *s,
2414 gfp_t gfpflags, unsigned long addr)
2415{
2416 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2417}
2418
Christoph Lameter81819f02007-05-06 14:49:36 -07002419void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2420{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002421 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002422
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002423 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002424
2425 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002426}
2427EXPORT_SYMBOL(kmem_cache_alloc);
2428
Li Zefan0f24f122009-12-11 15:45:30 +08002429#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002430void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002431{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002432 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002433 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2434 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002435}
Richard Kennedy4a923792010-10-21 10:29:19 +01002436EXPORT_SYMBOL(kmem_cache_alloc_trace);
2437
2438void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2439{
2440 void *ret = kmalloc_order(size, flags, order);
2441 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2442 return ret;
2443}
2444EXPORT_SYMBOL(kmalloc_order_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002445#endif
2446
Christoph Lameter81819f02007-05-06 14:49:36 -07002447#ifdef CONFIG_NUMA
2448void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2449{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002450 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002451
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002452 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002453 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002454
2455 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002456}
2457EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002458
Li Zefan0f24f122009-12-11 15:45:30 +08002459#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002460void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002461 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002462 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002463{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002464 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002465
2466 trace_kmalloc_node(_RET_IP_, ret,
2467 size, s->size, gfpflags, node);
2468 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002469}
Richard Kennedy4a923792010-10-21 10:29:19 +01002470EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002471#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002472#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002473
Christoph Lameter81819f02007-05-06 14:49:36 -07002474/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002475 * Slow patch handling. This may still be called frequently since objects
2476 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002477 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002478 * So we still attempt to reduce cache line usage. Just take the slab
2479 * lock and free the item. If there is no additional partial page
2480 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002481 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002482static void __slab_free(struct kmem_cache *s, struct page *page,
Christoph Lameterff120592009-12-18 16:26:22 -06002483 void *x, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07002484{
2485 void *prior;
2486 void **object = (void *)x;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002487 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002488 struct page new;
2489 unsigned long counters;
2490 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002491 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002492
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002493 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002494
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002495 if (kmem_cache_debug(s) &&
2496 !(n = free_debug_processing(s, page, x, addr, &flags)))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002497 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002498
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002499 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002500 if (unlikely(n)) {
2501 spin_unlock_irqrestore(&n->list_lock, flags);
2502 n = NULL;
2503 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002504 prior = page->freelist;
2505 counters = page->counters;
2506 set_freepointer(s, object, prior);
2507 new.counters = counters;
2508 was_frozen = new.frozen;
2509 new.inuse--;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002510 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002511
Joonsoo Kim345c9052013-06-19 14:05:52 +09002512 if (kmem_cache_has_cpu_partial(s) && !prior)
Christoph Lameter49e22582011-08-09 16:12:27 -05002513
2514 /*
2515 * Slab was on no list before and will be partially empty
2516 * We can defer the list move and instead freeze it.
2517 */
2518 new.frozen = 1;
2519
2520 else { /* Needs to be taken off a list */
2521
2522 n = get_node(s, page_to_nid(page));
2523 /*
2524 * Speculatively acquire the list_lock.
2525 * If the cmpxchg does not succeed then we may
2526 * drop the list_lock without any processing.
2527 *
2528 * Otherwise the list_lock will synchronize with
2529 * other processors updating the list of slabs.
2530 */
2531 spin_lock_irqsave(&n->list_lock, flags);
2532
2533 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002534 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002535
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002536 } while (!cmpxchg_double_slab(s, page,
2537 prior, counters,
2538 object, new.counters,
2539 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002540
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002541 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002542
2543 /*
2544 * If we just froze the page then put it onto the
2545 * per cpu partial list.
2546 */
Alex Shi8028dce2012-02-03 23:34:56 +08002547 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002548 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002549 stat(s, CPU_PARTIAL_FREE);
2550 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002551 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002552 * The list lock was not taken therefore no list
2553 * activity can be necessary.
2554 */
2555 if (was_frozen)
2556 stat(s, FREE_FROZEN);
Christoph Lameter80f08c12011-06-01 12:25:55 -05002557 return;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002558 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002559
Joonsoo Kim837d6782012-08-16 00:02:40 +09002560 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2561 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002562
Joonsoo Kim837d6782012-08-16 00:02:40 +09002563 /*
2564 * Objects left in the slab. If it was not on the partial list before
2565 * then add it.
2566 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002567 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2568 if (kmem_cache_debug(s))
2569 remove_full(s, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002570 add_partial(n, page, DEACTIVATE_TO_TAIL);
2571 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002572 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002573 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002574 return;
2575
2576slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002577 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002578 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002579 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002580 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002581 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002582 stat(s, FREE_REMOVE_PARTIAL);
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002583 } else
2584 /* Slab must be on the full list */
2585 remove_full(s, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002586
Christoph Lameter80f08c12011-06-01 12:25:55 -05002587 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002588 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002589 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002590}
2591
Christoph Lameter894b8782007-05-10 03:15:16 -07002592/*
2593 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2594 * can perform fastpath freeing without additional function calls.
2595 *
2596 * The fastpath is only possible if we are freeing to the current cpu slab
2597 * of this processor. This typically the case if we have just allocated
2598 * the item before.
2599 *
2600 * If fastpath is not possible then fall back to __slab_free where we deal
2601 * with all sorts of special processing.
2602 */
Pekka Enberg06428782008-01-07 23:20:27 -08002603static __always_inline void slab_free(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002604 struct page *page, void *x, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002605{
2606 void **object = (void *)x;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002607 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002608 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002609
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002610 slab_free_hook(s, x);
2611
Christoph Lametera24c5a02011-03-15 12:45:21 -05002612redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002613 /*
2614 * Determine the currently cpus per cpu slab.
2615 * The cpu may change afterward. However that does not matter since
2616 * data is retrieved via this pointer. If we are on the same cpu
2617 * during the cmpxchg then the free will succedd.
2618 */
Christoph Lameter7cccd802013-01-23 21:45:48 +00002619 preempt_disable();
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002620 c = __this_cpu_ptr(s->cpu_slab);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002621
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002622 tid = c->tid;
Christoph Lameter7cccd802013-01-23 21:45:48 +00002623 preempt_enable();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002624
Christoph Lameter442b06b2011-05-17 16:29:31 -05002625 if (likely(page == c->page)) {
Christoph Lameterff120592009-12-18 16:26:22 -06002626 set_freepointer(s, object, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002627
Christoph Lameter933393f2011-12-22 11:58:51 -06002628 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002629 s->cpu_slab->freelist, s->cpu_slab->tid,
2630 c->freelist, tid,
2631 object, next_tid(tid)))) {
2632
2633 note_cmpxchg_failure("slab_free", s, tid);
2634 goto redo;
2635 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002636 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002637 } else
Christoph Lameterff120592009-12-18 16:26:22 -06002638 __slab_free(s, page, x, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002639
Christoph Lameter894b8782007-05-10 03:15:16 -07002640}
2641
Christoph Lameter81819f02007-05-06 14:49:36 -07002642void kmem_cache_free(struct kmem_cache *s, void *x)
2643{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002644 s = cache_from_obj(s, x);
2645 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002646 return;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002647 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002648 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002649}
2650EXPORT_SYMBOL(kmem_cache_free);
2651
Christoph Lameter81819f02007-05-06 14:49:36 -07002652/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002653 * Object placement in a slab is made very easy because we always start at
2654 * offset 0. If we tune the size of the object to the alignment then we can
2655 * get the required alignment by putting one properly sized object after
2656 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002657 *
2658 * Notice that the allocation order determines the sizes of the per cpu
2659 * caches. Each processor has always one slab available for allocations.
2660 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07002661 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07002662 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07002663 */
2664
2665/*
2666 * Mininum / Maximum order of slab pages. This influences locking overhead
2667 * and slab fragmentation. A higher order reduces the number of partial slabs
2668 * and increases the number of allocations possible without having to
2669 * take the list_lock.
2670 */
2671static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03002672static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002673static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002674
2675/*
2676 * Merge control. If this is set then no merging of slab caches will occur.
Christoph Lameter672bba32007-05-09 02:32:39 -07002677 * (Could be removed. This was introduced to pacify the merge skeptics.)
Christoph Lameter81819f02007-05-06 14:49:36 -07002678 */
2679static int slub_nomerge;
2680
2681/*
Christoph Lameter81819f02007-05-06 14:49:36 -07002682 * Calculate the order of allocation given an slab object size.
2683 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002684 * The order of allocation has significant impact on performance and other
2685 * system components. Generally order 0 allocations should be preferred since
2686 * order 0 does not cause fragmentation in the page allocator. Larger objects
2687 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002688 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07002689 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07002690 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002691 * In order to reach satisfactory performance we must ensure that a minimum
2692 * number of objects is in one slab. Otherwise we may generate too much
2693 * activity on the partial lists which requires taking the list_lock. This is
2694 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07002695 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002696 * slub_max_order specifies the order where we begin to stop considering the
2697 * number of objects in a slab as critical. If we reach slub_max_order then
2698 * we try to keep the page order as low as possible. So we accept more waste
2699 * of space in favor of a small page order.
2700 *
2701 * Higher order allocations also allow the placement of more objects in a
2702 * slab and thereby reduce object handling overhead. If the user has
2703 * requested a higher mininum order then we start with that one instead of
2704 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07002705 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002706static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002707 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002708{
2709 int order;
2710 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07002711 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002712
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002713 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04002714 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03002715
Christoph Lameter6300ea72007-07-17 04:03:20 -07002716 for (order = max(min_order,
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002717 fls(min_objects * size - 1) - PAGE_SHIFT);
2718 order <= max_order; order++) {
2719
Christoph Lameter81819f02007-05-06 14:49:36 -07002720 unsigned long slab_size = PAGE_SIZE << order;
2721
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002722 if (slab_size < min_objects * size + reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002723 continue;
2724
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002725 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07002726
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002727 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07002728 break;
2729
2730 }
Christoph Lameter672bba32007-05-09 02:32:39 -07002731
Christoph Lameter81819f02007-05-06 14:49:36 -07002732 return order;
2733}
2734
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002735static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002736{
2737 int order;
2738 int min_objects;
2739 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002740 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002741
2742 /*
2743 * Attempt to find best configuration for a slab. This
2744 * works by first attempting to generate a layout with
2745 * the best configuration and backing off gradually.
