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 lockdep_assert_held(&n
->list_lock
);
1009 if (!(s
->flags
& SLAB_STORE_USER
))
1012 list_add(&page
->lru
, &n
->full
);
1015 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1017 lockdep_assert_held(&n
->list_lock
);
1019 if (!(s
->flags
& SLAB_STORE_USER
))
1022 list_del(&page
->lru
);
1025 /* Tracking of the number of slabs for debugging purposes */
1026 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1028 struct kmem_cache_node
*n
= get_node(s
, node
);
1030 return atomic_long_read(&n
->nr_slabs
);
1033 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1035 return atomic_long_read(&n
->nr_slabs
);
1038 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1040 struct kmem_cache_node
*n
= get_node(s
, node
);
1043 * May be called early in order to allocate a slab for the
1044 * kmem_cache_node structure. Solve the chicken-egg
1045 * dilemma by deferring the increment of the count during
1046 * bootstrap (see early_kmem_cache_node_alloc).
1049 atomic_long_inc(&n
->nr_slabs
);
1050 atomic_long_add(objects
, &n
->total_objects
);
1053 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1055 struct kmem_cache_node
*n
= get_node(s
, node
);
1057 atomic_long_dec(&n
->nr_slabs
);
1058 atomic_long_sub(objects
, &n
->total_objects
);
1061 /* Object debug checks for alloc/free paths */
1062 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1065 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1068 init_object(s
, object
, SLUB_RED_INACTIVE
);
1069 init_tracking(s
, object
);
1072 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1074 void *object
, unsigned long addr
)
1076 if (!check_slab(s
, page
))
1079 if (!check_valid_pointer(s
, page
, object
)) {
1080 object_err(s
, page
, object
, "Freelist Pointer check fails");
1084 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1087 /* Success perform special debug activities for allocs */
1088 if (s
->flags
& SLAB_STORE_USER
)
1089 set_track(s
, object
, TRACK_ALLOC
, addr
);
1090 trace(s
, page
, object
, 1);
1091 init_object(s
, object
, SLUB_RED_ACTIVE
);
1095 if (PageSlab(page
)) {
1097 * If this is a slab page then lets do the best we can
1098 * to avoid issues in the future. Marking all objects
1099 * as used avoids touching the remaining objects.
1101 slab_fix(s
, "Marking all objects used");
1102 page
->inuse
= page
->objects
;
1103 page
->freelist
= NULL
;
1108 static noinline
struct kmem_cache_node
*free_debug_processing(
1109 struct kmem_cache
*s
, struct page
*page
, void *object
,
1110 unsigned long addr
, unsigned long *flags
)
1112 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1114 spin_lock_irqsave(&n
->list_lock
, *flags
);
1117 if (!check_slab(s
, page
))
1120 if (!check_valid_pointer(s
, page
, object
)) {
1121 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1125 if (on_freelist(s
, page
, object
)) {
1126 object_err(s
, page
, object
, "Object already free");
1130 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1133 if (unlikely(s
!= page
->slab_cache
)) {
1134 if (!PageSlab(page
)) {
1135 slab_err(s
, page
, "Attempt to free object(0x%p) "
1136 "outside of slab", object
);
1137 } else if (!page
->slab_cache
) {
1139 "SLUB <none>: no slab for object 0x%p.\n",
1143 object_err(s
, page
, object
,
1144 "page slab pointer corrupt.");
1148 if (s
->flags
& SLAB_STORE_USER
)
1149 set_track(s
, object
, TRACK_FREE
, addr
);
1150 trace(s
, page
, object
, 0);
1151 init_object(s
, object
, SLUB_RED_INACTIVE
);
1155 * Keep node_lock to preserve integrity
1156 * until the object is actually freed
1162 spin_unlock_irqrestore(&n
->list_lock
, *flags
);
1163 slab_fix(s
, "Object at 0x%p not freed", object
);
1167 static int __init
setup_slub_debug(char *str
)
1169 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1170 if (*str
++ != '=' || !*str
)
1172 * No options specified. Switch on full debugging.
1178 * No options but restriction on slabs. This means full
1179 * debugging for slabs matching a pattern.
1183 if (tolower(*str
) == 'o') {
1185 * Avoid enabling debugging on caches if its minimum order
1186 * would increase as a result.
1188 disable_higher_order_debug
= 1;
1195 * Switch off all debugging measures.
1200 * Determine which debug features should be switched on
1202 for (; *str
&& *str
!= ','; str
++) {
1203 switch (tolower(*str
)) {
1205 slub_debug
|= SLAB_DEBUG_FREE
;
1208 slub_debug
|= SLAB_RED_ZONE
;
1211 slub_debug
|= SLAB_POISON
;
1214 slub_debug
|= SLAB_STORE_USER
;
1217 slub_debug
|= SLAB_TRACE
;
1220 slub_debug
|= SLAB_FAILSLAB
;
1223 printk(KERN_ERR
"slub_debug option '%c' "
1224 "unknown. skipped\n", *str
);
1230 slub_debug_slabs
= str
+ 1;
1235 __setup("slub_debug", setup_slub_debug
);
1237 static unsigned long kmem_cache_flags(unsigned long object_size
,
1238 unsigned long flags
, const char *name
,
1239 void (*ctor
)(void *))
1242 * Enable debugging if selected on the kernel commandline.
1244 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1245 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1246 flags
|= slub_debug
;
1251 static inline void setup_object_debug(struct kmem_cache
*s
,
1252 struct page
*page
, void *object
) {}
1254 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1255 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1257 static inline struct kmem_cache_node
*free_debug_processing(
1258 struct kmem_cache
*s
, struct page
*page
, void *object
,
1259 unsigned long addr
, unsigned long *flags
) { return NULL
; }
1261 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1263 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1264 void *object
, u8 val
) { return 1; }
1265 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1266 struct page
*page
) {}
1267 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1268 struct page
*page
) {}
1269 static inline unsigned long kmem_cache_flags(unsigned long object_size
,
1270 unsigned long flags
, const char *name
,
1271 void (*ctor
)(void *))
1275 #define slub_debug 0
1277 #define disable_higher_order_debug 0
1279 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1281 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1283 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1285 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1288 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1290 kmemleak_alloc(ptr
, size
, 1, flags
);
1293 static inline void kfree_hook(const void *x
)
1298 static inline int slab_pre_alloc_hook(struct kmem_cache
*s
, gfp_t flags
)
1301 static inline void slab_post_alloc_hook(struct kmem_cache
*s
, gfp_t flags
,
1304 kmemleak_alloc_recursive(object
, s
->object_size
, 1, s
->flags
,
1305 flags
& gfp_allowed_mask
);
1308 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
1310 kmemleak_free_recursive(x
, s
->flags
);
1313 #endif /* CONFIG_SLUB_DEBUG */
1316 * Slab allocation and freeing
1318 static inline struct page
*alloc_slab_page(gfp_t flags
, int node
,
1319 struct kmem_cache_order_objects oo
)
1321 int order
= oo_order(oo
);
1323 flags
|= __GFP_NOTRACK
;
1325 if (node
== NUMA_NO_NODE
)
1326 return alloc_pages(flags
, order
);
1328 return alloc_pages_exact_node(node
, flags
, order
);
1331 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1334 struct kmem_cache_order_objects oo
= s
->oo
;
1337 flags
&= gfp_allowed_mask
;
1339 if (flags
& __GFP_WAIT
)
1342 flags
|= s
->allocflags
;
1345 * Let the initial higher-order allocation fail under memory pressure
1346 * so we fall-back to the minimum order allocation.
1348 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1350 page
= alloc_slab_page(alloc_gfp
, node
, oo
);
1351 if (unlikely(!page
)) {
1354 * Allocation may have failed due to fragmentation.
1355 * Try a lower order alloc if possible
1357 page
= alloc_slab_page(flags
, node
, oo
);
1360 stat(s
, ORDER_FALLBACK
);
1363 if (kmemcheck_enabled
&& page
1364 && !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1365 int pages
= 1 << oo_order(oo
);
1367 kmemcheck_alloc_shadow(page
, oo_order(oo
), flags
, node
);
1370 * Objects from caches that have a constructor don't get
1371 * cleared when they're allocated, so we need to do it here.
