blob: 9cde663bbb1096fe544fb710ed3287dc33666e68 [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>
Andrey Ryabinina79316c2015-02-13 14:39:38 -080023#include <linux/kasan.h>
Vegard Nossum5a896d92008-04-04 00:54:48 +020024#include <linux/kmemcheck.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070025#include <linux/cpu.h>
26#include <linux/cpuset.h>
27#include <linux/mempolicy.h>
28#include <linux/ctype.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -070029#include <linux/debugobjects.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070030#include <linux/kallsyms.h>
Yasunori Gotob9049e22007-10-21 16:41:37 -070031#include <linux/memory.h>
Roman Zippelf8bd2252008-05-01 04:34:31 -070032#include <linux/math64.h>
Akinobu Mita773ff602008-12-23 19:37:01 +090033#include <linux/fault-inject.h>
Pekka Enbergbfa71452011-07-07 22:47:01 +030034#include <linux/stacktrace.h>
Christoph Lameter4de900b2012-01-30 15:53:51 -060035#include <linux/prefetch.h>
Glauber Costa2633d7a2012-12-18 14:22:34 -080036#include <linux/memcontrol.h>
Christoph Lameter81819f02007-05-06 14:49:36 -070037
Richard Kennedy4a923792010-10-21 10:29:19 +010038#include <trace/events/kmem.h>
39
Mel Gorman072bb0a2012-07-31 16:43:58 -070040#include "internal.h"
41
Christoph Lameter81819f02007-05-06 14:49:36 -070042/*
43 * Lock order:
Christoph Lameter18004c52012-07-06 15:25:12 -050044 * 1. slab_mutex (Global Mutex)
Christoph Lameter881db7f2011-06-01 12:25:53 -050045 * 2. node->list_lock
46 * 3. slab_lock(page) (Only on some arches and for debugging)
Christoph Lameter81819f02007-05-06 14:49:36 -070047 *
Christoph Lameter18004c52012-07-06 15:25:12 -050048 * slab_mutex
Christoph Lameter881db7f2011-06-01 12:25:53 -050049 *
Christoph Lameter18004c52012-07-06 15:25:12 -050050 * The role of the slab_mutex is to protect the list of all the slabs
Christoph Lameter881db7f2011-06-01 12:25:53 -050051 * and to synchronize major metadata changes to slab cache structures.
52 *
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
59 *
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -070065 *
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
71 *
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
76 * the list lock.
Christoph Lameter81819f02007-05-06 14:49:36 -070077 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
81 *
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
84 *
Christoph Lameter672bba32007-05-09 02:32:39 -070085 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
Christoph Lameter81819f02007-05-06 14:49:36 -070087 * freed then the slab will show up again on the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -070088 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
Christoph Lameter81819f02007-05-06 14:49:36 -070090 *
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
94 *
95 * Overloading of page flags that are otherwise used for LRU management.
96 *
Christoph Lameter4b6f0752007-05-16 22:10:53 -070097 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
105 *
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
Christoph Lameterdfb4f092007-10-16 01:26:05 -0700109 * freelist that allows lockless access to
Christoph Lameter894b8782007-05-10 03:15:16 -0700110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
Christoph Lameter81819f02007-05-06 14:49:36 -0700112 *
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
Christoph Lameter894b8782007-05-10 03:15:16 -0700115 * the fast path and disables lockless freelists.
Christoph Lameter81819f02007-05-06 14:49:36 -0700116 */
117
Christoph Lameteraf537b02010-07-09 14:07:14 -0500118static inline int kmem_cache_debug(struct kmem_cache *s)
119{
Christoph Lameter5577bd82007-05-16 22:10:56 -0700120#ifdef CONFIG_SLUB_DEBUG
Christoph Lameteraf537b02010-07-09 14:07:14 -0500121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
Christoph Lameter5577bd82007-05-16 22:10:56 -0700122#else
Christoph Lameteraf537b02010-07-09 14:07:14 -0500123 return 0;
Christoph Lameter5577bd82007-05-16 22:10:56 -0700124#endif
Christoph Lameteraf537b02010-07-09 14:07:14 -0500125}
Christoph Lameter5577bd82007-05-16 22:10:56 -0700126
Joonsoo Kim345c9052013-06-19 14:05:52 +0900127static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128{
129#ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
131#else
132 return false;
133#endif
134}
135
Christoph Lameter81819f02007-05-06 14:49:36 -0700136/*
137 * Issues still to be resolved:
138 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700141 * - Variable sizing of the per node arrays
142 */
143
144/* Enable to test recovery from slab corruption on boot */
145#undef SLUB_RESILIENCY_TEST
146
Christoph Lameterb789ef52011-06-01 12:25:49 -0500147/* Enable to log cmpxchg failures */
148#undef SLUB_DEBUG_CMPXCHG
149
Christoph Lameter81819f02007-05-06 14:49:36 -0700150/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800154#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700155
Christoph Lameter2086d262007-05-06 14:49:46 -0700156/*
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
Zhi Yong Wu721ae222013-11-08 20:47:37 +0800159 * sort the partial list by the number of objects in use.
Christoph Lameter2086d262007-05-06 14:49:46 -0700160 */
161#define MAX_PARTIAL 10
162
Laura Abbottbecfda62016-03-15 14:55:06 -0700163#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
Christoph Lameter81819f02007-05-06 14:49:36 -0700164 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700165
Christoph Lameter81819f02007-05-06 14:49:36 -0700166/*
David Rientjes3de47212009-07-27 18:30:35 -0700167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
169 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700170 */
David Rientjes3de47212009-07-27 18:30:35 -0700171#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700172
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400173#define OO_SHIFT 16
174#define OO_MASK ((1 << OO_SHIFT) - 1)
Christoph Lameter50d5c412011-06-01 12:25:45 -0500175#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400176
Christoph Lameter81819f02007-05-06 14:49:36 -0700177/* Internal SLUB flags */
Christoph Lameterf90ec392010-07-09 14:07:11 -0500178#define __OBJECT_POISON 0x80000000UL /* Poison object */
Christoph Lameterb789ef52011-06-01 12:25:49 -0500179#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
Christoph Lameter81819f02007-05-06 14:49:36 -0700180
Christoph Lameter81819f02007-05-06 14:49:36 -0700181#ifdef CONFIG_SMP
182static struct notifier_block slab_notifier;
183#endif
184
Christoph Lameter02cbc872007-05-09 02:32:43 -0700185/*
186 * Tracking user of a slab.
187 */
Ben Greeard6543e32011-07-07 11:36:36 -0700188#define TRACK_ADDRS_COUNT 16
Christoph Lameter02cbc872007-05-09 02:32:43 -0700189struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300190 unsigned long addr; /* Called from address */
Ben Greeard6543e32011-07-07 11:36:36 -0700191#ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193#endif
Christoph Lameter02cbc872007-05-09 02:32:43 -0700194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
197};
198
199enum track_item { TRACK_ALLOC, TRACK_FREE };
200
Christoph Lameterab4d5ed2010-10-05 13:57:26 -0500201#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -0700202static int sysfs_slab_add(struct kmem_cache *);
203static int sysfs_slab_alias(struct kmem_cache *, const char *);
Glauber Costa107dab52012-12-18 14:23:05 -0800204static void memcg_propagate_slab_attrs(struct kmem_cache *s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700205#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700206static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
208 { return 0; }
Glauber Costa107dab52012-12-18 14:23:05 -0800209static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
Christoph Lameter81819f02007-05-06 14:49:36 -0700210#endif
211
Christoph Lameter4fdccdf2011-03-22 13:35:00 -0500212static inline void stat(const struct kmem_cache *s, enum stat_item si)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800213{
214#ifdef CONFIG_SLUB_STATS
Christoph Lameter88da03a2014-04-07 15:39:42 -0700215 /*
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
218 */
219 raw_cpu_inc(s->cpu_slab->stat[si]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800220#endif
221}
222
Christoph Lameter81819f02007-05-06 14:49:36 -0700223/********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
226
Christoph Lameter6446faa2008-02-15 23:45:26 -0800227/* Verify that a pointer has an address that is valid within a slab page */
Christoph Lameter02cbc872007-05-09 02:32:43 -0700228static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
230{
231 void *base;
232
Christoph Lametera973e9d2008-03-01 13:40:44 -0800233 if (!object)
Christoph Lameter02cbc872007-05-09 02:32:43 -0700234 return 1;
235
Christoph Lametera973e9d2008-03-01 13:40:44 -0800236 base = page_address(page);
Christoph Lameter39b26462008-04-14 19:11:30 +0300237 if (object < base || object >= base + page->objects * s->size ||
Christoph Lameter02cbc872007-05-09 02:32:43 -0700238 (object - base) % s->size) {
239 return 0;
240 }
241
242 return 1;
243}
244
Christoph Lameter7656c722007-05-09 02:32:40 -0700245static inline void *get_freepointer(struct kmem_cache *s, void *object)
246{
247 return *(void **)(object + s->offset);
248}
249
Eric Dumazet0ad95002011-12-16 16:25:34 +0100250static void prefetch_freepointer(const struct kmem_cache *s, void *object)
251{
252 prefetch(object + s->offset);
253}
254
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500255static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
256{
257 void *p;
258
259#ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261#else
262 p = get_freepointer(s, object);
263#endif
264 return p;
265}
266
Christoph Lameter7656c722007-05-09 02:32:40 -0700267static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
268{
269 *(void **)(object + s->offset) = fp;
270}
271
272/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300273#define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
Christoph Lameter7656c722007-05-09 02:32:40 -0700275 __p += (__s)->size)
276
Wei Yang54266642014-08-06 16:04:42 -0700277#define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
280
Christoph Lameter7656c722007-05-09 02:32:40 -0700281/* Determine object index from a given position */
282static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283{
284 return (p - addr) / s->size;
285}
286
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800287static inline int order_objects(int order, unsigned long size, int reserved)
288{
289 return ((PAGE_SIZE << order) - reserved) / size;
290}
291
Christoph Lameter834f3d12008-04-14 19:11:31 +0300292static inline struct kmem_cache_order_objects oo_make(int order,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800293 unsigned long size, int reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300294{
295 struct kmem_cache_order_objects x = {
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800296 (order << OO_SHIFT) + order_objects(order, size, reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300297 };
298
299 return x;
300}
301
302static inline int oo_order(struct kmem_cache_order_objects x)
303{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400304 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300305}
306
307static inline int oo_objects(struct kmem_cache_order_objects x)
308{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400309 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300310}
311
Christoph Lameter881db7f2011-06-01 12:25:53 -0500312/*
313 * Per slab locking using the pagelock
314 */
315static __always_inline void slab_lock(struct page *page)
316{
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -0800317 VM_BUG_ON_PAGE(PageTail(page), page);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500318 bit_spin_lock(PG_locked, &page->flags);
319}
320
321static __always_inline void slab_unlock(struct page *page)
322{
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -0800323 VM_BUG_ON_PAGE(PageTail(page), page);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500324 __bit_spin_unlock(PG_locked, &page->flags);
325}
326
Dave Hansena0320862014-01-30 15:46:09 -0800327static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
328{
329 struct page tmp;
330 tmp.counters = counters_new;
331 /*
332 * page->counters can cover frozen/inuse/objects as well
333 * as page->_count. If we assign to ->counters directly
334 * we run the risk of losing updates to page->_count, so
335 * be careful and only assign to the fields we need.
336 */
337 page->frozen = tmp.frozen;
338 page->inuse = tmp.inuse;
339 page->objects = tmp.objects;
340}
341
Christoph Lameter1d071712011-07-14 12:49:12 -0500342/* Interrupts must be disabled (for the fallback code to work right) */
343static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500344 void *freelist_old, unsigned long counters_old,
345 void *freelist_new, unsigned long counters_new,
346 const char *n)
347{
Christoph Lameter1d071712011-07-14 12:49:12 -0500348 VM_BUG_ON(!irqs_disabled());
Heiko Carstens25654092012-01-12 17:17:33 -0800349#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
350 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -0500351 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000352 if (cmpxchg_double(&page->freelist, &page->counters,
Dan Carpenter0aa9a132014-08-06 16:04:48 -0700353 freelist_old, counters_old,
354 freelist_new, counters_new))
Joe Perches6f6528a2015-04-14 15:44:31 -0700355 return true;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500356 } else
357#endif
358 {
Christoph Lameter881db7f2011-06-01 12:25:53 -0500359 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800360 if (page->freelist == freelist_old &&
361 page->counters == counters_old) {
Christoph Lameterb789ef52011-06-01 12:25:49 -0500362 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800363 set_page_slub_counters(page, counters_new);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500364 slab_unlock(page);
Joe Perches6f6528a2015-04-14 15:44:31 -0700365 return true;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500366 }
Christoph Lameter881db7f2011-06-01 12:25:53 -0500367 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500368 }
369
370 cpu_relax();
371 stat(s, CMPXCHG_DOUBLE_FAIL);
372
373#ifdef SLUB_DEBUG_CMPXCHG
Fabian Frederickf9f58282014-06-04 16:06:34 -0700374 pr_info("%s %s: cmpxchg double redo ", n, s->name);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500375#endif
376
Joe Perches6f6528a2015-04-14 15:44:31 -0700377 return false;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500378}
379
Christoph Lameter1d071712011-07-14 12:49:12 -0500380static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
381 void *freelist_old, unsigned long counters_old,
382 void *freelist_new, unsigned long counters_new,
383 const char *n)
384{
Heiko Carstens25654092012-01-12 17:17:33 -0800385#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
386 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameter1d071712011-07-14 12:49:12 -0500387 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000388 if (cmpxchg_double(&page->freelist, &page->counters,
Dan Carpenter0aa9a132014-08-06 16:04:48 -0700389 freelist_old, counters_old,
390 freelist_new, counters_new))
Joe Perches6f6528a2015-04-14 15:44:31 -0700391 return true;
Christoph Lameter1d071712011-07-14 12:49:12 -0500392 } else
393#endif
394 {
395 unsigned long flags;
396
397 local_irq_save(flags);
398 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800399 if (page->freelist == freelist_old &&
400 page->counters == counters_old) {
Christoph Lameter1d071712011-07-14 12:49:12 -0500401 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800402 set_page_slub_counters(page, counters_new);
Christoph Lameter1d071712011-07-14 12:49:12 -0500403 slab_unlock(page);
404 local_irq_restore(flags);
Joe Perches6f6528a2015-04-14 15:44:31 -0700405 return true;
Christoph Lameter1d071712011-07-14 12:49:12 -0500406 }
407 slab_unlock(page);
408 local_irq_restore(flags);
409 }
410
411 cpu_relax();
412 stat(s, CMPXCHG_DOUBLE_FAIL);
413
414#ifdef SLUB_DEBUG_CMPXCHG
Fabian Frederickf9f58282014-06-04 16:06:34 -0700415 pr_info("%s %s: cmpxchg double redo ", n, s->name);
Christoph Lameter1d071712011-07-14 12:49:12 -0500416#endif
417
Joe Perches6f6528a2015-04-14 15:44:31 -0700418 return false;
Christoph Lameter1d071712011-07-14 12:49:12 -0500419}
420
Christoph Lameter41ecc552007-05-09 02:32:44 -0700421#ifdef CONFIG_SLUB_DEBUG
422/*
Christoph Lameter5f80b132011-04-15 14:48:13 -0500423 * Determine a map of object in use on a page.
424 *
Christoph Lameter881db7f2011-06-01 12:25:53 -0500425 * Node listlock must be held to guarantee that the page does
Christoph Lameter5f80b132011-04-15 14:48:13 -0500426 * not vanish from under us.
427 */
428static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
429{
430 void *p;
431 void *addr = page_address(page);
432
433 for (p = page->freelist; p; p = get_freepointer(s, p))
434 set_bit(slab_index(p, s, addr), map);
435}
436
Christoph Lameter41ecc552007-05-09 02:32:44 -0700437/*
438 * Debug settings:
439 */
Andrey Ryabinin89d3c872015-11-05 18:51:23 -0800440#if defined(CONFIG_SLUB_DEBUG_ON)
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700441static int slub_debug = DEBUG_DEFAULT_FLAGS;
Andrey Ryabinin89d3c872015-11-05 18:51:23 -0800442#elif defined(CONFIG_KASAN)
443static int slub_debug = SLAB_STORE_USER;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700444#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700445static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700446#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700447
448static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700449static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700450
Christoph Lameter7656c722007-05-09 02:32:40 -0700451/*
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800452 * slub is about to manipulate internal object metadata. This memory lies
453 * outside the range of the allocated object, so accessing it would normally
454 * be reported by kasan as a bounds error. metadata_access_enable() is used
455 * to tell kasan that these accesses are OK.
456 */
457static inline void metadata_access_enable(void)
458{
459 kasan_disable_current();
460}
461
462static inline void metadata_access_disable(void)
463{
464 kasan_enable_current();
465}
466
467/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700468 * Object debugging
469 */
470static void print_section(char *text, u8 *addr, unsigned int length)
471{
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800472 metadata_access_enable();
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200473 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
474 length, 1);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800475 metadata_access_disable();
Christoph Lameter81819f02007-05-06 14:49:36 -0700476}
477
Christoph Lameter81819f02007-05-06 14:49:36 -0700478static struct track *get_track(struct kmem_cache *s, void *object,
479 enum track_item alloc)
480{
481 struct track *p;
482
483 if (s->offset)
484 p = object + s->offset + sizeof(void *);
485 else
486 p = object + s->inuse;
487
488 return p + alloc;
489}
490
491static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300492 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700493{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900494 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700495
Christoph Lameter81819f02007-05-06 14:49:36 -0700496 if (addr) {
Ben Greeard6543e32011-07-07 11:36:36 -0700497#ifdef CONFIG_STACKTRACE
498 struct stack_trace trace;
499 int i;
500
501 trace.nr_entries = 0;
502 trace.max_entries = TRACK_ADDRS_COUNT;
503 trace.entries = p->addrs;
504 trace.skip = 3;
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800505 metadata_access_enable();
Ben Greeard6543e32011-07-07 11:36:36 -0700506 save_stack_trace(&trace);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800507 metadata_access_disable();
Ben Greeard6543e32011-07-07 11:36:36 -0700508
509 /* See rant in lockdep.c */
510 if (trace.nr_entries != 0 &&
511 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
512 trace.nr_entries--;
513
514 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
515 p->addrs[i] = 0;
516#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700517 p->addr = addr;
518 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400519 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700520 p->when = jiffies;
521 } else
522 memset(p, 0, sizeof(struct track));
523}
524
Christoph Lameter81819f02007-05-06 14:49:36 -0700525static void init_tracking(struct kmem_cache *s, void *object)
526{
Christoph Lameter24922682007-07-17 04:03:18 -0700527 if (!(s->flags & SLAB_STORE_USER))
528 return;
529
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300530 set_track(s, object, TRACK_FREE, 0UL);
531 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700532}
533
534static void print_track(const char *s, struct track *t)
535{
536 if (!t->addr)
537 return;
538
Fabian Frederickf9f58282014-06-04 16:06:34 -0700539 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
540 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Ben Greeard6543e32011-07-07 11:36:36 -0700541#ifdef CONFIG_STACKTRACE
542 {
543 int i;
544 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
545 if (t->addrs[i])
Fabian Frederickf9f58282014-06-04 16:06:34 -0700546 pr_err("\t%pS\n", (void *)t->addrs[i]);
Ben Greeard6543e32011-07-07 11:36:36 -0700547 else
548 break;
549 }
550#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700551}
552
Christoph Lameter24922682007-07-17 04:03:18 -0700553static void print_tracking(struct kmem_cache *s, void *object)
554{
555 if (!(s->flags & SLAB_STORE_USER))
556 return;
557
558 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
559 print_track("Freed", get_track(s, object, TRACK_FREE));
560}
561
562static void print_page_info(struct page *page)
563{
Fabian Frederickf9f58282014-06-04 16:06:34 -0700564 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800565 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700566
567}
568
569static void slab_bug(struct kmem_cache *s, char *fmt, ...)
570{
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700571 struct va_format vaf;
Christoph Lameter24922682007-07-17 04:03:18 -0700572 va_list args;
Christoph Lameter24922682007-07-17 04:03:18 -0700573
574 va_start(args, fmt);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700575 vaf.fmt = fmt;
576 vaf.va = &args;
Fabian Frederickf9f58282014-06-04 16:06:34 -0700577 pr_err("=============================================================================\n");
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700578 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
Fabian Frederickf9f58282014-06-04 16:06:34 -0700579 pr_err("-----------------------------------------------------------------------------\n\n");
Dave Jones645df232012-09-18 15:54:12 -0400580
Rusty Russell373d4d02013-01-21 17:17:39 +1030581 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700582 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700583}
584
585static void slab_fix(struct kmem_cache *s, char *fmt, ...)
