2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #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>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache
*s
)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s
);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier
;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr
; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
203 int cpu
; /* Was running on cpu */
204 int pid
; /* Pid context */
205 unsigned long when
; /* When did the operation occur */
208 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
211 static int sysfs_slab_add(struct kmem_cache
*);
212 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
213 static void sysfs_slab_remove(struct kmem_cache
*);
214 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
216 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
217 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
219 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
221 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
224 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
226 #ifdef CONFIG_SLUB_STATS
227 __this_cpu_inc(s
->cpu_slab
->stat
[si
]);
231 /********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
235 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
237 return s
->node
[node
];
240 /* Verify that a pointer has an address that is valid within a slab page */
241 static inline int check_valid_pointer(struct kmem_cache
*s
,
242 struct page
*page
, const void *object
)
249 base
= page_address(page
);
250 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
251 (object
- base
) % s
->size
) {
258 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
260 return *(void **)(object
+ s
->offset
);
263 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
265 prefetch(object
+ s
->offset
);
268 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
272 #ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p
, (void **)(object
+ s
->offset
), sizeof(p
));
275 p
= get_freepointer(s
, object
);
280 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
282 *(void **)(object
+ s
->offset
) = fp
;
285 /* Loop over all objects in a slab */
286 #define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
293 return (p
- addr
) / s
->size
;
296 static inline size_t slab_ksize(const struct kmem_cache
*s
)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
304 return s
->object_size
;
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order
, unsigned long size
, int reserved
)
322 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
325 static inline struct kmem_cache_order_objects
oo_make(int order
,
326 unsigned long size
, int reserved
)
328 struct kmem_cache_order_objects x
= {
329 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
335 static inline int oo_order(struct kmem_cache_order_objects x
)
337 return x
.x
>> OO_SHIFT
;
340 static inline int oo_objects(struct kmem_cache_order_objects x
)
342 return x
.x
& OO_MASK
;
346 * Per slab locking using the pagelock
348 static __always_inline
void slab_lock(struct page
*page
)
350 bit_spin_lock(PG_locked
, &page
->flags
);
353 static __always_inline
void slab_unlock(struct page
*page
)
355 __bit_spin_unlock(PG_locked
, &page
->flags
);
358 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
361 tmp
.counters
= counters_new
;
363 * page->counters can cover frozen/inuse/objects as well
364 * as page->_count. If we assign to ->counters directly
365 * we run the risk of losing updates to page->_count, so
366 * be careful and only assign to the fields we need.
368 page
->frozen
= tmp
.frozen
;
369 page
->inuse
= tmp
.inuse
;
370 page
->objects
= tmp
.objects
;
373 /* Interrupts must be disabled (for the fallback code to work right) */
374 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
375 void *freelist_old
, unsigned long counters_old
,
376 void *freelist_new
, unsigned long counters_new
,
379 VM_BUG_ON(!irqs_disabled());
380 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
382 if (s
->flags
& __CMPXCHG_DOUBLE
) {
383 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
384 freelist_old
, counters_old
,
385 freelist_new
, counters_new
))
391 if (page
->freelist
== freelist_old
&&
392 page
->counters
== counters_old
) {
393 page
->freelist
= freelist_new
;
394 set_page_slub_counters(page
, counters_new
);
402 stat(s
, CMPXCHG_DOUBLE_FAIL
);
404 #ifdef SLUB_DEBUG_CMPXCHG
405 printk(KERN_INFO
"%s %s: cmpxchg double redo ", n
, s
->name
);
411 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
412 void *freelist_old
, unsigned long counters_old
,
413 void *freelist_new
, unsigned long counters_new
,
416 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
417 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
418 if (s
->flags
& __CMPXCHG_DOUBLE
) {
419 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
420 freelist_old
, counters_old
,
421 freelist_new
, counters_new
))
428 local_irq_save(flags
);
430 if (page
->freelist
== freelist_old
&&
431 page
->counters
== counters_old
) {
432 page
->freelist
= freelist_new
;
433 set_page_slub_counters(page
, counters_new
);
435 local_irq_restore(flags
);
439 local_irq_restore(flags
);
443 stat(s
, CMPXCHG_DOUBLE_FAIL
);
445 #ifdef SLUB_DEBUG_CMPXCHG
446 printk(KERN_INFO
"%s %s: cmpxchg double redo ", n
, s
->name
);
452 #ifdef CONFIG_SLUB_DEBUG
454 * Determine a map of object in use on a page.
456 * Node listlock must be held to guarantee that the page does
457 * not vanish from under us.
459 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
462 void *addr
= page_address(page
);
464 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
465 set_bit(slab_index(p
, s
, addr
), map
);
471 #ifdef CONFIG_SLUB_DEBUG_ON
472 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
474 static int slub_debug
;
477 static char *slub_debug_slabs
;
478 static int disable_higher_order_debug
;
483 static void print_section(char *text
, u8
*addr
, unsigned int length
)
485 print_hex_dump(KERN_ERR
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
489 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
490 enum track_item alloc
)
495 p
= object
+ s
->offset
+ sizeof(void *);
497 p
= object
+ s
->inuse
;
502 static void set_track(struct kmem_cache
*s
, void *object
,
503 enum track_item alloc
, unsigned long addr
)
505 struct track
*p
= get_track(s
, object
, alloc
);
508 #ifdef CONFIG_STACKTRACE
509 struct stack_trace trace
;
512 trace
.nr_entries
= 0;
513 trace
.max_entries
= TRACK_ADDRS_COUNT
;
514 trace
.entries
= p
->addrs
;
516 save_stack_trace(&trace
);
518 /* See rant in lockdep.c */
519 if (trace
.nr_entries
!= 0 &&
520 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
523 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
527 p
->cpu
= smp_processor_id();
528 p
->pid
= current
->pid
;
531 memset(p
, 0, sizeof(struct track
));
534 static void init_tracking(struct kmem_cache
*s
, void *object
)
536 if (!(s
->flags
& SLAB_STORE_USER
))
539 set_track(s
, object
, TRACK_FREE
, 0UL);
540 set_track(s
, object
, TRACK_ALLOC
, 0UL);
543 static void print_track(const char *s
, struct track
*t
)
548 printk(KERN_ERR
"INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
549 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
550 #ifdef CONFIG_STACKTRACE
553 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
555 printk(KERN_ERR
"\t%pS\n", (void *)t
->addrs
[i
]);
562 static void print_tracking(struct kmem_cache
*s
, void *object
)
564 if (!(s
->flags
& SLAB_STORE_USER
))
567 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
568 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
571 static void print_page_info(struct page
*page
)
574 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
575 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
579 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
585 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
587 printk(KERN_ERR
"========================================"
588 "=====================================\n");
589 printk(KERN_ERR
"BUG %s (%s): %s\n", s
->name
, print_tainted(), buf
);
590 printk(KERN_ERR
"----------------------------------------"
591 "-------------------------------------\n\n");
593 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
596 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
602 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
604 printk(KERN_ERR
"FIX %s: %s\n", s
->name
, buf
);
607 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
609 unsigned int off
; /* Offset of last byte */
610 u8
*addr
= page_address(page
);
612 print_tracking(s
, p
);
614 print_page_info(page
);
616 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
617 p
, p
- addr
, get_freepointer(s
, p
));
620 print_section("Bytes b4 ", p
- 16, 16);
622 print_section("Object ", p
, min_t(unsigned long, s
->object_size
,
624 if (s
->flags
& SLAB_RED_ZONE
)
625 print_section("Redzone ", p
+ s
->object_size
,
626 s
->inuse
- s
->object_size
);
629 off
= s
->offset
+ sizeof(void *);
633 if (s
->flags
& SLAB_STORE_USER
)
634 off
+= 2 * sizeof(struct track
);
637 /* Beginning of the filler is the free pointer */
638 print_section("Padding ", p
+ off
, s
->size
- off
);
643 static void object_err(struct kmem_cache
*s
, struct page
*page
,
644 u8
*object
, char *reason
)
646 slab_bug(s
, "%s", reason
);
647 print_trailer(s
, page
, object
);
650 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
651 const char *fmt
, ...)
657 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
659 slab_bug(s
, "%s", buf
);
660 print_page_info(page
);
664 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
668 if (s
->flags
& __OBJECT_POISON
) {
669 memset(p
, POISON_FREE
, s
->object_size
- 1);
670 p
[s
->object_size
- 1] = POISON_END
;
673 if (s
->flags
& SLAB_RED_ZONE
)
674 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
677 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
678 void *from
, void *to
)
680 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
681 memset(from
, data
, to
- from
);
684 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
685 u8
*object
, char *what
,
686 u8
*start
, unsigned int value
, unsigned int bytes
)
691 fault
= memchr_inv(start
, value
, bytes
);
696 while (end
> fault
&& end
[-1] == value
)
699 slab_bug(s
, "%s overwritten", what
);
700 printk(KERN_ERR
"INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
701 fault
, end
- 1, fault
[0], value
);
702 print_trailer(s
, page
, object
);
704 restore_bytes(s
, what
, value
, fault
, end
);
712 * Bytes of the object to be managed.
713 * If the freepointer may overlay the object then the free
714 * pointer is the first word of the object.
716 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
719 * object + s->object_size
720 * Padding to reach word boundary. This is also used for Redzoning.
721 * Padding is extended by another word if Redzoning is enabled and
722 * object_size == inuse.
724 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
725 * 0xcc (RED_ACTIVE) for objects in use.
728 * Meta data starts here.
730 * A. Free pointer (if we cannot overwrite object on free)
731 * B. Tracking data for SLAB_STORE_USER
732 * C. Padding to reach required alignment boundary or at mininum
733 * one word if debugging is on to be able to detect writes
734 * before the word boundary.
736 * Padding is done using 0x5a (POISON_INUSE)
739 * Nothing is used beyond s->size.
741 * If slabcaches are merged then the object_size and inuse boundaries are mostly
742 * ignored. And therefore no slab options that rely on these boundaries
743 * may be used with merged slabcaches.
746 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
748 unsigned long off
= s
->inuse
; /* The end of info */
751 /* Freepointer is placed after the object. */
752 off
+= sizeof(void *);
754 if (s
->flags
& SLAB_STORE_USER
)
755 /* We also have user information there */
756 off
+= 2 * sizeof(struct track
);
761 return check_bytes_and_report(s
, page
, p
, "Object padding",
762 p
+ off
, POISON_INUSE
, s
->size
- off
);
765 /* Check the pad bytes at the end of a slab page */
766 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
774 if (!(s
->flags
& SLAB_POISON
))
777 start
= page_address(page
);
778 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
779 end
= start
+ length
;
780 remainder
= length
% s
->size
;
784 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
787 while (end
> fault
&& end
[-1] == POISON_INUSE
)
790 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
791 print_section("Padding ", end
- remainder
, remainder
);
793 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
797 static int check_object(struct kmem_cache
*s
, struct page
*page
,
798 void *object
, u8 val
)
801 u8
*endobject
= object
+ s
->object_size
;
803 if (s
->flags
& SLAB_RED_ZONE
) {
804 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
805 endobject
, val
, s
->inuse
- s
->object_size
))
808 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
809 check_bytes_and_report(s
, page
, p
, "Alignment padding",
810 endobject
, POISON_INUSE
,
811 s
->inuse
- s
->object_size
);
815 if (s
->flags
& SLAB_POISON
) {
816 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
817 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
818 POISON_FREE
, s
->object_size
- 1) ||
819 !check_bytes_and_report(s
, page
, p
, "Poison",
820 p
+ s
->object_size
- 1, POISON_END
, 1)))
823 * check_pad_bytes cleans up on its own.
825 check_pad_bytes(s
, page
, p
);
828 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
830 * Object and freepointer overlap. Cannot check
831 * freepointer while object is allocated.
835 /* Check free pointer validity */
836 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
837 object_err(s
, page
, p
, "Freepointer corrupt");
839 * No choice but to zap it and thus lose the remainder
840 * of the free objects in this slab. May cause
841 * another error because the object count is now wrong.
843 set_freepointer(s
, p
, NULL
);
849 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
853 VM_BUG_ON(!irqs_disabled());
855 if (!PageSlab(page
)) {
856 slab_err(s
, page
, "Not a valid slab page");
860 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
861 if (page
->objects
> maxobj
) {
862 slab_err(s
, page
, "objects %u > max %u",
863 s
->name
, page
->objects
, maxobj
);
866 if (page
->inuse
> page
->objects
) {
867 slab_err(s
, page
, "inuse %u > max %u",
868 s
->name
, page
->inuse
, page
->objects
);
871 /* Slab_pad_check fixes things up after itself */
872 slab_pad_check(s
, page
);
877 * Determine if a certain object on a page is on the freelist. Must hold the
878 * slab lock to guarantee that the chains are in a consistent state.