1374 kmemcheck_mark_uninitialized_pages(page
, pages
);
1376 kmemcheck_mark_unallocated_pages(page
, pages
);
1379 if (flags
& __GFP_WAIT
)
1380 local_irq_disable();
1384 page
->objects
= oo_objects(oo
);
1385 mod_zone_page_state(page_zone(page
),
1386 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1387 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1393 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1396 setup_object_debug(s
, page
, object
);
1397 if (unlikely(s
->ctor
))
1401 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1409 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1411 page
= allocate_slab(s
,
1412 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1416 order
= compound_order(page
);
1417 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1418 memcg_bind_pages(s
, order
);
1419 page
->slab_cache
= s
;
1420 __SetPageSlab(page
);
1421 if (page
->pfmemalloc
)
1422 SetPageSlabPfmemalloc(page
);
1424 start
= page_address(page
);
1426 if (unlikely(s
->flags
& SLAB_POISON
))
1427 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1430 for_each_object(p
, s
, start
, page
->objects
) {
1431 setup_object(s
, page
, last
);
1432 set_freepointer(s
, last
, p
);
1435 setup_object(s
, page
, last
);
1436 set_freepointer(s
, last
, NULL
);
1438 page
->freelist
= start
;
1439 page
->inuse
= page
->objects
;
1445 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1447 int order
= compound_order(page
);
1448 int pages
= 1 << order
;
1450 if (kmem_cache_debug(s
)) {
1453 slab_pad_check(s
, page
);
1454 for_each_object(p
, s
, page_address(page
),
1456 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1459 kmemcheck_free_shadow(page
, compound_order(page
));
1461 mod_zone_page_state(page_zone(page
),
1462 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1463 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1466 __ClearPageSlabPfmemalloc(page
);
1467 __ClearPageSlab(page
);
1469 memcg_release_pages(s
, order
);
1470 page_mapcount_reset(page
);
1471 if (current
->reclaim_state
)
1472 current
->reclaim_state
->reclaimed_slab
+= pages
;
1473 __free_memcg_kmem_pages(page
, order
);
1476 #define need_reserve_slab_rcu \
1477 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1479 static void rcu_free_slab(struct rcu_head
*h
)
1483 if (need_reserve_slab_rcu
)
1484 page
= virt_to_head_page(h
);
1486 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1488 __free_slab(page
->slab_cache
, page
);
1491 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1493 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1494 struct rcu_head
*head
;
1496 if (need_reserve_slab_rcu
) {
1497 int order
= compound_order(page
);
1498 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1500 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1501 head
= page_address(page
) + offset
;
1504 * RCU free overloads the RCU head over the LRU
1506 head
= (void *)&page
->lru
;
1509 call_rcu(head
, rcu_free_slab
);
1511 __free_slab(s
, page
);
1514 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1516 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1521 * Management of partially allocated slabs.
1523 static inline void add_partial(struct kmem_cache_node
*n
,
1524 struct page
*page
, int tail
)
1526 lockdep_assert_held(&n
->list_lock
);
1529 if (tail
== DEACTIVATE_TO_TAIL
)
1530 list_add_tail(&page
->lru
, &n
->partial
);
1532 list_add(&page
->lru
, &n
->partial
);
1535 static inline void remove_partial(struct kmem_cache_node
*n
,
1538 lockdep_assert_held(&n
->list_lock
);
1540 list_del(&page
->lru
);
1545 * Remove slab from the partial list, freeze it and
1546 * return the pointer to the freelist.
1548 * Returns a list of objects or NULL if it fails.
1550 static inline void *acquire_slab(struct kmem_cache
*s
,
1551 struct kmem_cache_node
*n
, struct page
*page
,
1552 int mode
, int *objects
)
1555 unsigned long counters
;
1558 lockdep_assert_held(&n
->list_lock
);
1561 * Zap the freelist and set the frozen bit.
1562 * The old freelist is the list of objects for the
1563 * per cpu allocation list.
1565 freelist
= page
->freelist
;
1566 counters
= page
->counters
;
1567 new.counters
= counters
;
1568 *objects
= new.objects
- new.inuse
;
1570 new.inuse
= page
->objects
;
1571 new.freelist
= NULL
;
1573 new.freelist
= freelist
;
1576 VM_BUG_ON(new.frozen
);
1579 if (!__cmpxchg_double_slab(s
, page
,
1581 new.freelist
, new.counters
,
1585 remove_partial(n
, page
);
1590 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1591 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1594 * Try to allocate a partial slab from a specific node.
1596 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1597 struct kmem_cache_cpu
*c
, gfp_t flags
)
1599 struct page
*page
, *page2
;
1600 void *object
= NULL
;
1605 * Racy check. If we mistakenly see no partial slabs then we
1606 * just allocate an empty slab. If we mistakenly try to get a
1607 * partial slab and there is none available then get_partials()
1610 if (!n
|| !n
->nr_partial
)
1613 spin_lock(&n
->list_lock
);
1614 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1617 if (!pfmemalloc_match(page
, flags
))
1620 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1624 available
+= objects
;
1627 stat(s
, ALLOC_FROM_PARTIAL
);
1630 put_cpu_partial(s
, page
, 0);
1631 stat(s
, CPU_PARTIAL_NODE
);
1633 if (!kmem_cache_has_cpu_partial(s
)
1634 || available
> s
->cpu_partial
/ 2)
1638 spin_unlock(&n
->list_lock
);
1643 * Get a page from somewhere. Search in increasing NUMA distances.
1645 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1646 struct kmem_cache_cpu
*c
)
1649 struct zonelist
*zonelist
;
1652 enum zone_type high_zoneidx
= gfp_zone(flags
);
1654 unsigned int cpuset_mems_cookie
;
1657 * The defrag ratio allows a configuration of the tradeoffs between
1658 * inter node defragmentation and node local allocations. A lower
1659 * defrag_ratio increases the tendency to do local allocations
1660 * instead of attempting to obtain partial slabs from other nodes.
1662 * If the defrag_ratio is set to 0 then kmalloc() always
1663 * returns node local objects. If the ratio is higher then kmalloc()
1664 * may return off node objects because partial slabs are obtained
1665 * from other nodes and filled up.
1667 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1668 * defrag_ratio = 1000) then every (well almost) allocation will
1669 * first attempt to defrag slab caches on other nodes. This means
1670 * scanning over all nodes to look for partial slabs which may be
1671 * expensive if we do it every time we are trying to find a slab
1672 * with available objects.
1674 if (!s
->remote_node_defrag_ratio
||
1675 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1679 cpuset_mems_cookie
= get_mems_allowed();
1680 zonelist
= node_zonelist(slab_node(), flags
);
1681 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1682 struct kmem_cache_node
*n
;
1684 n
= get_node(s
, zone_to_nid(zone
));
1686 if (n
&& cpuset_zone_allowed_hardwall(zone
, flags
) &&
1687 n
->nr_partial
> s
->min_partial
) {
1688 object
= get_partial_node(s
, n
, c
, flags
);
1691 * Return the object even if
1692 * put_mems_allowed indicated that
1693 * the cpuset mems_allowed was
1694 * updated in parallel. It's a
1695 * harmless race between the alloc
1696 * and the cpuset update.
1698 put_mems_allowed(cpuset_mems_cookie
);
1703 } while (!put_mems_allowed(cpuset_mems_cookie
));
1709 * Get a partial page, lock it and return it.
1711 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1712 struct kmem_cache_cpu
*c
)
1715 int searchnode
= (node
== NUMA_NO_NODE
) ? numa_node_id() : node
;
1717 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1718 if (object
|| node
!= NUMA_NO_NODE
)
1721 return get_any_partial(s
, flags
, c
);
1724 #ifdef CONFIG_PREEMPT
1726 * Calculate the next globally unique transaction for disambiguiation
1727 * during cmpxchg. The transactions start with the cpu number and are then
1728 * incremented by CONFIG_NR_CPUS.
1730 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1733 * No preemption supported therefore also no need to check for
1739 static inline unsigned long next_tid(unsigned long tid
)
1741 return tid
+ TID_STEP
;
1744 static inline unsigned int tid_to_cpu(unsigned long tid
)
1746 return tid
% TID_STEP
;
1749 static inline unsigned long tid_to_event(unsigned long tid
)
1751 return tid
/ TID_STEP
;
1754 static inline unsigned int init_tid(int cpu
)
1759 static inline void note_cmpxchg_failure(const char *n
,
1760 const struct kmem_cache
*s
, unsigned long tid
)
1762 #ifdef SLUB_DEBUG_CMPXCHG
1763 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1765 printk(KERN_INFO
"%s %s: cmpxchg redo ", n
, s
->name
);
1767 #ifdef CONFIG_PREEMPT
1768 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1769 printk("due to cpu change %d -> %d\n",
1770 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1773 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1774 printk("due to cpu running other code. Event %ld->%ld\n",
1775 tid_to_event(tid
), tid_to_event(actual_tid
));
1777 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1778 actual_tid
, tid
, next_tid(tid
));
1780 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1783 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1787 for_each_possible_cpu(cpu
)
1788 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1792 * Remove the cpu slab
1794 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1797 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1798 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1800 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1802 int tail
= DEACTIVATE_TO_HEAD
;
1806 if (page
->freelist
) {
1807 stat(s
, DEACTIVATE_REMOTE_FREES
);
1808 tail
= DEACTIVATE_TO_TAIL
;
1812 * Stage one: Free all available per cpu objects back
1813 * to the page freelist while it is still frozen. Leave the
1816 * There is no need to take the list->lock because the page
1819 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
1821 unsigned long counters
;
1824 prior
= page
->freelist
;
1825 counters
= page
->counters
;
1826 set_freepointer(s
, freelist
, prior
);
1827 new.counters
= counters
;
1829 VM_BUG_ON(!new.frozen
);
1831 } while (!__cmpxchg_double_slab(s
, page
,
1833 freelist
, new.counters
,
1834 "drain percpu freelist"));
1836 freelist
= nextfree
;
1840 * Stage two: Ensure that the page is unfrozen while the
1841 * list presence reflects the actual number of objects
1844 * We setup the list membership and then perform a cmpxchg
1845 * with the count. If there is a mismatch then the page
1846 * is not unfrozen but the page is on the wrong list.