586{
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700587 struct va_format vaf;
Christoph Lameter24922682007-07-17 04:03:18 -0700588 va_list args;
Christoph Lameter24922682007-07-17 04:03:18 -0700589
590 va_start(args, fmt);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700591 vaf.fmt = fmt;
592 vaf.va = &args;
593 pr_err("FIX %s: %pV\n", s->name, &vaf);
Christoph Lameter24922682007-07-17 04:03:18 -0700594 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700595}
596
597static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700598{
599 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800600 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700601
602 print_tracking(s, p);
603
604 print_page_info(page);
605
Fabian Frederickf9f58282014-06-04 16:06:34 -0700606 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
Christoph Lameter24922682007-07-17 04:03:18 -0700608
609 if (p > addr + 16)
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200610 print_section("Bytes b4 ", p - 16, 16);
Christoph Lameter24922682007-07-17 04:03:18 -0700611
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500612 print_section("Object ", p, min_t(unsigned long, s->object_size,
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200613 PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700614 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500615 print_section("Redzone ", p + s->object_size,
616 s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700617
Christoph Lameter81819f02007-05-06 14:49:36 -0700618 if (s->offset)
619 off = s->offset + sizeof(void *);
620 else
621 off = s->inuse;
622
Christoph Lameter24922682007-07-17 04:03:18 -0700623 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700624 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700625
626 if (off != s->size)
627 /* Beginning of the filler is the free pointer */
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200628 print_section("Padding ", p + off, s->size - off);
Christoph Lameter24922682007-07-17 04:03:18 -0700629
630 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700631}
632
Andrey Ryabinin75c66de2015-02-13 14:39:35 -0800633void object_err(struct kmem_cache *s, struct page *page,
Christoph Lameter81819f02007-05-06 14:49:36 -0700634 u8 *object, char *reason)
635{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700636 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700637 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700638}
639
Chen Gangd0e0ac92013-07-15 09:05:29 +0800640static void slab_err(struct kmem_cache *s, struct page *page,
641 const char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700642{
643 va_list args;
644 char buf[100];
645
Christoph Lameter24922682007-07-17 04:03:18 -0700646 va_start(args, fmt);
647 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700648 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700649 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700650 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700651 dump_stack();
652}
653
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500654static void init_object(struct kmem_cache *s, void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700655{
656 u8 *p = object;
657
658 if (s->flags & __OBJECT_POISON) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500659 memset(p, POISON_FREE, s->object_size - 1);
660 p[s->object_size - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700661 }
662
663 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500664 memset(p + s->object_size, val, s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700665}
666
Christoph Lameter24922682007-07-17 04:03:18 -0700667static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
668 void *from, void *to)
669{
670 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
671 memset(from, data, to - from);
672}
673
674static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
675 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800676 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700677{
678 u8 *fault;
679 u8 *end;
680
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800681 metadata_access_enable();
Akinobu Mita798248202011-10-31 17:08:07 -0700682 fault = memchr_inv(start, value, bytes);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800683 metadata_access_disable();
Christoph Lameter24922682007-07-17 04:03:18 -0700684 if (!fault)
685 return 1;
686
687 end = start + bytes;
688 while (end > fault && end[-1] == value)
689 end--;
690
691 slab_bug(s, "%s overwritten", what);
Fabian Frederickf9f58282014-06-04 16:06:34 -0700692 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
Christoph Lameter24922682007-07-17 04:03:18 -0700693 fault, end - 1, fault[0], value);
694 print_trailer(s, page, object);
695
696 restore_bytes(s, what, value, fault, end);
697 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700698}
699
Christoph Lameter81819f02007-05-06 14:49:36 -0700700/*
701 * Object layout:
702 *
703 * object address
704 * Bytes of the object to be managed.
705 * If the freepointer may overlay the object then the free
706 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700707 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700708 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
709 * 0xa5 (POISON_END)
710 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500711 * object + s->object_size
Christoph Lameter81819f02007-05-06 14:49:36 -0700712 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700713 * Padding is extended by another word if Redzoning is enabled and
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500714 * object_size == inuse.
Christoph Lameter672bba32007-05-09 02:32:39 -0700715 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700716 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
717 * 0xcc (RED_ACTIVE) for objects in use.
718 *
719 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700720 * Meta data starts here.
721 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700722 * A. Free pointer (if we cannot overwrite object on free)
723 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700724 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800725 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700726 * before the word boundary.
727 *
728 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700729 *
730 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700731 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700732 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500733 * If slabcaches are merged then the object_size and inuse boundaries are mostly
Christoph Lameter672bba32007-05-09 02:32:39 -0700734 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700735 * may be used with merged slabcaches.
736 */
737
Christoph Lameter81819f02007-05-06 14:49:36 -0700738static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
739{
740 unsigned long off = s->inuse; /* The end of info */
741
742 if (s->offset)
743 /* Freepointer is placed after the object. */
744 off += sizeof(void *);
745
746 if (s->flags & SLAB_STORE_USER)
747 /* We also have user information there */
748 off += 2 * sizeof(struct track);
749
750 if (s->size == off)
751 return 1;
752
Christoph Lameter24922682007-07-17 04:03:18 -0700753 return check_bytes_and_report(s, page, p, "Object padding",
754 p + off, POISON_INUSE, s->size - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700755}
756
Christoph Lameter39b26462008-04-14 19:11:30 +0300757/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700758static int slab_pad_check(struct kmem_cache *s, struct page *page)
759{
Christoph Lameter24922682007-07-17 04:03:18 -0700760 u8 *start;
761 u8 *fault;
762 u8 *end;
763 int length;
764 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700765
766 if (!(s->flags & SLAB_POISON))
767 return 1;
768
Christoph Lametera973e9d2008-03-01 13:40:44 -0800769 start = page_address(page);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800770 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
Christoph Lameter39b26462008-04-14 19:11:30 +0300771 end = start + length;
772 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700773 if (!remainder)
774 return 1;
775
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800776 metadata_access_enable();
Akinobu Mita798248202011-10-31 17:08:07 -0700777 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800778 metadata_access_disable();
Christoph Lameter24922682007-07-17 04:03:18 -0700779 if (!fault)
780 return 1;
781 while (end > fault && end[-1] == POISON_INUSE)
782 end--;
783
784 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200785 print_section("Padding ", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700786
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200787 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700788 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700789}
790
791static int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500792 void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700793{
794 u8 *p = object;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500795 u8 *endobject = object + s->object_size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700796
797 if (s->flags & SLAB_RED_ZONE) {
Christoph Lameter24922682007-07-17 04:03:18 -0700798 if (!check_bytes_and_report(s, page, object, "Redzone",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500799 endobject, val, s->inuse - s->object_size))
Christoph Lameter81819f02007-05-06 14:49:36 -0700800 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700801 } else {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500802 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800803 check_bytes_and_report(s, page, p, "Alignment padding",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800804 endobject, POISON_INUSE,
805 s->inuse - s->object_size);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800806 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700807 }
808
809 if (s->flags & SLAB_POISON) {
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500810 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700811 (!check_bytes_and_report(s, page, p, "Poison", p,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500812 POISON_FREE, s->object_size - 1) ||
Christoph Lameter24922682007-07-17 04:03:18 -0700813 !check_bytes_and_report(s, page, p, "Poison",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500814 p + s->object_size - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700815 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700816 /*
817 * check_pad_bytes cleans up on its own.
818 */
819 check_pad_bytes(s, page, p);
820 }
821
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500822 if (!s->offset && val == SLUB_RED_ACTIVE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700823 /*
824 * Object and freepointer overlap. Cannot check
825 * freepointer while object is allocated.
826 */
827 return 1;
828
829 /* Check free pointer validity */
830 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
831 object_err(s, page, p, "Freepointer corrupt");
832 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100833 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700834 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700835 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700836 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800837 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700838 return 0;
839 }
840 return 1;
841}
842
843static int check_slab(struct kmem_cache *s, struct page *page)
844{
Christoph Lameter39b26462008-04-14 19:11:30 +0300845 int maxobj;
846
Christoph Lameter81819f02007-05-06 14:49:36 -0700847 VM_BUG_ON(!irqs_disabled());
848
849 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700850 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700851 return 0;
852 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300853
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800854 maxobj = order_objects(compound_order(page), s->size, s->reserved);
Christoph Lameter39b26462008-04-14 19:11:30 +0300855 if (page->objects > maxobj) {
856 slab_err(s, page, "objects %u > max %u",
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800857 page->objects, maxobj);
Christoph Lameter39b26462008-04-14 19:11:30 +0300858 return 0;
859 }
860 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700861 slab_err(s, page, "inuse %u > max %u",
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800862 page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700863 return 0;
864 }
865 /* Slab_pad_check fixes things up after itself */
866 slab_pad_check(s, page);
867 return 1;
868}
869
870/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700871 * Determine if a certain object on a page is on the freelist. Must hold the
872 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700873 */
874static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
875{
876 int nr = 0;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500877 void *fp;
Christoph Lameter81819f02007-05-06 14:49:36 -0700878 void *object = NULL;
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800879 int max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700880
Christoph Lameter881db7f2011-06-01 12:25:53 -0500881 fp = page->freelist;
Christoph Lameter39b26462008-04-14 19:11:30 +0300882 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700883 if (fp == search)
884 return 1;
885 if (!check_valid_pointer(s, page, fp)) {
886 if (object) {
887 object_err(s, page, object,
888 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800889 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700890 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700891 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800892 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300893 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700894 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700895 return 0;
896 }
897 break;
898 }
899 object = fp;
900 fp = get_freepointer(s, object);
901 nr++;
902 }
903
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800904 max_objects = order_objects(compound_order(page), s->size, s->reserved);
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400905 if (max_objects > MAX_OBJS_PER_PAGE)
906 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300907
908 if (page->objects != max_objects) {
909 slab_err(s, page, "Wrong number of objects. Found %d but "
910 "should be %d", page->objects, max_objects);
911 page->objects = max_objects;
912 slab_fix(s, "Number of objects adjusted.");
913 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300914 if (page->inuse != page->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700915 slab_err(s, page, "Wrong object count. Counter is %d but "
Christoph Lameter39b26462008-04-14 19:11:30 +0300916 "counted were %d", page->inuse, page->objects - nr);
917 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700918 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700919 }
920 return search == NULL;
921}
922
Christoph Lameter0121c6192008-04-29 16:11:12 -0700923static void trace(struct kmem_cache *s, struct page *page, void *object,
924 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700925{
926 if (s->flags & SLAB_TRACE) {
Fabian Frederickf9f58282014-06-04 16:06:34 -0700927 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
Christoph Lameter3ec09742007-05-16 22:11:00 -0700928 s->name,
929 alloc ? "alloc" : "free",
930 object, page->inuse,
931 page->freelist);
932
933 if (!alloc)
Chen Gangd0e0ac92013-07-15 09:05:29 +0800934 print_section("Object ", (void *)object,
935 s->object_size);
Christoph Lameter3ec09742007-05-16 22:11:00 -0700936
937 dump_stack();
938 }
939}
940
Christoph Lameter643b1132007-05-06 14:49:42 -0700941/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700942 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700943 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500944static void add_full(struct kmem_cache *s,
945 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700946{
Christoph Lameter643b1132007-05-06 14:49:42 -0700947 if (!(s->flags & SLAB_STORE_USER))
948 return;
949
David Rientjes255d0882014-02-10 14:25:39 -0800950 lockdep_assert_held(&n->list_lock);
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500951 list_add(&page->lru, &n->full);
952}
Christoph Lameter643b1132007-05-06 14:49:42 -0700953
Peter Zijlstrac65c1872014-01-10 13:23:49 +0100954static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500955{
956 if (!(s->flags & SLAB_STORE_USER))
957 return;
958
David Rientjes255d0882014-02-10 14:25:39 -0800959 lockdep_assert_held(&n->list_lock);
Christoph Lameter643b1132007-05-06 14:49:42 -0700960 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -0700961}
962
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300963/* Tracking of the number of slabs for debugging purposes */
964static inline unsigned long slabs_node(struct kmem_cache *s, int node)
965{
966 struct kmem_cache_node *n = get_node(s, node);
967
968 return atomic_long_read(&n->nr_slabs);
969}
970
Alexander Beregalov26c02cf2009-06-11 14:08:48 +0400971static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
972{
973 return atomic_long_read(&n->nr_slabs);
974}
975
Christoph Lameter205ab992008-04-14 19:11:40 +0300976static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300977{
978 struct kmem_cache_node *n = get_node(s, node);
979
980 /*
981 * May be called early in order to allocate a slab for the
982 * kmem_cache_node structure. Solve the chicken-egg
983 * dilemma by deferring the increment of the count during
984 * bootstrap (see early_kmem_cache_node_alloc).
985 */
Joonsoo Kim338b2642013-01-21 17:01:27 +0900986 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300987 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300988 atomic_long_add(objects, &n->total_objects);
989 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300990}
Christoph Lameter205ab992008-04-14 19:11:40 +0300991static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300992{
993 struct kmem_cache_node *n = get_node(s, node);
994
995 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +0300996 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +0300997}
998
999/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001000static void setup_object_debug(struct kmem_cache *s, struct page *page,
1001 void *object)
1002{
1003 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1004 return;
1005
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001006 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001007 init_tracking(s, object);
1008}
1009
Laura Abbottbecfda62016-03-15 14:55:06 -07001010static inline int alloc_consistency_checks(struct kmem_cache *s,
Chen Gangd0e0ac92013-07-15 09:05:29 +08001011 struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001012 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001013{
1014 if (!check_slab(s, page))
Laura Abbottbecfda62016-03-15 14:55:06 -07001015 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001016
Christoph Lameter81819f02007-05-06 14:49:36 -07001017 if (!check_valid_pointer(s, page, object)) {
1018 object_err(s, page, object, "Freelist Pointer check fails");
Laura Abbottbecfda62016-03-15 14:55:06 -07001019 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001020 }
1021
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001022 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Laura Abbottbecfda62016-03-15 14:55:06 -07001023 return 0;
1024
1025 return 1;
1026}
1027
1028static noinline int alloc_debug_processing(struct kmem_cache *s,
1029 struct page *page,
1030 void *object, unsigned long addr)
1031{
1032 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1033 if (!alloc_consistency_checks(s, page, object, addr))
1034 goto bad;
1035 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001036
Christoph Lameter3ec09742007-05-16 22:11:00 -07001037 /* Success perform special debug activities for allocs */
1038 if (s->flags & SLAB_STORE_USER)
1039 set_track(s, object, TRACK_ALLOC, addr);
1040 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001041 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001042 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001043
Christoph Lameter81819f02007-05-06 14:49:36 -07001044bad:
1045 if (PageSlab(page)) {
1046 /*
1047 * If this is a slab page then lets do the best we can
1048 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001049 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001050 */
Christoph Lameter24922682007-07-17 04:03:18 -07001051 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001052 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001053 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001054 }
1055 return 0;
1056}
1057
Laura Abbottbecfda62016-03-15 14:55:06 -07001058static inline int free_consistency_checks(struct kmem_cache *s,
1059 struct page *page, void *object, unsigned long addr)
1060{
1061 if (!check_valid_pointer(s, page, object)) {
1062 slab_err(s, page, "Invalid object pointer 0x%p", object);
1063 return 0;
1064 }
1065
1066 if (on_freelist(s, page, object)) {
1067 object_err(s, page, object, "Object already free");
1068 return 0;
1069 }
1070
1071 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1072 return 0;
1073
1074 if (unlikely(s != page->slab_cache)) {
1075 if (!PageSlab(page)) {
1076 slab_err(s, page, "Attempt to free object(0x%p) "
1077 "outside of slab", object);
1078 } else if (!page->slab_cache) {
1079 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1080 object);
1081 dump_stack();
1082 } else
1083 object_err(s, page, object,
1084 "page slab pointer corrupt.");
1085 return 0;
1086 }
1087 return 1;
1088}
1089
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001090/* Supports checking bulk free of a constructed freelist */
Laura Abbott282acb42016-03-15 14:54:59 -07001091static noinline int free_debug_processing(
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001092 struct kmem_cache *s, struct page *page,
1093 void *head, void *tail, int bulk_cnt,
Laura Abbott282acb42016-03-15 14:54:59 -07001094 unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001095{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001096 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001097 void *object = head;
1098 int cnt = 0;
Laura Abbott282acb42016-03-15 14:54:59 -07001099 unsigned long uninitialized_var(flags);
Laura Abbott804aa132016-03-15 14:55:02 -07001100 int ret = 0;
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001101
Laura Abbott282acb42016-03-15 14:54:59 -07001102 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001103 slab_lock(page);
1104
Laura Abbottbecfda62016-03-15 14:55:06 -07001105 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1106 if (!check_slab(s, page))
1107 goto out;
1108 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001109
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001110next_object:
1111 cnt++;
1112
Laura Abbottbecfda62016-03-15 14:55:06 -07001113 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1114 if (!free_consistency_checks(s, page, object, addr))
1115 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001116 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001117
Christoph Lameter3ec09742007-05-16 22:11:00 -07001118 if (s->flags & SLAB_STORE_USER)
1119 set_track(s, object, TRACK_FREE, addr);
1120 trace(s, page, object, 0);
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001121 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001122 init_object(s, object, SLUB_RED_INACTIVE);
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001123
1124 /* Reached end of constructed freelist yet? */
1125 if (object != tail) {
1126 object = get_freepointer(s, object);
1127 goto next_object;
1128 }
Laura Abbott804aa132016-03-15 14:55:02 -07001129 ret = 1;
1130
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001131out:
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001132 if (cnt != bulk_cnt)
1133 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1134 bulk_cnt, cnt);
1135
Christoph Lameter881db7f2011-06-01 12:25:53 -05001136 slab_unlock(page);
Laura Abbott282acb42016-03-15 14:54:59 -07001137 spin_unlock_irqrestore(&n->list_lock, flags);
Laura Abbott804aa132016-03-15 14:55:02 -07001138 if (!ret)
1139 slab_fix(s, "Object at 0x%p not freed", object);
1140 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07001141}
1142
Christoph Lameter41ecc552007-05-09 02:32:44 -07001143static int __init setup_slub_debug(char *str)
1144{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001145 slub_debug = DEBUG_DEFAULT_FLAGS;
1146 if (*str++ != '=' || !*str)
1147 /*
1148 * No options specified. Switch on full debugging.
1149 */
1150 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001151
1152 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001153 /*
1154 * No options but restriction on slabs. This means full
1155 * debugging for slabs matching a pattern.
1156 */
1157 goto check_slabs;
1158
1159 slub_debug = 0;
1160 if (*str == '-')
1161 /*
1162 * Switch off all debugging measures.
1163 */
1164 goto out;
1165
1166 /*
1167 * Determine which debug features should be switched on
1168 */
Pekka Enberg06428782008-01-07 23:20:27 -08001169 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001170 switch (tolower(*str)) {
1171 case 'f':
Laura Abbottbecfda62016-03-15 14:55:06 -07001172 slub_debug |= SLAB_CONSISTENCY_CHECKS;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001173 break;
1174 case 'z':
1175 slub_debug |= SLAB_RED_ZONE;
1176 break;
1177 case 'p':
1178 slub_debug |= SLAB_POISON;
1179 break;
1180 case 'u':
1181 slub_debug |= SLAB_STORE_USER;
1182 break;
1183 case 't':
1184 slub_debug |= SLAB_TRACE;
1185 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001186 case 'a':
1187 slub_debug |= SLAB_FAILSLAB;
1188 break;
Chris J Arges08303a72015-04-14 15:44:25 -07001189 case 'o':
1190 /*
1191 * Avoid enabling debugging on caches if its minimum
1192 * order would increase as a result.
1193 */
1194 disable_higher_order_debug = 1;
1195 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001196 default:
Fabian Frederickf9f58282014-06-04 16:06:34 -07001197 pr_err("slub_debug option '%c' unknown. skipped\n",
1198 *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001199 }
1200 }
1201
1202check_slabs:
1203 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001204 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001205out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001206 return 1;
1207}
1208
1209__setup("slub_debug", setup_slub_debug);
1210
Joonsoo Kim423c9292014-10-09 15:26:22 -07001211unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001212 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001213 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001214{
1215 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001216 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001217 */
Christoph Lameterc6f58d92013-11-07 16:29:15 +00001218 if (slub_debug && (!slub_debug_slabs || (name &&
1219 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
David Rientjes3de47212009-07-27 18:30:35 -07001220 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001221
1222 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001223}
Jesper Dangaard Brouerb4a64712015-11-20 15:57:41 -08001224#else /* !CONFIG_SLUB_DEBUG */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001225static inline void setup_object_debug(struct kmem_cache *s,
1226 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001227
Christoph Lameter3ec09742007-05-16 22:11:00 -07001228static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001229 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001230
Laura Abbott282acb42016-03-15 14:54:59 -07001231static inline int free_debug_processing(
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001232 struct kmem_cache *s, struct page *page,
1233 void *head, void *tail, int bulk_cnt,
Laura Abbott282acb42016-03-15 14:54:59 -07001234 unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001235
Christoph Lameter41ecc552007-05-09 02:32:44 -07001236static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1237 { return 1; }
1238static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001239 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001240static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1241 struct page *page) {}
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001242static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1243 struct page *page) {}
Joonsoo Kim423c9292014-10-09 15:26:22 -07001244unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001245 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001246 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001247{
1248 return flags;
1249}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001250#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001251
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001252#define disable_higher_order_debug 0
1253
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001254static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1255 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001256static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1257 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001258static inline void inc_slabs_node(struct kmem_cache *s, int node,
1259 int objects) {}
1260static inline void dec_slabs_node(struct kmem_cache *s, int node,
1261 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001262
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001263#endif /* CONFIG_SLUB_DEBUG */
1264
1265/*
1266 * Hooks for other subsystems that check memory allocations. In a typical
1267 * production configuration these hooks all should produce no code at all.