880 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
885 unsigned long max_objects
;
888 while (fp
&& nr
<= page
->objects
) {
891 if (!check_valid_pointer(s
, page
, fp
)) {
893 object_err(s
, page
, object
,
894 "Freechain corrupt");
895 set_freepointer(s
, object
, NULL
);
897 slab_err(s
, page
, "Freepointer corrupt");
898 page
->freelist
= NULL
;
899 page
->inuse
= page
->objects
;
900 slab_fix(s
, "Freelist cleared");
906 fp
= get_freepointer(s
, object
);
910 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
911 if (max_objects
> MAX_OBJS_PER_PAGE
)
912 max_objects
= MAX_OBJS_PER_PAGE
;
914 if (page
->objects
!= max_objects
) {
915 slab_err(s
, page
, "Wrong number of objects. Found %d but "
916 "should be %d", page
->objects
, max_objects
);
917 page
->objects
= max_objects
;
918 slab_fix(s
, "Number of objects adjusted.");
920 if (page
->inuse
!= page
->objects
- nr
) {
921 slab_err(s
, page
, "Wrong object count. Counter is %d but "
922 "counted were %d", page
->inuse
, page
->objects
- nr
);
923 page
->inuse
= page
->objects
- nr
;
924 slab_fix(s
, "Object count adjusted.");
926 return search
== NULL
;
929 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
932 if (s
->flags
& SLAB_TRACE
) {
933 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
935 alloc
? "alloc" : "free",
940 print_section("Object ", (void *)object
,
948 * Hooks for other subsystems that check memory allocations. In a typical
949 * production configuration these hooks all should produce no code at all.
951 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
953 kmemleak_alloc(ptr
, size
, 1, flags
);
956 static inline void kfree_hook(const void *x
)
961 static inline int slab_pre_alloc_hook(struct kmem_cache
*s
, gfp_t flags
)
963 flags
&= gfp_allowed_mask
;
964 lockdep_trace_alloc(flags
);
965 might_sleep_if(flags
& __GFP_WAIT
);
967 return should_failslab(s
->object_size
, flags
, s
->flags
);
970 static inline void slab_post_alloc_hook(struct kmem_cache
*s
,
971 gfp_t flags
, void *object
)
973 flags
&= gfp_allowed_mask
;
974 kmemcheck_slab_alloc(s
, flags
, object
, slab_ksize(s
));
975 kmemleak_alloc_recursive(object
, s
->object_size
, 1, s
->flags
, flags
);
978 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
980 kmemleak_free_recursive(x
, s
->flags
);
983 * Trouble is that we may no longer disable interrupts in the fast path
984 * So in order to make the debug calls that expect irqs to be
985 * disabled we need to disable interrupts temporarily.
987 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
991 local_irq_save(flags
);
992 kmemcheck_slab_free(s
, x
, s
->object_size
);
993 debug_check_no_locks_freed(x
, s
->object_size
);
994 local_irq_restore(flags
);
997 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
998 debug_check_no_obj_freed(x
, s
->object_size
);
1002 * Tracking of fully allocated slabs for debugging purposes.
1004 static void add_full(struct kmem_cache
*s
,
1005 struct kmem_cache_node
*n
, struct page
*page
)
1007 if (!(s
->flags
& SLAB_STORE_USER
))
1010 lockdep_assert_held(&n
->list_lock
);
1011 list_add(&page
->lru
, &n
->full
);
1014 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1016 if (!(s
->flags
& SLAB_STORE_USER
))
1019 lockdep_assert_held(&n
->list_lock
);
1020 list_del(&page
->lru
);
1023 /* Tracking of the number of slabs for debugging purposes */
1024 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1026 struct kmem_cache_node
*n
= get_node(s
, node
);
1028 return atomic_long_read(&n
->nr_slabs
);
1031 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1033 return atomic_long_read(&n
->nr_slabs
);
1036 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1038 struct kmem_cache_node
*n
= get_node(s
, node
);
1041 * May be called early in order to allocate a slab for the
1042 * kmem_cache_node structure. Solve the chicken-egg
1043 * dilemma by deferring the increment of the count during
1044 * bootstrap (see early_kmem_cache_node_alloc).
1047 atomic_long_inc(&n
->nr_slabs
);
1048 atomic_long_add(objects
, &n
->total_objects
);
1051 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1053 struct kmem_cache_node
*n
= get_node(s
, node
);
1055 atomic_long_dec(&n
->nr_slabs
);
1056 atomic_long_sub(objects
, &n
->total_objects
);
1059 /* Object debug checks for alloc/free paths */
1060 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1063 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1066 init_object(s
, object
, SLUB_RED_INACTIVE
);
1067 init_tracking(s
, object
);
1070 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1072 void *object
, unsigned long addr
)
1074 if (!check_slab(s
, page
))
1077 if (!check_valid_pointer(s
, page
, object
)) {
1078 object_err(s
, page
, object
, "Freelist Pointer check fails");
1082 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1085 /* Success perform special debug activities for allocs */
1086 if (s
->flags
& SLAB_STORE_USER
)
1087 set_track(s
, object
, TRACK_ALLOC
, addr
);
1088 trace(s
, page
, object
, 1);
1089 init_object(s
, object
, SLUB_RED_ACTIVE
);
1093 if (PageSlab(page
)) {
1095 * If this is a slab page then lets do the best we can
1096 * to avoid issues in the future. Marking all objects
1097 * as used avoids touching the remaining objects.
1099 slab_fix(s
, "Marking all objects used");
1100 page
->inuse
= page
->objects
;
1101 page
->freelist
= NULL
;
1106 static noinline
struct kmem_cache_node
*free_debug_processing(
1107 struct kmem_cache
*s
, struct page
*page
, void *object
,
1108 unsigned long addr
, unsigned long *flags
)
1110 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1112 spin_lock_irqsave(&n
->list_lock
, *flags
);
1115 if (!check_slab(s
, page
))
1118 if (!check_valid_pointer(s
, page
, object
)) {
1119 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1123 if (on_freelist(s
, page
, object
)) {
1124 object_err(s
, page
, object
, "Object already free");
1128 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1131 if (unlikely(s
!= page
->slab_cache
)) {
1132 if (!PageSlab(page
)) {
1133 slab_err(s
, page
, "Attempt to free object(0x%p) "
1134 "outside of slab", object
);
1135 } else if (!page
->slab_cache
) {
1137 "SLUB <none>: no slab for object 0x%p.\n",
1141 object_err(s
, page
, object
,
1142 "page slab pointer corrupt.");
1146 if (s
->flags
& SLAB_STORE_USER
)
1147 set_track(s
, object
, TRACK_FREE
, addr
);
1148 trace(s
, page
, object
, 0);
1149 init_object(s
, object
, SLUB_RED_INACTIVE
);
1153 * Keep node_lock to preserve integrity
1154 * until the object is actually freed
1160 spin_unlock_irqrestore(&n
->list_lock
, *flags
);
1161 slab_fix(s
, "Object at 0x%p not freed", object
);
1165 static int __init
setup_slub_debug(char *str
)
1167 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1168 if (*str
++ != '=' || !*str
)
1170 * No options specified. Switch on full debugging.
1176 * No options but restriction on slabs. This means full
1177 * debugging for slabs matching a pattern.
1181 if (tolower(*str
) == 'o') {
1183 * Avoid enabling debugging on caches if its minimum order
1184 * would increase as a result.
1186 disable_higher_order_debug
= 1;
1193 * Switch off all debugging measures.
1198 * Determine which debug features should be switched on
1200 for (; *str
&& *str
!= ','; str
++) {
1201 switch (tolower(*str
)) {
1203 slub_debug
|= SLAB_DEBUG_FREE
;
1206 slub_debug
|= SLAB_RED_ZONE
;
1209 slub_debug
|= SLAB_POISON
;
1212 slub_debug
|= SLAB_STORE_USER
;
1215 slub_debug
|= SLAB_TRACE
;
1218 slub_debug
|= SLAB_FAILSLAB
;
1221 printk(KERN_ERR
"slub_debug option '%c' "
1222 "unknown. skipped\n", *str
);
1228 slub_debug_slabs
= str
+ 1;
1233 __setup("slub_debug", setup_slub_debug
);
1235 static unsigned long kmem_cache_flags(unsigned long object_size
,
1236 unsigned long flags
, const char *name
,
1237 void (*ctor
)(void *))
1240 * Enable debugging if selected on the kernel commandline.
1242 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1243 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1244 flags
|= slub_debug
;
1249 static inline void setup_object_debug(struct kmem_cache
*s
,
1250 struct page
*page
, void *object
) {}
1252 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1253 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1255 static inline struct kmem_cache_node
*free_debug_processing(
1256 struct kmem_cache
*s
, struct page
*page
, void *object
,
1257 unsigned long addr
, unsigned long *flags
) { return NULL
; }
1259 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1261 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1262 void *object
, u8 val
) { return 1; }
1263 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1264 struct page
*page
) {}
1265 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1266 struct page
*page
) {}
1267 static inline unsigned long kmem_cache_flags(unsigned long object_size
,
1268 unsigned long flags
, const char *name
,
1269 void (*ctor
)(void *))
1273 #define slub_debug 0
1275 #define disable_higher_order_debug 0
1277 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1279 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1281 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1283 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1286 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1288 kmemleak_alloc(ptr
, size
, 1, flags
);
1291 static inline void kfree_hook(const void *x
)
1296 static inline int slab_pre_alloc_hook(struct kmem_cache
*s
, gfp_t flags
)
1299 static inline void slab_post_alloc_hook(struct kmem_cache
*s
, gfp_t flags
,
1302 kmemleak_alloc_recursive(object
, s
->object_size
, 1, s
->flags
,
1303 flags
& gfp_allowed_mask
);
1306 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
1308 kmemleak_free_recursive(x
, s
->flags
);
1311 #endif /* CONFIG_SLUB_DEBUG */
1314 * Slab allocation and freeing
1316 static inline struct page
*alloc_slab_page(gfp_t flags
, int node
,
1317 struct kmem_cache_order_objects oo
)
1319 int order
= oo_order(oo
);
1321 flags
|= __GFP_NOTRACK
;
1323 if (node
== NUMA_NO_NODE
)
1324 return alloc_pages(flags
, order
);
1326 return alloc_pages_exact_node(node
, flags
, order
);
1329 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1332 struct kmem_cache_order_objects oo
= s
->oo
;
1335 flags
&= gfp_allowed_mask
;
1337 if (flags
& __GFP_WAIT
)
1340 flags
|= s
->allocflags
;
1343 * Let the initial higher-order allocation fail under memory pressure
1344 * so we fall-back to the minimum order allocation.
1346 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1348 page
= alloc_slab_page(alloc_gfp
, node
, oo
);
1349 if (unlikely(!page
)) {
1352 * Allocation may have failed due to fragmentation.
1353 * Try a lower order alloc if possible
1355 page
= alloc_slab_page(flags
, node
, oo
);
1358 stat(s
, ORDER_FALLBACK
);
1361 if (kmemcheck_enabled
&& page
1362 && !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1363 int pages
= 1 << oo_order(oo
);
1365 kmemcheck_alloc_shadow(page
, oo_order(oo
), flags
, node
);
1368 * Objects from caches that have a constructor don't get
1369 * cleared when they're allocated, so we need to do it here.
1372 kmemcheck_mark_uninitialized_pages(page
, pages
);
1374 kmemcheck_mark_unallocated_pages(page
, pages
);
1377 if (flags
& __GFP_WAIT
)
1378 local_irq_disable();
1382 page
->objects
= oo_objects(oo
);
1383 mod_zone_page_state(page_zone(page
),
1384 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1385 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1391 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1394 setup_object_debug(s
, page
, object
);
1395 if (unlikely(s
->ctor
))
1399 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1407 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1409 page
= allocate_slab(s
,
1410 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1414 order
= compound_order(page
);
1415 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1416 memcg_bind_pages(s
, order
);
1417 page
->slab_cache
= s
;
1418 __SetPageSlab(page
);
1419 if (page
->pfmemalloc
)
1420 SetPageSlabPfmemalloc(page
);
1422 start
= page_address(page
);
1424 if (unlikely(s
->flags
& SLAB_POISON
))
1425 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1428 for_each_object(p
, s
, start
, page
->objects
) {
1429 setup_object(s
, page
, last
);
1430 set_freepointer(s
, last
, p
);
1433 setup_object(s
, page
, last
);
1434 set_freepointer(s
, last
, NULL
);
1436 page
->freelist
= start
;
1437 page
->inuse
= page
->objects
;
1443 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1445 int order
= compound_order(page
);
1446 int pages
= 1 << order
;
1448 if (kmem_cache_debug(s
)) {
1451 slab_pad_check(s
, page
);
1452 for_each_object(p
, s
, page_address(page
),
1454 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1457 kmemcheck_free_shadow(page
, compound_order(page
));
1459 mod_zone_page_state(page_zone(page
),
1460 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1461 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1464 __ClearPageSlabPfmemalloc(page
);
1465 __ClearPageSlab(page
);
1467 memcg_release_pages(s
, order
);
1468 page_mapcount_reset(page
);
1469 if (current
->reclaim_state
)
1470 current
->reclaim_state
->reclaimed_slab
+= pages
;
1471 __free_memcg_kmem_pages(page
, order
);
1474 #define need_reserve_slab_rcu \
1475 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1477 static void rcu_free_slab(struct rcu_head
*h
)
1481 if (need_reserve_slab_rcu
)
1482 page
= virt_to_head_page(h
);
1484 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1486 __free_slab(page
->slab_cache
, page
);
1489 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1491 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1492 struct rcu_head
*head
;
1494 if (need_reserve_slab_rcu
) {
1495 int order
= compound_order(page
);
1496 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1498 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1499 head
= page_address(page
) + offset
;
1502 * RCU free overloads the RCU head over the LRU
1504 head
= (void *)&page
->lru
;
1507 call_rcu(head
, rcu_free_slab
);
1509 __free_slab(s
, page
);
1512 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1514 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1519 * Management of partially allocated slabs.
