blob: 4dbb109eb8cd1f5eb03bd82e90285be1090eacd7 [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 Kimd86bd1b2016-03-15 14:55:12 -0700127static inline void *fixup_red_left(struct kmem_cache *s, void *p)
128{
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
131
132 return p;
133}
134
Joonsoo Kim345c9052013-06-19 14:05:52 +0900135static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136{
137#ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
139#else
140 return false;
141#endif
142}
143
Christoph Lameter81819f02007-05-06 14:49:36 -0700144/*
145 * Issues still to be resolved:
146 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700149 * - Variable sizing of the per node arrays
150 */
151
152/* Enable to test recovery from slab corruption on boot */
153#undef SLUB_RESILIENCY_TEST
154
Christoph Lameterb789ef52011-06-01 12:25:49 -0500155/* Enable to log cmpxchg failures */
156#undef SLUB_DEBUG_CMPXCHG
157
Christoph Lameter81819f02007-05-06 14:49:36 -0700158/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 */
Christoph Lameter76be8952007-12-21 14:37:37 -0800162#define MIN_PARTIAL 5
Christoph Lametere95eed52007-05-06 14:49:44 -0700163
Christoph Lameter2086d262007-05-06 14:49:46 -0700164/*
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
Zhi Yong Wu721ae222013-11-08 20:47:37 +0800167 * sort the partial list by the number of objects in use.
Christoph Lameter2086d262007-05-06 14:49:46 -0700168 */
169#define MAX_PARTIAL 10
170
Laura Abbottbecfda62016-03-15 14:55:06 -0700171#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
Christoph Lameter81819f02007-05-06 14:49:36 -0700172 SLAB_POISON | SLAB_STORE_USER)
Christoph Lameter672bba32007-05-09 02:32:39 -0700173
Christoph Lameter81819f02007-05-06 14:49:36 -0700174/*
Laura Abbott149daaf2016-03-15 14:55:09 -0700175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
177 */
178#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
179 SLAB_TRACE)
180
181
182/*
David Rientjes3de47212009-07-27 18:30:35 -0700183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
185 * metadata.
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700186 */
David Rientjes3de47212009-07-27 18:30:35 -0700187#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700188
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400189#define OO_SHIFT 16
190#define OO_MASK ((1 << OO_SHIFT) - 1)
Christoph Lameter50d5c412011-06-01 12:25:45 -0500191#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400192
Christoph Lameter81819f02007-05-06 14:49:36 -0700193/* Internal SLUB flags */
Christoph Lameterf90ec392010-07-09 14:07:11 -0500194#define __OBJECT_POISON 0x80000000UL /* Poison object */
Christoph Lameterb789ef52011-06-01 12:25:49 -0500195#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
Christoph Lameter81819f02007-05-06 14:49:36 -0700196
Christoph Lameter81819f02007-05-06 14:49:36 -0700197#ifdef CONFIG_SMP
198static struct notifier_block slab_notifier;
199#endif
200
Christoph Lameter02cbc872007-05-09 02:32:43 -0700201/*
202 * Tracking user of a slab.
203 */
Ben Greeard6543e32011-07-07 11:36:36 -0700204#define TRACK_ADDRS_COUNT 16
Christoph Lameter02cbc872007-05-09 02:32:43 -0700205struct track {
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300206 unsigned long addr; /* Called from address */
Ben Greeard6543e32011-07-07 11:36:36 -0700207#ifdef CONFIG_STACKTRACE
208 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209#endif
Christoph Lameter02cbc872007-05-09 02:32:43 -0700210 int cpu; /* Was running on cpu */
211 int pid; /* Pid context */
212 unsigned long when; /* When did the operation occur */
213};
214
215enum track_item { TRACK_ALLOC, TRACK_FREE };
216
Christoph Lameterab4d5ed2010-10-05 13:57:26 -0500217#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -0700218static int sysfs_slab_add(struct kmem_cache *);
219static int sysfs_slab_alias(struct kmem_cache *, const char *);
Glauber Costa107dab52012-12-18 14:23:05 -0800220static void memcg_propagate_slab_attrs(struct kmem_cache *s);
Christoph Lameter81819f02007-05-06 14:49:36 -0700221#else
Christoph Lameter0c710012007-07-17 04:03:24 -0700222static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 { return 0; }
Glauber Costa107dab52012-12-18 14:23:05 -0800225static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
Christoph Lameter81819f02007-05-06 14:49:36 -0700226#endif
227
Christoph Lameter4fdccdf2011-03-22 13:35:00 -0500228static inline void stat(const struct kmem_cache *s, enum stat_item si)
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800229{
230#ifdef CONFIG_SLUB_STATS
Christoph Lameter88da03a2014-04-07 15:39:42 -0700231 /*
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
234 */
235 raw_cpu_inc(s->cpu_slab->stat[si]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -0800236#endif
237}
238
Christoph Lameter81819f02007-05-06 14:49:36 -0700239/********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
242
Christoph Lameter7656c722007-05-09 02:32:40 -0700243static inline void *get_freepointer(struct kmem_cache *s, void *object)
244{
245 return *(void **)(object + s->offset);
246}
247
Eric Dumazet0ad95002011-12-16 16:25:34 +0100248static void prefetch_freepointer(const struct kmem_cache *s, void *object)
249{
250 prefetch(object + s->offset);
251}
252
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500253static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
254{
255 void *p;
256
Joonsoo Kim922d5662016-03-17 14:17:53 -0700257 if (!debug_pagealloc_enabled())
258 return get_freepointer(s, object);
259
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
Christoph Lameter1393d9a2011-05-16 15:26:08 -0500261 return p;
262}
263
Christoph Lameter7656c722007-05-09 02:32:40 -0700264static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
265{
266 *(void **)(object + s->offset) = fp;
267}
268
269/* Loop over all objects in a slab */
Christoph Lameter224a88b2008-04-14 19:11:31 +0300270#define for_each_object(__p, __s, __addr, __objects) \
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700271 for (__p = fixup_red_left(__s, __addr); \
272 __p < (__addr) + (__objects) * (__s)->size; \
273 __p += (__s)->size)
Christoph Lameter7656c722007-05-09 02:32:40 -0700274
Wei Yang54266642014-08-06 16:04:42 -0700275#define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700276 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
277 __idx <= __objects; \
278 __p += (__s)->size, __idx++)
Wei Yang54266642014-08-06 16:04:42 -0700279
Christoph Lameter7656c722007-05-09 02:32:40 -0700280/* Determine object index from a given position */
281static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282{
283 return (p - addr) / s->size;
284}
285
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800286static inline int order_objects(int order, unsigned long size, int reserved)
287{
288 return ((PAGE_SIZE << order) - reserved) / size;
289}
290
Christoph Lameter834f3d12008-04-14 19:11:31 +0300291static inline struct kmem_cache_order_objects oo_make(int order,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800292 unsigned long size, int reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300293{
294 struct kmem_cache_order_objects x = {
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800295 (order << OO_SHIFT) + order_objects(order, size, reserved)
Christoph Lameter834f3d12008-04-14 19:11:31 +0300296 };
297
298 return x;
299}
300
301static inline int oo_order(struct kmem_cache_order_objects x)
302{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400303 return x.x >> OO_SHIFT;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300304}
305
306static inline int oo_objects(struct kmem_cache_order_objects x)
307{
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400308 return x.x & OO_MASK;
Christoph Lameter834f3d12008-04-14 19:11:31 +0300309}
310
Christoph Lameter881db7f2011-06-01 12:25:53 -0500311/*
312 * Per slab locking using the pagelock
313 */
314static __always_inline void slab_lock(struct page *page)
315{
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -0800316 VM_BUG_ON_PAGE(PageTail(page), page);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500317 bit_spin_lock(PG_locked, &page->flags);
318}
319
320static __always_inline void slab_unlock(struct page *page)
321{
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -0800322 VM_BUG_ON_PAGE(PageTail(page), page);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500323 __bit_spin_unlock(PG_locked, &page->flags);
324}
325
Dave Hansena0320862014-01-30 15:46:09 -0800326static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
327{
328 struct page tmp;
329 tmp.counters = counters_new;
330 /*
331 * page->counters can cover frozen/inuse/objects as well
332 * as page->_count. If we assign to ->counters directly
333 * we run the risk of losing updates to page->_count, so
334 * be careful and only assign to the fields we need.
335 */
336 page->frozen = tmp.frozen;
337 page->inuse = tmp.inuse;
338 page->objects = tmp.objects;
339}
340
Christoph Lameter1d071712011-07-14 12:49:12 -0500341/* Interrupts must be disabled (for the fallback code to work right) */
342static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
Christoph Lameterb789ef52011-06-01 12:25:49 -0500343 void *freelist_old, unsigned long counters_old,
344 void *freelist_new, unsigned long counters_new,
345 const char *n)
346{
Christoph Lameter1d071712011-07-14 12:49:12 -0500347 VM_BUG_ON(!irqs_disabled());
Heiko Carstens25654092012-01-12 17:17:33 -0800348#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
349 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameterb789ef52011-06-01 12:25:49 -0500350 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000351 if (cmpxchg_double(&page->freelist, &page->counters,
Dan Carpenter0aa9a132014-08-06 16:04:48 -0700352 freelist_old, counters_old,
353 freelist_new, counters_new))
Joe Perches6f6528a2015-04-14 15:44:31 -0700354 return true;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500355 } else
356#endif
357 {
Christoph Lameter881db7f2011-06-01 12:25:53 -0500358 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800359 if (page->freelist == freelist_old &&
360 page->counters == counters_old) {
Christoph Lameterb789ef52011-06-01 12:25:49 -0500361 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800362 set_page_slub_counters(page, counters_new);
Christoph Lameter881db7f2011-06-01 12:25:53 -0500363 slab_unlock(page);
Joe Perches6f6528a2015-04-14 15:44:31 -0700364 return true;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500365 }
Christoph Lameter881db7f2011-06-01 12:25:53 -0500366 slab_unlock(page);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500367 }
368
369 cpu_relax();
370 stat(s, CMPXCHG_DOUBLE_FAIL);
371
372#ifdef SLUB_DEBUG_CMPXCHG
Fabian Frederickf9f58282014-06-04 16:06:34 -0700373 pr_info("%s %s: cmpxchg double redo ", n, s->name);
Christoph Lameterb789ef52011-06-01 12:25:49 -0500374#endif
375
Joe Perches6f6528a2015-04-14 15:44:31 -0700376 return false;
Christoph Lameterb789ef52011-06-01 12:25:49 -0500377}
378
Christoph Lameter1d071712011-07-14 12:49:12 -0500379static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
380 void *freelist_old, unsigned long counters_old,
381 void *freelist_new, unsigned long counters_new,
382 const char *n)
383{
Heiko Carstens25654092012-01-12 17:17:33 -0800384#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
385 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Christoph Lameter1d071712011-07-14 12:49:12 -0500386 if (s->flags & __CMPXCHG_DOUBLE) {
Jan Beulichcdcd6292012-01-02 17:02:18 +0000387 if (cmpxchg_double(&page->freelist, &page->counters,
Dan Carpenter0aa9a132014-08-06 16:04:48 -0700388 freelist_old, counters_old,
389 freelist_new, counters_new))
Joe Perches6f6528a2015-04-14 15:44:31 -0700390 return true;
Christoph Lameter1d071712011-07-14 12:49:12 -0500391 } else
392#endif
393 {
394 unsigned long flags;
395
396 local_irq_save(flags);
397 slab_lock(page);
Chen Gangd0e0ac92013-07-15 09:05:29 +0800398 if (page->freelist == freelist_old &&
399 page->counters == counters_old) {
Christoph Lameter1d071712011-07-14 12:49:12 -0500400 page->freelist = freelist_new;
Dave Hansena0320862014-01-30 15:46:09 -0800401 set_page_slub_counters(page, counters_new);
Christoph Lameter1d071712011-07-14 12:49:12 -0500402 slab_unlock(page);
403 local_irq_restore(flags);
Joe Perches6f6528a2015-04-14 15:44:31 -0700404 return true;
Christoph Lameter1d071712011-07-14 12:49:12 -0500405 }
406 slab_unlock(page);
407 local_irq_restore(flags);
408 }
409
410 cpu_relax();
411 stat(s, CMPXCHG_DOUBLE_FAIL);
412
413#ifdef SLUB_DEBUG_CMPXCHG
Fabian Frederickf9f58282014-06-04 16:06:34 -0700414 pr_info("%s %s: cmpxchg double redo ", n, s->name);
Christoph Lameter1d071712011-07-14 12:49:12 -0500415#endif
416
Joe Perches6f6528a2015-04-14 15:44:31 -0700417 return false;
Christoph Lameter1d071712011-07-14 12:49:12 -0500418}
419
Christoph Lameter41ecc552007-05-09 02:32:44 -0700420#ifdef CONFIG_SLUB_DEBUG
421/*
Christoph Lameter5f80b132011-04-15 14:48:13 -0500422 * Determine a map of object in use on a page.
423 *
Christoph Lameter881db7f2011-06-01 12:25:53 -0500424 * Node listlock must be held to guarantee that the page does
Christoph Lameter5f80b132011-04-15 14:48:13 -0500425 * not vanish from under us.
426 */
427static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
428{
429 void *p;
430 void *addr = page_address(page);
431
432 for (p = page->freelist; p; p = get_freepointer(s, p))
433 set_bit(slab_index(p, s, addr), map);
434}
435
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700436static inline int size_from_object(struct kmem_cache *s)
437{
438 if (s->flags & SLAB_RED_ZONE)
439 return s->size - s->red_left_pad;
440
441 return s->size;
442}
443
444static inline void *restore_red_left(struct kmem_cache *s, void *p)
445{
446 if (s->flags & SLAB_RED_ZONE)
447 p -= s->red_left_pad;
448
449 return p;
450}
451
Christoph Lameter41ecc552007-05-09 02:32:44 -0700452/*
453 * Debug settings:
454 */
Andrey Ryabinin89d3c872015-11-05 18:51:23 -0800455#if defined(CONFIG_SLUB_DEBUG_ON)
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700456static int slub_debug = DEBUG_DEFAULT_FLAGS;
Andrey Ryabinin89d3c872015-11-05 18:51:23 -0800457#elif defined(CONFIG_KASAN)
458static int slub_debug = SLAB_STORE_USER;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700459#else
Christoph Lameter41ecc552007-05-09 02:32:44 -0700460static int slub_debug;
Christoph Lameterf0630ff2007-07-15 23:38:14 -0700461#endif
Christoph Lameter41ecc552007-05-09 02:32:44 -0700462
463static char *slub_debug_slabs;
David Rientjesfa5ec8a2009-07-07 00:14:14 -0700464static int disable_higher_order_debug;
Christoph Lameter41ecc552007-05-09 02:32:44 -0700465
Christoph Lameter7656c722007-05-09 02:32:40 -0700466/*
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800467 * slub is about to manipulate internal object metadata. This memory lies
468 * outside the range of the allocated object, so accessing it would normally
469 * be reported by kasan as a bounds error. metadata_access_enable() is used
470 * to tell kasan that these accesses are OK.
471 */
472static inline void metadata_access_enable(void)
473{
474 kasan_disable_current();
475}
476
477static inline void metadata_access_disable(void)
478{
479 kasan_enable_current();
480}
481
482/*
Christoph Lameter81819f02007-05-06 14:49:36 -0700483 * Object debugging
484 */
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700485
486/* Verify that a pointer has an address that is valid within a slab page */
487static inline int check_valid_pointer(struct kmem_cache *s,
488 struct page *page, void *object)
489{
490 void *base;
491
492 if (!object)
493 return 1;
494
495 base = page_address(page);
496 object = restore_red_left(s, object);
497 if (object < base || object >= base + page->objects * s->size ||
498 (object - base) % s->size) {
499 return 0;
500 }
501
502 return 1;
503}
504
Christoph Lameter81819f02007-05-06 14:49:36 -0700505static void print_section(char *text, u8 *addr, unsigned int length)
506{
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800507 metadata_access_enable();
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200508 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
509 length, 1);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800510 metadata_access_disable();
Christoph Lameter81819f02007-05-06 14:49:36 -0700511}
512
Christoph Lameter81819f02007-05-06 14:49:36 -0700513static struct track *get_track(struct kmem_cache *s, void *object,
514 enum track_item alloc)
515{
516 struct track *p;
517
518 if (s->offset)
519 p = object + s->offset + sizeof(void *);
520 else
521 p = object + s->inuse;
522
523 return p + alloc;
524}
525
526static void set_track(struct kmem_cache *s, void *object,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300527 enum track_item alloc, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -0700528{
Akinobu Mita1a00df42009-03-07 00:36:21 +0900529 struct track *p = get_track(s, object, alloc);
Christoph Lameter81819f02007-05-06 14:49:36 -0700530
Christoph Lameter81819f02007-05-06 14:49:36 -0700531 if (addr) {
Ben Greeard6543e32011-07-07 11:36:36 -0700532#ifdef CONFIG_STACKTRACE
533 struct stack_trace trace;
534 int i;
535
536 trace.nr_entries = 0;
537 trace.max_entries = TRACK_ADDRS_COUNT;
538 trace.entries = p->addrs;
539 trace.skip = 3;
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800540 metadata_access_enable();
Ben Greeard6543e32011-07-07 11:36:36 -0700541 save_stack_trace(&trace);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800542 metadata_access_disable();
Ben Greeard6543e32011-07-07 11:36:36 -0700543
544 /* See rant in lockdep.c */
545 if (trace.nr_entries != 0 &&
546 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
547 trace.nr_entries--;
548
549 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
550 p->addrs[i] = 0;
551#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700552 p->addr = addr;
553 p->cpu = smp_processor_id();
Alexey Dobriyan88e4ccf2008-06-23 02:58:37 +0400554 p->pid = current->pid;
Christoph Lameter81819f02007-05-06 14:49:36 -0700555 p->when = jiffies;
556 } else
557 memset(p, 0, sizeof(struct track));
558}
559
Christoph Lameter81819f02007-05-06 14:49:36 -0700560static void init_tracking(struct kmem_cache *s, void *object)
561{
Christoph Lameter24922682007-07-17 04:03:18 -0700562 if (!(s->flags & SLAB_STORE_USER))
563 return;
564
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +0300565 set_track(s, object, TRACK_FREE, 0UL);
566 set_track(s, object, TRACK_ALLOC, 0UL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700567}
568
569static void print_track(const char *s, struct track *t)
570{
571 if (!t->addr)
572 return;
573
Fabian Frederickf9f58282014-06-04 16:06:34 -0700574 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
575 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
Ben Greeard6543e32011-07-07 11:36:36 -0700576#ifdef CONFIG_STACKTRACE
577 {
578 int i;
579 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
580 if (t->addrs[i])
Fabian Frederickf9f58282014-06-04 16:06:34 -0700581 pr_err("\t%pS\n", (void *)t->addrs[i]);
Ben Greeard6543e32011-07-07 11:36:36 -0700582 else
583 break;
584 }
585#endif
Christoph Lameter81819f02007-05-06 14:49:36 -0700586}
587
Christoph Lameter24922682007-07-17 04:03:18 -0700588static void print_tracking(struct kmem_cache *s, void *object)
589{
590 if (!(s->flags & SLAB_STORE_USER))
591 return;
592
593 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
594 print_track("Freed", get_track(s, object, TRACK_FREE));
595}
596
597static void print_page_info(struct page *page)
598{
Fabian Frederickf9f58282014-06-04 16:06:34 -0700599 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800600 page, page->objects, page->inuse, page->freelist, page->flags);
Christoph Lameter24922682007-07-17 04:03:18 -0700601
602}
603
604static void slab_bug(struct kmem_cache *s, char *fmt, ...)
605{
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700606 struct va_format vaf;
Christoph Lameter24922682007-07-17 04:03:18 -0700607 va_list args;
Christoph Lameter24922682007-07-17 04:03:18 -0700608
609 va_start(args, fmt);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700610 vaf.fmt = fmt;
611 vaf.va = &args;
Fabian Frederickf9f58282014-06-04 16:06:34 -0700612 pr_err("=============================================================================\n");
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700613 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
Fabian Frederickf9f58282014-06-04 16:06:34 -0700614 pr_err("-----------------------------------------------------------------------------\n\n");
Dave Jones645df232012-09-18 15:54:12 -0400615
Rusty Russell373d4d02013-01-21 17:17:39 +1030616 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700617 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700618}
619
620static void slab_fix(struct kmem_cache *s, char *fmt, ...)
621{
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700622 struct va_format vaf;
Christoph Lameter24922682007-07-17 04:03:18 -0700623 va_list args;
Christoph Lameter24922682007-07-17 04:03:18 -0700624
625 va_start(args, fmt);
Fabian Frederickecc42fb2014-06-04 16:06:35 -0700626 vaf.fmt = fmt;
627 vaf.va = &args;
628 pr_err("FIX %s: %pV\n", s->name, &vaf);
Christoph Lameter24922682007-07-17 04:03:18 -0700629 va_end(args);
Christoph Lameter24922682007-07-17 04:03:18 -0700630}
631
632static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
Christoph Lameter81819f02007-05-06 14:49:36 -0700633{
634 unsigned int off; /* Offset of last byte */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800635 u8 *addr = page_address(page);
Christoph Lameter24922682007-07-17 04:03:18 -0700636
637 print_tracking(s, p);
638
639 print_page_info(page);
640
Fabian Frederickf9f58282014-06-04 16:06:34 -0700641 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
642 p, p - addr, get_freepointer(s, p));
Christoph Lameter24922682007-07-17 04:03:18 -0700643
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700644 if (s->flags & SLAB_RED_ZONE)
645 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
646 else if (p > addr + 16)
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200647 print_section("Bytes b4 ", p - 16, 16);
Christoph Lameter24922682007-07-17 04:03:18 -0700648
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500649 print_section("Object ", p, min_t(unsigned long, s->object_size,
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200650 PAGE_SIZE));
Christoph Lameter81819f02007-05-06 14:49:36 -0700651 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500652 print_section("Redzone ", p + s->object_size,
653 s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700654
Christoph Lameter81819f02007-05-06 14:49:36 -0700655 if (s->offset)
656 off = s->offset + sizeof(void *);
657 else
658 off = s->inuse;
659
Christoph Lameter24922682007-07-17 04:03:18 -0700660 if (s->flags & SLAB_STORE_USER)
Christoph Lameter81819f02007-05-06 14:49:36 -0700661 off += 2 * sizeof(struct track);
Christoph Lameter81819f02007-05-06 14:49:36 -0700662
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700663 if (off != size_from_object(s))
Christoph Lameter81819f02007-05-06 14:49:36 -0700664 /* Beginning of the filler is the free pointer */
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700665 print_section("Padding ", p + off, size_from_object(s) - off);
Christoph Lameter24922682007-07-17 04:03:18 -0700666
667 dump_stack();
Christoph Lameter81819f02007-05-06 14:49:36 -0700668}
669
Andrey Ryabinin75c66de2015-02-13 14:39:35 -0800670void object_err(struct kmem_cache *s, struct page *page,
Christoph Lameter81819f02007-05-06 14:49:36 -0700671 u8 *object, char *reason)
672{
Christoph Lameter3dc50632008-04-23 12:28:01 -0700673 slab_bug(s, "%s", reason);
Christoph Lameter24922682007-07-17 04:03:18 -0700674 print_trailer(s, page, object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700675}
676
Chen Gangd0e0ac92013-07-15 09:05:29 +0800677static void slab_err(struct kmem_cache *s, struct page *page,
678 const char *fmt, ...)
Christoph Lameter81819f02007-05-06 14:49:36 -0700679{
680 va_list args;
681 char buf[100];
682
Christoph Lameter24922682007-07-17 04:03:18 -0700683 va_start(args, fmt);
684 vsnprintf(buf, sizeof(buf), fmt, args);
Christoph Lameter81819f02007-05-06 14:49:36 -0700685 va_end(args);
Christoph Lameter3dc50632008-04-23 12:28:01 -0700686 slab_bug(s, "%s", buf);
Christoph Lameter24922682007-07-17 04:03:18 -0700687 print_page_info(page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700688 dump_stack();
689}
690
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500691static void init_object(struct kmem_cache *s, void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700692{
693 u8 *p = object;
694
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700695 if (s->flags & SLAB_RED_ZONE)
696 memset(p - s->red_left_pad, val, s->red_left_pad);
697
Christoph Lameter81819f02007-05-06 14:49:36 -0700698 if (s->flags & __OBJECT_POISON) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500699 memset(p, POISON_FREE, s->object_size - 1);
700 p[s->object_size - 1] = POISON_END;
Christoph Lameter81819f02007-05-06 14:49:36 -0700701 }
702
703 if (s->flags & SLAB_RED_ZONE)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500704 memset(p + s->object_size, val, s->inuse - s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -0700705}
706
Christoph Lameter24922682007-07-17 04:03:18 -0700707static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
708 void *from, void *to)
709{
710 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
711 memset(from, data, to - from);
712}
713
714static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
715 u8 *object, char *what,
Pekka Enberg06428782008-01-07 23:20:27 -0800716 u8 *start, unsigned int value, unsigned int bytes)
Christoph Lameter24922682007-07-17 04:03:18 -0700717{
718 u8 *fault;
719 u8 *end;
720
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800721 metadata_access_enable();
Akinobu Mita798248202011-10-31 17:08:07 -0700722 fault = memchr_inv(start, value, bytes);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800723 metadata_access_disable();
Christoph Lameter24922682007-07-17 04:03:18 -0700724 if (!fault)
725 return 1;
726
727 end = start + bytes;
728 while (end > fault && end[-1] == value)
729 end--;
730
731 slab_bug(s, "%s overwritten", what);
Fabian Frederickf9f58282014-06-04 16:06:34 -0700732 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
Christoph Lameter24922682007-07-17 04:03:18 -0700733 fault, end - 1, fault[0], value);
734 print_trailer(s, page, object);
735
736 restore_bytes(s, what, value, fault, end);
737 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700738}
739
Christoph Lameter81819f02007-05-06 14:49:36 -0700740/*
741 * Object layout:
742 *
743 * object address
744 * Bytes of the object to be managed.
745 * If the freepointer may overlay the object then the free
746 * pointer is the first word of the object.
Christoph Lameter672bba32007-05-09 02:32:39 -0700747 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700748 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
749 * 0xa5 (POISON_END)
750 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500751 * object + s->object_size
Christoph Lameter81819f02007-05-06 14:49:36 -0700752 * Padding to reach word boundary. This is also used for Redzoning.
Christoph Lameter672bba32007-05-09 02:32:39 -0700753 * Padding is extended by another word if Redzoning is enabled and
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500754 * object_size == inuse.
Christoph Lameter672bba32007-05-09 02:32:39 -0700755 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700756 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
757 * 0xcc (RED_ACTIVE) for objects in use.
758 *
759 * object + s->inuse
Christoph Lameter672bba32007-05-09 02:32:39 -0700760 * Meta data starts here.
761 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700762 * A. Free pointer (if we cannot overwrite object on free)
763 * B. Tracking data for SLAB_STORE_USER
Christoph Lameter672bba32007-05-09 02:32:39 -0700764 * C. Padding to reach required alignment boundary or at mininum
Christoph Lameter6446faa2008-02-15 23:45:26 -0800765 * one word if debugging is on to be able to detect writes
Christoph Lameter672bba32007-05-09 02:32:39 -0700766 * before the word boundary.
767 *
768 * Padding is done using 0x5a (POISON_INUSE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700769 *
770 * object + s->size
Christoph Lameter672bba32007-05-09 02:32:39 -0700771 * Nothing is used beyond s->size.
Christoph Lameter81819f02007-05-06 14:49:36 -0700772 *
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500773 * If slabcaches are merged then the object_size and inuse boundaries are mostly
Christoph Lameter672bba32007-05-09 02:32:39 -0700774 * ignored. And therefore no slab options that rely on these boundaries
Christoph Lameter81819f02007-05-06 14:49:36 -0700775 * may be used with merged slabcaches.
776 */
777
Christoph Lameter81819f02007-05-06 14:49:36 -0700778static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
779{
780 unsigned long off = s->inuse; /* The end of info */
781
782 if (s->offset)
783 /* Freepointer is placed after the object. */
784 off += sizeof(void *);
785
786 if (s->flags & SLAB_STORE_USER)
787 /* We also have user information there */
788 off += 2 * sizeof(struct track);
789
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700790 if (size_from_object(s) == off)
Christoph Lameter81819f02007-05-06 14:49:36 -0700791 return 1;
792
Christoph Lameter24922682007-07-17 04:03:18 -0700793 return check_bytes_and_report(s, page, p, "Object padding",
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700794 p + off, POISON_INUSE, size_from_object(s) - off);
Christoph Lameter81819f02007-05-06 14:49:36 -0700795}
796
Christoph Lameter39b26462008-04-14 19:11:30 +0300797/* Check the pad bytes at the end of a slab page */
Christoph Lameter81819f02007-05-06 14:49:36 -0700798static int slab_pad_check(struct kmem_cache *s, struct page *page)
799{
Christoph Lameter24922682007-07-17 04:03:18 -0700800 u8 *start;
801 u8 *fault;
802 u8 *end;
803 int length;
804 int remainder;
Christoph Lameter81819f02007-05-06 14:49:36 -0700805
806 if (!(s->flags & SLAB_POISON))
807 return 1;
808
Christoph Lametera973e9d2008-03-01 13:40:44 -0800809 start = page_address(page);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800810 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
Christoph Lameter39b26462008-04-14 19:11:30 +0300811 end = start + length;
812 remainder = length % s->size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700813 if (!remainder)
814 return 1;
815
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800816 metadata_access_enable();
Akinobu Mita798248202011-10-31 17:08:07 -0700817 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
Andrey Ryabinina79316c2015-02-13 14:39:38 -0800818 metadata_access_disable();
Christoph Lameter24922682007-07-17 04:03:18 -0700819 if (!fault)
820 return 1;
821 while (end > fault && end[-1] == POISON_INUSE)
822 end--;
823
824 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
Sebastian Andrzej Siewiorffc79d22011-07-29 14:10:20 +0200825 print_section("Padding ", end - remainder, remainder);
Christoph Lameter24922682007-07-17 04:03:18 -0700826
Eric Dumazet8a3d2712009-09-03 16:08:06 +0200827 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
Christoph Lameter24922682007-07-17 04:03:18 -0700828 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700829}
830
831static int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500832 void *object, u8 val)
Christoph Lameter81819f02007-05-06 14:49:36 -0700833{
834 u8 *p = object;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500835 u8 *endobject = object + s->object_size;
Christoph Lameter81819f02007-05-06 14:49:36 -0700836
837 if (s->flags & SLAB_RED_ZONE) {
Christoph Lameter24922682007-07-17 04:03:18 -0700838 if (!check_bytes_and_report(s, page, object, "Redzone",
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -0700839 object - s->red_left_pad, val, s->red_left_pad))
840 return 0;
841
842 if (!check_bytes_and_report(s, page, object, "Redzone",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500843 endobject, val, s->inuse - s->object_size))
Christoph Lameter81819f02007-05-06 14:49:36 -0700844 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700845 } else {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500846 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800847 check_bytes_and_report(s, page, p, "Alignment padding",
Chen Gangd0e0ac92013-07-15 09:05:29 +0800848 endobject, POISON_INUSE,
849 s->inuse - s->object_size);
Ingo Molnar3adbefe2008-02-05 17:57:39 -0800850 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700851 }
852
853 if (s->flags & SLAB_POISON) {
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500854 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
Christoph Lameter24922682007-07-17 04:03:18 -0700855 (!check_bytes_and_report(s, page, p, "Poison", p,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500856 POISON_FREE, s->object_size - 1) ||
Christoph Lameter24922682007-07-17 04:03:18 -0700857 !check_bytes_and_report(s, page, p, "Poison",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500858 p + s->object_size - 1, POISON_END, 1)))
Christoph Lameter81819f02007-05-06 14:49:36 -0700859 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -0700860 /*
861 * check_pad_bytes cleans up on its own.