1268 */
Roman Bobnievd56791b2013-10-08 15:58:57 -07001269static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1270{
1271 kmemleak_alloc(ptr, size, 1, flags);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001272 kasan_kmalloc_large(ptr, size);
Roman Bobnievd56791b2013-10-08 15:58:57 -07001273}
1274
1275static inline void kfree_hook(const void *x)
1276{
1277 kmemleak_free(x);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001278 kasan_kfree_large(x);
Roman Bobnievd56791b2013-10-08 15:58:57 -07001279}
1280
Roman Bobnievd56791b2013-10-08 15:58:57 -07001281static inline void slab_free_hook(struct kmem_cache *s, void *x)
1282{
1283 kmemleak_free_recursive(x, s->flags);
Christoph Lameter7d550c52010-08-25 14:07:16 -05001284
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001285 /*
1286 * Trouble is that we may no longer disable interrupts in the fast path
1287 * So in order to make the debug calls that expect irqs to be
1288 * disabled we need to disable interrupts temporarily.
1289 */
1290#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1291 {
1292 unsigned long flags;
1293
1294 local_irq_save(flags);
1295 kmemcheck_slab_free(s, x, s->object_size);
1296 debug_check_no_locks_freed(x, s->object_size);
1297 local_irq_restore(flags);
1298 }
1299#endif
1300 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1301 debug_check_no_obj_freed(x, s->object_size);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001302
1303 kasan_slab_free(s, x);
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001304}
Christoph Lameter205ab992008-04-14 19:11:40 +03001305
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001306static inline void slab_free_freelist_hook(struct kmem_cache *s,
1307 void *head, void *tail)
1308{
1309/*
1310 * Compiler cannot detect this function can be removed if slab_free_hook()
1311 * evaluates to nothing. Thus, catch all relevant config debug options here.
1312 */
1313#if defined(CONFIG_KMEMCHECK) || \
1314 defined(CONFIG_LOCKDEP) || \
1315 defined(CONFIG_DEBUG_KMEMLEAK) || \
1316 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1317 defined(CONFIG_KASAN)
1318
1319 void *object = head;
1320 void *tail_obj = tail ? : head;
1321
1322 do {
1323 slab_free_hook(s, object);
1324 } while ((object != tail_obj) &&
1325 (object = get_freepointer(s, object)));
1326#endif
1327}
1328
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001329static void setup_object(struct kmem_cache *s, struct page *page,
1330 void *object)
1331{
1332 setup_object_debug(s, page, object);
1333 if (unlikely(s->ctor)) {
1334 kasan_unpoison_object_data(s, object);
1335 s->ctor(object);
1336 kasan_poison_object_data(s, object);
1337 }
1338}
1339
Christoph Lameter81819f02007-05-06 14:49:36 -07001340/*
1341 * Slab allocation and freeing
1342 */
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001343static inline struct page *alloc_slab_page(struct kmem_cache *s,
1344 gfp_t flags, int node, struct kmem_cache_order_objects oo)
Christoph Lameter65c33762008-04-14 19:11:40 +03001345{
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001346 struct page *page;
Christoph Lameter65c33762008-04-14 19:11:40 +03001347 int order = oo_order(oo);
1348
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001349 flags |= __GFP_NOTRACK;
1350
Christoph Lameter2154a332010-07-09 14:07:10 -05001351 if (node == NUMA_NO_NODE)
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001352 page = alloc_pages(flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001353 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001354 page = __alloc_pages_node(node, flags, order);
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001355
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001356 if (page && memcg_charge_slab(page, flags, order, s)) {
1357 __free_pages(page, order);
1358 page = NULL;
1359 }
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001360
1361 return page;
Christoph Lameter65c33762008-04-14 19:11:40 +03001362}
1363
Christoph Lameter81819f02007-05-06 14:49:36 -07001364static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1365{
Pekka Enberg06428782008-01-07 23:20:27 -08001366 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001367 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001368 gfp_t alloc_gfp;
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001369 void *start, *p;
1370 int idx, order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001371
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001372 flags &= gfp_allowed_mask;
1373
Mel Gormand0164ad2015-11-06 16:28:21 -08001374 if (gfpflags_allow_blocking(flags))
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001375 local_irq_enable();
1376
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001377 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001378
Pekka Enbergba522702009-06-24 21:59:51 +03001379 /*
1380 * Let the initial higher-order allocation fail under memory pressure
1381 * so we fall-back to the minimum order allocation.
1382 */
1383 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
Mel Gormand0164ad2015-11-06 16:28:21 -08001384 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1385 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
Pekka Enbergba522702009-06-24 21:59:51 +03001386
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001387 page = alloc_slab_page(s, alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001388 if (unlikely(!page)) {
1389 oo = s->min;
Joonsoo Kim80c3a992014-03-12 17:26:20 +09001390 alloc_gfp = flags;
Christoph Lameter65c33762008-04-14 19:11:40 +03001391 /*
1392 * Allocation may have failed due to fragmentation.
1393 * Try a lower order alloc if possible
1394 */
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001395 page = alloc_slab_page(s, alloc_gfp, node, oo);
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001396 if (unlikely(!page))
1397 goto out;
1398 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001399 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001400
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001401 if (kmemcheck_enabled &&
1402 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001403 int pages = 1 << oo_order(oo);
1404
Joonsoo Kim80c3a992014-03-12 17:26:20 +09001405 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001406
1407 /*
1408 * Objects from caches that have a constructor don't get
1409 * cleared when they're allocated, so we need to do it here.
1410 */
1411 if (s->ctor)
1412 kmemcheck_mark_uninitialized_pages(page, pages);
1413 else
1414 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001415 }
1416
Christoph Lameter834f3d12008-04-14 19:11:31 +03001417 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001418
Glauber Costa1f458cb2012-12-18 14:22:50 -08001419 order = compound_order(page);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001420 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001421 __SetPageSlab(page);
Michal Hocko2f064f32015-08-21 14:11:51 -07001422 if (page_is_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -07001423 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001424
1425 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001426
1427 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001428 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001429
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001430 kasan_poison_slab(page);
1431
Wei Yang54266642014-08-06 16:04:42 -07001432 for_each_object_idx(p, idx, s, start, page->objects) {
1433 setup_object(s, page, p);
1434 if (likely(idx < page->objects))
1435 set_freepointer(s, p, p + s->size);
1436 else
1437 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001438 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001439
1440 page->freelist = start;
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001441 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001442 page->frozen = 1;
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001443
Christoph Lameter81819f02007-05-06 14:49:36 -07001444out:
Mel Gormand0164ad2015-11-06 16:28:21 -08001445 if (gfpflags_allow_blocking(flags))
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001446 local_irq_disable();
1447 if (!page)
1448 return NULL;
1449
1450 mod_zone_page_state(page_zone(page),
1451 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1452 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1453 1 << oo_order(oo));
1454
1455 inc_slabs_node(s, page_to_nid(page), page->objects);
1456
Christoph Lameter81819f02007-05-06 14:49:36 -07001457 return page;
1458}
1459
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001460static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1461{
1462 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1463 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1464 BUG();
1465 }
1466
1467 return allocate_slab(s,
1468 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1469}
1470
Christoph Lameter81819f02007-05-06 14:49:36 -07001471static void __free_slab(struct kmem_cache *s, struct page *page)
1472{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001473 int order = compound_order(page);
1474 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001475
Laura Abbottbecfda62016-03-15 14:55:06 -07001476 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001477 void *p;
1478
1479 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001480 for_each_object(p, s, page_address(page),
1481 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001482 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001483 }
1484
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001485 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001486
Christoph Lameter81819f02007-05-06 14:49:36 -07001487 mod_zone_page_state(page_zone(page),
1488 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1489 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001490 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001491
Mel Gorman072bb0a2012-07-31 16:43:58 -07001492 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001493 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001494
Mel Gorman22b751c2013-02-22 16:34:59 -08001495 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001496 if (current->reclaim_state)
1497 current->reclaim_state->reclaimed_slab += pages;
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001498 __free_kmem_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001499}
1500
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001501#define need_reserve_slab_rcu \
1502 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1503
Christoph Lameter81819f02007-05-06 14:49:36 -07001504static void rcu_free_slab(struct rcu_head *h)
1505{
1506 struct page *page;
1507
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001508 if (need_reserve_slab_rcu)
1509 page = virt_to_head_page(h);
1510 else
1511 page = container_of((struct list_head *)h, struct page, lru);
1512
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001513 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001514}
1515
1516static void free_slab(struct kmem_cache *s, struct page *page)
1517{
1518 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001519 struct rcu_head *head;
1520
1521 if (need_reserve_slab_rcu) {
1522 int order = compound_order(page);
1523 int offset = (PAGE_SIZE << order) - s->reserved;
1524
1525 VM_BUG_ON(s->reserved != sizeof(*head));
1526 head = page_address(page) + offset;
1527 } else {
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001528 head = &page->rcu_head;
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001529 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001530
1531 call_rcu(head, rcu_free_slab);
1532 } else
1533 __free_slab(s, page);
1534}
1535
1536static void discard_slab(struct kmem_cache *s, struct page *page)
1537{
Christoph Lameter205ab992008-04-14 19:11:40 +03001538 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001539 free_slab(s, page);
1540}
1541
1542/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001543 * Management of partially allocated slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07001544 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001545static inline void
1546__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001547{
Christoph Lametere95eed52007-05-06 14:49:44 -07001548 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001549 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001550 list_add_tail(&page->lru, &n->partial);
1551 else
1552 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001553}
1554
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001555static inline void add_partial(struct kmem_cache_node *n,
1556 struct page *page, int tail)
1557{
1558 lockdep_assert_held(&n->list_lock);
1559 __add_partial(n, page, tail);
1560}
1561
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001562static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001563 struct page *page)
1564{
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001565 lockdep_assert_held(&n->list_lock);
Dmitry Safonov52b4b952016-02-17 13:11:37 -08001566 list_del(&page->lru);
1567 n->nr_partial--;
Christoph Lameter62e346a2010-09-28 08:10:28 -05001568}
1569
Christoph Lameter81819f02007-05-06 14:49:36 -07001570/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001571 * Remove slab from the partial list, freeze it and
1572 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001573 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001574 * Returns a list of objects or NULL if it fails.
Christoph Lameter81819f02007-05-06 14:49:36 -07001575 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001576static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001577 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001578 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001579{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001580 void *freelist;
1581 unsigned long counters;
1582 struct page new;
1583
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001584 lockdep_assert_held(&n->list_lock);
1585
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001586 /*
1587 * Zap the freelist and set the frozen bit.
1588 * The old freelist is the list of objects for the
1589 * per cpu allocation list.
1590 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001591 freelist = page->freelist;
1592 counters = page->counters;
1593 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001594 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001595 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001596 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001597 new.freelist = NULL;
1598 } else {
1599 new.freelist = freelist;
1600 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001601
Dave Hansena0132ac2014-01-29 14:05:50 -08001602 VM_BUG_ON(new.frozen);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001603 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001604
Christoph Lameter7ced3712012-05-09 10:09:53 -05001605 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001606 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001607 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001608 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001609 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001610
1611 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001612 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001613 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001614}
1615
Joonsoo Kim633b0762013-01-21 17:01:25 +09001616static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001617static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001618
Christoph Lameter81819f02007-05-06 14:49:36 -07001619/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001620 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001621 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001622static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1623 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001624{
Christoph Lameter49e22582011-08-09 16:12:27 -05001625 struct page *page, *page2;
1626 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001627 int available = 0;
1628 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001629
1630 /*
1631 * Racy check. If we mistakenly see no partial slabs then we
1632 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001633 * partial slab and there is none available then get_partials()
1634 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001635 */
1636 if (!n || !n->nr_partial)
1637 return NULL;
1638
1639 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001640 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001641 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001642
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001643 if (!pfmemalloc_match(page, flags))
1644 continue;
1645
Joonsoo Kim633b0762013-01-21 17:01:25 +09001646 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001647 if (!t)
1648 break;
1649
Joonsoo Kim633b0762013-01-21 17:01:25 +09001650 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001651 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001652 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001653 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001654 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001655 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001656 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001657 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001658 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001659 if (!kmem_cache_has_cpu_partial(s)
1660 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001661 break;
1662
Christoph Lameter497b66f2011-08-09 16:12:26 -05001663 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001664 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001665 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001666}
1667
1668/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001669 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001670 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001671static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001672 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001673{
1674#ifdef CONFIG_NUMA
1675 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001676 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001677 struct zone *zone;
1678 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001679 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001680 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001681
1682 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001683 * The defrag ratio allows a configuration of the tradeoffs between
1684 * inter node defragmentation and node local allocations. A lower
1685 * defrag_ratio increases the tendency to do local allocations
1686 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001687 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001688 * If the defrag_ratio is set to 0 then kmalloc() always
1689 * returns node local objects. If the ratio is higher then kmalloc()
1690 * may return off node objects because partial slabs are obtained
1691 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001692 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001693 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001694 * defrag_ratio = 1000) then every (well almost) allocation will
1695 * first attempt to defrag slab caches on other nodes. This means
1696 * scanning over all nodes to look for partial slabs which may be
1697 * expensive if we do it every time we are trying to find a slab
1698 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001699 */
Christoph Lameter98246012008-01-07 23:20:26 -08001700 if (!s->remote_node_defrag_ratio ||
1701 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001702 return NULL;
1703
Mel Gormancc9a6c82012-03-21 16:34:11 -07001704 do {
Mel Gormand26914d2014-04-03 14:47:24 -07001705 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07001706 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001707 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1708 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001709
Mel Gormancc9a6c82012-03-21 16:34:11 -07001710 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001711
Vladimir Davydovdee2f8a2014-12-12 16:58:28 -08001712 if (n && cpuset_zone_allowed(zone, flags) &&
Mel Gormancc9a6c82012-03-21 16:34:11 -07001713 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001714 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001715 if (object) {
1716 /*
Mel Gormand26914d2014-04-03 14:47:24 -07001717 * Don't check read_mems_allowed_retry()
1718 * here - if mems_allowed was updated in
1719 * parallel, that was a harmless race
1720 * between allocation and the cpuset
1721 * update
Mel Gormancc9a6c82012-03-21 16:34:11 -07001722 */
Mel Gormancc9a6c82012-03-21 16:34:11 -07001723 return object;
1724 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001725 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001726 }
Mel Gormand26914d2014-04-03 14:47:24 -07001727 } while (read_mems_allowed_retry(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001728#endif
1729 return NULL;
1730}
1731
1732/*
1733 * Get a partial page, lock it and return it.
1734 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001735static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001736 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001737{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001738 void *object;
Joonsoo Kima561ce02014-10-09 15:26:15 -07001739 int searchnode = node;
1740
1741 if (node == NUMA_NO_NODE)
1742 searchnode = numa_mem_id();
1743 else if (!node_present_pages(node))
1744 searchnode = node_to_mem_node(node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001745
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001746 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001747 if (object || node != NUMA_NO_NODE)
1748 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001749
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001750 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001751}
1752
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001753#ifdef CONFIG_PREEMPT
1754/*
1755 * Calculate the next globally unique transaction for disambiguiation
1756 * during cmpxchg. The transactions start with the cpu number and are then
1757 * incremented by CONFIG_NR_CPUS.
1758 */
1759#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1760#else
1761/*
1762 * No preemption supported therefore also no need to check for
1763 * different cpus.
1764 */
1765#define TID_STEP 1
1766#endif
1767
1768static inline unsigned long next_tid(unsigned long tid)
1769{
1770 return tid + TID_STEP;
1771}
1772
1773static inline unsigned int tid_to_cpu(unsigned long tid)
1774{
1775 return tid % TID_STEP;
1776}
1777
1778static inline unsigned long tid_to_event(unsigned long tid)
1779{
1780 return tid / TID_STEP;
1781}
1782
1783static inline unsigned int init_tid(int cpu)
1784{
1785 return cpu;
1786}
1787
1788static inline void note_cmpxchg_failure(const char *n,
1789 const struct kmem_cache *s, unsigned long tid)
1790{
1791#ifdef SLUB_DEBUG_CMPXCHG
1792 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1793
Fabian Frederickf9f58282014-06-04 16:06:34 -07001794 pr_info("%s %s: cmpxchg redo ", n, s->name);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001795
1796#ifdef CONFIG_PREEMPT
1797 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
Fabian Frederickf9f58282014-06-04 16:06:34 -07001798 pr_warn("due to cpu change %d -> %d\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001799 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1800 else
1801#endif
1802 if (tid_to_event(tid) != tid_to_event(actual_tid))
Fabian Frederickf9f58282014-06-04 16:06:34 -07001803 pr_warn("due to cpu running other code. Event %ld->%ld\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001804 tid_to_event(tid), tid_to_event(actual_tid));
1805 else
Fabian Frederickf9f58282014-06-04 16:06:34 -07001806 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001807 actual_tid, tid, next_tid(tid));
1808#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001809 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001810}
1811
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001812static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001813{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001814 int cpu;
1815
1816 for_each_possible_cpu(cpu)
1817 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001818}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001819
1820/*
1821 * Remove the cpu slab
1822 */
Chen Gangd0e0ac92013-07-15 09:05:29 +08001823static void deactivate_slab(struct kmem_cache *s, struct page *page,
1824 void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001825{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001826 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001827 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1828 int lock = 0;
1829 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001830 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001831 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001832 struct page new;
1833 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001834
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001835 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001836 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001837 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001838 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001839
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001840 /*
1841 * Stage one: Free all available per cpu objects back
1842 * to the page freelist while it is still frozen. Leave the
1843 * last one.
1844 *
1845 * There is no need to take the list->lock because the page
1846 * is still frozen.
1847 */
1848 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1849 void *prior;
1850 unsigned long counters;
1851
1852 do {
1853 prior = page->freelist;
1854 counters = page->counters;
1855 set_freepointer(s, freelist, prior);
1856 new.counters = counters;
1857 new.inuse--;
Dave Hansena0132ac2014-01-29 14:05:50 -08001858 VM_BUG_ON(!new.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001859
Christoph Lameter1d071712011-07-14 12:49:12 -05001860 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001861 prior, counters,
1862 freelist, new.counters,
1863 "drain percpu freelist"));
1864
1865 freelist = nextfree;
1866 }
1867
1868 /*
1869 * Stage two: Ensure that the page is unfrozen while the
1870 * list presence reflects the actual number of objects
1871 * during unfreeze.
1872 *
1873 * We setup the list membership and then perform a cmpxchg
1874 * with the count. If there is a mismatch then the page
1875 * is not unfrozen but the page is on the wrong list.
1876 *
1877 * Then we restart the process which may have to remove
1878 * the page from the list that we just put it on again
1879 * because the number of objects in the slab may have
1880 * changed.
1881 */
1882redo:
1883
1884 old.freelist = page->freelist;
1885 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08001886 VM_BUG_ON(!old.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001887
1888 /* Determine target state of the slab */
1889 new.counters = old.counters;
1890 if (freelist) {
1891 new.inuse--;
1892 set_freepointer(s, freelist, old.freelist);
1893 new.freelist = freelist;
1894 } else
1895 new.freelist = old.freelist;
1896
1897 new.frozen = 0;
1898
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07001899 if (!new.inuse && n->nr_partial >= s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001900 m = M_FREE;
1901 else if (new.freelist) {
1902 m = M_PARTIAL;
1903 if (!lock) {
1904 lock = 1;
1905 /*
1906 * Taking the spinlock removes the possiblity
1907 * that acquire_slab() will see a slab page that
1908 * is frozen
1909 */
1910 spin_lock(&n->list_lock);
1911 }
1912 } else {
1913 m = M_FULL;
1914 if (kmem_cache_debug(s) && !lock) {
1915 lock = 1;
1916 /*
1917 * This also ensures that the scanning of full
1918 * slabs from diagnostic functions will not see
1919 * any frozen slabs.
1920 */
1921 spin_lock(&n->list_lock);
1922 }
1923 }
1924
1925 if (l != m) {
1926
1927 if (l == M_PARTIAL)
1928
1929 remove_partial(n, page);
1930
1931 else if (l == M_FULL)
1932
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001933 remove_full(s, n, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001934
1935 if (m == M_PARTIAL) {
1936
1937 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001938 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001939
1940 } else if (m == M_FULL) {
1941
1942 stat(s, DEACTIVATE_FULL);
1943 add_full(s, n, page);
1944
1945 }
1946 }
1947
1948 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001949 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001950 old.freelist, old.counters,
1951 new.freelist, new.counters,
1952 "unfreezing slab"))
1953 goto redo;
1954
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001955 if (lock)
1956 spin_unlock(&n->list_lock);
1957
1958 if (m == M_FREE) {
1959 stat(s, DEACTIVATE_EMPTY);
1960 discard_slab(s, page);
1961 stat(s, FREE_SLAB);
1962 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001963}
1964
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001965/*
1966 * Unfreeze all the cpu partial slabs.