1522 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1525 if (tail
== DEACTIVATE_TO_TAIL
)
1526 list_add_tail(&page
->lru
, &n
->partial
);
1528 list_add(&page
->lru
, &n
->partial
);
1531 static inline void add_partial(struct kmem_cache_node
*n
,
1532 struct page
*page
, int tail
)
1534 lockdep_assert_held(&n
->list_lock
);
1535 __add_partial(n
, page
, tail
);
1539 __remove_partial(struct kmem_cache_node
*n
, struct page
*page
)
1541 list_del(&page
->lru
);
1545 static inline void remove_partial(struct kmem_cache_node
*n
,
1548 lockdep_assert_held(&n
->list_lock
);
1549 __remove_partial(n
, page
);
1553 * Remove slab from the partial list, freeze it and
1554 * return the pointer to the freelist.
1556 * Returns a list of objects or NULL if it fails.
1558 static inline void *acquire_slab(struct kmem_cache
*s
,
1559 struct kmem_cache_node
*n
, struct page
*page
,
1560 int mode
, int *objects
)
1563 unsigned long counters
;
1566 lockdep_assert_held(&n
->list_lock
);
1569 * Zap the freelist and set the frozen bit.
1570 * The old freelist is the list of objects for the
1571 * per cpu allocation list.
1573 freelist
= page
->freelist
;
1574 counters
= page
->counters
;
1575 new.counters
= counters
;
1576 *objects
= new.objects
- new.inuse
;
1578 new.inuse
= page
->objects
;
1579 new.freelist
= NULL
;
1581 new.freelist
= freelist
;
1584 VM_BUG_ON(new.frozen
);
1587 if (!__cmpxchg_double_slab(s
, page
,
1589 new.freelist
, new.counters
,
1593 remove_partial(n
, page
);
1598 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1599 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1602 * Try to allocate a partial slab from a specific node.
1604 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1605 struct kmem_cache_cpu
*c
, gfp_t flags
)
1607 struct page
*page
, *page2
;
1608 void *object
= NULL
;
1613 * Racy check. If we mistakenly see no partial slabs then we
1614 * just allocate an empty slab. If we mistakenly try to get a
1615 * partial slab and there is none available then get_partials()
1618 if (!n
|| !n
->nr_partial
)
1621 spin_lock(&n
->list_lock
);
1622 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1625 if (!pfmemalloc_match(page
, flags
))
1628 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1632 available
+= objects
;
1635 stat(s
, ALLOC_FROM_PARTIAL
);
1638 put_cpu_partial(s
, page
, 0);
1639 stat(s
, CPU_PARTIAL_NODE
);
1641 if (!kmem_cache_has_cpu_partial(s
)
1642 || available
> s
->cpu_partial
/ 2)
1646 spin_unlock(&n
->list_lock
);
1651 * Get a page from somewhere. Search in increasing NUMA distances.
1653 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1654 struct kmem_cache_cpu
*c
)
1657 struct zonelist
*zonelist
;
1660 enum zone_type high_zoneidx
= gfp_zone(flags
);
1662 unsigned int cpuset_mems_cookie
;
1665 * The defrag ratio allows a configuration of the tradeoffs between
1666 * inter node defragmentation and node local allocations. A lower
1667 * defrag_ratio increases the tendency to do local allocations
1668 * instead of attempting to obtain partial slabs from other nodes.
1670 * If the defrag_ratio is set to 0 then kmalloc() always
1671 * returns node local objects. If the ratio is higher then kmalloc()
1672 * may return off node objects because partial slabs are obtained
1673 * from other nodes and filled up.
1675 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1676 * defrag_ratio = 1000) then every (well almost) allocation will
1677 * first attempt to defrag slab caches on other nodes. This means
1678 * scanning over all nodes to look for partial slabs which may be
1679 * expensive if we do it every time we are trying to find a slab
1680 * with available objects.
1682 if (!s
->remote_node_defrag_ratio
||
1683 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1687 cpuset_mems_cookie
= read_mems_allowed_begin();
1688 zonelist
= node_zonelist(slab_node(), flags
);
1689 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1690 struct kmem_cache_node
*n
;
1692 n
= get_node(s
, zone_to_nid(zone
));
1694 if (n
&& cpuset_zone_allowed_hardwall(zone
, flags
) &&
1695 n
->nr_partial
> s
->min_partial
) {
1696 object
= get_partial_node(s
, n
, c
, flags
);
1699 * Don't check read_mems_allowed_retry()
1700 * here - if mems_allowed was updated in
1701 * parallel, that was a harmless race
1702 * between allocation and the cpuset
1709 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1715 * Get a partial page, lock it and return it.
1717 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1718 struct kmem_cache_cpu
*c
)
1721 int searchnode
= (node
== NUMA_NO_NODE
) ? numa_node_id() : node
;
1723 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1724 if (object
|| node
!= NUMA_NO_NODE
)
1727 return get_any_partial(s
, flags
, c
);
1730 #ifdef CONFIG_PREEMPT
1732 * Calculate the next globally unique transaction for disambiguiation
1733 * during cmpxchg. The transactions start with the cpu number and are then
1734 * incremented by CONFIG_NR_CPUS.
1736 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1739 * No preemption supported therefore also no need to check for
1745 static inline unsigned long next_tid(unsigned long tid
)
1747 return tid
+ TID_STEP
;
1750 static inline unsigned int tid_to_cpu(unsigned long tid
)
1752 return tid
% TID_STEP
;
1755 static inline unsigned long tid_to_event(unsigned long tid
)
1757 return tid
/ TID_STEP
;
1760 static inline unsigned int init_tid(int cpu
)
1765 static inline void note_cmpxchg_failure(const char *n
,
1766 const struct kmem_cache
*s
, unsigned long tid
)
1768 #ifdef SLUB_DEBUG_CMPXCHG
1769 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1771 printk(KERN_INFO
"%s %s: cmpxchg redo ", n
, s
->name
);
1773 #ifdef CONFIG_PREEMPT
1774 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1775 printk("due to cpu change %d -> %d\n",
1776 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1779 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1780 printk("due to cpu running other code. Event %ld->%ld\n",
1781 tid_to_event(tid
), tid_to_event(actual_tid
));
1783 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1784 actual_tid
, tid
, next_tid(tid
));
1786 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1789 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1793 for_each_possible_cpu(cpu
)
1794 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1798 * Remove the cpu slab
1800 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1803 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1804 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1806 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1808 int tail
= DEACTIVATE_TO_HEAD
;
1812 if (page
->freelist
) {
1813 stat(s
, DEACTIVATE_REMOTE_FREES
);
1814 tail
= DEACTIVATE_TO_TAIL
;
1818 * Stage one: Free all available per cpu objects back
1819 * to the page freelist while it is still frozen. Leave the
1822 * There is no need to take the list->lock because the page
1825 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
1827 unsigned long counters
;
1830 prior
= page
->freelist
;
1831 counters
= page
->counters
;
1832 set_freepointer(s
, freelist
, prior
);
1833 new.counters
= counters
;
1835 VM_BUG_ON(!new.frozen
);
1837 } while (!__cmpxchg_double_slab(s
, page
,
1839 freelist
, new.counters
,
1840 "drain percpu freelist"));
1842 freelist
= nextfree
;
1846 * Stage two: Ensure that the page is unfrozen while the
1847 * list presence reflects the actual number of objects
1850 * We setup the list membership and then perform a cmpxchg
1851 * with the count. If there is a mismatch then the page
1852 * is not unfrozen but the page is on the wrong list.
1854 * Then we restart the process which may have to remove
1855 * the page from the list that we just put it on again
1856 * because the number of objects in the slab may have
1861 old
.freelist
= page
->freelist
;
1862 old
.counters
= page
->counters
;
1863 VM_BUG_ON(!old
.frozen
);
1865 /* Determine target state of the slab */
1866 new.counters
= old
.counters
;
1869 set_freepointer(s
, freelist
, old
.freelist
);
1870 new.freelist
= freelist
;
1872 new.freelist
= old
.freelist
;
1876 if (!new.inuse
&& n
->nr_partial
> s
->min_partial
)
1878 else if (new.freelist
) {
1883 * Taking the spinlock removes the possiblity
1884 * that acquire_slab() will see a slab page that
1887 spin_lock(&n
->list_lock
);
1891 if (kmem_cache_debug(s
) && !lock
) {
1894 * This also ensures that the scanning of full
1895 * slabs from diagnostic functions will not see
1898 spin_lock(&n
->list_lock
);
1906 remove_partial(n
, page
);
1908 else if (l
== M_FULL
)
1910 remove_full(s
, n
, page
);
1912 if (m
== M_PARTIAL
) {
1914 add_partial(n
, page
, tail
);
1917 } else if (m
== M_FULL
) {
1919 stat(s
, DEACTIVATE_FULL
);
1920 add_full(s
, n
, page
);
1926 if (!__cmpxchg_double_slab(s
, page
,
1927 old
.freelist
, old
.counters
,
1928 new.freelist
, new.counters
,
1933 spin_unlock(&n
->list_lock
);
1936 stat(s
, DEACTIVATE_EMPTY
);
1937 discard_slab(s
, page
);
1943 * Unfreeze all the cpu partial slabs.
1945 * This function must be called with interrupts disabled
1946 * for the cpu using c (or some other guarantee must be there
1947 * to guarantee no concurrent accesses).
1949 static void unfreeze_partials(struct kmem_cache
*s
,
1950 struct kmem_cache_cpu
*c
)
1952 #ifdef CONFIG_SLUB_CPU_PARTIAL
1953 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
1954 struct page
*page
, *discard_page
= NULL
;
1956 while ((page
= c
->partial
)) {
1960 c
->partial
= page
->next
;
1962 n2
= get_node(s
, page_to_nid(page
));
1965 spin_unlock(&n
->list_lock
);
1968 spin_lock(&n
->list_lock
);
1973 old
.freelist
= page
->freelist
;
1974 old
.counters
= page
->counters
;
1975 VM_BUG_ON(!old
.frozen
);
1977 new.counters
= old
.counters
;
1978 new.freelist
= old
.freelist
;
1982 } while (!__cmpxchg_double_slab(s
, page
,
1983 old
.freelist
, old
.counters
,
1984 new.freelist
, new.counters
,
1985 "unfreezing slab"));
1987 if (unlikely(!new.inuse
&& n
->nr_partial
> s
->min_partial
)) {
1988 page
->next
= discard_page
;
1989 discard_page
= page
;
1991 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
1992 stat(s
, FREE_ADD_PARTIAL
);
1997 spin_unlock(&n
->list_lock
);
1999 while (discard_page
) {
2000 page
= discard_page
;
2001 discard_page
= discard_page
->next
;
2003 stat(s
, DEACTIVATE_EMPTY
);
2004 discard_slab(s
, page
);
2011 * Put a page that was just frozen (in __slab_free) into a partial page
2012 * slot if available. This is done without interrupts disabled and without
2013 * preemption disabled. The cmpxchg is racy and may put the partial page
2014 * onto a random cpus partial slot.
2016 * If we did not find a slot then simply move all the partials to the
2017 * per node partial list.
2019 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2021 #ifdef CONFIG_SLUB_CPU_PARTIAL
2022 struct page
*oldpage
;
2029 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2032 pobjects
= oldpage
->pobjects
;
2033 pages
= oldpage
->pages
;
2034 if (drain
&& pobjects
> s
->cpu_partial
) {
2035 unsigned long flags
;
2037 * partial array is full. Move the existing
2038 * set to the per node partial list.
2040 local_irq_save(flags
);
2041 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2042 local_irq_restore(flags
);
2046 stat(s
, CPU_PARTIAL_DRAIN
);
2051 pobjects
+= page
->objects
- page
->inuse
;
2053 page
->pages
= pages
;
2054 page
->pobjects
= pobjects
;
2055 page
->next
= oldpage
;
2057 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2062 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2064 stat(s
, CPUSLAB_FLUSH
);
2065 deactivate_slab(s
, c
->page
, c
->freelist
);
2067 c
->tid
= next_tid(c
->tid
);
2075 * Called from IPI handler with interrupts disabled.
2077 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2079 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2085 unfreeze_partials(s
, c
);
2089 static void flush_cpu_slab(void *d
)
2091 struct kmem_cache
*s
= d
;
2093 __flush_cpu_slab(s
, smp_processor_id());
2096 static bool has_cpu_slab(int cpu
, void *info
)
2098 struct kmem_cache
*s
= info
;
2099 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2101 return c
->page
|| c
->partial
;
2104 static void flush_all(struct kmem_cache
*s
)
2106 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2110 * Check if the objects in a per cpu structure fit numa
2111 * locality expectations.