1848 * Then we restart the process which may have to remove
1849 * the page from the list that we just put it on again
1850 * because the number of objects in the slab may have
1855 old
.freelist
= page
->freelist
;
1856 old
.counters
= page
->counters
;
1857 VM_BUG_ON(!old
.frozen
);
1859 /* Determine target state of the slab */
1860 new.counters
= old
.counters
;
1863 set_freepointer(s
, freelist
, old
.freelist
);
1864 new.freelist
= freelist
;
1866 new.freelist
= old
.freelist
;
1870 if (!new.inuse
&& n
->nr_partial
> s
->min_partial
)
1872 else if (new.freelist
) {
1877 * Taking the spinlock removes the possiblity
1878 * that acquire_slab() will see a slab page that
1881 spin_lock(&n
->list_lock
);
1885 if (kmem_cache_debug(s
) && !lock
) {
1888 * This also ensures that the scanning of full
1889 * slabs from diagnostic functions will not see
1892 spin_lock(&n
->list_lock
);
1900 remove_partial(n
, page
);
1902 else if (l
== M_FULL
)
1904 remove_full(s
, n
, page
);
1906 if (m
== M_PARTIAL
) {
1908 add_partial(n
, page
, tail
);
1911 } else if (m
== M_FULL
) {
1913 stat(s
, DEACTIVATE_FULL
);
1914 add_full(s
, n
, page
);
1920 if (!__cmpxchg_double_slab(s
, page
,
1921 old
.freelist
, old
.counters
,
1922 new.freelist
, new.counters
,
1927 spin_unlock(&n
->list_lock
);
1930 stat(s
, DEACTIVATE_EMPTY
);
1931 discard_slab(s
, page
);
1937 * Unfreeze all the cpu partial slabs.
1939 * This function must be called with interrupts disabled
1940 * for the cpu using c (or some other guarantee must be there
1941 * to guarantee no concurrent accesses).
1943 static void unfreeze_partials(struct kmem_cache
*s
,
1944 struct kmem_cache_cpu
*c
)
1946 #ifdef CONFIG_SLUB_CPU_PARTIAL
1947 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
1948 struct page
*page
, *discard_page
= NULL
;
1950 while ((page
= c
->partial
)) {
1954 c
->partial
= page
->next
;
1956 n2
= get_node(s
, page_to_nid(page
));
1959 spin_unlock(&n
->list_lock
);
1962 spin_lock(&n
->list_lock
);
1967 old
.freelist
= page
->freelist
;
1968 old
.counters
= page
->counters
;
1969 VM_BUG_ON(!old
.frozen
);
1971 new.counters
= old
.counters
;
1972 new.freelist
= old
.freelist
;
1976 } while (!__cmpxchg_double_slab(s
, page
,
1977 old
.freelist
, old
.counters
,
1978 new.freelist
, new.counters
,
1979 "unfreezing slab"));
1981 if (unlikely(!new.inuse
&& n
->nr_partial
> s
->min_partial
)) {
1982 page
->next
= discard_page
;
1983 discard_page
= page
;
1985 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
1986 stat(s
, FREE_ADD_PARTIAL
);
1991 spin_unlock(&n
->list_lock
);
1993 while (discard_page
) {
1994 page
= discard_page
;
1995 discard_page
= discard_page
->next
;
1997 stat(s
, DEACTIVATE_EMPTY
);
1998 discard_slab(s
, page
);
2005 * Put a page that was just frozen (in __slab_free) into a partial page
2006 * slot if available. This is done without interrupts disabled and without
2007 * preemption disabled. The cmpxchg is racy and may put the partial page
2008 * onto a random cpus partial slot.
2010 * If we did not find a slot then simply move all the partials to the
2011 * per node partial list.
2013 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2015 #ifdef CONFIG_SLUB_CPU_PARTIAL
2016 struct page
*oldpage
;
2023 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2026 pobjects
= oldpage
->pobjects
;
2027 pages
= oldpage
->pages
;
2028 if (drain
&& pobjects
> s
->cpu_partial
) {
2029 unsigned long flags
;
2031 * partial array is full. Move the existing
2032 * set to the per node partial list.
2034 local_irq_save(flags
);
2035 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2036 local_irq_restore(flags
);
2040 stat(s
, CPU_PARTIAL_DRAIN
);
2045 pobjects
+= page
->objects
- page
->inuse
;
2047 page
->pages
= pages
;
2048 page
->pobjects
= pobjects
;
2049 page
->next
= oldpage
;
2051 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2056 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2058 stat(s
, CPUSLAB_FLUSH
);
2059 deactivate_slab(s
, c
->page
, c
->freelist
);
2061 c
->tid
= next_tid(c
->tid
);
2069 * Called from IPI handler with interrupts disabled.
2071 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2073 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2079 unfreeze_partials(s
, c
);
2083 static void flush_cpu_slab(void *d
)
2085 struct kmem_cache
*s
= d
;
2087 __flush_cpu_slab(s
, smp_processor_id());
2090 static bool has_cpu_slab(int cpu
, void *info
)
2092 struct kmem_cache
*s
= info
;
2093 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2095 return c
->page
|| c
->partial
;
2098 static void flush_all(struct kmem_cache
*s
)
2100 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2104 * Check if the objects in a per cpu structure fit numa
2105 * locality expectations.
2107 static inline int node_match(struct page
*page
, int node
)
2110 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2116 static int count_free(struct page
*page
)
2118 return page
->objects
- page
->inuse
;
2121 static unsigned long count_partial(struct kmem_cache_node
*n
,
2122 int (*get_count
)(struct page
*))
2124 unsigned long flags
;
2125 unsigned long x
= 0;
2128 spin_lock_irqsave(&n
->list_lock
, flags
);
2129 list_for_each_entry(page
, &n
->partial
, lru
)
2130 x
+= get_count(page
);
2131 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2135 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2137 #ifdef CONFIG_SLUB_DEBUG
2138 return atomic_long_read(&n
->total_objects
);
2144 static noinline
void
2145 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2150 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2152 printk(KERN_WARNING
" cache: %s, object size: %d, buffer size: %d, "
2153 "default order: %d, min order: %d\n", s
->name
, s
->object_size
,
2154 s
->size
, oo_order(s
->oo
), oo_order(s
->min
));
2156 if (oo_order(s
->min
) > get_order(s
->object_size
))
2157 printk(KERN_WARNING
" %s debugging increased min order, use "
2158 "slub_debug=O to disable.\n", s
->name
);
2160 for_each_online_node(node
) {
2161 struct kmem_cache_node
*n
= get_node(s
, node
);
2162 unsigned long nr_slabs
;
2163 unsigned long nr_objs
;
2164 unsigned long nr_free
;
2169 nr_free
= count_partial(n
, count_free
);
2170 nr_slabs
= node_nr_slabs(n
);
2171 nr_objs
= node_nr_objs(n
);
2174 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2175 node
, nr_slabs
, nr_objs
, nr_free
);
2179 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2180 int node
, struct kmem_cache_cpu
**pc
)
2183 struct kmem_cache_cpu
*c
= *pc
;
2186 freelist
= get_partial(s
, flags
, node
, c
);
2191 page
= new_slab(s
, flags
, node
);
2193 c
= __this_cpu_ptr(s
->cpu_slab
);
2198 * No other reference to the page yet so we can
2199 * muck around with it freely without cmpxchg
2201 freelist
= page
->freelist
;
2202 page
->freelist
= NULL
;
2204 stat(s
, ALLOC_SLAB
);
2213 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2215 if (unlikely(PageSlabPfmemalloc(page
)))
2216 return gfp_pfmemalloc_allowed(gfpflags
);
2222 * Check the page->freelist of a page and either transfer the freelist to the
2223 * per cpu freelist or deactivate the page.
2225 * The page is still frozen if the return value is not NULL.
2227 * If this function returns NULL then the page has been unfrozen.
2229 * This function must be called with interrupt disabled.