862 */
863 check_pad_bytes(s, page, p);
864 }
865
Christoph Lameterf7cb1932010-09-29 07:15:01 -0500866 if (!s->offset && val == SLUB_RED_ACTIVE)
Christoph Lameter81819f02007-05-06 14:49:36 -0700867 /*
868 * Object and freepointer overlap. Cannot check
869 * freepointer while object is allocated.
870 */
871 return 1;
872
873 /* Check free pointer validity */
874 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
875 object_err(s, page, p, "Freepointer corrupt");
876 /*
Nick Andrew9f6c708e2008-12-05 14:08:08 +1100877 * No choice but to zap it and thus lose the remainder
Christoph Lameter81819f02007-05-06 14:49:36 -0700878 * of the free objects in this slab. May cause
Christoph Lameter672bba32007-05-09 02:32:39 -0700879 * another error because the object count is now wrong.
Christoph Lameter81819f02007-05-06 14:49:36 -0700880 */
Christoph Lametera973e9d2008-03-01 13:40:44 -0800881 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700882 return 0;
883 }
884 return 1;
885}
886
887static int check_slab(struct kmem_cache *s, struct page *page)
888{
Christoph Lameter39b26462008-04-14 19:11:30 +0300889 int maxobj;
890
Christoph Lameter81819f02007-05-06 14:49:36 -0700891 VM_BUG_ON(!irqs_disabled());
892
893 if (!PageSlab(page)) {
Christoph Lameter24922682007-07-17 04:03:18 -0700894 slab_err(s, page, "Not a valid slab page");
Christoph Lameter81819f02007-05-06 14:49:36 -0700895 return 0;
896 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300897
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800898 maxobj = order_objects(compound_order(page), s->size, s->reserved);
Christoph Lameter39b26462008-04-14 19:11:30 +0300899 if (page->objects > maxobj) {
900 slab_err(s, page, "objects %u > max %u",
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800901 page->objects, maxobj);
Christoph Lameter39b26462008-04-14 19:11:30 +0300902 return 0;
903 }
904 if (page->inuse > page->objects) {
Christoph Lameter24922682007-07-17 04:03:18 -0700905 slab_err(s, page, "inuse %u > max %u",
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800906 page->inuse, page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -0700907 return 0;
908 }
909 /* Slab_pad_check fixes things up after itself */
910 slab_pad_check(s, page);
911 return 1;
912}
913
914/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700915 * Determine if a certain object on a page is on the freelist. Must hold the
916 * slab lock to guarantee that the chains are in a consistent state.
Christoph Lameter81819f02007-05-06 14:49:36 -0700917 */
918static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
919{
920 int nr = 0;
Christoph Lameter881db7f2011-06-01 12:25:53 -0500921 void *fp;
Christoph Lameter81819f02007-05-06 14:49:36 -0700922 void *object = NULL;
Andrey Ryabininf6edde92014-12-10 15:42:22 -0800923 int max_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -0700924
Christoph Lameter881db7f2011-06-01 12:25:53 -0500925 fp = page->freelist;
Christoph Lameter39b26462008-04-14 19:11:30 +0300926 while (fp && nr <= page->objects) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700927 if (fp == search)
928 return 1;
929 if (!check_valid_pointer(s, page, fp)) {
930 if (object) {
931 object_err(s, page, object,
932 "Freechain corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800933 set_freepointer(s, object, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -0700934 } else {
Christoph Lameter24922682007-07-17 04:03:18 -0700935 slab_err(s, page, "Freepointer corrupt");
Christoph Lametera973e9d2008-03-01 13:40:44 -0800936 page->freelist = NULL;
Christoph Lameter39b26462008-04-14 19:11:30 +0300937 page->inuse = page->objects;
Christoph Lameter24922682007-07-17 04:03:18 -0700938 slab_fix(s, "Freelist cleared");
Christoph Lameter81819f02007-05-06 14:49:36 -0700939 return 0;
940 }
941 break;
942 }
943 object = fp;
944 fp = get_freepointer(s, object);
945 nr++;
946 }
947
Lai Jiangshanab9a0f12011-03-10 15:21:48 +0800948 max_objects = order_objects(compound_order(page), s->size, s->reserved);
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +0400949 if (max_objects > MAX_OBJS_PER_PAGE)
950 max_objects = MAX_OBJS_PER_PAGE;
Christoph Lameter224a88b2008-04-14 19:11:31 +0300951
952 if (page->objects != max_objects) {
Joe Perches756a0252016-03-17 14:19:47 -0700953 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
954 page->objects, max_objects);
Christoph Lameter224a88b2008-04-14 19:11:31 +0300955 page->objects = max_objects;
956 slab_fix(s, "Number of objects adjusted.");
957 }
Christoph Lameter39b26462008-04-14 19:11:30 +0300958 if (page->inuse != page->objects - nr) {
Joe Perches756a0252016-03-17 14:19:47 -0700959 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
960 page->inuse, page->objects - nr);
Christoph Lameter39b26462008-04-14 19:11:30 +0300961 page->inuse = page->objects - nr;
Christoph Lameter24922682007-07-17 04:03:18 -0700962 slab_fix(s, "Object count adjusted.");
Christoph Lameter81819f02007-05-06 14:49:36 -0700963 }
964 return search == NULL;
965}
966
Christoph Lameter0121c6192008-04-29 16:11:12 -0700967static void trace(struct kmem_cache *s, struct page *page, void *object,
968 int alloc)
Christoph Lameter3ec09742007-05-16 22:11:00 -0700969{
970 if (s->flags & SLAB_TRACE) {
Fabian Frederickf9f58282014-06-04 16:06:34 -0700971 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
Christoph Lameter3ec09742007-05-16 22:11:00 -0700972 s->name,
973 alloc ? "alloc" : "free",
974 object, page->inuse,
975 page->freelist);
976
977 if (!alloc)
Chen Gangd0e0ac92013-07-15 09:05:29 +0800978 print_section("Object ", (void *)object,
979 s->object_size);
Christoph Lameter3ec09742007-05-16 22:11:00 -0700980
981 dump_stack();
982 }
983}
984
Christoph Lameter643b1132007-05-06 14:49:42 -0700985/*
Christoph Lameter672bba32007-05-09 02:32:39 -0700986 * Tracking of fully allocated slabs for debugging purposes.
Christoph Lameter643b1132007-05-06 14:49:42 -0700987 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500988static void add_full(struct kmem_cache *s,
989 struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700990{
Christoph Lameter643b1132007-05-06 14:49:42 -0700991 if (!(s->flags & SLAB_STORE_USER))
992 return;
993
David Rientjes255d0882014-02-10 14:25:39 -0800994 lockdep_assert_held(&n->list_lock);
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500995 list_add(&page->lru, &n->full);
996}
Christoph Lameter643b1132007-05-06 14:49:42 -0700997
Peter Zijlstrac65c1872014-01-10 13:23:49 +0100998static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
Christoph Lameter5cc6eee2011-06-01 12:25:50 -0500999{
1000 if (!(s->flags & SLAB_STORE_USER))
1001 return;
1002
David Rientjes255d0882014-02-10 14:25:39 -08001003 lockdep_assert_held(&n->list_lock);
Christoph Lameter643b1132007-05-06 14:49:42 -07001004 list_del(&page->lru);
Christoph Lameter643b1132007-05-06 14:49:42 -07001005}
1006
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001007/* Tracking of the number of slabs for debugging purposes */
1008static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1009{
1010 struct kmem_cache_node *n = get_node(s, node);
1011
1012 return atomic_long_read(&n->nr_slabs);
1013}
1014
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001015static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1016{
1017 return atomic_long_read(&n->nr_slabs);
1018}
1019
Christoph Lameter205ab992008-04-14 19:11:40 +03001020static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001021{
1022 struct kmem_cache_node *n = get_node(s, node);
1023
1024 /*
1025 * May be called early in order to allocate a slab for the
1026 * kmem_cache_node structure. Solve the chicken-egg
1027 * dilemma by deferring the increment of the count during
1028 * bootstrap (see early_kmem_cache_node_alloc).
1029 */
Joonsoo Kim338b2642013-01-21 17:01:27 +09001030 if (likely(n)) {
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001031 atomic_long_inc(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001032 atomic_long_add(objects, &n->total_objects);
1033 }
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001034}
Christoph Lameter205ab992008-04-14 19:11:40 +03001035static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001036{
1037 struct kmem_cache_node *n = get_node(s, node);
1038
1039 atomic_long_dec(&n->nr_slabs);
Christoph Lameter205ab992008-04-14 19:11:40 +03001040 atomic_long_sub(objects, &n->total_objects);
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001041}
1042
1043/* Object debug checks for alloc/free paths */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001044static void setup_object_debug(struct kmem_cache *s, struct page *page,
1045 void *object)
1046{
1047 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1048 return;
1049
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001050 init_object(s, object, SLUB_RED_INACTIVE);
Christoph Lameter3ec09742007-05-16 22:11:00 -07001051 init_tracking(s, object);
1052}
1053
Laura Abbottbecfda62016-03-15 14:55:06 -07001054static inline int alloc_consistency_checks(struct kmem_cache *s,
Chen Gangd0e0ac92013-07-15 09:05:29 +08001055 struct page *page,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001056 void *object, unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001057{
1058 if (!check_slab(s, page))
Laura Abbottbecfda62016-03-15 14:55:06 -07001059 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001060
Christoph Lameter81819f02007-05-06 14:49:36 -07001061 if (!check_valid_pointer(s, page, object)) {
1062 object_err(s, page, object, "Freelist Pointer check fails");
Laura Abbottbecfda62016-03-15 14:55:06 -07001063 return 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07001064 }
1065
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001066 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
Laura Abbottbecfda62016-03-15 14:55:06 -07001067 return 0;
1068
1069 return 1;
1070}
1071
1072static noinline int alloc_debug_processing(struct kmem_cache *s,
1073 struct page *page,
1074 void *object, unsigned long addr)
1075{
1076 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1077 if (!alloc_consistency_checks(s, page, object, addr))
1078 goto bad;
1079 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001080
Christoph Lameter3ec09742007-05-16 22:11:00 -07001081 /* Success perform special debug activities for allocs */
1082 if (s->flags & SLAB_STORE_USER)
1083 set_track(s, object, TRACK_ALLOC, addr);
1084 trace(s, page, object, 1);
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001085 init_object(s, object, SLUB_RED_ACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001086 return 1;
Christoph Lameter3ec09742007-05-16 22:11:00 -07001087
Christoph Lameter81819f02007-05-06 14:49:36 -07001088bad:
1089 if (PageSlab(page)) {
1090 /*
1091 * If this is a slab page then lets do the best we can
1092 * to avoid issues in the future. Marking all objects
Christoph Lameter672bba32007-05-09 02:32:39 -07001093 * as used avoids touching the remaining objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001094 */
Christoph Lameter24922682007-07-17 04:03:18 -07001095 slab_fix(s, "Marking all objects used");
Christoph Lameter39b26462008-04-14 19:11:30 +03001096 page->inuse = page->objects;
Christoph Lametera973e9d2008-03-01 13:40:44 -08001097 page->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07001098 }
1099 return 0;
1100}
1101
Laura Abbottbecfda62016-03-15 14:55:06 -07001102static inline int free_consistency_checks(struct kmem_cache *s,
1103 struct page *page, void *object, unsigned long addr)
1104{
1105 if (!check_valid_pointer(s, page, object)) {
1106 slab_err(s, page, "Invalid object pointer 0x%p", object);
1107 return 0;
1108 }
1109
1110 if (on_freelist(s, page, object)) {
1111 object_err(s, page, object, "Object already free");
1112 return 0;
1113 }
1114
1115 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1116 return 0;
1117
1118 if (unlikely(s != page->slab_cache)) {
1119 if (!PageSlab(page)) {
Joe Perches756a0252016-03-17 14:19:47 -07001120 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1121 object);
Laura Abbottbecfda62016-03-15 14:55:06 -07001122 } else if (!page->slab_cache) {
1123 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1124 object);
1125 dump_stack();
1126 } else
1127 object_err(s, page, object,
1128 "page slab pointer corrupt.");
1129 return 0;
1130 }
1131 return 1;
1132}
1133
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001134/* Supports checking bulk free of a constructed freelist */
Laura Abbott282acb42016-03-15 14:54:59 -07001135static noinline int free_debug_processing(
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001136 struct kmem_cache *s, struct page *page,
1137 void *head, void *tail, int bulk_cnt,
Laura Abbott282acb42016-03-15 14:54:59 -07001138 unsigned long addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001139{
Christoph Lameter19c7ff92012-05-30 12:54:46 -05001140 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001141 void *object = head;
1142 int cnt = 0;
Laura Abbott282acb42016-03-15 14:54:59 -07001143 unsigned long uninitialized_var(flags);
Laura Abbott804aa132016-03-15 14:55:02 -07001144 int ret = 0;
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001145
Laura Abbott282acb42016-03-15 14:54:59 -07001146 spin_lock_irqsave(&n->list_lock, flags);
Christoph Lameter881db7f2011-06-01 12:25:53 -05001147 slab_lock(page);
1148
Laura Abbottbecfda62016-03-15 14:55:06 -07001149 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1150 if (!check_slab(s, page))
1151 goto out;
1152 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001153
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001154next_object:
1155 cnt++;
1156
Laura Abbottbecfda62016-03-15 14:55:06 -07001157 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1158 if (!free_consistency_checks(s, page, object, addr))
1159 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07001160 }
Christoph Lameter3ec09742007-05-16 22:11:00 -07001161
Christoph Lameter3ec09742007-05-16 22:11:00 -07001162 if (s->flags & SLAB_STORE_USER)
1163 set_track(s, object, TRACK_FREE, addr);
1164 trace(s, page, object, 0);
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001165 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001166 init_object(s, object, SLUB_RED_INACTIVE);
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001167
1168 /* Reached end of constructed freelist yet? */
1169 if (object != tail) {
1170 object = get_freepointer(s, object);
1171 goto next_object;
1172 }
Laura Abbott804aa132016-03-15 14:55:02 -07001173 ret = 1;
1174
Christoph Lameter5c2e4bb2011-06-01 12:25:54 -05001175out:
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001176 if (cnt != bulk_cnt)
1177 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1178 bulk_cnt, cnt);
1179
Christoph Lameter881db7f2011-06-01 12:25:53 -05001180 slab_unlock(page);
Laura Abbott282acb42016-03-15 14:54:59 -07001181 spin_unlock_irqrestore(&n->list_lock, flags);
Laura Abbott804aa132016-03-15 14:55:02 -07001182 if (!ret)
1183 slab_fix(s, "Object at 0x%p not freed", object);
1184 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07001185}
1186
Christoph Lameter41ecc552007-05-09 02:32:44 -07001187static int __init setup_slub_debug(char *str)
1188{
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001189 slub_debug = DEBUG_DEFAULT_FLAGS;
1190 if (*str++ != '=' || !*str)
1191 /*
1192 * No options specified. Switch on full debugging.
1193 */
1194 goto out;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001195
1196 if (*str == ',')
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001197 /*
1198 * No options but restriction on slabs. This means full
1199 * debugging for slabs matching a pattern.
1200 */
1201 goto check_slabs;
1202
1203 slub_debug = 0;
1204 if (*str == '-')
1205 /*
1206 * Switch off all debugging measures.
1207 */
1208 goto out;
1209
1210 /*
1211 * Determine which debug features should be switched on
1212 */
Pekka Enberg06428782008-01-07 23:20:27 -08001213 for (; *str && *str != ','; str++) {
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001214 switch (tolower(*str)) {
1215 case 'f':
Laura Abbottbecfda62016-03-15 14:55:06 -07001216 slub_debug |= SLAB_CONSISTENCY_CHECKS;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001217 break;
1218 case 'z':
1219 slub_debug |= SLAB_RED_ZONE;
1220 break;
1221 case 'p':
1222 slub_debug |= SLAB_POISON;
1223 break;
1224 case 'u':
1225 slub_debug |= SLAB_STORE_USER;
1226 break;
1227 case 't':
1228 slub_debug |= SLAB_TRACE;
1229 break;
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03001230 case 'a':
1231 slub_debug |= SLAB_FAILSLAB;
1232 break;
Chris J Arges08303a72015-04-14 15:44:25 -07001233 case 'o':
1234 /*
1235 * Avoid enabling debugging on caches if its minimum
1236 * order would increase as a result.
1237 */
1238 disable_higher_order_debug = 1;
1239 break;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001240 default:
Fabian Frederickf9f58282014-06-04 16:06:34 -07001241 pr_err("slub_debug option '%c' unknown. skipped\n",
1242 *str);
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001243 }
1244 }
1245
1246check_slabs:
1247 if (*str == ',')
Christoph Lameter41ecc552007-05-09 02:32:44 -07001248 slub_debug_slabs = str + 1;
Christoph Lameterf0630ff2007-07-15 23:38:14 -07001249out:
Christoph Lameter41ecc552007-05-09 02:32:44 -07001250 return 1;
1251}
1252
1253__setup("slub_debug", setup_slub_debug);
1254
Joonsoo Kim423c9292014-10-09 15:26:22 -07001255unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001256 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001257 void (*ctor)(void *))
Christoph Lameter41ecc552007-05-09 02:32:44 -07001258{
1259 /*
Christoph Lametere1533622008-02-15 23:45:24 -08001260 * Enable debugging if selected on the kernel commandline.
Christoph Lameter41ecc552007-05-09 02:32:44 -07001261 */
Christoph Lameterc6f58d92013-11-07 16:29:15 +00001262 if (slub_debug && (!slub_debug_slabs || (name &&
1263 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
David Rientjes3de47212009-07-27 18:30:35 -07001264 flags |= slub_debug;
Christoph Lameterba0268a2007-09-11 15:24:11 -07001265
1266 return flags;
Christoph Lameter41ecc552007-05-09 02:32:44 -07001267}
Jesper Dangaard Brouerb4a64712015-11-20 15:57:41 -08001268#else /* !CONFIG_SLUB_DEBUG */
Christoph Lameter3ec09742007-05-16 22:11:00 -07001269static inline void setup_object_debug(struct kmem_cache *s,
1270 struct page *page, void *object) {}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001271
Christoph Lameter3ec09742007-05-16 22:11:00 -07001272static inline int alloc_debug_processing(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03001273 struct page *page, void *object, unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001274
Laura Abbott282acb42016-03-15 14:54:59 -07001275static inline int free_debug_processing(
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001276 struct kmem_cache *s, struct page *page,
1277 void *head, void *tail, int bulk_cnt,
Laura Abbott282acb42016-03-15 14:54:59 -07001278 unsigned long addr) { return 0; }
Christoph Lameter41ecc552007-05-09 02:32:44 -07001279
Christoph Lameter41ecc552007-05-09 02:32:44 -07001280static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1281 { return 1; }
1282static inline int check_object(struct kmem_cache *s, struct page *page,
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001283 void *object, u8 val) { return 1; }
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001284static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1285 struct page *page) {}
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001286static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1287 struct page *page) {}
Joonsoo Kim423c9292014-10-09 15:26:22 -07001288unsigned long kmem_cache_flags(unsigned long object_size,
Christoph Lameterba0268a2007-09-11 15:24:11 -07001289 unsigned long flags, const char *name,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07001290 void (*ctor)(void *))
Christoph Lameterba0268a2007-09-11 15:24:11 -07001291{
1292 return flags;
1293}
Christoph Lameter41ecc552007-05-09 02:32:44 -07001294#define slub_debug 0
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001295
Ingo Molnarfdaa45e2009-09-15 11:00:26 +02001296#define disable_higher_order_debug 0
1297
Christoph Lameter0f389ec2008-04-14 18:53:02 +03001298static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1299 { return 0; }
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04001300static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1301 { return 0; }
Christoph Lameter205ab992008-04-14 19:11:40 +03001302static inline void inc_slabs_node(struct kmem_cache *s, int node,
1303 int objects) {}
1304static inline void dec_slabs_node(struct kmem_cache *s, int node,
1305 int objects) {}
Christoph Lameter7d550c52010-08-25 14:07:16 -05001306
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001307#endif /* CONFIG_SLUB_DEBUG */
1308
1309/*
1310 * Hooks for other subsystems that check memory allocations. In a typical
1311 * production configuration these hooks all should produce no code at all.
1312 */
Roman Bobnievd56791b2013-10-08 15:58:57 -07001313static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1314{
1315 kmemleak_alloc(ptr, size, 1, flags);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07001316 kasan_kmalloc_large(ptr, size, flags);
Roman Bobnievd56791b2013-10-08 15:58:57 -07001317}
1318
1319static inline void kfree_hook(const void *x)
1320{
1321 kmemleak_free(x);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001322 kasan_kfree_large(x);
Roman Bobnievd56791b2013-10-08 15:58:57 -07001323}
1324
Roman Bobnievd56791b2013-10-08 15:58:57 -07001325static inline void slab_free_hook(struct kmem_cache *s, void *x)
1326{
1327 kmemleak_free_recursive(x, s->flags);
Christoph Lameter7d550c52010-08-25 14:07:16 -05001328
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001329 /*
1330 * Trouble is that we may no longer disable interrupts in the fast path
1331 * So in order to make the debug calls that expect irqs to be
1332 * disabled we need to disable interrupts temporarily.
1333 */
1334#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1335 {
1336 unsigned long flags;
1337
1338 local_irq_save(flags);
1339 kmemcheck_slab_free(s, x, s->object_size);
1340 debug_check_no_locks_freed(x, s->object_size);
1341 local_irq_restore(flags);
1342 }
1343#endif
1344 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1345 debug_check_no_obj_freed(x, s->object_size);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001346
1347 kasan_slab_free(s, x);
Andrey Ryabinin02e72cc2014-08-06 16:04:18 -07001348}
Christoph Lameter205ab992008-04-14 19:11:40 +03001349
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08001350static inline void slab_free_freelist_hook(struct kmem_cache *s,
1351 void *head, void *tail)
1352{
1353/*
1354 * Compiler cannot detect this function can be removed if slab_free_hook()
1355 * evaluates to nothing. Thus, catch all relevant config debug options here.
1356 */
1357#if defined(CONFIG_KMEMCHECK) || \
1358 defined(CONFIG_LOCKDEP) || \
1359 defined(CONFIG_DEBUG_KMEMLEAK) || \
1360 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1361 defined(CONFIG_KASAN)
1362
1363 void *object = head;
1364 void *tail_obj = tail ? : head;
1365
1366 do {
1367 slab_free_hook(s, object);
1368 } while ((object != tail_obj) &&
1369 (object = get_freepointer(s, object)));
1370#endif
1371}
1372
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001373static void setup_object(struct kmem_cache *s, struct page *page,
1374 void *object)
1375{
1376 setup_object_debug(s, page, object);
1377 if (unlikely(s->ctor)) {
1378 kasan_unpoison_object_data(s, object);
1379 s->ctor(object);
1380 kasan_poison_object_data(s, object);
1381 }
1382}
1383
Christoph Lameter81819f02007-05-06 14:49:36 -07001384/*
1385 * Slab allocation and freeing
1386 */
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001387static inline struct page *alloc_slab_page(struct kmem_cache *s,
1388 gfp_t flags, int node, struct kmem_cache_order_objects oo)
Christoph Lameter65c33762008-04-14 19:11:40 +03001389{
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001390 struct page *page;
Christoph Lameter65c33762008-04-14 19:11:40 +03001391 int order = oo_order(oo);
1392
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001393 flags |= __GFP_NOTRACK;
1394
Christoph Lameter2154a332010-07-09 14:07:10 -05001395 if (node == NUMA_NO_NODE)
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001396 page = alloc_pages(flags, order);
Christoph Lameter65c33762008-04-14 19:11:40 +03001397 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001398 page = __alloc_pages_node(node, flags, order);
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001399
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001400 if (page && memcg_charge_slab(page, flags, order, s)) {
1401 __free_pages(page, order);
1402 page = NULL;
1403 }
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001404
1405 return page;
Christoph Lameter65c33762008-04-14 19:11:40 +03001406}
1407
Christoph Lameter81819f02007-05-06 14:49:36 -07001408static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1409{
Pekka Enberg06428782008-01-07 23:20:27 -08001410 struct page *page;
Christoph Lameter834f3d12008-04-14 19:11:31 +03001411 struct kmem_cache_order_objects oo = s->oo;
Pekka Enbergba522702009-06-24 21:59:51 +03001412 gfp_t alloc_gfp;
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001413 void *start, *p;
1414 int idx, order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001415
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001416 flags &= gfp_allowed_mask;
1417
Mel Gormand0164ad2015-11-06 16:28:21 -08001418 if (gfpflags_allow_blocking(flags))
Christoph Lameter7e0528d2011-06-01 12:25:44 -05001419 local_irq_enable();
1420
Christoph Lameterb7a49f02008-02-14 14:21:32 -08001421 flags |= s->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001422
Pekka Enbergba522702009-06-24 21:59:51 +03001423 /*
1424 * Let the initial higher-order allocation fail under memory pressure
1425 * so we fall-back to the minimum order allocation.
1426 */
1427 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
Mel Gormand0164ad2015-11-06 16:28:21 -08001428 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
Mel Gorman444eb2a42016-03-17 14:19:23 -07001429 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
Pekka Enbergba522702009-06-24 21:59:51 +03001430
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001431 page = alloc_slab_page(s, alloc_gfp, node, oo);
Christoph Lameter65c33762008-04-14 19:11:40 +03001432 if (unlikely(!page)) {
1433 oo = s->min;
Joonsoo Kim80c3a992014-03-12 17:26:20 +09001434 alloc_gfp = flags;
Christoph Lameter65c33762008-04-14 19:11:40 +03001435 /*
1436 * Allocation may have failed due to fragmentation.
1437 * Try a lower order alloc if possible
1438 */
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001439 page = alloc_slab_page(s, alloc_gfp, node, oo);
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001440 if (unlikely(!page))
1441 goto out;
1442 stat(s, ORDER_FALLBACK);
Christoph Lameter65c33762008-04-14 19:11:40 +03001443 }
Vegard Nossum5a896d92008-04-04 00:54:48 +02001444
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001445 if (kmemcheck_enabled &&
1446 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001447 int pages = 1 << oo_order(oo);
1448
Joonsoo Kim80c3a992014-03-12 17:26:20 +09001449 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001450
1451 /*
1452 * Objects from caches that have a constructor don't get
1453 * cleared when they're allocated, so we need to do it here.