2113 static inline int node_match(struct page
*page
, int node
)
2116 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2122 static int count_free(struct page
*page
)
2124 return page
->objects
- page
->inuse
;
2127 static unsigned long count_partial(struct kmem_cache_node
*n
,
2128 int (*get_count
)(struct page
*))
2130 unsigned long flags
;
2131 unsigned long x
= 0;
2134 spin_lock_irqsave(&n
->list_lock
, flags
);
2135 list_for_each_entry(page
, &n
->partial
, lru
)
2136 x
+= get_count(page
);
2137 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2141 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2143 #ifdef CONFIG_SLUB_DEBUG
2144 return atomic_long_read(&n
->total_objects
);
2150 static noinline
void
2151 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2156 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2158 printk(KERN_WARNING
" cache: %s, object size: %d, buffer size: %d, "
2159 "default order: %d, min order: %d\n", s
->name
, s
->object_size
,
2160 s
->size
, oo_order(s
->oo
), oo_order(s
->min
));
2162 if (oo_order(s
->min
) > get_order(s
->object_size
))
2163 printk(KERN_WARNING
" %s debugging increased min order, use "
2164 "slub_debug=O to disable.\n", s
->name
);
2166 for_each_online_node(node
) {
2167 struct kmem_cache_node
*n
= get_node(s
, node
);
2168 unsigned long nr_slabs
;
2169 unsigned long nr_objs
;
2170 unsigned long nr_free
;
2175 nr_free
= count_partial(n
, count_free
);
2176 nr_slabs
= node_nr_slabs(n
);
2177 nr_objs
= node_nr_objs(n
);
2180 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2181 node
, nr_slabs
, nr_objs
, nr_free
);
2185 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2186 int node
, struct kmem_cache_cpu
**pc
)
2189 struct kmem_cache_cpu
*c
= *pc
;
2192 freelist
= get_partial(s
, flags
, node
, c
);
2197 page
= new_slab(s
, flags
, node
);
2199 c
= __this_cpu_ptr(s
->cpu_slab
);
2204 * No other reference to the page yet so we can
2205 * muck around with it freely without cmpxchg
2207 freelist
= page
->freelist
;
2208 page
->freelist
= NULL
;
2210 stat(s
, ALLOC_SLAB
);
2219 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2221 if (unlikely(PageSlabPfmemalloc(page
)))
2222 return gfp_pfmemalloc_allowed(gfpflags
);
2228 * Check the page->freelist of a page and either transfer the freelist to the
2229 * per cpu freelist or deactivate the page.
2231 * The page is still frozen if the return value is not NULL.
2233 * If this function returns NULL then the page has been unfrozen.
2235 * This function must be called with interrupt disabled.
2237 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2240 unsigned long counters
;
2244 freelist
= page
->freelist
;
2245 counters
= page
->counters
;
2247 new.counters
= counters
;
2248 VM_BUG_ON(!new.frozen
);
2250 new.inuse
= page
->objects
;
2251 new.frozen
= freelist
!= NULL
;
2253 } while (!__cmpxchg_double_slab(s
, page
,
2262 * Slow path. The lockless freelist is empty or we need to perform
2265 * Processing is still very fast if new objects have been freed to the
2266 * regular freelist. In that case we simply take over the regular freelist
2267 * as the lockless freelist and zap the regular freelist.
2269 * If that is not working then we fall back to the partial lists. We take the
2270 * first element of the freelist as the object to allocate now and move the
2271 * rest of the freelist to the lockless freelist.
2273 * And if we were unable to get a new slab from the partial slab lists then
2274 * we need to allocate a new slab. This is the slowest path since it involves
2275 * a call to the page allocator and the setup of a new slab.
2277 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2278 unsigned long addr
, struct kmem_cache_cpu
*c
)
2282 unsigned long flags
;
2284 local_irq_save(flags
);
2285 #ifdef CONFIG_PREEMPT
2287 * We may have been preempted and rescheduled on a different
2288 * cpu before disabling interrupts. Need to reload cpu area
2291 c
= this_cpu_ptr(s
->cpu_slab
);
2299 if (unlikely(!node_match(page
, node
))) {
2300 stat(s
, ALLOC_NODE_MISMATCH
);
2301 deactivate_slab(s
, page
, c
->freelist
);
2308 * By rights, we should be searching for a slab page that was
2309 * PFMEMALLOC but right now, we are losing the pfmemalloc
2310 * information when the page leaves the per-cpu allocator
2312 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2313 deactivate_slab(s
, page
, c
->freelist
);
2319 /* must check again c->freelist in case of cpu migration or IRQ */
2320 freelist
= c
->freelist
;
2324 stat(s
, ALLOC_SLOWPATH
);
2326 freelist
= get_freelist(s
, page
);
2330 stat(s
, DEACTIVATE_BYPASS
);
2334 stat(s
, ALLOC_REFILL
);
2338 * freelist is pointing to the list of objects to be used.
2339 * page is pointing to the page from which the objects are obtained.
2340 * That page must be frozen for per cpu allocations to work.
2342 VM_BUG_ON(!c
->page
->frozen
);
2343 c
->freelist
= get_freepointer(s
, freelist
);
2344 c
->tid
= next_tid(c
->tid
);
2345 local_irq_restore(flags
);
2351 page
= c
->page
= c
->partial
;
2352 c
->partial
= page
->next
;
2353 stat(s
, CPU_PARTIAL_ALLOC
);
2358 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2360 if (unlikely(!freelist
)) {
2361 if (!(gfpflags
& __GFP_NOWARN
) && printk_ratelimit())
2362 slab_out_of_memory(s
, gfpflags
, node
);
2364 local_irq_restore(flags
);
2369 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2372 /* Only entered in the debug case */
2373 if (kmem_cache_debug(s
) &&
2374 !alloc_debug_processing(s
, page
, freelist
, addr
))
2375 goto new_slab
; /* Slab failed checks. Next slab needed */
2377 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
2380 local_irq_restore(flags
);
2385 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2386 * have the fastpath folded into their functions. So no function call
2387 * overhead for requests that can be satisfied on the fastpath.
2389 * The fastpath works by first checking if the lockless freelist can be used.
2390 * If not then __slab_alloc is called for slow processing.
2392 * Otherwise we can simply pick the next object from the lockless free list.
2394 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2395 gfp_t gfpflags
, int node
, unsigned long addr
)
2398 struct kmem_cache_cpu
*c
;
2402 if (slab_pre_alloc_hook(s
, gfpflags
))
2405 s
= memcg_kmem_get_cache(s
, gfpflags
);
2408 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2409 * enabled. We may switch back and forth between cpus while
2410 * reading from one cpu area. That does not matter as long
2411 * as we end up on the original cpu again when doing the cmpxchg.
2413 * Preemption is disabled for the retrieval of the tid because that
2414 * must occur from the current processor. We cannot allow rescheduling
2415 * on a different processor between the determination of the pointer
2416 * and the retrieval of the tid.
2419 c
= __this_cpu_ptr(s
->cpu_slab
);
2422 * The transaction ids are globally unique per cpu and per operation on
2423 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2424 * occurs on the right processor and that there was no operation on the
2425 * linked list in between.
2430 object
= c
->freelist
;
2432 if (unlikely(!object
|| !node_match(page
, node
)))
2433 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2436 void *next_object
= get_freepointer_safe(s
, object
);
2439 * The cmpxchg will only match if there was no additional
2440 * operation and if we are on the right processor.
2442 * The cmpxchg does the following atomically (without lock
2444 * 1. Relocate first pointer to the current per cpu area.
2445 * 2. Verify that tid and freelist have not been changed
2446 * 3. If they were not changed replace tid and freelist
2448 * Since this is without lock semantics the protection is only
2449 * against code executing on this cpu *not* from access by
2452 if (unlikely(!this_cpu_cmpxchg_double(
2453 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2455 next_object
, next_tid(tid
)))) {
2457 note_cmpxchg_failure("slab_alloc", s
, tid
);
2460 prefetch_freepointer(s
, next_object
);
2461 stat(s
, ALLOC_FASTPATH
);
2464 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2465 memset(object
, 0, s
->object_size
);
2467 slab_post_alloc_hook(s
, gfpflags
, object
);
2472 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2473 gfp_t gfpflags
, unsigned long addr
)
2475 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2478 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2480 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2482 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2487 EXPORT_SYMBOL(kmem_cache_alloc
);
2489 #ifdef CONFIG_TRACING
2490 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2492 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2493 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2496 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2500 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2502 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2504 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2505 s
->object_size
, s
->size
, gfpflags
, node
);
2509 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2511 #ifdef CONFIG_TRACING
2512 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2514 int node
, size_t size
)
2516 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2518 trace_kmalloc_node(_RET_IP_
, ret
,
2519 size
, s
->size
, gfpflags
, node
);
2522 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2527 * Slow patch handling. This may still be called frequently since objects
2528 * have a longer lifetime than the cpu slabs in most processing loads.
2530 * So we still attempt to reduce cache line usage. Just take the slab
2531 * lock and free the item. If there is no additional partial page
2532 * handling required then we can return immediately.
2534 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2535 void *x
, unsigned long addr
)
2538 void **object
= (void *)x
;
2541 unsigned long counters
;
2542 struct kmem_cache_node
*n
= NULL
;
2543 unsigned long uninitialized_var(flags
);
2545 stat(s
, FREE_SLOWPATH
);
2547 if (kmem_cache_debug(s
) &&
2548 !(n
= free_debug_processing(s
, page
, x
, addr
, &flags
)))
2553 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2556 prior
= page
->freelist
;
2557 counters
= page
->counters
;
2558 set_freepointer(s
, object
, prior
);
2559 new.counters
= counters
;
2560 was_frozen
= new.frozen
;
2562 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2564 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2567 * Slab was on no list before and will be
2569 * We can defer the list move and instead
2574 } else { /* Needs to be taken off a list */
2576 n
= get_node(s
, page_to_nid(page
));
2578 * Speculatively acquire the list_lock.
2579 * If the cmpxchg does not succeed then we may
2580 * drop the list_lock without any processing.
2582 * Otherwise the list_lock will synchronize with
2583 * other processors updating the list of slabs.
2585 spin_lock_irqsave(&n
->list_lock
, flags
);
2590 } while (!cmpxchg_double_slab(s
, page
,
2592 object
, new.counters
,
2598 * If we just froze the page then put it onto the
2599 * per cpu partial list.
2601 if (new.frozen
&& !was_frozen
) {
2602 put_cpu_partial(s
, page
, 1);
2603 stat(s
, CPU_PARTIAL_FREE
);
2606 * The list lock was not taken therefore no list
2607 * activity can be necessary.
2610 stat(s
, FREE_FROZEN
);
2614 if (unlikely(!new.inuse
&& n
->nr_partial
> s
->min_partial
))
2618 * Objects left in the slab. If it was not on the partial list before
2621 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2622 if (kmem_cache_debug(s
))
2623 remove_full(s
, n
, page
);
2624 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2625 stat(s
, FREE_ADD_PARTIAL
);
2627 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2633 * Slab on the partial list.
2635 remove_partial(n
, page
);
2636 stat(s
, FREE_REMOVE_PARTIAL
);
2638 /* Slab must be on the full list */
2639 remove_full(s
, n
, page
);
2642 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2644 discard_slab(s
, page
);
2648 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2649 * can perform fastpath freeing without additional function calls.
2651 * The fastpath is only possible if we are freeing to the current cpu slab
2652 * of this processor. This typically the case if we have just allocated
2655 * If fastpath is not possible then fall back to __slab_free where we deal
2656 * with all sorts of special processing.
2658 static __always_inline
void slab_free(struct kmem_cache
*s
,
2659 struct page
*page
, void *x
, unsigned long addr
)
2661 void **object
= (void *)x
;
2662 struct kmem_cache_cpu
*c
;
2665 slab_free_hook(s
, x
);
2669 * Determine the currently cpus per cpu slab.
2670 * The cpu may change afterward. However that does not matter since
2671 * data is retrieved via this pointer. If we are on the same cpu
2672 * during the cmpxchg then the free will succedd.
2675 c
= __this_cpu_ptr(s
->cpu_slab
);
2680 if (likely(page
== c
->page
)) {
2681 set_freepointer(s
, object
, c
->freelist
);
2683 if (unlikely(!this_cpu_cmpxchg_double(
2684 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2686 object
, next_tid(tid
)))) {
2688 note_cmpxchg_failure("slab_free", s
, tid
);
2691 stat(s
, FREE_FASTPATH
);
2693 __slab_free(s
, page
, x
, addr
);
2697 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2699 s
= cache_from_obj(s
, x
);
2702 slab_free(s
, virt_to_head_page(x
), x
, _RET_IP_
);
2703 trace_kmem_cache_free(_RET_IP_
, x
);
2705 EXPORT_SYMBOL(kmem_cache_free
);
2708 * Object placement in a slab is made very easy because we always start at
2709 * offset 0. If we tune the size of the object to the alignment then we can
2710 * get the required alignment by putting one properly sized object after
2713 * Notice that the allocation order determines the sizes of the per cpu
2714 * caches. Each processor has always one slab available for allocations.
2715 * Increasing the allocation order reduces the number of times that slabs
2716 * must be moved on and off the partial lists and is therefore a factor in
2721 * Mininum / Maximum order of slab pages. This influences locking overhead
2722 * and slab fragmentation. A higher order reduces the number of partial slabs
2723 * and increases the number of allocations possible without having to
2724 * take the list_lock.
2726 static int slub_min_order
;
2727 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
2728 static int slub_min_objects
;
2731 * Merge control. If this is set then no merging of slab caches will occur.
2732 * (Could be removed. This was introduced to pacify the merge skeptics.)
2734 static int slub_nomerge
;
2737 * Calculate the order of allocation given an slab object size.
2739 * The order of allocation has significant impact on performance and other
2740 * system components. Generally order 0 allocations should be preferred since
2741 * order 0 does not cause fragmentation in the page allocator. Larger objects
2742 * be problematic to put into order 0 slabs because there may be too much
2743 * unused space left. We go to a higher order if more than 1/16th of the slab
2746 * In order to reach satisfactory performance we must ensure that a minimum
2747 * number of objects is in one slab. Otherwise we may generate too much
2748 * activity on the partial lists which requires taking the list_lock. This is
2749 * less a concern for large slabs though which are rarely used.