2231 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2234 unsigned long counters
;
2238 freelist
= page
->freelist
;
2239 counters
= page
->counters
;
2241 new.counters
= counters
;
2242 VM_BUG_ON(!new.frozen
);
2244 new.inuse
= page
->objects
;
2245 new.frozen
= freelist
!= NULL
;
2247 } while (!__cmpxchg_double_slab(s
, page
,
2256 * Slow path. The lockless freelist is empty or we need to perform
2259 * Processing is still very fast if new objects have been freed to the
2260 * regular freelist. In that case we simply take over the regular freelist
2261 * as the lockless freelist and zap the regular freelist.
2263 * If that is not working then we fall back to the partial lists. We take the
2264 * first element of the freelist as the object to allocate now and move the
2265 * rest of the freelist to the lockless freelist.
2267 * And if we were unable to get a new slab from the partial slab lists then
2268 * we need to allocate a new slab. This is the slowest path since it involves
2269 * a call to the page allocator and the setup of a new slab.
2271 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2272 unsigned long addr
, struct kmem_cache_cpu
*c
)
2276 unsigned long flags
;
2278 local_irq_save(flags
);
2279 #ifdef CONFIG_PREEMPT
2281 * We may have been preempted and rescheduled on a different
2282 * cpu before disabling interrupts. Need to reload cpu area
2285 c
= this_cpu_ptr(s
->cpu_slab
);
2293 if (unlikely(!node_match(page
, node
))) {
2294 stat(s
, ALLOC_NODE_MISMATCH
);
2295 deactivate_slab(s
, page
, c
->freelist
);
2302 * By rights, we should be searching for a slab page that was
2303 * PFMEMALLOC but right now, we are losing the pfmemalloc
2304 * information when the page leaves the per-cpu allocator
2306 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2307 deactivate_slab(s
, page
, c
->freelist
);
2313 /* must check again c->freelist in case of cpu migration or IRQ */
2314 freelist
= c
->freelist
;
2318 stat(s
, ALLOC_SLOWPATH
);
2320 freelist
= get_freelist(s
, page
);
2324 stat(s
, DEACTIVATE_BYPASS
);
2328 stat(s
, ALLOC_REFILL
);
2332 * freelist is pointing to the list of objects to be used.
2333 * page is pointing to the page from which the objects are obtained.
2334 * That page must be frozen for per cpu allocations to work.
2336 VM_BUG_ON(!c
->page
->frozen
);
2337 c
->freelist
= get_freepointer(s
, freelist
);
2338 c
->tid
= next_tid(c
->tid
);
2339 local_irq_restore(flags
);
2345 page
= c
->page
= c
->partial
;
2346 c
->partial
= page
->next
;
2347 stat(s
, CPU_PARTIAL_ALLOC
);
2352 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2354 if (unlikely(!freelist
)) {
2355 if (!(gfpflags
& __GFP_NOWARN
) && printk_ratelimit())
2356 slab_out_of_memory(s
, gfpflags
, node
);
2358 local_irq_restore(flags
);
2363 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2366 /* Only entered in the debug case */
2367 if (kmem_cache_debug(s
) &&
2368 !alloc_debug_processing(s
, page
, freelist
, addr
))
2369 goto new_slab
; /* Slab failed checks. Next slab needed */
2371 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
2374 local_irq_restore(flags
);
2379 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2380 * have the fastpath folded into their functions. So no function call
2381 * overhead for requests that can be satisfied on the fastpath.
2383 * The fastpath works by first checking if the lockless freelist can be used.
2384 * If not then __slab_alloc is called for slow processing.
2386 * Otherwise we can simply pick the next object from the lockless free list.
2388 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2389 gfp_t gfpflags
, int node
, unsigned long addr
)
2392 struct kmem_cache_cpu
*c
;
2396 if (slab_pre_alloc_hook(s
, gfpflags
))
2399 s
= memcg_kmem_get_cache(s
, gfpflags
);
2402 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2403 * enabled. We may switch back and forth between cpus while
2404 * reading from one cpu area. That does not matter as long
2405 * as we end up on the original cpu again when doing the cmpxchg.
2407 * Preemption is disabled for the retrieval of the tid because that
2408 * must occur from the current processor. We cannot allow rescheduling
2409 * on a different processor between the determination of the pointer
2410 * and the retrieval of the tid.
2413 c
= __this_cpu_ptr(s
->cpu_slab
);
2416 * The transaction ids are globally unique per cpu and per operation on
2417 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2418 * occurs on the right processor and that there was no operation on the
2419 * linked list in between.
2424 object
= c
->freelist
;
2426 if (unlikely(!object
|| !node_match(page
, node
)))
2427 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2430 void *next_object
= get_freepointer_safe(s
, object
);
2433 * The cmpxchg will only match if there was no additional
2434 * operation and if we are on the right processor.
2436 * The cmpxchg does the following atomically (without lock
2438 * 1. Relocate first pointer to the current per cpu area.
2439 * 2. Verify that tid and freelist have not been changed
2440 * 3. If they were not changed replace tid and freelist
2442 * Since this is without lock semantics the protection is only
2443 * against code executing on this cpu *not* from access by
2446 if (unlikely(!this_cpu_cmpxchg_double(
2447 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2449 next_object
, next_tid(tid
)))) {
2451 note_cmpxchg_failure("slab_alloc", s
, tid
);
2454 prefetch_freepointer(s
, next_object
);
2455 stat(s
, ALLOC_FASTPATH
);
2458 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2459 memset(object
, 0, s
->object_size
);
2461 slab_post_alloc_hook(s
, gfpflags
, object
);
2466 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2467 gfp_t gfpflags
, unsigned long addr
)
2469 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2472 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2474 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2476 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2481 EXPORT_SYMBOL(kmem_cache_alloc
);
2483 #ifdef CONFIG_TRACING
2484 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2486 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2487 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2490 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2494 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2496 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2498 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2499 s
->object_size
, s
->size
, gfpflags
, node
);
2503 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2505 #ifdef CONFIG_TRACING
2506 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2508 int node
, size_t size
)
2510 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2512 trace_kmalloc_node(_RET_IP_
, ret
,
2513 size
, s
->size
, gfpflags
, node
);
2516 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2521 * Slow patch handling. This may still be called frequently since objects
2522 * have a longer lifetime than the cpu slabs in most processing loads.
2524 * So we still attempt to reduce cache line usage. Just take the slab
2525 * lock and free the item. If there is no additional partial page
2526 * handling required then we can return immediately.
2528 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2529 void *x
, unsigned long addr
)
2532 void **object
= (void *)x
;
2535 unsigned long counters
;
2536 struct kmem_cache_node
*n
= NULL
;
2537 unsigned long uninitialized_var(flags
);
2539 stat(s
, FREE_SLOWPATH
);
2541 if (kmem_cache_debug(s
) &&
2542 !(n
= free_debug_processing(s
, page
, x
, addr
, &flags
)))
2547 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2550 prior
= page
->freelist
;
2551 counters
= page
->counters
;
2552 set_freepointer(s
, object
, prior
);
2553 new.counters
= counters
;
2554 was_frozen
= new.frozen
;
2556 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2558 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2561 * Slab was on no list before and will be
2563 * We can defer the list move and instead
2568 } else { /* Needs to be taken off a list */
2570 n
= get_node(s
, page_to_nid(page
));
2572 * Speculatively acquire the list_lock.
2573 * If the cmpxchg does not succeed then we may
2574 * drop the list_lock without any processing.
2576 * Otherwise the list_lock will synchronize with
2577 * other processors updating the list of slabs.
2579 spin_lock_irqsave(&n
->list_lock
, flags
);
2584 } while (!cmpxchg_double_slab(s
, page
,
2586 object
, new.counters
,
2592 * If we just froze the page then put it onto the
2593 * per cpu partial list.
2595 if (new.frozen
&& !was_frozen
) {
2596 put_cpu_partial(s
, page
, 1);
2597 stat(s
, CPU_PARTIAL_FREE
);
2600 * The list lock was not taken therefore no list
2601 * activity can be necessary.
2604 stat(s
, FREE_FROZEN
);
2608 if (unlikely(!new.inuse
&& n
->nr_partial
> s
->min_partial
))
2612 * Objects left in the slab. If it was not on the partial list before
2615 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2616 if (kmem_cache_debug(s
))
2617 remove_full(s
, n
, page
);
2618 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2619 stat(s
, FREE_ADD_PARTIAL
);
2621 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2627 * Slab on the partial list.
2629 remove_partial(n
, page
);
2630 stat(s
, FREE_REMOVE_PARTIAL
);
2632 /* Slab must be on the full list */
2633 remove_full(s
, n
, page
);
2636 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2638 discard_slab(s
, page
);
2642 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2643 * can perform fastpath freeing without additional function calls.