1454 */
1455 if (s->ctor)
1456 kmemcheck_mark_uninitialized_pages(page, pages);
1457 else
1458 kmemcheck_mark_unallocated_pages(page, pages);
Vegard Nossum5a896d92008-04-04 00:54:48 +02001459 }
1460
Christoph Lameter834f3d12008-04-14 19:11:31 +03001461 page->objects = oo_objects(oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07001462
Glauber Costa1f458cb2012-12-18 14:22:50 -08001463 order = compound_order(page);
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001464 page->slab_cache = s;
Joonsoo Kimc03f94c2012-05-18 00:47:47 +09001465 __SetPageSlab(page);
Michal Hocko2f064f32015-08-21 14:11:51 -07001466 if (page_is_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -07001467 SetPageSlabPfmemalloc(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001468
1469 start = page_address(page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001470
1471 if (unlikely(s->flags & SLAB_POISON))
Glauber Costa1f458cb2012-12-18 14:22:50 -08001472 memset(start, POISON_INUSE, PAGE_SIZE << order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001473
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08001474 kasan_poison_slab(page);
1475
Wei Yang54266642014-08-06 16:04:42 -07001476 for_each_object_idx(p, idx, s, start, page->objects) {
1477 setup_object(s, page, p);
1478 if (likely(idx < page->objects))
1479 set_freepointer(s, p, p + s->size);
1480 else
1481 set_freepointer(s, p, NULL);
Christoph Lameter81819f02007-05-06 14:49:36 -07001482 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001483
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -07001484 page->freelist = fixup_red_left(s, start);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05001485 page->inuse = page->objects;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05001486 page->frozen = 1;
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001487
Christoph Lameter81819f02007-05-06 14:49:36 -07001488out:
Mel Gormand0164ad2015-11-06 16:28:21 -08001489 if (gfpflags_allow_blocking(flags))
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001490 local_irq_disable();
1491 if (!page)
1492 return NULL;
1493
1494 mod_zone_page_state(page_zone(page),
1495 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1496 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1497 1 << oo_order(oo));
1498
1499 inc_slabs_node(s, page_to_nid(page), page->objects);
1500
Christoph Lameter81819f02007-05-06 14:49:36 -07001501 return page;
1502}
1503
Thomas Gleixner588f8ba2015-09-04 15:45:48 -07001504static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1505{
1506 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1507 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1508 BUG();
1509 }
1510
1511 return allocate_slab(s,
1512 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1513}
1514
Christoph Lameter81819f02007-05-06 14:49:36 -07001515static void __free_slab(struct kmem_cache *s, struct page *page)
1516{
Christoph Lameter834f3d12008-04-14 19:11:31 +03001517 int order = compound_order(page);
1518 int pages = 1 << order;
Christoph Lameter81819f02007-05-06 14:49:36 -07001519
Laura Abbottbecfda62016-03-15 14:55:06 -07001520 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001521 void *p;
1522
1523 slab_pad_check(s, page);
Christoph Lameter224a88b2008-04-14 19:11:31 +03001524 for_each_object(p, s, page_address(page),
1525 page->objects)
Christoph Lameterf7cb1932010-09-29 07:15:01 -05001526 check_object(s, page, p, SLUB_RED_INACTIVE);
Christoph Lameter81819f02007-05-06 14:49:36 -07001527 }
1528
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001529 kmemcheck_free_shadow(page, compound_order(page));
Vegard Nossum5a896d92008-04-04 00:54:48 +02001530
Christoph Lameter81819f02007-05-06 14:49:36 -07001531 mod_zone_page_state(page_zone(page),
1532 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1533 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
Pekka Enberg06428782008-01-07 23:20:27 -08001534 -pages);
Christoph Lameter81819f02007-05-06 14:49:36 -07001535
Mel Gorman072bb0a2012-07-31 16:43:58 -07001536 __ClearPageSlabPfmemalloc(page);
Christoph Lameter49bd5222008-04-14 18:52:18 +03001537 __ClearPageSlab(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001538
Mel Gorman22b751c2013-02-22 16:34:59 -08001539 page_mapcount_reset(page);
Nick Piggin1eb5ac62009-05-05 19:13:44 +10001540 if (current->reclaim_state)
1541 current->reclaim_state->reclaimed_slab += pages;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001542 memcg_uncharge_slab(page, order, s);
1543 __free_pages(page, order);
Christoph Lameter81819f02007-05-06 14:49:36 -07001544}
1545
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001546#define need_reserve_slab_rcu \
1547 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1548
Christoph Lameter81819f02007-05-06 14:49:36 -07001549static void rcu_free_slab(struct rcu_head *h)
1550{
1551 struct page *page;
1552
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001553 if (need_reserve_slab_rcu)
1554 page = virt_to_head_page(h);
1555 else
1556 page = container_of((struct list_head *)h, struct page, lru);
1557
Glauber Costa1b4f59e32012-10-22 18:05:36 +04001558 __free_slab(page->slab_cache, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001559}
1560
1561static void free_slab(struct kmem_cache *s, struct page *page)
1562{
1563 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001564 struct rcu_head *head;
1565
1566 if (need_reserve_slab_rcu) {
1567 int order = compound_order(page);
1568 int offset = (PAGE_SIZE << order) - s->reserved;
1569
1570 VM_BUG_ON(s->reserved != sizeof(*head));
1571 head = page_address(page) + offset;
1572 } else {
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001573 head = &page->rcu_head;
Lai Jiangshanda9a6382011-03-10 15:22:00 +08001574 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001575
1576 call_rcu(head, rcu_free_slab);
1577 } else
1578 __free_slab(s, page);
1579}
1580
1581static void discard_slab(struct kmem_cache *s, struct page *page)
1582{
Christoph Lameter205ab992008-04-14 19:11:40 +03001583 dec_slabs_node(s, page_to_nid(page), page->objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07001584 free_slab(s, page);
1585}
1586
1587/*
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001588 * Management of partially allocated slabs.
Christoph Lameter81819f02007-05-06 14:49:36 -07001589 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001590static inline void
1591__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
Christoph Lameter81819f02007-05-06 14:49:36 -07001592{
Christoph Lametere95eed52007-05-06 14:49:44 -07001593 n->nr_partial++;
Shaohua Li136333d2011-08-24 08:57:52 +08001594 if (tail == DEACTIVATE_TO_TAIL)
Christoph Lameter7c2e1322008-01-07 23:20:27 -08001595 list_add_tail(&page->lru, &n->partial);
1596 else
1597 list_add(&page->lru, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001598}
1599
Steven Rostedt1e4dd942014-02-10 14:25:46 -08001600static inline void add_partial(struct kmem_cache_node *n,
1601 struct page *page, int tail)
1602{
1603 lockdep_assert_held(&n->list_lock);
1604 __add_partial(n, page, tail);
1605}
1606
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05001607static inline void remove_partial(struct kmem_cache_node *n,
Christoph Lameter62e346a2010-09-28 08:10:28 -05001608 struct page *page)
1609{
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001610 lockdep_assert_held(&n->list_lock);
Dmitry Safonov52b4b952016-02-17 13:11:37 -08001611 list_del(&page->lru);
1612 n->nr_partial--;
Christoph Lameter62e346a2010-09-28 08:10:28 -05001613}
1614
Christoph Lameter81819f02007-05-06 14:49:36 -07001615/*
Christoph Lameter7ced3712012-05-09 10:09:53 -05001616 * Remove slab from the partial list, freeze it and
1617 * return the pointer to the freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07001618 *
Christoph Lameter497b66f2011-08-09 16:12:26 -05001619 * Returns a list of objects or NULL if it fails.
Christoph Lameter81819f02007-05-06 14:49:36 -07001620 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001621static inline void *acquire_slab(struct kmem_cache *s,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001622 struct kmem_cache_node *n, struct page *page,
Joonsoo Kim633b0762013-01-21 17:01:25 +09001623 int mode, int *objects)
Christoph Lameter81819f02007-05-06 14:49:36 -07001624{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001625 void *freelist;
1626 unsigned long counters;
1627 struct page new;
1628
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001629 lockdep_assert_held(&n->list_lock);
1630
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001631 /*
1632 * Zap the freelist and set the frozen bit.
1633 * The old freelist is the list of objects for the
1634 * per cpu allocation list.
1635 */
Christoph Lameter7ced3712012-05-09 10:09:53 -05001636 freelist = page->freelist;
1637 counters = page->counters;
1638 new.counters = counters;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001639 *objects = new.objects - new.inuse;
Pekka Enberg23910c52012-06-04 10:14:58 +03001640 if (mode) {
Christoph Lameter7ced3712012-05-09 10:09:53 -05001641 new.inuse = page->objects;
Pekka Enberg23910c52012-06-04 10:14:58 +03001642 new.freelist = NULL;
1643 } else {
1644 new.freelist = freelist;
1645 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001646
Dave Hansena0132ac2014-01-29 14:05:50 -08001647 VM_BUG_ON(new.frozen);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001648 new.frozen = 1;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001649
Christoph Lameter7ced3712012-05-09 10:09:53 -05001650 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001651 freelist, counters,
Joonsoo Kim02d76332012-05-17 00:13:02 +09001652 new.freelist, new.counters,
Christoph Lameter7ced3712012-05-09 10:09:53 -05001653 "acquire_slab"))
Christoph Lameter7ced3712012-05-09 10:09:53 -05001654 return NULL;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001655
1656 remove_partial(n, page);
Christoph Lameter7ced3712012-05-09 10:09:53 -05001657 WARN_ON(!freelist);
Christoph Lameter49e22582011-08-09 16:12:27 -05001658 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07001659}
1660
Joonsoo Kim633b0762013-01-21 17:01:25 +09001661static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001662static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
Christoph Lameter49e22582011-08-09 16:12:27 -05001663
Christoph Lameter81819f02007-05-06 14:49:36 -07001664/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001665 * Try to allocate a partial slab from a specific node.
Christoph Lameter81819f02007-05-06 14:49:36 -07001666 */
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001667static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1668 struct kmem_cache_cpu *c, gfp_t flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07001669{
Christoph Lameter49e22582011-08-09 16:12:27 -05001670 struct page *page, *page2;
1671 void *object = NULL;
Joonsoo Kim633b0762013-01-21 17:01:25 +09001672 int available = 0;
1673 int objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07001674
1675 /*
1676 * Racy check. If we mistakenly see no partial slabs then we
1677 * just allocate an empty slab. If we mistakenly try to get a
Christoph Lameter672bba32007-05-09 02:32:39 -07001678 * partial slab and there is none available then get_partials()
1679 * will return NULL.
Christoph Lameter81819f02007-05-06 14:49:36 -07001680 */
1681 if (!n || !n->nr_partial)
1682 return NULL;
1683
1684 spin_lock(&n->list_lock);
Christoph Lameter49e22582011-08-09 16:12:27 -05001685 list_for_each_entry_safe(page, page2, &n->partial, lru) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001686 void *t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001687
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001688 if (!pfmemalloc_match(page, flags))
1689 continue;
1690
Joonsoo Kim633b0762013-01-21 17:01:25 +09001691 t = acquire_slab(s, n, page, object == NULL, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05001692 if (!t)
1693 break;
1694
Joonsoo Kim633b0762013-01-21 17:01:25 +09001695 available += objects;
Alex,Shi12d79632011-09-07 10:26:36 +08001696 if (!object) {
Christoph Lameter49e22582011-08-09 16:12:27 -05001697 c->page = page;
Christoph Lameter49e22582011-08-09 16:12:27 -05001698 stat(s, ALLOC_FROM_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05001699 object = t;
Christoph Lameter49e22582011-08-09 16:12:27 -05001700 } else {
Joonsoo Kim633b0762013-01-21 17:01:25 +09001701 put_cpu_partial(s, page, 0);
Alex Shi8028dce2012-02-03 23:34:56 +08001702 stat(s, CPU_PARTIAL_NODE);
Christoph Lameter49e22582011-08-09 16:12:27 -05001703 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09001704 if (!kmem_cache_has_cpu_partial(s)
1705 || available > s->cpu_partial / 2)
Christoph Lameter49e22582011-08-09 16:12:27 -05001706 break;
1707
Christoph Lameter497b66f2011-08-09 16:12:26 -05001708 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001709 spin_unlock(&n->list_lock);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001710 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001711}
1712
1713/*
Christoph Lameter672bba32007-05-09 02:32:39 -07001714 * Get a page from somewhere. Search in increasing NUMA distances.
Christoph Lameter81819f02007-05-06 14:49:36 -07001715 */
Joonsoo Kimde3ec032012-01-27 00:12:23 -08001716static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001717 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001718{
1719#ifdef CONFIG_NUMA
1720 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07001721 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001722 struct zone *zone;
1723 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001724 void *object;
Mel Gormancc9a6c82012-03-21 16:34:11 -07001725 unsigned int cpuset_mems_cookie;
Christoph Lameter81819f02007-05-06 14:49:36 -07001726
1727 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07001728 * The defrag ratio allows a configuration of the tradeoffs between
1729 * inter node defragmentation and node local allocations. A lower
1730 * defrag_ratio increases the tendency to do local allocations
1731 * instead of attempting to obtain partial slabs from other nodes.
Christoph Lameter81819f02007-05-06 14:49:36 -07001732 *
Christoph Lameter672bba32007-05-09 02:32:39 -07001733 * If the defrag_ratio is set to 0 then kmalloc() always
1734 * returns node local objects. If the ratio is higher then kmalloc()
1735 * may return off node objects because partial slabs are obtained
1736 * from other nodes and filled up.
Christoph Lameter81819f02007-05-06 14:49:36 -07001737 *
Christoph Lameter6446faa2008-02-15 23:45:26 -08001738 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
Christoph Lameter672bba32007-05-09 02:32:39 -07001739 * defrag_ratio = 1000) then every (well almost) allocation will
1740 * first attempt to defrag slab caches on other nodes. This means
1741 * scanning over all nodes to look for partial slabs which may be
1742 * expensive if we do it every time we are trying to find a slab
1743 * with available objects.
Christoph Lameter81819f02007-05-06 14:49:36 -07001744 */
Christoph Lameter98246012008-01-07 23:20:26 -08001745 if (!s->remote_node_defrag_ratio ||
1746 get_cycles() % 1024 > s->remote_node_defrag_ratio)
Christoph Lameter81819f02007-05-06 14:49:36 -07001747 return NULL;
1748
Mel Gormancc9a6c82012-03-21 16:34:11 -07001749 do {
Mel Gormand26914d2014-04-03 14:47:24 -07001750 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07001751 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001752 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1753 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07001754
Mel Gormancc9a6c82012-03-21 16:34:11 -07001755 n = get_node(s, zone_to_nid(zone));
Christoph Lameter81819f02007-05-06 14:49:36 -07001756
Vladimir Davydovdee2f8a2014-12-12 16:58:28 -08001757 if (n && cpuset_zone_allowed(zone, flags) &&
Mel Gormancc9a6c82012-03-21 16:34:11 -07001758 n->nr_partial > s->min_partial) {
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001759 object = get_partial_node(s, n, c, flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001760 if (object) {
1761 /*
Mel Gormand26914d2014-04-03 14:47:24 -07001762 * Don't check read_mems_allowed_retry()
1763 * here - if mems_allowed was updated in
1764 * parallel, that was a harmless race
1765 * between allocation and the cpuset
1766 * update
Mel Gormancc9a6c82012-03-21 16:34:11 -07001767 */
Mel Gormancc9a6c82012-03-21 16:34:11 -07001768 return object;
1769 }
Miao Xiec0ff7452010-05-24 14:32:08 -07001770 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001771 }
Mel Gormand26914d2014-04-03 14:47:24 -07001772 } while (read_mems_allowed_retry(cpuset_mems_cookie));
Christoph Lameter81819f02007-05-06 14:49:36 -07001773#endif
1774 return NULL;
1775}
1776
1777/*
1778 * Get a partial page, lock it and return it.
1779 */
Christoph Lameter497b66f2011-08-09 16:12:26 -05001780static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001781 struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07001782{
Christoph Lameter497b66f2011-08-09 16:12:26 -05001783 void *object;
Joonsoo Kima561ce02014-10-09 15:26:15 -07001784 int searchnode = node;
1785
1786 if (node == NUMA_NO_NODE)
1787 searchnode = numa_mem_id();
1788 else if (!node_present_pages(node))
1789 searchnode = node_to_mem_node(node);
Christoph Lameter81819f02007-05-06 14:49:36 -07001790
Joonsoo Kim8ba00bb2012-09-17 14:09:09 -07001791 object = get_partial_node(s, get_node(s, searchnode), c, flags);
Christoph Lameter497b66f2011-08-09 16:12:26 -05001792 if (object || node != NUMA_NO_NODE)
1793 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07001794
Christoph Lameteracd19fd2011-08-09 16:12:25 -05001795 return get_any_partial(s, flags, c);
Christoph Lameter81819f02007-05-06 14:49:36 -07001796}
1797
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001798#ifdef CONFIG_PREEMPT
1799/*
1800 * Calculate the next globally unique transaction for disambiguiation
1801 * during cmpxchg. The transactions start with the cpu number and are then
1802 * incremented by CONFIG_NR_CPUS.
1803 */
1804#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1805#else
1806/*
1807 * No preemption supported therefore also no need to check for
1808 * different cpus.
1809 */
1810#define TID_STEP 1
1811#endif
1812
1813static inline unsigned long next_tid(unsigned long tid)
1814{
1815 return tid + TID_STEP;
1816}
1817
1818static inline unsigned int tid_to_cpu(unsigned long tid)
1819{
1820 return tid % TID_STEP;
1821}
1822
1823static inline unsigned long tid_to_event(unsigned long tid)
1824{
1825 return tid / TID_STEP;
1826}
1827
1828static inline unsigned int init_tid(int cpu)
1829{
1830 return cpu;
1831}
1832
1833static inline void note_cmpxchg_failure(const char *n,
1834 const struct kmem_cache *s, unsigned long tid)
1835{
1836#ifdef SLUB_DEBUG_CMPXCHG
1837 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1838
Fabian Frederickf9f58282014-06-04 16:06:34 -07001839 pr_info("%s %s: cmpxchg redo ", n, s->name);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001840
1841#ifdef CONFIG_PREEMPT
1842 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
Fabian Frederickf9f58282014-06-04 16:06:34 -07001843 pr_warn("due to cpu change %d -> %d\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001844 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1845 else
1846#endif
1847 if (tid_to_event(tid) != tid_to_event(actual_tid))
Fabian Frederickf9f58282014-06-04 16:06:34 -07001848 pr_warn("due to cpu running other code. Event %ld->%ld\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001849 tid_to_event(tid), tid_to_event(actual_tid));
1850 else
Fabian Frederickf9f58282014-06-04 16:06:34 -07001851 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001852 actual_tid, tid, next_tid(tid));
1853#endif
Christoph Lameter4fdccdf2011-03-22 13:35:00 -05001854 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001855}
1856
Fengguang Wu788e1aa2012-09-28 16:34:05 +08001857static void init_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001858{
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001859 int cpu;
1860
1861 for_each_possible_cpu(cpu)
1862 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06001863}
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001864
1865/*
1866 * Remove the cpu slab
1867 */
Chen Gangd0e0ac92013-07-15 09:05:29 +08001868static void deactivate_slab(struct kmem_cache *s, struct page *page,
1869 void *freelist)
Christoph Lameter81819f02007-05-06 14:49:36 -07001870{
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001871 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001872 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1873 int lock = 0;
1874 enum slab_modes l = M_NONE, m = M_NONE;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001875 void *nextfree;
Shaohua Li136333d2011-08-24 08:57:52 +08001876 int tail = DEACTIVATE_TO_HEAD;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001877 struct page new;
1878 struct page old;
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08001879
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001880 if (page->freelist) {
Christoph Lameter84e554e62009-12-18 16:26:23 -06001881 stat(s, DEACTIVATE_REMOTE_FREES);
Shaohua Li136333d2011-08-24 08:57:52 +08001882 tail = DEACTIVATE_TO_TAIL;
Christoph Lameter894b8782007-05-10 03:15:16 -07001883 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001884
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001885 /*
1886 * Stage one: Free all available per cpu objects back
1887 * to the page freelist while it is still frozen. Leave the
1888 * last one.
1889 *
1890 * There is no need to take the list->lock because the page
1891 * is still frozen.
1892 */
1893 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1894 void *prior;
1895 unsigned long counters;
1896
1897 do {
1898 prior = page->freelist;
1899 counters = page->counters;
1900 set_freepointer(s, freelist, prior);
1901 new.counters = counters;
1902 new.inuse--;
Dave Hansena0132ac2014-01-29 14:05:50 -08001903 VM_BUG_ON(!new.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001904
Christoph Lameter1d071712011-07-14 12:49:12 -05001905 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001906 prior, counters,
1907 freelist, new.counters,
1908 "drain percpu freelist"));
1909
1910 freelist = nextfree;
1911 }
1912
1913 /*
1914 * Stage two: Ensure that the page is unfrozen while the
1915 * list presence reflects the actual number of objects
1916 * during unfreeze.
1917 *
1918 * We setup the list membership and then perform a cmpxchg
1919 * with the count. If there is a mismatch then the page
1920 * is not unfrozen but the page is on the wrong list.
1921 *
1922 * Then we restart the process which may have to remove
1923 * the page from the list that we just put it on again
1924 * because the number of objects in the slab may have
1925 * changed.
1926 */
1927redo:
1928
1929 old.freelist = page->freelist;
1930 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08001931 VM_BUG_ON(!old.frozen);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001932
1933 /* Determine target state of the slab */
1934 new.counters = old.counters;
1935 if (freelist) {
1936 new.inuse--;
1937 set_freepointer(s, freelist, old.freelist);
1938 new.freelist = freelist;
1939 } else
1940 new.freelist = old.freelist;
1941
1942 new.frozen = 0;
1943
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07001944 if (!new.inuse && n->nr_partial >= s->min_partial)
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001945 m = M_FREE;
1946 else if (new.freelist) {
1947 m = M_PARTIAL;
1948 if (!lock) {
1949 lock = 1;
1950 /*
1951 * Taking the spinlock removes the possiblity
1952 * that acquire_slab() will see a slab page that
1953 * is frozen
1954 */
1955 spin_lock(&n->list_lock);
1956 }
1957 } else {
1958 m = M_FULL;
1959 if (kmem_cache_debug(s) && !lock) {
1960 lock = 1;
1961 /*
1962 * This also ensures that the scanning of full
1963 * slabs from diagnostic functions will not see
1964 * any frozen slabs.
1965 */
1966 spin_lock(&n->list_lock);
1967 }
1968 }
1969
1970 if (l != m) {
1971
1972 if (l == M_PARTIAL)
1973
1974 remove_partial(n, page);
1975
1976 else if (l == M_FULL)
1977
Peter Zijlstrac65c1872014-01-10 13:23:49 +01001978 remove_full(s, n, page);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001979
1980 if (m == M_PARTIAL) {
1981
1982 add_partial(n, page, tail);
Shaohua Li136333d2011-08-24 08:57:52 +08001983 stat(s, tail);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001984
1985 } else if (m == M_FULL) {
1986
1987 stat(s, DEACTIVATE_FULL);
1988 add_full(s, n, page);
1989
1990 }
1991 }
1992
1993 l = m;
Christoph Lameter1d071712011-07-14 12:49:12 -05001994 if (!__cmpxchg_double_slab(s, page,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05001995 old.freelist, old.counters,
1996 new.freelist, new.counters,
1997 "unfreezing slab"))
1998 goto redo;
1999
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002000 if (lock)
2001 spin_unlock(&n->list_lock);
2002
2003 if (m == M_FREE) {
2004 stat(s, DEACTIVATE_EMPTY);
2005 discard_slab(s, page);
2006 stat(s, FREE_SLAB);
2007 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002008}
2009
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002010/*
2011 * Unfreeze all the cpu partial slabs.
2012 *
Christoph Lameter59a09912012-11-28 16:23:00 +00002013 * This function must be called with interrupts disabled
2014 * for the cpu using c (or some other guarantee must be there
2015 * to guarantee no concurrent accesses).
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002016 */
Christoph Lameter59a09912012-11-28 16:23:00 +00002017static void unfreeze_partials(struct kmem_cache *s,
2018 struct kmem_cache_cpu *c)
Christoph Lameter49e22582011-08-09 16:12:27 -05002019{
Joonsoo Kim345c9052013-06-19 14:05:52 +09002020#ifdef CONFIG_SLUB_CPU_PARTIAL
Joonsoo Kim43d77862012-06-09 02:23:16 +09002021 struct kmem_cache_node *n = NULL, *n2 = NULL;
Shaohua Li9ada1932011-11-14 13:34:13 +08002022 struct page *page, *discard_page = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05002023
2024 while ((page = c->partial)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002025 struct page new;
2026 struct page old;
2027
2028 c->partial = page->next;
Joonsoo Kim43d77862012-06-09 02:23:16 +09002029
2030 n2 = get_node(s, page_to_nid(page));
2031 if (n != n2) {
2032 if (n)
2033 spin_unlock(&n->list_lock);
2034
2035 n = n2;
2036 spin_lock(&n->list_lock);
2037 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002038
2039 do {
2040
2041 old.freelist = page->freelist;
2042 old.counters = page->counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08002043 VM_BUG_ON(!old.frozen);
Christoph Lameter49e22582011-08-09 16:12:27 -05002044
2045 new.counters = old.counters;
2046 new.freelist = old.freelist;
2047
2048 new.frozen = 0;
2049
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002050 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter49e22582011-08-09 16:12:27 -05002051 old.freelist, old.counters,
2052 new.freelist, new.counters,
2053 "unfreezing slab"));
2054
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07002055 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
Shaohua Li9ada1932011-11-14 13:34:13 +08002056 page->next = discard_page;
2057 discard_page = page;
Joonsoo Kim43d77862012-06-09 02:23:16 +09002058 } else {
2059 add_partial(n, page, DEACTIVATE_TO_TAIL);
2060 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter49e22582011-08-09 16:12:27 -05002061 }
2062 }
2063
2064 if (n)
2065 spin_unlock(&n->list_lock);
Shaohua Li9ada1932011-11-14 13:34:13 +08002066
2067 while (discard_page) {
2068 page = discard_page;
2069 discard_page = discard_page->next;
2070
2071 stat(s, DEACTIVATE_EMPTY);
2072 discard_slab(s, page);
2073 stat(s, FREE_SLAB);
2074 }
Joonsoo Kim345c9052013-06-19 14:05:52 +09002075#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002076}
2077
2078/*
2079 * Put a page that was just frozen (in __slab_free) into a partial page
2080 * slot if available. This is done without interrupts disabled and without
2081 * preemption disabled. The cmpxchg is racy and may put the partial page
2082 * onto a random cpus partial slot.
2083 *
2084 * If we did not find a slot then simply move all the partials to the
2085 * per node partial list.
2086 */
Joonsoo Kim633b0762013-01-21 17:01:25 +09002087static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
Christoph Lameter49e22582011-08-09 16:12:27 -05002088{
Joonsoo Kim345c9052013-06-19 14:05:52 +09002089#ifdef CONFIG_SLUB_CPU_PARTIAL
Christoph Lameter49e22582011-08-09 16:12:27 -05002090 struct page *oldpage;
2091 int pages;
2092 int pobjects;
2093
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002094 preempt_disable();
Christoph Lameter49e22582011-08-09 16:12:27 -05002095 do {
2096 pages = 0;
2097 pobjects = 0;
2098 oldpage = this_cpu_read(s->cpu_slab->partial);
2099
2100 if (oldpage) {
2101 pobjects = oldpage->pobjects;
2102 pages = oldpage->pages;
2103 if (drain && pobjects > s->cpu_partial) {
2104 unsigned long flags;
2105 /*
2106 * partial array is full. Move the existing
2107 * set to the per node partial list.
2108 */
2109 local_irq_save(flags);
Christoph Lameter59a09912012-11-28 16:23:00 +00002110 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
Christoph Lameter49e22582011-08-09 16:12:27 -05002111 local_irq_restore(flags);
Joonsoo Kime24fc412012-06-23 03:22:38 +09002112 oldpage = NULL;
Christoph Lameter49e22582011-08-09 16:12:27 -05002113 pobjects = 0;
2114 pages = 0;
Alex Shi8028dce2012-02-03 23:34:56 +08002115 stat(s, CPU_PARTIAL_DRAIN);
Christoph Lameter49e22582011-08-09 16:12:27 -05002116 }
2117 }
2118
2119 pages++;
2120 pobjects += page->objects - page->inuse;
2121
2122 page->pages = pages;
2123 page->pobjects = pobjects;
2124 page->next = oldpage;
2125
Chen Gangd0e0ac92013-07-15 09:05:29 +08002126 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2127 != oldpage);
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002128 if (unlikely(!s->cpu_partial)) {
2129 unsigned long flags;
2130
2131 local_irq_save(flags);
2132 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2133 local_irq_restore(flags);
2134 }
2135 preempt_enable();
Joonsoo Kim345c9052013-06-19 14:05:52 +09002136#endif
Christoph Lameter49e22582011-08-09 16:12:27 -05002137}
2138
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002139static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002140{
Christoph Lameter84e554e62009-12-18 16:26:23 -06002141 stat(s, CPUSLAB_FLUSH);
Christoph Lameterc17dda42012-05-09 10:09:57 -05002142 deactivate_slab(s, c->page, c->freelist);
2143
2144 c->tid = next_tid(c->tid);
2145 c->page = NULL;
2146 c->freelist = NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002147}
2148
2149/*
2150 * Flush cpu slab.
Christoph Lameter6446faa2008-02-15 23:45:26 -08002151 *
Christoph Lameter81819f02007-05-06 14:49:36 -07002152 * Called from IPI handler with interrupts disabled.
2153 */
Christoph Lameter0c710012007-07-17 04:03:24 -07002154static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
Christoph Lameter81819f02007-05-06 14:49:36 -07002155{
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002156 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
Christoph Lameter81819f02007-05-06 14:49:36 -07002157
Christoph Lameter49e22582011-08-09 16:12:27 -05002158 if (likely(c)) {
2159 if (c->page)
2160 flush_slab(s, c);
2161
Christoph Lameter59a09912012-11-28 16:23:00 +00002162 unfreeze_partials(s, c);
Christoph Lameter49e22582011-08-09 16:12:27 -05002163 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002164}
2165
2166static void flush_cpu_slab(void *d)
2167{
2168 struct kmem_cache *s = d;
Christoph Lameter81819f02007-05-06 14:49:36 -07002169
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002170 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameter81819f02007-05-06 14:49:36 -07002171}
2172
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002173static bool has_cpu_slab(int cpu, void *info)
2174{
2175 struct kmem_cache *s = info;
2176 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2177
majianpeng02e1a9c2012-05-17 17:03:26 -07002178 return c->page || c->partial;
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002179}
2180
Christoph Lameter81819f02007-05-06 14:49:36 -07002181static void flush_all(struct kmem_cache *s)
2182{
Gilad Ben-Yossefa8364d52012-03-28 14:42:44 -07002183 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
Christoph Lameter81819f02007-05-06 14:49:36 -07002184}
2185
2186/*
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002187 * Check if the objects in a per cpu structure fit numa
2188 * locality expectations.