1967 *
Christoph Lameter59a09912012-11-28 16:23:00 +00001968 * This function must be called with interrupts disabled
1969 * for the cpu using c (or some other guarantee must be there
1970 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09001971 */
Christoph Lameter59a09912012-11-28 16:23:00 +00001972static void unfreeze_partials(struct kmem_cache *s,
1973 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05001974{
Joonsoo Kim345c9052013-06-19 14:05:52 +09001975#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09001976 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08001977 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05001978
1979 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001980 struct page new;
1981 struct page old;
1982
1983 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09001984
1985 n2 = get_node(s, page_to_nid(page));
1986 if (n != n2) {
1987 if (n)
1988 spin_unlock(&n->list_lock);
1989
1990 n = n2;
1991 spin_lock(&n->list_lock);
1992 }
Christoph Lameter49e22582011-08-09 16:12:27 -05001993
1994 do {
1995
1996 old.freelist = page->freelist;
1997 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08001998 VM_BUG_ON(!old.frozen);
Christoph Lameter49e22582011-08-09 16:12:27 -05001999
2000 new.counters = old.counters;
2001 new.freelist = old.freelist;
2002
2003 new.frozen = 0;
2004
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002005 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05002006 old.freelist, old.counters,
2007 new.freelist, new.counters,
2008 "unfreezing slab"));
2009
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07002010 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08002011 page->next = discard_page;
2012 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09002013 } else {
2014 add_partial(n, page, DEACTIVATE_TO_TAIL);
2015 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05002016 }
2017 }
2018
2019 if (n)
2020 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08002021
2022 while (discard_page) {
2023 page = discard_page;
2024 discard_page = discard_page->next;
2025
2026 stat(s, DEACTIVATE_EMPTY);
2027 discard_slab(s, page);
2028 stat(s, FREE_SLAB);
2029 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09002030#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002031}
2032
2033/*
2034 * Put a page that was just frozen (in __slab_free) into a partial page
2035 * slot if available. This is done without interrupts disabled and without
2036 * preemption disabled. The cmpxchg is racy and may put the partial page
2037 * onto a random cpus partial slot.
2038 *
2039 * If we did not find a slot then simply move all the partials to the
2040 * per node partial list.
2041 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09002042static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05002043{
Joonsoo Kim345c9052013-06-19 14:05:52 +09002044#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05002045 struct page *oldpage;
2046 int pages;
2047 int pobjects;
2048
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002049 preempt_disable();
Christoph Lameter49e22582011-08-09 16:12:27 -05002050 do {
2051 pages = 0;
2052 pobjects = 0;
2053 oldpage = this_cpu_read(s->cpu_slab->partial);
2054
2055 if (oldpage) {
2056 pobjects = oldpage->pobjects;
2057 pages = oldpage->pages;
2058 if (drain && pobjects > s->cpu_partial) {
2059 unsigned long flags;
2060 /*
2061 * partial array is full. Move the existing
2062 * set to the per node partial list.
2063 */
2064 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00002065 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05002066 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09002067 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05002068 pobjects = 0;
2069 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08002070 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05002071 }
2072 }
2073
2074 pages++;
2075 pobjects += page->objects - page->inuse;
2076
2077 page->pages = pages;
2078 page->pobjects = pobjects;
2079 page->next = oldpage;
2080
Chen Gangd0e0ac92013-07-15 09:05:29 +08002081 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2082 != oldpage);
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002083 if (unlikely(!s->cpu_partial)) {
2084 unsigned long flags;
2085
2086 local_irq_save(flags);
2087 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2088 local_irq_restore(flags);
2089 }
2090 preempt_enable();
Joonsoo Kim345c9052013-06-19 14:05:52 +09002091#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002092}
2093
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002094static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002095{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002096 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002097 deactivate_slab(s, c->page, c->freelist);
2098
2099 c->tid = next_tid(c->tid);
2100 c->page = NULL;
2101 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002102}
2103
2104/*
2105 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002106 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002107 * Called from IPI handler with interrupts disabled.
2108 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002109static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002110{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002111 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002112
Christoph Lameter49e22582011-08-09 16:12:27 -05002113 if (likely(c)) {
2114 if (c->page)
2115 flush_slab(s, c);
2116
Christoph Lameter59a09912012-11-28 16:23:00 +00002117 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002118 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002119}
2120
2121static void flush_cpu_slab(void *d)
2122{
2123 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002124
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002125 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002126}
2127
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002128static bool has_cpu_slab(int cpu, void *info)
2129{
2130 struct kmem_cache *s = info;
2131 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2132
majianpeng02e1a9c2012-05-17 17:03:26 -07002133 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002134}
2135
Christoph Lameter81819f02007-05-06 14:49:36 -07002136static void flush_all(struct kmem_cache *s)
2137{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002138 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002139}
2140
2141/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002142 * Check if the objects in a per cpu structure fit numa
2143 * locality expectations.
2144 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002145static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002146{
2147#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002148 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002149 return 0;
2150#endif
2151 return 1;
2152}
2153
David Rientjes9a02d692014-06-04 16:06:36 -07002154#ifdef CONFIG_SLUB_DEBUG
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002155static int count_free(struct page *page)
2156{
2157 return page->objects - page->inuse;
2158}
2159
David Rientjes9a02d692014-06-04 16:06:36 -07002160static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2161{
2162 return atomic_long_read(&n->total_objects);
2163}
2164#endif /* CONFIG_SLUB_DEBUG */
2165
2166#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002167static unsigned long count_partial(struct kmem_cache_node *n,
2168 int (*get_count)(struct page *))
2169{
2170 unsigned long flags;
2171 unsigned long x = 0;
2172 struct page *page;
2173
2174 spin_lock_irqsave(&n->list_lock, flags);
2175 list_for_each_entry(page, &n->partial, lru)
2176 x += get_count(page);
2177 spin_unlock_irqrestore(&n->list_lock, flags);
2178 return x;
2179}
David Rientjes9a02d692014-06-04 16:06:36 -07002180#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002181
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002182static noinline void
2183slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2184{
David Rientjes9a02d692014-06-04 16:06:36 -07002185#ifdef CONFIG_SLUB_DEBUG
2186 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2187 DEFAULT_RATELIMIT_BURST);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002188 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07002189 struct kmem_cache_node *n;
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002190
David Rientjes9a02d692014-06-04 16:06:36 -07002191 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2192 return;
2193
Fabian Frederickf9f58282014-06-04 16:06:34 -07002194 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002195 nid, gfpflags);
Fabian Frederickf9f58282014-06-04 16:06:34 -07002196 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2197 s->name, s->object_size, s->size, oo_order(s->oo),
2198 oo_order(s->min));
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002199
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002200 if (oo_order(s->min) > get_order(s->object_size))
Fabian Frederickf9f58282014-06-04 16:06:34 -07002201 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2202 s->name);
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002203
Christoph Lameterfa45dc22014-08-06 16:04:09 -07002204 for_each_kmem_cache_node(s, node, n) {
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002205 unsigned long nr_slabs;
2206 unsigned long nr_objs;
2207 unsigned long nr_free;
2208
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002209 nr_free = count_partial(n, count_free);
2210 nr_slabs = node_nr_slabs(n);
2211 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002212
Fabian Frederickf9f58282014-06-04 16:06:34 -07002213 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002214 node, nr_slabs, nr_objs, nr_free);
2215 }
David Rientjes9a02d692014-06-04 16:06:36 -07002216#endif
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002217}
2218
Christoph Lameter497b66f2011-08-09 16:12:26 -05002219static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2220 int node, struct kmem_cache_cpu **pc)
2221{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002222 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002223 struct kmem_cache_cpu *c = *pc;
2224 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002225
Christoph Lameter188fd062012-05-09 10:09:55 -05002226 freelist = get_partial(s, flags, node, c);
2227
2228 if (freelist)
2229 return freelist;
2230
2231 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002232 if (page) {
Christoph Lameter7c8e0182014-06-04 16:07:56 -07002233 c = raw_cpu_ptr(s->cpu_slab);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002234 if (c->page)
2235 flush_slab(s, c);
2236
2237 /*
2238 * No other reference to the page yet so we can
2239 * muck around with it freely without cmpxchg
2240 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002241 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002242 page->freelist = NULL;
2243
2244 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002245 c->page = page;
2246 *pc = c;
2247 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002248 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002249
Christoph Lameter6faa6832012-05-09 10:09:51 -05002250 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002251}
2252
Mel Gorman072bb0a2012-07-31 16:43:58 -07002253static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2254{
2255 if (unlikely(PageSlabPfmemalloc(page)))
2256 return gfp_pfmemalloc_allowed(gfpflags);
2257
2258 return true;
2259}
2260
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002261/*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002262 * Check the page->freelist of a page and either transfer the freelist to the
2263 * per cpu freelist or deactivate the page.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002264 *
2265 * The page is still frozen if the return value is not NULL.
2266 *
2267 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002268 *
2269 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002270 */
2271static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2272{
2273 struct page new;
2274 unsigned long counters;
2275 void *freelist;
2276
2277 do {
2278 freelist = page->freelist;
2279 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002280
Christoph Lameter213eeb92011-11-11 14:07:14 -06002281 new.counters = counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08002282 VM_BUG_ON(!new.frozen);
Christoph Lameter213eeb92011-11-11 14:07:14 -06002283
2284 new.inuse = page->objects;
2285 new.frozen = freelist != NULL;
2286
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002287 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002288 freelist, counters,
2289 NULL, new.counters,
2290 "get_freelist"));
2291
2292 return freelist;
2293}
2294
2295/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002296 * Slow path. The lockless freelist is empty or we need to perform
2297 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002298 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002299 * Processing is still very fast if new objects have been freed to the
2300 * regular freelist. In that case we simply take over the regular freelist
2301 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002302 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002303 * If that is not working then we fall back to the partial lists. We take the
2304 * first element of the freelist as the object to allocate now and move the
2305 * rest of the freelist to the lockless freelist.
2306 *
2307 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002308 * we need to allocate a new slab. This is the slowest path since it involves
2309 * a call to the page allocator and the setup of a new slab.
Christoph Lametera380a3c2015-11-20 15:57:35 -08002310 *
2311 * Version of __slab_alloc to use when we know that interrupts are
2312 * already disabled (which is the case for bulk allocation).
Christoph Lameter81819f02007-05-06 14:49:36 -07002313 */
Christoph Lametera380a3c2015-11-20 15:57:35 -08002314static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002315 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002316{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002317 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002318 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002319
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002320 page = c->page;
2321 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002322 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002323redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002324
Christoph Lameter57d437d2012-05-09 10:09:59 -05002325 if (unlikely(!node_match(page, node))) {
Joonsoo Kima561ce02014-10-09 15:26:15 -07002326 int searchnode = node;
2327
2328 if (node != NUMA_NO_NODE && !node_present_pages(node))
2329 searchnode = node_to_mem_node(node);
2330
2331 if (unlikely(!node_match(page, searchnode))) {
2332 stat(s, ALLOC_NODE_MISMATCH);
2333 deactivate_slab(s, page, c->freelist);
2334 c->page = NULL;
2335 c->freelist = NULL;
2336 goto new_slab;
2337 }
Christoph Lameterfc59c052011-06-01 12:25:56 -05002338 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002339
Mel Gorman072bb0a2012-07-31 16:43:58 -07002340 /*
2341 * By rights, we should be searching for a slab page that was
2342 * PFMEMALLOC but right now, we are losing the pfmemalloc
2343 * information when the page leaves the per-cpu allocator
2344 */
2345 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2346 deactivate_slab(s, page, c->freelist);
2347 c->page = NULL;
2348 c->freelist = NULL;
2349 goto new_slab;
2350 }
2351
Eric Dumazet73736e02011-12-13 04:57:06 +01002352 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002353 freelist = c->freelist;
2354 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002355 goto load_freelist;
2356
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002357 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002358
Christoph Lameter6faa6832012-05-09 10:09:51 -05002359 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002360 c->page = NULL;
2361 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002362 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002363 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002364
Christoph Lameter81819f02007-05-06 14:49:36 -07002365 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002366
Christoph Lameter894b8782007-05-10 03:15:16 -07002367load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002368 /*
2369 * freelist is pointing to the list of objects to be used.
2370 * page is pointing to the page from which the objects are obtained.
2371 * That page must be frozen for per cpu allocations to work.
2372 */
Dave Hansena0132ac2014-01-29 14:05:50 -08002373 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002374 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002375 c->tid = next_tid(c->tid);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002376 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002377
Christoph Lameter81819f02007-05-06 14:49:36 -07002378new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002379
Christoph Lameter49e22582011-08-09 16:12:27 -05002380 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002381 page = c->page = c->partial;
2382 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002383 stat(s, CPU_PARTIAL_ALLOC);
2384 c->freelist = NULL;
2385 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002386 }
2387
Christoph Lameter188fd062012-05-09 10:09:55 -05002388 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002389
Christoph Lameterf46974362012-05-09 10:09:54 -05002390 if (unlikely(!freelist)) {
David Rientjes9a02d692014-06-04 16:06:36 -07002391 slab_out_of_memory(s, gfpflags, node);
Christoph Lameterf46974362012-05-09 10:09:54 -05002392 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002393 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002394
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002395 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002396 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002397 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002398
Christoph Lameter497b66f2011-08-09 16:12:26 -05002399 /* Only entered in the debug case */
Chen Gangd0e0ac92013-07-15 09:05:29 +08002400 if (kmem_cache_debug(s) &&
2401 !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002402 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002403
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002404 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002405 c->page = NULL;
2406 c->freelist = NULL;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002407 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002408}
2409
2410/*
Christoph Lametera380a3c2015-11-20 15:57:35 -08002411 * Another one that disabled interrupt and compensates for possible
2412 * cpu changes by refetching the per cpu area pointer.
2413 */
2414static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2415 unsigned long addr, struct kmem_cache_cpu *c)
2416{
2417 void *p;
2418 unsigned long flags;
2419
2420 local_irq_save(flags);
2421#ifdef CONFIG_PREEMPT
2422 /*
2423 * We may have been preempted and rescheduled on a different
2424 * cpu before disabling interrupts. Need to reload cpu area
2425 * pointer.
2426 */
2427 c = this_cpu_ptr(s->cpu_slab);
2428#endif
2429
2430 p = ___slab_alloc(s, gfpflags, node, addr, c);
2431 local_irq_restore(flags);
2432 return p;
2433}
2434
2435/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002436 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2437 * have the fastpath folded into their functions. So no function call
2438 * overhead for requests that can be satisfied on the fastpath.
2439 *
2440 * The fastpath works by first checking if the lockless freelist can be used.
2441 * If not then __slab_alloc is called for slow processing.
2442 *
2443 * Otherwise we can simply pick the next object from the lockless free list.
2444 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002445static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002446 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002447{
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002448 void *object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002449 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002450 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002451 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002452
Vladimir Davydov8135be52014-12-12 16:56:38 -08002453 s = slab_pre_alloc_hook(s, gfpflags);
2454 if (!s)
Akinobu Mita773ff602008-12-23 19:37:01 +09002455 return NULL;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002456redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002457 /*
2458 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2459 * enabled. We may switch back and forth between cpus while
2460 * reading from one cpu area. That does not matter as long
2461 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002462 *
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002463 * We should guarantee that tid and kmem_cache are retrieved on
2464 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2465 * to check if it is matched or not.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002466 */
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002467 do {
2468 tid = this_cpu_read(s->cpu_slab->tid);
2469 c = raw_cpu_ptr(s->cpu_slab);
Mark Rutland859b7a02015-03-25 15:55:23 -07002470 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2471 unlikely(tid != READ_ONCE(c->tid)));
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002472
2473 /*
2474 * Irqless object alloc/free algorithm used here depends on sequence
2475 * of fetching cpu_slab's data. tid should be fetched before anything
2476 * on c to guarantee that object and page associated with previous tid
2477 * won't be used with current tid. If we fetch tid first, object and
2478 * page could be one associated with next tid and our alloc/free
2479 * request will be failed. In this case, we will retry. So, no problem.
2480 */
2481 barrier();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002482
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002483 /*
2484 * The transaction ids are globally unique per cpu and per operation on
2485 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2486 * occurs on the right processor and that there was no operation on the
2487 * linked list in between.
2488 */
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002489
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002490 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002491 page = c->page;
Dave Hansen8eae1492014-06-04 16:06:37 -07002492 if (unlikely(!object || !node_match(page, node))) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002493 object = __slab_alloc(s, gfpflags, node, addr, c);
Dave Hansen8eae1492014-06-04 16:06:37 -07002494 stat(s, ALLOC_SLOWPATH);
2495 } else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002496 void *next_object = get_freepointer_safe(s, object);
2497
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002498 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002499 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002500 * operation and if we are on the right processor.
2501 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002502 * The cmpxchg does the following atomically (without lock
2503 * semantics!)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002504 * 1. Relocate first pointer to the current per cpu area.
2505 * 2. Verify that tid and freelist have not been changed
2506 * 3. If they were not changed replace tid and freelist
2507 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002508 * Since this is without lock semantics the protection is only
2509 * against code executing on this cpu *not* from access by
2510 * other cpus.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002511 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002512 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002513 s->cpu_slab->freelist, s->cpu_slab->tid,
2514 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002515 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002516
2517 note_cmpxchg_failure("slab_alloc", s, tid);
2518 goto redo;
2519 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002520 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002521 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002522 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002523
Pekka Enberg74e21342009-11-25 20:14:48 +02002524 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002525 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002526
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002527 slab_post_alloc_hook(s, gfpflags, 1, &object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002528
Christoph Lameter894b8782007-05-10 03:15:16 -07002529 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002530}
2531
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002532static __always_inline void *slab_alloc(struct kmem_cache *s,
2533 gfp_t gfpflags, unsigned long addr)
2534{
2535 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2536}
2537
Christoph Lameter81819f02007-05-06 14:49:36 -07002538void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2539{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002540 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002541
Chen Gangd0e0ac92013-07-15 09:05:29 +08002542 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2543 s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002544
2545 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002546}
2547EXPORT_SYMBOL(kmem_cache_alloc);
2548
Li Zefan0f24f122009-12-11 15:45:30 +08002549#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002550void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002551{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002552 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002553 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08002554 kasan_kmalloc(s, ret, size);
Richard Kennedy4a923792010-10-21 10:29:19 +01002555 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002556}
Richard Kennedy4a923792010-10-21 10:29:19 +01002557EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002558#endif
2559
Christoph Lameter81819f02007-05-06 14:49:36 -07002560#ifdef CONFIG_NUMA
2561void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2562{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002563 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002564
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002565 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002566 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002567
2568 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002569}
2570EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002571
Li Zefan0f24f122009-12-11 15:45:30 +08002572#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002573void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002574 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002575 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002576{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002577 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002578
2579 trace_kmalloc_node(_RET_IP_, ret,
2580 size, s->size, gfpflags, node);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08002581
2582 kasan_kmalloc(s, ret, size);
Richard Kennedy4a923792010-10-21 10:29:19 +01002583 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002584}
Richard Kennedy4a923792010-10-21 10:29:19 +01002585EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002586#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002587#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002588
Christoph Lameter81819f02007-05-06 14:49:36 -07002589/*
Kim Phillips94e4d712015-02-10 14:09:37 -08002590 * Slow path handling. This may still be called frequently since objects
Christoph Lameter894b8782007-05-10 03:15:16 -07002591 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002592 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002593 * So we still attempt to reduce cache line usage. Just take the slab
2594 * lock and free the item. If there is no additional partial page
2595 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002596 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002597static void __slab_free(struct kmem_cache *s, struct page *page,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002598 void *head, void *tail, int cnt,
2599 unsigned long addr)
2600
Christoph Lameter81819f02007-05-06 14:49:36 -07002601{
2602 void *prior;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002603 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002604 struct page new;
2605 unsigned long counters;
2606 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002607 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002608
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002609 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002610
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002611 if (kmem_cache_debug(s) &&
Laura Abbott282acb42016-03-15 14:54:59 -07002612 !free_debug_processing(s, page, head, tail, cnt, addr))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002613 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002614
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002615 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002616 if (unlikely(n)) {
2617 spin_unlock_irqrestore(&n->list_lock, flags);
2618 n = NULL;
2619 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002620 prior = page->freelist;
2621 counters = page->counters;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002622 set_freepointer(s, tail, prior);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002623 new.counters = counters;
2624 was_frozen = new.frozen;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002625 new.inuse -= cnt;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002626 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002627
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002628 if (kmem_cache_has_cpu_partial(s) && !prior) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002629
2630 /*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002631 * Slab was on no list before and will be
2632 * partially empty
2633 * We can defer the list move and instead
2634 * freeze it.
Christoph Lameter49e22582011-08-09 16:12:27 -05002635 */
2636 new.frozen = 1;
2637
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002638 } else { /* Needs to be taken off a list */
Christoph Lameter49e22582011-08-09 16:12:27 -05002639
LQYMGTb455def2014-12-10 15:42:13 -08002640 n = get_node(s, page_to_nid(page));
Christoph Lameter49e22582011-08-09 16:12:27 -05002641 /*
2642 * Speculatively acquire the list_lock.
2643 * If the cmpxchg does not succeed then we may
2644 * drop the list_lock without any processing.
2645 *
2646 * Otherwise the list_lock will synchronize with
2647 * other processors updating the list of slabs.
2648 */
2649 spin_lock_irqsave(&n->list_lock, flags);
2650
2651 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002652 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002653
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002654 } while (!cmpxchg_double_slab(s, page,
2655 prior, counters,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002656 head, new.counters,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002657 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002658
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002659 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002660
2661 /*
2662 * If we just froze the page then put it onto the
2663 * per cpu partial list.
2664 */
Alex Shi8028dce2012-02-03 23:34:56 +08002665 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002666 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002667 stat(s, CPU_PARTIAL_FREE);
2668 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002669 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002670 * The list lock was not taken therefore no list
2671 * activity can be necessary.