2751 * slub_max_order specifies the order where we begin to stop considering the
2752 * number of objects in a slab as critical. If we reach slub_max_order then
2753 * we try to keep the page order as low as possible. So we accept more waste
2754 * of space in favor of a small page order.
2756 * Higher order allocations also allow the placement of more objects in a
2757 * slab and thereby reduce object handling overhead. If the user has
2758 * requested a higher mininum order then we start with that one instead of
2759 * the smallest order which will fit the object.
2761 static inline int slab_order(int size
, int min_objects
,
2762 int max_order
, int fract_leftover
, int reserved
)
2766 int min_order
= slub_min_order
;
2768 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
2769 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
2771 for (order
= max(min_order
,
2772 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
2773 order
<= max_order
; order
++) {
2775 unsigned long slab_size
= PAGE_SIZE
<< order
;
2777 if (slab_size
< min_objects
* size
+ reserved
)
2780 rem
= (slab_size
- reserved
) % size
;
2782 if (rem
<= slab_size
/ fract_leftover
)
2790 static inline int calculate_order(int size
, int reserved
)
2798 * Attempt to find best configuration for a slab. This
2799 * works by first attempting to generate a layout with
2800 * the best configuration and backing off gradually.
2802 * First we reduce the acceptable waste in a slab. Then
2803 * we reduce the minimum objects required in a slab.
2805 min_objects
= slub_min_objects
;
2807 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
2808 max_objects
= order_objects(slub_max_order
, size
, reserved
);
2809 min_objects
= min(min_objects
, max_objects
);
2811 while (min_objects
> 1) {
2813 while (fraction
>= 4) {
2814 order
= slab_order(size
, min_objects
,
2815 slub_max_order
, fraction
, reserved
);
2816 if (order
<= slub_max_order
)
2824 * We were unable to place multiple objects in a slab. Now
2825 * lets see if we can place a single object there.
2827 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
2828 if (order
<= slub_max_order
)
2832 * Doh this slab cannot be placed using slub_max_order.
2834 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
2835 if (order
< MAX_ORDER
)
2841 init_kmem_cache_node(struct kmem_cache_node
*n
)
2844 spin_lock_init(&n
->list_lock
);
2845 INIT_LIST_HEAD(&n
->partial
);
2846 #ifdef CONFIG_SLUB_DEBUG
2847 atomic_long_set(&n
->nr_slabs
, 0);
2848 atomic_long_set(&n
->total_objects
, 0);
2849 INIT_LIST_HEAD(&n
->full
);
2853 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
2855 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
2856 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
2859 * Must align to double word boundary for the double cmpxchg
2860 * instructions to work; see __pcpu_double_call_return_bool().
2862 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
2863 2 * sizeof(void *));
2868 init_kmem_cache_cpus(s
);
2873 static struct kmem_cache
*kmem_cache_node
;
2876 * No kmalloc_node yet so do it by hand. We know that this is the first
2877 * slab on the node for this slabcache. There are no concurrent accesses
2880 * Note that this function only works on the kmem_cache_node
2881 * when allocating for the kmem_cache_node. This is used for bootstrapping
2882 * memory on a fresh node that has no slab structures yet.
2884 static void early_kmem_cache_node_alloc(int node
)
2887 struct kmem_cache_node
*n
;
2889 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
2891 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
2894 if (page_to_nid(page
) != node
) {
2895 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
2897 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
2898 "in order to be able to continue\n");
2903 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
2906 kmem_cache_node
->node
[node
] = n
;
2907 #ifdef CONFIG_SLUB_DEBUG
2908 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
2909 init_tracking(kmem_cache_node
, n
);
2911 init_kmem_cache_node(n
);
2912 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
2915 * No locks need to be taken here as it has just been
2916 * initialized and there is no concurrent access.
2918 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
2921 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2925 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2926 struct kmem_cache_node
*n
= s
->node
[node
];
2929 kmem_cache_free(kmem_cache_node
, n
);
2931 s
->node
[node
] = NULL
;
2935 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
2939 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2940 struct kmem_cache_node
*n
;
2942 if (slab_state
== DOWN
) {
2943 early_kmem_cache_node_alloc(node
);
2946 n
= kmem_cache_alloc_node(kmem_cache_node
,
2950 free_kmem_cache_nodes(s
);
2955 init_kmem_cache_node(n
);
2960 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
2962 if (min
< MIN_PARTIAL
)
2964 else if (min
> MAX_PARTIAL
)
2966 s
->min_partial
= min
;
2970 * calculate_sizes() determines the order and the distribution of data within
2973 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
2975 unsigned long flags
= s
->flags
;
2976 unsigned long size
= s
->object_size
;
2980 * Round up object size to the next word boundary. We can only
2981 * place the free pointer at word boundaries and this determines
2982 * the possible location of the free pointer.
2984 size
= ALIGN(size
, sizeof(void *));
2986 #ifdef CONFIG_SLUB_DEBUG
2988 * Determine if we can poison the object itself. If the user of
2989 * the slab may touch the object after free or before allocation
2990 * then we should never poison the object itself.
2992 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2994 s
->flags
|= __OBJECT_POISON
;
2996 s
->flags
&= ~__OBJECT_POISON
;
3000 * If we are Redzoning then check if there is some space between the
3001 * end of the object and the free pointer. If not then add an
3002 * additional word to have some bytes to store Redzone information.
3004 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3005 size
+= sizeof(void *);
3009 * With that we have determined the number of bytes in actual use
3010 * by the object. This is the potential offset to the free pointer.
3014 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
3017 * Relocate free pointer after the object if it is not
3018 * permitted to overwrite the first word of the object on
3021 * This is the case if we do RCU, have a constructor or
3022 * destructor or are poisoning the objects.
3025 size
+= sizeof(void *);
3028 #ifdef CONFIG_SLUB_DEBUG
3029 if (flags
& SLAB_STORE_USER
)
3031 * Need to store information about allocs and frees after
3034 size
+= 2 * sizeof(struct track
);
3036 if (flags
& SLAB_RED_ZONE
)
3038 * Add some empty padding so that we can catch
3039 * overwrites from earlier objects rather than let
3040 * tracking information or the free pointer be
3041 * corrupted if a user writes before the start
3044 size
+= sizeof(void *);
3048 * SLUB stores one object immediately after another beginning from
3049 * offset 0. In order to align the objects we have to simply size
3050 * each object to conform to the alignment.
3052 size
= ALIGN(size
, s
->align
);
3054 if (forced_order
>= 0)
3055 order
= forced_order
;
3057 order
= calculate_order(size
, s
->reserved
);
3064 s
->allocflags
|= __GFP_COMP
;
3066 if (s
->flags
& SLAB_CACHE_DMA
)
3067 s
->allocflags
|= GFP_DMA
;
3069 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3070 s
->allocflags
|= __GFP_RECLAIMABLE
;
3073 * Determine the number of objects per slab
3075 s
->oo
= oo_make(order
, size
, s
->reserved
);
3076 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3077 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3080 return !!oo_objects(s
->oo
);
3083 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3085 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3088 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
3089 s
->reserved
= sizeof(struct rcu_head
);
3091 if (!calculate_sizes(s
, -1))
3093 if (disable_higher_order_debug
) {
3095 * Disable debugging flags that store metadata if the min slab
3098 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3099 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3101 if (!calculate_sizes(s
, -1))
3106 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3107 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3108 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_DEBUG_FLAGS
) == 0)
3109 /* Enable fast mode */
3110 s
->flags
|= __CMPXCHG_DOUBLE
;
3114 * The larger the object size is, the more pages we want on the partial
3115 * list to avoid pounding the page allocator excessively.
3117 set_min_partial(s
, ilog2(s
->size
) / 2);
3120 * cpu_partial determined the maximum number of objects kept in the
3121 * per cpu partial lists of a processor.
3123 * Per cpu partial lists mainly contain slabs that just have one
3124 * object freed. If they are used for allocation then they can be
3125 * filled up again with minimal effort. The slab will never hit the
3126 * per node partial lists and therefore no locking will be required.
3128 * This setting also determines
3130 * A) The number of objects from per cpu partial slabs dumped to the
3131 * per node list when we reach the limit.
3132 * B) The number of objects in cpu partial slabs to extract from the
3133 * per node list when we run out of per cpu objects. We only fetch
3134 * 50% to keep some capacity around for frees.
3136 if (!kmem_cache_has_cpu_partial(s
))
3138 else if (s
->size
>= PAGE_SIZE
)
3140 else if (s
->size
>= 1024)
3142 else if (s
->size
>= 256)
3143 s
->cpu_partial
= 13;
3145 s
->cpu_partial
= 30;
3148 s
->remote_node_defrag_ratio
= 1000;
3150 if (!init_kmem_cache_nodes(s
))
3153 if (alloc_kmem_cache_cpus(s
))
3156 free_kmem_cache_nodes(s
);
3158 if (flags
& SLAB_PANIC
)
3159 panic("Cannot create slab %s size=%lu realsize=%u "
3160 "order=%u offset=%u flags=%lx\n",
3161 s
->name
, (unsigned long)s
->size
, s
->size
,
3162 oo_order(s
->oo
), s
->offset
, flags
);
3166 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3169 #ifdef CONFIG_SLUB_DEBUG
3170 void *addr
= page_address(page
);
3172 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3173 sizeof(long), GFP_ATOMIC
);
3176 slab_err(s
, page
, text
, s
->name
);
3179 get_map(s
, page
, map
);
3180 for_each_object(p
, s
, addr
, page
->objects
) {
3182 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3183 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu\n",
3185 print_tracking(s
, p
);
3194 * Attempt to free all partial slabs on a node.
3195 * This is called from kmem_cache_close(). We must be the last thread
3196 * using the cache and therefore we do not need to lock anymore.
3198 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3200 struct page
*page
, *h
;
3202 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3204 __remove_partial(n
, page
);
3205 discard_slab(s
, page
);
3207 list_slab_objects(s
, page
,
3208 "Objects remaining in %s on kmem_cache_close()");
3214 * Release all resources used by a slab cache.
3216 static inline int kmem_cache_close(struct kmem_cache
*s
)
3221 /* Attempt to free all objects */
3222 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3223 struct kmem_cache_node
*n
= get_node(s
, node
);
3226 if (n
->nr_partial
|| slabs_node(s
, node
))
3229 free_percpu(s
->cpu_slab
);
3230 free_kmem_cache_nodes(s
);
3234 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3236 int rc
= kmem_cache_close(s
);
3240 * We do the same lock strategy around sysfs_slab_add, see
3241 * __kmem_cache_create. Because this is pretty much the last
3242 * operation we do and the lock will be released shortly after
3243 * that in slab_common.c, we could just move sysfs_slab_remove
3244 * to a later point in common code. We should do that when we
3245 * have a common sysfs framework for all allocators.
3247 mutex_unlock(&slab_mutex
);
3248 sysfs_slab_remove(s
);
3249 mutex_lock(&slab_mutex
);
3255 /********************************************************************
3257 *******************************************************************/
3259 static int __init
setup_slub_min_order(char *str
)
3261 get_option(&str
, &slub_min_order
);
3266 __setup("slub_min_order=", setup_slub_min_order
);
3268 static int __init
setup_slub_max_order(char *str
)
3270 get_option(&str
, &slub_max_order
);
3271 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3276 __setup("slub_max_order=", setup_slub_max_order
);
3278 static int __init
setup_slub_min_objects(char *str
)
3280 get_option(&str
, &slub_min_objects
);
3285 __setup("slub_min_objects=", setup_slub_min_objects
);
3287 static int __init
setup_slub_nomerge(char *str
)
3293 __setup("slub_nomerge", setup_slub_nomerge
);
3295 void *__kmalloc(size_t size
, gfp_t flags
)
3297 struct kmem_cache
*s
;
3300 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3301 return kmalloc_large(size
, flags
);
3303 s
= kmalloc_slab(size
, flags
);
3305 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3308 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3310 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3314 EXPORT_SYMBOL(__kmalloc
);
3317 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3322 flags
|= __GFP_COMP
| __GFP_NOTRACK
| __GFP_KMEMCG
;
3323 page
= alloc_pages_node(node
, flags
, get_order(size
));
3325 ptr
= page_address(page
);
3327 kmalloc_large_node_hook(ptr
, size
, flags
);
3331 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3333 struct kmem_cache
*s
;
3336 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3337 ret
= kmalloc_large_node(size
, flags
, node
);
3339 trace_kmalloc_node(_RET_IP_
, ret
,
3340 size
, PAGE_SIZE
<< get_order(size
),
3346 s
= kmalloc_slab(size
, flags
);
3348 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3351 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3353 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3357 EXPORT_SYMBOL(__kmalloc_node
);
3360 size_t ksize(const void *object
)
3364 if (unlikely(object
== ZERO_SIZE_PTR
))
3367 page
= virt_to_head_page(object
);
3369 if (unlikely(!PageSlab(page
))) {
3370 WARN_ON(!PageCompound(page
));
3371 return PAGE_SIZE
<< compound_order(page
);
3374 return slab_ksize(page
->slab_cache
);
3376 EXPORT_SYMBOL(ksize
);
3378 void kfree(const void *x
)
3381 void *object
= (void *)x
;
3383 trace_kfree(_RET_IP_
, x
);
3385 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3388 page
= virt_to_head_page(x
);
3389 if (unlikely(!PageSlab(page
))) {
3390 BUG_ON(!PageCompound(page
));
3392 __free_memcg_kmem_pages(page
, compound_order(page
));
3395 slab_free(page
->slab_cache
, page
, object
, _RET_IP_
);
3397 EXPORT_SYMBOL(kfree
);
3400 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3401 * the remaining slabs by the number of items in use. The slabs with the
3402 * most items in use come first. New allocations will then fill those up
3403 * and thus they can be removed from the partial lists.