2189 */
Christoph Lameter57d437d2012-05-09 10:09:59 -05002190static inline int node_match(struct page *page, int node)
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002191{
2192#ifdef CONFIG_NUMA
Christoph Lameter4d7868e2013-01-23 21:45:47 +00002193 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002194 return 0;
2195#endif
2196 return 1;
2197}
2198
David Rientjes9a02d692014-06-04 16:06:36 -07002199#ifdef CONFIG_SLUB_DEBUG
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002200static int count_free(struct page *page)
2201{
2202 return page->objects - page->inuse;
2203}
2204
David Rientjes9a02d692014-06-04 16:06:36 -07002205static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2206{
2207 return atomic_long_read(&n->total_objects);
2208}
2209#endif /* CONFIG_SLUB_DEBUG */
2210
2211#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002212static unsigned long count_partial(struct kmem_cache_node *n,
2213 int (*get_count)(struct page *))
2214{
2215 unsigned long flags;
2216 unsigned long x = 0;
2217 struct page *page;
2218
2219 spin_lock_irqsave(&n->list_lock, flags);
2220 list_for_each_entry(page, &n->partial, lru)
2221 x += get_count(page);
2222 spin_unlock_irqrestore(&n->list_lock, flags);
2223 return x;
2224}
David Rientjes9a02d692014-06-04 16:06:36 -07002225#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002226
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002227static noinline void
2228slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2229{
David Rientjes9a02d692014-06-04 16:06:36 -07002230#ifdef CONFIG_SLUB_DEBUG
2231 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2232 DEFAULT_RATELIMIT_BURST);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002233 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07002234 struct kmem_cache_node *n;
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002235
David Rientjes9a02d692014-06-04 16:06:36 -07002236 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2237 return;
2238
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07002239 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2240 nid, gfpflags, &gfpflags);
Fabian Frederickf9f58282014-06-04 16:06:34 -07002241 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2242 s->name, s->object_size, s->size, oo_order(s->oo),
2243 oo_order(s->min));
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002244
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002245 if (oo_order(s->min) > get_order(s->object_size))
Fabian Frederickf9f58282014-06-04 16:06:34 -07002246 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2247 s->name);
David Rientjesfa5ec8a2009-07-07 00:14:14 -07002248
Christoph Lameterfa45dc22014-08-06 16:04:09 -07002249 for_each_kmem_cache_node(s, node, n) {
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002250 unsigned long nr_slabs;
2251 unsigned long nr_objs;
2252 unsigned long nr_free;
2253
Alexander Beregalov26c02cf2009-06-11 14:08:48 +04002254 nr_free = count_partial(n, count_free);
2255 nr_slabs = node_nr_slabs(n);
2256 nr_objs = node_nr_objs(n);
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002257
Fabian Frederickf9f58282014-06-04 16:06:34 -07002258 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002259 node, nr_slabs, nr_objs, nr_free);
2260 }
David Rientjes9a02d692014-06-04 16:06:36 -07002261#endif
Pekka Enberg781b2ba2009-06-10 18:50:32 +03002262}
2263
Christoph Lameter497b66f2011-08-09 16:12:26 -05002264static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2265 int node, struct kmem_cache_cpu **pc)
2266{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002267 void *freelist;
Christoph Lameter188fd062012-05-09 10:09:55 -05002268 struct kmem_cache_cpu *c = *pc;
2269 struct page *page;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002270
Christoph Lameter188fd062012-05-09 10:09:55 -05002271 freelist = get_partial(s, flags, node, c);
2272
2273 if (freelist)
2274 return freelist;
2275
2276 page = new_slab(s, flags, node);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002277 if (page) {
Christoph Lameter7c8e0182014-06-04 16:07:56 -07002278 c = raw_cpu_ptr(s->cpu_slab);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002279 if (c->page)
2280 flush_slab(s, c);
2281
2282 /*
2283 * No other reference to the page yet so we can
2284 * muck around with it freely without cmpxchg
2285 */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002286 freelist = page->freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002287 page->freelist = NULL;
2288
2289 stat(s, ALLOC_SLAB);
Christoph Lameter497b66f2011-08-09 16:12:26 -05002290 c->page = page;
2291 *pc = c;
2292 } else
Christoph Lameter6faa6832012-05-09 10:09:51 -05002293 freelist = NULL;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002294
Christoph Lameter6faa6832012-05-09 10:09:51 -05002295 return freelist;
Christoph Lameter497b66f2011-08-09 16:12:26 -05002296}
2297
Mel Gorman072bb0a2012-07-31 16:43:58 -07002298static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2299{
2300 if (unlikely(PageSlabPfmemalloc(page)))
2301 return gfp_pfmemalloc_allowed(gfpflags);
2302
2303 return true;
2304}
2305
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002306/*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002307 * Check the page->freelist of a page and either transfer the freelist to the
2308 * per cpu freelist or deactivate the page.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002309 *
2310 * The page is still frozen if the return value is not NULL.
2311 *
2312 * If this function returns NULL then the page has been unfrozen.
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002313 *
2314 * This function must be called with interrupt disabled.
Christoph Lameter213eeb92011-11-11 14:07:14 -06002315 */
2316static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2317{
2318 struct page new;
2319 unsigned long counters;
2320 void *freelist;
2321
2322 do {
2323 freelist = page->freelist;
2324 counters = page->counters;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002325
Christoph Lameter213eeb92011-11-11 14:07:14 -06002326 new.counters = counters;
Dave Hansena0132ac2014-01-29 14:05:50 -08002327 VM_BUG_ON(!new.frozen);
Christoph Lameter213eeb92011-11-11 14:07:14 -06002328
2329 new.inuse = page->objects;
2330 new.frozen = freelist != NULL;
2331
Joonsoo Kimd24ac772012-05-18 22:01:17 +09002332 } while (!__cmpxchg_double_slab(s, page,
Christoph Lameter213eeb92011-11-11 14:07:14 -06002333 freelist, counters,
2334 NULL, new.counters,
2335 "get_freelist"));
2336
2337 return freelist;
2338}
2339
2340/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002341 * Slow path. The lockless freelist is empty or we need to perform
2342 * debugging duties.
Christoph Lameter81819f02007-05-06 14:49:36 -07002343 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002344 * Processing is still very fast if new objects have been freed to the
2345 * regular freelist. In that case we simply take over the regular freelist
2346 * as the lockless freelist and zap the regular freelist.
Christoph Lameter81819f02007-05-06 14:49:36 -07002347 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002348 * If that is not working then we fall back to the partial lists. We take the
2349 * first element of the freelist as the object to allocate now and move the
2350 * rest of the freelist to the lockless freelist.
2351 *
2352 * And if we were unable to get a new slab from the partial slab lists then
Christoph Lameter6446faa2008-02-15 23:45:26 -08002353 * we need to allocate a new slab. This is the slowest path since it involves
2354 * a call to the page allocator and the setup of a new slab.
Christoph Lametera380a3c2015-11-20 15:57:35 -08002355 *
2356 * Version of __slab_alloc to use when we know that interrupts are
2357 * already disabled (which is the case for bulk allocation).
Christoph Lameter81819f02007-05-06 14:49:36 -07002358 */
Christoph Lametera380a3c2015-11-20 15:57:35 -08002359static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002360 unsigned long addr, struct kmem_cache_cpu *c)
Christoph Lameter81819f02007-05-06 14:49:36 -07002361{
Christoph Lameter6faa6832012-05-09 10:09:51 -05002362 void *freelist;
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002363 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07002364
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002365 page = c->page;
2366 if (!page)
Christoph Lameter81819f02007-05-06 14:49:36 -07002367 goto new_slab;
Christoph Lameter49e22582011-08-09 16:12:27 -05002368redo:
Christoph Lameter6faa6832012-05-09 10:09:51 -05002369
Christoph Lameter57d437d2012-05-09 10:09:59 -05002370 if (unlikely(!node_match(page, node))) {
Joonsoo Kima561ce02014-10-09 15:26:15 -07002371 int searchnode = node;
2372
2373 if (node != NUMA_NO_NODE && !node_present_pages(node))
2374 searchnode = node_to_mem_node(node);
2375
2376 if (unlikely(!node_match(page, searchnode))) {
2377 stat(s, ALLOC_NODE_MISMATCH);
2378 deactivate_slab(s, page, c->freelist);
2379 c->page = NULL;
2380 c->freelist = NULL;
2381 goto new_slab;
2382 }
Christoph Lameterfc59c052011-06-01 12:25:56 -05002383 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08002384
Mel Gorman072bb0a2012-07-31 16:43:58 -07002385 /*
2386 * By rights, we should be searching for a slab page that was
2387 * PFMEMALLOC but right now, we are losing the pfmemalloc
2388 * information when the page leaves the per-cpu allocator
2389 */
2390 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2391 deactivate_slab(s, page, c->freelist);
2392 c->page = NULL;
2393 c->freelist = NULL;
2394 goto new_slab;
2395 }
2396
Eric Dumazet73736e02011-12-13 04:57:06 +01002397 /* must check again c->freelist in case of cpu migration or IRQ */
Christoph Lameter6faa6832012-05-09 10:09:51 -05002398 freelist = c->freelist;
2399 if (freelist)
Eric Dumazet73736e02011-12-13 04:57:06 +01002400 goto load_freelist;
2401
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002402 freelist = get_freelist(s, page);
Christoph Lameter6446faa2008-02-15 23:45:26 -08002403
Christoph Lameter6faa6832012-05-09 10:09:51 -05002404 if (!freelist) {
Christoph Lameter03e404a2011-06-01 12:25:58 -05002405 c->page = NULL;
2406 stat(s, DEACTIVATE_BYPASS);
Christoph Lameterfc59c052011-06-01 12:25:56 -05002407 goto new_slab;
Christoph Lameter03e404a2011-06-01 12:25:58 -05002408 }
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002409
Christoph Lameter81819f02007-05-06 14:49:36 -07002410 stat(s, ALLOC_REFILL);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08002411
Christoph Lameter894b8782007-05-10 03:15:16 -07002412load_freelist:
Christoph Lameter507effe2012-05-09 10:09:52 -05002413 /*
2414 * freelist is pointing to the list of objects to be used.
2415 * page is pointing to the page from which the objects are obtained.
2416 * That page must be frozen for per cpu allocations to work.
2417 */
Dave Hansena0132ac2014-01-29 14:05:50 -08002418 VM_BUG_ON(!c->page->frozen);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002419 c->freelist = get_freepointer(s, freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002420 c->tid = next_tid(c->tid);
Christoph Lameter6faa6832012-05-09 10:09:51 -05002421 return freelist;
Christoph Lameter81819f02007-05-06 14:49:36 -07002422
Christoph Lameter81819f02007-05-06 14:49:36 -07002423new_slab:
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002424
Christoph Lameter49e22582011-08-09 16:12:27 -05002425 if (c->partial) {
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002426 page = c->page = c->partial;
2427 c->partial = page->next;
Christoph Lameter49e22582011-08-09 16:12:27 -05002428 stat(s, CPU_PARTIAL_ALLOC);
2429 c->freelist = NULL;
2430 goto redo;
Christoph Lameter81819f02007-05-06 14:49:36 -07002431 }
2432
Christoph Lameter188fd062012-05-09 10:09:55 -05002433 freelist = new_slab_objects(s, gfpflags, node, &c);
Christoph Lameterb811c202007-10-16 23:25:51 -07002434
Christoph Lameterf46974362012-05-09 10:09:54 -05002435 if (unlikely(!freelist)) {
David Rientjes9a02d692014-06-04 16:06:36 -07002436 slab_out_of_memory(s, gfpflags, node);
Christoph Lameterf46974362012-05-09 10:09:54 -05002437 return NULL;
Christoph Lameter81819f02007-05-06 14:49:36 -07002438 }
Christoph Lameter894b8782007-05-10 03:15:16 -07002439
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002440 page = c->page;
Christoph Lameter5091b742012-07-31 16:44:00 -07002441 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
Christoph Lameter81819f02007-05-06 14:49:36 -07002442 goto load_freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002443
Christoph Lameter497b66f2011-08-09 16:12:26 -05002444 /* Only entered in the debug case */
Chen Gangd0e0ac92013-07-15 09:05:29 +08002445 if (kmem_cache_debug(s) &&
2446 !alloc_debug_processing(s, page, freelist, addr))
Christoph Lameter497b66f2011-08-09 16:12:26 -05002447 goto new_slab; /* Slab failed checks. Next slab needed */
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002448
Christoph Lameterf6e7def2012-05-09 10:09:58 -05002449 deactivate_slab(s, page, get_freepointer(s, freelist));
Christoph Lameterc17dda42012-05-09 10:09:57 -05002450 c->page = NULL;
2451 c->freelist = NULL;
Christoph Lameter6faa6832012-05-09 10:09:51 -05002452 return freelist;
Christoph Lameter894b8782007-05-10 03:15:16 -07002453}
2454
2455/*
Christoph Lametera380a3c2015-11-20 15:57:35 -08002456 * Another one that disabled interrupt and compensates for possible
2457 * cpu changes by refetching the per cpu area pointer.
2458 */
2459static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2460 unsigned long addr, struct kmem_cache_cpu *c)
2461{
2462 void *p;
2463 unsigned long flags;
2464
2465 local_irq_save(flags);
2466#ifdef CONFIG_PREEMPT
2467 /*
2468 * We may have been preempted and rescheduled on a different
2469 * cpu before disabling interrupts. Need to reload cpu area
2470 * pointer.
2471 */
2472 c = this_cpu_ptr(s->cpu_slab);
2473#endif
2474
2475 p = ___slab_alloc(s, gfpflags, node, addr, c);
2476 local_irq_restore(flags);
2477 return p;
2478}
2479
2480/*
Christoph Lameter894b8782007-05-10 03:15:16 -07002481 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2482 * have the fastpath folded into their functions. So no function call
2483 * overhead for requests that can be satisfied on the fastpath.
2484 *
2485 * The fastpath works by first checking if the lockless freelist can be used.
2486 * If not then __slab_alloc is called for slow processing.
2487 *
2488 * Otherwise we can simply pick the next object from the lockless free list.
2489 */
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002490static __always_inline void *slab_alloc_node(struct kmem_cache *s,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03002491 gfp_t gfpflags, int node, unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002492{
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002493 void *object;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002494 struct kmem_cache_cpu *c;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002495 struct page *page;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002496 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002497
Vladimir Davydov8135be52014-12-12 16:56:38 -08002498 s = slab_pre_alloc_hook(s, gfpflags);
2499 if (!s)
Akinobu Mita773ff602008-12-23 19:37:01 +09002500 return NULL;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002501redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002502 /*
2503 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2504 * enabled. We may switch back and forth between cpus while
2505 * reading from one cpu area. That does not matter as long
2506 * as we end up on the original cpu again when doing the cmpxchg.
Christoph Lameter7cccd802013-01-23 21:45:48 +00002507 *
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002508 * We should guarantee that tid and kmem_cache are retrieved on
2509 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2510 * to check if it is matched or not.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002511 */
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002512 do {
2513 tid = this_cpu_read(s->cpu_slab->tid);
2514 c = raw_cpu_ptr(s->cpu_slab);
Mark Rutland859b7a02015-03-25 15:55:23 -07002515 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2516 unlikely(tid != READ_ONCE(c->tid)));
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002517
2518 /*
2519 * Irqless object alloc/free algorithm used here depends on sequence
2520 * of fetching cpu_slab's data. tid should be fetched before anything
2521 * on c to guarantee that object and page associated with previous tid
2522 * won't be used with current tid. If we fetch tid first, object and
2523 * page could be one associated with next tid and our alloc/free
2524 * request will be failed. In this case, we will retry. So, no problem.
2525 */
2526 barrier();
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002527
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002528 /*
2529 * The transaction ids are globally unique per cpu and per operation on
2530 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2531 * occurs on the right processor and that there was no operation on the
2532 * linked list in between.
2533 */
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002534
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06002535 object = c->freelist;
Christoph Lameter57d437d2012-05-09 10:09:59 -05002536 page = c->page;
Dave Hansen8eae1492014-06-04 16:06:37 -07002537 if (unlikely(!object || !node_match(page, node))) {
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002538 object = __slab_alloc(s, gfpflags, node, addr, c);
Dave Hansen8eae1492014-06-04 16:06:37 -07002539 stat(s, ALLOC_SLOWPATH);
2540 } else {
Eric Dumazet0ad95002011-12-16 16:25:34 +01002541 void *next_object = get_freepointer_safe(s, object);
2542
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002543 /*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002544 * The cmpxchg will only match if there was no additional
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002545 * operation and if we are on the right processor.
2546 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002547 * The cmpxchg does the following atomically (without lock
2548 * semantics!)
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002549 * 1. Relocate first pointer to the current per cpu area.
2550 * 2. Verify that tid and freelist have not been changed
2551 * 3. If they were not changed replace tid and freelist
2552 *
Chen Gangd0e0ac92013-07-15 09:05:29 +08002553 * Since this is without lock semantics the protection is only
2554 * against code executing on this cpu *not* from access by
2555 * other cpus.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002556 */
Christoph Lameter933393f2011-12-22 11:58:51 -06002557 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002558 s->cpu_slab->freelist, s->cpu_slab->tid,
2559 object, tid,
Eric Dumazet0ad95002011-12-16 16:25:34 +01002560 next_object, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002561
2562 note_cmpxchg_failure("slab_alloc", s, tid);
2563 goto redo;
2564 }
Eric Dumazet0ad95002011-12-16 16:25:34 +01002565 prefetch_freepointer(s, next_object);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002566 stat(s, ALLOC_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002567 }
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002568
Pekka Enberg74e21342009-11-25 20:14:48 +02002569 if (unlikely(gfpflags & __GFP_ZERO) && object)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002570 memset(object, 0, s->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002571
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002572 slab_post_alloc_hook(s, gfpflags, 1, &object);
Vegard Nossum5a896d92008-04-04 00:54:48 +02002573
Christoph Lameter894b8782007-05-10 03:15:16 -07002574 return object;
Christoph Lameter81819f02007-05-06 14:49:36 -07002575}
2576
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002577static __always_inline void *slab_alloc(struct kmem_cache *s,
2578 gfp_t gfpflags, unsigned long addr)
2579{
2580 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2581}
2582
Christoph Lameter81819f02007-05-06 14:49:36 -07002583void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2584{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002585 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002586
Chen Gangd0e0ac92013-07-15 09:05:29 +08002587 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2588 s->size, gfpflags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002589
2590 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002591}
2592EXPORT_SYMBOL(kmem_cache_alloc);
2593
Li Zefan0f24f122009-12-11 15:45:30 +08002594#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002595void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002596{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002597 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002598 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07002599 kasan_kmalloc(s, ret, size, gfpflags);
Richard Kennedy4a923792010-10-21 10:29:19 +01002600 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002601}
Richard Kennedy4a923792010-10-21 10:29:19 +01002602EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002603#endif
2604
Christoph Lameter81819f02007-05-06 14:49:36 -07002605#ifdef CONFIG_NUMA
2606void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2607{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002608 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002609
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002610 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002611 s->object_size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002612
2613 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07002614}
2615EXPORT_SYMBOL(kmem_cache_alloc_node);
Christoph Lameter81819f02007-05-06 14:49:36 -07002616
Li Zefan0f24f122009-12-11 15:45:30 +08002617#ifdef CONFIG_TRACING
Richard Kennedy4a923792010-10-21 10:29:19 +01002618void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002619 gfp_t gfpflags,
Richard Kennedy4a923792010-10-21 10:29:19 +01002620 int node, size_t size)
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002621{
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03002622 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
Richard Kennedy4a923792010-10-21 10:29:19 +01002623
2624 trace_kmalloc_node(_RET_IP_, ret,
2625 size, s->size, gfpflags, node);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08002626
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07002627 kasan_kmalloc(s, ret, size, gfpflags);
Richard Kennedy4a923792010-10-21 10:29:19 +01002628 return ret;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002629}
Richard Kennedy4a923792010-10-21 10:29:19 +01002630EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002631#endif
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09002632#endif
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03002633
Christoph Lameter81819f02007-05-06 14:49:36 -07002634/*
Kim Phillips94e4d712015-02-10 14:09:37 -08002635 * Slow path handling. This may still be called frequently since objects
Christoph Lameter894b8782007-05-10 03:15:16 -07002636 * have a longer lifetime than the cpu slabs in most processing loads.
Christoph Lameter81819f02007-05-06 14:49:36 -07002637 *
Christoph Lameter894b8782007-05-10 03:15:16 -07002638 * So we still attempt to reduce cache line usage. Just take the slab
2639 * lock and free the item. If there is no additional partial page
2640 * handling required then we can return immediately.
Christoph Lameter81819f02007-05-06 14:49:36 -07002641 */
Christoph Lameter894b8782007-05-10 03:15:16 -07002642static void __slab_free(struct kmem_cache *s, struct page *page,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002643 void *head, void *tail, int cnt,
2644 unsigned long addr)
2645
Christoph Lameter81819f02007-05-06 14:49:36 -07002646{
2647 void *prior;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002648 int was_frozen;
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002649 struct page new;
2650 unsigned long counters;
2651 struct kmem_cache_node *n = NULL;
Christoph Lameter61728d12011-06-01 12:25:51 -05002652 unsigned long uninitialized_var(flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002653
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002654 stat(s, FREE_SLOWPATH);
Christoph Lameter81819f02007-05-06 14:49:36 -07002655
Christoph Lameter19c7ff92012-05-30 12:54:46 -05002656 if (kmem_cache_debug(s) &&
Laura Abbott282acb42016-03-15 14:54:59 -07002657 !free_debug_processing(s, page, head, tail, cnt, addr))
Christoph Lameter80f08c12011-06-01 12:25:55 -05002658 return;
Christoph Lameter6446faa2008-02-15 23:45:26 -08002659
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002660 do {
Joonsoo Kim837d6782012-08-16 00:02:40 +09002661 if (unlikely(n)) {
2662 spin_unlock_irqrestore(&n->list_lock, flags);
2663 n = NULL;
2664 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002665 prior = page->freelist;
2666 counters = page->counters;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002667 set_freepointer(s, tail, prior);
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002668 new.counters = counters;
2669 was_frozen = new.frozen;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002670 new.inuse -= cnt;
Joonsoo Kim837d6782012-08-16 00:02:40 +09002671 if ((!new.inuse || !prior) && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002672
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002673 if (kmem_cache_has_cpu_partial(s) && !prior) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002674
2675 /*
Chen Gangd0e0ac92013-07-15 09:05:29 +08002676 * Slab was on no list before and will be
2677 * partially empty
2678 * We can defer the list move and instead
2679 * freeze it.
Christoph Lameter49e22582011-08-09 16:12:27 -05002680 */
2681 new.frozen = 1;
2682
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002683 } else { /* Needs to be taken off a list */
Christoph Lameter49e22582011-08-09 16:12:27 -05002684
LQYMGTb455def2014-12-10 15:42:13 -08002685 n = get_node(s, page_to_nid(page));
Christoph Lameter49e22582011-08-09 16:12:27 -05002686 /*
2687 * Speculatively acquire the list_lock.
2688 * If the cmpxchg does not succeed then we may
2689 * drop the list_lock without any processing.
2690 *
2691 * Otherwise the list_lock will synchronize with
2692 * other processors updating the list of slabs.
2693 */
2694 spin_lock_irqsave(&n->list_lock, flags);
2695
2696 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002697 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002698
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002699 } while (!cmpxchg_double_slab(s, page,
2700 prior, counters,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002701 head, new.counters,
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002702 "__slab_free"));
Christoph Lameter81819f02007-05-06 14:49:36 -07002703
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002704 if (likely(!n)) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002705
2706 /*
2707 * If we just froze the page then put it onto the
2708 * per cpu partial list.
2709 */
Alex Shi8028dce2012-02-03 23:34:56 +08002710 if (new.frozen && !was_frozen) {
Christoph Lameter49e22582011-08-09 16:12:27 -05002711 put_cpu_partial(s, page, 1);
Alex Shi8028dce2012-02-03 23:34:56 +08002712 stat(s, CPU_PARTIAL_FREE);
2713 }
Christoph Lameter49e22582011-08-09 16:12:27 -05002714 /*
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002715 * The list lock was not taken therefore no list
2716 * activity can be necessary.
2717 */
LQYMGTb455def2014-12-10 15:42:13 -08002718 if (was_frozen)
2719 stat(s, FREE_FROZEN);
2720 return;
2721 }
Christoph Lameter81819f02007-05-06 14:49:36 -07002722
Joonsoo Kim8a5b20a2014-07-02 15:22:35 -07002723 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
Joonsoo Kim837d6782012-08-16 00:02:40 +09002724 goto slab_empty;
Christoph Lameter81819f02007-05-06 14:49:36 -07002725
Joonsoo Kim837d6782012-08-16 00:02:40 +09002726 /*
2727 * Objects left in the slab. If it was not on the partial list before
2728 * then add it.
2729 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09002730 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2731 if (kmem_cache_debug(s))
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002732 remove_full(s, n, page);
Joonsoo Kim837d6782012-08-16 00:02:40 +09002733 add_partial(n, page, DEACTIVATE_TO_TAIL);
2734 stat(s, FREE_ADD_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07002735 }
Christoph Lameter80f08c12011-06-01 12:25:55 -05002736 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002737 return;
2738
2739slab_empty:
Christoph Lametera973e9d2008-03-01 13:40:44 -08002740 if (prior) {
Christoph Lameter81819f02007-05-06 14:49:36 -07002741 /*
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002742 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07002743 */
Christoph Lameter5cc6eee2011-06-01 12:25:50 -05002744 remove_partial(n, page);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002745 stat(s, FREE_REMOVE_PARTIAL);
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002746 } else {
Christoph Lameter6fbabb22011-08-08 11:16:56 -05002747 /* Slab must be on the full list */
Peter Zijlstrac65c1872014-01-10 13:23:49 +01002748 remove_full(s, n, page);
2749 }
Christoph Lameter2cfb7452011-06-01 12:25:52 -05002750
Christoph Lameter80f08c12011-06-01 12:25:55 -05002751 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter84e554e62009-12-18 16:26:23 -06002752 stat(s, FREE_SLAB);
Christoph Lameter81819f02007-05-06 14:49:36 -07002753 discard_slab(s, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07002754}
2755
Christoph Lameter894b8782007-05-10 03:15:16 -07002756/*
2757 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2758 * can perform fastpath freeing without additional function calls.
2759 *
2760 * The fastpath is only possible if we are freeing to the current cpu slab
2761 * of this processor. This typically the case if we have just allocated
2762 * the item before.
2763 *
2764 * If fastpath is not possible then fall back to __slab_free where we deal
2765 * with all sorts of special processing.
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002766 *
2767 * Bulk free of a freelist with several objects (all pointing to the
2768 * same page) possible by specifying head and tail ptr, plus objects
2769 * count (cnt). Bulk free indicated by tail pointer being set.
Christoph Lameter894b8782007-05-10 03:15:16 -07002770 */
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002771static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2772 void *head, void *tail, int cnt,
2773 unsigned long addr)
Christoph Lameter894b8782007-05-10 03:15:16 -07002774{
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002775 void *tail_obj = tail ? : head;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07002776 struct kmem_cache_cpu *c;
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002777 unsigned long tid;
Christoph Lameter1f842602008-01-07 23:20:30 -08002778
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002779 slab_free_freelist_hook(s, head, tail);
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002780
Christoph Lametera24c5a02011-03-15 12:45:21 -05002781redo:
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002782 /*
2783 * Determine the currently cpus per cpu slab.
2784 * The cpu may change afterward. However that does not matter since
2785 * data is retrieved via this pointer. If we are on the same cpu
Jesper Dangaard Brouer2ae44002015-09-04 15:45:31 -07002786 * during the cmpxchg then the free will succeed.
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002787 */
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002788 do {
2789 tid = this_cpu_read(s->cpu_slab->tid);
2790 c = raw_cpu_ptr(s->cpu_slab);
Mark Rutland859b7a02015-03-25 15:55:23 -07002791 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2792 unlikely(tid != READ_ONCE(c->tid)));
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002793
Joonsoo Kim9aabf812015-02-10 14:09:32 -08002794 /* Same with comment on barrier() in slab_alloc_node() */
2795 barrier();
Christoph Lameterc016b0b2010-08-20 12:37:16 -05002796
Christoph Lameter442b06b2011-05-17 16:29:31 -05002797 if (likely(page == c->page)) {
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002798 set_freepointer(s, tail_obj, c->freelist);
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002799
Christoph Lameter933393f2011-12-22 11:58:51 -06002800 if (unlikely(!this_cpu_cmpxchg_double(
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002801 s->cpu_slab->freelist, s->cpu_slab->tid,
2802 c->freelist, tid,
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002803 head, next_tid(tid)))) {
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06002804
2805 note_cmpxchg_failure("slab_free", s, tid);
2806 goto redo;
2807 }
Christoph Lameter84e554e62009-12-18 16:26:23 -06002808 stat(s, FREE_FASTPATH);
Christoph Lameter894b8782007-05-10 03:15:16 -07002809 } else
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002810 __slab_free(s, page, head, tail_obj, cnt, addr);
Christoph Lameter894b8782007-05-10 03:15:16 -07002811
Christoph Lameter894b8782007-05-10 03:15:16 -07002812}
2813
Christoph Lameter81819f02007-05-06 14:49:36 -07002814void kmem_cache_free(struct kmem_cache *s, void *x)
2815{
Glauber Costab9ce5ef2012-12-18 14:22:46 -08002816 s = cache_from_obj(s, x);
2817 if (!s)
Christoph Lameter79576102012-09-04 23:06:14 +00002818 return;
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08002819 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02002820 trace_kmem_cache_free(_RET_IP_, x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002821}
2822EXPORT_SYMBOL(kmem_cache_free);
2823
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002824struct detached_freelist {
2825 struct page *page;
2826 void *tail;
2827 void *freelist;
2828 int cnt;
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002829 struct kmem_cache *s;
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002830};
2831
2832/*
2833 * This function progressively scans the array with free objects (with
2834 * a limited look ahead) and extract objects belonging to the same
2835 * page. It builds a detached freelist directly within the given
2836 * page/objects. This can happen without any need for
2837 * synchronization, because the objects are owned by running process.
2838 * The freelist is build up as a single linked list in the objects.
2839 * The idea is, that this detached freelist can then be bulk
2840 * transferred to the real freelist(s), but only requiring a single
2841 * synchronization primitive. Look ahead in the array is limited due
2842 * to performance reasons.