2746 *
2747 * First we reduce the acceptable waste in a slab. Then
2748 * we reduce the minimum objects required in a slab.
2749 */
2750 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002751 if (!min_objects)
2752 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002753 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02002754 min_objects = min(min_objects, max_objects);
2755
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002756 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002757 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002758 while (fraction >= 4) {
2759 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002760 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002761 if (order <= slub_max_order)
2762 return order;
2763 fraction /= 2;
2764 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03002765 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002766 }
2767
2768 /*
2769 * We were unable to place multiple objects in a slab. Now
2770 * lets see if we can place a single object there.
2771 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002772 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002773 if (order <= slub_max_order)
2774 return order;
2775
2776 /*
2777 * Doh this slab cannot be placed using slub_max_order.
2778 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002779 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03002780 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002781 return order;
2782 return -ENOSYS;
2783}
2784
Pekka Enberg5595cff2008-08-05 09:28:47 +03002785static void
Joonsoo Kim40534972012-05-11 00:50:47 +09002786init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07002787{
2788 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07002789 spin_lock_init(&n->list_lock);
2790 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002791#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03002792 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07002793 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07002794 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002795#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002796}
2797
Christoph Lameter55136592010-08-20 12:37:13 -05002798static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002799{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05002800 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00002801 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002802
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002803 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002804 * Must align to double word boundary for the double cmpxchg
2805 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002806 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04002807 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2808 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002809
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002810 if (!s->cpu_slab)
2811 return 0;
2812
2813 init_kmem_cache_cpus(s);
2814
2815 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002816}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002817
Christoph Lameter51df1142010-08-20 12:37:15 -05002818static struct kmem_cache *kmem_cache_node;
2819
Christoph Lameter81819f02007-05-06 14:49:36 -07002820/*
2821 * No kmalloc_node yet so do it by hand. We know that this is the first
2822 * slab on the node for this slabcache. There are no concurrent accesses
2823 * possible.
2824 *
2825 * Note that this function only works on the kmalloc_node_cache
Christoph Lameter4c93c3552007-10-16 01:26:08 -07002826 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2827 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07002828 */
Christoph Lameter55136592010-08-20 12:37:13 -05002829static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07002830{
2831 struct page *page;
2832 struct kmem_cache_node *n;
2833
Christoph Lameter51df1142010-08-20 12:37:15 -05002834 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07002835
Christoph Lameter51df1142010-08-20 12:37:15 -05002836 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002837
2838 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07002839 if (page_to_nid(page) != node) {
2840 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2841 "node %d\n", node);
2842 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2843 "in order to be able to continue\n");
2844 }
2845
Christoph Lameter81819f02007-05-06 14:49:36 -07002846 n = page->freelist;
2847 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002848 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05002849 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05002850 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05002851 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07002852#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05002853 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05002854 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07002855#endif
Joonsoo Kim40534972012-05-11 00:50:47 +09002856 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05002857 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002858
Shaohua Li136333d2011-08-24 08:57:52 +08002859 add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07002860}
2861
2862static void free_kmem_cache_nodes(struct kmem_cache *s)
2863{
2864 int node;
2865
Christoph Lameterf64dc582007-10-16 01:25:33 -07002866 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002867 struct kmem_cache_node *n = s->node[node];
Christoph Lameter51df1142010-08-20 12:37:15 -05002868
Alexander Duyck73367bd2010-05-21 14:41:35 -07002869 if (n)
Christoph Lameter51df1142010-08-20 12:37:15 -05002870 kmem_cache_free(kmem_cache_node, n);
2871
Christoph Lameter81819f02007-05-06 14:49:36 -07002872 s->node[node] = NULL;
2873 }
2874}
2875
Christoph Lameter55136592010-08-20 12:37:13 -05002876static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07002877{
2878 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07002879
Christoph Lameterf64dc582007-10-16 01:25:33 -07002880 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002881 struct kmem_cache_node *n;
2882
Alexander Duyck73367bd2010-05-21 14:41:35 -07002883 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05002884 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002885 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07002886 }
Christoph Lameter51df1142010-08-20 12:37:15 -05002887 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05002888 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07002889
2890 if (!n) {
2891 free_kmem_cache_nodes(s);
2892 return 0;
2893 }
2894
Christoph Lameter81819f02007-05-06 14:49:36 -07002895 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09002896 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07002897 }
2898 return 1;
2899}
Christoph Lameter81819f02007-05-06 14:49:36 -07002900
David Rientjesc0bdb232009-02-25 09:16:35 +02002901static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08002902{
2903 if (min < MIN_PARTIAL)
2904 min = MIN_PARTIAL;
2905 else if (min > MAX_PARTIAL)
2906 min = MAX_PARTIAL;
2907 s->min_partial = min;
2908}
2909
Christoph Lameter81819f02007-05-06 14:49:36 -07002910/*
2911 * calculate_sizes() determines the order and the distribution of data within
2912 * a slab object.
2913 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03002914static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07002915{
2916 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002917 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03002918 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07002919
2920 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08002921 * Round up object size to the next word boundary. We can only
2922 * place the free pointer at word boundaries and this determines
2923 * the possible location of the free pointer.
2924 */
2925 size = ALIGN(size, sizeof(void *));
2926
2927#ifdef CONFIG_SLUB_DEBUG
2928 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002929 * Determine if we can poison the object itself. If the user of
2930 * the slab may touch the object after free or before allocation
2931 * then we should never poison the object itself.
2932 */
2933 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07002934 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07002935 s->flags |= __OBJECT_POISON;
2936 else
2937 s->flags &= ~__OBJECT_POISON;
2938
Christoph Lameter81819f02007-05-06 14:49:36 -07002939
2940 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002941 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07002942 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07002943 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07002944 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002945 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07002946 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002947#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07002948
2949 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07002950 * With that we have determined the number of bytes in actual use
2951 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07002952 */
2953 s->inuse = size;
2954
2955 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07002956 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002957 /*
2958 * Relocate free pointer after the object if it is not
2959 * permitted to overwrite the first word of the object on
2960 * kmem_cache_free.
2961 *
2962 * This is the case if we do RCU, have a constructor or
2963 * destructor or are poisoning the objects.
2964 */
2965 s->offset = size;
2966 size += sizeof(void *);
2967 }
2968
Christoph Lameterc12b3c62007-05-23 13:57:31 -07002969#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07002970 if (flags & SLAB_STORE_USER)
2971 /*
2972 * Need to store information about allocs and frees after
2973 * the object.
2974 */
2975 size += 2 * sizeof(struct track);
2976
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07002977 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07002978 /*
2979 * Add some empty padding so that we can catch
2980 * overwrites from earlier objects rather than let
2981 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01002982 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07002983 * of the object.
2984 */
2985 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07002986#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07002987
Christoph Lameter81819f02007-05-06 14:49:36 -07002988 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07002989 * SLUB stores one object immediately after another beginning from
2990 * offset 0. In order to align the objects we have to simply size
2991 * each object to conform to the alignment.
2992 */
Christoph Lameter45906852012-11-28 16:23:16 +00002993 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07002994 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03002995 if (forced_order >= 0)
2996 order = forced_order;
2997 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002998 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07002999
Christoph Lameter834f3d12008-04-14 19:11:31 +03003000 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003001 return 0;
3002
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003003 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003004 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003005 s->allocflags |= __GFP_COMP;
3006
3007 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003008 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003009
3010 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3011 s->allocflags |= __GFP_RECLAIMABLE;
3012
Christoph Lameter81819f02007-05-06 14:49:36 -07003013 /*
3014 * Determine the number of objects per slab
3015 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003016 s->oo = oo_make(order, size, s->reserved);
3017 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003018 if (oo_objects(s->oo) > oo_objects(s->max))
3019 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003020
Christoph Lameter834f3d12008-04-14 19:11:31 +03003021 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003022}
3023
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003024static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003025{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003026 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003027 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003028
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003029 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3030 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003031
Christoph Lameter06b285d2008-04-14 19:11:41 +03003032 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003033 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003034 if (disable_higher_order_debug) {
3035 /*
3036 * Disable debugging flags that store metadata if the min slab
3037 * order increased.
3038 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003039 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003040 s->flags &= ~DEBUG_METADATA_FLAGS;
3041 s->offset = 0;
3042 if (!calculate_sizes(s, -1))
3043 goto error;
3044 }
3045 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003046
Heiko Carstens25654092012-01-12 17:17:33 -08003047#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3048 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003049 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3050 /* Enable fast mode */
3051 s->flags |= __CMPXCHG_DOUBLE;
3052#endif
3053
David Rientjes3b89d7d2009-02-22 17:40:07 -08003054 /*
3055 * The larger the object size is, the more pages we want on the partial
3056 * list to avoid pounding the page allocator excessively.
3057 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003058 set_min_partial(s, ilog2(s->size) / 2);
3059
3060 /*
3061 * cpu_partial determined the maximum number of objects kept in the
3062 * per cpu partial lists of a processor.
3063 *
3064 * Per cpu partial lists mainly contain slabs that just have one
3065 * object freed. If they are used for allocation then they can be
3066 * filled up again with minimal effort. The slab will never hit the
3067 * per node partial lists and therefore no locking will be required.
3068 *
3069 * This setting also determines
3070 *
3071 * A) The number of objects from per cpu partial slabs dumped to the
3072 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003073 * B) The number of objects in cpu partial slabs to extract from the
Christoph Lameter49e22582011-08-09 16:12:27 -05003074 * per node list when we run out of per cpu objects. We only fetch 50%
3075 * to keep some capacity around for frees.