3405 * The slabs with the least items are placed last. This results in them
3406 * being allocated from last increasing the chance that the last objects
3407 * are freed in them.
3409 int kmem_cache_shrink(struct kmem_cache
*s
)
3413 struct kmem_cache_node
*n
;
3416 int objects
= oo_objects(s
->max
);
3417 struct list_head
*slabs_by_inuse
=
3418 kmalloc(sizeof(struct list_head
) * objects
, GFP_KERNEL
);
3419 unsigned long flags
;
3421 if (!slabs_by_inuse
)
3425 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3426 n
= get_node(s
, node
);
3431 for (i
= 0; i
< objects
; i
++)
3432 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
3434 spin_lock_irqsave(&n
->list_lock
, flags
);
3437 * Build lists indexed by the items in use in each slab.
3439 * Note that concurrent frees may occur while we hold the
3440 * list_lock. page->inuse here is the upper limit.
3442 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3443 list_move(&page
->lru
, slabs_by_inuse
+ page
->inuse
);
3449 * Rebuild the partial list with the slabs filled up most
3450 * first and the least used slabs at the end.
3452 for (i
= objects
- 1; i
> 0; i
--)
3453 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
3455 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3457 /* Release empty slabs */
3458 list_for_each_entry_safe(page
, t
, slabs_by_inuse
, lru
)
3459 discard_slab(s
, page
);
3462 kfree(slabs_by_inuse
);
3465 EXPORT_SYMBOL(kmem_cache_shrink
);
3467 static int slab_mem_going_offline_callback(void *arg
)
3469 struct kmem_cache
*s
;
3471 mutex_lock(&slab_mutex
);
3472 list_for_each_entry(s
, &slab_caches
, list
)
3473 kmem_cache_shrink(s
);
3474 mutex_unlock(&slab_mutex
);
3479 static void slab_mem_offline_callback(void *arg
)
3481 struct kmem_cache_node
*n
;
3482 struct kmem_cache
*s
;
3483 struct memory_notify
*marg
= arg
;
3486 offline_node
= marg
->status_change_nid_normal
;
3489 * If the node still has available memory. we need kmem_cache_node
3492 if (offline_node
< 0)
3495 mutex_lock(&slab_mutex
);
3496 list_for_each_entry(s
, &slab_caches
, list
) {
3497 n
= get_node(s
, offline_node
);
3500 * if n->nr_slabs > 0, slabs still exist on the node
3501 * that is going down. We were unable to free them,
3502 * and offline_pages() function shouldn't call this
3503 * callback. So, we must fail.
3505 BUG_ON(slabs_node(s
, offline_node
));
3507 s
->node
[offline_node
] = NULL
;
3508 kmem_cache_free(kmem_cache_node
, n
);
3511 mutex_unlock(&slab_mutex
);
3514 static int slab_mem_going_online_callback(void *arg
)
3516 struct kmem_cache_node
*n
;
3517 struct kmem_cache
*s
;
3518 struct memory_notify
*marg
= arg
;
3519 int nid
= marg
->status_change_nid_normal
;
3523 * If the node's memory is already available, then kmem_cache_node is
3524 * already created. Nothing to do.
3530 * We are bringing a node online. No memory is available yet. We must
3531 * allocate a kmem_cache_node structure in order to bring the node
3534 mutex_lock(&slab_mutex
);
3535 list_for_each_entry(s
, &slab_caches
, list
) {
3537 * XXX: kmem_cache_alloc_node will fallback to other nodes
3538 * since memory is not yet available from the node that
3541 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3546 init_kmem_cache_node(n
);
3550 mutex_unlock(&slab_mutex
);
3554 static int slab_memory_callback(struct notifier_block
*self
,
3555 unsigned long action
, void *arg
)
3560 case MEM_GOING_ONLINE
:
3561 ret
= slab_mem_going_online_callback(arg
);
3563 case MEM_GOING_OFFLINE
:
3564 ret
= slab_mem_going_offline_callback(arg
);
3567 case MEM_CANCEL_ONLINE
:
3568 slab_mem_offline_callback(arg
);
3571 case MEM_CANCEL_OFFLINE
:
3575 ret
= notifier_from_errno(ret
);
3581 static struct notifier_block slab_memory_callback_nb
= {
3582 .notifier_call
= slab_memory_callback
,
3583 .priority
= SLAB_CALLBACK_PRI
,
3586 /********************************************************************
3587 * Basic setup of slabs
3588 *******************************************************************/
3591 * Used for early kmem_cache structures that were allocated using
3592 * the page allocator. Allocate them properly then fix up the pointers
3593 * that may be pointing to the wrong kmem_cache structure.
3596 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
3599 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
3601 memcpy(s
, static_cache
, kmem_cache
->object_size
);
3604 * This runs very early, and only the boot processor is supposed to be
3605 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3608 __flush_cpu_slab(s
, smp_processor_id());
3609 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3610 struct kmem_cache_node
*n
= get_node(s
, node
);
3614 list_for_each_entry(p
, &n
->partial
, lru
)
3617 #ifdef CONFIG_SLUB_DEBUG
3618 list_for_each_entry(p
, &n
->full
, lru
)
3623 list_add(&s
->list
, &slab_caches
);
3627 void __init
kmem_cache_init(void)
3629 static __initdata
struct kmem_cache boot_kmem_cache
,
3630 boot_kmem_cache_node
;
3632 if (debug_guardpage_minorder())
3635 kmem_cache_node
= &boot_kmem_cache_node
;
3636 kmem_cache
= &boot_kmem_cache
;
3638 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
3639 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
3641 register_hotmemory_notifier(&slab_memory_callback_nb
);
3643 /* Able to allocate the per node structures */
3644 slab_state
= PARTIAL
;
3646 create_boot_cache(kmem_cache
, "kmem_cache",
3647 offsetof(struct kmem_cache
, node
) +
3648 nr_node_ids
* sizeof(struct kmem_cache_node
*),
3649 SLAB_HWCACHE_ALIGN
);
3651 kmem_cache
= bootstrap(&boot_kmem_cache
);
3654 * Allocate kmem_cache_node properly from the kmem_cache slab.
3655 * kmem_cache_node is separately allocated so no need to
3656 * update any list pointers.
3658 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
3660 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3661 create_kmalloc_caches(0);
3664 register_cpu_notifier(&slab_notifier
);
3668 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3669 " CPUs=%d, Nodes=%d\n",
3671 slub_min_order
, slub_max_order
, slub_min_objects
,
3672 nr_cpu_ids
, nr_node_ids
);
3675 void __init
kmem_cache_init_late(void)
3680 * Find a mergeable slab cache
3682 static int slab_unmergeable(struct kmem_cache
*s
)
3684 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
3691 * We may have set a slab to be unmergeable during bootstrap.
3693 if (s
->refcount
< 0)
3699 static struct kmem_cache
*find_mergeable(struct mem_cgroup
*memcg
, size_t size
,
3700 size_t align
, unsigned long flags
, const char *name
,
3701 void (*ctor
)(void *))
3703 struct kmem_cache
*s
;
3705 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
3711 size
= ALIGN(size
, sizeof(void *));
3712 align
= calculate_alignment(flags
, align
, size
);
3713 size
= ALIGN(size
, align
);
3714 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
3716 list_for_each_entry(s
, &slab_caches
, list
) {
3717 if (slab_unmergeable(s
))
3723 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
3726 * Check if alignment is compatible.
3727 * Courtesy of Adrian Drzewiecki
3729 if ((s
->size
& ~(align
- 1)) != s
->size
)
3732 if (s
->size
- size
>= sizeof(void *))
3735 if (!cache_match_memcg(s
, memcg
))
3744 __kmem_cache_alias(struct mem_cgroup
*memcg
, const char *name
, size_t size
,
3745 size_t align
, unsigned long flags
, void (*ctor
)(void *))
3747 struct kmem_cache
*s
;
3749 s
= find_mergeable(memcg
, size
, align
, flags
, name
, ctor
);
3753 * Adjust the object sizes so that we clear
3754 * the complete object on kzalloc.
3756 s
->object_size
= max(s
->object_size
, (int)size
);
3757 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3759 if (sysfs_slab_alias(s
, name
)) {
3768 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
3772 err
= kmem_cache_open(s
, flags
);
3776 /* Mutex is not taken during early boot */
3777 if (slab_state
<= UP
)
3780 memcg_propagate_slab_attrs(s
);
3781 mutex_unlock(&slab_mutex
);
3782 err
= sysfs_slab_add(s
);
3783 mutex_lock(&slab_mutex
);
3786 kmem_cache_close(s
);
3793 * Use the cpu notifier to insure that the cpu slabs are flushed when
3796 static int slab_cpuup_callback(struct notifier_block
*nfb
,
3797 unsigned long action
, void *hcpu
)
3799 long cpu
= (long)hcpu
;
3800 struct kmem_cache
*s
;
3801 unsigned long flags
;
3804 case CPU_UP_CANCELED
:
3805 case CPU_UP_CANCELED_FROZEN
:
3807 case CPU_DEAD_FROZEN
:
3808 mutex_lock(&slab_mutex
);
3809 list_for_each_entry(s
, &slab_caches
, list
) {
3810 local_irq_save(flags
);
3811 __flush_cpu_slab(s
, cpu
);
3812 local_irq_restore(flags
);
3814 mutex_unlock(&slab_mutex
);
3822 static struct notifier_block slab_notifier
= {
3823 .notifier_call
= slab_cpuup_callback
3828 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
3830 struct kmem_cache
*s
;
3833 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3834 return kmalloc_large(size
, gfpflags
);
3836 s
= kmalloc_slab(size
, gfpflags
);
3838 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3841 ret
= slab_alloc(s
, gfpflags
, caller
);
3843 /* Honor the call site pointer we received. */
3844 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
3850 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3851 int node
, unsigned long caller
)
3853 struct kmem_cache
*s
;
3856 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3857 ret
= kmalloc_large_node(size
, gfpflags
, node
);
3859 trace_kmalloc_node(caller
, ret
,
3860 size
, PAGE_SIZE
<< get_order(size
),
3866 s
= kmalloc_slab(size
, gfpflags
);
3868 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3871 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
3873 /* Honor the call site pointer we received. */
3874 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
3881 static int count_inuse(struct page
*page
)
3886 static int count_total(struct page
*page
)
3888 return page
->objects
;
3892 #ifdef CONFIG_SLUB_DEBUG
3893 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3897 void *addr
= page_address(page
);
3899 if (!check_slab(s
, page
) ||
3900 !on_freelist(s
, page
, NULL
))
3903 /* Now we know that a valid freelist exists */
3904 bitmap_zero(map
, page
->objects
);
3906 get_map(s
, page
, map
);
3907 for_each_object(p
, s
, addr
, page
->objects
) {
3908 if (test_bit(slab_index(p
, s
, addr
), map
))
3909 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
3913 for_each_object(p
, s
, addr
, page
->objects
)
3914 if (!test_bit(slab_index(p
, s
, addr
), map
))
3915 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
3920 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
3924 validate_slab(s
, page
, map
);
3928 static int validate_slab_node(struct kmem_cache
*s
,
3929 struct kmem_cache_node
*n
, unsigned long *map
)
3931 unsigned long count
= 0;
3933 unsigned long flags
;
3935 spin_lock_irqsave(&n
->list_lock
, flags
);
3937 list_for_each_entry(page
, &n
->partial
, lru
) {
3938 validate_slab_slab(s
, page
, map
);
3941 if (count
!= n
->nr_partial
)
3942 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
3943 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
3945 if (!(s
->flags
& SLAB_STORE_USER
))
3948 list_for_each_entry(page
, &n
->full
, lru
) {
3949 validate_slab_slab(s
, page
, map
);
3952 if (count
!= atomic_long_read(&n
->nr_slabs
))
3953 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
3954 "counter=%ld\n", s
->name
, count
,
3955 atomic_long_read(&n
->nr_slabs
));
3958 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3962 static long validate_slab_cache(struct kmem_cache
*s
)
3965 unsigned long count
= 0;
3966 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
3967 sizeof(unsigned long), GFP_KERNEL
);
3973 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3974 struct kmem_cache_node
*n
= get_node(s
, node
);
3976 count
+= validate_slab_node(s
, n
, map
);
3982 * Generate lists of code addresses where slabcache objects are allocated
3987 unsigned long count
;
3994 DECLARE_BITMAP(cpus
, NR_CPUS
);
4000 unsigned long count
;
4001 struct location
*loc
;
4004 static void free_loc_track(struct loc_track
*t
)
4007 free_pages((unsigned long)t
->loc
,
4008 get_order(sizeof(struct location
) * t
->max
));
4011 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4016 order
= get_order(sizeof(struct location
) * max
);
4018 l
= (void *)__get_free_pages(flags
, order
);
4023 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4031 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4032 const struct track
*track
)
4034 long start
, end
, pos
;
4036 unsigned long caddr
;
4037 unsigned long age
= jiffies
- track
->when
;
4043 pos
= start
+ (end
- start
+ 1) / 2;
4046 * There is nothing at "end". If we end up there
4047 * we need to add something to before end.