2843 */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002844static inline
2845int build_detached_freelist(struct kmem_cache *s, size_t size,
2846 void **p, struct detached_freelist *df)
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002847{
2848 size_t first_skipped_index = 0;
2849 int lookahead = 3;
2850 void *object;
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002851 struct page *page;
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002852
2853 /* Always re-init detached_freelist */
2854 df->page = NULL;
2855
2856 do {
2857 object = p[--size];
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002858 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002859 } while (!object && size);
2860
2861 if (!object)
2862 return 0;
2863
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002864 page = virt_to_head_page(object);
2865 if (!s) {
2866 /* Handle kalloc'ed objects */
2867 if (unlikely(!PageSlab(page))) {
2868 BUG_ON(!PageCompound(page));
2869 kfree_hook(object);
2870 __free_kmem_pages(page, compound_order(page));
2871 p[size] = NULL; /* mark object processed */
2872 return size;
2873 }
2874 /* Derive kmem_cache from object */
2875 df->s = page->slab_cache;
2876 } else {
2877 df->s = cache_from_obj(s, object); /* Support for memcg */
2878 }
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002879
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002880 /* Start new detached freelist */
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07002881 df->page = page;
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002882 set_freepointer(df->s, object, NULL);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002883 df->tail = object;
2884 df->freelist = object;
2885 p[size] = NULL; /* mark object processed */
2886 df->cnt = 1;
2887
2888 while (size) {
2889 object = p[--size];
2890 if (!object)
2891 continue; /* Skip processed objects */
2892
2893 /* df->page is always set at this point */
2894 if (df->page == virt_to_head_page(object)) {
2895 /* Opportunity build freelist */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002896 set_freepointer(df->s, object, df->freelist);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002897 df->freelist = object;
2898 df->cnt++;
2899 p[size] = NULL; /* mark object processed */
2900
2901 continue;
2902 }
2903
2904 /* Limit look ahead search */
2905 if (!--lookahead)
2906 break;
2907
2908 if (!first_skipped_index)
2909 first_skipped_index = size + 1;
2910 }
2911
2912 return first_skipped_index;
2913}
2914
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002915/* Note that interrupts must be enabled when calling this function. */
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002916void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07002917{
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002918 if (WARN_ON(!size))
2919 return;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002920
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002921 do {
2922 struct detached_freelist df;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002923
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002924 size = build_detached_freelist(s, size, p, &df);
2925 if (unlikely(!df.page))
2926 continue;
Jesper Dangaard Brouerfbd02632015-09-04 15:45:43 -07002927
Jesper Dangaard Brouer376bf122016-03-15 14:53:32 -07002928 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
Jesper Dangaard Brouerd0ecd892015-11-20 15:57:49 -08002929 } while (likely(size));
Christoph Lameter484748f2015-09-04 15:45:34 -07002930}
2931EXPORT_SYMBOL(kmem_cache_free_bulk);
2932
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002933/* Note that interrupts must be enabled when calling this function. */
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002934int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2935 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07002936{
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002937 struct kmem_cache_cpu *c;
2938 int i;
2939
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002940 /* memcg and kmem_cache debug support */
2941 s = slab_pre_alloc_hook(s, flags);
2942 if (unlikely(!s))
2943 return false;
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002944 /*
2945 * Drain objects in the per cpu slab, while disabling local
2946 * IRQs, which protects against PREEMPT and interrupts
2947 * handlers invoking normal fastpath.
2948 */
2949 local_irq_disable();
2950 c = this_cpu_ptr(s->cpu_slab);
2951
2952 for (i = 0; i < size; i++) {
2953 void *object = c->freelist;
2954
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002955 if (unlikely(!object)) {
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002956 /*
2957 * Invoking slow path likely have side-effect
2958 * of re-populating per CPU c->freelist
2959 */
Christoph Lameter87098372015-11-20 15:57:38 -08002960 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002961 _RET_IP_, c);
Christoph Lameter87098372015-11-20 15:57:38 -08002962 if (unlikely(!p[i]))
2963 goto error;
2964
Jesper Dangaard Brouerebe909e2015-09-04 15:45:40 -07002965 c = this_cpu_ptr(s->cpu_slab);
2966 continue; /* goto for-loop */
2967 }
Jesper Dangaard Brouer994eb762015-09-04 15:45:37 -07002968 c->freelist = get_freepointer(s, object);
2969 p[i] = object;
2970 }
2971 c->tid = next_tid(c->tid);
2972 local_irq_enable();
2973
2974 /* Clear memory outside IRQ disabled fastpath loop */
2975 if (unlikely(flags & __GFP_ZERO)) {
2976 int j;
2977
2978 for (j = 0; j < i; j++)
2979 memset(p[j], 0, s->object_size);
2980 }
2981
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002982 /* memcg and kmem_cache debug support */
2983 slab_post_alloc_hook(s, flags, size, p);
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002984 return i;
Christoph Lameter87098372015-11-20 15:57:38 -08002985error:
Christoph Lameter87098372015-11-20 15:57:38 -08002986 local_irq_enable();
Jesper Dangaard Brouer03ec0ed2015-11-20 15:57:52 -08002987 slab_post_alloc_hook(s, flags, i, p);
2988 __kmem_cache_free_bulk(s, i, p);
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08002989 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07002990}
2991EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2992
2993
Christoph Lameter81819f02007-05-06 14:49:36 -07002994/*
Christoph Lameter672bba32007-05-09 02:32:39 -07002995 * Object placement in a slab is made very easy because we always start at
2996 * offset 0. If we tune the size of the object to the alignment then we can
2997 * get the required alignment by putting one properly sized object after
2998 * another.
Christoph Lameter81819f02007-05-06 14:49:36 -07002999 *
3000 * Notice that the allocation order determines the sizes of the per cpu
3001 * caches. Each processor has always one slab available for allocations.
3002 * Increasing the allocation order reduces the number of times that slabs
Christoph Lameter672bba32007-05-09 02:32:39 -07003003 * must be moved on and off the partial lists and is therefore a factor in
Christoph Lameter81819f02007-05-06 14:49:36 -07003004 * locking overhead.
Christoph Lameter81819f02007-05-06 14:49:36 -07003005 */
3006
3007/*
3008 * Mininum / Maximum order of slab pages. This influences locking overhead
3009 * and slab fragmentation. A higher order reduces the number of partial slabs
3010 * and increases the number of allocations possible without having to
3011 * take the list_lock.
3012 */
3013static int slub_min_order;
Christoph Lameter114e9e82008-04-14 19:11:41 +03003014static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03003015static int slub_min_objects;
Christoph Lameter81819f02007-05-06 14:49:36 -07003016
3017/*
Christoph Lameter81819f02007-05-06 14:49:36 -07003018 * Calculate the order of allocation given an slab object size.
3019 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003020 * The order of allocation has significant impact on performance and other
3021 * system components. Generally order 0 allocations should be preferred since
3022 * order 0 does not cause fragmentation in the page allocator. Larger objects
3023 * be problematic to put into order 0 slabs because there may be too much
Christoph Lameterc124f5b2008-04-14 19:13:29 +03003024 * unused space left. We go to a higher order if more than 1/16th of the slab
Christoph Lameter672bba32007-05-09 02:32:39 -07003025 * would be wasted.
Christoph Lameter81819f02007-05-06 14:49:36 -07003026 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003027 * In order to reach satisfactory performance we must ensure that a minimum
3028 * number of objects is in one slab. Otherwise we may generate too much
3029 * activity on the partial lists which requires taking the list_lock. This is
3030 * less a concern for large slabs though which are rarely used.
Christoph Lameter81819f02007-05-06 14:49:36 -07003031 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003032 * slub_max_order specifies the order where we begin to stop considering the
3033 * number of objects in a slab as critical. If we reach slub_max_order then
3034 * we try to keep the page order as low as possible. So we accept more waste
3035 * of space in favor of a small page order.
3036 *
3037 * Higher order allocations also allow the placement of more objects in a
3038 * slab and thereby reduce object handling overhead. If the user has
3039 * requested a higher mininum order then we start with that one instead of
3040 * the smallest order which will fit the object.
Christoph Lameter81819f02007-05-06 14:49:36 -07003041 */
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003042static inline int slab_order(int size, int min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003043 int max_order, int fract_leftover, int reserved)
Christoph Lameter81819f02007-05-06 14:49:36 -07003044{
3045 int order;
3046 int rem;
Christoph Lameter6300ea72007-07-17 04:03:20 -07003047 int min_order = slub_min_order;
Christoph Lameter81819f02007-05-06 14:49:36 -07003048
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003049 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
Cyrill Gorcunov210b5c02008-10-22 23:00:38 +04003050 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
Christoph Lameter39b26462008-04-14 19:11:30 +03003051
Wei Yang9f835702015-11-05 18:45:51 -08003052 for (order = max(min_order, get_order(min_objects * size + reserved));
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003053 order <= max_order; order++) {
3054
Christoph Lameter81819f02007-05-06 14:49:36 -07003055 unsigned long slab_size = PAGE_SIZE << order;
3056
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003057 rem = (slab_size - reserved) % size;
Christoph Lameter81819f02007-05-06 14:49:36 -07003058
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003059 if (rem <= slab_size / fract_leftover)
Christoph Lameter81819f02007-05-06 14:49:36 -07003060 break;
Christoph Lameter81819f02007-05-06 14:49:36 -07003061 }
Christoph Lameter672bba32007-05-09 02:32:39 -07003062
Christoph Lameter81819f02007-05-06 14:49:36 -07003063 return order;
3064}
3065
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003066static inline int calculate_order(int size, int reserved)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003067{
3068 int order;
3069 int min_objects;
3070 int fraction;
Zhang Yanmine8120ff2009-02-12 18:00:17 +02003071 int max_objects;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003072
3073 /*
3074 * Attempt to find best configuration for a slab. This
3075 * works by first attempting to generate a layout with
3076 * the best configuration and backing off gradually.
3077 *
Wei Yang422ff4d2015-11-05 18:45:46 -08003078 * First we increase the acceptable waste in a slab. Then
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003079 * we reduce the minimum objects required in a slab.
3080 */
3081 min_objects = slub_min_objects;
Christoph Lameter9b2cd502008-04-14 19:11:41 +03003082 if (!min_objects)
3083 min_objects = 4 * (fls(nr_cpu_ids) + 1);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003084 max_objects = order_objects(slub_max_order, size, reserved);
Zhang Yanmine8120ff2009-02-12 18:00:17 +02003085 min_objects = min(min_objects, max_objects);
3086
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003087 while (min_objects > 1) {
Christoph Lameterc124f5b2008-04-14 19:13:29 +03003088 fraction = 16;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003089 while (fraction >= 4) {
3090 order = slab_order(size, min_objects,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003091 slub_max_order, fraction, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003092 if (order <= slub_max_order)
3093 return order;
3094 fraction /= 2;
3095 }
Amerigo Wang5086c389c2009-08-19 21:44:13 +03003096 min_objects--;
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003097 }
3098
3099 /*
3100 * We were unable to place multiple objects in a slab. Now
3101 * lets see if we can place a single object there.
3102 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003103 order = slab_order(size, 1, slub_max_order, 1, reserved);
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003104 if (order <= slub_max_order)
3105 return order;
3106
3107 /*
3108 * Doh this slab cannot be placed using slub_max_order.
3109 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003110 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
David Rientjes818cf592009-04-23 09:58:22 +03003111 if (order < MAX_ORDER)
Christoph Lameter5e6d4442007-05-09 02:32:46 -07003112 return order;
3113 return -ENOSYS;
3114}
3115
Pekka Enberg5595cff2008-08-05 09:28:47 +03003116static void
Joonsoo Kim40534972012-05-11 00:50:47 +09003117init_kmem_cache_node(struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003118{
3119 n->nr_partial = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003120 spin_lock_init(&n->list_lock);
3121 INIT_LIST_HEAD(&n->partial);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003122#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003123 atomic_long_set(&n->nr_slabs, 0);
Salman Qazi02b71b72008-09-11 12:25:41 -07003124 atomic_long_set(&n->total_objects, 0);
Christoph Lameter643b1132007-05-06 14:49:42 -07003125 INIT_LIST_HEAD(&n->full);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003126#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003127}
3128
Christoph Lameter55136592010-08-20 12:37:13 -05003129static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003130{
Christoph Lameter6c182dc2010-08-20 12:37:14 -05003131 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
Christoph Lameter95a05b42013-01-10 19:14:19 +00003132 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003133
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003134 /*
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04003135 * Must align to double word boundary for the double cmpxchg
3136 * instructions to work; see __pcpu_double_call_return_bool().
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003137 */
Chris Metcalfd4d84fe2011-06-02 10:19:41 -04003138 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3139 2 * sizeof(void *));
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003140
Christoph Lameter8a5ec0b2011-02-25 11:38:54 -06003141 if (!s->cpu_slab)
3142 return 0;
3143
3144 init_kmem_cache_cpus(s);
3145
3146 return 1;
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003147}
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003148
Christoph Lameter51df1142010-08-20 12:37:15 -05003149static struct kmem_cache *kmem_cache_node;
3150
Christoph Lameter81819f02007-05-06 14:49:36 -07003151/*
3152 * No kmalloc_node yet so do it by hand. We know that this is the first
3153 * slab on the node for this slabcache. There are no concurrent accesses
3154 * possible.
3155 *
Zhi Yong Wu721ae222013-11-08 20:47:37 +08003156 * Note that this function only works on the kmem_cache_node
3157 * when allocating for the kmem_cache_node. This is used for bootstrapping
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003158 * memory on a fresh node that has no slab structures yet.
Christoph Lameter81819f02007-05-06 14:49:36 -07003159 */
Christoph Lameter55136592010-08-20 12:37:13 -05003160static void early_kmem_cache_node_alloc(int node)
Christoph Lameter81819f02007-05-06 14:49:36 -07003161{
3162 struct page *page;
3163 struct kmem_cache_node *n;
3164
Christoph Lameter51df1142010-08-20 12:37:15 -05003165 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
Christoph Lameter81819f02007-05-06 14:49:36 -07003166
Christoph Lameter51df1142010-08-20 12:37:15 -05003167 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
Christoph Lameter81819f02007-05-06 14:49:36 -07003168
3169 BUG_ON(!page);
Christoph Lametera2f92ee2007-08-22 14:01:57 -07003170 if (page_to_nid(page) != node) {
Fabian Frederickf9f58282014-06-04 16:06:34 -07003171 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3172 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
Christoph Lametera2f92ee2007-08-22 14:01:57 -07003173 }
3174
Christoph Lameter81819f02007-05-06 14:49:36 -07003175 n = page->freelist;
3176 BUG_ON(!n);
Christoph Lameter51df1142010-08-20 12:37:15 -05003177 page->freelist = get_freepointer(kmem_cache_node, n);
Christoph Lametere6e82ea2011-08-09 16:12:24 -05003178 page->inuse = 1;
Christoph Lameter8cb0a502011-06-01 12:25:46 -05003179 page->frozen = 0;
Christoph Lameter51df1142010-08-20 12:37:15 -05003180 kmem_cache_node->node[node] = n;
Christoph Lameter8ab13722007-07-17 04:03:32 -07003181#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterf7cb1932010-09-29 07:15:01 -05003182 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
Christoph Lameter51df1142010-08-20 12:37:15 -05003183 init_tracking(kmem_cache_node, n);
Christoph Lameter8ab13722007-07-17 04:03:32 -07003184#endif
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003185 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3186 GFP_KERNEL);
Joonsoo Kim40534972012-05-11 00:50:47 +09003187 init_kmem_cache_node(n);
Christoph Lameter51df1142010-08-20 12:37:15 -05003188 inc_slabs_node(kmem_cache_node, node, page->objects);
Christoph Lameter6446faa2008-02-15 23:45:26 -08003189
Dave Hansen67b6c902014-01-24 07:20:23 -08003190 /*
Steven Rostedt1e4dd942014-02-10 14:25:46 -08003191 * No locks need to be taken here as it has just been
3192 * initialized and there is no concurrent access.
Dave Hansen67b6c902014-01-24 07:20:23 -08003193 */
Steven Rostedt1e4dd942014-02-10 14:25:46 -08003194 __add_partial(n, page, DEACTIVATE_TO_HEAD);
Christoph Lameter81819f02007-05-06 14:49:36 -07003195}
3196
3197static void free_kmem_cache_nodes(struct kmem_cache *s)
3198{
3199 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003200 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07003201
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003202 for_each_kmem_cache_node(s, node, n) {
3203 kmem_cache_free(kmem_cache_node, n);
Christoph Lameter81819f02007-05-06 14:49:36 -07003204 s->node[node] = NULL;
3205 }
3206}
3207
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003208void __kmem_cache_release(struct kmem_cache *s)
3209{
3210 free_percpu(s->cpu_slab);
3211 free_kmem_cache_nodes(s);
3212}
3213
Christoph Lameter55136592010-08-20 12:37:13 -05003214static int init_kmem_cache_nodes(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003215{
3216 int node;
Christoph Lameter81819f02007-05-06 14:49:36 -07003217
Christoph Lameterf64dc582007-10-16 01:25:33 -07003218 for_each_node_state(node, N_NORMAL_MEMORY) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003219 struct kmem_cache_node *n;
3220
Alexander Duyck73367bd2010-05-21 14:41:35 -07003221 if (slab_state == DOWN) {
Christoph Lameter55136592010-08-20 12:37:13 -05003222 early_kmem_cache_node_alloc(node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07003223 continue;
Christoph Lameter81819f02007-05-06 14:49:36 -07003224 }
Christoph Lameter51df1142010-08-20 12:37:15 -05003225 n = kmem_cache_alloc_node(kmem_cache_node,
Christoph Lameter55136592010-08-20 12:37:13 -05003226 GFP_KERNEL, node);
Alexander Duyck73367bd2010-05-21 14:41:35 -07003227
3228 if (!n) {
3229 free_kmem_cache_nodes(s);
3230 return 0;
3231 }
3232
Christoph Lameter81819f02007-05-06 14:49:36 -07003233 s->node[node] = n;
Joonsoo Kim40534972012-05-11 00:50:47 +09003234 init_kmem_cache_node(n);
Christoph Lameter81819f02007-05-06 14:49:36 -07003235 }
3236 return 1;
3237}
Christoph Lameter81819f02007-05-06 14:49:36 -07003238
David Rientjesc0bdb232009-02-25 09:16:35 +02003239static void set_min_partial(struct kmem_cache *s, unsigned long min)
David Rientjes3b89d7d2009-02-22 17:40:07 -08003240{
3241 if (min < MIN_PARTIAL)
3242 min = MIN_PARTIAL;
3243 else if (min > MAX_PARTIAL)
3244 min = MAX_PARTIAL;
3245 s->min_partial = min;
3246}
3247
Christoph Lameter81819f02007-05-06 14:49:36 -07003248/*
3249 * calculate_sizes() determines the order and the distribution of data within
3250 * a slab object.
3251 */
Christoph Lameter06b285d2008-04-14 19:11:41 +03003252static int calculate_sizes(struct kmem_cache *s, int forced_order)
Christoph Lameter81819f02007-05-06 14:49:36 -07003253{
3254 unsigned long flags = s->flags;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003255 unsigned long size = s->object_size;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003256 int order;
Christoph Lameter81819f02007-05-06 14:49:36 -07003257
3258 /*
Christoph Lameterd8b42bf2008-02-15 23:45:25 -08003259 * Round up object size to the next word boundary. We can only
3260 * place the free pointer at word boundaries and this determines
3261 * the possible location of the free pointer.
3262 */
3263 size = ALIGN(size, sizeof(void *));
3264
3265#ifdef CONFIG_SLUB_DEBUG
3266 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003267 * Determine if we can poison the object itself. If the user of
3268 * the slab may touch the object after free or before allocation
3269 * then we should never poison the object itself.
3270 */
3271 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
Christoph Lameterc59def92007-05-16 22:10:50 -07003272 !s->ctor)
Christoph Lameter81819f02007-05-06 14:49:36 -07003273 s->flags |= __OBJECT_POISON;
3274 else
3275 s->flags &= ~__OBJECT_POISON;
3276
Christoph Lameter81819f02007-05-06 14:49:36 -07003277
3278 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003279 * If we are Redzoning then check if there is some space between the
Christoph Lameter81819f02007-05-06 14:49:36 -07003280 * end of the object and the free pointer. If not then add an
Christoph Lameter672bba32007-05-09 02:32:39 -07003281 * additional word to have some bytes to store Redzone information.
Christoph Lameter81819f02007-05-06 14:49:36 -07003282 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003283 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07003284 size += sizeof(void *);
Christoph Lameter41ecc552007-05-09 02:32:44 -07003285#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003286
3287 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07003288 * With that we have determined the number of bytes in actual use
3289 * by the object. This is the potential offset to the free pointer.
Christoph Lameter81819f02007-05-06 14:49:36 -07003290 */
3291 s->inuse = size;
3292
3293 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
Christoph Lameterc59def92007-05-16 22:10:50 -07003294 s->ctor)) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003295 /*
3296 * Relocate free pointer after the object if it is not
3297 * permitted to overwrite the first word of the object on
3298 * kmem_cache_free.
3299 *
3300 * This is the case if we do RCU, have a constructor or
3301 * destructor or are poisoning the objects.
3302 */
3303 s->offset = size;
3304 size += sizeof(void *);
3305 }
3306
Christoph Lameterc12b3c62007-05-23 13:57:31 -07003307#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07003308 if (flags & SLAB_STORE_USER)
3309 /*
3310 * Need to store information about allocs and frees after
3311 * the object.
3312 */
3313 size += 2 * sizeof(struct track);
3314
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -07003315 if (flags & SLAB_RED_ZONE) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003316 /*
3317 * Add some empty padding so that we can catch
3318 * overwrites from earlier objects rather than let
3319 * tracking information or the free pointer be
Frederik Schwarzer0211a9c2008-12-29 22:14:56 +01003320 * corrupted if a user writes before the start
Christoph Lameter81819f02007-05-06 14:49:36 -07003321 * of the object.
3322 */
3323 size += sizeof(void *);
Joonsoo Kimd86bd1b2016-03-15 14:55:12 -07003324
3325 s->red_left_pad = sizeof(void *);
3326 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3327 size += s->red_left_pad;
3328 }
Christoph Lameter41ecc552007-05-09 02:32:44 -07003329#endif
Christoph Lameter672bba32007-05-09 02:32:39 -07003330
Christoph Lameter81819f02007-05-06 14:49:36 -07003331 /*
Christoph Lameter81819f02007-05-06 14:49:36 -07003332 * SLUB stores one object immediately after another beginning from
3333 * offset 0. In order to align the objects we have to simply size
3334 * each object to conform to the alignment.
3335 */
Christoph Lameter45906852012-11-28 16:23:16 +00003336 size = ALIGN(size, s->align);
Christoph Lameter81819f02007-05-06 14:49:36 -07003337 s->size = size;
Christoph Lameter06b285d2008-04-14 19:11:41 +03003338 if (forced_order >= 0)
3339 order = forced_order;
3340 else
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003341 order = calculate_order(size, s->reserved);
Christoph Lameter81819f02007-05-06 14:49:36 -07003342
Christoph Lameter834f3d12008-04-14 19:11:31 +03003343 if (order < 0)
Christoph Lameter81819f02007-05-06 14:49:36 -07003344 return 0;
3345
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003346 s->allocflags = 0;
Christoph Lameter834f3d12008-04-14 19:11:31 +03003347 if (order)
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003348 s->allocflags |= __GFP_COMP;
3349
3350 if (s->flags & SLAB_CACHE_DMA)
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003351 s->allocflags |= GFP_DMA;
Christoph Lameterb7a49f02008-02-14 14:21:32 -08003352
3353 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3354 s->allocflags |= __GFP_RECLAIMABLE;
3355
Christoph Lameter81819f02007-05-06 14:49:36 -07003356 /*
3357 * Determine the number of objects per slab
3358 */
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003359 s->oo = oo_make(order, size, s->reserved);
3360 s->min = oo_make(get_order(size), size, s->reserved);
Christoph Lameter205ab992008-04-14 19:11:40 +03003361 if (oo_objects(s->oo) > oo_objects(s->max))
3362 s->max = s->oo;
Christoph Lameter81819f02007-05-06 14:49:36 -07003363
Christoph Lameter834f3d12008-04-14 19:11:31 +03003364 return !!oo_objects(s->oo);
Christoph Lameter81819f02007-05-06 14:49:36 -07003365}
3366
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003367static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07003368{
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003369 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08003370 s->reserved = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07003371
Lai Jiangshanda9a6382011-03-10 15:22:00 +08003372 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3373 s->reserved = sizeof(struct rcu_head);
Christoph Lameter81819f02007-05-06 14:49:36 -07003374
Christoph Lameter06b285d2008-04-14 19:11:41 +03003375 if (!calculate_sizes(s, -1))
Christoph Lameter81819f02007-05-06 14:49:36 -07003376 goto error;
David Rientjes3de47212009-07-27 18:30:35 -07003377 if (disable_higher_order_debug) {
3378 /*
3379 * Disable debugging flags that store metadata if the min slab
3380 * order increased.
3381 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003382 if (get_order(s->size) > get_order(s->object_size)) {
David Rientjes3de47212009-07-27 18:30:35 -07003383 s->flags &= ~DEBUG_METADATA_FLAGS;
3384 s->offset = 0;
3385 if (!calculate_sizes(s, -1))
3386 goto error;
3387 }
3388 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003389
Heiko Carstens25654092012-01-12 17:17:33 -08003390#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
Laura Abbott149daaf2016-03-15 14:55:09 -07003392 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
Christoph Lameterb789ef52011-06-01 12:25:49 -05003393 /* Enable fast mode */
3394 s->flags |= __CMPXCHG_DOUBLE;
3395#endif
3396
David Rientjes3b89d7d2009-02-22 17:40:07 -08003397 /*
3398 * The larger the object size is, the more pages we want on the partial
3399 * list to avoid pounding the page allocator excessively.
3400 */
Christoph Lameter49e22582011-08-09 16:12:27 -05003401 set_min_partial(s, ilog2(s->size) / 2);
3402
3403 /*
3404 * cpu_partial determined the maximum number of objects kept in the
3405 * per cpu partial lists of a processor.
3406 *
3407 * Per cpu partial lists mainly contain slabs that just have one
3408 * object freed. If they are used for allocation then they can be
3409 * filled up again with minimal effort. The slab will never hit the
3410 * per node partial lists and therefore no locking will be required.
3411 *
3412 * This setting also determines
3413 *
3414 * A) The number of objects from per cpu partial slabs dumped to the
3415 * per node list when we reach the limit.
Alex Shi9f264902011-09-01 11:32:18 +08003416 * B) The number of objects in cpu partial slabs to extract from the
Chen Gangd0e0ac92013-07-15 09:05:29 +08003417 * per node list when we run out of per cpu objects. We only fetch
3418 * 50% to keep some capacity around for frees.
Christoph Lameter49e22582011-08-09 16:12:27 -05003419 */
Joonsoo Kim345c9052013-06-19 14:05:52 +09003420 if (!kmem_cache_has_cpu_partial(s))
Christoph Lameter8f1e33d2011-11-23 09:24:27 -06003421 s->cpu_partial = 0;
3422 else if (s->size >= PAGE_SIZE)
Christoph Lameter49e22582011-08-09 16:12:27 -05003423 s->cpu_partial = 2;
3424 else if (s->size >= 1024)
3425 s->cpu_partial = 6;
3426 else if (s->size >= 256)
3427 s->cpu_partial = 13;
3428 else
3429 s->cpu_partial = 30;
3430
Christoph Lameter81819f02007-05-06 14:49:36 -07003431#ifdef CONFIG_NUMA
Christoph Lametere2cb96b2008-08-19 08:51:22 -05003432 s->remote_node_defrag_ratio = 1000;
Christoph Lameter81819f02007-05-06 14:49:36 -07003433#endif
Christoph Lameter55136592010-08-20 12:37:13 -05003434 if (!init_kmem_cache_nodes(s))
Christoph Lameterdfb4f092007-10-16 01:26:05 -07003435 goto error;
Christoph Lameter81819f02007-05-06 14:49:36 -07003436
Christoph Lameter55136592010-08-20 12:37:13 -05003437 if (alloc_kmem_cache_cpus(s))
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003438 return 0;
Christoph Lameterff120592009-12-18 16:26:22 -06003439
Christoph Lameter4c93c3552007-10-16 01:26:08 -07003440 free_kmem_cache_nodes(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003441error:
3442 if (flags & SLAB_PANIC)
Joe Perches756a0252016-03-17 14:19:47 -07003443 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3444 s->name, (unsigned long)s->size, s->size,
3445 oo_order(s->oo), s->offset, flags);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00003446 return -EINVAL;
Christoph Lameter81819f02007-05-06 14:49:36 -07003447}
Christoph Lameter81819f02007-05-06 14:49:36 -07003448
Christoph Lameter33b12c32008-04-25 12:22:43 -07003449static void list_slab_objects(struct kmem_cache *s, struct page *page,
3450 const char *text)
Christoph Lameter81819f02007-05-06 14:49:36 -07003451{
Christoph Lameter33b12c32008-04-25 12:22:43 -07003452#ifdef CONFIG_SLUB_DEBUG
3453 void *addr = page_address(page);
3454 void *p;
Namhyung Kima5dd5c12010-09-29 21:02:13 +09003455 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3456 sizeof(long), GFP_ATOMIC);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003457 if (!map)
3458 return;
Christoph Lameter945cf2b2012-09-04 23:18:33 +00003459 slab_err(s, page, text, s->name);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003460 slab_lock(page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003461
Christoph Lameter5f80b132011-04-15 14:48:13 -05003462 get_map(s, page, map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003463 for_each_object(p, s, addr, page->objects) {
3464
3465 if (!test_bit(slab_index(p, s, addr), map)) {
Fabian Frederickf9f58282014-06-04 16:06:34 -07003466 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003467 print_tracking(s, p);
3468 }
3469 }
3470 slab_unlock(page);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01003471 kfree(map);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003472#endif
3473}
3474
Christoph Lameter81819f02007-05-06 14:49:36 -07003475/*
Christoph Lameter599870b2008-04-23 12:36:52 -07003476 * Attempt to free all partial slabs on a node.
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003477 * This is called from __kmem_cache_shutdown(). We must take list_lock
3478 * because sysfs file might still access partial list after the shutdowning.