3076 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003077 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003078 s->cpu_partial = 0;
3079 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003080 s->cpu_partial = 2;
3081 else if (s->size >= 1024)
3082 s->cpu_partial = 6;
3083 else if (s->size >= 256)
3084 s->cpu_partial = 13;
3085 else
3086 s->cpu_partial = 30;
3087
Christoph Lameter81819f02007-05-06 14:49:36 -07003088#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003089 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003090#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003091 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003092 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003093
Christoph Lameter55136592010-08-20 12:37:13 -05003094 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003095 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003096
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003097 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003098error:
3099 if (flags & SLAB_PANIC)
3100 panic("Cannot create slab %s size=%lu realsize=%u "
3101 "order=%u offset=%u flags=%lx\n",
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003102 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
Christoph Lameter81819f02007-05-06 14:49:36 -07003103 s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003104 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003105}
Christoph Lameter81819f02007-05-06 14:49:36 -07003106
Christoph Lameter33b12c32008-04-25 12:22:43 -07003107static void list_slab_objects(struct kmem_cache *s, struct page *page,
3108 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003109{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003110#ifdef CONFIG_SLUB_DEBUG
3111 void *addr = page_address(page);
3112 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003113 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3114 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003115 if (!map)
3116 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003117 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003118 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003119
Christoph Lameter5f80b132011-04-15 14:48:13 -05003120 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003121 for_each_object(p, s, addr, page->objects) {
3122
3123 if (!test_bit(slab_index(p, s, addr), map)) {
3124 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3125 p, p - addr);
3126 print_tracking(s, p);
3127 }
3128 }
3129 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003130 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003131#endif
3132}
3133
Christoph Lameter81819f02007-05-06 14:49:36 -07003134/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003135 * Attempt to free all partial slabs on a node.
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003136 * This is called from kmem_cache_close(). We must be the last thread
3137 * using the cache and therefore we do not need to lock anymore.
Christoph Lameter81819f02007-05-06 14:49:36 -07003138 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003139static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003140{
Christoph Lameter81819f02007-05-06 14:49:36 -07003141 struct page *page, *h;
3142
Christoph Lameter33b12c32008-04-25 12:22:43 -07003143 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003144 if (!page->inuse) {
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05003145 remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003146 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003147 } else {
3148 list_slab_objects(s, page,
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003149 "Objects remaining in %s on kmem_cache_close()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003150 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003151 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003152}
3153
3154/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003155 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003156 */
Christoph Lameter0c710012007-07-17 04:03:24 -07003157static inline int kmem_cache_close(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003158{
3159 int node;
3160
3161 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003162 /* Attempt to free all objects */
Christoph Lameterf64dc582007-10-16 01:25:33 -07003163 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003164 struct kmem_cache_node *n = get_node(s, node);
3165
Christoph Lameter599870b2008-04-23 12:36:52 -07003166 free_partial(s, n);
3167 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003168 return 1;
3169 }
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003170 free_percpu(s->cpu_slab);
Christoph Lameter81819f02007-05-06 14:49:36 -07003171 free_kmem_cache_nodes(s);
3172 return 0;
3173}
3174
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003175int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003176{
Christoph Lameter12c36672012-09-04 23:38:33 +00003177 int rc = kmem_cache_close(s);
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003178
Glauber Costa5413dfb2012-12-18 14:23:13 -08003179 if (!rc) {
3180 /*
3181 * We do the same lock strategy around sysfs_slab_add, see
3182 * __kmem_cache_create. Because this is pretty much the last
3183 * operation we do and the lock will be released shortly after
3184 * that in slab_common.c, we could just move sysfs_slab_remove
3185 * to a later point in common code. We should do that when we
3186 * have a common sysfs framework for all allocators.
3187 */
3188 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003189 sysfs_slab_remove(s);
Glauber Costa5413dfb2012-12-18 14:23:13 -08003190 mutex_lock(&slab_mutex);
3191 }
Christoph Lameter12c36672012-09-04 23:38:33 +00003192
3193 return rc;
Christoph Lameter81819f02007-05-06 14:49:36 -07003194}
Christoph Lameter81819f02007-05-06 14:49:36 -07003195
3196/********************************************************************
3197 * Kmalloc subsystem
3198 *******************************************************************/
3199
Christoph Lameter81819f02007-05-06 14:49:36 -07003200static int __init setup_slub_min_order(char *str)
3201{
Pekka Enberg06428782008-01-07 23:20:27 -08003202 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003203
3204 return 1;
3205}
3206
3207__setup("slub_min_order=", setup_slub_min_order);
3208
3209static int __init setup_slub_max_order(char *str)
3210{
Pekka Enberg06428782008-01-07 23:20:27 -08003211 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003212 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003213
3214 return 1;
3215}
3216
3217__setup("slub_max_order=", setup_slub_max_order);
3218
3219static int __init setup_slub_min_objects(char *str)
3220{
Pekka Enberg06428782008-01-07 23:20:27 -08003221 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003222
3223 return 1;
3224}
3225
3226__setup("slub_min_objects=", setup_slub_min_objects);
3227
3228static int __init setup_slub_nomerge(char *str)
3229{
3230 slub_nomerge = 1;
3231 return 1;
3232}
3233
3234__setup("slub_nomerge", setup_slub_nomerge);
3235
Christoph Lameter81819f02007-05-06 14:49:36 -07003236void *__kmalloc(size_t size, gfp_t flags)
3237{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003238 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003239 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003240
Christoph Lameter95a05b42013-01-10 19:14:19 +00003241 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003242 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003243
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003244 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003245
3246 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003247 return s;
3248
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003249 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003250
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003251 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003252
3253 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003254}
3255EXPORT_SYMBOL(__kmalloc);
3256
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003257#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003258static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3259{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003260 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003261 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003262
Glauber Costad79923f2012-12-18 14:22:48 -08003263 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003264 page = alloc_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003265 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003266 ptr = page_address(page);
3267
3268 kmemleak_alloc(ptr, size, 1, flags);
3269 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003270}
3271
Christoph Lameter81819f02007-05-06 14:49:36 -07003272void *__kmalloc_node(size_t size, gfp_t flags, int node)
3273{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003274 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003275 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003276
Christoph Lameter95a05b42013-01-10 19:14:19 +00003277 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003278 ret = kmalloc_large_node(size, flags, node);
3279
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003280 trace_kmalloc_node(_RET_IP_, ret,
3281 size, PAGE_SIZE << get_order(size),
3282 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003283
3284 return ret;
3285 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003286
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003287 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003288
3289 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003290 return s;
3291
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003292 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003293
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003294 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003295
3296 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003297}
3298EXPORT_SYMBOL(__kmalloc_node);
3299#endif
3300
3301size_t ksize(const void *object)
3302{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003303 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003304
Christoph Lameteref8b4522007-10-16 01:24:46 -07003305 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003306 return 0;
3307
Vegard Nossum294a80a2007-12-04 23:45:30 -08003308 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003309
Pekka Enberg76994412008-05-22 19:22:25 +03003310 if (unlikely(!PageSlab(page))) {
3311 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003312 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003313 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003314
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003315 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003316}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003317EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003318
Ben Greeard18a90d2011-07-07 11:36:37 -07003319#ifdef CONFIG_SLUB_DEBUG
3320bool verify_mem_not_deleted(const void *x)
3321{
3322 struct page *page;
3323 void *object = (void *)x;
3324 unsigned long flags;
3325 bool rv;
3326
3327 if (unlikely(ZERO_OR_NULL_PTR(x)))
3328 return false;
3329
3330 local_irq_save(flags);
3331
3332 page = virt_to_head_page(x);
3333 if (unlikely(!PageSlab(page))) {
3334 /* maybe it was from stack? */
3335 rv = true;
3336 goto out_unlock;
3337 }
3338
3339 slab_lock(page);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003340 if (on_freelist(page->slab_cache, page, object)) {
3341 object_err(page->slab_cache, page, object, "Object is on free-list");
Ben Greeard18a90d2011-07-07 11:36:37 -07003342 rv = false;
3343 } else {
3344 rv = true;
3345 }
3346 slab_unlock(page);
3347
3348out_unlock:
3349 local_irq_restore(flags);
3350 return rv;
3351}
3352EXPORT_SYMBOL(verify_mem_not_deleted);
3353#endif
3354
Christoph Lameter81819f02007-05-06 14:49:36 -07003355void kfree(const void *x)
3356{
Christoph Lameter81819f02007-05-06 14:49:36 -07003357 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003358 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003359
Pekka Enberg2121db72009-03-25 11:05:57 +02003360 trace_kfree(_RET_IP_, x);
3361
Satyam Sharma2408c552007-10-16 01:24:44 -07003362 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003363 return;
3364
Christoph Lameterb49af682007-05-06 14:49:41 -07003365 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003366 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003367 BUG_ON(!PageCompound(page));
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003368 kmemleak_free(x);
Glauber Costad79923f2012-12-18 14:22:48 -08003369 __free_memcg_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003370 return;
3371 }
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003372 slab_free(page->slab_cache, page, object, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003373}
3374EXPORT_SYMBOL(kfree);
3375
Christoph Lameter2086d262007-05-06 14:49:46 -07003376/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003377 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3378 * the remaining slabs by the number of items in use. The slabs with the
3379 * most items in use come first. New allocations will then fill those up
3380 * and thus they can be removed from the partial lists.
3381 *
3382 * The slabs with the least items are placed last. This results in them
3383 * being allocated from last increasing the chance that the last objects
3384 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003385 */
3386int kmem_cache_shrink(struct kmem_cache *s)
3387{
3388 int node;
3389 int i;
3390 struct kmem_cache_node *n;
3391 struct page *page;
3392 struct page *t;
Christoph Lameter205ab992008-04-14 19:11:40 +03003393 int objects = oo_objects(s->max);
Christoph Lameter2086d262007-05-06 14:49:46 -07003394 struct list_head *slabs_by_inuse =
Christoph Lameter834f3d12008-04-14 19:11:31 +03003395 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
Christoph Lameter2086d262007-05-06 14:49:46 -07003396 unsigned long flags;
3397
3398 if (!slabs_by_inuse)
3399 return -ENOMEM;
3400
3401 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003402 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter2086d262007-05-06 14:49:46 -07003403 n = get_node(s, node);
3404
3405 if (!n->nr_partial)
3406 continue;
3407
Christoph Lameter834f3d12008-04-14 19:11:31 +03003408 for (i = 0; i < objects; i++)
Christoph Lameter2086d262007-05-06 14:49:46 -07003409 INIT_LIST_HEAD(slabs_by_inuse + i);
3410
3411 spin_lock_irqsave(&n->list_lock, flags);
3412
3413 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003414 * Build lists indexed by the items in use in each slab.