4052 caddr
= t
->loc
[pos
].addr
;
4053 if (track
->addr
== caddr
) {
4059 if (age
< l
->min_time
)
4061 if (age
> l
->max_time
)
4064 if (track
->pid
< l
->min_pid
)
4065 l
->min_pid
= track
->pid
;
4066 if (track
->pid
> l
->max_pid
)
4067 l
->max_pid
= track
->pid
;
4069 cpumask_set_cpu(track
->cpu
,
4070 to_cpumask(l
->cpus
));
4072 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4076 if (track
->addr
< caddr
)
4083 * Not found. Insert new tracking element.
4085 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4091 (t
->count
- pos
) * sizeof(struct location
));
4094 l
->addr
= track
->addr
;
4098 l
->min_pid
= track
->pid
;
4099 l
->max_pid
= track
->pid
;
4100 cpumask_clear(to_cpumask(l
->cpus
));
4101 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4102 nodes_clear(l
->nodes
);
4103 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4107 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4108 struct page
*page
, enum track_item alloc
,
4111 void *addr
= page_address(page
);
4114 bitmap_zero(map
, page
->objects
);
4115 get_map(s
, page
, map
);
4117 for_each_object(p
, s
, addr
, page
->objects
)
4118 if (!test_bit(slab_index(p
, s
, addr
), map
))
4119 add_location(t
, s
, get_track(s
, p
, alloc
));
4122 static int list_locations(struct kmem_cache
*s
, char *buf
,
4123 enum track_item alloc
)
4127 struct loc_track t
= { 0, 0, NULL
};
4129 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4130 sizeof(unsigned long), GFP_KERNEL
);
4132 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4135 return sprintf(buf
, "Out of memory\n");
4137 /* Push back cpu slabs */
4140 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4141 struct kmem_cache_node
*n
= get_node(s
, node
);
4142 unsigned long flags
;
4145 if (!atomic_long_read(&n
->nr_slabs
))
4148 spin_lock_irqsave(&n
->list_lock
, flags
);
4149 list_for_each_entry(page
, &n
->partial
, lru
)
4150 process_slab(&t
, s
, page
, alloc
, map
);
4151 list_for_each_entry(page
, &n
->full
, lru
)
4152 process_slab(&t
, s
, page
, alloc
, map
);
4153 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4156 for (i
= 0; i
< t
.count
; i
++) {
4157 struct location
*l
= &t
.loc
[i
];
4159 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4161 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4164 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4166 len
+= sprintf(buf
+ len
, "<not-available>");
4168 if (l
->sum_time
!= l
->min_time
) {
4169 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4171 (long)div_u64(l
->sum_time
, l
->count
),
4174 len
+= sprintf(buf
+ len
, " age=%ld",
4177 if (l
->min_pid
!= l
->max_pid
)
4178 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4179 l
->min_pid
, l
->max_pid
);
4181 len
+= sprintf(buf
+ len
, " pid=%ld",
4184 if (num_online_cpus() > 1 &&
4185 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4186 len
< PAGE_SIZE
- 60) {
4187 len
+= sprintf(buf
+ len
, " cpus=");
4188 len
+= cpulist_scnprintf(buf
+ len
,
4189 PAGE_SIZE
- len
- 50,
4190 to_cpumask(l
->cpus
));
4193 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4194 len
< PAGE_SIZE
- 60) {
4195 len
+= sprintf(buf
+ len
, " nodes=");
4196 len
+= nodelist_scnprintf(buf
+ len
,
4197 PAGE_SIZE
- len
- 50,
4201 len
+= sprintf(buf
+ len
, "\n");
4207 len
+= sprintf(buf
, "No data\n");
4212 #ifdef SLUB_RESILIENCY_TEST
4213 static void resiliency_test(void)
4217 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4219 printk(KERN_ERR
"SLUB resiliency testing\n");
4220 printk(KERN_ERR
"-----------------------\n");
4221 printk(KERN_ERR
"A. Corruption after allocation\n");
4223 p
= kzalloc(16, GFP_KERNEL
);
4225 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
4226 " 0x12->0x%p\n\n", p
+ 16);
4228 validate_slab_cache(kmalloc_caches
[4]);
4230 /* Hmmm... The next two are dangerous */
4231 p
= kzalloc(32, GFP_KERNEL
);
4232 p
[32 + sizeof(void *)] = 0x34;
4233 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
4234 " 0x34 -> -0x%p\n", p
);
4236 "If allocated object is overwritten then not detectable\n\n");
4238 validate_slab_cache(kmalloc_caches
[5]);
4239 p
= kzalloc(64, GFP_KERNEL
);
4240 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4242 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4245 "If allocated object is overwritten then not detectable\n\n");
4246 validate_slab_cache(kmalloc_caches
[6]);
4248 printk(KERN_ERR
"\nB. Corruption after free\n");
4249 p
= kzalloc(128, GFP_KERNEL
);
4252 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4253 validate_slab_cache(kmalloc_caches
[7]);
4255 p
= kzalloc(256, GFP_KERNEL
);
4258 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4260 validate_slab_cache(kmalloc_caches
[8]);
4262 p
= kzalloc(512, GFP_KERNEL
);
4265 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4266 validate_slab_cache(kmalloc_caches
[9]);
4270 static void resiliency_test(void) {};
4275 enum slab_stat_type
{
4276 SL_ALL
, /* All slabs */
4277 SL_PARTIAL
, /* Only partially allocated slabs */
4278 SL_CPU
, /* Only slabs used for cpu caches */
4279 SL_OBJECTS
, /* Determine allocated objects not slabs */
4280 SL_TOTAL
/* Determine object capacity not slabs */
4283 #define SO_ALL (1 << SL_ALL)
4284 #define SO_PARTIAL (1 << SL_PARTIAL)
4285 #define SO_CPU (1 << SL_CPU)
4286 #define SO_OBJECTS (1 << SL_OBJECTS)
4287 #define SO_TOTAL (1 << SL_TOTAL)
4289 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4290 char *buf
, unsigned long flags
)
4292 unsigned long total
= 0;
4295 unsigned long *nodes
;
4297 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4301 if (flags
& SO_CPU
) {
4304 for_each_possible_cpu(cpu
) {
4305 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4310 page
= ACCESS_ONCE(c
->page
);
4314 node
= page_to_nid(page
);
4315 if (flags
& SO_TOTAL
)
4317 else if (flags
& SO_OBJECTS
)
4325 page
= ACCESS_ONCE(c
->partial
);
4327 node
= page_to_nid(page
);
4328 if (flags
& SO_TOTAL
)
4330 else if (flags
& SO_OBJECTS
)
4340 lock_memory_hotplug();
4341 #ifdef CONFIG_SLUB_DEBUG
4342 if (flags
& SO_ALL
) {
4343 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4344 struct kmem_cache_node
*n
= get_node(s
, node
);
4346 if (flags
& SO_TOTAL
)
4347 x
= atomic_long_read(&n
->total_objects
);
4348 else if (flags
& SO_OBJECTS
)
4349 x
= atomic_long_read(&n
->total_objects
) -
4350 count_partial(n
, count_free
);
4352 x
= atomic_long_read(&n
->nr_slabs
);
4359 if (flags
& SO_PARTIAL
) {
4360 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4361 struct kmem_cache_node
*n
= get_node(s
, node
);
4363 if (flags
& SO_TOTAL
)
4364 x
= count_partial(n
, count_total
);
4365 else if (flags
& SO_OBJECTS
)
4366 x
= count_partial(n
, count_inuse
);
4373 x
= sprintf(buf
, "%lu", total
);
4375 for_each_node_state(node
, N_NORMAL_MEMORY
)
4377 x
+= sprintf(buf
+ x
, " N%d=%lu",
4380 unlock_memory_hotplug();
4382 return x
+ sprintf(buf
+ x
, "\n");
4385 #ifdef CONFIG_SLUB_DEBUG
4386 static int any_slab_objects(struct kmem_cache
*s
)
4390 for_each_online_node(node
) {
4391 struct kmem_cache_node
*n
= get_node(s
, node
);
4396 if (atomic_long_read(&n
->total_objects
))
4403 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4404 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4406 struct slab_attribute
{
4407 struct attribute attr
;
4408 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4409 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4412 #define SLAB_ATTR_RO(_name) \
4413 static struct slab_attribute _name##_attr = \
4414 __ATTR(_name, 0400, _name##_show, NULL)
4416 #define SLAB_ATTR(_name) \
4417 static struct slab_attribute _name##_attr = \
4418 __ATTR(_name, 0600, _name##_show, _name##_store)
4420 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4422 return sprintf(buf
, "%d\n", s
->size
);
4424 SLAB_ATTR_RO(slab_size
);
4426 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4428 return sprintf(buf
, "%d\n", s
->align
);
4430 SLAB_ATTR_RO(align
);
4432 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4434 return sprintf(buf
, "%d\n", s
->object_size
);
4436 SLAB_ATTR_RO(object_size
);
4438 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4440 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4442 SLAB_ATTR_RO(objs_per_slab
);
4444 static ssize_t
order_store(struct kmem_cache
*s
,
4445 const char *buf
, size_t length
)
4447 unsigned long order
;
4450 err
= kstrtoul(buf
, 10, &order
);
4454 if (order
> slub_max_order
|| order
< slub_min_order
)
4457 calculate_sizes(s
, order
);
4461 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4463 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4467 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4469 return sprintf(buf
, "%lu\n", s
->min_partial
);
4472 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4478 err
= kstrtoul(buf
, 10, &min
);
4482 set_min_partial(s
, min
);
4485 SLAB_ATTR(min_partial
);
4487 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4489 return sprintf(buf
, "%u\n", s
->cpu_partial
);
4492 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4495 unsigned long objects
;
4498 err
= kstrtoul(buf
, 10, &objects
);
4501 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4504 s
->cpu_partial
= objects
;
4508 SLAB_ATTR(cpu_partial
);
4510 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4514 return sprintf(buf
, "%pS\n", s
->ctor
);
4518 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4520 return sprintf(buf
, "%d\n", s
->refcount
- 1);
4522 SLAB_ATTR_RO(aliases
);
4524 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4526 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4528 SLAB_ATTR_RO(partial
);
4530 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4532 return show_slab_objects(s
, buf
, SO_CPU
);
4534 SLAB_ATTR_RO(cpu_slabs
);
4536 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4538 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4540 SLAB_ATTR_RO(objects
);
4542 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4544 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4546 SLAB_ATTR_RO(objects_partial
);
4548 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4555 for_each_online_cpu(cpu
) {
4556 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
)->partial
;
4559 pages
+= page
->pages
;
4560 objects
+= page
->pobjects
;
4564 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
4567 for_each_online_cpu(cpu
) {
4568 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
) ->partial
;
4570 if (page
&& len
< PAGE_SIZE
- 20)
4571 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
4572 page
->pobjects
, page
->pages
);
4575 return len
+ sprintf(buf
+ len
, "\n");
4577 SLAB_ATTR_RO(slabs_cpu_partial
);
4579 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4581 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4584 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4585 const char *buf
, size_t length
)
4587 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4589 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4592 SLAB_ATTR(reclaim_account
);
4594 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4596 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4598 SLAB_ATTR_RO(hwcache_align
);
4600 #ifdef CONFIG_ZONE_DMA
4601 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4603 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4605 SLAB_ATTR_RO(cache_dma
);
4608 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4610 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4612 SLAB_ATTR_RO(destroy_by_rcu
);
4614 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4616 return sprintf(buf
, "%d\n", s
->reserved
);
4618 SLAB_ATTR_RO(reserved
);
4620 #ifdef CONFIG_SLUB_DEBUG
4621 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4623 return show_slab_objects(s
, buf
, SO_ALL
);
4625 SLAB_ATTR_RO(slabs
);
4627 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4629 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4631 SLAB_ATTR_RO(total_objects
);
4633 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4635 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
4638 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
4639 const char *buf
, size_t length
)
4641 s
->flags
&= ~SLAB_DEBUG_FREE
;
4642 if (buf
[0] == '1') {
4643 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4644 s
->flags
|= SLAB_DEBUG_FREE
;
4648 SLAB_ATTR(sanity_checks
);
4650 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
4652 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
4655 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
4658 s
->flags
&= ~SLAB_TRACE
;
4659 if (buf
[0] == '1') {
4660 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4661 s
->flags
|= SLAB_TRACE
;
4667 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
4669 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
4672 static ssize_t
red_zone_store(struct kmem_cache
*s
,
4673 const char *buf
, size_t length
)
4675 if (any_slab_objects(s
))
4678 s
->flags
&= ~SLAB_RED_ZONE
;
4679 if (buf
[0] == '1') {
4680 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4681 s
->flags
|= SLAB_RED_ZONE
;
4683 calculate_sizes(s
, -1);
4686 SLAB_ATTR(red_zone
);
4688 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
4690 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
4693 static ssize_t
poison_store(struct kmem_cache
*s
,
4694 const char *buf
, size_t length
)
4696 if (any_slab_objects(s
))
4699 s
->flags
&= ~SLAB_POISON
;
4700 if (buf
[0] == '1') {
4701 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4702 s
->flags
|= SLAB_POISON
;
4704 calculate_sizes(s
, -1);
4709 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