Christoph Lameter81819f02007-05-06 14:49:36 -07003479 */
Christoph Lameter599870b2008-04-23 12:36:52 -07003480static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
Christoph Lameter81819f02007-05-06 14:49:36 -07003481{
Christoph Lameter81819f02007-05-06 14:49:36 -07003482 struct page *page, *h;
3483
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003484 BUG_ON(irqs_disabled());
3485 spin_lock_irq(&n->list_lock);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003486 list_for_each_entry_safe(page, h, &n->partial, lru) {
Christoph Lameter81819f02007-05-06 14:49:36 -07003487 if (!page->inuse) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003488 remove_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07003489 discard_slab(s, page);
Christoph Lameter33b12c32008-04-25 12:22:43 -07003490 } else {
3491 list_slab_objects(s, page,
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003492 "Objects remaining in %s on __kmem_cache_shutdown()");
Christoph Lameter599870b2008-04-23 12:36:52 -07003493 }
Christoph Lameter33b12c32008-04-25 12:22:43 -07003494 }
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003495 spin_unlock_irq(&n->list_lock);
Christoph Lameter81819f02007-05-06 14:49:36 -07003496}
3497
3498/*
Christoph Lameter672bba32007-05-09 02:32:39 -07003499 * Release all resources used by a slab cache.
Christoph Lameter81819f02007-05-06 14:49:36 -07003500 */
Dmitry Safonov52b4b952016-02-17 13:11:37 -08003501int __kmem_cache_shutdown(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07003502{
3503 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003504 struct kmem_cache_node *n;
Christoph Lameter81819f02007-05-06 14:49:36 -07003505
3506 flush_all(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07003507 /* Attempt to free all objects */
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003508 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter599870b2008-04-23 12:36:52 -07003509 free_partial(s, n);
3510 if (n->nr_partial || slabs_node(s, node))
Christoph Lameter81819f02007-05-06 14:49:36 -07003511 return 1;
3512 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003513 return 0;
3514}
3515
Christoph Lameter81819f02007-05-06 14:49:36 -07003516/********************************************************************
3517 * Kmalloc subsystem
3518 *******************************************************************/
3519
Christoph Lameter81819f02007-05-06 14:49:36 -07003520static int __init setup_slub_min_order(char *str)
3521{
Pekka Enberg06428782008-01-07 23:20:27 -08003522 get_option(&str, &slub_min_order);
Christoph Lameter81819f02007-05-06 14:49:36 -07003523
3524 return 1;
3525}
3526
3527__setup("slub_min_order=", setup_slub_min_order);
3528
3529static int __init setup_slub_max_order(char *str)
3530{
Pekka Enberg06428782008-01-07 23:20:27 -08003531 get_option(&str, &slub_max_order);
David Rientjes818cf592009-04-23 09:58:22 +03003532 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07003533
3534 return 1;
3535}
3536
3537__setup("slub_max_order=", setup_slub_max_order);
3538
3539static int __init setup_slub_min_objects(char *str)
3540{
Pekka Enberg06428782008-01-07 23:20:27 -08003541 get_option(&str, &slub_min_objects);
Christoph Lameter81819f02007-05-06 14:49:36 -07003542
3543 return 1;
3544}
3545
3546__setup("slub_min_objects=", setup_slub_min_objects);
3547
Christoph Lameter81819f02007-05-06 14:49:36 -07003548void *__kmalloc(size_t size, gfp_t flags)
3549{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003550 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003551 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003552
Christoph Lameter95a05b42013-01-10 19:14:19 +00003553 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02003554 return kmalloc_large(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003555
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003556 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003557
3558 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003559 return s;
3560
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003561 ret = slab_alloc(s, flags, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003562
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003563 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003564
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003565 kasan_kmalloc(s, ret, size, flags);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003566
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003567 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003568}
3569EXPORT_SYMBOL(__kmalloc);
3570
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09003571#ifdef CONFIG_NUMA
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003572static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3573{
Vegard Nossumb1eeab62008-11-25 16:55:53 +01003574 struct page *page;
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003575 void *ptr = NULL;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003576
Vladimir Davydov52383432014-06-04 16:06:39 -07003577 flags |= __GFP_COMP | __GFP_NOTRACK;
3578 page = alloc_kmem_pages_node(node, flags, get_order(size));
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003579 if (page)
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003580 ptr = page_address(page);
3581
Roman Bobnievd56791b2013-10-08 15:58:57 -07003582 kmalloc_large_node_hook(ptr, size, flags);
Catalin Marinase4f7c0b42009-07-07 10:32:59 +01003583 return ptr;
Christoph Lameterf619cfe2008-03-01 13:56:40 -08003584}
3585
Christoph Lameter81819f02007-05-06 14:49:36 -07003586void *__kmalloc_node(size_t size, gfp_t flags, int node)
3587{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003588 struct kmem_cache *s;
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003589 void *ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003590
Christoph Lameter95a05b42013-01-10 19:14:19 +00003591 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003592 ret = kmalloc_large_node(size, flags, node);
3593
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003594 trace_kmalloc_node(_RET_IP_, ret,
3595 size, PAGE_SIZE << get_order(size),
3596 flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003597
3598 return ret;
3599 }
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003600
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003601 s = kmalloc_slab(size, flags);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003602
3603 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003604 return s;
3605
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03003606 ret = slab_alloc_node(s, flags, node, _RET_IP_);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003607
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003608 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003609
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003610 kasan_kmalloc(s, ret, size, flags);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003611
Eduard - Gabriel Munteanu5b882be2008-08-19 20:43:26 +03003612 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07003613}
3614EXPORT_SYMBOL(__kmalloc_node);
3615#endif
3616
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003617static size_t __ksize(const void *object)
Christoph Lameter81819f02007-05-06 14:49:36 -07003618{
Christoph Lameter272c1d22007-06-08 13:46:49 -07003619 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07003620
Christoph Lameteref8b4522007-10-16 01:24:46 -07003621 if (unlikely(object == ZERO_SIZE_PTR))
Christoph Lameter272c1d22007-06-08 13:46:49 -07003622 return 0;
3623
Vegard Nossum294a80a2007-12-04 23:45:30 -08003624 page = virt_to_head_page(object);
Vegard Nossum294a80a2007-12-04 23:45:30 -08003625
Pekka Enberg76994412008-05-22 19:22:25 +03003626 if (unlikely(!PageSlab(page))) {
3627 WARN_ON(!PageCompound(page));
Vegard Nossum294a80a2007-12-04 23:45:30 -08003628 return PAGE_SIZE << compound_order(page);
Pekka Enberg76994412008-05-22 19:22:25 +03003629 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003630
Glauber Costa1b4f59e32012-10-22 18:05:36 +04003631 return slab_ksize(page->slab_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003632}
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003633
3634size_t ksize(const void *object)
3635{
3636 size_t size = __ksize(object);
3637 /* We assume that ksize callers could use whole allocated area,
3638 so we need unpoison this area. */
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003639 kasan_krealloc(object, size, GFP_NOWAIT);
Andrey Ryabinin0316bec2015-02-13 14:39:42 -08003640 return size;
3641}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02003642EXPORT_SYMBOL(ksize);
Christoph Lameter81819f02007-05-06 14:49:36 -07003643
3644void kfree(const void *x)
3645{
Christoph Lameter81819f02007-05-06 14:49:36 -07003646 struct page *page;
Christoph Lameter5bb983b2008-02-07 17:47:41 -08003647 void *object = (void *)x;
Christoph Lameter81819f02007-05-06 14:49:36 -07003648
Pekka Enberg2121db72009-03-25 11:05:57 +02003649 trace_kfree(_RET_IP_, x);
3650
Satyam Sharma2408c552007-10-16 01:24:44 -07003651 if (unlikely(ZERO_OR_NULL_PTR(x)))
Christoph Lameter81819f02007-05-06 14:49:36 -07003652 return;
3653
Christoph Lameterb49af682007-05-06 14:49:41 -07003654 page = virt_to_head_page(x);
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003655 if (unlikely(!PageSlab(page))) {
Christoph Lameter09375022008-05-28 10:32:22 -07003656 BUG_ON(!PageCompound(page));
Roman Bobnievd56791b2013-10-08 15:58:57 -07003657 kfree_hook(x);
Vladimir Davydov52383432014-06-04 16:06:39 -07003658 __free_kmem_pages(page, compound_order(page));
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07003659 return;
3660 }
Jesper Dangaard Brouer81084652015-11-20 15:57:46 -08003661 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
Christoph Lameter81819f02007-05-06 14:49:36 -07003662}
3663EXPORT_SYMBOL(kfree);
3664
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003665#define SHRINK_PROMOTE_MAX 32
3666
Christoph Lameter2086d262007-05-06 14:49:46 -07003667/*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003668 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3669 * up most to the head of the partial lists. New allocations will then
3670 * fill those up and thus they can be removed from the partial lists.
Christoph Lameter672bba32007-05-09 02:32:39 -07003671 *
3672 * The slabs with the least items are placed last. This results in them
3673 * being allocated from last increasing the chance that the last objects
3674 * are freed in them.
Christoph Lameter2086d262007-05-06 14:49:46 -07003675 */
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003676int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
Christoph Lameter2086d262007-05-06 14:49:46 -07003677{
3678 int node;
3679 int i;
3680 struct kmem_cache_node *n;
3681 struct page *page;
3682 struct page *t;
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003683 struct list_head discard;
3684 struct list_head promote[SHRINK_PROMOTE_MAX];
Christoph Lameter2086d262007-05-06 14:49:46 -07003685 unsigned long flags;
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003686 int ret = 0;
Christoph Lameter2086d262007-05-06 14:49:46 -07003687
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003688 if (deactivate) {
3689 /*
3690 * Disable empty slabs caching. Used to avoid pinning offline
3691 * memory cgroups by kmem pages that can be freed.
3692 */
3693 s->cpu_partial = 0;
3694 s->min_partial = 0;
3695
3696 /*
3697 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3698 * so we have to make sure the change is visible.
3699 */
3700 kick_all_cpus_sync();
3701 }
3702
Christoph Lameter2086d262007-05-06 14:49:46 -07003703 flush_all(s);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003704 for_each_kmem_cache_node(s, node, n) {
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003705 INIT_LIST_HEAD(&discard);
3706 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3707 INIT_LIST_HEAD(promote + i);
Christoph Lameter2086d262007-05-06 14:49:46 -07003708
3709 spin_lock_irqsave(&n->list_lock, flags);
3710
3711 /*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003712 * Build lists of slabs to discard or promote.
Christoph Lameter2086d262007-05-06 14:49:46 -07003713 *
Christoph Lameter672bba32007-05-09 02:32:39 -07003714 * Note that concurrent frees may occur while we hold the
3715 * list_lock. page->inuse here is the upper limit.
Christoph Lameter2086d262007-05-06 14:49:46 -07003716 */
3717 list_for_each_entry_safe(page, t, &n->partial, lru) {
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003718 int free = page->objects - page->inuse;
3719
3720 /* Do not reread page->inuse */
3721 barrier();
3722
3723 /* We do not keep full slabs on the list */
3724 BUG_ON(free <= 0);
3725
3726 if (free == page->objects) {
3727 list_move(&page->lru, &discard);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003728 n->nr_partial--;
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003729 } else if (free <= SHRINK_PROMOTE_MAX)
3730 list_move(&page->lru, promote + free - 1);
Christoph Lameter2086d262007-05-06 14:49:46 -07003731 }
3732
Christoph Lameter2086d262007-05-06 14:49:46 -07003733 /*
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003734 * Promote the slabs filled up most to the head of the
3735 * partial list.
Christoph Lameter2086d262007-05-06 14:49:46 -07003736 */
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003737 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3738 list_splice(promote + i, &n->partial);
Christoph Lameter2086d262007-05-06 14:49:46 -07003739
Christoph Lameter2086d262007-05-06 14:49:46 -07003740 spin_unlock_irqrestore(&n->list_lock, flags);
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003741
3742 /* Release empty slabs */
Vladimir Davydov832f37f2015-02-12 14:59:41 -08003743 list_for_each_entry_safe(page, t, &discard, lru)
Christoph Lameter69cb8e62011-08-09 16:12:22 -05003744 discard_slab(s, page);
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003745
3746 if (slabs_node(s, node))
3747 ret = 1;
Christoph Lameter2086d262007-05-06 14:49:46 -07003748 }
3749
Vladimir Davydovce3712d2015-02-12 14:59:44 -08003750 return ret;
Christoph Lameter2086d262007-05-06 14:49:46 -07003751}
Christoph Lameter2086d262007-05-06 14:49:46 -07003752
Yasunori Gotob9049e22007-10-21 16:41:37 -07003753static int slab_mem_going_offline_callback(void *arg)
3754{
3755 struct kmem_cache *s;
3756
Christoph Lameter18004c52012-07-06 15:25:12 -05003757 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003758 list_for_each_entry(s, &slab_caches, list)
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08003759 __kmem_cache_shrink(s, false);
Christoph Lameter18004c52012-07-06 15:25:12 -05003760 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003761
3762 return 0;
3763}
3764
3765static void slab_mem_offline_callback(void *arg)
3766{
3767 struct kmem_cache_node *n;
3768 struct kmem_cache *s;
3769 struct memory_notify *marg = arg;
3770 int offline_node;
3771
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003772 offline_node = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003773
3774 /*
3775 * If the node still has available memory. we need kmem_cache_node
3776 * for it yet.
3777 */
3778 if (offline_node < 0)
3779 return;
3780
Christoph Lameter18004c52012-07-06 15:25:12 -05003781 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003782 list_for_each_entry(s, &slab_caches, list) {
3783 n = get_node(s, offline_node);
3784 if (n) {
3785 /*
3786 * if n->nr_slabs > 0, slabs still exist on the node
3787 * that is going down. We were unable to free them,
Adam Buchbinderc9404c92009-12-18 15:40:42 -05003788 * and offline_pages() function shouldn't call this
Yasunori Gotob9049e22007-10-21 16:41:37 -07003789 * callback. So, we must fail.
3790 */
Christoph Lameter0f389ec2008-04-14 18:53:02 +03003791 BUG_ON(slabs_node(s, offline_node));
Yasunori Gotob9049e22007-10-21 16:41:37 -07003792
3793 s->node[offline_node] = NULL;
Christoph Lameter8de66a02010-08-25 14:51:14 -05003794 kmem_cache_free(kmem_cache_node, n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003795 }
3796 }
Christoph Lameter18004c52012-07-06 15:25:12 -05003797 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003798}
3799
3800static int slab_mem_going_online_callback(void *arg)
3801{
3802 struct kmem_cache_node *n;
3803 struct kmem_cache *s;
3804 struct memory_notify *marg = arg;
Lai Jiangshanb9d5ab22012-12-11 16:01:05 -08003805 int nid = marg->status_change_nid_normal;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003806 int ret = 0;
3807
3808 /*
3809 * If the node's memory is already available, then kmem_cache_node is
3810 * already created. Nothing to do.
3811 */
3812 if (nid < 0)
3813 return 0;
3814
3815 /*
Christoph Lameter0121c6192008-04-29 16:11:12 -07003816 * We are bringing a node online. No memory is available yet. We must
Yasunori Gotob9049e22007-10-21 16:41:37 -07003817 * allocate a kmem_cache_node structure in order to bring the node
3818 * online.
3819 */
Christoph Lameter18004c52012-07-06 15:25:12 -05003820 mutex_lock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003821 list_for_each_entry(s, &slab_caches, list) {
3822 /*
3823 * XXX: kmem_cache_alloc_node will fallback to other nodes
3824 * since memory is not yet available from the node that
3825 * is brought up.
3826 */
Christoph Lameter8de66a02010-08-25 14:51:14 -05003827 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003828 if (!n) {
3829 ret = -ENOMEM;
3830 goto out;
3831 }
Joonsoo Kim40534972012-05-11 00:50:47 +09003832 init_kmem_cache_node(n);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003833 s->node[nid] = n;
3834 }
3835out:
Christoph Lameter18004c52012-07-06 15:25:12 -05003836 mutex_unlock(&slab_mutex);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003837 return ret;
3838}
3839
3840static int slab_memory_callback(struct notifier_block *self,
3841 unsigned long action, void *arg)
3842{
3843 int ret = 0;
3844
3845 switch (action) {
3846 case MEM_GOING_ONLINE:
3847 ret = slab_mem_going_online_callback(arg);
3848 break;
3849 case MEM_GOING_OFFLINE:
3850 ret = slab_mem_going_offline_callback(arg);
3851 break;
3852 case MEM_OFFLINE:
3853 case MEM_CANCEL_ONLINE:
3854 slab_mem_offline_callback(arg);
3855 break;
3856 case MEM_ONLINE:
3857 case MEM_CANCEL_OFFLINE:
3858 break;
3859 }
KAMEZAWA Hiroyukidc19f9d2008-12-01 13:13:48 -08003860 if (ret)
3861 ret = notifier_from_errno(ret);
3862 else
3863 ret = NOTIFY_OK;
Yasunori Gotob9049e22007-10-21 16:41:37 -07003864 return ret;
3865}
3866
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003867static struct notifier_block slab_memory_callback_nb = {
3868 .notifier_call = slab_memory_callback,
3869 .priority = SLAB_CALLBACK_PRI,
3870};
Yasunori Gotob9049e22007-10-21 16:41:37 -07003871
Christoph Lameter81819f02007-05-06 14:49:36 -07003872/********************************************************************
3873 * Basic setup of slabs
3874 *******************************************************************/
3875
Christoph Lameter51df1142010-08-20 12:37:15 -05003876/*
3877 * Used for early kmem_cache structures that were allocated using
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003878 * the page allocator. Allocate them properly then fix up the pointers
3879 * that may be pointing to the wrong kmem_cache structure.
Christoph Lameter51df1142010-08-20 12:37:15 -05003880 */
3881
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003882static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
Christoph Lameter51df1142010-08-20 12:37:15 -05003883{
3884 int node;
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003885 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003886 struct kmem_cache_node *n;
Christoph Lameter51df1142010-08-20 12:37:15 -05003887
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003888 memcpy(s, static_cache, kmem_cache->object_size);
Christoph Lameter51df1142010-08-20 12:37:15 -05003889
Glauber Costa7d557b32013-02-22 20:20:00 +04003890 /*
3891 * This runs very early, and only the boot processor is supposed to be
3892 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3893 * IPIs around.
3894 */
3895 __flush_cpu_slab(s, smp_processor_id());
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003896 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter51df1142010-08-20 12:37:15 -05003897 struct page *p;
3898
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003899 list_for_each_entry(p, &n->partial, lru)
3900 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003901
Li Zefan607bf322011-04-12 15:22:26 +08003902#ifdef CONFIG_SLUB_DEBUG
Christoph Lameterfa45dc22014-08-06 16:04:09 -07003903 list_for_each_entry(p, &n->full, lru)
3904 p->slab_cache = s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003905#endif
Christoph Lameter51df1142010-08-20 12:37:15 -05003906 }
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08003907 slab_init_memcg_params(s);
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003908 list_add(&s->list, &slab_caches);
3909 return s;
Christoph Lameter51df1142010-08-20 12:37:15 -05003910}
3911
Christoph Lameter81819f02007-05-06 14:49:36 -07003912void __init kmem_cache_init(void)
3913{
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003914 static __initdata struct kmem_cache boot_kmem_cache,
3915 boot_kmem_cache_node;
Christoph Lameter51df1142010-08-20 12:37:15 -05003916
Stanislaw Gruszkafc8d8622012-01-10 15:07:32 -08003917 if (debug_guardpage_minorder())
3918 slub_max_order = 0;
3919
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003920 kmem_cache_node = &boot_kmem_cache_node;
3921 kmem_cache = &boot_kmem_cache;
Christoph Lameter51df1142010-08-20 12:37:15 -05003922
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003923 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3924 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
Yasunori Gotob9049e22007-10-21 16:41:37 -07003925
Andrew Morton3ac38fa2013-04-29 15:08:06 -07003926 register_hotmemory_notifier(&slab_memory_callback_nb);
Christoph Lameter81819f02007-05-06 14:49:36 -07003927
3928 /* Able to allocate the per node structures */
3929 slab_state = PARTIAL;
3930
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003931 create_boot_cache(kmem_cache, "kmem_cache",
3932 offsetof(struct kmem_cache, node) +
3933 nr_node_ids * sizeof(struct kmem_cache_node *),
3934 SLAB_HWCACHE_ALIGN);
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003935
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003936 kmem_cache = bootstrap(&boot_kmem_cache);
Christoph Lameter81819f02007-05-06 14:49:36 -07003937
Christoph Lameter51df1142010-08-20 12:37:15 -05003938 /*
3939 * Allocate kmem_cache_node properly from the kmem_cache slab.
3940 * kmem_cache_node is separately allocated so no need to
3941 * update any list pointers.
3942 */
Christoph Lameterdffb4d62012-11-28 16:23:07 +00003943 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
Christoph Lameter51df1142010-08-20 12:37:15 -05003944
3945 /* Now we can use the kmem_cache to allocate kmalloc slabs */
Daniel Sanders34cc6992015-06-24 16:55:57 -07003946 setup_kmalloc_cache_index_table();
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003947 create_kmalloc_caches(0);
Christoph Lameter81819f02007-05-06 14:49:36 -07003948
3949#ifdef CONFIG_SMP
3950 register_cpu_notifier(&slab_notifier);
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06003951#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07003952
Fabian Frederickf9f58282014-06-04 16:06:34 -07003953 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
Christoph Lameterf97d5f62013-01-10 19:12:17 +00003954 cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07003955 slub_min_order, slub_max_order, slub_min_objects,
3956 nr_cpu_ids, nr_node_ids);
3957}
3958
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003959void __init kmem_cache_init_late(void)
3960{
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003961}
3962
Glauber Costa2633d7a2012-12-18 14:22:34 -08003963struct kmem_cache *
Vladimir Davydova44cb942014-04-07 15:39:23 -07003964__kmem_cache_alias(const char *name, size_t size, size_t align,
3965 unsigned long flags, void (*ctor)(void *))
Christoph Lameter81819f02007-05-06 14:49:36 -07003966{
Vladimir Davydov426589f2015-02-12 14:59:23 -08003967 struct kmem_cache *s, *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07003968
Vladimir Davydova44cb942014-04-07 15:39:23 -07003969 s = find_mergeable(size, align, flags, name, ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07003970 if (s) {
3971 s->refcount++;
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003972
Christoph Lameter81819f02007-05-06 14:49:36 -07003973 /*
3974 * Adjust the object sizes so that we clear
3975 * the complete object on kzalloc.
3976 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003977 s->object_size = max(s->object_size, (int)size);
Christoph Lameter81819f02007-05-06 14:49:36 -07003978 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
Christoph Lameter6446faa2008-02-15 23:45:26 -08003979
Vladimir Davydov426589f2015-02-12 14:59:23 -08003980 for_each_memcg_cache(c, s) {
Vladimir Davydov84d0ddd2014-04-07 15:39:29 -07003981 c->object_size = s->object_size;
3982 c->inuse = max_t(int, c->inuse,
3983 ALIGN(size, sizeof(void *)));
3984 }
3985
David Rientjes7b8f3b62008-12-17 22:09:46 -08003986 if (sysfs_slab_alias(s, name)) {
David Rientjes7b8f3b62008-12-17 22:09:46 -08003987 s->refcount--;
Christoph Lametercbb79692012-09-05 00:18:32 +00003988 s = NULL;
David Rientjes7b8f3b62008-12-17 22:09:46 -08003989 }
Christoph Lametera0e1d1b2007-07-17 04:03:31 -07003990 }
Christoph Lameter6446faa2008-02-15 23:45:26 -08003991
Christoph Lametercbb79692012-09-05 00:18:32 +00003992 return s;
3993}
Pekka Enberg84c1cf62010-09-14 23:21:12 +03003994
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00003995int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
Christoph Lametercbb79692012-09-05 00:18:32 +00003996{
Pekka Enbergaac3a162012-09-05 12:07:44 +03003997 int err;
Christoph Lameter20cea962012-07-06 15:25:13 -05003998
Pekka Enbergaac3a162012-09-05 12:07:44 +03003999 err = kmem_cache_open(s, flags);
4000 if (err)
4001 return err;
Christoph Lameter20cea962012-07-06 15:25:13 -05004002
Christoph Lameter45530c42012-11-28 16:23:07 +00004003 /* Mutex is not taken during early boot */
4004 if (slab_state <= UP)
4005 return 0;
4006
Glauber Costa107dab52012-12-18 14:23:05 -08004007 memcg_propagate_slab_attrs(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03004008 err = sysfs_slab_add(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03004009 if (err)
Dmitry Safonov52b4b952016-02-17 13:11:37 -08004010 __kmem_cache_release(s);
Pekka Enbergaac3a162012-09-05 12:07:44 +03004011
4012 return err;
Christoph Lameter81819f02007-05-06 14:49:36 -07004013}
Christoph Lameter81819f02007-05-06 14:49:36 -07004014
Christoph Lameter81819f02007-05-06 14:49:36 -07004015#ifdef CONFIG_SMP
Christoph Lameter27390bc2007-06-01 00:47:09 -07004016/*
Christoph Lameter672bba32007-05-09 02:32:39 -07004017 * Use the cpu notifier to insure that the cpu slabs are flushed when
4018 * necessary.
Christoph Lameter81819f02007-05-06 14:49:36 -07004019 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004020static int slab_cpuup_callback(struct notifier_block *nfb,
Christoph Lameter81819f02007-05-06 14:49:36 -07004021 unsigned long action, void *hcpu)
4022{
4023 long cpu = (long)hcpu;
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004024 struct kmem_cache *s;
4025 unsigned long flags;
Christoph Lameter81819f02007-05-06 14:49:36 -07004026
4027 switch (action) {
4028 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07004029 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter81819f02007-05-06 14:49:36 -07004030 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07004031 case CPU_DEAD_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05004032 mutex_lock(&slab_mutex);
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07004033 list_for_each_entry(s, &slab_caches, list) {
4034 local_irq_save(flags);
4035 __flush_cpu_slab(s, cpu);
4036 local_irq_restore(flags);
4037 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004038 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07004039 break;
4040 default:
4041 break;
4042 }
4043 return NOTIFY_OK;
4044}
4045
Paul Gortmaker0db06282013-06-19 14:53:51 -04004046static struct notifier_block slab_notifier = {
Ingo Molnar3adbefe2008-02-05 17:57:39 -08004047 .notifier_call = slab_cpuup_callback
Pekka Enberg06428782008-01-07 23:20:27 -08004048};
Christoph Lameter81819f02007-05-06 14:49:36 -07004049
4050#endif
4051
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004052void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07004053{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004054 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004055 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004056
Christoph Lameter95a05b42013-01-10 19:14:19 +00004057 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
Pekka Enbergeada35e2008-02-11 22:47:46 +02004058 return kmalloc_large(size, gfpflags);
4059
Christoph Lameter2c59dd62013-01-10 19:14:19 +00004060 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07004061
Satyam Sharma2408c552007-10-16 01:24:44 -07004062 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004063 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004064
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03004065 ret = slab_alloc(s, gfpflags, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004066
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004067 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02004068 trace_kmalloc(caller, ret, size, s->size, gfpflags);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004069
4070 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07004071}
4072
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09004073#ifdef CONFIG_NUMA
Christoph Lameter81819f02007-05-06 14:49:36 -07004074void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004075 int node, unsigned long caller)
Christoph Lameter81819f02007-05-06 14:49:36 -07004076{
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004077 struct kmem_cache *s;
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004078 void *ret;
Christoph Lameteraadb4bc2007-10-16 01:24:38 -07004079
Christoph Lameter95a05b42013-01-10 19:14:19 +00004080 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
Xiaotian Fengd3e14aa2010-04-08 17:26:44 +08004081 ret = kmalloc_large_node(size, gfpflags, node);
4082
4083 trace_kmalloc_node(caller, ret,
4084 size, PAGE_SIZE << get_order(size),
4085 gfpflags, node);
4086
4087 return ret;
4088 }
Pekka Enbergeada35e2008-02-11 22:47:46 +02004089
Christoph Lameter2c59dd62013-01-10 19:14:19 +00004090 s = kmalloc_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07004091
Satyam Sharma2408c552007-10-16 01:24:44 -07004092 if (unlikely(ZERO_OR_NULL_PTR(s)))
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004093 return s;
Christoph Lameter81819f02007-05-06 14:49:36 -07004094
Ezequiel Garcia2b847c32012-09-08 17:47:58 -03004095 ret = slab_alloc_node(s, gfpflags, node, caller);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004096
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004097 /* Honor the call site pointer we received. */
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02004098 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
Eduard - Gabriel Munteanu94b528d2008-08-24 20:49:35 +03004099
4100 return ret;
Christoph Lameter81819f02007-05-06 14:49:36 -07004101}
Namhyung Kim5d1f57e2010-09-29 21:02:15 +09004102#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004103
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004104#ifdef CONFIG_SYSFS
Christoph Lameter205ab992008-04-14 19:11:40 +03004105static int count_inuse(struct page *page)
4106{
4107 return page->inuse;
4108}
4109
4110static int count_total(struct page *page)
4111{
4112 return page->objects;
4113}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004114#endif
Christoph Lameter205ab992008-04-14 19:11:40 +03004115
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004116#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter434e2452007-07-17 04:03:30 -07004117static int validate_slab(struct kmem_cache *s, struct page *page,
4118 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004119{
4120 void *p;
Christoph Lametera973e9d2008-03-01 13:40:44 -08004121 void *addr = page_address(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004122
4123 if (!check_slab(s, page) ||
4124 !on_freelist(s, page, NULL))
4125 return 0;
4126
4127 /* Now we know that a valid freelist exists */
Christoph Lameter39b26462008-04-14 19:11:30 +03004128 bitmap_zero(map, page->objects);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004129
Christoph Lameter5f80b132011-04-15 14:48:13 -05004130 get_map(s, page, map);
4131 for_each_object(p, s, addr, page->objects) {
4132 if (test_bit(slab_index(p, s, addr), map))
4133 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4134 return 0;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004135 }
4136
Christoph Lameter224a88b2008-04-14 19:11:31 +03004137 for_each_object(p, s, addr, page->objects)
Christoph Lameter7656c722007-05-09 02:32:40 -07004138 if (!test_bit(slab_index(p, s, addr), map))
Tero Roponen37d57442010-12-01 20:04:20 +02004139 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
Christoph Lameter53e15af2007-05-06 14:49:43 -07004140 return 0;
4141 return 1;
4142}
4143
Christoph Lameter434e2452007-07-17 04:03:30 -07004144static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4145 unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004146{
Christoph Lameter881db7f2011-06-01 12:25:53 -05004147 slab_lock(page);
4148 validate_slab(s, page, map);
4149 slab_unlock(page);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004150}
4151
Christoph Lameter434e2452007-07-17 04:03:30 -07004152static int validate_slab_node(struct kmem_cache *s,
4153 struct kmem_cache_node *n, unsigned long *map)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004154{
4155 unsigned long count = 0;
4156 struct page *page;
4157 unsigned long flags;
4158
4159 spin_lock_irqsave(&n->list_lock, flags);
4160
4161 list_for_each_entry(page, &n->partial, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07004162 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004163 count++;
4164 }
4165 if (count != n->nr_partial)
Fabian Frederickf9f58282014-06-04 16:06:34 -07004166 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4167 s->name, count, n->nr_partial);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004168
4169 if (!(s->flags & SLAB_STORE_USER))
4170 goto out;
4171
4172 list_for_each_entry(page, &n->full, lru) {
Christoph Lameter434e2452007-07-17 04:03:30 -07004173 validate_slab_slab(s, page, map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004174 count++;
4175 }
4176 if (count != atomic_long_read(&n->nr_slabs))
Fabian Frederickf9f58282014-06-04 16:06:34 -07004177 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4178 s->name, count, atomic_long_read(&n->nr_slabs));
Christoph Lameter53e15af2007-05-06 14:49:43 -07004179
4180out:
4181 spin_unlock_irqrestore(&n->list_lock, flags);
4182 return count;
4183}
4184
Christoph Lameter434e2452007-07-17 04:03:30 -07004185static long validate_slab_cache(struct kmem_cache *s)
Christoph Lameter53e15af2007-05-06 14:49:43 -07004186{
4187 int node;
4188 unsigned long count = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03004189 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
Christoph Lameter434e2452007-07-17 04:03:30 -07004190 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004191 struct kmem_cache_node *n;
Christoph Lameter434e2452007-07-17 04:03:30 -07004192
4193 if (!map)
4194 return -ENOMEM;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004195
4196 flush_all(s);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004197 for_each_kmem_cache_node(s, node, n)
Christoph Lameter434e2452007-07-17 04:03:30 -07004198 count += validate_slab_node(s, n, map);
Christoph Lameter434e2452007-07-17 04:03:30 -07004199 kfree(map);
Christoph Lameter53e15af2007-05-06 14:49:43 -07004200 return count;
4201}
Christoph Lameter88a420e2007-05-06 14:49:45 -07004202/*
Christoph Lameter672bba32007-05-09 02:32:39 -07004203 * Generate lists of code addresses where slabcache objects are allocated
Christoph Lameter88a420e2007-05-06 14:49:45 -07004204 * and freed.