2672 */
LQYMGTb455def2014-12-10 15:42:13 -08002673 if (was_frozen)
2674 stat(s, FREE_FROZEN);
2675 return;
2676 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002677
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07002678 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
Joonsoo Kim837d6782012-08-16 00:02:40 +09002679 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002680
Joonsoo Kim837d6782012-08-16 00:02:40 +09002681 /*
2682 * Objects left in the slab. If it was not on the partial list before
2683 * then add it.
2684 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002685 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2686 if (kmem_cache_debug(s))
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002687 remove_full(s, n, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002688 add_partial(n, page, DEACTIVATE_TO_TAIL);
2689 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002690 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002691 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002692 return;
2693
2694slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002695 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002696 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002697 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002698 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002699 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002700 stat(s, FREE_REMOVE_PARTIAL);
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002701 } else {
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002702 /* Slab must be on the full list */
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002703 remove_full(s, n, page);
2704 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002705
Christoph Lameter80f08c12011-06-01 12:25:55 -05002706 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002707 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002708 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002709}
2710
Christoph Lameter894b8782007-05-10 03:15:16 -07002711/*
2712 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2713 * can perform fastpath freeing without additional function calls.
2714 *
2715 * The fastpath is only possible if we are freeing to the current cpu slab
2716 * of this processor. This typically the case if we have just allocated
2717 * the item before.
2718 *
2719 * If fastpath is not possible then fall back to __slab_free where we deal
2720 * with all sorts of special processing.
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002721 *
2722 * Bulk free of a freelist with several objects (all pointing to the
2723 * same page) possible by specifying head and tail ptr, plus objects
2724 * count (cnt). Bulk free indicated by tail pointer being set.
Christoph Lameter894b8782007-05-10 03:15:16 -07002725 */
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002726static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2727 void *head, void *tail, int cnt,
2728 unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002729{
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002730 void *tail_obj = tail ? : head;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002731 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002732 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002733
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002734 slab_free_freelist_hook(s, head, tail);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002735
Christoph Lametera24c5a02011-03-15 12:45:21 -05002736redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002737 /*
2738 * Determine the currently cpus per cpu slab.
2739 * The cpu may change afterward. However that does not matter since
2740 * data is retrieved via this pointer. If we are on the same cpu
Jesper Dangaard Brouer2ae44002015-09-04 15:45:31 -07002741 * during the cmpxchg then the free will succeed.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002742 */
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002743 do {
2744 tid = this_cpu_read(s->cpu_slab->tid);
2745 c = raw_cpu_ptr(s->cpu_slab);
Mark Rutland859b7a02015-03-25 15:55:23 -07002746 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2747 unlikely(tid != READ_ONCE(c->tid)));
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002748
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002749 /* Same with comment on barrier() in slab_alloc_node() */
2750 barrier();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002751
Christoph Lameter442b06b2011-05-17 16:29:31 -05002752 if (likely(page == c->page)) {
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002753 set_freepointer(s, tail_obj, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002754
Christoph Lameter933393f2011-12-22 11:58:51 -06002755 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002756 s->cpu_slab->freelist, s->cpu_slab->tid,
2757 c->freelist, tid,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002758 head, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002759
2760 note_cmpxchg_failure("slab_free", s, tid);
2761 goto redo;
2762 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002763 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002764 } else
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002765 __slab_free(s, page, head, tail_obj, cnt, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002766
Christoph Lameter894b8782007-05-10 03:15:16 -07002767}
2768
Christoph Lameter81819f02007-05-06 14:49:36 -07002769void kmem_cache_free(struct kmem_cache *s, void *x)
2770{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002771 s = cache_from_obj(s, x);
2772 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002773 return;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002774 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002775 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002776}
2777EXPORT_SYMBOL(kmem_cache_free);
2778
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002779struct detached_freelist {
2780 struct page *page;
2781 void *tail;
2782 void *freelist;
2783 int cnt;
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002784 struct kmem_cache *s;
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002785};
2786
2787/*
2788 * This function progressively scans the array with free objects (with
2789 * a limited look ahead) and extract objects belonging to the same
2790 * page. It builds a detached freelist directly within the given
2791 * page/objects. This can happen without any need for
2792 * synchronization, because the objects are owned by running process.
2793 * The freelist is build up as a single linked list in the objects.
2794 * The idea is, that this detached freelist can then be bulk
2795 * transferred to the real freelist(s), but only requiring a single
2796 * synchronization primitive. Look ahead in the array is limited due
2797 * to performance reasons.
2798 */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002799static inline
2800int build_detached_freelist(struct kmem_cache *s, size_t size,
2801 void **p, struct detached_freelist *df)
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002802{
2803 size_t first_skipped_index = 0;
2804 int lookahead = 3;
2805 void *object;
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002806 struct page *page;
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002807
2808 /* Always re-init detached_freelist */
2809 df->page = NULL;
2810
2811 do {
2812 object = p[--size];
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002813 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002814 } while (!object && size);
2815
2816 if (!object)
2817 return 0;
2818
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002819 page = virt_to_head_page(object);
2820 if (!s) {
2821 /* Handle kalloc'ed objects */
2822 if (unlikely(!PageSlab(page))) {
2823 BUG_ON(!PageCompound(page));
2824 kfree_hook(object);
2825 __free_kmem_pages(page, compound_order(page));
2826 p[size] = NULL; /* mark object processed */
2827 return size;
2828 }
2829 /* Derive kmem_cache from object */
2830 df->s = page->slab_cache;
2831 } else {
2832 df->s = cache_from_obj(s, object); /* Support for memcg */
2833 }
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002834
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002835 /* Start new detached freelist */
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002836 df->page = page;
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002837 set_freepointer(df->s, object, NULL);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002838 df->tail = object;
2839 df->freelist = object;
2840 p[size] = NULL; /* mark object processed */
2841 df->cnt = 1;
2842
2843 while (size) {
2844 object = p[--size];
2845 if (!object)
2846 continue; /* Skip processed objects */
2847
2848 /* df->page is always set at this point */
2849 if (df->page == virt_to_head_page(object)) {
2850 /* Opportunity build freelist */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002851 set_freepointer(df->s, object, df->freelist);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002852 df->freelist = object;
2853 df->cnt++;
2854 p[size] = NULL; /* mark object processed */
2855
2856 continue;
2857 }
2858
2859 /* Limit look ahead search */
2860 if (!--lookahead)
2861 break;
2862
2863 if (!first_skipped_index)
2864 first_skipped_index = size + 1;
2865 }
2866
2867 return first_skipped_index;
2868}
2869
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002870/* Note that interrupts must be enabled when calling this function. */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002871void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07002872{
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002873 if (WARN_ON(!size))
2874 return;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002875
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002876 do {
2877 struct detached_freelist df;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002878
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002879 size = build_detached_freelist(s, size, p, &df);
2880 if (unlikely(!df.page))
2881 continue;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002882
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002883 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002884 } while (likely(size));
Christoph Lameter484748f2015-09-04 15:45:34 -07002885}
2886EXPORT_SYMBOL(kmem_cache_free_bulk);
2887
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002888/* Note that interrupts must be enabled when calling this function. */
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002889int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2890 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07002891{
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002892 struct kmem_cache_cpu *c;
2893 int i;
2894
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002895 /* memcg and kmem_cache debug support */
2896 s = slab_pre_alloc_hook(s, flags);
2897 if (unlikely(!s))
2898 return false;
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002899 /*
2900 * Drain objects in the per cpu slab, while disabling local
2901 * IRQs, which protects against PREEMPT and interrupts
2902 * handlers invoking normal fastpath.
2903 */
2904 local_irq_disable();
2905 c = this_cpu_ptr(s->cpu_slab);
2906
2907 for (i = 0; i < size; i++) {
2908 void *object = c->freelist;
2909
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002910 if (unlikely(!object)) {
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002911 /*
2912 * Invoking slow path likely have side-effect
2913 * of re-populating per CPU c->freelist
2914 */
Christoph Lameter87098372015-11-20 15:57:38 -08002915 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002916 _RET_IP_, c);
Christoph Lameter87098372015-11-20 15:57:38 -08002917 if (unlikely(!p[i]))
2918 goto error;
2919
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002920 c = this_cpu_ptr(s->cpu_slab);
2921 continue; /* goto for-loop */
2922 }
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002923 c->freelist = get_freepointer(s, object);
2924 p[i] = object;
2925 }
2926 c->tid = next_tid(c->tid);
2927 local_irq_enable();
2928
2929 /* Clear memory outside IRQ disabled fastpath loop */
2930 if (unlikely(flags & __GFP_ZERO)) {
2931 int j;
2932
2933 for (j = 0; j < i; j++)
2934 memset(p[j], 0, s->object_size);
2935 }
2936
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002937 /* memcg and kmem_cache debug support */
2938 slab_post_alloc_hook(s, flags, size, p);
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002939 return i;
Christoph Lameter87098372015-11-20 15:57:38 -08002940error:
Christoph Lameter87098372015-11-20 15:57:38 -08002941 local_irq_enable();
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002942 slab_post_alloc_hook(s, flags, i, p);
2943 __kmem_cache_free_bulk(s, i, p);
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002944 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07002945}
2946EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2947
2948
Christoph Lameter81819f02007-05-06 14:49:36 -07002949/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002950 * Object placement in a slab is made very easy because we always start at
2951 * offset 0. If we tune the size of the object to the alignment then we can
2952 * get the required alignment by putting one properly sized object after
2953 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002954 *
2955 * Notice that the allocation order determines the sizes of the per cpu
2956 * caches. Each processor has always one slab available for allocations.
2957 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07002958 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07002959 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07002960 */
2961
2962/*
2963 * Mininum / Maximum order of slab pages. This influences locking overhead
2964 * and slab fragmentation. A higher order reduces the number of partial slabs
2965 * and increases the number of allocations possible without having to
2966 * take the list_lock.
2967 */
2968static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03002969static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03002970static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07002971
2972/*
Christoph Lameter81819f02007-05-06 14:49:36 -07002973 * Calculate the order of allocation given an slab object size.
2974 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002975 * The order of allocation has significant impact on performance and other
2976 * system components. Generally order 0 allocations should be preferred since
2977 * order 0 does not cause fragmentation in the page allocator. Larger objects
2978 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03002979 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07002980 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07002981 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002982 * In order to reach satisfactory performance we must ensure that a minimum
2983 * number of objects is in one slab. Otherwise we may generate too much
2984 * activity on the partial lists which requires taking the list_lock. This is
2985 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07002986 *
Christoph Lameter672bba32007-05-09 02:32:39 -07002987 * slub_max_order specifies the order where we begin to stop considering the
2988 * number of objects in a slab as critical. If we reach slub_max_order then
2989 * we try to keep the page order as low as possible. So we accept more waste
2990 * of space in favor of a small page order.
2991 *
2992 * Higher order allocations also allow the placement of more objects in a
2993 * slab and thereby reduce object handling overhead. If the user has
2994 * requested a higher mininum order then we start with that one instead of
2995 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07002996 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07002997static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08002998 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07002999{
3000 int order;
3001 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07003002 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07003003
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003004 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04003005 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03003006
Wei Yang9f835702015-11-05 18:45:51 -08003007 for (order = max(min_order, get_order(min_objects * size + reserved));
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003008 order <= max_order; order++) {
3009
Christoph Lameter81819f02007-05-06 14:49:36 -07003010 unsigned long slab_size = PAGE_SIZE << order;
3011
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003012 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07003013
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003014 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07003015 break;
Christoph Lameter81819f02007-05-06 14:49:36 -07003016 }
Christoph Lameter672bba32007-05-09 02:32:39 -07003017
Christoph Lameter81819f02007-05-06 14:49:36 -07003018 return order;
3019}
3020
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003021static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003022{
3023 int order;
3024 int min_objects;
3025 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02003026 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003027
3028 /*
3029 * Attempt to find best configuration for a slab. This
3030 * works by first attempting to generate a layout with
3031 * the best configuration and backing off gradually.
3032 *
Wei Yang422ff4d2015-11-05 18:45:46 -08003033 * First we increase the acceptable waste in a slab. Then
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003034 * we reduce the minimum objects required in a slab.
3035 */
3036 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03003037 if (!min_objects)
3038 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003039 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02003040 min_objects = min(min_objects, max_objects);
3041
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003042 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03003043 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003044 while (fraction >= 4) {
3045 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003046 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003047 if (order <= slub_max_order)
3048 return order;
3049 fraction /= 2;
3050 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03003051 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003052 }
3053
3054 /*
3055 * We were unable to place multiple objects in a slab. Now
3056 * lets see if we can place a single object there.
3057 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003058 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003059 if (order <= slub_max_order)
3060 return order;
3061
3062 /*
3063 * Doh this slab cannot be placed using slub_max_order.
3064 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003065 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03003066 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003067 return order;
3068 return -ENOSYS;
3069}
3070
Pekka Enberg5595cff2008-08-05 09:28:47 +03003071static void
Joonsoo Kim40534972012-05-11 00:50:47 +09003072init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003073{
3074 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003075 spin_lock_init(&n->list_lock);
3076 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003077#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003078 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07003079 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07003080 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003081#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003082}
3083
Christoph Lameter55136592010-08-20 12:37:13 -05003084static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003085{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05003086 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00003087 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003088
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003089 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04003090 * Must align to double word boundary for the double cmpxchg
3091 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003092 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04003093 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3094 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003095
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003096 if (!s->cpu_slab)
3097 return 0;
3098
3099 init_kmem_cache_cpus(s);
3100
3101 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003102}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003103
Christoph Lameter51df1142010-08-20 12:37:15 -05003104static struct kmem_cache *kmem_cache_node;
3105
Christoph Lameter81819f02007-05-06 14:49:36 -07003106/*
3107 * No kmalloc_node yet so do it by hand. We know that this is the first
3108 * slab on the node for this slabcache. There are no concurrent accesses
3109 * possible.
3110 *
Zhi Yong Wu721ae222013-11-08 20:47:37 +08003111 * Note that this function only works on the kmem_cache_node
3112 * when allocating for the kmem_cache_node. This is used for bootstrapping
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003113 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07003114 */
Christoph Lameter55136592010-08-20 12:37:13 -05003115static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07003116{
3117 struct page *page;
3118 struct kmem_cache_node *n;
3119
Christoph Lameter51df1142010-08-20 12:37:15 -05003120 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07003121
Christoph Lameter51df1142010-08-20 12:37:15 -05003122 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003123
3124 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07003125 if (page_to_nid(page) != node) {
Fabian Frederickf9f58282014-06-04 16:06:34 -07003126 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3127 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
Christoph Lametera2f92ee2007-08-22 14:01:57 -07003128 }
3129
Christoph Lameter81819f02007-05-06 14:49:36 -07003130 n = page->freelist;
3131 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05003132 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05003133 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05003134 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05003135 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07003136#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05003137 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05003138 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003139#endif
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003140 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
Joonsoo Kim40534972012-05-11 00:50:47 +09003141 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05003142 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003143
Dave Hansen67b6c902014-01-24 07:20:23 -08003144 /*
Steven Rostedt1e4dd942014-02-10 14:25:46 -08003145 * No locks need to be taken here as it has just been
3146 * initialized and there is no concurrent access.
Dave Hansen67b6c902014-01-24 07:20:23 -08003147 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08003148 __add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07003149}
3150
3151static void free_kmem_cache_nodes(struct kmem_cache *s)
3152{
3153 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003154 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07003155
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003156 for_each_kmem_cache_node(s, node, n) {
3157 kmem_cache_free(kmem_cache_node, n);
Christoph Lameter81819f02007-05-06 14:49:36 -07003158 s->node[node] = NULL;
3159 }
3160}
3161
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003162void __kmem_cache_release(struct kmem_cache *s)
3163{
3164 free_percpu(s->cpu_slab);
3165 free_kmem_cache_nodes(s);
3166}
3167
Christoph Lameter55136592010-08-20 12:37:13 -05003168static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003169{
3170 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07003171
Christoph Lameterf64dc582007-10-16 01:25:33 -07003172 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003173 struct kmem_cache_node *n;
3174
Alexander Duyck73367bd2010-05-21 14:41:35 -07003175 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05003176 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07003177 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003178 }
Christoph Lameter51df1142010-08-20 12:37:15 -05003179 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05003180 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07003181
3182 if (!n) {
3183 free_kmem_cache_nodes(s);
3184 return 0;
3185 }
3186
Christoph Lameter81819f02007-05-06 14:49:36 -07003187 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09003188 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07003189 }
3190 return 1;
3191}
Christoph Lameter81819f02007-05-06 14:49:36 -07003192
David Rientjesc0bdb232009-02-25 09:16:35 +02003193static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08003194{
3195 if (min < MIN_PARTIAL)
3196 min = MIN_PARTIAL;
3197 else if (min > MAX_PARTIAL)
3198 min = MAX_PARTIAL;
3199 s->min_partial = min;
3200}
3201
Christoph Lameter81819f02007-05-06 14:49:36 -07003202/*
3203 * calculate_sizes() determines the order and the distribution of data within
3204 * a slab object.
3205 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03003206static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07003207{
3208 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003209 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003210 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07003211
3212 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08003213 * Round up object size to the next word boundary. We can only
3214 * place the free pointer at word boundaries and this determines
3215 * the possible location of the free pointer.
3216 */
3217 size = ALIGN(size, sizeof(void *));
3218
3219#ifdef CONFIG_SLUB_DEBUG
3220 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003221 * Determine if we can poison the object itself. If the user of
3222 * the slab may touch the object after free or before allocation
3223 * then we should never poison the object itself.
3224 */
3225 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07003226 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003227 s->flags |= __OBJECT_POISON;
3228 else
3229 s->flags &= ~__OBJECT_POISON;
3230
Christoph Lameter81819f02007-05-06 14:49:36 -07003231
3232 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003233 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07003234 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07003235 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07003236 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003237 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003238 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003239#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003240
3241 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003242 * With that we have determined the number of bytes in actual use
3243 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07003244 */
3245 s->inuse = size;
3246
3247 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07003248 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003249 /*
3250 * Relocate free pointer after the object if it is not
3251 * permitted to overwrite the first word of the object on
3252 * kmem_cache_free.
3253 *
3254 * This is the case if we do RCU, have a constructor or
3255 * destructor or are poisoning the objects.
3256 */
3257 s->offset = size;
3258 size += sizeof(void *);
3259 }
3260
Christoph Lameterc12b3c62007-05-23 13:57:31 -07003261#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07003262 if (flags & SLAB_STORE_USER)
3263 /*
3264 * Need to store information about allocs and frees after
3265 * the object.
3266 */
3267 size += 2 * sizeof(struct track);
3268
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07003269 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07003270 /*
3271 * Add some empty padding so that we can catch
3272 * overwrites from earlier objects rather than let
3273 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01003274 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07003275 * of the object.
3276 */
3277 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003278#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07003279
Christoph Lameter81819f02007-05-06 14:49:36 -07003280 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003281 * SLUB stores one object immediately after another beginning from
3282 * offset 0. In order to align the objects we have to simply size
3283 * each object to conform to the alignment.
3284 */
Christoph Lameter45906852012-11-28 16:23:16 +00003285 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07003286 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003287 if (forced_order >= 0)
3288 order = forced_order;
3289 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003290 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07003291
Christoph Lameter834f3d12008-04-14 19:11:31 +03003292 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003293 return 0;
3294
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003295 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003296 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003297 s->allocflags |= __GFP_COMP;
3298
3299 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003300 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003301
3302 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3303 s->allocflags |= __GFP_RECLAIMABLE;
3304
Christoph Lameter81819f02007-05-06 14:49:36 -07003305 /*
3306 * Determine the number of objects per slab
3307 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003308 s->oo = oo_make(order, size, s->reserved);
3309 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003310 if (oo_objects(s->oo) > oo_objects(s->max))
3311 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003312
Christoph Lameter834f3d12008-04-14 19:11:31 +03003313 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003314}
3315
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003316static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003317{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003318 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003319 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003320
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003321 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3322 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003323
Christoph Lameter06b285d2008-04-14 19:11:41 +03003324 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003325 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003326 if (disable_higher_order_debug) {
3327 /*
3328 * Disable debugging flags that store metadata if the min slab
3329 * order increased.
3330 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003331 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003332 s->flags &= ~DEBUG_METADATA_FLAGS;
3333 s->offset = 0;
3334 if (!calculate_sizes(s, -1))
3335 goto error;
3336 }
3337 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003338
Heiko Carstens25654092012-01-12 17:17:33 -08003339#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3340 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003341 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3342 /* Enable fast mode */
3343 s->flags |= __CMPXCHG_DOUBLE;
3344#endif
3345
David Rientjes3b89d7d2009-02-22 17:40:07 -08003346 /*
3347 * The larger the object size is, the more pages we want on the partial
3348 * list to avoid pounding the page allocator excessively.
3349 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003350 set_min_partial(s, ilog2(s->size) / 2);
3351
3352 /*
3353 * cpu_partial determined the maximum number of objects kept in the
3354 * per cpu partial lists of a processor.
3355 *
3356 * Per cpu partial lists mainly contain slabs that just have one
3357 * object freed. If they are used for allocation then they can be
3358 * filled up again with minimal effort. The slab will never hit the
3359 * per node partial lists and therefore no locking will be required.
3360 *
3361 * This setting also determines
3362 *
3363 * A) The number of objects from per cpu partial slabs dumped to the
3364 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003365 * B) The number of objects in cpu partial slabs to extract from the
Chen Gangd0e0ac92013-07-15 09:05:29 +08003366 * per node list when we run out of per cpu objects. We only fetch
3367 * 50% to keep some capacity around for frees.