Christoph Lameter2086d262007-05-06 14:49:46 -07003415 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003416 * Note that concurrent frees may occur while we hold the
3417 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003418 */
3419 list_for_each_entry_safe(page, t, &n->partial, lru) {
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003420 list_move(&page->lru, slabs_by_inuse + page->inuse);
3421 if (!page->inuse)
3422 n->nr_partial--;
Christoph Lameter2086d262007-05-06 14:49:46 -07003423 }
3424
Christoph Lameter2086d262007-05-06 14:49:46 -07003425 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003426 * Rebuild the partial list with the slabs filled up most
3427 * first and the least used slabs at the end.
Christoph Lameter2086d262007-05-06 14:49:46 -07003428 */
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003429 for (i = objects - 1; i > 0; i--)
Christoph Lameter2086d262007-05-06 14:49:46 -07003430 list_splice(slabs_by_inuse + i, n->partial.prev);
3431
Christoph Lameter2086d262007-05-06 14:49:46 -07003432 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003433
3434 /* Release empty slabs */
3435 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3436 discard_slab(s, page);
Christoph Lameter2086d262007-05-06 14:49:46 -07003437 }
3438
3439 kfree(slabs_by_inuse);
3440 return 0;
3441}
3442EXPORT_SYMBOL(kmem_cache_shrink);
3443
Yasunori Gotob9049e22007-10-21 16:41:37 -07003444static int slab_mem_going_offline_callback(void *arg)
3445{
3446 struct kmem_cache *s;
3447
Christoph Lameter18004c52012-07-06 15:25:12 -05003448 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003449 list_for_each_entry(s, &slab_caches, list)
3450 kmem_cache_shrink(s);
Christoph Lameter18004c52012-07-06 15:25:12 -05003451 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003452
3453 return 0;
3454}
3455
3456static void slab_mem_offline_callback(void *arg)
3457{
3458 struct kmem_cache_node *n;
3459 struct kmem_cache *s;
3460 struct memory_notify *marg = arg;
3461 int offline_node;
3462
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003463 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003464
3465 /*
3466 * If the node still has available memory. we need kmem_cache_node
3467 * for it yet.
3468 */
3469 if (offline_node < 0)
3470 return;
3471
Christoph Lameter18004c52012-07-06 15:25:12 -05003472 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003473 list_for_each_entry(s, &slab_caches, list) {
3474 n = get_node(s, offline_node);
3475 if (n) {
3476 /*
3477 * if n->nr_slabs > 0, slabs still exist on the node
3478 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003479 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003480 * callback. So, we must fail.
3481 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003482 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003483
3484 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003485 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003486 }
3487 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003488 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003489}
3490
3491static int slab_mem_going_online_callback(void *arg)
3492{
3493 struct kmem_cache_node *n;
3494 struct kmem_cache *s;
3495 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003496 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003497 int ret = 0;
3498
3499 /*
3500 * If the node's memory is already available, then kmem_cache_node is
3501 * already created. Nothing to do.
3502 */
3503 if (nid < 0)
3504 return 0;
3505
3506 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003507 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003508 * allocate a kmem_cache_node structure in order to bring the node
3509 * online.
3510 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003511 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003512 list_for_each_entry(s, &slab_caches, list) {
3513 /*
3514 * XXX: kmem_cache_alloc_node will fallback to other nodes
3515 * since memory is not yet available from the node that
3516 * is brought up.
3517 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003518 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003519 if (!n) {
3520 ret = -ENOMEM;
3521 goto out;
3522 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003523 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003524 s->node[nid] = n;
3525 }
3526out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003527 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003528 return ret;
3529}
3530
3531static int slab_memory_callback(struct notifier_block *self,
3532 unsigned long action, void *arg)
3533{
3534 int ret = 0;
3535
3536 switch (action) {
3537 case MEM_GOING_ONLINE:
3538 ret = slab_mem_going_online_callback(arg);
3539 break;
3540 case MEM_GOING_OFFLINE:
3541 ret = slab_mem_going_offline_callback(arg);
3542 break;
3543 case MEM_OFFLINE:
3544 case MEM_CANCEL_ONLINE:
3545 slab_mem_offline_callback(arg);
3546 break;
3547 case MEM_ONLINE:
3548 case MEM_CANCEL_OFFLINE:
3549 break;
3550 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003551 if (ret)
3552 ret = notifier_from_errno(ret);
3553 else
3554 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003555 return ret;
3556}
3557
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003558static struct notifier_block slab_memory_callback_nb = {
3559 .notifier_call = slab_memory_callback,
3560 .priority = SLAB_CALLBACK_PRI,
3561};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003562
Christoph Lameter81819f02007-05-06 14:49:36 -07003563/********************************************************************
3564 * Basic setup of slabs
3565 *******************************************************************/
3566
Christoph Lameter51df1142010-08-20 12:37:15 -05003567/*
3568 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003569 * the page allocator. Allocate them properly then fix up the pointers
3570 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003571 */
3572
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003573static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003574{
3575 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003576 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameter51df1142010-08-20 12:37:15 -05003577
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003578 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003579
Glauber Costa7d557b32013-02-22 20:20:00 +04003580 /*
3581 * This runs very early, and only the boot processor is supposed to be
3582 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3583 * IPIs around.
3584 */
3585 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter51df1142010-08-20 12:37:15 -05003586 for_each_node_state(node, N_NORMAL_MEMORY) {
3587 struct kmem_cache_node *n = get_node(s, node);
3588 struct page *p;
3589
3590 if (n) {
3591 list_for_each_entry(p, &n->partial, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003592 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003593
Li Zefan607bf322011-04-12 15:22:26 +08003594#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter51df1142010-08-20 12:37:15 -05003595 list_for_each_entry(p, &n->full, lru)
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003596 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003597#endif
3598 }
3599 }
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003600 list_add(&s->list, &slab_caches);
3601 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003602}
3603
Christoph Lameter81819f02007-05-06 14:49:36 -07003604void __init kmem_cache_init(void)
3605{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003606 static __initdata struct kmem_cache boot_kmem_cache,
3607 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003608
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003609 if (debug_guardpage_minorder())
3610 slub_max_order = 0;
3611
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003612 kmem_cache_node = &boot_kmem_cache_node;
3613 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003614
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003615 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3616 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003617
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003618 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003619
3620 /* Able to allocate the per node structures */
3621 slab_state = PARTIAL;
3622
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003623 create_boot_cache(kmem_cache, "kmem_cache",
3624 offsetof(struct kmem_cache, node) +
3625 nr_node_ids * sizeof(struct kmem_cache_node *),
3626 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003627
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003628 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003629
Christoph Lameter51df1142010-08-20 12:37:15 -05003630 /*
3631 * Allocate kmem_cache_node properly from the kmem_cache slab.
3632 * kmem_cache_node is separately allocated so no need to
3633 * update any list pointers.
3634 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003635 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003636
3637 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003638 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003639
3640#ifdef CONFIG_SMP
3641 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003642#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003643
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003644 printk(KERN_INFO
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003645 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
Christoph Lameter4b356be2007-06-16 10:16:13 -07003646 " CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003647 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003648 slub_min_order, slub_max_order, slub_min_objects,
3649 nr_cpu_ids, nr_node_ids);
3650}
3651
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003652void __init kmem_cache_init_late(void)
3653{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003654}
3655
Christoph Lameter81819f02007-05-06 14:49:36 -07003656/*
3657 * Find a mergeable slab cache
3658 */
3659static int slab_unmergeable(struct kmem_cache *s)
3660{
3661 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3662 return 1;
3663
Christoph Lameterc59def92007-05-16 22:10:50 -07003664 if (s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003665 return 1;
3666
Christoph Lameter8ffa6872007-05-31 00:40:51 -07003667 /*
3668 * We may have set a slab to be unmergeable during bootstrap.
3669 */
3670 if (s->refcount < 0)
3671 return 1;
3672
Christoph Lameter81819f02007-05-06 14:49:36 -07003673 return 0;
3674}
3675
Glauber Costa2633d7a2012-12-18 14:22:34 -08003676static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07003677 size_t align, unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003678 void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003679{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003680 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003681
3682 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3683 return NULL;
3684
Christoph Lameterc59def92007-05-16 22:10:50 -07003685 if (ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003686 return NULL;
3687
3688 size = ALIGN(size, sizeof(void *));
3689 align = calculate_alignment(flags, align, size);
3690 size = ALIGN(size, align);
Christoph Lameterba0268a2007-09-11 15:24:11 -07003691 flags = kmem_cache_flags(size, flags, name, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07003692
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003693 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003694 if (slab_unmergeable(s))
3695 continue;
3696
3697 if (size > s->size)
3698 continue;
3699
Christoph Lameterba0268a2007-09-11 15:24:11 -07003700 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
Christoph Lameter81819f02007-05-06 14:49:36 -07003701 continue;
3702 /*
3703 * Check if alignment is compatible.
3704 * Courtesy of Adrian Drzewiecki
3705 */
Pekka Enberg06428782008-01-07 23:20:27 -08003706 if ((s->size & ~(align - 1)) != s->size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003707 continue;
3708
3709 if (s->size - size >= sizeof(void *))
3710 continue;
3711
Glauber Costa2633d7a2012-12-18 14:22:34 -08003712 if (!cache_match_memcg(s, memcg))
3713 continue;
3714
Christoph Lameter81819f02007-05-06 14:49:36 -07003715 return s;
3716 }
3717 return NULL;
3718}
3719
Glauber Costa2633d7a2012-12-18 14:22:34 -08003720struct kmem_cache *
3721__kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3722 size_t align, unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003723{
3724 struct kmem_cache *s;
3725
Glauber Costa2633d7a2012-12-18 14:22:34 -08003726 s = find_mergeable(memcg, size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003727 if (s) {
3728 s->refcount++;
3729 /*
3730 * Adjust the object sizes so that we clear
3731 * the complete object on kzalloc.