4205 */
4206
4207struct location {
4208 unsigned long count;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004209 unsigned long addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004210 long long sum_time;
4211 long min_time;
4212 long max_time;
4213 long min_pid;
4214 long max_pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304215 DECLARE_BITMAP(cpus, NR_CPUS);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004216 nodemask_t nodes;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004217};
4218
4219struct loc_track {
4220 unsigned long max;
4221 unsigned long count;
4222 struct location *loc;
4223};
4224
4225static void free_loc_track(struct loc_track *t)
4226{
4227 if (t->max)
4228 free_pages((unsigned long)t->loc,
4229 get_order(sizeof(struct location) * t->max));
4230}
4231
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004232static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004233{
4234 struct location *l;
4235 int order;
4236
Christoph Lameter88a420e2007-05-06 14:49:45 -07004237 order = get_order(sizeof(struct location) * max);
4238
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004239 l = (void *)__get_free_pages(flags, order);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004240 if (!l)
4241 return 0;
4242
4243 if (t->count) {
4244 memcpy(l, t->loc, sizeof(struct location) * t->count);
4245 free_loc_track(t);
4246 }
4247 t->max = max;
4248 t->loc = l;
4249 return 1;
4250}
4251
4252static int add_location(struct loc_track *t, struct kmem_cache *s,
Christoph Lameter45edfa52007-05-09 02:32:45 -07004253 const struct track *track)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004254{
4255 long start, end, pos;
4256 struct location *l;
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03004257 unsigned long caddr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004258 unsigned long age = jiffies - track->when;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004259
4260 start = -1;
4261 end = t->count;
4262
4263 for ( ; ; ) {
4264 pos = start + (end - start + 1) / 2;
4265
4266 /*
4267 * There is nothing at "end". If we end up there
4268 * we need to add something to before end.
4269 */
4270 if (pos == end)
4271 break;
4272
4273 caddr = t->loc[pos].addr;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004274 if (track->addr == caddr) {
4275
4276 l = &t->loc[pos];
4277 l->count++;
4278 if (track->when) {
4279 l->sum_time += age;
4280 if (age < l->min_time)
4281 l->min_time = age;
4282 if (age > l->max_time)
4283 l->max_time = age;
4284
4285 if (track->pid < l->min_pid)
4286 l->min_pid = track->pid;
4287 if (track->pid > l->max_pid)
4288 l->max_pid = track->pid;
4289
Rusty Russell174596a2009-01-01 10:12:29 +10304290 cpumask_set_cpu(track->cpu,
4291 to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004292 }
4293 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004294 return 1;
4295 }
4296
Christoph Lameter45edfa52007-05-09 02:32:45 -07004297 if (track->addr < caddr)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004298 end = pos;
4299 else
4300 start = pos;
4301 }
4302
4303 /*
Christoph Lameter672bba32007-05-09 02:32:39 -07004304 * Not found. Insert new tracking element.
Christoph Lameter88a420e2007-05-06 14:49:45 -07004305 */
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004306 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004307 return 0;
4308
4309 l = t->loc + pos;
4310 if (pos < t->count)
4311 memmove(l + 1, l,
4312 (t->count - pos) * sizeof(struct location));
4313 t->count++;
4314 l->count = 1;
Christoph Lameter45edfa52007-05-09 02:32:45 -07004315 l->addr = track->addr;
4316 l->sum_time = age;
4317 l->min_time = age;
4318 l->max_time = age;
4319 l->min_pid = track->pid;
4320 l->max_pid = track->pid;
Rusty Russell174596a2009-01-01 10:12:29 +10304321 cpumask_clear(to_cpumask(l->cpus));
4322 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004323 nodes_clear(l->nodes);
4324 node_set(page_to_nid(virt_to_page(track)), l->nodes);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004325 return 1;
4326}
4327
4328static void process_slab(struct loc_track *t, struct kmem_cache *s,
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004329 struct page *page, enum track_item alloc,
Namhyung Kima5dd5c12010-09-29 21:02:13 +09004330 unsigned long *map)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004331{
Christoph Lametera973e9d2008-03-01 13:40:44 -08004332 void *addr = page_address(page);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004333 void *p;
4334
Christoph Lameter39b26462008-04-14 19:11:30 +03004335 bitmap_zero(map, page->objects);
Christoph Lameter5f80b132011-04-15 14:48:13 -05004336 get_map(s, page, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004337
Christoph Lameter224a88b2008-04-14 19:11:31 +03004338 for_each_object(p, s, addr, page->objects)
Christoph Lameter45edfa52007-05-09 02:32:45 -07004339 if (!test_bit(slab_index(p, s, addr), map))
4340 add_location(t, s, get_track(s, p, alloc));
Christoph Lameter88a420e2007-05-06 14:49:45 -07004341}
4342
4343static int list_locations(struct kmem_cache *s, char *buf,
4344 enum track_item alloc)
4345{
Harvey Harrisone374d482008-01-31 15:20:50 -08004346 int len = 0;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004347 unsigned long i;
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004348 struct loc_track t = { 0, 0, NULL };
Christoph Lameter88a420e2007-05-06 14:49:45 -07004349 int node;
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004350 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4351 sizeof(unsigned long), GFP_KERNEL);
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004352 struct kmem_cache_node *n;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004353
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004354 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4355 GFP_TEMPORARY)) {
4356 kfree(map);
Christoph Lameter68dff6a2007-07-17 04:03:20 -07004357 return sprintf(buf, "Out of memory\n");
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004358 }
Christoph Lameter88a420e2007-05-06 14:49:45 -07004359 /* Push back cpu slabs */
4360 flush_all(s);
4361
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004362 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter88a420e2007-05-06 14:49:45 -07004363 unsigned long flags;
4364 struct page *page;
4365
Christoph Lameter9e869432007-08-22 14:01:56 -07004366 if (!atomic_long_read(&n->nr_slabs))
Christoph Lameter88a420e2007-05-06 14:49:45 -07004367 continue;
4368
4369 spin_lock_irqsave(&n->list_lock, flags);
4370 list_for_each_entry(page, &n->partial, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004371 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004372 list_for_each_entry(page, &n->full, lru)
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004373 process_slab(&t, s, page, alloc, map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004374 spin_unlock_irqrestore(&n->list_lock, flags);
4375 }
4376
4377 for (i = 0; i < t.count; i++) {
Christoph Lameter45edfa52007-05-09 02:32:45 -07004378 struct location *l = &t.loc[i];
Christoph Lameter88a420e2007-05-06 14:49:45 -07004379
Hugh Dickins9c246242008-12-09 13:14:27 -08004380 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
Christoph Lameter88a420e2007-05-06 14:49:45 -07004381 break;
Harvey Harrisone374d482008-01-31 15:20:50 -08004382 len += sprintf(buf + len, "%7ld ", l->count);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004383
4384 if (l->addr)
Joe Perches62c70bc2011-01-13 15:45:52 -08004385 len += sprintf(buf + len, "%pS", (void *)l->addr);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004386 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004387 len += sprintf(buf + len, "<not-available>");
Christoph Lameter45edfa52007-05-09 02:32:45 -07004388
4389 if (l->sum_time != l->min_time) {
Harvey Harrisone374d482008-01-31 15:20:50 -08004390 len += sprintf(buf + len, " age=%ld/%ld/%ld",
Roman Zippelf8bd2252008-05-01 04:34:31 -07004391 l->min_time,
4392 (long)div_u64(l->sum_time, l->count),
4393 l->max_time);
Christoph Lameter45edfa52007-05-09 02:32:45 -07004394 } else
Harvey Harrisone374d482008-01-31 15:20:50 -08004395 len += sprintf(buf + len, " age=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004396 l->min_time);
4397
4398 if (l->min_pid != l->max_pid)
Harvey Harrisone374d482008-01-31 15:20:50 -08004399 len += sprintf(buf + len, " pid=%ld-%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004400 l->min_pid, l->max_pid);
4401 else
Harvey Harrisone374d482008-01-31 15:20:50 -08004402 len += sprintf(buf + len, " pid=%ld",
Christoph Lameter45edfa52007-05-09 02:32:45 -07004403 l->min_pid);
4404
Rusty Russell174596a2009-01-01 10:12:29 +10304405 if (num_online_cpus() > 1 &&
4406 !cpumask_empty(to_cpumask(l->cpus)) &&
Tejun Heo5024c1d2015-02-13 14:37:59 -08004407 len < PAGE_SIZE - 60)
4408 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4409 " cpus=%*pbl",
4410 cpumask_pr_args(to_cpumask(l->cpus)));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004411
Christoph Lameter62bc62a2009-06-16 15:32:15 -07004412 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
Tejun Heo5024c1d2015-02-13 14:37:59 -08004413 len < PAGE_SIZE - 60)
4414 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4415 " nodes=%*pbl",
4416 nodemask_pr_args(&l->nodes));
Christoph Lameter45edfa52007-05-09 02:32:45 -07004417
Harvey Harrisone374d482008-01-31 15:20:50 -08004418 len += sprintf(buf + len, "\n");
Christoph Lameter88a420e2007-05-06 14:49:45 -07004419 }
4420
4421 free_loc_track(&t);
Eric Dumazetbbd7d572010-03-24 22:25:47 +01004422 kfree(map);
Christoph Lameter88a420e2007-05-06 14:49:45 -07004423 if (!t.count)
Harvey Harrisone374d482008-01-31 15:20:50 -08004424 len += sprintf(buf, "No data\n");
4425 return len;
Christoph Lameter88a420e2007-05-06 14:49:45 -07004426}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004427#endif
Christoph Lameter88a420e2007-05-06 14:49:45 -07004428
Christoph Lametera5a84752010-10-05 13:57:27 -05004429#ifdef SLUB_RESILIENCY_TEST
David Rientjesc07b8182014-08-06 16:04:16 -07004430static void __init resiliency_test(void)
Christoph Lametera5a84752010-10-05 13:57:27 -05004431{
4432 u8 *p;
4433
Christoph Lameter95a05b42013-01-10 19:14:19 +00004434 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
Christoph Lametera5a84752010-10-05 13:57:27 -05004435
Fabian Frederickf9f58282014-06-04 16:06:34 -07004436 pr_err("SLUB resiliency testing\n");
4437 pr_err("-----------------------\n");
4438 pr_err("A. Corruption after allocation\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004439
4440 p = kzalloc(16, GFP_KERNEL);
4441 p[16] = 0x12;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004442 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4443 p + 16);
Christoph Lametera5a84752010-10-05 13:57:27 -05004444
4445 validate_slab_cache(kmalloc_caches[4]);
4446
4447 /* Hmmm... The next two are dangerous */
4448 p = kzalloc(32, GFP_KERNEL);
4449 p[32 + sizeof(void *)] = 0x34;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004450 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4451 p);
4452 pr_err("If allocated object is overwritten then not detectable\n\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004453
4454 validate_slab_cache(kmalloc_caches[5]);
4455 p = kzalloc(64, GFP_KERNEL);
4456 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4457 *p = 0x56;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004458 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4459 p);
4460 pr_err("If allocated object is overwritten then not detectable\n\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004461 validate_slab_cache(kmalloc_caches[6]);
4462
Fabian Frederickf9f58282014-06-04 16:06:34 -07004463 pr_err("\nB. Corruption after free\n");
Christoph Lametera5a84752010-10-05 13:57:27 -05004464 p = kzalloc(128, GFP_KERNEL);
4465 kfree(p);
4466 *p = 0x78;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004467 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004468 validate_slab_cache(kmalloc_caches[7]);
4469
4470 p = kzalloc(256, GFP_KERNEL);
4471 kfree(p);
4472 p[50] = 0x9a;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004473 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004474 validate_slab_cache(kmalloc_caches[8]);
4475
4476 p = kzalloc(512, GFP_KERNEL);
4477 kfree(p);
4478 p[512] = 0xab;
Fabian Frederickf9f58282014-06-04 16:06:34 -07004479 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
Christoph Lametera5a84752010-10-05 13:57:27 -05004480 validate_slab_cache(kmalloc_caches[9]);
4481}
4482#else
4483#ifdef CONFIG_SYSFS
4484static void resiliency_test(void) {};
4485#endif
4486#endif
4487
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004488#ifdef CONFIG_SYSFS
Christoph Lameter81819f02007-05-06 14:49:36 -07004489enum slab_stat_type {
Christoph Lameter205ab992008-04-14 19:11:40 +03004490 SL_ALL, /* All slabs */
4491 SL_PARTIAL, /* Only partially allocated slabs */
4492 SL_CPU, /* Only slabs used for cpu caches */
4493 SL_OBJECTS, /* Determine allocated objects not slabs */
4494 SL_TOTAL /* Determine object capacity not slabs */
Christoph Lameter81819f02007-05-06 14:49:36 -07004495};
4496
Christoph Lameter205ab992008-04-14 19:11:40 +03004497#define SO_ALL (1 << SL_ALL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004498#define SO_PARTIAL (1 << SL_PARTIAL)
4499#define SO_CPU (1 << SL_CPU)
4500#define SO_OBJECTS (1 << SL_OBJECTS)
Christoph Lameter205ab992008-04-14 19:11:40 +03004501#define SO_TOTAL (1 << SL_TOTAL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004502
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004503static ssize_t show_slab_objects(struct kmem_cache *s,
4504 char *buf, unsigned long flags)
Christoph Lameter81819f02007-05-06 14:49:36 -07004505{
4506 unsigned long total = 0;
Christoph Lameter81819f02007-05-06 14:49:36 -07004507 int node;
4508 int x;
4509 unsigned long *nodes;
Christoph Lameter81819f02007-05-06 14:49:36 -07004510
Chen Gange35e1a92013-07-12 08:23:48 +08004511 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
Cyrill Gorcunov62e5c4b2008-03-02 23:28:24 +03004512 if (!nodes)
4513 return -ENOMEM;
Christoph Lameter81819f02007-05-06 14:49:36 -07004514
Christoph Lameter205ab992008-04-14 19:11:40 +03004515 if (flags & SO_CPU) {
4516 int cpu;
Christoph Lameter81819f02007-05-06 14:49:36 -07004517
Christoph Lameter205ab992008-04-14 19:11:40 +03004518 for_each_possible_cpu(cpu) {
Chen Gangd0e0ac92013-07-15 09:05:29 +08004519 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4520 cpu);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004521 int node;
Christoph Lameter49e22582011-08-09 16:12:27 -05004522 struct page *page;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004523
Jason Low4db0c3c2015-04-15 16:14:08 -07004524 page = READ_ONCE(c->page);
Christoph Lameterec3ab082012-05-09 10:09:56 -05004525 if (!page)
4526 continue;
Christoph Lameter205ab992008-04-14 19:11:40 +03004527
Christoph Lameterec3ab082012-05-09 10:09:56 -05004528 node = page_to_nid(page);
4529 if (flags & SO_TOTAL)
4530 x = page->objects;
4531 else if (flags & SO_OBJECTS)
4532 x = page->inuse;
4533 else
4534 x = 1;
Christoph Lameter49e22582011-08-09 16:12:27 -05004535
Christoph Lameterec3ab082012-05-09 10:09:56 -05004536 total += x;
4537 nodes[node] += x;
4538
Jason Low4db0c3c2015-04-15 16:14:08 -07004539 page = READ_ONCE(c->partial);
Christoph Lameter49e22582011-08-09 16:12:27 -05004540 if (page) {
Li Zefan8afb1472013-09-10 11:43:37 +08004541 node = page_to_nid(page);
4542 if (flags & SO_TOTAL)
4543 WARN_ON_ONCE(1);
4544 else if (flags & SO_OBJECTS)
4545 WARN_ON_ONCE(1);
4546 else
4547 x = page->pages;
Eric Dumazetbc6697d2011-11-22 16:02:02 +01004548 total += x;
4549 nodes[node] += x;
Christoph Lameter49e22582011-08-09 16:12:27 -05004550 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004551 }
4552 }
4553
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004554 get_online_mems();
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004555#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter205ab992008-04-14 19:11:40 +03004556 if (flags & SO_ALL) {
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004557 struct kmem_cache_node *n;
4558
4559 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter81819f02007-05-06 14:49:36 -07004560
Chen Gangd0e0ac92013-07-15 09:05:29 +08004561 if (flags & SO_TOTAL)
4562 x = atomic_long_read(&n->total_objects);
4563 else if (flags & SO_OBJECTS)
4564 x = atomic_long_read(&n->total_objects) -
4565 count_partial(n, count_free);
Christoph Lameter205ab992008-04-14 19:11:40 +03004566 else
4567 x = atomic_long_read(&n->nr_slabs);
4568 total += x;
4569 nodes[node] += x;
4570 }
4571
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004572 } else
4573#endif
4574 if (flags & SO_PARTIAL) {
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004575 struct kmem_cache_node *n;
Christoph Lameter205ab992008-04-14 19:11:40 +03004576
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004577 for_each_kmem_cache_node(s, node, n) {
Christoph Lameter205ab992008-04-14 19:11:40 +03004578 if (flags & SO_TOTAL)
4579 x = count_partial(n, count_total);
4580 else if (flags & SO_OBJECTS)
4581 x = count_partial(n, count_inuse);
Christoph Lameter81819f02007-05-06 14:49:36 -07004582 else
4583 x = n->nr_partial;
4584 total += x;
4585 nodes[node] += x;
4586 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004587 }
Christoph Lameter81819f02007-05-06 14:49:36 -07004588 x = sprintf(buf, "%lu", total);
4589#ifdef CONFIG_NUMA
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004590 for (node = 0; node < nr_node_ids; node++)
Christoph Lameter81819f02007-05-06 14:49:36 -07004591 if (nodes[node])
4592 x += sprintf(buf + x, " N%d=%lu",
4593 node, nodes[node]);
4594#endif
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07004595 put_online_mems();
Christoph Lameter81819f02007-05-06 14:49:36 -07004596 kfree(nodes);
4597 return x + sprintf(buf + x, "\n");
4598}
4599
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004600#ifdef CONFIG_SLUB_DEBUG
Christoph Lameter81819f02007-05-06 14:49:36 -07004601static int any_slab_objects(struct kmem_cache *s)
4602{
4603 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004604 struct kmem_cache_node *n;
Christoph Lameterdfb4f092007-10-16 01:26:05 -07004605
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004606 for_each_kmem_cache_node(s, node, n)
Benjamin Herrenschmidt4ea33e22008-05-06 20:42:39 -07004607 if (atomic_long_read(&n->total_objects))
Christoph Lameter81819f02007-05-06 14:49:36 -07004608 return 1;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07004609
Christoph Lameter81819f02007-05-06 14:49:36 -07004610 return 0;
4611}
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004612#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07004613
4614#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
Phil Carmody497888c2011-07-14 15:07:13 +03004615#define to_slab(n) container_of(n, struct kmem_cache, kobj)
Christoph Lameter81819f02007-05-06 14:49:36 -07004616
4617struct slab_attribute {
4618 struct attribute attr;
4619 ssize_t (*show)(struct kmem_cache *s, char *buf);
4620 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4621};
4622
4623#define SLAB_ATTR_RO(_name) \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004624 static struct slab_attribute _name##_attr = \
4625 __ATTR(_name, 0400, _name##_show, NULL)
Christoph Lameter81819f02007-05-06 14:49:36 -07004626
4627#define SLAB_ATTR(_name) \
4628 static struct slab_attribute _name##_attr = \
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004629 __ATTR(_name, 0600, _name##_show, _name##_store)
Christoph Lameter81819f02007-05-06 14:49:36 -07004630
Christoph Lameter81819f02007-05-06 14:49:36 -07004631static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4632{
4633 return sprintf(buf, "%d\n", s->size);
4634}
4635SLAB_ATTR_RO(slab_size);
4636
4637static ssize_t align_show(struct kmem_cache *s, char *buf)
4638{
4639 return sprintf(buf, "%d\n", s->align);
4640}
4641SLAB_ATTR_RO(align);
4642
4643static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4644{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004645 return sprintf(buf, "%d\n", s->object_size);
Christoph Lameter81819f02007-05-06 14:49:36 -07004646}
4647SLAB_ATTR_RO(object_size);
4648
4649static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4650{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004651 return sprintf(buf, "%d\n", oo_objects(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004652}
4653SLAB_ATTR_RO(objs_per_slab);
4654
Christoph Lameter06b285d2008-04-14 19:11:41 +03004655static ssize_t order_store(struct kmem_cache *s,
4656 const char *buf, size_t length)
4657{
Christoph Lameter0121c6192008-04-29 16:11:12 -07004658 unsigned long order;
4659 int err;
4660
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004661 err = kstrtoul(buf, 10, &order);
Christoph Lameter0121c6192008-04-29 16:11:12 -07004662 if (err)
4663 return err;
Christoph Lameter06b285d2008-04-14 19:11:41 +03004664
4665 if (order > slub_max_order || order < slub_min_order)
4666 return -EINVAL;
4667
4668 calculate_sizes(s, order);
4669 return length;
4670}
4671
Christoph Lameter81819f02007-05-06 14:49:36 -07004672static ssize_t order_show(struct kmem_cache *s, char *buf)
4673{
Christoph Lameter834f3d12008-04-14 19:11:31 +03004674 return sprintf(buf, "%d\n", oo_order(s->oo));
Christoph Lameter81819f02007-05-06 14:49:36 -07004675}
Christoph Lameter06b285d2008-04-14 19:11:41 +03004676SLAB_ATTR(order);
Christoph Lameter81819f02007-05-06 14:49:36 -07004677
David Rientjes73d342b2009-02-22 17:40:09 -08004678static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4679{
4680 return sprintf(buf, "%lu\n", s->min_partial);
4681}
4682
4683static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4684 size_t length)
4685{
4686 unsigned long min;
4687 int err;
4688
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004689 err = kstrtoul(buf, 10, &min);
David Rientjes73d342b2009-02-22 17:40:09 -08004690 if (err)
4691 return err;
4692
David Rientjesc0bdb232009-02-25 09:16:35 +02004693 set_min_partial(s, min);
David Rientjes73d342b2009-02-22 17:40:09 -08004694 return length;
4695}
4696SLAB_ATTR(min_partial);
4697
Christoph Lameter49e22582011-08-09 16:12:27 -05004698static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4699{
4700 return sprintf(buf, "%u\n", s->cpu_partial);
4701}
4702
4703static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4704 size_t length)
4705{
4706 unsigned long objects;
4707 int err;
4708
Jingoo Han3dbb95f2013-09-11 14:20:25 -07004709 err = kstrtoul(buf, 10, &objects);
Christoph Lameter49e22582011-08-09 16:12:27 -05004710 if (err)
4711 return err;
Joonsoo Kim345c9052013-06-19 14:05:52 +09004712 if (objects && !kmem_cache_has_cpu_partial(s))
David Rientjes74ee4ef2012-01-09 13:19:45 -08004713 return -EINVAL;
Christoph Lameter49e22582011-08-09 16:12:27 -05004714
4715 s->cpu_partial = objects;
4716 flush_all(s);
4717 return length;
4718}
4719SLAB_ATTR(cpu_partial);
4720
Christoph Lameter81819f02007-05-06 14:49:36 -07004721static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4722{
Joe Perches62c70bc2011-01-13 15:45:52 -08004723 if (!s->ctor)
4724 return 0;
4725 return sprintf(buf, "%pS\n", s->ctor);
Christoph Lameter81819f02007-05-06 14:49:36 -07004726}
4727SLAB_ATTR_RO(ctor);
4728
Christoph Lameter81819f02007-05-06 14:49:36 -07004729static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4730{
Gu Zheng4307c142014-08-06 16:04:51 -07004731 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004732}
4733SLAB_ATTR_RO(aliases);
4734
Christoph Lameter81819f02007-05-06 14:49:36 -07004735static ssize_t partial_show(struct kmem_cache *s, char *buf)
4736{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004737 return show_slab_objects(s, buf, SO_PARTIAL);
Christoph Lameter81819f02007-05-06 14:49:36 -07004738}
4739SLAB_ATTR_RO(partial);
4740
4741static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4742{
Christoph Lameterd9acf4b2008-02-15 15:22:21 -08004743 return show_slab_objects(s, buf, SO_CPU);
Christoph Lameter81819f02007-05-06 14:49:36 -07004744}
4745SLAB_ATTR_RO(cpu_slabs);
4746
4747static ssize_t objects_show(struct kmem_cache *s, char *buf)
4748{
Christoph Lameter205ab992008-04-14 19:11:40 +03004749 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
Christoph Lameter81819f02007-05-06 14:49:36 -07004750}
4751SLAB_ATTR_RO(objects);
4752
Christoph Lameter205ab992008-04-14 19:11:40 +03004753static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4754{
4755 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4756}
4757SLAB_ATTR_RO(objects_partial);
4758
Christoph Lameter49e22582011-08-09 16:12:27 -05004759static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4760{
4761 int objects = 0;
4762 int pages = 0;
4763 int cpu;
4764 int len;
4765
4766 for_each_online_cpu(cpu) {
4767 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4768
4769 if (page) {
4770 pages += page->pages;
4771 objects += page->pobjects;
4772 }
4773 }
4774
4775 len = sprintf(buf, "%d(%d)", objects, pages);
4776
4777#ifdef CONFIG_SMP
4778 for_each_online_cpu(cpu) {
4779 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4780
4781 if (page && len < PAGE_SIZE - 20)
4782 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4783 page->pobjects, page->pages);
4784 }
4785#endif
4786 return len + sprintf(buf + len, "\n");
4787}
4788SLAB_ATTR_RO(slabs_cpu_partial);
4789
Christoph Lameter81819f02007-05-06 14:49:36 -07004790static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4791{
4792 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4793}
4794
4795static ssize_t reclaim_account_store(struct kmem_cache *s,
4796 const char *buf, size_t length)
4797{
4798 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4799 if (buf[0] == '1')
4800 s->flags |= SLAB_RECLAIM_ACCOUNT;
4801 return length;
4802}
4803SLAB_ATTR(reclaim_account);
4804
4805static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4806{
Christoph Lameter5af60832007-05-06 14:49:56 -07004807 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07004808}
4809SLAB_ATTR_RO(hwcache_align);
4810
4811#ifdef CONFIG_ZONE_DMA
4812static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4813{
4814 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4815}
4816SLAB_ATTR_RO(cache_dma);
4817#endif
4818
4819static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4820{
4821 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4822}
4823SLAB_ATTR_RO(destroy_by_rcu);
4824
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08004825static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4826{
4827 return sprintf(buf, "%d\n", s->reserved);
4828}
4829SLAB_ATTR_RO(reserved);
4830
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05004831#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05004832static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4833{
4834 return show_slab_objects(s, buf, SO_ALL);
4835}
4836SLAB_ATTR_RO(slabs);
4837
4838static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4839{
4840 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4841}
4842SLAB_ATTR_RO(total_objects);
4843
4844static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4845{
Laura Abbottbecfda62016-03-15 14:55:06 -07004846 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
Christoph Lametera5a84752010-10-05 13:57:27 -05004847}
4848
4849static ssize_t sanity_checks_store(struct kmem_cache *s,
4850 const char *buf, size_t length)
4851{
Laura Abbottbecfda62016-03-15 14:55:06 -07004852 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004853 if (buf[0] == '1') {
4854 s->flags &= ~__CMPXCHG_DOUBLE;
Laura Abbottbecfda62016-03-15 14:55:06 -07004855 s->flags |= SLAB_CONSISTENCY_CHECKS;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004856 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004857 return length;
4858}
4859SLAB_ATTR(sanity_checks);
4860
4861static ssize_t trace_show(struct kmem_cache *s, char *buf)
4862{
4863 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4864}
4865
4866static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4867 size_t length)
4868{
Christoph Lameterc9e16132014-10-09 15:26:11 -07004869 /*
4870 * Tracing a merged cache is going to give confusing results
4871 * as well as cause other issues like converting a mergeable
4872 * cache into an umergeable one.