Christoph Lameter49e22582011-08-09 16:12:27 -05003368 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003369 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003370 s->cpu_partial = 0;
3371 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003372 s->cpu_partial = 2;
3373 else if (s->size >= 1024)
3374 s->cpu_partial = 6;
3375 else if (s->size >= 256)
3376 s->cpu_partial = 13;
3377 else
3378 s->cpu_partial = 30;
3379
Christoph Lameter81819f02007-05-06 14:49:36 -07003380#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003381 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003382#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003383 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003384 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003385
Christoph Lameter55136592010-08-20 12:37:13 -05003386 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003387 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003388
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003389 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003390error:
3391 if (flags & SLAB_PANIC)
3392 panic("Cannot create slab %s size=%lu realsize=%u "
3393 "order=%u offset=%u flags=%lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +08003394 s->name, (unsigned long)s->size, s->size,
3395 oo_order(s->oo), s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003396 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003397}
Christoph Lameter81819f02007-05-06 14:49:36 -07003398
Christoph Lameter33b12c32008-04-25 12:22:43 -07003399static void list_slab_objects(struct kmem_cache *s, struct page *page,
3400 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003401{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003402#ifdef CONFIG_SLUB_DEBUG
3403 void *addr = page_address(page);
3404 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003405 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3406 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003407 if (!map)
3408 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003409 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003410 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003411
Christoph Lameter5f80b132011-04-15 14:48:13 -05003412 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003413 for_each_object(p, s, addr, page->objects) {
3414
3415 if (!test_bit(slab_index(p, s, addr), map)) {
Fabian Frederickf9f58282014-06-04 16:06:34 -07003416 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003417 print_tracking(s, p);
3418 }
3419 }
3420 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003421 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003422#endif
3423}
3424
Christoph Lameter81819f02007-05-06 14:49:36 -07003425/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003426 * Attempt to free all partial slabs on a node.
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003427 * This is called from __kmem_cache_shutdown(). We must take list_lock
3428 * because sysfs file might still access partial list after the shutdowning.
Christoph Lameter81819f02007-05-06 14:49:36 -07003429 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003430static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003431{
Christoph Lameter81819f02007-05-06 14:49:36 -07003432 struct page *page, *h;
3433
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003434 BUG_ON(irqs_disabled());
3435 spin_lock_irq(&n->list_lock);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003436 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003437 if (!page->inuse) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003438 remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003439 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003440 } else {
3441 list_slab_objects(s, page,
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003442 "Objects remaining in %s on __kmem_cache_shutdown()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003443 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003444 }
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003445 spin_unlock_irq(&n->list_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003446}
3447
3448/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003449 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003450 */
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003451int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003452{
3453 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003454 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07003455
3456 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003457 /* Attempt to free all objects */
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003458 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter599870b2008-04-23 12:36:52 -07003459 free_partial(s, n);
3460 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003461 return 1;
3462 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003463 return 0;
3464}
3465
Christoph Lameter81819f02007-05-06 14:49:36 -07003466/********************************************************************
3467 * Kmalloc subsystem
3468 *******************************************************************/
3469
Christoph Lameter81819f02007-05-06 14:49:36 -07003470static int __init setup_slub_min_order(char *str)
3471{
Pekka Enberg06428782008-01-07 23:20:27 -08003472 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003473
3474 return 1;
3475}
3476
3477__setup("slub_min_order=", setup_slub_min_order);
3478
3479static int __init setup_slub_max_order(char *str)
3480{
Pekka Enberg06428782008-01-07 23:20:27 -08003481 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003482 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003483
3484 return 1;
3485}
3486
3487__setup("slub_max_order=", setup_slub_max_order);
3488
3489static int __init setup_slub_min_objects(char *str)
3490{
Pekka Enberg06428782008-01-07 23:20:27 -08003491 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003492
3493 return 1;
3494}
3495
3496__setup("slub_min_objects=", setup_slub_min_objects);
3497
Christoph Lameter81819f02007-05-06 14:49:36 -07003498void *__kmalloc(size_t size, gfp_t flags)
3499{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003500 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003501 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003502
Christoph Lameter95a05b42013-01-10 19:14:19 +00003503 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003504 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003505
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003506 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003507
3508 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003509 return s;
3510
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003511 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003512
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003513 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003514
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003515 kasan_kmalloc(s, ret, size);
3516
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003517 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003518}
3519EXPORT_SYMBOL(__kmalloc);
3520
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003521#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003522static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3523{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003524 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003525 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003526
Vladimir Davydov52383432014-06-04 16:06:39 -07003527 flags |= __GFP_COMP | __GFP_NOTRACK;
3528 page = alloc_kmem_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003529 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003530 ptr = page_address(page);
3531
Roman Bobnievd56791b2013-10-08 15:58:57 -07003532 kmalloc_large_node_hook(ptr, size, flags);
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003533 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003534}
3535
Christoph Lameter81819f02007-05-06 14:49:36 -07003536void *__kmalloc_node(size_t size, gfp_t flags, int node)
3537{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003538 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003539 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003540
Christoph Lameter95a05b42013-01-10 19:14:19 +00003541 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003542 ret = kmalloc_large_node(size, flags, node);
3543
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003544 trace_kmalloc_node(_RET_IP_, ret,
3545 size, PAGE_SIZE << get_order(size),
3546 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003547
3548 return ret;
3549 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003550
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003551 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003552
3553 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003554 return s;
3555
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003556 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003557
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003558 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003559
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003560 kasan_kmalloc(s, ret, size);
3561
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003562 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003563}
3564EXPORT_SYMBOL(__kmalloc_node);
3565#endif
3566
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003567static size_t __ksize(const void *object)
Christoph Lameter81819f02007-05-06 14:49:36 -07003568{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003569 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003570
Christoph Lameteref8b4522007-10-16 01:24:46 -07003571 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003572 return 0;
3573
Vegard Nossum294a80a2007-12-04 23:45:30 -08003574 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003575
Pekka Enberg76994412008-05-22 19:22:25 +03003576 if (unlikely(!PageSlab(page))) {
3577 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003578 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003579 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003580
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003581 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003582}
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003583
3584size_t ksize(const void *object)
3585{
3586 size_t size = __ksize(object);
3587 /* We assume that ksize callers could use whole allocated area,
3588 so we need unpoison this area. */
3589 kasan_krealloc(object, size);
3590 return size;
3591}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003592EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003593
3594void kfree(const void *x)
3595{
Christoph Lameter81819f02007-05-06 14:49:36 -07003596 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003597 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003598
Pekka Enberg2121db72009-03-25 11:05:57 +02003599 trace_kfree(_RET_IP_, x);
3600
Satyam Sharma2408c552007-10-16 01:24:44 -07003601 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003602 return;
3603
Christoph Lameterb49af682007-05-06 14:49:41 -07003604 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003605 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003606 BUG_ON(!PageCompound(page));
Roman Bobnievd56791b2013-10-08 15:58:57 -07003607 kfree_hook(x);
Vladimir Davydov52383432014-06-04 16:06:39 -07003608 __free_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003609 return;
3610 }
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08003611 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003612}
3613EXPORT_SYMBOL(kfree);
3614
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003615#define SHRINK_PROMOTE_MAX 32
3616
Christoph Lameter2086d262007-05-06 14:49:46 -07003617/*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003618 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3619 * up most to the head of the partial lists. New allocations will then
3620 * fill those up and thus they can be removed from the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -07003621 *
3622 * The slabs with the least items are placed last. This results in them
3623 * being allocated from last increasing the chance that the last objects
3624 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003625 */
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003626int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
Christoph Lameter2086d262007-05-06 14:49:46 -07003627{
3628 int node;
3629 int i;
3630 struct kmem_cache_node *n;
3631 struct page *page;
3632 struct page *t;
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003633 struct list_head discard;
3634 struct list_head promote[SHRINK_PROMOTE_MAX];
Christoph Lameter2086d262007-05-06 14:49:46 -07003635 unsigned long flags;
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003636 int ret = 0;
Christoph Lameter2086d262007-05-06 14:49:46 -07003637
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003638 if (deactivate) {
3639 /*
3640 * Disable empty slabs caching. Used to avoid pinning offline
3641 * memory cgroups by kmem pages that can be freed.
3642 */
3643 s->cpu_partial = 0;
3644 s->min_partial = 0;
3645
3646 /*
3647 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3648 * so we have to make sure the change is visible.
3649 */
3650 kick_all_cpus_sync();
3651 }
3652
Christoph Lameter2086d262007-05-06 14:49:46 -07003653 flush_all(s);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003654 for_each_kmem_cache_node(s, node, n) {
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003655 INIT_LIST_HEAD(&discard);
3656 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3657 INIT_LIST_HEAD(promote + i);
Christoph Lameter2086d262007-05-06 14:49:46 -07003658
3659 spin_lock_irqsave(&n->list_lock, flags);
3660
3661 /*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003662 * Build lists of slabs to discard or promote.
Christoph Lameter2086d262007-05-06 14:49:46 -07003663 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003664 * Note that concurrent frees may occur while we hold the
3665 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003666 */
3667 list_for_each_entry_safe(page, t, &n->partial, lru) {
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003668 int free = page->objects - page->inuse;
3669
3670 /* Do not reread page->inuse */
3671 barrier();
3672
3673 /* We do not keep full slabs on the list */
3674 BUG_ON(free <= 0);
3675
3676 if (free == page->objects) {
3677 list_move(&page->lru, &discard);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003678 n->nr_partial--;
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003679 } else if (free <= SHRINK_PROMOTE_MAX)
3680 list_move(&page->lru, promote + free - 1);
Christoph Lameter2086d262007-05-06 14:49:46 -07003681 }
3682
Christoph Lameter2086d262007-05-06 14:49:46 -07003683 /*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003684 * Promote the slabs filled up most to the head of the
3685 * partial list.
Christoph Lameter2086d262007-05-06 14:49:46 -07003686 */
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003687 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3688 list_splice(promote + i, &n->partial);
Christoph Lameter2086d262007-05-06 14:49:46 -07003689
Christoph Lameter2086d262007-05-06 14:49:46 -07003690 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003691
3692 /* Release empty slabs */
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003693 list_for_each_entry_safe(page, t, &discard, lru)
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003694 discard_slab(s, page);
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003695
3696 if (slabs_node(s, node))
3697 ret = 1;
Christoph Lameter2086d262007-05-06 14:49:46 -07003698 }
3699
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003700 return ret;
Christoph Lameter2086d262007-05-06 14:49:46 -07003701}
Christoph Lameter2086d262007-05-06 14:49:46 -07003702
Yasunori Gotob9049e22007-10-21 16:41:37 -07003703static int slab_mem_going_offline_callback(void *arg)
3704{
3705 struct kmem_cache *s;
3706
Christoph Lameter18004c52012-07-06 15:25:12 -05003707 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003708 list_for_each_entry(s, &slab_caches, list)
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003709 __kmem_cache_shrink(s, false);
Christoph Lameter18004c52012-07-06 15:25:12 -05003710 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003711
3712 return 0;
3713}
3714
3715static void slab_mem_offline_callback(void *arg)
3716{
3717 struct kmem_cache_node *n;
3718 struct kmem_cache *s;
3719 struct memory_notify *marg = arg;
3720 int offline_node;
3721
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003722 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003723
3724 /*
3725 * If the node still has available memory. we need kmem_cache_node
3726 * for it yet.
3727 */
3728 if (offline_node < 0)
3729 return;
3730
Christoph Lameter18004c52012-07-06 15:25:12 -05003731 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003732 list_for_each_entry(s, &slab_caches, list) {
3733 n = get_node(s, offline_node);
3734 if (n) {
3735 /*
3736 * if n->nr_slabs > 0, slabs still exist on the node
3737 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003738 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003739 * callback. So, we must fail.
3740 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003741 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003742
3743 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003744 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003745 }
3746 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003747 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003748}
3749
3750static int slab_mem_going_online_callback(void *arg)
3751{
3752 struct kmem_cache_node *n;
3753 struct kmem_cache *s;
3754 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003755 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003756 int ret = 0;
3757
3758 /*
3759 * If the node's memory is already available, then kmem_cache_node is
3760 * already created. Nothing to do.
3761 */
3762 if (nid < 0)
3763 return 0;
3764
3765 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003766 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003767 * allocate a kmem_cache_node structure in order to bring the node
3768 * online.
3769 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003770 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003771 list_for_each_entry(s, &slab_caches, list) {
3772 /*
3773 * XXX: kmem_cache_alloc_node will fallback to other nodes
3774 * since memory is not yet available from the node that
3775 * is brought up.
3776 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003777 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003778 if (!n) {
3779 ret = -ENOMEM;
3780 goto out;
3781 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003782 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003783 s->node[nid] = n;
3784 }
3785out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003786 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003787 return ret;
3788}
3789
3790static int slab_memory_callback(struct notifier_block *self,
3791 unsigned long action, void *arg)
3792{
3793 int ret = 0;
3794
3795 switch (action) {
3796 case MEM_GOING_ONLINE:
3797 ret = slab_mem_going_online_callback(arg);
3798 break;
3799 case MEM_GOING_OFFLINE:
3800 ret = slab_mem_going_offline_callback(arg);
3801 break;
3802 case MEM_OFFLINE:
3803 case MEM_CANCEL_ONLINE:
3804 slab_mem_offline_callback(arg);
3805 break;
3806 case MEM_ONLINE:
3807 case MEM_CANCEL_OFFLINE:
3808 break;
3809 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003810 if (ret)
3811 ret = notifier_from_errno(ret);
3812 else
3813 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003814 return ret;
3815}
3816
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003817static struct notifier_block slab_memory_callback_nb = {
3818 .notifier_call = slab_memory_callback,
3819 .priority = SLAB_CALLBACK_PRI,
3820};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003821
Christoph Lameter81819f02007-05-06 14:49:36 -07003822/********************************************************************
3823 * Basic setup of slabs
3824 *******************************************************************/
3825
Christoph Lameter51df1142010-08-20 12:37:15 -05003826/*
3827 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003828 * the page allocator. Allocate them properly then fix up the pointers
3829 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003830 */
3831
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003832static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003833{
3834 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003835 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003836 struct kmem_cache_node *n;
Christoph Lameter51df1142010-08-20 12:37:15 -05003837
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003838 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003839
Glauber Costa7d557b32013-02-22 20:20:00 +04003840 /*
3841 * This runs very early, and only the boot processor is supposed to be
3842 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3843 * IPIs around.
3844 */
3845 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003846 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter51df1142010-08-20 12:37:15 -05003847 struct page *p;
3848
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003849 list_for_each_entry(p, &n->partial, lru)
3850 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003851
Li Zefan607bf322011-04-12 15:22:26 +08003852#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003853 list_for_each_entry(p, &n->full, lru)
3854 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003855#endif
Christoph Lameter51df1142010-08-20 12:37:15 -05003856 }
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08003857 slab_init_memcg_params(s);
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003858 list_add(&s->list, &slab_caches);
3859 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003860}
3861
Christoph Lameter81819f02007-05-06 14:49:36 -07003862void __init kmem_cache_init(void)
3863{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003864 static __initdata struct kmem_cache boot_kmem_cache,
3865 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003866
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003867 if (debug_guardpage_minorder())
3868 slub_max_order = 0;
3869
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003870 kmem_cache_node = &boot_kmem_cache_node;
3871 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003872
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003873 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3874 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003875
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003876 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003877
3878 /* Able to allocate the per node structures */
3879 slab_state = PARTIAL;
3880
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003881 create_boot_cache(kmem_cache, "kmem_cache",
3882 offsetof(struct kmem_cache, node) +
3883 nr_node_ids * sizeof(struct kmem_cache_node *),
3884 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003885
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003886 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003887
Christoph Lameter51df1142010-08-20 12:37:15 -05003888 /*
3889 * Allocate kmem_cache_node properly from the kmem_cache slab.
3890 * kmem_cache_node is separately allocated so no need to
3891 * update any list pointers.
3892 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003893 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003894
3895 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Daniel Sanders34cc6992015-06-24 16:55:57 -07003896 setup_kmalloc_cache_index_table();
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003897 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003898
3899#ifdef CONFIG_SMP
3900 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003901#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003902
Fabian Frederickf9f58282014-06-04 16:06:34 -07003903 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003904 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003905 slub_min_order, slub_max_order, slub_min_objects,
3906 nr_cpu_ids, nr_node_ids);
3907}
3908
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003909void __init kmem_cache_init_late(void)
3910{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003911}
3912
Glauber Costa2633d7a2012-12-18 14:22:34 -08003913struct kmem_cache *
Vladimir Davydova44cb942014-04-07 15:39:23 -07003914__kmem_cache_alias(const char *name, size_t size, size_t align,
3915 unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003916{
Vladimir Davydov426589f2015-02-12 14:59:23 -08003917 struct kmem_cache *s, *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07003918
Vladimir Davydova44cb942014-04-07 15:39:23 -07003919 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003920 if (s) {
3921 s->refcount++;
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003922
Christoph Lameter81819f02007-05-06 14:49:36 -07003923 /*
3924 * Adjust the object sizes so that we clear
3925 * the complete object on kzalloc.
3926 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003927 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003928 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003929
Vladimir Davydov426589f2015-02-12 14:59:23 -08003930 for_each_memcg_cache(c, s) {
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003931 c->object_size = s->object_size;
3932 c->inuse = max_t(int, c->inuse,
3933 ALIGN(size, sizeof(void *)));
3934 }
3935
David Rientjes7b8f3b62008-12-17 22:09:46 -08003936 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003937 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003938 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003939 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003940 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003941
Christoph Lametercbb79692012-09-05 00:18:32 +00003942 return s;
3943}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003944
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003945int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003946{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003947 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003948
Pekka Enbergaac3a162012-09-05 12:07:44 +03003949 err = kmem_cache_open(s, flags);
3950 if (err)
3951 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003952
Christoph Lameter45530c42012-11-28 16:23:07 +00003953 /* Mutex is not taken during early boot */
3954 if (slab_state <= UP)
3955 return 0;
3956
Glauber Costa107dab52012-12-18 14:23:05 -08003957 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003958 err = sysfs_slab_add(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003959 if (err)
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003960 __kmem_cache_release(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03003961
3962 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07003963}
Christoph Lameter81819f02007-05-06 14:49:36 -07003964
Christoph Lameter81819f02007-05-06 14:49:36 -07003965#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07003966/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003967 * Use the cpu notifier to insure that the cpu slabs are flushed when
3968 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07003969 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04003970static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07003971 unsigned long action, void *hcpu)
3972{
3973 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003974 struct kmem_cache *s;
3975 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07003976
3977 switch (action) {
3978 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003979 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07003980 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003981 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05003982 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07003983 list_for_each_entry(s, &slab_caches, list) {
3984 local_irq_save(flags);
3985 __flush_cpu_slab(s, cpu);
3986 local_irq_restore(flags);
3987 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003988 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07003989 break;
3990 default:
3991 break;
3992 }
3993 return NOTIFY_OK;
3994}
3995
Paul Gortmaker0db06282013-06-19 14:53:51 -04003996static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08003997 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08003998};
Christoph Lameter81819f02007-05-06 14:49:36 -07003999
4000#endif
4001
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004002void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07004003{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004004 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004005 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004006
Christoph Lameter95a05b42013-01-10 19:14:19 +00004007 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02004008 return kmalloc_large(size, gfpflags);
4009
Christoph Lameter2c59dd62013-01-10 19:14:19 +00004010 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07004011
Satyam Sharma2408c552007-10-16 01:24:44 -07004012 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004013 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004014
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03004015 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004016
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004017 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02004018 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004019
4020 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07004021}
4022
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09004023#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07004024void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004025 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07004026{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004027 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004028 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004029
Christoph Lameter95a05b42013-01-10 19:14:19 +00004030 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08004031 ret = kmalloc_large_node(size, gfpflags, node);
4032
4033 trace_kmalloc_node(caller, ret,
4034 size, PAGE_SIZE << get_order(size),
4035 gfpflags, node);
4036
4037 return ret;
4038 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02004039
Christoph Lameter2c59dd62013-01-10 19:14:19 +00004040 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07004041
Satyam Sharma2408c552007-10-16 01:24:44 -07004042 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004043 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004044
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03004045 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004046
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004047 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02004048 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004049
4050 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07004051}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09004052#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004053
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004054#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03004055static int count_inuse(struct page *page)
4056{
4057 return page->inuse;
4058}
4059
4060static int count_total(struct page *page)
4061{
4062 return page->objects;
4063}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004064#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03004065
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004066#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07004067static int validate_slab(struct kmem_cache *s, struct page *page,
4068 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004069{
4070 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08004071 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004072
4073 if (!check_slab(s, page) ||
4074 !on_freelist(s, page, NULL))
4075 return 0;
4076
4077 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03004078 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004079
Christoph Lameter5f80b132011-04-15 14:48:13 -05004080 get_map(s, page, map);
4081 for_each_object(p, s, addr, page->objects) {
4082 if (test_bit(slab_index(p, s, addr), map))
4083 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4084 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004085 }
4086
Christoph Lameter224a88b2008-04-14 19:11:31 +03004087 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07004088 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02004089 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07004090 return 0;
4091 return 1;
4092}
4093
Christoph Lameter434e2452007-07-17 04:03:30 -07004094static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4095 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004096{
Christoph Lameter881db7f2011-06-01 12:25:53 -05004097 slab_lock(page);
4098 validate_slab(s, page, map);
4099 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004100}
4101
Christoph Lameter434e2452007-07-17 04:03:30 -07004102static int validate_slab_node(struct kmem_cache *s,
4103 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004104{
4105 unsigned long count = 0;
4106 struct page *page;
4107 unsigned long flags;
4108
4109 spin_lock_irqsave(&n->list_lock, flags);
4110
4111 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07004112 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004113 count++;
4114 }
4115 if (count != n->nr_partial)
Fabian Frederickf9f58282014-06-04 16:06:34 -07004116 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4117 s->name, count, n->nr_partial);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004118
4119 if (!(s->flags & SLAB_STORE_USER))
4120 goto out;
4121
4122 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07004123 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004124 count++;
4125 }
4126 if (count != atomic_long_read(&n->nr_slabs))
Fabian Frederickf9f58282014-06-04 16:06:34 -07004127 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4128 s->name, count, atomic_long_read(&n->nr_slabs));
Christoph Lameter53e15af2007-05-06 14:49:43 -07004129
4130out:
4131 spin_unlock_irqrestore(&n->list_lock, flags);
4132 return count;
4133}
4134
Christoph Lameter434e2452007-07-17 04:03:30 -07004135static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004136{
4137 int node;
4138 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004139 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07004140 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004141 struct kmem_cache_node *n;
Christoph Lameter434e2452007-07-17 04:03:30 -07004142
4143 if (!map)
4144 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004145
4146 flush_all(s);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004147 for_each_kmem_cache_node(s, node, n)
Christoph Lameter434e2452007-07-17 04:03:30 -07004148 count += validate_slab_node(s, n, map);
Christoph Lameter434e2452007-07-17 04:03:30 -07004149 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004150 return count;
4151}
Christoph Lameter88a420e2007-05-06 14:49:45 -07004152/*
Christoph Lameter672bba32007-05-09 02:32:39 -07004153 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07004154 * and freed.