3732 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003733 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003734 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003735
David Rientjes7b8f3b62008-12-17 22:09:46 -08003736 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003737 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003738 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003739 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003740 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003741
Christoph Lametercbb79692012-09-05 00:18:32 +00003742 return s;
3743}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003744
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003745int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003746{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003747 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003748
Pekka Enbergaac3a162012-09-05 12:07:44 +03003749 err = kmem_cache_open(s, flags);
3750 if (err)
3751 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003752
Christoph Lameter45530c42012-11-28 16:23:07 +00003753 /* Mutex is not taken during early boot */
3754 if (slab_state <= UP)
3755 return 0;
3756
Glauber Costa107dab52012-12-18 14:23:05 -08003757 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003758 mutex_unlock(&slab_mutex);
3759 err = sysfs_slab_add(s);
3760 mutex_lock(&slab_mutex);
Christoph Lameter20cea962012-07-06 15:25:13 -05003761
Pekka Enbergaac3a162012-09-05 12:07:44 +03003762 if (err)
3763 kmem_cache_close(s);
3764
3765 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07003766}
Christoph Lameter81819f02007-05-06 14:49:36 -07003767
Christoph Lameter81819f02007-05-06 14:49:36 -07003768#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003769/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003770 * Use the cpu notifier to insure that the cpu slabs are flushed when
3771 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003772 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04003773static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07003774 unsigned long action, void *hcpu)
3775{
3776 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003777 struct kmem_cache *s;
3778 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003779
3780 switch (action) {
3781 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003782 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003783 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003784 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05003785 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003786 list_for_each_entry(s, &slab_caches, list) {
3787 local_irq_save(flags);
3788 __flush_cpu_slab(s, cpu);
3789 local_irq_restore(flags);
3790 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003791 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003792 break;
3793 default:
3794 break;
3795 }
3796 return NOTIFY_OK;
3797}
3798
Paul Gortmaker0db06282013-06-19 14:53:51 -04003799static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003800 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003801};
Christoph Lameter81819f02007-05-06 14:49:36 -07003802
3803#endif
3804
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003805void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003806{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003807 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003808 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003809
Christoph Lameter95a05b42013-01-10 19:14:19 +00003810 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003811 return kmalloc_large(size, gfpflags);
3812
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003813 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003814
Satyam Sharma2408c552007-10-16 01:24:44 -07003815 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003816 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003817
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003818 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003819
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003820 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003821 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003822
3823 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003824}
3825
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003826#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07003827void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003828 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07003829{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003830 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003831 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003832
Christoph Lameter95a05b42013-01-10 19:14:19 +00003833 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08003834 ret = kmalloc_large_node(size, gfpflags, node);
3835
3836 trace_kmalloc_node(caller, ret,
3837 size, PAGE_SIZE << get_order(size),
3838 gfpflags, node);
3839
3840 return ret;
3841 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02003842
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003843 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07003844
Satyam Sharma2408c552007-10-16 01:24:44 -07003845 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003846 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07003847
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003848 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003849
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003850 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003851 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03003852
3853 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003854}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003855#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003856
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003857#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03003858static int count_inuse(struct page *page)
3859{
3860 return page->inuse;
3861}
3862
3863static int count_total(struct page *page)
3864{
3865 return page->objects;
3866}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003867#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03003868
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05003869#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07003870static int validate_slab(struct kmem_cache *s, struct page *page,
3871 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003872{
3873 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08003874 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003875
3876 if (!check_slab(s, page) ||
3877 !on_freelist(s, page, NULL))
3878 return 0;
3879
3880 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03003881 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003882
Christoph Lameter5f80b132011-04-15 14:48:13 -05003883 get_map(s, page, map);
3884 for_each_object(p, s, addr, page->objects) {
3885 if (test_bit(slab_index(p, s, addr), map))
3886 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3887 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003888 }
3889
Christoph Lameter224a88b2008-04-14 19:11:31 +03003890 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07003891 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02003892 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07003893 return 0;
3894 return 1;
3895}
3896
Christoph Lameter434e2452007-07-17 04:03:30 -07003897static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3898 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003899{
Christoph Lameter881db7f2011-06-01 12:25:53 -05003900 slab_lock(page);
3901 validate_slab(s, page, map);
3902 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003903}
3904
Christoph Lameter434e2452007-07-17 04:03:30 -07003905static int validate_slab_node(struct kmem_cache *s,
3906 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003907{
3908 unsigned long count = 0;
3909 struct page *page;
3910 unsigned long flags;
3911
3912 spin_lock_irqsave(&n->list_lock, flags);
3913
3914 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003915 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003916 count++;
3917 }
3918 if (count != n->nr_partial)
3919 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3920 "counter=%ld\n", s->name, count, n->nr_partial);
3921
3922 if (!(s->flags & SLAB_STORE_USER))
3923 goto out;
3924
3925 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07003926 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003927 count++;
3928 }
3929 if (count != atomic_long_read(&n->nr_slabs))
3930 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3931 "counter=%ld\n", s->name, count,
3932 atomic_long_read(&n->nr_slabs));
3933
3934out:
3935 spin_unlock_irqrestore(&n->list_lock, flags);
3936 return count;
3937}
3938
Christoph Lameter434e2452007-07-17 04:03:30 -07003939static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07003940{
3941 int node;
3942 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03003943 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07003944 sizeof(unsigned long), GFP_KERNEL);
3945
3946 if (!map)
3947 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07003948
3949 flush_all(s);
Christoph Lameterf64dc582007-10-16 01:25:33 -07003950 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter53e15af2007-05-06 14:49:43 -07003951 struct kmem_cache_node *n = get_node(s, node);
3952
Christoph Lameter434e2452007-07-17 04:03:30 -07003953 count += validate_slab_node(s, n, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003954 }
Christoph Lameter434e2452007-07-17 04:03:30 -07003955 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07003956 return count;
3957}
Christoph Lameter88a420e2007-05-06 14:49:45 -07003958/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003959 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07003960 * and freed.
3961 */
3962
3963struct location {
3964 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003965 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07003966 long long sum_time;
3967 long min_time;
3968 long max_time;
3969 long min_pid;
3970 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10303971 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07003972 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07003973};
3974
3975struct loc_track {
3976 unsigned long max;
3977 unsigned long count;
3978 struct location *loc;
3979};
3980
3981static void free_loc_track(struct loc_track *t)
3982{
3983 if (t->max)
3984 free_pages((unsigned long)t->loc,
3985 get_order(sizeof(struct location) * t->max));
3986}
3987
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003988static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07003989{
3990 struct location *l;
3991 int order;
3992
Christoph Lameter88a420e2007-05-06 14:49:45 -07003993 order = get_order(sizeof(struct location) * max);
3994
Christoph Lameter68dff6a2007-07-17 04:03:20 -07003995 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07003996 if (!l)
3997 return 0;
3998
3999 if (t->count) {
4000 memcpy(l, t->loc, sizeof(struct location) * t->count);
4001 free_loc_track(t);
4002 }
4003 t->max = max;
4004 t->loc = l;
4005 return 1;
4006}
4007
4008static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004009 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004010{
4011 long start, end, pos;
4012 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004013 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004014 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004015
4016 start = -1;
4017 end = t->count;
4018
4019 for ( ; ; ) {
4020 pos = start + (end - start + 1) / 2;
4021
4022 /*
4023 * There is nothing at "end". If we end up there
4024 * we need to add something to before end.