4873 */
4874 if (s->refcount > 1)
4875 return -EINVAL;
4876
Christoph Lametera5a84752010-10-05 13:57:27 -05004877 s->flags &= ~SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004878 if (buf[0] == '1') {
4879 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lametera5a84752010-10-05 13:57:27 -05004880 s->flags |= SLAB_TRACE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004881 }
Christoph Lametera5a84752010-10-05 13:57:27 -05004882 return length;
4883}
4884SLAB_ATTR(trace);
4885
Christoph Lameter81819f02007-05-06 14:49:36 -07004886static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4887{
4888 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4889}
4890
4891static ssize_t red_zone_store(struct kmem_cache *s,
4892 const char *buf, size_t length)
4893{
4894 if (any_slab_objects(s))
4895 return -EBUSY;
4896
4897 s->flags &= ~SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004898 if (buf[0] == '1') {
Christoph Lameter81819f02007-05-06 14:49:36 -07004899 s->flags |= SLAB_RED_ZONE;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004900 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004901 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004902 return length;
4903}
4904SLAB_ATTR(red_zone);
4905
4906static ssize_t poison_show(struct kmem_cache *s, char *buf)
4907{
4908 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4909}
4910
4911static ssize_t poison_store(struct kmem_cache *s,
4912 const char *buf, size_t length)
4913{
4914 if (any_slab_objects(s))
4915 return -EBUSY;
4916
4917 s->flags &= ~SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004918 if (buf[0] == '1') {
Christoph Lameter81819f02007-05-06 14:49:36 -07004919 s->flags |= SLAB_POISON;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004920 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004921 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004922 return length;
4923}
4924SLAB_ATTR(poison);
4925
4926static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4927{
4928 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4929}
4930
4931static ssize_t store_user_store(struct kmem_cache *s,
4932 const char *buf, size_t length)
4933{
4934 if (any_slab_objects(s))
4935 return -EBUSY;
4936
4937 s->flags &= ~SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004938 if (buf[0] == '1') {
4939 s->flags &= ~__CMPXCHG_DOUBLE;
Christoph Lameter81819f02007-05-06 14:49:36 -07004940 s->flags |= SLAB_STORE_USER;
Christoph Lameterb789ef52011-06-01 12:25:49 -05004941 }
Christoph Lameter06b285d2008-04-14 19:11:41 +03004942 calculate_sizes(s, -1);
Christoph Lameter81819f02007-05-06 14:49:36 -07004943 return length;
4944}
4945SLAB_ATTR(store_user);
4946
Christoph Lameter53e15af2007-05-06 14:49:43 -07004947static ssize_t validate_show(struct kmem_cache *s, char *buf)
4948{
4949 return 0;
4950}
4951
4952static ssize_t validate_store(struct kmem_cache *s,
4953 const char *buf, size_t length)
4954{
Christoph Lameter434e2452007-07-17 04:03:30 -07004955 int ret = -EINVAL;
4956
4957 if (buf[0] == '1') {
4958 ret = validate_slab_cache(s);
4959 if (ret >= 0)
4960 ret = length;
4961 }
4962 return ret;
Christoph Lameter53e15af2007-05-06 14:49:43 -07004963}
4964SLAB_ATTR(validate);
Christoph Lametera5a84752010-10-05 13:57:27 -05004965
4966static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4967{
4968 if (!(s->flags & SLAB_STORE_USER))
4969 return -ENOSYS;
4970 return list_locations(s, buf, TRACK_ALLOC);
4971}
4972SLAB_ATTR_RO(alloc_calls);
4973
4974static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4975{
4976 if (!(s->flags & SLAB_STORE_USER))
4977 return -ENOSYS;
4978 return list_locations(s, buf, TRACK_FREE);
4979}
4980SLAB_ATTR_RO(free_calls);
4981#endif /* CONFIG_SLUB_DEBUG */
4982
4983#ifdef CONFIG_FAILSLAB
4984static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4985{
4986 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4987}
4988
4989static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4990 size_t length)
4991{
Christoph Lameterc9e16132014-10-09 15:26:11 -07004992 if (s->refcount > 1)
4993 return -EINVAL;
4994
Christoph Lametera5a84752010-10-05 13:57:27 -05004995 s->flags &= ~SLAB_FAILSLAB;
4996 if (buf[0] == '1')
4997 s->flags |= SLAB_FAILSLAB;
4998 return length;
4999}
5000SLAB_ATTR(failslab);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005001#endif
Christoph Lameter53e15af2007-05-06 14:49:43 -07005002
Christoph Lameter2086d262007-05-06 14:49:46 -07005003static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5004{
5005 return 0;
5006}
5007
5008static ssize_t shrink_store(struct kmem_cache *s,
5009 const char *buf, size_t length)
5010{
Vladimir Davydov832f37f2015-02-12 14:59:41 -08005011 if (buf[0] == '1')
5012 kmem_cache_shrink(s);
5013 else
Christoph Lameter2086d262007-05-06 14:49:46 -07005014 return -EINVAL;
5015 return length;
5016}
5017SLAB_ATTR(shrink);
5018
Christoph Lameter81819f02007-05-06 14:49:36 -07005019#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08005020static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
Christoph Lameter81819f02007-05-06 14:49:36 -07005021{
Christoph Lameter98246012008-01-07 23:20:26 -08005022 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
Christoph Lameter81819f02007-05-06 14:49:36 -07005023}
5024
Christoph Lameter98246012008-01-07 23:20:26 -08005025static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
Christoph Lameter81819f02007-05-06 14:49:36 -07005026 const char *buf, size_t length)
5027{
Christoph Lameter0121c6192008-04-29 16:11:12 -07005028 unsigned long ratio;
5029 int err;
Christoph Lameter81819f02007-05-06 14:49:36 -07005030
Jingoo Han3dbb95f2013-09-11 14:20:25 -07005031 err = kstrtoul(buf, 10, &ratio);
Christoph Lameter0121c6192008-04-29 16:11:12 -07005032 if (err)
5033 return err;
5034
Christoph Lametere2cb96b2008-08-19 08:51:22 -05005035 if (ratio <= 100)
Christoph Lameter0121c6192008-04-29 16:11:12 -07005036 s->remote_node_defrag_ratio = ratio * 10;
5037
Christoph Lameter81819f02007-05-06 14:49:36 -07005038 return length;
5039}
Christoph Lameter98246012008-01-07 23:20:26 -08005040SLAB_ATTR(remote_node_defrag_ratio);
Christoph Lameter81819f02007-05-06 14:49:36 -07005041#endif
5042
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005043#ifdef CONFIG_SLUB_STATS
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005044static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5045{
5046 unsigned long sum = 0;
5047 int cpu;
5048 int len;
5049 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5050
5051 if (!data)
5052 return -ENOMEM;
5053
5054 for_each_online_cpu(cpu) {
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06005055 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005056
5057 data[cpu] = x;
5058 sum += x;
5059 }
5060
5061 len = sprintf(buf, "%lu", sum);
5062
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005063#ifdef CONFIG_SMP
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005064 for_each_online_cpu(cpu) {
5065 if (data[cpu] && len < PAGE_SIZE - 20)
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005066 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005067 }
Christoph Lameter50ef37b2008-04-14 18:52:05 +03005068#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005069 kfree(data);
5070 return len + sprintf(buf + len, "\n");
5071}
5072
David Rientjes78eb00c2009-10-15 02:20:22 -07005073static void clear_stat(struct kmem_cache *s, enum stat_item si)
5074{
5075 int cpu;
5076
5077 for_each_online_cpu(cpu)
Christoph Lameter9dfc6e62009-12-18 16:26:20 -06005078 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
David Rientjes78eb00c2009-10-15 02:20:22 -07005079}
5080
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005081#define STAT_ATTR(si, text) \
5082static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5083{ \
5084 return show_stat(s, buf, si); \
5085} \
David Rientjes78eb00c2009-10-15 02:20:22 -07005086static ssize_t text##_store(struct kmem_cache *s, \
5087 const char *buf, size_t length) \
5088{ \
5089 if (buf[0] != '0') \
5090 return -EINVAL; \
5091 clear_stat(s, si); \
5092 return length; \
5093} \
5094SLAB_ATTR(text); \
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005095
5096STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5097STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5098STAT_ATTR(FREE_FASTPATH, free_fastpath);
5099STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5100STAT_ATTR(FREE_FROZEN, free_frozen);
5101STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5102STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5103STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5104STAT_ATTR(ALLOC_SLAB, alloc_slab);
5105STAT_ATTR(ALLOC_REFILL, alloc_refill);
Christoph Lametere36a2652011-06-01 12:25:57 -05005106STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005107STAT_ATTR(FREE_SLAB, free_slab);
5108STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5109STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5110STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5111STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5112STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5113STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
Christoph Lameter03e404a2011-06-01 12:25:58 -05005114STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
Christoph Lameter65c33762008-04-14 19:11:40 +03005115STAT_ATTR(ORDER_FALLBACK, order_fallback);
Christoph Lameterb789ef52011-06-01 12:25:49 -05005116STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5117STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
Christoph Lameter49e22582011-08-09 16:12:27 -05005118STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5119STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
Alex Shi8028dce2012-02-03 23:34:56 +08005120STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5121STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005122#endif
5123
Pekka Enberg06428782008-01-07 23:20:27 -08005124static struct attribute *slab_attrs[] = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005125 &slab_size_attr.attr,
5126 &object_size_attr.attr,
5127 &objs_per_slab_attr.attr,
5128 &order_attr.attr,
David Rientjes73d342b2009-02-22 17:40:09 -08005129 &min_partial_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005130 &cpu_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005131 &objects_attr.attr,
Christoph Lameter205ab992008-04-14 19:11:40 +03005132 &objects_partial_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005133 &partial_attr.attr,
5134 &cpu_slabs_attr.attr,
5135 &ctor_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005136 &aliases_attr.attr,
5137 &align_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005138 &hwcache_align_attr.attr,
5139 &reclaim_account_attr.attr,
5140 &destroy_by_rcu_attr.attr,
Christoph Lametera5a84752010-10-05 13:57:27 -05005141 &shrink_attr.attr,
Lai Jiangshanab9a0f12011-03-10 15:21:48 +08005142 &reserved_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005143 &slabs_cpu_partial_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005144#ifdef CONFIG_SLUB_DEBUG
Christoph Lametera5a84752010-10-05 13:57:27 -05005145 &total_objects_attr.attr,
5146 &slabs_attr.attr,
5147 &sanity_checks_attr.attr,
5148 &trace_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005149 &red_zone_attr.attr,
5150 &poison_attr.attr,
5151 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07005152 &validate_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07005153 &alloc_calls_attr.attr,
5154 &free_calls_attr.attr,
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005155#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005156#ifdef CONFIG_ZONE_DMA
5157 &cache_dma_attr.attr,
5158#endif
5159#ifdef CONFIG_NUMA
Christoph Lameter98246012008-01-07 23:20:26 -08005160 &remote_node_defrag_ratio_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07005161#endif
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005162#ifdef CONFIG_SLUB_STATS
5163 &alloc_fastpath_attr.attr,
5164 &alloc_slowpath_attr.attr,
5165 &free_fastpath_attr.attr,
5166 &free_slowpath_attr.attr,
5167 &free_frozen_attr.attr,
5168 &free_add_partial_attr.attr,
5169 &free_remove_partial_attr.attr,
5170 &alloc_from_partial_attr.attr,
5171 &alloc_slab_attr.attr,
5172 &alloc_refill_attr.attr,
Christoph Lametere36a2652011-06-01 12:25:57 -05005173 &alloc_node_mismatch_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005174 &free_slab_attr.attr,
5175 &cpuslab_flush_attr.attr,
5176 &deactivate_full_attr.attr,
5177 &deactivate_empty_attr.attr,
5178 &deactivate_to_head_attr.attr,
5179 &deactivate_to_tail_attr.attr,
5180 &deactivate_remote_frees_attr.attr,
Christoph Lameter03e404a2011-06-01 12:25:58 -05005181 &deactivate_bypass_attr.attr,
Christoph Lameter65c33762008-04-14 19:11:40 +03005182 &order_fallback_attr.attr,
Christoph Lameterb789ef52011-06-01 12:25:49 -05005183 &cmpxchg_double_fail_attr.attr,
5184 &cmpxchg_double_cpu_fail_attr.attr,
Christoph Lameter49e22582011-08-09 16:12:27 -05005185 &cpu_partial_alloc_attr.attr,
5186 &cpu_partial_free_attr.attr,
Alex Shi8028dce2012-02-03 23:34:56 +08005187 &cpu_partial_node_attr.attr,
5188 &cpu_partial_drain_attr.attr,
Christoph Lameter8ff12cf2008-02-07 17:47:41 -08005189#endif
Dmitry Monakhov4c13dd32010-02-26 09:36:12 +03005190#ifdef CONFIG_FAILSLAB
5191 &failslab_attr.attr,
5192#endif
5193
Christoph Lameter81819f02007-05-06 14:49:36 -07005194 NULL
5195};
5196
5197static struct attribute_group slab_attr_group = {
5198 .attrs = slab_attrs,
5199};
5200
5201static ssize_t slab_attr_show(struct kobject *kobj,
5202 struct attribute *attr,
5203 char *buf)
5204{
5205 struct slab_attribute *attribute;
5206 struct kmem_cache *s;
5207 int err;
5208
5209 attribute = to_slab_attr(attr);
5210 s = to_slab(kobj);
5211
5212 if (!attribute->show)
5213 return -EIO;
5214
5215 err = attribute->show(s, buf);
5216
5217 return err;
5218}
5219
5220static ssize_t slab_attr_store(struct kobject *kobj,
5221 struct attribute *attr,
5222 const char *buf, size_t len)
5223{
5224 struct slab_attribute *attribute;
5225 struct kmem_cache *s;
5226 int err;
5227
5228 attribute = to_slab_attr(attr);
5229 s = to_slab(kobj);
5230
5231 if (!attribute->store)
5232 return -EIO;
5233
5234 err = attribute->store(s, buf, len);
Johannes Weiner127424c2016-01-20 15:02:32 -08005235#ifdef CONFIG_MEMCG
Glauber Costa107dab52012-12-18 14:23:05 -08005236 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
Vladimir Davydov426589f2015-02-12 14:59:23 -08005237 struct kmem_cache *c;
Christoph Lameter81819f02007-05-06 14:49:36 -07005238
Glauber Costa107dab52012-12-18 14:23:05 -08005239 mutex_lock(&slab_mutex);
5240 if (s->max_attr_size < len)
5241 s->max_attr_size = len;
5242
Glauber Costaebe945c2012-12-18 14:23:10 -08005243 /*
5244 * This is a best effort propagation, so this function's return
5245 * value will be determined by the parent cache only. This is
5246 * basically because not all attributes will have a well
5247 * defined semantics for rollbacks - most of the actions will
5248 * have permanent effects.
5249 *
5250 * Returning the error value of any of the children that fail
5251 * is not 100 % defined, in the sense that users seeing the
5252 * error code won't be able to know anything about the state of
5253 * the cache.
5254 *
5255 * Only returning the error code for the parent cache at least
5256 * has well defined semantics. The cache being written to
5257 * directly either failed or succeeded, in which case we loop
5258 * through the descendants with best-effort propagation.
5259 */
Vladimir Davydov426589f2015-02-12 14:59:23 -08005260 for_each_memcg_cache(c, s)
5261 attribute->store(c, buf, len);
Glauber Costa107dab52012-12-18 14:23:05 -08005262 mutex_unlock(&slab_mutex);
5263 }
5264#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005265 return err;
5266}
5267
Glauber Costa107dab52012-12-18 14:23:05 -08005268static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5269{
Johannes Weiner127424c2016-01-20 15:02:32 -08005270#ifdef CONFIG_MEMCG
Glauber Costa107dab52012-12-18 14:23:05 -08005271 int i;
5272 char *buffer = NULL;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005273 struct kmem_cache *root_cache;
Glauber Costa107dab52012-12-18 14:23:05 -08005274
Vladimir Davydov93030d82014-05-06 12:49:59 -07005275 if (is_root_cache(s))
Glauber Costa107dab52012-12-18 14:23:05 -08005276 return;
5277
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08005278 root_cache = s->memcg_params.root_cache;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005279
Glauber Costa107dab52012-12-18 14:23:05 -08005280 /*
5281 * This mean this cache had no attribute written. Therefore, no point
5282 * in copying default values around
5283 */
Vladimir Davydov93030d82014-05-06 12:49:59 -07005284 if (!root_cache->max_attr_size)
Glauber Costa107dab52012-12-18 14:23:05 -08005285 return;
5286
5287 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5288 char mbuf[64];
5289 char *buf;
5290 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5291
5292 if (!attr || !attr->store || !attr->show)
5293 continue;
5294
5295 /*
5296 * It is really bad that we have to allocate here, so we will
5297 * do it only as a fallback. If we actually allocate, though,
5298 * we can just use the allocated buffer until the end.
5299 *
5300 * Most of the slub attributes will tend to be very small in
5301 * size, but sysfs allows buffers up to a page, so they can
5302 * theoretically happen.
5303 */
5304 if (buffer)
5305 buf = buffer;
Vladimir Davydov93030d82014-05-06 12:49:59 -07005306 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
Glauber Costa107dab52012-12-18 14:23:05 -08005307 buf = mbuf;
5308 else {
5309 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5310 if (WARN_ON(!buffer))
5311 continue;
5312 buf = buffer;
5313 }
5314
Vladimir Davydov93030d82014-05-06 12:49:59 -07005315 attr->show(root_cache, buf);
Glauber Costa107dab52012-12-18 14:23:05 -08005316 attr->store(s, buf, strlen(buf));
5317 }
5318
5319 if (buffer)
5320 free_page((unsigned long)buffer);
5321#endif
5322}
5323
Christoph Lameter41a21282014-05-06 12:50:08 -07005324static void kmem_cache_release(struct kobject *k)
5325{
5326 slab_kmem_cache_release(to_slab(k));
5327}
5328
Emese Revfy52cf25d2010-01-19 02:58:23 +01005329static const struct sysfs_ops slab_sysfs_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005330 .show = slab_attr_show,
5331 .store = slab_attr_store,
5332};
5333
5334static struct kobj_type slab_ktype = {
5335 .sysfs_ops = &slab_sysfs_ops,
Christoph Lameter41a21282014-05-06 12:50:08 -07005336 .release = kmem_cache_release,
Christoph Lameter81819f02007-05-06 14:49:36 -07005337};
5338
5339static int uevent_filter(struct kset *kset, struct kobject *kobj)
5340{
5341 struct kobj_type *ktype = get_ktype(kobj);
5342
5343 if (ktype == &slab_ktype)
5344 return 1;
5345 return 0;
5346}
5347
Emese Revfy9cd43612009-12-31 14:52:51 +01005348static const struct kset_uevent_ops slab_uevent_ops = {
Christoph Lameter81819f02007-05-06 14:49:36 -07005349 .filter = uevent_filter,
5350};
5351
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005352static struct kset *slab_kset;
Christoph Lameter81819f02007-05-06 14:49:36 -07005353
Vladimir Davydov9a417072014-04-07 15:39:31 -07005354static inline struct kset *cache_kset(struct kmem_cache *s)
5355{
Johannes Weiner127424c2016-01-20 15:02:32 -08005356#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005357 if (!is_root_cache(s))
Vladimir Davydovf7ce3192015-02-12 14:59:20 -08005358 return s->memcg_params.root_cache->memcg_kset;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005359#endif
5360 return slab_kset;
5361}
5362
Christoph Lameter81819f02007-05-06 14:49:36 -07005363#define ID_STR_LENGTH 64
5364
5365/* Create a unique string id for a slab cache:
Christoph Lameter6446faa2008-02-15 23:45:26 -08005366 *
5367 * Format :[flags-]size
Christoph Lameter81819f02007-05-06 14:49:36 -07005368 */
5369static char *create_unique_id(struct kmem_cache *s)
5370{
5371 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5372 char *p = name;
5373
5374 BUG_ON(!name);
5375
5376 *p++ = ':';
5377 /*
5378 * First flags affecting slabcache operations. We will only
5379 * get here for aliasable slabs so we do not need to support
5380 * too many flags. The flags here must cover all flags that
5381 * are matched during merging to guarantee that the id is
5382 * unique.
5383 */
5384 if (s->flags & SLAB_CACHE_DMA)
5385 *p++ = 'd';
5386 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5387 *p++ = 'a';
Laura Abbottbecfda62016-03-15 14:55:06 -07005388 if (s->flags & SLAB_CONSISTENCY_CHECKS)
Christoph Lameter81819f02007-05-06 14:49:36 -07005389 *p++ = 'F';
Vegard Nossum5a896d92008-04-04 00:54:48 +02005390 if (!(s->flags & SLAB_NOTRACK))
5391 *p++ = 't';
Vladimir Davydov230e9fc2016-01-14 15:18:15 -08005392 if (s->flags & SLAB_ACCOUNT)
5393 *p++ = 'A';
Christoph Lameter81819f02007-05-06 14:49:36 -07005394 if (p != name + 1)
5395 *p++ = '-';
5396 p += sprintf(p, "%07d", s->size);
Glauber Costa2633d7a2012-12-18 14:22:34 -08005397
Christoph Lameter81819f02007-05-06 14:49:36 -07005398 BUG_ON(p > name + ID_STR_LENGTH - 1);
5399 return name;
5400}
5401
5402static int sysfs_slab_add(struct kmem_cache *s)
5403{
5404 int err;
5405 const char *name;
Christoph Lameter45530c42012-11-28 16:23:07 +00005406 int unmergeable = slab_unmergeable(s);
Christoph Lameter81819f02007-05-06 14:49:36 -07005407
Christoph Lameter81819f02007-05-06 14:49:36 -07005408 if (unmergeable) {
5409 /*
5410 * Slabcache can never be merged so we can use the name proper.
5411 * This is typically the case for debug situations. In that
5412 * case we can catch duplicate names easily.
5413 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005414 sysfs_remove_link(&slab_kset->kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005415 name = s->name;
5416 } else {
5417 /*
5418 * Create a unique name for the slab as a target
5419 * for the symlinks.
5420 */
5421 name = create_unique_id(s);
5422 }
5423
Vladimir Davydov9a417072014-04-07 15:39:31 -07005424 s->kobj.kset = cache_kset(s);
Tetsuo Handa26e4f202014-01-04 16:32:31 +09005425 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
Dave Jones54b6a732014-04-07 15:39:32 -07005426 if (err)
Konstantin Khlebnikov80da0262015-09-04 15:45:51 -07005427 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07005428
5429 err = sysfs_create_group(&s->kobj, &slab_attr_group);
Dave Jones54b6a732014-04-07 15:39:32 -07005430 if (err)
5431 goto out_del_kobj;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005432
Johannes Weiner127424c2016-01-20 15:02:32 -08005433#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005434 if (is_root_cache(s)) {
5435 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5436 if (!s->memcg_kset) {
Dave Jones54b6a732014-04-07 15:39:32 -07005437 err = -ENOMEM;
5438 goto out_del_kobj;
Vladimir Davydov9a417072014-04-07 15:39:31 -07005439 }
5440 }
5441#endif
5442
Christoph Lameter81819f02007-05-06 14:49:36 -07005443 kobject_uevent(&s->kobj, KOBJ_ADD);
5444 if (!unmergeable) {
5445 /* Setup first alias */
5446 sysfs_slab_alias(s, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005447 }
Dave Jones54b6a732014-04-07 15:39:32 -07005448out:
5449 if (!unmergeable)
5450 kfree(name);
5451 return err;
5452out_del_kobj:
5453 kobject_del(&s->kobj);
Dave Jones54b6a732014-04-07 15:39:32 -07005454 goto out;
Christoph Lameter81819f02007-05-06 14:49:36 -07005455}
5456
Christoph Lameter41a21282014-05-06 12:50:08 -07005457void sysfs_slab_remove(struct kmem_cache *s)
Christoph Lameter81819f02007-05-06 14:49:36 -07005458{
Christoph Lameter97d06602012-07-06 15:25:11 -05005459 if (slab_state < FULL)
Christoph Lameter2bce6482010-07-19 11:39:11 -05005460 /*
5461 * Sysfs has not been setup yet so no need to remove the
5462 * cache from sysfs.
5463 */
5464 return;
5465
Johannes Weiner127424c2016-01-20 15:02:32 -08005466#ifdef CONFIG_MEMCG
Vladimir Davydov9a417072014-04-07 15:39:31 -07005467 kset_unregister(s->memcg_kset);
5468#endif
Christoph Lameter81819f02007-05-06 14:49:36 -07005469 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5470 kobject_del(&s->kobj);
Christoph Lameter151c6022008-01-07 22:29:05 -08005471 kobject_put(&s->kobj);
Christoph Lameter81819f02007-05-06 14:49:36 -07005472}
5473
5474/*
5475 * Need to buffer aliases during bootup until sysfs becomes
Nick Andrew9f6c708e2008-12-05 14:08:08 +11005476 * available lest we lose that information.
Christoph Lameter81819f02007-05-06 14:49:36 -07005477 */
5478struct saved_alias {
5479 struct kmem_cache *s;
5480 const char *name;
5481 struct saved_alias *next;
5482};
5483
Adrian Bunk5af328a2007-07-17 04:03:27 -07005484static struct saved_alias *alias_list;
Christoph Lameter81819f02007-05-06 14:49:36 -07005485
5486static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5487{
5488 struct saved_alias *al;
5489
Christoph Lameter97d06602012-07-06 15:25:11 -05005490 if (slab_state == FULL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07005491 /*
5492 * If we have a leftover link then remove it.
5493 */
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005494 sysfs_remove_link(&slab_kset->kobj, name);
5495 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005496 }
5497
5498 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5499 if (!al)
5500 return -ENOMEM;
5501
5502 al->s = s;
5503 al->name = name;
5504 al->next = alias_list;
5505 alias_list = al;
5506 return 0;
5507}
5508
5509static int __init slab_sysfs_init(void)
5510{
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005511 struct kmem_cache *s;
Christoph Lameter81819f02007-05-06 14:49:36 -07005512 int err;
5513
Christoph Lameter18004c52012-07-06 15:25:12 -05005514 mutex_lock(&slab_mutex);
Christoph Lameter2bce6482010-07-19 11:39:11 -05005515
Greg Kroah-Hartman0ff21e42007-11-06 10:36:58 -08005516 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
Greg Kroah-Hartman27c3a312007-11-01 09:29:06 -06005517 if (!slab_kset) {
Christoph Lameter18004c52012-07-06 15:25:12 -05005518 mutex_unlock(&slab_mutex);
Fabian Frederickf9f58282014-06-04 16:06:34 -07005519 pr_err("Cannot register slab subsystem.\n");
Christoph Lameter81819f02007-05-06 14:49:36 -07005520 return -ENOSYS;
5521 }
5522
Christoph Lameter97d06602012-07-06 15:25:11 -05005523 slab_state = FULL;
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005524
Christoph Lameter5b95a4ac2007-07-17 04:03:19 -07005525 list_for_each_entry(s, &slab_caches, list) {
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005526 err = sysfs_slab_add(s);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005527 if (err)
Fabian Frederickf9f58282014-06-04 16:06:34 -07005528 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5529 s->name);
Christoph Lameter26a7bd02007-05-09 02:32:39 -07005530 }
Christoph Lameter81819f02007-05-06 14:49:36 -07005531
5532 while (alias_list) {
5533 struct saved_alias *al = alias_list;
5534
5535 alias_list = alias_list->next;
5536 err = sysfs_slab_alias(al->s, al->name);
Christoph Lameter5d540fb2007-08-30 23:56:26 -07005537 if (err)
Fabian Frederickf9f58282014-06-04 16:06:34 -07005538 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5539 al->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07005540 kfree(al);
5541 }
5542
Christoph Lameter18004c52012-07-06 15:25:12 -05005543 mutex_unlock(&slab_mutex);
Christoph Lameter81819f02007-05-06 14:49:36 -07005544 resiliency_test();
5545 return 0;
5546}
5547
5548__initcall(slab_sysfs_init);
Christoph Lameterab4d5ed2010-10-05 13:57:26 -05005549#endif /* CONFIG_SYSFS */
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005550
5551/*
5552 * The /proc/slabinfo ABI
5553 */
Linus Torvalds158a9622008-01-02 13:04:48 -08005554#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04005555void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005556{
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005557 unsigned long nr_slabs = 0;
Christoph Lameter205ab992008-04-14 19:11:40 +03005558 unsigned long nr_objs = 0;
5559 unsigned long nr_free = 0;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005560 int node;
Christoph Lameterfa45dc22014-08-06 16:04:09 -07005561 struct kmem_cache_node *n;
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005562
Christoph Lameterfa45dc22014-08-06 16:04:09 -07005563 for_each_kmem_cache_node(s, node, n) {
Wanpeng Lic17fd132013-07-04 08:33:26 +08005564 nr_slabs += node_nr_slabs(n);
5565 nr_objs += node_nr_objs(n);
Christoph Lameter205ab992008-04-14 19:11:40 +03005566 nr_free += count_partial(n, count_free);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005567 }
5568
Glauber Costa0d7561c2012-10-19 18:20:27 +04005569 sinfo->active_objs = nr_objs - nr_free;
5570 sinfo->num_objs = nr_objs;
5571 sinfo->active_slabs = nr_slabs;
5572 sinfo->num_slabs = nr_slabs;
5573 sinfo->objects_per_slab = oo_objects(s->oo);
5574 sinfo->cache_order = oo_order(s->oo);
Pekka J Enberg57ed3ed2008-01-01 17:23:28 +01005575}
5576
Glauber Costa0d7561c2012-10-19 18:20:27 +04005577void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005578{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005579}
5580
Glauber Costab7454ad2012-10-19 18:20:25 +04005581ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5582 size_t count, loff_t *ppos)
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005583{
Glauber Costab7454ad2012-10-19 18:20:25 +04005584 return -EIO;
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04005585}
Linus Torvalds158a9622008-01-02 13:04:48 -08005586#endif /* CONFIG_SLABINFO */