2645 * The fastpath is only possible if we are freeing to the current cpu slab
2646 * of this processor. This typically the case if we have just allocated
2649 * If fastpath is not possible then fall back to __slab_free where we deal
2650 * with all sorts of special processing.
2652 static __always_inline
void slab_free(struct kmem_cache
*s
,
2653 struct page
*page
, void *x
, unsigned long addr
)
2655 void **object
= (void *)x
;
2656 struct kmem_cache_cpu
*c
;
2659 slab_free_hook(s
, x
);
2663 * Determine the currently cpus per cpu slab.
2664 * The cpu may change afterward. However that does not matter since
2665 * data is retrieved via this pointer. If we are on the same cpu
2666 * during the cmpxchg then the free will succedd.
2669 c
= __this_cpu_ptr(s
->cpu_slab
);
2674 if (likely(page
== c
->page
)) {
2675 set_freepointer(s
, object
, c
->freelist
);
2677 if (unlikely(!this_cpu_cmpxchg_double(
2678 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2680 object
, next_tid(tid
)))) {
2682 note_cmpxchg_failure("slab_free", s
, tid
);
2685 stat(s
, FREE_FASTPATH
);
2687 __slab_free(s
, page
, x
, addr
);
2691 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2693 s
= cache_from_obj(s
, x
);
2696 slab_free(s
, virt_to_head_page(x
), x
, _RET_IP_
);
2697 trace_kmem_cache_free(_RET_IP_
, x
);
2699 EXPORT_SYMBOL(kmem_cache_free
);
2702 * Object placement in a slab is made very easy because we always start at
2703 * offset 0. If we tune the size of the object to the alignment then we can
2704 * get the required alignment by putting one properly sized object after
2707 * Notice that the allocation order determines the sizes of the per cpu
2708 * caches. Each processor has always one slab available for allocations.
2709 * Increasing the allocation order reduces the number of times that slabs
2710 * must be moved on and off the partial lists and is therefore a factor in
2715 * Mininum / Maximum order of slab pages. This influences locking overhead
2716 * and slab fragmentation. A higher order reduces the number of partial slabs
2717 * and increases the number of allocations possible without having to
2718 * take the list_lock.
2720 static int slub_min_order
;
2721 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
2722 static int slub_min_objects
;
2725 * Merge control. If this is set then no merging of slab caches will occur.
2726 * (Could be removed. This was introduced to pacify the merge skeptics.)
2728 static int slub_nomerge
;
2731 * Calculate the order of allocation given an slab object size.
2733 * The order of allocation has significant impact on performance and other
2734 * system components. Generally order 0 allocations should be preferred since
2735 * order 0 does not cause fragmentation in the page allocator. Larger objects
2736 * be problematic to put into order 0 slabs because there may be too much
2737 * unused space left. We go to a higher order if more than 1/16th of the slab
2740 * In order to reach satisfactory performance we must ensure that a minimum
2741 * number of objects is in one slab. Otherwise we may generate too much
2742 * activity on the partial lists which requires taking the list_lock. This is
2743 * less a concern for large slabs though which are rarely used.
2745 * slub_max_order specifies the order where we begin to stop considering the
2746 * number of objects in a slab as critical. If we reach slub_max_order then
2747 * we try to keep the page order as low as possible. So we accept more waste
2748 * of space in favor of a small page order.
2750 * Higher order allocations also allow the placement of more objects in a
2751 * slab and thereby reduce object handling overhead. If the user has
2752 * requested a higher mininum order then we start with that one instead of
2753 * the smallest order which will fit the object.
2755 static inline int slab_order(int size
, int min_objects
,
2756 int max_order
, int fract_leftover
, int reserved
)
2760 int min_order
= slub_min_order
;
2762 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
2763 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
2765 for (order
= max(min_order
,
2766 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
2767 order
<= max_order
; order
++) {
2769 unsigned long slab_size
= PAGE_SIZE
<< order
;
2771 if (slab_size
< min_objects
* size
+ reserved
)
2774 rem
= (slab_size
- reserved
) % size
;
2776 if (rem
<= slab_size
/ fract_leftover
)
2784 static inline int calculate_order(int size
, int reserved
)
2792 * Attempt to find best configuration for a slab. This
2793 * works by first attempting to generate a layout with
2794 * the best configuration and backing off gradually.
2796 * First we reduce the acceptable waste in a slab. Then
2797 * we reduce the minimum objects required in a slab.
2799 min_objects
= slub_min_objects
;
2801 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
2802 max_objects
= order_objects(slub_max_order
, size
, reserved
);
2803 min_objects
= min(min_objects
, max_objects
);
2805 while (min_objects
> 1) {
2807 while (fraction
>= 4) {
2808 order
= slab_order(size
, min_objects
,
2809 slub_max_order
, fraction
, reserved
);
2810 if (order
<= slub_max_order
)
2818 * We were unable to place multiple objects in a slab. Now
2819 * lets see if we can place a single object there.
2821 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
2822 if (order
<= slub_max_order
)
2826 * Doh this slab cannot be placed using slub_max_order.
2828 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
2829 if (order
< MAX_ORDER
)
2835 init_kmem_cache_node(struct kmem_cache_node
*n
)
2838 spin_lock_init(&n
->list_lock
);
2839 INIT_LIST_HEAD(&n
->partial
);
2840 #ifdef CONFIG_SLUB_DEBUG
2841 atomic_long_set(&n
->nr_slabs
, 0);
2842 atomic_long_set(&n
->total_objects
, 0);
2843 INIT_LIST_HEAD(&n
->full
);
2847 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
2849 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
2850 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
2853 * Must align to double word boundary for the double cmpxchg
2854 * instructions to work; see __pcpu_double_call_return_bool().
2856 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
2857 2 * sizeof(void *));
2862 init_kmem_cache_cpus(s
);
2867 static struct kmem_cache
*kmem_cache_node
;
2870 * No kmalloc_node yet so do it by hand. We know that this is the first
2871 * slab on the node for this slabcache. There are no concurrent accesses
2874 * Note that this function only works on the kmem_cache_node
2875 * when allocating for the kmem_cache_node. This is used for bootstrapping
2876 * memory on a fresh node that has no slab structures yet.
2878 static void early_kmem_cache_node_alloc(int node
)
2881 struct kmem_cache_node
*n
;
2883 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
2885 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
2888 if (page_to_nid(page
) != node
) {
2889 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
2891 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
2892 "in order to be able to continue\n");
2897 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
2900 kmem_cache_node
->node
[node
] = n
;
2901 #ifdef CONFIG_SLUB_DEBUG
2902 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
2903 init_tracking(kmem_cache_node
, n
);
2905 init_kmem_cache_node(n
);
2906 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
2909 * the lock is for lockdep's sake, not for any actual
2912 spin_lock(&n
->list_lock
);
2913 add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
2914 spin_unlock(&n
->list_lock
);
2917 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2921 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2922 struct kmem_cache_node
*n
= s
->node
[node
];
2925 kmem_cache_free(kmem_cache_node
, n
);
2927 s
->node
[node
] = NULL
;
2931 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
2935 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2936 struct kmem_cache_node
*n
;
2938 if (slab_state
== DOWN
) {
2939 early_kmem_cache_node_alloc(node
);
2942 n
= kmem_cache_alloc_node(kmem_cache_node
,
2946 free_kmem_cache_nodes(s
);
2951 init_kmem_cache_node(n
);
2956 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
2958 if (min
< MIN_PARTIAL
)
2960 else if (min
> MAX_PARTIAL
)
2962 s
->min_partial
= min
;
2966 * calculate_sizes() determines the order and the distribution of data within
2969 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
2971 unsigned long flags
= s
->flags
;
2972 unsigned long size
= s
->object_size
;
2976 * Round up object size to the next word boundary. We can only
2977 * place the free pointer at word boundaries and this determines
2978 * the possible location of the free pointer.
2980 size
= ALIGN(size
, sizeof(void *));
2982 #ifdef CONFIG_SLUB_DEBUG
2984 * Determine if we can poison the object itself. If the user of
2985 * the slab may touch the object after free or before allocation
2986 * then we should never poison the object itself.
2988 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2990 s
->flags
|= __OBJECT_POISON
;
2992 s
->flags
&= ~__OBJECT_POISON
;
2996 * If we are Redzoning then check if there is some space between the
2997 * end of the object and the free pointer. If not then add an
2998 * additional word to have some bytes to store Redzone information.
3000 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3001 size
+= sizeof(void *);
3005 * With that we have determined the number of bytes in actual use
3006 * by the object. This is the potential offset to the free pointer.
3010 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
3013 * Relocate free pointer after the object if it is not
3014 * permitted to overwrite the first word of the object on
3017 * This is the case if we do RCU, have a constructor or
3018 * destructor or are poisoning the objects.