4155 */
4156
4157struct location {
4158 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004159 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004160 long long sum_time;
4161 long min_time;
4162 long max_time;
4163 long min_pid;
4164 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304165 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004166 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004167};
4168
4169struct loc_track {
4170 unsigned long max;
4171 unsigned long count;
4172 struct location *loc;
4173};
4174
4175static void free_loc_track(struct loc_track *t)
4176{
4177 if (t->max)
4178 free_pages((unsigned long)t->loc,
4179 get_order(sizeof(struct location) * t->max));
4180}
4181
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004182static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004183{
4184 struct location *l;
4185 int order;
4186
Christoph Lameter88a420e2007-05-06 14:49:45 -07004187 order = get_order(sizeof(struct location) * max);
4188
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004189 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004190 if (!l)
4191 return 0;
4192
4193 if (t->count) {
4194 memcpy(l, t->loc, sizeof(struct location) * t->count);
4195 free_loc_track(t);
4196 }
4197 t->max = max;
4198 t->loc = l;
4199 return 1;
4200}
4201
4202static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004203 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004204{
4205 long start, end, pos;
4206 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004207 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004208 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004209
4210 start = -1;
4211 end = t->count;
4212
4213 for ( ; ; ) {
4214 pos = start + (end - start + 1) / 2;
4215
4216 /*
4217 * There is nothing at "end". If we end up there
4218 * we need to add something to before end.
4219 */
4220 if (pos == end)
4221 break;
4222
4223 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004224 if (track->addr == caddr) {
4225
4226 l = &t->loc[pos];
4227 l->count++;
4228 if (track->when) {
4229 l->sum_time += age;
4230 if (age < l->min_time)
4231 l->min_time = age;
4232 if (age > l->max_time)
4233 l->max_time = age;
4234
4235 if (track->pid < l->min_pid)
4236 l->min_pid = track->pid;
4237 if (track->pid > l->max_pid)
4238 l->max_pid = track->pid;
4239
Rusty Russell174596a2009-01-01 10:12:29 +10304240 cpumask_set_cpu(track->cpu,
4241 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004242 }
4243 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004244 return 1;
4245 }
4246
Christoph Lameter45edfa52007-05-09 02:32:45 -07004247 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004248 end = pos;
4249 else
4250 start = pos;
4251 }
4252
4253 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004254 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004255 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004256 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004257 return 0;
4258
4259 l = t->loc + pos;
4260 if (pos < t->count)
4261 memmove(l + 1, l,
4262 (t->count - pos) * sizeof(struct location));
4263 t->count++;
4264 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004265 l->addr = track->addr;
4266 l->sum_time = age;
4267 l->min_time = age;
4268 l->max_time = age;
4269 l->min_pid = track->pid;
4270 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304271 cpumask_clear(to_cpumask(l->cpus));
4272 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004273 nodes_clear(l->nodes);
4274 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004275 return 1;
4276}
4277
4278static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004279 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004280 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004281{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004282 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004283 void *p;
4284
Christoph Lameter39b26462008-04-14 19:11:30 +03004285 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004286 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004287
Christoph Lameter224a88b2008-04-14 19:11:31 +03004288 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004289 if (!test_bit(slab_index(p, s, addr), map))
4290 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004291}
4292
4293static int list_locations(struct kmem_cache *s, char *buf,
4294 enum track_item alloc)
4295{
Harvey Harrisone374d482008-01-31 15:20:50 -08004296 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004297 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004298 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004299 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004300 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4301 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004302 struct kmem_cache_node *n;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004303
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004304 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4305 GFP_TEMPORARY)) {
4306 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004307 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004308 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004309 /* Push back cpu slabs */
4310 flush_all(s);
4311
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004312 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004313 unsigned long flags;
4314 struct page *page;
4315
Christoph Lameter9e869432007-08-22 14:01:56 -07004316 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004317 continue;
4318
4319 spin_lock_irqsave(&n->list_lock, flags);
4320 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004321 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004322 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004323 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004324 spin_unlock_irqrestore(&n->list_lock, flags);
4325 }
4326
4327 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004328 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004329
Hugh Dickins9c246242008-12-09 13:14:27 -08004330 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004331 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004332 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004333
4334 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004335 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004336 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004337 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004338
4339 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004340 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004341 l->min_time,
4342 (long)div_u64(l->sum_time, l->count),
4343 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004344 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004345 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004346 l->min_time);
4347
4348 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004349 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004350 l->min_pid, l->max_pid);
4351 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004352 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004353 l->min_pid);
4354
Rusty Russell174596a2009-01-01 10:12:29 +10304355 if (num_online_cpus() > 1 &&
4356 !cpumask_empty(to_cpumask(l->cpus)) &&
Tejun Heo5024c1d2015-02-13 14:37:59 -08004357 len < PAGE_SIZE - 60)
4358 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4359 " cpus=%*pbl",
4360 cpumask_pr_args(to_cpumask(l->cpus)));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004361
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004362 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Tejun Heo5024c1d2015-02-13 14:37:59 -08004363 len < PAGE_SIZE - 60)
4364 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4365 " nodes=%*pbl",
4366 nodemask_pr_args(&l->nodes));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004367
Harvey Harrisone374d482008-01-31 15:20:50 -08004368 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07004369 }
4370
4371 free_loc_track(&t);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004372 kfree(map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004373 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08004374 len += sprintf(buf, "No data\n");
4375 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004376}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004377#endif
Christoph Lameter88a420e2007-05-06 14:49:45 -07004378
Christoph Lametera5a84752010-10-05 13:57:27 -05004379#ifdef SLUB_RESILIENCY_TEST
David Rientjesc07b8182014-08-06 16:04:16 -07004380static void __init resiliency_test(void)
Christoph Lametera5a84752010-10-05 13:57:27 -05004381{
4382 u8 *p;
4383
Christoph Lameter95a05b42013-01-10 19:14:19 +00004384 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
Christoph Lametera5a84752010-10-05 13:57:27 -05004385
Fabian Frederickf9f58282014-06-04 16:06:34 -07004386 pr_err("SLUB resiliency testing\n");
4387 pr_err("-----------------------\n");
4388 pr_err("A. Corruption after allocation\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004389
4390 p = kzalloc(16, GFP_KERNEL);
4391 p[16] = 0x12;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004392 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4393 p + 16);
Christoph Lametera5a84752010-10-05 13:57:27 -05004394
4395 validate_slab_cache(kmalloc_caches[4]);
4396
4397 /* Hmmm... The next two are dangerous */
4398 p = kzalloc(32, GFP_KERNEL);
4399 p[32 + sizeof(void *)] = 0x34;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004400 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4401 p);
4402 pr_err("If allocated object is overwritten then not detectable\n\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004403
4404 validate_slab_cache(kmalloc_caches[5]);
4405 p = kzalloc(64, GFP_KERNEL);
4406 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4407 *p = 0x56;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004408 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4409 p);
4410 pr_err("If allocated object is overwritten then not detectable\n\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004411 validate_slab_cache(kmalloc_caches[6]);
4412
Fabian Frederickf9f58282014-06-04 16:06:34 -07004413 pr_err("\nB. Corruption after free\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004414 p = kzalloc(128, GFP_KERNEL);
4415 kfree(p);
4416 *p = 0x78;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004417 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004418 validate_slab_cache(kmalloc_caches[7]);
4419
4420 p = kzalloc(256, GFP_KERNEL);
4421 kfree(p);
4422 p[50] = 0x9a;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004423 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004424 validate_slab_cache(kmalloc_caches[8]);
4425
4426 p = kzalloc(512, GFP_KERNEL);
4427 kfree(p);
4428 p[512] = 0xab;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004429 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004430 validate_slab_cache(kmalloc_caches[9]);
4431}
4432#else
4433#ifdef CONFIG_SYSFS
4434static void resiliency_test(void) {};
4435#endif
4436#endif
4437
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004438#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -07004439enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03004440 SL_ALL, /* All slabs */
4441 SL_PARTIAL, /* Only partially allocated slabs */
4442 SL_CPU, /* Only slabs used for cpu caches */
4443 SL_OBJECTS, /* Determine allocated objects not slabs */
4444 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07004445};
4446
Christoph Lameter205ab992008-04-14 19:11:40 +03004447#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004448#define SO_PARTIAL (1 << SL_PARTIAL)
4449#define SO_CPU (1 << SL_CPU)
4450#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03004451#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004452
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004453static ssize_t show_slab_objects(struct kmem_cache *s,
4454 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07004455{
4456 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07004457 int node;
4458 int x;
4459 unsigned long *nodes;
Christoph Lameter81819f02007-05-06 14:49:36 -07004460
Chen Gange35e1a92013-07-12 08:23:48 +08004461 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004462 if (!nodes)
4463 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004464
Christoph Lameter205ab992008-04-14 19:11:40 +03004465 if (flags & SO_CPU) {
4466 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004467
Christoph Lameter205ab992008-04-14 19:11:40 +03004468 for_each_possible_cpu(cpu) {
Chen Gangd0e0ac92013-07-15 09:05:29 +08004469 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4470 cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004471 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004472 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004473
Jason Low4db0c3c2015-04-15 16:14:08 -07004474 page = READ_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004475 if (!page)
4476 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004477
Christoph Lameterec3ab082012-05-09 10:09:56 -05004478 node = page_to_nid(page);
4479 if (flags & SO_TOTAL)
4480 x = page->objects;
4481 else if (flags & SO_OBJECTS)
4482 x = page->inuse;
4483 else
4484 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004485
Christoph Lameterec3ab082012-05-09 10:09:56 -05004486 total += x;
4487 nodes[node] += x;
4488
Jason Low4db0c3c2015-04-15 16:14:08 -07004489 page = READ_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004490 if (page) {
Li Zefan8afb1472013-09-10 11:43:37 +08004491 node = page_to_nid(page);
4492 if (flags & SO_TOTAL)
4493 WARN_ON_ONCE(1);
4494 else if (flags & SO_OBJECTS)
4495 WARN_ON_ONCE(1);
4496 else
4497 x = page->pages;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004498 total += x;
4499 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004500 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004501 }
4502 }
4503
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004504 get_online_mems();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004505#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004506 if (flags & SO_ALL) {
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004507 struct kmem_cache_node *n;
4508
4509 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004510
Chen Gangd0e0ac92013-07-15 09:05:29 +08004511 if (flags & SO_TOTAL)
4512 x = atomic_long_read(&n->total_objects);
4513 else if (flags & SO_OBJECTS)
4514 x = atomic_long_read(&n->total_objects) -
4515 count_partial(n, count_free);
Christoph Lameter205ab992008-04-14 19:11:40 +03004516 else
4517 x = atomic_long_read(&n->nr_slabs);
4518 total += x;
4519 nodes[node] += x;
4520 }
4521
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004522 } else
4523#endif
4524 if (flags & SO_PARTIAL) {
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004525 struct kmem_cache_node *n;
Christoph Lameter205ab992008-04-14 19:11:40 +03004526
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004527 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004528 if (flags & SO_TOTAL)
4529 x = count_partial(n, count_total);
4530 else if (flags & SO_OBJECTS)
4531 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004532 else
4533 x = n->nr_partial;
4534 total += x;
4535 nodes[node] += x;
4536 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004537 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004538 x = sprintf(buf, "%lu", total);
4539#ifdef CONFIG_NUMA
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004540 for (node = 0; node < nr_node_ids; node++)
Christoph Lameter81819f02007-05-06 14:49:36 -07004541 if (nodes[node])
4542 x += sprintf(buf + x, " N%d=%lu",
4543 node, nodes[node]);
4544#endif
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004545 put_online_mems();
Christoph Lameter81819f02007-05-06 14:49:36 -07004546 kfree(nodes);
4547 return x + sprintf(buf + x, "\n");
4548}
4549
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004550#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004551static int any_slab_objects(struct kmem_cache *s)
4552{
4553 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004554 struct kmem_cache_node *n;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004555
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004556 for_each_kmem_cache_node(s, node, n)
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004557 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004558 return 1;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004559
Christoph Lameter81819f02007-05-06 14:49:36 -07004560 return 0;
4561}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004562#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004563
4564#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004565#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004566
4567struct slab_attribute {
4568 struct attribute attr;
4569 ssize_t (*show)(struct kmem_cache *s, char *buf);
4570 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4571};
4572
4573#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004574 static struct slab_attribute _name##_attr = \
4575 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004576
4577#define SLAB_ATTR(_name) \
4578 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004579 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004580
Christoph Lameter81819f02007-05-06 14:49:36 -07004581static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4582{
4583 return sprintf(buf, "%d\n", s->size);
4584}
4585SLAB_ATTR_RO(slab_size);
4586
4587static ssize_t align_show(struct kmem_cache *s, char *buf)
4588{
4589 return sprintf(buf, "%d\n", s->align);
4590}
4591SLAB_ATTR_RO(align);
4592
4593static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4594{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004595 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004596}
4597SLAB_ATTR_RO(object_size);
4598
4599static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4600{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004601 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004602}
4603SLAB_ATTR_RO(objs_per_slab);
4604
Christoph Lameter06b285d2008-04-14 19:11:41 +03004605static ssize_t order_store(struct kmem_cache *s,
4606 const char *buf, size_t length)
4607{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004608 unsigned long order;
4609 int err;
4610
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004611 err = kstrtoul(buf, 10, &order);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004612 if (err)
4613 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004614
4615 if (order > slub_max_order || order < slub_min_order)
4616 return -EINVAL;
4617
4618 calculate_sizes(s, order);
4619 return length;
4620}
4621
Christoph Lameter81819f02007-05-06 14:49:36 -07004622static ssize_t order_show(struct kmem_cache *s, char *buf)
4623{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004624 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004625}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004626SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004627
David Rientjes73d342b2009-02-22 17:40:09 -08004628static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4629{
4630 return sprintf(buf, "%lu\n", s->min_partial);
4631}
4632
4633static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4634 size_t length)
4635{
4636 unsigned long min;
4637 int err;
4638
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004639 err = kstrtoul(buf, 10, &min);
David Rientjes73d342b2009-02-22 17:40:09 -08004640 if (err)
4641 return err;
4642
David Rientjesc0bdb232009-02-25 09:16:35 +02004643 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004644 return length;
4645}
4646SLAB_ATTR(min_partial);
4647
Christoph Lameter49e22582011-08-09 16:12:27 -05004648static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4649{
4650 return sprintf(buf, "%u\n", s->cpu_partial);
4651}
4652
4653static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4654 size_t length)
4655{
4656 unsigned long objects;
4657 int err;
4658
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004659 err = kstrtoul(buf, 10, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05004660 if (err)
4661 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004662 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004663 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004664
4665 s->cpu_partial = objects;
4666 flush_all(s);
4667 return length;
4668}
4669SLAB_ATTR(cpu_partial);
4670
Christoph Lameter81819f02007-05-06 14:49:36 -07004671static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4672{
Joe Perches62c70bc2011-01-13 15:45:52 -08004673 if (!s->ctor)
4674 return 0;
4675 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004676}
4677SLAB_ATTR_RO(ctor);
4678
Christoph Lameter81819f02007-05-06 14:49:36 -07004679static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4680{
Gu Zheng4307c142014-08-06 16:04:51 -07004681 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004682}
4683SLAB_ATTR_RO(aliases);
4684
Christoph Lameter81819f02007-05-06 14:49:36 -07004685static ssize_t partial_show(struct kmem_cache *s, char *buf)
4686{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004687 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004688}
4689SLAB_ATTR_RO(partial);
4690
4691static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4692{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004693 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004694}
4695SLAB_ATTR_RO(cpu_slabs);
4696
4697static ssize_t objects_show(struct kmem_cache *s, char *buf)
4698{
Christoph Lameter205ab992008-04-14 19:11:40 +03004699 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004700}
4701SLAB_ATTR_RO(objects);
4702
Christoph Lameter205ab992008-04-14 19:11:40 +03004703static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4704{
4705 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4706}
4707SLAB_ATTR_RO(objects_partial);
4708
Christoph Lameter49e22582011-08-09 16:12:27 -05004709static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4710{
4711 int objects = 0;
4712 int pages = 0;
4713 int cpu;
4714 int len;
4715
4716 for_each_online_cpu(cpu) {
4717 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4718
4719 if (page) {
4720 pages += page->pages;
4721 objects += page->pobjects;
4722 }
4723 }
4724
4725 len = sprintf(buf, "%d(%d)", objects, pages);
4726
4727#ifdef CONFIG_SMP
4728 for_each_online_cpu(cpu) {
4729 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4730
4731 if (page && len < PAGE_SIZE - 20)
4732 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4733 page->pobjects, page->pages);
4734 }
4735#endif
4736 return len + sprintf(buf + len, "\n");
4737}
4738SLAB_ATTR_RO(slabs_cpu_partial);
4739
Christoph Lameter81819f02007-05-06 14:49:36 -07004740static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4741{
4742 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4743}
4744
4745static ssize_t reclaim_account_store(struct kmem_cache *s,
4746 const char *buf, size_t length)
4747{
4748 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4749 if (buf[0] == '1')
4750 s->flags |= SLAB_RECLAIM_ACCOUNT;
4751 return length;
4752}
4753SLAB_ATTR(reclaim_account);
4754
4755static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4756{
Christoph Lameter5af60832007-05-06 14:49:56 -07004757 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004758}
4759SLAB_ATTR_RO(hwcache_align);
4760
4761#ifdef CONFIG_ZONE_DMA
4762static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4763{
4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4765}
4766SLAB_ATTR_RO(cache_dma);
4767#endif
4768
4769static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4770{
4771 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4772}
4773SLAB_ATTR_RO(destroy_by_rcu);
4774
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004775static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4776{
4777 return sprintf(buf, "%d\n", s->reserved);
4778}
4779SLAB_ATTR_RO(reserved);
4780
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004781#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004782static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4783{
4784 return show_slab_objects(s, buf, SO_ALL);
4785}
4786SLAB_ATTR_RO(slabs);
4787
4788static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4789{
4790 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4791}
4792SLAB_ATTR_RO(total_objects);
4793
4794static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4795{
Laura Abbottbecfda62016-03-15 14:55:06 -07004796 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
Christoph Lametera5a84752010-10-05 13:57:27 -05004797}
4798
4799static ssize_t sanity_checks_store(struct kmem_cache *s,
4800 const char *buf, size_t length)
4801{
Laura Abbottbecfda62016-03-15 14:55:06 -07004802 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004803 if (buf[0] == '1') {
4804 s->flags &= ~__CMPXCHG_DOUBLE;
Laura Abbottbecfda62016-03-15 14:55:06 -07004805 s->flags |= SLAB_CONSISTENCY_CHECKS;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004806 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004807 return length;
4808}
4809SLAB_ATTR(sanity_checks);
4810
4811static ssize_t trace_show(struct kmem_cache *s, char *buf)
4812{
4813 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4814}
4815
4816static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4817 size_t length)
4818{
Christoph Lameterc9e16132014-10-09 15:26:11 -07004819 /*
4820 * Tracing a merged cache is going to give confusing results
4821 * as well as cause other issues like converting a mergeable
4822 * cache into an umergeable one.