4711 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
4714 static ssize_t
store_user_store(struct kmem_cache
*s
,
4715 const char *buf
, size_t length
)
4717 if (any_slab_objects(s
))
4720 s
->flags
&= ~SLAB_STORE_USER
;
4721 if (buf
[0] == '1') {
4722 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4723 s
->flags
|= SLAB_STORE_USER
;
4725 calculate_sizes(s
, -1);
4728 SLAB_ATTR(store_user
);
4730 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
4735 static ssize_t
validate_store(struct kmem_cache
*s
,
4736 const char *buf
, size_t length
)
4740 if (buf
[0] == '1') {
4741 ret
= validate_slab_cache(s
);
4747 SLAB_ATTR(validate
);
4749 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
4751 if (!(s
->flags
& SLAB_STORE_USER
))
4753 return list_locations(s
, buf
, TRACK_ALLOC
);
4755 SLAB_ATTR_RO(alloc_calls
);
4757 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
4759 if (!(s
->flags
& SLAB_STORE_USER
))
4761 return list_locations(s
, buf
, TRACK_FREE
);
4763 SLAB_ATTR_RO(free_calls
);
4764 #endif /* CONFIG_SLUB_DEBUG */
4766 #ifdef CONFIG_FAILSLAB
4767 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
4769 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
4772 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
4775 s
->flags
&= ~SLAB_FAILSLAB
;
4777 s
->flags
|= SLAB_FAILSLAB
;
4780 SLAB_ATTR(failslab
);
4783 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
4788 static ssize_t
shrink_store(struct kmem_cache
*s
,
4789 const char *buf
, size_t length
)
4791 if (buf
[0] == '1') {
4792 int rc
= kmem_cache_shrink(s
);
4803 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
4805 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
4808 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
4809 const char *buf
, size_t length
)
4811 unsigned long ratio
;
4814 err
= kstrtoul(buf
, 10, &ratio
);
4819 s
->remote_node_defrag_ratio
= ratio
* 10;
4823 SLAB_ATTR(remote_node_defrag_ratio
);
4826 #ifdef CONFIG_SLUB_STATS
4827 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
4829 unsigned long sum
= 0;
4832 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
4837 for_each_online_cpu(cpu
) {
4838 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
4844 len
= sprintf(buf
, "%lu", sum
);
4847 for_each_online_cpu(cpu
) {
4848 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
4849 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
4853 return len
+ sprintf(buf
+ len
, "\n");
4856 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
4860 for_each_online_cpu(cpu
)
4861 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
4864 #define STAT_ATTR(si, text) \
4865 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4867 return show_stat(s, buf, si); \
4869 static ssize_t text##_store(struct kmem_cache *s, \
4870 const char *buf, size_t length) \
4872 if (buf[0] != '0') \
4874 clear_stat(s, si); \
4879 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4880 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4881 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4882 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4883 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4884 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4885 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4886 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4887 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4888 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4889 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
4890 STAT_ATTR(FREE_SLAB
, free_slab
);
4891 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4892 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4893 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4894 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4895 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4896 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4897 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
4898 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
4899 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
4900 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
4901 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
4902 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
4903 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
4904 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
4907 static struct attribute
*slab_attrs
[] = {
4908 &slab_size_attr
.attr
,
4909 &object_size_attr
.attr
,
4910 &objs_per_slab_attr
.attr
,
4912 &min_partial_attr
.attr
,
4913 &cpu_partial_attr
.attr
,
4915 &objects_partial_attr
.attr
,
4917 &cpu_slabs_attr
.attr
,
4921 &hwcache_align_attr
.attr
,
4922 &reclaim_account_attr
.attr
,
4923 &destroy_by_rcu_attr
.attr
,
4925 &reserved_attr
.attr
,
4926 &slabs_cpu_partial_attr
.attr
,
4927 #ifdef CONFIG_SLUB_DEBUG
4928 &total_objects_attr
.attr
,
4930 &sanity_checks_attr
.attr
,
4932 &red_zone_attr
.attr
,
4934 &store_user_attr
.attr
,
4935 &validate_attr
.attr
,
4936 &alloc_calls_attr
.attr
,
4937 &free_calls_attr
.attr
,
4939 #ifdef CONFIG_ZONE_DMA
4940 &cache_dma_attr
.attr
,
4943 &remote_node_defrag_ratio_attr
.attr
,
4945 #ifdef CONFIG_SLUB_STATS
4946 &alloc_fastpath_attr
.attr
,
4947 &alloc_slowpath_attr
.attr
,
4948 &free_fastpath_attr
.attr
,
4949 &free_slowpath_attr
.attr
,
4950 &free_frozen_attr
.attr
,
4951 &free_add_partial_attr
.attr
,
4952 &free_remove_partial_attr
.attr
,
4953 &alloc_from_partial_attr
.attr
,
4954 &alloc_slab_attr
.attr
,
4955 &alloc_refill_attr
.attr
,
4956 &alloc_node_mismatch_attr
.attr
,
4957 &free_slab_attr
.attr
,
4958 &cpuslab_flush_attr
.attr
,
4959 &deactivate_full_attr
.attr
,
4960 &deactivate_empty_attr
.attr
,
4961 &deactivate_to_head_attr
.attr
,
4962 &deactivate_to_tail_attr
.attr
,
4963 &deactivate_remote_frees_attr
.attr
,
4964 &deactivate_bypass_attr
.attr
,
4965 &order_fallback_attr
.attr
,
4966 &cmpxchg_double_fail_attr
.attr
,
4967 &cmpxchg_double_cpu_fail_attr
.attr
,
4968 &cpu_partial_alloc_attr
.attr
,
4969 &cpu_partial_free_attr
.attr
,
4970 &cpu_partial_node_attr
.attr
,
4971 &cpu_partial_drain_attr
.attr
,
4973 #ifdef CONFIG_FAILSLAB
4974 &failslab_attr
.attr
,
4980 static struct attribute_group slab_attr_group
= {
4981 .attrs
= slab_attrs
,
4984 static ssize_t
slab_attr_show(struct kobject
*kobj
,
4985 struct attribute
*attr
,
4988 struct slab_attribute
*attribute
;
4989 struct kmem_cache
*s
;
4992 attribute
= to_slab_attr(attr
);
4995 if (!attribute
->show
)
4998 err
= attribute
->show(s
, buf
);
5003 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5004 struct attribute
*attr
,
5005 const char *buf
, size_t len
)
5007 struct slab_attribute
*attribute
;
5008 struct kmem_cache
*s
;
5011 attribute
= to_slab_attr(attr
);
5014 if (!attribute
->store
)
5017 err
= attribute
->store(s
, buf
, len
);
5018 #ifdef CONFIG_MEMCG_KMEM
5019 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5022 mutex_lock(&slab_mutex
);
5023 if (s
->max_attr_size
< len
)
5024 s
->max_attr_size
= len
;
5027 * This is a best effort propagation, so this function's return
5028 * value will be determined by the parent cache only. This is
5029 * basically because not all attributes will have a well
5030 * defined semantics for rollbacks - most of the actions will
5031 * have permanent effects.
5033 * Returning the error value of any of the children that fail
5034 * is not 100 % defined, in the sense that users seeing the
5035 * error code won't be able to know anything about the state of
5038 * Only returning the error code for the parent cache at least
5039 * has well defined semantics. The cache being written to
5040 * directly either failed or succeeded, in which case we loop
5041 * through the descendants with best-effort propagation.
5043 for_each_memcg_cache_index(i
) {
5044 struct kmem_cache
*c
= cache_from_memcg_idx(s
, i
);
5046 attribute
->store(c
, buf
, len
);
5048 mutex_unlock(&slab_mutex
);
5054 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5056 #ifdef CONFIG_MEMCG_KMEM
5058 char *buffer
= NULL
;
5060 if (!is_root_cache(s
))
5064 * This mean this cache had no attribute written. Therefore, no point
5065 * in copying default values around
5067 if (!s
->max_attr_size
)
5070 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5073 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5075 if (!attr
|| !attr
->store
|| !attr
->show
)
5079 * It is really bad that we have to allocate here, so we will
5080 * do it only as a fallback. If we actually allocate, though,
5081 * we can just use the allocated buffer until the end.
5083 * Most of the slub attributes will tend to be very small in
5084 * size, but sysfs allows buffers up to a page, so they can
5085 * theoretically happen.
5089 else if (s
->max_attr_size
< ARRAY_SIZE(mbuf
))
5092 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5093 if (WARN_ON(!buffer
))
5098 attr
->show(s
->memcg_params
->root_cache
, buf
);
5099 attr
->store(s
, buf
, strlen(buf
));
5103 free_page((unsigned long)buffer
);
5107 static const struct sysfs_ops slab_sysfs_ops
= {
5108 .show
= slab_attr_show
,
5109 .store
= slab_attr_store
,
5112 static struct kobj_type slab_ktype
= {
5113 .sysfs_ops
= &slab_sysfs_ops
,
5116 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5118 struct kobj_type
*ktype
= get_ktype(kobj
);
5120 if (ktype
== &slab_ktype
)
5125 static const struct kset_uevent_ops slab_uevent_ops
= {
5126 .filter
= uevent_filter
,
5129 static struct kset
*slab_kset
;
5131 #define ID_STR_LENGTH 64
5133 /* Create a unique string id for a slab cache:
5135 * Format :[flags-]size
5137 static char *create_unique_id(struct kmem_cache
*s
)
5139 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5146 * First flags affecting slabcache operations. We will only
5147 * get here for aliasable slabs so we do not need to support
5148 * too many flags. The flags here must cover all flags that
5149 * are matched during merging to guarantee that the id is
5152 if (s
->flags
& SLAB_CACHE_DMA
)
5154 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5156 if (s
->flags
& SLAB_DEBUG_FREE
)
5158 if (!(s
->flags
& SLAB_NOTRACK
))
5162 p
+= sprintf(p
, "%07d", s
->size
);
5164 #ifdef CONFIG_MEMCG_KMEM
5165 if (!is_root_cache(s
))
5166 p
+= sprintf(p
, "-%08d",
5167 memcg_cache_id(s
->memcg_params
->memcg
));
5170 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5174 static int sysfs_slab_add(struct kmem_cache
*s
)
5178 int unmergeable
= slab_unmergeable(s
);
5182 * Slabcache can never be merged so we can use the name proper.
5183 * This is typically the case for debug situations. In that
5184 * case we can catch duplicate names easily.
5186 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5190 * Create a unique name for the slab as a target
5193 name
= create_unique_id(s
);
5196 s
->kobj
.kset
= slab_kset
;
5197 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5199 kobject_put(&s
->kobj
);
5203 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5205 kobject_del(&s
->kobj
);
5206 kobject_put(&s
->kobj
);
5209 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5211 /* Setup first alias */
5212 sysfs_slab_alias(s
, s
->name
);
5218 static void sysfs_slab_remove(struct kmem_cache
*s
)
5220 if (slab_state
< FULL
)
5222 * Sysfs has not been setup yet so no need to remove the
5227 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5228 kobject_del(&s
->kobj
);
5229 kobject_put(&s
->kobj
);
5233 * Need to buffer aliases during bootup until sysfs becomes
5234 * available lest we lose that information.
5236 struct saved_alias
{
5237 struct kmem_cache
*s
;
5239 struct saved_alias
*next
;
5242 static struct saved_alias
*alias_list
;
5244 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5246 struct saved_alias
*al
;
5248 if (slab_state
== FULL
) {
5250 * If we have a leftover link then remove it.
5252 sysfs_remove_link(&slab_kset
->kobj
, name
);
5253 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5256 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5262 al
->next
= alias_list
;
5267 static int __init
slab_sysfs_init(void)
5269 struct kmem_cache
*s
;
5272 mutex_lock(&slab_mutex
);
5274 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5276 mutex_unlock(&slab_mutex
);
5277 printk(KERN_ERR
"Cannot register slab subsystem.\n");
5283 list_for_each_entry(s
, &slab_caches
, list
) {
5284 err
= sysfs_slab_add(s
);
5286 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
5287 " to sysfs\n", s
->name
);
5290 while (alias_list
) {
5291 struct saved_alias
*al
= alias_list
;
5293 alias_list
= alias_list
->next
;
5294 err
= sysfs_slab_alias(al
->s
, al
->name
);
5296 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
5297 " %s to sysfs\n", al
->name
);
5301 mutex_unlock(&slab_mutex
);
5306 __initcall(slab_sysfs_init
);
5307 #endif /* CONFIG_SYSFS */
5310 * The /proc/slabinfo ABI
5312 #ifdef CONFIG_SLABINFO
5313 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5315 unsigned long nr_slabs
= 0;
5316 unsigned long nr_objs
= 0;
5317 unsigned long nr_free
= 0;
5320 for_each_online_node(node
) {
5321 struct kmem_cache_node
*n
= get_node(s
, node
);
5326 nr_slabs
+= node_nr_slabs(n
);
5327 nr_objs
+= node_nr_objs(n
);
5328 nr_free
+= count_partial(n
, count_free
);
5331 sinfo
->active_objs
= nr_objs
- nr_free
;
5332 sinfo
->num_objs
= nr_objs
;
5333 sinfo
->active_slabs
= nr_slabs
;
5334 sinfo
->num_slabs
= nr_slabs
;
5335 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5336 sinfo
->cache_order
= oo_order(s
->oo
);
5339 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5343 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5344 size_t count
, loff_t
*ppos
)
5348 #endif /* CONFIG_SLABINFO */