3021 size
+= sizeof(void *);
3024 #ifdef CONFIG_SLUB_DEBUG
3025 if (flags
& SLAB_STORE_USER
)
3027 * Need to store information about allocs and frees after
3030 size
+= 2 * sizeof(struct track
);
3032 if (flags
& SLAB_RED_ZONE
)
3034 * Add some empty padding so that we can catch
3035 * overwrites from earlier objects rather than let
3036 * tracking information or the free pointer be
3037 * corrupted if a user writes before the start
3040 size
+= sizeof(void *);
3044 * SLUB stores one object immediately after another beginning from
3045 * offset 0. In order to align the objects we have to simply size
3046 * each object to conform to the alignment.
3048 size
= ALIGN(size
, s
->align
);
3050 if (forced_order
>= 0)
3051 order
= forced_order
;
3053 order
= calculate_order(size
, s
->reserved
);
3060 s
->allocflags
|= __GFP_COMP
;
3062 if (s
->flags
& SLAB_CACHE_DMA
)
3063 s
->allocflags
|= GFP_DMA
;
3065 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3066 s
->allocflags
|= __GFP_RECLAIMABLE
;
3069 * Determine the number of objects per slab
3071 s
->oo
= oo_make(order
, size
, s
->reserved
);
3072 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3073 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3076 return !!oo_objects(s
->oo
);
3079 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3081 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3084 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
3085 s
->reserved
= sizeof(struct rcu_head
);
3087 if (!calculate_sizes(s
, -1))
3089 if (disable_higher_order_debug
) {
3091 * Disable debugging flags that store metadata if the min slab
3094 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3095 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3097 if (!calculate_sizes(s
, -1))
3102 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3103 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3104 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_DEBUG_FLAGS
) == 0)
3105 /* Enable fast mode */
3106 s
->flags
|= __CMPXCHG_DOUBLE
;
3110 * The larger the object size is, the more pages we want on the partial
3111 * list to avoid pounding the page allocator excessively.
3113 set_min_partial(s
, ilog2(s
->size
) / 2);
3116 * cpu_partial determined the maximum number of objects kept in the
3117 * per cpu partial lists of a processor.
3119 * Per cpu partial lists mainly contain slabs that just have one
3120 * object freed. If they are used for allocation then they can be
3121 * filled up again with minimal effort. The slab will never hit the
3122 * per node partial lists and therefore no locking will be required.
3124 * This setting also determines
3126 * A) The number of objects from per cpu partial slabs dumped to the
3127 * per node list when we reach the limit.
3128 * B) The number of objects in cpu partial slabs to extract from the
3129 * per node list when we run out of per cpu objects. We only fetch
3130 * 50% to keep some capacity around for frees.
3132 if (!kmem_cache_has_cpu_partial(s
))
3134 else if (s
->size
>= PAGE_SIZE
)
3136 else if (s
->size
>= 1024)
3138 else if (s
->size
>= 256)
3139 s
->cpu_partial
= 13;
3141 s
->cpu_partial
= 30;
3144 s
->remote_node_defrag_ratio
= 1000;
3146 if (!init_kmem_cache_nodes(s
))
3149 if (alloc_kmem_cache_cpus(s
))
3152 free_kmem_cache_nodes(s
);
3154 if (flags
& SLAB_PANIC
)
3155 panic("Cannot create slab %s size=%lu realsize=%u "
3156 "order=%u offset=%u flags=%lx\n",
3157 s
->name
, (unsigned long)s
->size
, s
->size
,
3158 oo_order(s
->oo
), s
->offset
, flags
);
3162 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3165 #ifdef CONFIG_SLUB_DEBUG
3166 void *addr
= page_address(page
);
3168 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3169 sizeof(long), GFP_ATOMIC
);
3172 slab_err(s
, page
, text
, s
->name
);
3175 get_map(s
, page
, map
);
3176 for_each_object(p
, s
, addr
, page
->objects
) {
3178 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3179 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu\n",
3181 print_tracking(s
, p
);
3190 * Attempt to free all partial slabs on a node.
3191 * This is called from kmem_cache_close(). We must be the last thread
3192 * using the cache and therefore we do not need to lock anymore.
3194 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3196 struct page
*page
, *h
;
3198 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3200 remove_partial(n
, page
);
3201 discard_slab(s
, page
);
3203 list_slab_objects(s
, page
,
3204 "Objects remaining in %s on kmem_cache_close()");
3210 * Release all resources used by a slab cache.
3212 static inline int kmem_cache_close(struct kmem_cache
*s
)
3217 /* Attempt to free all objects */
3218 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3219 struct kmem_cache_node
*n
= get_node(s
, node
);
3222 if (n
->nr_partial
|| slabs_node(s
, node
))
3225 free_percpu(s
->cpu_slab
);
3226 free_kmem_cache_nodes(s
);
3230 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3232 int rc
= kmem_cache_close(s
);
3236 * We do the same lock strategy around sysfs_slab_add, see
3237 * __kmem_cache_create. Because this is pretty much the last
3238 * operation we do and the lock will be released shortly after
3239 * that in slab_common.c, we could just move sysfs_slab_remove
3240 * to a later point in common code. We should do that when we
3241 * have a common sysfs framework for all allocators.
3243 mutex_unlock(&slab_mutex
);
3244 sysfs_slab_remove(s
);
3245 mutex_lock(&slab_mutex
);
3251 /********************************************************************
3253 *******************************************************************/
3255 static int __init
setup_slub_min_order(char *str
)
3257 get_option(&str
, &slub_min_order
);
3262 __setup("slub_min_order=", setup_slub_min_order
);
3264 static int __init
setup_slub_max_order(char *str
)
3266 get_option(&str
, &slub_max_order
);
3267 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3272 __setup("slub_max_order=", setup_slub_max_order
);
3274 static int __init
setup_slub_min_objects(char *str
)
3276 get_option(&str
, &slub_min_objects
);
3281 __setup("slub_min_objects=", setup_slub_min_objects
);
3283 static int __init
setup_slub_nomerge(char *str
)
3289 __setup("slub_nomerge", setup_slub_nomerge
);
3291 void *__kmalloc(size_t size
, gfp_t flags
)
3293 struct kmem_cache
*s
;
3296 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3297 return kmalloc_large(size
, flags
);
3299 s
= kmalloc_slab(size
, flags
);
3301 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3304 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3306 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3310 EXPORT_SYMBOL(__kmalloc
);
3313 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3318 flags
|= __GFP_COMP
| __GFP_NOTRACK
| __GFP_KMEMCG
;
3319 page
= alloc_pages_node(node
, flags
, get_order(size
));
3321 ptr
= page_address(page
);
3323 kmalloc_large_node_hook(ptr
, size
, flags
);
3327 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3329 struct kmem_cache
*s
;
3332 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3333 ret
= kmalloc_large_node(size
, flags
, node
);
3335 trace_kmalloc_node(_RET_IP_
, ret
,
3336 size
, PAGE_SIZE
<< get_order(size
),
3342 s
= kmalloc_slab(size
, flags
);
3344 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3347 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3349 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3353 EXPORT_SYMBOL(__kmalloc_node
);
3356 size_t ksize(const void *object
)
3360 if (unlikely(object
== ZERO_SIZE_PTR
))
3363 page
= virt_to_head_page(object
);
3365 if (unlikely(!PageSlab(page
))) {
3366 WARN_ON(!PageCompound(page
));
3367 return PAGE_SIZE
<< compound_order(page
);
3370 return slab_ksize(page
->slab_cache
);
3372 EXPORT_SYMBOL(ksize
);
3374 void kfree(const void *x
)
3377 void *object
= (void *)x
;
3379 trace_kfree(_RET_IP_
, x
);
3381 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3384 page
= virt_to_head_page(x
);
3385 if (unlikely(!PageSlab(page
))) {
3386 BUG_ON(!PageCompound(page
));
3388 __free_memcg_kmem_pages(page
, compound_order(page
));
3391 slab_free(page
->slab_cache
, page
, object
, _RET_IP_
);
3393 EXPORT_SYMBOL(kfree
);
3396 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3397 * the remaining slabs by the number of items in use. The slabs with the
3398 * most items in use come first. New allocations will then fill those up
3399 * and thus they can be removed from the partial lists.
3401 * The slabs with the least items are placed last. This results in them
3402 * being allocated from last increasing the chance that the last objects
3403 * are freed in them.
3405 int kmem_cache_shrink(struct kmem_cache
*s
)
3409 struct kmem_cache_node
*n
;
3412 int objects
= oo_objects(s
->max
);
3413 struct list_head
*slabs_by_inuse
=
3414 kmalloc(sizeof(struct list_head
) * objects
, GFP_KERNEL
);
3415 unsigned long flags
;
3417 if (!slabs_by_inuse
)
3421 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3422 n
= get_node(s
, node
);
3427 for (i
= 0; i
< objects
; i
++)
3428 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
3430 spin_lock_irqsave(&n
->list_lock
, flags
);
3433 * Build lists indexed by the items in use in each slab.