4823 */
4824 if (s->refcount > 1)
4825 return -EINVAL;
4826
Christoph Lametera5a84752010-10-05 13:57:27 -05004827 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004828 if (buf[0] == '1') {
4829 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004830 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004831 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004832 return length;
4833}
4834SLAB_ATTR(trace);
4835
Christoph Lameter81819f02007-05-06 14:49:36 -07004836static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4837{
4838 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4839}
4840
4841static ssize_t red_zone_store(struct kmem_cache *s,
4842 const char *buf, size_t length)
4843{
4844 if (any_slab_objects(s))
4845 return -EBUSY;
4846
4847 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004848 if (buf[0] == '1') {
4849 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004850 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004851 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004852 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004853 return length;
4854}
4855SLAB_ATTR(red_zone);
4856
4857static ssize_t poison_show(struct kmem_cache *s, char *buf)
4858{
4859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4860}
4861
4862static ssize_t poison_store(struct kmem_cache *s,
4863 const char *buf, size_t length)
4864{
4865 if (any_slab_objects(s))
4866 return -EBUSY;
4867
4868 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004869 if (buf[0] == '1') {
4870 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004871 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004872 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004873 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004874 return length;
4875}
4876SLAB_ATTR(poison);
4877
4878static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4879{
4880 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4881}
4882
4883static ssize_t store_user_store(struct kmem_cache *s,
4884 const char *buf, size_t length)
4885{
4886 if (any_slab_objects(s))
4887 return -EBUSY;
4888
4889 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004890 if (buf[0] == '1') {
4891 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004892 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004893 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004894 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004895 return length;
4896}
4897SLAB_ATTR(store_user);
4898
Christoph Lameter53e15af2007-05-06 14:49:43 -07004899static ssize_t validate_show(struct kmem_cache *s, char *buf)
4900{
4901 return 0;
4902}
4903
4904static ssize_t validate_store(struct kmem_cache *s,
4905 const char *buf, size_t length)
4906{
Christoph Lameter434e2452007-07-17 04:03:30 -07004907 int ret = -EINVAL;
4908
4909 if (buf[0] == '1') {
4910 ret = validate_slab_cache(s);
4911 if (ret >= 0)
4912 ret = length;
4913 }
4914 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004915}
4916SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004917
4918static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4919{
4920 if (!(s->flags & SLAB_STORE_USER))
4921 return -ENOSYS;
4922 return list_locations(s, buf, TRACK_ALLOC);
4923}
4924SLAB_ATTR_RO(alloc_calls);
4925
4926static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4927{
4928 if (!(s->flags & SLAB_STORE_USER))
4929 return -ENOSYS;
4930 return list_locations(s, buf, TRACK_FREE);
4931}
4932SLAB_ATTR_RO(free_calls);
4933#endif /* CONFIG_SLUB_DEBUG */
4934
4935#ifdef CONFIG_FAILSLAB
4936static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4937{
4938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4939}
4940
4941static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4942 size_t length)
4943{
Christoph Lameterc9e16132014-10-09 15:26:11 -07004944 if (s->refcount > 1)
4945 return -EINVAL;
4946
Christoph Lametera5a84752010-10-05 13:57:27 -05004947 s->flags &= ~SLAB_FAILSLAB;
4948 if (buf[0] == '1')
4949 s->flags |= SLAB_FAILSLAB;
4950 return length;
4951}
4952SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004953#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07004954
Christoph Lameter2086d262007-05-06 14:49:46 -07004955static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4956{
4957 return 0;
4958}
4959
4960static ssize_t shrink_store(struct kmem_cache *s,
4961 const char *buf, size_t length)
4962{
Vladimir Davydov832f37f2015-02-12 14:59:41 -08004963 if (buf[0] == '1')
4964 kmem_cache_shrink(s);
4965 else
Christoph Lameter2086d262007-05-06 14:49:46 -07004966 return -EINVAL;
4967 return length;
4968}
4969SLAB_ATTR(shrink);
4970
Christoph Lameter81819f02007-05-06 14:49:36 -07004971#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08004972static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07004973{
Christoph Lameter98246012008-01-07 23:20:26 -08004974 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07004975}
4976
Christoph Lameter98246012008-01-07 23:20:26 -08004977static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07004978 const char *buf, size_t length)
4979{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004980 unsigned long ratio;
4981 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004982
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004983 err = kstrtoul(buf, 10, &ratio);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004984 if (err)
4985 return err;
4986
Christoph Lametere2cb96b2008-08-19 08:51:22 -05004987 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07004988 s->remote_node_defrag_ratio = ratio * 10;
4989
Christoph Lameter81819f02007-05-06 14:49:36 -07004990 return length;
4991}
Christoph Lameter98246012008-01-07 23:20:26 -08004992SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07004993#endif
4994
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004995#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08004996static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4997{
4998 unsigned long sum = 0;
4999 int cpu;
5000 int len;
5001 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5002
5003 if (!data)
5004 return -ENOMEM;
5005
5006 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06005007 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005008
5009 data[cpu] = x;
5010 sum += x;
5011 }
5012
5013 len = sprintf(buf, "%lu", sum);
5014
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005015#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005016 for_each_online_cpu(cpu) {
5017 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005018 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005019 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005020#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005021 kfree(data);
5022 return len + sprintf(buf + len, "\n");
5023}
5024
David Rientjes78eb00c2009-10-15 02:20:22 -07005025static void clear_stat(struct kmem_cache *s, enum stat_item si)
5026{
5027 int cpu;
5028
5029 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06005030 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07005031}
5032
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005033#define STAT_ATTR(si, text) \
5034static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5035{ \
5036 return show_stat(s, buf, si); \
5037} \
David Rientjes78eb00c2009-10-15 02:20:22 -07005038static ssize_t text##_store(struct kmem_cache *s, \
5039 const char *buf, size_t length) \
5040{ \
5041 if (buf[0] != '0') \
5042 return -EINVAL; \
5043 clear_stat(s, si); \
5044 return length; \
5045} \
5046SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005047
5048STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5049STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5050STAT_ATTR(FREE_FASTPATH, free_fastpath);
5051STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5052STAT_ATTR(FREE_FROZEN, free_frozen);
5053STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5054STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5055STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5056STAT_ATTR(ALLOC_SLAB, alloc_slab);
5057STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05005058STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005059STAT_ATTR(FREE_SLAB, free_slab);
5060STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5061STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5062STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5063STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5064STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5065STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05005066STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03005067STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05005068STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5069STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05005070STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5071STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08005072STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5073STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005074#endif
5075
Pekka Enberg06428782008-01-07 23:20:27 -08005076static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005077 &slab_size_attr.attr,
5078 &object_size_attr.attr,
5079 &objs_per_slab_attr.attr,
5080 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08005081 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005082 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005083 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03005084 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005085 &partial_attr.attr,
5086 &cpu_slabs_attr.attr,
5087 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005088 &aliases_attr.attr,
5089 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005090 &hwcache_align_attr.attr,
5091 &reclaim_account_attr.attr,
5092 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05005093 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08005094 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005095 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005096#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05005097 &total_objects_attr.attr,
5098 &slabs_attr.attr,
5099 &sanity_checks_attr.attr,
5100 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005101 &red_zone_attr.attr,
5102 &poison_attr.attr,
5103 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07005104 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07005105 &alloc_calls_attr.attr,
5106 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005107#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005108#ifdef CONFIG_ZONE_DMA
5109 &cache_dma_attr.attr,
5110#endif
5111#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08005112 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005113#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005114#ifdef CONFIG_SLUB_STATS
5115 &alloc_fastpath_attr.attr,
5116 &alloc_slowpath_attr.attr,
5117 &free_fastpath_attr.attr,
5118 &free_slowpath_attr.attr,
5119 &free_frozen_attr.attr,
5120 &free_add_partial_attr.attr,
5121 &free_remove_partial_attr.attr,
5122 &alloc_from_partial_attr.attr,
5123 &alloc_slab_attr.attr,
5124 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05005125 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005126 &free_slab_attr.attr,
5127 &cpuslab_flush_attr.attr,
5128 &deactivate_full_attr.attr,
5129 &deactivate_empty_attr.attr,
5130 &deactivate_to_head_attr.attr,
5131 &deactivate_to_tail_attr.attr,
5132 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05005133 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03005134 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05005135 &cmpxchg_double_fail_attr.attr,
5136 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005137 &cpu_partial_alloc_attr.attr,
5138 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08005139 &cpu_partial_node_attr.attr,
5140 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005141#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03005142#ifdef CONFIG_FAILSLAB
5143 &failslab_attr.attr,
5144#endif
5145
Christoph Lameter81819f02007-05-06 14:49:36 -07005146 NULL
5147};
5148
5149static struct attribute_group slab_attr_group = {
5150 .attrs = slab_attrs,
5151};
5152
5153static ssize_t slab_attr_show(struct kobject *kobj,
5154 struct attribute *attr,
5155 char *buf)
5156{
5157 struct slab_attribute *attribute;
5158 struct kmem_cache *s;
5159 int err;
5160
5161 attribute = to_slab_attr(attr);
5162 s = to_slab(kobj);
5163
5164 if (!attribute->show)
5165 return -EIO;
5166
5167 err = attribute->show(s, buf);
5168
5169 return err;
5170}
5171
5172static ssize_t slab_attr_store(struct kobject *kobj,
5173 struct attribute *attr,
5174 const char *buf, size_t len)
5175{
5176 struct slab_attribute *attribute;
5177 struct kmem_cache *s;
5178 int err;
5179
5180 attribute = to_slab_attr(attr);
5181 s = to_slab(kobj);
5182
5183 if (!attribute->store)
5184 return -EIO;
5185
5186 err = attribute->store(s, buf, len);
Johannes Weiner127424c2016-01-20 15:02:32 -08005187#ifdef CONFIG_MEMCG
Glauber Costa107dab52012-12-18 14:23:05 -08005188 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
Vladimir Davydov426589f2015-02-12 14:59:23 -08005189 struct kmem_cache *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07005190
Glauber Costa107dab52012-12-18 14:23:05 -08005191 mutex_lock(&slab_mutex);
5192 if (s->max_attr_size < len)
5193 s->max_attr_size = len;
5194
Glauber Costaebe945c2012-12-18 14:23:10 -08005195 /*
5196 * This is a best effort propagation, so this function's return
5197 * value will be determined by the parent cache only. This is
5198 * basically because not all attributes will have a well
5199 * defined semantics for rollbacks - most of the actions will
5200 * have permanent effects.
5201 *
5202 * Returning the error value of any of the children that fail
5203 * is not 100 % defined, in the sense that users seeing the
5204 * error code won't be able to know anything about the state of
5205 * the cache.
5206 *
5207 * Only returning the error code for the parent cache at least
5208 * has well defined semantics. The cache being written to
5209 * directly either failed or succeeded, in which case we loop
5210 * through the descendants with best-effort propagation.
5211 */
Vladimir Davydov426589f2015-02-12 14:59:23 -08005212 for_each_memcg_cache(c, s)
5213 attribute->store(c, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08005214 mutex_unlock(&slab_mutex);
5215 }
5216#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005217 return err;
5218}
5219
Glauber Costa107dab52012-12-18 14:23:05 -08005220static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5221{
Johannes Weiner127424c2016-01-20 15:02:32 -08005222#ifdef CONFIG_MEMCG
Glauber Costa107dab52012-12-18 14:23:05 -08005223 int i;
5224 char *buffer = NULL;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005225 struct kmem_cache *root_cache;
Glauber Costa107dab52012-12-18 14:23:05 -08005226
Vladimir Davydov93030d82014-05-06 12:49:59 -07005227 if (is_root_cache(s))
Glauber Costa107dab52012-12-18 14:23:05 -08005228 return;
5229
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08005230 root_cache = s->memcg_params.root_cache;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005231
Glauber Costa107dab52012-12-18 14:23:05 -08005232 /*
5233 * This mean this cache had no attribute written. Therefore, no point
5234 * in copying default values around
5235 */
Vladimir Davydov93030d82014-05-06 12:49:59 -07005236 if (!root_cache->max_attr_size)
Glauber Costa107dab52012-12-18 14:23:05 -08005237 return;
5238
5239 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5240 char mbuf[64];
5241 char *buf;
5242 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5243
5244 if (!attr || !attr->store || !attr->show)
5245 continue;
5246
5247 /*
5248 * It is really bad that we have to allocate here, so we will
5249 * do it only as a fallback. If we actually allocate, though,
5250 * we can just use the allocated buffer until the end.
5251 *
5252 * Most of the slub attributes will tend to be very small in
5253 * size, but sysfs allows buffers up to a page, so they can
5254 * theoretically happen.
5255 */
5256 if (buffer)
5257 buf = buffer;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005258 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
Glauber Costa107dab52012-12-18 14:23:05 -08005259 buf = mbuf;
5260 else {
5261 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5262 if (WARN_ON(!buffer))
5263 continue;
5264 buf = buffer;
5265 }
5266
Vladimir Davydov93030d82014-05-06 12:49:59 -07005267 attr->show(root_cache, buf);
Glauber Costa107dab52012-12-18 14:23:05 -08005268 attr->store(s, buf, strlen(buf));
5269 }
5270
5271 if (buffer)
5272 free_page((unsigned long)buffer);
5273#endif
5274}
5275
Christoph Lameter41a21282014-05-06 12:50:08 -07005276static void kmem_cache_release(struct kobject *k)
5277{
5278 slab_kmem_cache_release(to_slab(k));
5279}
5280
Emese Revfy52cf25d2010-01-19 02:58:23 +01005281static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005282 .show = slab_attr_show,
5283 .store = slab_attr_store,
5284};
5285
5286static struct kobj_type slab_ktype = {
5287 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter41a21282014-05-06 12:50:08 -07005288 .release = kmem_cache_release,
Christoph Lameter81819f02007-05-06 14:49:36 -07005289};
5290
5291static int uevent_filter(struct kset *kset, struct kobject *kobj)
5292{
5293 struct kobj_type *ktype = get_ktype(kobj);
5294
5295 if (ktype == &slab_ktype)
5296 return 1;
5297 return 0;
5298}
5299
Emese Revfy9cd43612009-12-31 14:52:51 +01005300static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005301 .filter = uevent_filter,
5302};
5303
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005304static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005305
Vladimir Davydov9a417072014-04-07 15:39:31 -07005306static inline struct kset *cache_kset(struct kmem_cache *s)
5307{
Johannes Weiner127424c2016-01-20 15:02:32 -08005308#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005309 if (!is_root_cache(s))
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08005310 return s->memcg_params.root_cache->memcg_kset;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005311#endif
5312 return slab_kset;
5313}
5314
Christoph Lameter81819f02007-05-06 14:49:36 -07005315#define ID_STR_LENGTH 64
5316
5317/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005318 *
5319 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005320 */
5321static char *create_unique_id(struct kmem_cache *s)
5322{
5323 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5324 char *p = name;
5325
5326 BUG_ON(!name);
5327
5328 *p++ = ':';
5329 /*
5330 * First flags affecting slabcache operations. We will only
5331 * get here for aliasable slabs so we do not need to support
5332 * too many flags. The flags here must cover all flags that
5333 * are matched during merging to guarantee that the id is
5334 * unique.
5335 */
5336 if (s->flags & SLAB_CACHE_DMA)
5337 *p++ = 'd';
5338 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5339 *p++ = 'a';
Laura Abbottbecfda62016-03-15 14:55:06 -07005340 if (s->flags & SLAB_CONSISTENCY_CHECKS)
Christoph Lameter81819f02007-05-06 14:49:36 -07005341 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005342 if (!(s->flags & SLAB_NOTRACK))
5343 *p++ = 't';
Vladimir Davydov230e9fc2016-01-14 15:18:15 -08005344 if (s->flags & SLAB_ACCOUNT)
5345 *p++ = 'A';
Christoph Lameter81819f02007-05-06 14:49:36 -07005346 if (p != name + 1)
5347 *p++ = '-';
5348 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005349
Christoph Lameter81819f02007-05-06 14:49:36 -07005350 BUG_ON(p > name + ID_STR_LENGTH - 1);
5351 return name;
5352}
5353
5354static int sysfs_slab_add(struct kmem_cache *s)
5355{
5356 int err;
5357 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005358 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005359
Christoph Lameter81819f02007-05-06 14:49:36 -07005360 if (unmergeable) {
5361 /*
5362 * Slabcache can never be merged so we can use the name proper.
5363 * This is typically the case for debug situations. In that
5364 * case we can catch duplicate names easily.
5365 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005366 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005367 name = s->name;
5368 } else {
5369 /*
5370 * Create a unique name for the slab as a target
5371 * for the symlinks.
5372 */
5373 name = create_unique_id(s);
5374 }
5375
Vladimir Davydov9a417072014-04-07 15:39:31 -07005376 s->kobj.kset = cache_kset(s);
Tetsuo Handa26e4f202014-01-04 16:32:31 +09005377 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
Dave Jones54b6a732014-04-07 15:39:32 -07005378 if (err)
Konstantin Khlebnikov80da0262015-09-04 15:45:51 -07005379 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07005380
5381 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Dave Jones54b6a732014-04-07 15:39:32 -07005382 if (err)
5383 goto out_del_kobj;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005384
Johannes Weiner127424c2016-01-20 15:02:32 -08005385#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005386 if (is_root_cache(s)) {
5387 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5388 if (!s->memcg_kset) {
Dave Jones54b6a732014-04-07 15:39:32 -07005389 err = -ENOMEM;
5390 goto out_del_kobj;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005391 }
5392 }
5393#endif
5394
Christoph Lameter81819f02007-05-06 14:49:36 -07005395 kobject_uevent(&s->kobj, KOBJ_ADD);
5396 if (!unmergeable) {
5397 /* Setup first alias */
5398 sysfs_slab_alias(s, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005399 }
Dave Jones54b6a732014-04-07 15:39:32 -07005400out:
5401 if (!unmergeable)
5402 kfree(name);
5403 return err;
5404out_del_kobj:
5405 kobject_del(&s->kobj);
Dave Jones54b6a732014-04-07 15:39:32 -07005406 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07005407}
5408
Christoph Lameter41a21282014-05-06 12:50:08 -07005409void sysfs_slab_remove(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07005410{
Christoph Lameter97d06602012-07-06 15:25:11 -05005411 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005412 /*
5413 * Sysfs has not been setup yet so no need to remove the
5414 * cache from sysfs.
5415 */
5416 return;
5417
Johannes Weiner127424c2016-01-20 15:02:32 -08005418#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005419 kset_unregister(s->memcg_kset);
5420#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005421 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5422 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005423 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005424}
5425
5426/*
5427 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005428 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005429 */
5430struct saved_alias {
5431 struct kmem_cache *s;
5432 const char *name;
5433 struct saved_alias *next;
5434};
5435
Adrian Bunk5af328a2007-07-17 04:03:27 -07005436static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005437
5438static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5439{
5440 struct saved_alias *al;
5441
Christoph Lameter97d06602012-07-06 15:25:11 -05005442 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005443 /*
5444 * If we have a leftover link then remove it.
5445 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005446 sysfs_remove_link(&slab_kset->kobj, name);
5447 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005448 }
5449
5450 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5451 if (!al)
5452 return -ENOMEM;
5453
5454 al->s = s;
5455 al->name = name;
5456 al->next = alias_list;
5457 alias_list = al;
5458 return 0;
5459}
5460
5461static int __init slab_sysfs_init(void)
5462{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005463 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005464 int err;
5465
Christoph Lameter18004c52012-07-06 15:25:12 -05005466 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005467
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005468 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005469 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005470 mutex_unlock(&slab_mutex);
Fabian Frederickf9f58282014-06-04 16:06:34 -07005471 pr_err("Cannot register slab subsystem.\n");
Christoph Lameter81819f02007-05-06 14:49:36 -07005472 return -ENOSYS;
5473 }
5474
Christoph Lameter97d06602012-07-06 15:25:11 -05005475 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005476
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005477 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005478 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005479 if (err)
Fabian Frederickf9f58282014-06-04 16:06:34 -07005480 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5481 s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005482 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005483
5484 while (alias_list) {
5485 struct saved_alias *al = alias_list;
5486
5487 alias_list = alias_list->next;
5488 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005489 if (err)
Fabian Frederickf9f58282014-06-04 16:06:34 -07005490 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5491 al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005492 kfree(al);
5493 }
5494
Christoph Lameter18004c52012-07-06 15:25:12 -05005495 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005496 resiliency_test();
5497 return 0;
5498}
5499
5500__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005501#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005502
5503/*
5504 * The /proc/slabinfo ABI
5505 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005506#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005507void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005508{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005509 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005510 unsigned long nr_objs = 0;
5511 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005512 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07005513 struct kmem_cache_node *n;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005514
Christoph Lameterfa45dc22014-08-06 16:04:09 -07005515 for_each_kmem_cache_node(s, node, n) {
Wanpeng Lic17fd132013-07-04 08:33:26 +08005516 nr_slabs += node_nr_slabs(n);
5517 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005518 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005519 }
5520
Glauber Costa0d7561c2012-10-19 18:20:27 +04005521 sinfo->active_objs = nr_objs - nr_free;
5522 sinfo->num_objs = nr_objs;
5523 sinfo->active_slabs = nr_slabs;
5524 sinfo->num_slabs = nr_slabs;
5525 sinfo->objects_per_slab = oo_objects(s->oo);
5526 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005527}
5528
Glauber Costa0d7561c2012-10-19 18:20:27 +04005529void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005530{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005531}
5532
Glauber Costab7454ad2012-10-19 18:20:25 +04005533ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5534 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005535{
Glauber Costab7454ad2012-10-19 18:20:25 +04005536 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005537}
Linus Torvalds158a9622008-01-02 13:04:48 -08005538#endif /* CONFIG_SLABINFO */