4025 */
4026 if (pos == end)
4027 break;
4028
4029 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004030 if (track->addr == caddr) {
4031
4032 l = &t->loc[pos];
4033 l->count++;
4034 if (track->when) {
4035 l->sum_time += age;
4036 if (age < l->min_time)
4037 l->min_time = age;
4038 if (age > l->max_time)
4039 l->max_time = age;
4040
4041 if (track->pid < l->min_pid)
4042 l->min_pid = track->pid;
4043 if (track->pid > l->max_pid)
4044 l->max_pid = track->pid;
4045
Rusty Russell174596a2009-01-01 10:12:29 +10304046 cpumask_set_cpu(track->cpu,
4047 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004048 }
4049 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004050 return 1;
4051 }
4052
Christoph Lameter45edfa52007-05-09 02:32:45 -07004053 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004054 end = pos;
4055 else
4056 start = pos;
4057 }
4058
4059 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004060 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004061 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004062 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004063 return 0;
4064
4065 l = t->loc + pos;
4066 if (pos < t->count)
4067 memmove(l + 1, l,
4068 (t->count - pos) * sizeof(struct location));
4069 t->count++;
4070 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004071 l->addr = track->addr;
4072 l->sum_time = age;
4073 l->min_time = age;
4074 l->max_time = age;
4075 l->min_pid = track->pid;
4076 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304077 cpumask_clear(to_cpumask(l->cpus));
4078 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004079 nodes_clear(l->nodes);
4080 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004081 return 1;
4082}
4083
4084static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004085 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004086 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004087{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004088 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004089 void *p;
4090
Christoph Lameter39b26462008-04-14 19:11:30 +03004091 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004092 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004093
Christoph Lameter224a88b2008-04-14 19:11:31 +03004094 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004095 if (!test_bit(slab_index(p, s, addr), map))
4096 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004097}
4098
4099static int list_locations(struct kmem_cache *s, char *buf,
4100 enum track_item alloc)
4101{
Harvey Harrisone374d482008-01-31 15:20:50 -08004102 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004103 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004104 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004105 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004106 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4107 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004108
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004109 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4110 GFP_TEMPORARY)) {
4111 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004112 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004113 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004114 /* Push back cpu slabs */
4115 flush_all(s);
4116
Christoph Lameterf64dc582007-10-16 01:25:33 -07004117 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004118 struct kmem_cache_node *n = get_node(s, node);
4119 unsigned long flags;
4120 struct page *page;
4121
Christoph Lameter9e869432007-08-22 14:01:56 -07004122 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004123 continue;
4124
4125 spin_lock_irqsave(&n->list_lock, flags);
4126 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004127 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004128 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004129 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004130 spin_unlock_irqrestore(&n->list_lock, flags);
4131 }
4132
4133 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004134 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004135
Hugh Dickins9c246242008-12-09 13:14:27 -08004136 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004137 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004138 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004139
4140 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004141 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004142 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004143 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004144
4145 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004146 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004147 l->min_time,
4148 (long)div_u64(l->sum_time, l->count),
4149 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004150 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004151 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004152 l->min_time);
4153
4154 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004155 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004156 l->min_pid, l->max_pid);
4157 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004158 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004159 l->min_pid);
4160
Rusty Russell174596a2009-01-01 10:12:29 +10304161 if (num_online_cpus() > 1 &&
4162 !cpumask_empty(to_cpumask(l->cpus)) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004163 len < PAGE_SIZE - 60) {
4164 len += sprintf(buf + len, " cpus=");
4165 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Rusty Russell174596a2009-01-01 10:12:29 +10304166 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004167 }
4168
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004169 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Harvey Harrisone374d482008-01-31 15:20:50 -08004170 len < PAGE_SIZE - 60) {
4171 len += sprintf(buf + len, " nodes=");
4172 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004173 l->nodes);
4174 }
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;
4271 unsigned long *per_cpu;
4272
4273 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004274 if (!nodes)
4275 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004276 per_cpu = nodes + nr_node_ids;
4277
Christoph Lameter205ab992008-04-14 19:11:40 +03004278 if (flags & SO_CPU) {
4279 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004280
Christoph Lameter205ab992008-04-14 19:11:40 +03004281 for_each_possible_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004282 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004283 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004284 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004285
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004286 page = ACCESS_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004287 if (!page)
4288 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004289
Christoph Lameterec3ab082012-05-09 10:09:56 -05004290 node = page_to_nid(page);
4291 if (flags & SO_TOTAL)
4292 x = page->objects;
4293 else if (flags & SO_OBJECTS)
4294 x = page->inuse;
4295 else
4296 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004297
Christoph Lameterec3ab082012-05-09 10:09:56 -05004298 total += x;
4299 nodes[node] += x;
4300
4301 page = ACCESS_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004302 if (page) {
4303 x = page->pobjects;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004304 total += x;
4305 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004306 }
Christoph Lameterec3ab082012-05-09 10:09:56 -05004307
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004308 per_cpu[node]++;
Christoph Lameter81819f02007-05-06 14:49:36 -07004309 }
4310 }
4311
Christoph Lameter04d94872011-01-10 10:15:15 -06004312 lock_memory_hotplug();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004313#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004314 if (flags & SO_ALL) {
4315 for_each_node_state(node, N_NORMAL_MEMORY) {
4316 struct kmem_cache_node *n = get_node(s, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07004317
Christoph Lameter205ab992008-04-14 19:11:40 +03004318 if (flags & SO_TOTAL)
4319 x = atomic_long_read(&n->total_objects);
4320 else if (flags & SO_OBJECTS)
4321 x = atomic_long_read(&n->total_objects) -
4322 count_partial(n, count_free);
4323
4324 else
4325 x = atomic_long_read(&n->nr_slabs);
4326 total += x;
4327 nodes[node] += x;
4328 }
4329
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004330 } else
4331#endif
4332 if (flags & SO_PARTIAL) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004333 for_each_node_state(node, N_NORMAL_MEMORY) {
4334 struct kmem_cache_node *n = get_node(s, node);
4335
4336 if (flags & SO_TOTAL)
4337 x = count_partial(n, count_total);
4338 else if (flags & SO_OBJECTS)
4339 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004340 else
4341 x = n->nr_partial;
4342 total += x;
4343 nodes[node] += x;
4344 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004345 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004346 x = sprintf(buf, "%lu", total);
4347#ifdef CONFIG_NUMA
Christoph Lameterf64dc582007-10-16 01:25:33 -07004348 for_each_node_state(node, N_NORMAL_MEMORY)
Christoph Lameter81819f02007-05-06 14:49:36 -07004349 if (nodes[node])
4350 x += sprintf(buf + x, " N%d=%lu",
4351 node, nodes[node]);
4352#endif
Christoph Lameter04d94872011-01-10 10:15:15 -06004353 unlock_memory_hotplug();
Christoph Lameter81819f02007-05-06 14:49:36 -07004354 kfree(nodes);
4355 return x + sprintf(buf + x, "\n");
4356}
4357
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004358#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004359static int any_slab_objects(struct kmem_cache *s)
4360{
4361 int node;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004362
4363 for_each_online_node(node) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004364 struct kmem_cache_node *n = get_node(s, node);
4365
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004366 if (!n)
4367 continue;
4368
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004369 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004370 return 1;
4371 }
4372 return 0;
4373}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004374#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004375
4376#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004377#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004378
4379struct slab_attribute {
4380 struct attribute attr;
4381 ssize_t (*show)(struct kmem_cache *s, char *buf);
4382 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4383};
4384
4385#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004386 static struct slab_attribute _name##_attr = \
4387 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004388
4389#define SLAB_ATTR(_name) \
4390 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004391 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004392
Christoph Lameter81819f02007-05-06 14:49:36 -07004393static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4394{
4395 return sprintf(buf, "%d\n", s->size);
4396}
4397SLAB_ATTR_RO(slab_size);
4398
4399static ssize_t align_show(struct kmem_cache *s, char *buf)
4400{
4401 return sprintf(buf, "%d\n", s->align);
4402}
4403SLAB_ATTR_RO(align);
4404
4405static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4406{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004407 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004408}
4409SLAB_ATTR_RO(object_size);
4410
4411static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4412{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004413 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004414}
4415SLAB_ATTR_RO(objs_per_slab);
4416
Christoph Lameter06b285d2008-04-14 19:11:41 +03004417static ssize_t order_store(struct kmem_cache *s,
4418 const char *buf, size_t length)
4419{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004420 unsigned long order;
4421 int err;
4422
4423 err = strict_strtoul(buf, 10, &order);
4424 if (err)
4425 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004426
4427 if (order > slub_max_order || order < slub_min_order)
4428 return -EINVAL;
4429
4430 calculate_sizes(s, order);
4431 return length;
4432}
4433
Christoph Lameter81819f02007-05-06 14:49:36 -07004434static ssize_t order_show(struct kmem_cache *s, char *buf)
4435{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004436 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004437}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004438SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004439
David Rientjes73d342b2009-02-22 17:40:09 -08004440static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4441{
4442 return sprintf(buf, "%lu\n", s->min_partial);
4443}
4444
4445static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4446 size_t length)
4447{
4448 unsigned long min;
4449 int err;
4450
4451 err = strict_strtoul(buf, 10, &min);
4452 if (err)
4453 return err;
4454
David Rientjesc0bdb232009-02-25 09:16:35 +02004455 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004456 return length;
4457}
4458SLAB_ATTR(min_partial);
4459
Christoph Lameter49e22582011-08-09 16:12:27 -05004460static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4461{
4462 return sprintf(buf, "%u\n", s->cpu_partial);
4463}
4464
4465static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4466 size_t length)
4467{
4468 unsigned long objects;
4469 int err;
4470
4471 err = strict_strtoul(buf, 10, &objects);
4472 if (err)
4473 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004474 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004475 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004476
4477 s->cpu_partial = objects;
4478 flush_all(s);
4479 return length;
4480}
4481SLAB_ATTR(cpu_partial);
4482
Christoph Lameter81819f02007-05-06 14:49:36 -07004483static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4484{
Joe Perches62c70bc2011-01-13 15:45:52 -08004485 if (!s->ctor)
4486 return 0;
4487 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004488}
4489SLAB_ATTR_RO(ctor);
4490
Christoph Lameter81819f02007-05-06 14:49:36 -07004491static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4492{
4493 return sprintf(buf, "%d\n", s->refcount - 1);
4494}
4495SLAB_ATTR_RO(aliases);
4496
Christoph Lameter81819f02007-05-06 14:49:36 -07004497static ssize_t partial_show(struct kmem_cache *s, char *buf)
4498{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004499 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004500}
4501SLAB_ATTR_RO(partial);
4502
4503static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4504{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004505 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004506}
4507SLAB_ATTR_RO(cpu_slabs);
4508
4509static ssize_t objects_show(struct kmem_cache *s, char *buf)
4510{
Christoph Lameter205ab992008-04-14 19:11:40 +03004511 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004512}
4513SLAB_ATTR_RO(objects);
4514
Christoph Lameter205ab992008-04-14 19:11:40 +03004515static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4516{
4517 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4518}
4519SLAB_ATTR_RO(objects_partial);
4520
Christoph Lameter49e22582011-08-09 16:12:27 -05004521static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4522{
4523 int objects = 0;
4524 int pages = 0;
4525 int cpu;
4526 int len;
4527
4528 for_each_online_cpu(cpu) {
4529 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4530
4531 if (page) {
4532 pages += page->pages;
4533 objects += page->pobjects;
4534 }
4535 }
4536
4537 len = sprintf(buf, "%d(%d)", objects, pages);
4538
4539#ifdef CONFIG_SMP
4540 for_each_online_cpu(cpu) {
4541 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4542
4543 if (page && len < PAGE_SIZE - 20)
4544 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4545 page->pobjects, page->pages);
4546 }