3435 * Note that concurrent frees may occur while we hold the
3436 * list_lock. page->inuse here is the upper limit.
3438 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3439 list_move(&page
->lru
, slabs_by_inuse
+ page
->inuse
);
3445 * Rebuild the partial list with the slabs filled up most
3446 * first and the least used slabs at the end.
3448 for (i
= objects
- 1; i
> 0; i
--)
3449 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
3451 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3453 /* Release empty slabs */
3454 list_for_each_entry_safe(page
, t
, slabs_by_inuse
, lru
)
3455 discard_slab(s
, page
);
3458 kfree(slabs_by_inuse
);
3461 EXPORT_SYMBOL(kmem_cache_shrink
);
3463 static int slab_mem_going_offline_callback(void *arg
)
3465 struct kmem_cache
*s
;
3467 mutex_lock(&slab_mutex
);
3468 list_for_each_entry(s
, &slab_caches
, list
)
3469 kmem_cache_shrink(s
);
3470 mutex_unlock(&slab_mutex
);
3475 static void slab_mem_offline_callback(void *arg
)
3477 struct kmem_cache_node
*n
;
3478 struct kmem_cache
*s
;
3479 struct memory_notify
*marg
= arg
;
3482 offline_node
= marg
->status_change_nid_normal
;
3485 * If the node still has available memory. we need kmem_cache_node
3488 if (offline_node
< 0)
3491 mutex_lock(&slab_mutex
);
3492 list_for_each_entry(s
, &slab_caches
, list
) {
3493 n
= get_node(s
, offline_node
);
3496 * if n->nr_slabs > 0, slabs still exist on the node
3497 * that is going down. We were unable to free them,
3498 * and offline_pages() function shouldn't call this
3499 * callback. So, we must fail.
3501 BUG_ON(slabs_node(s
, offline_node
));
3503 s
->node
[offline_node
] = NULL
;
3504 kmem_cache_free(kmem_cache_node
, n
);
3507 mutex_unlock(&slab_mutex
);
3510 static int slab_mem_going_online_callback(void *arg
)
3512 struct kmem_cache_node
*n
;
3513 struct kmem_cache
*s
;
3514 struct memory_notify
*marg
= arg
;
3515 int nid
= marg
->status_change_nid_normal
;
3519 * If the node's memory is already available, then kmem_cache_node is
3520 * already created. Nothing to do.
3526 * We are bringing a node online. No memory is available yet. We must
3527 * allocate a kmem_cache_node structure in order to bring the node
3530 mutex_lock(&slab_mutex
);
3531 list_for_each_entry(s
, &slab_caches
, list
) {
3533 * XXX: kmem_cache_alloc_node will fallback to other nodes
3534 * since memory is not yet available from the node that
3537 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3542 init_kmem_cache_node(n
);
3546 mutex_unlock(&slab_mutex
);
3550 static int slab_memory_callback(struct notifier_block
*self
,
3551 unsigned long action
, void *arg
)
3556 case MEM_GOING_ONLINE
:
3557 ret
= slab_mem_going_online_callback(arg
);
3559 case MEM_GOING_OFFLINE
:
3560 ret
= slab_mem_going_offline_callback(arg
);
3563 case MEM_CANCEL_ONLINE
:
3564 slab_mem_offline_callback(arg
);
3567 case MEM_CANCEL_OFFLINE
:
3571 ret
= notifier_from_errno(ret
);
3577 static struct notifier_block slab_memory_callback_nb
= {
3578 .notifier_call
= slab_memory_callback
,
3579 .priority
= SLAB_CALLBACK_PRI
,
3582 /********************************************************************
3583 * Basic setup of slabs
3584 *******************************************************************/
3587 * Used for early kmem_cache structures that were allocated using
3588 * the page allocator. Allocate them properly then fix up the pointers
3589 * that may be pointing to the wrong kmem_cache structure.
3592 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
3595 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
3597 memcpy(s
, static_cache
, kmem_cache
->object_size
);
3600 * This runs very early, and only the boot processor is supposed to be
3601 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3604 __flush_cpu_slab(s
, smp_processor_id());
3605 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3606 struct kmem_cache_node
*n
= get_node(s
, node
);
3610 list_for_each_entry(p
, &n
->partial
, lru
)
3613 #ifdef CONFIG_SLUB_DEBUG
3614 list_for_each_entry(p
, &n
->full
, lru
)
3619 list_add(&s
->list
, &slab_caches
);
3623 void __init
kmem_cache_init(void)
3625 static __initdata
struct kmem_cache boot_kmem_cache
,
3626 boot_kmem_cache_node
;
3628 if (debug_guardpage_minorder())
3631 kmem_cache_node
= &boot_kmem_cache_node
;
3632 kmem_cache
= &boot_kmem_cache
;
3634 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
3635 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
3637 register_hotmemory_notifier(&slab_memory_callback_nb
);
3639 /* Able to allocate the per node structures */
3640 slab_state
= PARTIAL
;
3642 create_boot_cache(kmem_cache
, "kmem_cache",
3643 offsetof(struct kmem_cache
, node
) +
3644 nr_node_ids
* sizeof(struct kmem_cache_node
*),
3645 SLAB_HWCACHE_ALIGN
);
3647 kmem_cache
= bootstrap(&boot_kmem_cache
);
3650 * Allocate kmem_cache_node properly from the kmem_cache slab.
3651 * kmem_cache_node is separately allocated so no need to
3652 * update any list pointers.
3654 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
3656 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3657 create_kmalloc_caches(0);
3660 register_cpu_notifier(&slab_notifier
);
3664 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3665 " CPUs=%d, Nodes=%d\n",
3667 slub_min_order
, slub_max_order
, slub_min_objects
,
3668 nr_cpu_ids
, nr_node_ids
);
3671 void __init
kmem_cache_init_late(void)
3676 * Find a mergeable slab cache
3678 static int slab_unmergeable(struct kmem_cache
*s
)
3680 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
3687 * We may have set a slab to be unmergeable during bootstrap.
3689 if (s
->refcount
< 0)
3695 static struct kmem_cache
*find_mergeable(struct mem_cgroup
*memcg
, size_t size
,
3696 size_t align
, unsigned long flags
, const char *name
,
3697 void (*ctor
)(void *))
3699 struct kmem_cache
*s
;
3701 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
3707 size
= ALIGN(size
, sizeof(void *));
3708 align
= calculate_alignment(flags
, align
, size
);
3709 size
= ALIGN(size
, align
);
3710 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
3712 list_for_each_entry(s
, &slab_caches
, list
) {
3713 if (slab_unmergeable(s
))
3719 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
3722 * Check if alignment is compatible.
3723 * Courtesy of Adrian Drzewiecki
3725 if ((s
->size
& ~(align
- 1)) != s
->size
)
3728 if (s
->size
- size
>= sizeof(void *))
3731 if (!cache_match_memcg(s
, memcg
))
3740 __kmem_cache_alias(struct mem_cgroup
*memcg
, const char *name
, size_t size
,
3741 size_t align
, unsigned long flags
, void (*ctor
)(void *))
3743 struct kmem_cache
*s
;
3745 s
= find_mergeable(memcg
, size
, align
, flags
, name
, ctor
);
3749 * Adjust the object sizes so that we clear
3750 * the complete object on kzalloc.