4547#endif
4548 return len + sprintf(buf + len, "\n");
4549}
4550SLAB_ATTR_RO(slabs_cpu_partial);
4551
Christoph Lameter81819f02007-05-06 14:49:36 -07004552static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4553{
4554 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4555}
4556
4557static ssize_t reclaim_account_store(struct kmem_cache *s,
4558 const char *buf, size_t length)
4559{
4560 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4561 if (buf[0] == '1')
4562 s->flags |= SLAB_RECLAIM_ACCOUNT;
4563 return length;
4564}
4565SLAB_ATTR(reclaim_account);
4566
4567static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4568{
Christoph Lameter5af60832007-05-06 14:49:56 -07004569 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004570}
4571SLAB_ATTR_RO(hwcache_align);
4572
4573#ifdef CONFIG_ZONE_DMA
4574static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4575{
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4577}
4578SLAB_ATTR_RO(cache_dma);
4579#endif
4580
4581static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4582{
4583 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4584}
4585SLAB_ATTR_RO(destroy_by_rcu);
4586
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004587static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4588{
4589 return sprintf(buf, "%d\n", s->reserved);
4590}
4591SLAB_ATTR_RO(reserved);
4592
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004593#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004594static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4595{
4596 return show_slab_objects(s, buf, SO_ALL);
4597}
4598SLAB_ATTR_RO(slabs);
4599
4600static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4601{
4602 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4603}
4604SLAB_ATTR_RO(total_objects);
4605
4606static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4607{
4608 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4609}
4610
4611static ssize_t sanity_checks_store(struct kmem_cache *s,
4612 const char *buf, size_t length)
4613{
4614 s->flags &= ~SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004615 if (buf[0] == '1') {
4616 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004617 s->flags |= SLAB_DEBUG_FREE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004618 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004619 return length;
4620}
4621SLAB_ATTR(sanity_checks);
4622
4623static ssize_t trace_show(struct kmem_cache *s, char *buf)
4624{
4625 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4626}
4627
4628static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4629 size_t length)
4630{
4631 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004632 if (buf[0] == '1') {
4633 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004634 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004635 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004636 return length;
4637}
4638SLAB_ATTR(trace);
4639
Christoph Lameter81819f02007-05-06 14:49:36 -07004640static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4641{
4642 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4643}
4644
4645static ssize_t red_zone_store(struct kmem_cache *s,
4646 const char *buf, size_t length)
4647{
4648 if (any_slab_objects(s))
4649 return -EBUSY;
4650
4651 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004652 if (buf[0] == '1') {
4653 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004654 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004655 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004656 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004657 return length;
4658}
4659SLAB_ATTR(red_zone);
4660
4661static ssize_t poison_show(struct kmem_cache *s, char *buf)
4662{
4663 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4664}
4665
4666static ssize_t poison_store(struct kmem_cache *s,
4667 const char *buf, size_t length)
4668{
4669 if (any_slab_objects(s))
4670 return -EBUSY;
4671
4672 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004673 if (buf[0] == '1') {
4674 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004675 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004676 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004677 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004678 return length;
4679}
4680SLAB_ATTR(poison);
4681
4682static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4683{
4684 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4685}
4686
4687static ssize_t store_user_store(struct kmem_cache *s,
4688 const char *buf, size_t length)
4689{
4690 if (any_slab_objects(s))
4691 return -EBUSY;
4692
4693 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004694 if (buf[0] == '1') {
4695 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004696 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004697 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004698 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004699 return length;
4700}
4701SLAB_ATTR(store_user);
4702
Christoph Lameter53e15af2007-05-06 14:49:43 -07004703static ssize_t validate_show(struct kmem_cache *s, char *buf)
4704{
4705 return 0;
4706}
4707
4708static ssize_t validate_store(struct kmem_cache *s,
4709 const char *buf, size_t length)
4710{
Christoph Lameter434e2452007-07-17 04:03:30 -07004711 int ret = -EINVAL;
4712
4713 if (buf[0] == '1') {
4714 ret = validate_slab_cache(s);
4715 if (ret >= 0)
4716 ret = length;
4717 }
4718 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004719}
4720SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004721
4722static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4723{
4724 if (!(s->flags & SLAB_STORE_USER))
4725 return -ENOSYS;
4726 return list_locations(s, buf, TRACK_ALLOC);
4727}
4728SLAB_ATTR_RO(alloc_calls);
4729
4730static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4731{
4732 if (!(s->flags & SLAB_STORE_USER))
4733 return -ENOSYS;
4734 return list_locations(s, buf, TRACK_FREE);
4735}
4736SLAB_ATTR_RO(free_calls);
4737#endif /* CONFIG_SLUB_DEBUG */
4738
4739#ifdef CONFIG_FAILSLAB
4740static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4741{
4742 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4743}
4744
4745static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4746 size_t length)
4747{
4748 s->flags &= ~SLAB_FAILSLAB;
4749 if (buf[0] == '1')
4750 s->flags |= SLAB_FAILSLAB;
4751 return length;
4752}
4753SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004754#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07004755
Christoph Lameter2086d262007-05-06 14:49:46 -07004756static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4757{
4758 return 0;
4759}
4760
4761static ssize_t shrink_store(struct kmem_cache *s,
4762 const char *buf, size_t length)
4763{
4764 if (buf[0] == '1') {
4765 int rc = kmem_cache_shrink(s);
4766
4767 if (rc)
4768 return rc;
4769 } else
4770 return -EINVAL;
4771 return length;
4772}
4773SLAB_ATTR(shrink);
4774
Christoph Lameter81819f02007-05-06 14:49:36 -07004775#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004776static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004777{
Christoph Lameter98246012008-01-07 23:20:26 -08004778 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004779}
4780
Christoph Lameter98246012008-01-07 23:20:26 -08004781static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004782 const char *buf, size_t length)
4783{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004784 unsigned long ratio;
4785 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004786
Christoph Lameter0121c6192008-04-29 16:11:12 -07004787 err = strict_strtoul(buf, 10, &ratio);
4788 if (err)
4789 return err;
4790
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004791 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004792 s->remote_node_defrag_ratio = ratio * 10;
4793
Christoph Lameter81819f02007-05-06 14:49:36 -07004794 return length;
4795}
Christoph Lameter98246012008-01-07 23:20:26 -08004796SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004797#endif
4798
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004799#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004800static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4801{
4802 unsigned long sum = 0;
4803 int cpu;
4804 int len;
4805 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4806
4807 if (!data)
4808 return -ENOMEM;
4809
4810 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004811 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004812
4813 data[cpu] = x;
4814 sum += x;
4815 }
4816
4817 len = sprintf(buf, "%lu", sum);
4818
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004819#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004820 for_each_online_cpu(cpu) {
4821 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004822 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004823 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03004824#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004825 kfree(data);
4826 return len + sprintf(buf + len, "\n");
4827}
4828
David Rientjes78eb00c2009-10-15 02:20:22 -07004829static void clear_stat(struct kmem_cache *s, enum stat_item si)
4830{
4831 int cpu;
4832
4833 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06004834 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07004835}
4836
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004837#define STAT_ATTR(si, text) \
4838static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4839{ \
4840 return show_stat(s, buf, si); \
4841} \
David Rientjes78eb00c2009-10-15 02:20:22 -07004842static ssize_t text##_store(struct kmem_cache *s, \
4843 const char *buf, size_t length) \
4844{ \
4845 if (buf[0] != '0') \
4846 return -EINVAL; \
4847 clear_stat(s, si); \
4848 return length; \
4849} \
4850SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004851
4852STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4853STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4854STAT_ATTR(FREE_FASTPATH, free_fastpath);
4855STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4856STAT_ATTR(FREE_FROZEN, free_frozen);
4857STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4858STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4859STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4860STAT_ATTR(ALLOC_SLAB, alloc_slab);
4861STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05004862STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004863STAT_ATTR(FREE_SLAB, free_slab);
4864STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4865STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4866STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4867STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4868STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4869STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05004870STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03004871STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05004872STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4873STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05004874STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4875STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08004876STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4877STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004878#endif
4879
Pekka Enberg06428782008-01-07 23:20:27 -08004880static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07004881 &slab_size_attr.attr,
4882 &object_size_attr.attr,
4883 &objs_per_slab_attr.attr,
4884 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08004885 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004886 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004887 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03004888 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004889 &partial_attr.attr,
4890 &cpu_slabs_attr.attr,
4891 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004892 &aliases_attr.attr,
4893 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004894 &hwcache_align_attr.attr,
4895 &reclaim_account_attr.attr,
4896 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05004897 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004898 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004899 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004900#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004901 &total_objects_attr.attr,
4902 &slabs_attr.attr,
4903 &sanity_checks_attr.attr,
4904 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004905 &red_zone_attr.attr,
4906 &poison_attr.attr,
4907 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07004908 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07004909 &alloc_calls_attr.attr,
4910 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004911#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004912#ifdef CONFIG_ZONE_DMA
4913 &cache_dma_attr.attr,
4914#endif
4915#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004916 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07004917#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004918#ifdef CONFIG_SLUB_STATS
4919 &alloc_fastpath_attr.attr,
4920 &alloc_slowpath_attr.attr,
4921 &free_fastpath_attr.attr,
4922 &free_slowpath_attr.attr,
4923 &free_frozen_attr.attr,
4924 &free_add_partial_attr.attr,
4925 &free_remove_partial_attr.attr,
4926 &alloc_from_partial_attr.attr,
4927 &alloc_slab_attr.attr,
4928 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05004929 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004930 &free_slab_attr.attr,
4931 &cpuslab_flush_attr.attr,
4932 &deactivate_full_attr.attr,
4933 &deactivate_empty_attr.attr,
4934 &deactivate_to_head_attr.attr,
4935 &deactivate_to_tail_attr.attr,
4936 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05004937 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03004938 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05004939 &cmpxchg_double_fail_attr.attr,
4940 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05004941 &cpu_partial_alloc_attr.attr,
4942 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08004943 &cpu_partial_node_attr.attr,
4944 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004945#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03004946#ifdef CONFIG_FAILSLAB
4947 &failslab_attr.attr,
4948#endif
4949
Christoph Lameter81819f02007-05-06 14:49:36 -07004950 NULL
4951};
4952
4953static struct attribute_group slab_attr_group = {
4954 .attrs = slab_attrs,
4955};
4956
4957static ssize_t slab_attr_show(struct kobject *kobj,
4958 struct attribute *attr,
4959 char *buf)
4960{
4961 struct slab_attribute *attribute;
4962 struct kmem_cache *s;
4963 int err;
4964
4965 attribute = to_slab_attr(attr);
4966 s = to_slab(kobj);
4967
4968 if (!attribute->show)
4969 return -EIO;
4970
4971 err = attribute->show(s, buf);
4972
4973 return err;
4974}
4975
4976static ssize_t slab_attr_store(struct kobject *kobj,
4977 struct attribute *attr,
4978 const char *buf, size_t len)
4979{
4980 struct slab_attribute *attribute;
4981 struct kmem_cache *s;
4982 int err;
4983
4984 attribute = to_slab_attr(attr);
4985 s = to_slab(kobj);
4986
4987 if (!attribute->store)
4988 return -EIO;
4989
4990 err = attribute->store(s, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08004991#ifdef CONFIG_MEMCG_KMEM
4992 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4993 int i;
Christoph Lameter81819f02007-05-06 14:49:36 -07004994
Glauber Costa107dab52012-12-18 14:23:05 -08004995 mutex_lock(&slab_mutex);
4996 if (s->max_attr_size < len)
4997 s->max_attr_size = len;
4998
Glauber Costaebe945c2012-12-18 14:23:10 -08004999 /*
5000 * This is a best effort propagation, so this function's return
5001 * value will be determined by the parent cache only. This is
5002 * basically because not all attributes will have a well
5003 * defined semantics for rollbacks - most of the actions will
5004 * have permanent effects.
5005 *
5006 * Returning the error value of any of the children that fail
5007 * is not 100 % defined, in the sense that users seeing the
5008 * error code won't be able to know anything about the state of
5009 * the cache.
5010 *
5011 * Only returning the error code for the parent cache at least
5012 * has well defined semantics. The cache being written to
5013 * directly either failed or succeeded, in which case we loop
5014 * through the descendants with best-effort propagation.
5015 */
Glauber Costa107dab52012-12-18 14:23:05 -08005016 for_each_memcg_cache_index(i) {
5017 struct kmem_cache *c = cache_from_memcg(s, i);
Glauber Costa107dab52012-12-18 14:23:05 -08005018 if (c)
5019 attribute->store(c, buf, len);
5020 }
5021 mutex_unlock(&slab_mutex);
5022 }
5023#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005024 return err;
5025}
5026
Glauber Costa107dab52012-12-18 14:23:05 -08005027static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5028{
5029#ifdef CONFIG_MEMCG_KMEM
5030 int i;
5031 char *buffer = NULL;
5032
5033 if (!is_root_cache(s))
5034 return;
5035
5036 /*
5037 * This mean this cache had no attribute written. Therefore, no point
5038 * in copying default values around
5039 */
5040 if (!s->max_attr_size)
5041 return;
5042
5043 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5044 char mbuf[64];
5045 char *buf;
5046 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5047
5048 if (!attr || !attr->store || !attr->show)
5049 continue;
5050
5051 /*
5052 * It is really bad that we have to allocate here, so we will
5053 * do it only as a fallback. If we actually allocate, though,
5054 * we can just use the allocated buffer until the end.
5055 *
5056 * Most of the slub attributes will tend to be very small in
5057 * size, but sysfs allows buffers up to a page, so they can
5058 * theoretically happen.
5059 */
5060 if (buffer)
5061 buf = buffer;
5062 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5063 buf = mbuf;
5064 else {
5065 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5066 if (WARN_ON(!buffer))
5067 continue;
5068 buf = buffer;
5069 }
5070
5071 attr->show(s->memcg_params->root_cache, buf);
5072 attr->store(s, buf, strlen(buf));
5073 }
5074
5075 if (buffer)
5076 free_page((unsigned long)buffer);
5077#endif
5078}
5079
Emese Revfy52cf25d2010-01-19 02:58:23 +01005080static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005081 .show = slab_attr_show,
5082 .store = slab_attr_store,
5083};
5084
5085static struct kobj_type slab_ktype = {
5086 .sysfs_ops = &slab_sysfs_ops,
5087};
5088
5089static int uevent_filter(struct kset *kset, struct kobject *kobj)
5090{
5091 struct kobj_type *ktype = get_ktype(kobj);
5092
5093 if (ktype == &slab_ktype)
5094 return 1;
5095 return 0;
5096}
5097
Emese Revfy9cd43612009-12-31 14:52:51 +01005098static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005099 .filter = uevent_filter,
5100};
5101
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005102static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005103
5104#define ID_STR_LENGTH 64
5105
5106/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005107 *
5108 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005109 */
5110static char *create_unique_id(struct kmem_cache *s)
5111{
5112 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5113 char *p = name;
5114
5115 BUG_ON(!name);
5116
5117 *p++ = ':';
5118 /*
5119 * First flags affecting slabcache operations. We will only
5120 * get here for aliasable slabs so we do not need to support
5121 * too many flags. The flags here must cover all flags that
5122 * are matched during merging to guarantee that the id is
5123 * unique.
5124 */
5125 if (s->flags & SLAB_CACHE_DMA)
5126 *p++ = 'd';
5127 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5128 *p++ = 'a';
5129 if (s->flags & SLAB_DEBUG_FREE)
5130 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005131 if (!(s->flags & SLAB_NOTRACK))
5132 *p++ = 't';
Christoph Lameter81819f02007-05-06 14:49:36 -07005133 if (p != name + 1)
5134 *p++ = '-';
5135 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005136
5137#ifdef CONFIG_MEMCG_KMEM
5138 if (!is_root_cache(s))
5139 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5140#endif
5141
Christoph Lameter81819f02007-05-06 14:49:36 -07005142 BUG_ON(p > name + ID_STR_LENGTH - 1);
5143 return name;
5144}
5145
5146static int sysfs_slab_add(struct kmem_cache *s)
5147{
5148 int err;
5149 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005150 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005151
Christoph Lameter81819f02007-05-06 14:49:36 -07005152 if (unmergeable) {
5153 /*
5154 * Slabcache can never be merged so we can use the name proper.
5155 * This is typically the case for debug situations. In that
5156 * case we can catch duplicate names easily.
5157 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005158 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005159 name = s->name;
5160 } else {
5161 /*
5162 * Create a unique name for the slab as a target
5163 * for the symlinks.
5164 */
5165 name = create_unique_id(s);
5166 }
5167
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005168 s->kobj.kset = slab_kset;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005169 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5170 if (err) {
5171 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005172 return err;
Greg Kroah-Hartman1eada112007-12-17 23:05:35 -07005173 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005174
5175 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005176 if (err) {
5177 kobject_del(&s->kobj);
5178 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005179 return err;
Xiaotian Feng5788d8a2009-07-22 11:28:53 +08005180 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005181 kobject_uevent(&s->kobj, KOBJ_ADD);
5182 if (!unmergeable) {
5183 /* Setup first alias */
5184 sysfs_slab_alias(s, s->name);
5185 kfree(name);
5186 }
5187 return 0;
5188}
5189
5190static void sysfs_slab_remove(struct kmem_cache *s)
5191{
Christoph Lameter97d06602012-07-06 15:25:11 -05005192 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005193 /*
5194 * Sysfs has not been setup yet so no need to remove the
5195 * cache from sysfs.
5196 */
5197 return;
5198
Christoph Lameter81819f02007-05-06 14:49:36 -07005199 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5200 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005201 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005202}
5203
5204/*
5205 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005206 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005207 */
5208struct saved_alias {
5209 struct kmem_cache *s;
5210 const char *name;
5211 struct saved_alias *next;
5212};
5213
Adrian Bunk5af328a2007-07-17 04:03:27 -07005214static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005215
5216static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5217{
5218 struct saved_alias *al;
5219
Christoph Lameter97d06602012-07-06 15:25:11 -05005220 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005221 /*
5222 * If we have a leftover link then remove it.
5223 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005224 sysfs_remove_link(&slab_kset->kobj, name);
5225 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005226 }
5227
5228 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5229 if (!al)
5230 return -ENOMEM;
5231
5232 al->s = s;
5233 al->name = name;
5234 al->next = alias_list;
5235 alias_list = al;
5236 return 0;
5237}
5238
5239static int __init slab_sysfs_init(void)
5240{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005241 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005242 int err;
5243
Christoph Lameter18004c52012-07-06 15:25:12 -05005244 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005245
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005246 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005247 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005248 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005249 printk(KERN_ERR "Cannot register slab subsystem.\n");
5250 return -ENOSYS;
5251 }
5252
Christoph Lameter97d06602012-07-06 15:25:11 -05005253 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005254
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005255 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005256 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005257 if (err)
5258 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5259 " to sysfs\n", s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005260 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005261
5262 while (alias_list) {
5263 struct saved_alias *al = alias_list;
5264
5265 alias_list = alias_list->next;
5266 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005267 if (err)
5268 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
Julia Lawall068ce412012-07-08 13:37:40 +02005269 " %s to sysfs\n", al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005270 kfree(al);
5271 }
5272
Christoph Lameter18004c52012-07-06 15:25:12 -05005273 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005274 resiliency_test();
5275 return 0;
5276}
5277
5278__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005279#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005280
5281/*
5282 * The /proc/slabinfo ABI
5283 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005284#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005285void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005286{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005287 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005288 unsigned long nr_objs = 0;
5289 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005290 int node;
5291
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005292 for_each_online_node(node) {
5293 struct kmem_cache_node *n = get_node(s, node);
5294
5295 if (!n)
5296 continue;
5297
Wanpeng Lic17fd132013-07-04 08:33:26 +08005298 nr_slabs += node_nr_slabs(n);
5299 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005300 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005301 }
5302
Glauber Costa0d7561c2012-10-19 18:20:27 +04005303 sinfo->active_objs = nr_objs - nr_free;
5304 sinfo->num_objs = nr_objs;
5305 sinfo->active_slabs = nr_slabs;
5306 sinfo->num_slabs = nr_slabs;
5307 sinfo->objects_per_slab = oo_objects(s->oo);
5308 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005309}
5310
Glauber Costa0d7561c2012-10-19 18:20:27 +04005311void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005312{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005313}
5314
Glauber Costab7454ad2012-10-19 18:20:25 +04005315ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5316 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005317{
Glauber Costab7454ad2012-10-19 18:20:25 +04005318 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005319}
Linus Torvalds158a9622008-01-02 13:04:48 -08005320#endif /* CONFIG_SLABINFO */