3752 s
->object_size
= max(s
->object_size
, (int)size
);
3753 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3755 if (sysfs_slab_alias(s
, name
)) {
3764 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
3768 err
= kmem_cache_open(s
, flags
);
3772 /* Mutex is not taken during early boot */
3773 if (slab_state
<= UP
)
3776 memcg_propagate_slab_attrs(s
);
3777 mutex_unlock(&slab_mutex
);
3778 err
= sysfs_slab_add(s
);
3779 mutex_lock(&slab_mutex
);
3782 kmem_cache_close(s
);
3789 * Use the cpu notifier to insure that the cpu slabs are flushed when
3792 static int slab_cpuup_callback(struct notifier_block
*nfb
,
3793 unsigned long action
, void *hcpu
)
3795 long cpu
= (long)hcpu
;
3796 struct kmem_cache
*s
;
3797 unsigned long flags
;
3800 case CPU_UP_CANCELED
:
3801 case CPU_UP_CANCELED_FROZEN
:
3803 case CPU_DEAD_FROZEN
:
3804 mutex_lock(&slab_mutex
);
3805 list_for_each_entry(s
, &slab_caches
, list
) {
3806 local_irq_save(flags
);
3807 __flush_cpu_slab(s
, cpu
);
3808 local_irq_restore(flags
);
3810 mutex_unlock(&slab_mutex
);
3818 static struct notifier_block slab_notifier
= {
3819 .notifier_call
= slab_cpuup_callback
3824 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
3826 struct kmem_cache
*s
;
3829 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3830 return kmalloc_large(size
, gfpflags
);
3832 s
= kmalloc_slab(size
, gfpflags
);
3834 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3837 ret
= slab_alloc(s
, gfpflags
, caller
);
3839 /* Honor the call site pointer we received. */
3840 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
3846 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3847 int node
, unsigned long caller
)
3849 struct kmem_cache
*s
;
3852 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3853 ret
= kmalloc_large_node(size
, gfpflags
, node
);
3855 trace_kmalloc_node(caller
, ret
,
3856 size
, PAGE_SIZE
<< get_order(size
),
3862 s
= kmalloc_slab(size
, gfpflags
);
3864 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3867 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
3869 /* Honor the call site pointer we received. */
3870 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
3877 static int count_inuse(struct page
*page
)
3882 static int count_total(struct page
*page
)
3884 return page
->objects
;
3888 #ifdef CONFIG_SLUB_DEBUG
3889 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3893 void *addr
= page_address(page
);
3895 if (!check_slab(s
, page
) ||
3896 !on_freelist(s
, page
, NULL
))
3899 /* Now we know that a valid freelist exists */
3900 bitmap_zero(map
, page
->objects
);
3902 get_map(s
, page
, map
);
3903 for_each_object(p
, s
, addr
, page
->objects
) {
3904 if (test_bit(slab_index(p
, s
, addr
), map
))
3905 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
3909 for_each_object(p
, s
, addr
, page
->objects
)
3910 if (!test_bit(slab_index(p
, s
, addr
), map
))
3911 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
3916 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
3920 validate_slab(s
, page
, map
);
3924 static int validate_slab_node(struct kmem_cache
*s
,
3925 struct kmem_cache_node
*n
, unsigned long *map
)
3927 unsigned long count
= 0;
3929 unsigned long flags
;
3931 spin_lock_irqsave(&n
->list_lock
, flags
);
3933 list_for_each_entry(page
, &n
->partial
, lru
) {
3934 validate_slab_slab(s
, page
, map
);
3937 if (count
!= n
->nr_partial
)
3938 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
3939 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
3941 if (!(s
->flags
& SLAB_STORE_USER
))
3944 list_for_each_entry(page
, &n
->full
, lru
) {
3945 validate_slab_slab(s
, page
, map
);
3948 if (count
!= atomic_long_read(&n
->nr_slabs
))
3949 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
3950 "counter=%ld\n", s
->name
, count
,
3951 atomic_long_read(&n
->nr_slabs
));
3954 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3958 static long validate_slab_cache(struct kmem_cache
*s
)
3961 unsigned long count
= 0;
3962 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
3963 sizeof(unsigned long), GFP_KERNEL
);
3969 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3970 struct kmem_cache_node
*n
= get_node(s
, node
);
3972 count
+= validate_slab_node(s
, n
, map
);
3978 * Generate lists of code addresses where slabcache objects are allocated
3983 unsigned long count
;
3990 DECLARE_BITMAP(cpus
, NR_CPUS
);
3996 unsigned long count
;
3997 struct location
*loc
;
4000 static void free_loc_track(struct loc_track
*t
)
4003 free_pages((unsigned long)t
->loc
,
4004 get_order(sizeof(struct location
) * t
->max
));
4007 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4012 order
= get_order(sizeof(struct location
) * max
);
4014 l
= (void *)__get_free_pages(flags
, order
);
4019 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4027 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4028 const struct track
*track
)
4030 long start
, end
, pos
;
4032 unsigned long caddr
;
4033 unsigned long age
= jiffies
- track
->when
;
4039 pos
= start
+ (end
- start
+ 1) / 2;
4042 * There is nothing at "end". If we end up there
4043 * we need to add something to before end.
4048 caddr
= t
->loc
[pos
].addr
;
4049 if (track
->addr
== caddr
) {
4055 if (age
< l
->min_time
)
4057 if (age
> l
->max_time
)
4060 if (track
->pid
< l
->min_pid
)
4061 l
->min_pid
= track
->pid
;
4062 if (track
->pid
> l
->max_pid
)
4063 l
->max_pid
= track
->pid
;
4065 cpumask_set_cpu(track
->cpu
,
4066 to_cpumask(l
->cpus
));
4068 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4072 if (track
->addr
< caddr
)
4079 * Not found. Insert new tracking element.
4081 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4087 (t
->count
- pos
) * sizeof(struct location
));
4090 l
->addr
= track
->addr
;
4094 l
->min_pid
= track
->pid
;
4095 l
->max_pid
= track
->pid
;
4096 cpumask_clear(to_cpumask(l
->cpus
));
4097 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4098 nodes_clear(l
->nodes
);
4099 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4103 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4104 struct page
*page
, enum track_item alloc
,
4107 void *addr
= page_address(page
);
4110 bitmap_zero(map
, page
->objects
);
4111 get_map(s
, page
, map
);
4113 for_each_object(p
, s
, addr
, page
->objects
)
4114 if (!test_bit(slab_index(p
, s
, addr
), map
))
4115 add_location(t
, s
, get_track(s
, p
, alloc
));
4118 static int list_locations(struct kmem_cache
*s
, char *buf
,
4119 enum track_item alloc
)
4123 struct loc_track t
= { 0, 0, NULL
};
4125 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4126 sizeof(unsigned long), GFP_KERNEL
);
4128 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4131 return sprintf(buf
, "Out of memory\n");
4133 /* Push back cpu slabs */
4136 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4137 struct kmem_cache_node
*n
= get_node(s
, node
);
4138 unsigned long flags
;
4141 if (!atomic_long_read(&n
->nr_slabs
))
4144 spin_lock_irqsave(&n
->list_lock
, flags
);
4145 list_for_each_entry(page
, &n
->partial
, lru
)
4146 process_slab(&t
, s
, page
, alloc
, map
);
4147 list_for_each_entry(page
, &n
->full
, lru
)
4148 process_slab(&t
, s
, page
, alloc
, map
);
4149 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4152 for (i
= 0; i
< t
.count
; i
++) {
4153 struct location
*l
= &t
.loc
[i
];
4155 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4157 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4160 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4162 len
+= sprintf(buf
+ len
, "<not-available>");
4164 if (l
->sum_time
!= l
->min_time
) {
4165 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4167 (long)div_u64(l
->sum_time
, l
->count
),
4170 len
+= sprintf(buf
+ len
, " age=%ld",
4173 if (l
->min_pid
!= l
->max_pid
)
4174 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4175 l
->min_pid
, l
->max_pid
);
4177 len
+= sprintf(buf
+ len
, " pid=%ld",
4180 if (num_online_cpus() > 1 &&
4181 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4182 len
< PAGE_SIZE
- 60) {
4183 len
+= sprintf(buf
+ len
, " cpus=");
4184 len
+= cpulist_scnprintf(buf
+ len
,
4185 PAGE_SIZE
- len
- 50,
4186 to_cpumask(l
->cpus
));
4189 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4190 len
< PAGE_SIZE
- 60) {
4191 len
+= sprintf(buf
+ len
, " nodes=");
4192 len
+= nodelist_scnprintf(buf
+ len
,
4193 PAGE_SIZE
- len
- 50,
4197 len
+= sprintf(buf
+ len
, "\n");
4203 len
+= sprintf(buf
, "No data\n");
4208 #ifdef SLUB_RESILIENCY_TEST
4209 static void resiliency_test(void)
4213 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4215 printk(KERN_ERR
"SLUB resiliency testing\n");
4216 printk(KERN_ERR
"-----------------------\n");
4217 printk(KERN_ERR
"A. Corruption after allocation\n");
4219 p
= kzalloc(16, GFP_KERNEL
);
4221 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
4222 " 0x12->0x%p\n\n", p
+ 16);
4224 validate_slab_cache(kmalloc_caches
[4]);
4226 /* Hmmm... The next two are dangerous */
4227 p
= kzalloc(32, GFP_KERNEL
);
4228 p
[32 + sizeof(void *)] = 0x34;
4229 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
4230 " 0x34 -> -0x%p\n", p
);
4232 "If allocated object is overwritten then not detectable\n\n");
4234 validate_slab_cache(kmalloc_caches
[5]);
4235 p
= kzalloc(64, GFP_KERNEL
);
4236 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4238 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4241 "If allocated object is overwritten then not detectable\n\n");
4242 validate_slab_cache(kmalloc_caches
[6]);
4244 printk(KERN_ERR
"\nB. Corruption after free\n");
4245 p
= kzalloc(128, GFP_KERNEL
);
4248 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4249 validate_slab_cache(kmalloc_caches
[7]);
4251 p
= kzalloc(256, GFP_KERNEL
);
4254 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4256 validate_slab_cache(kmalloc_caches
[8]);
4258 p
= kzalloc(512, GFP_KERNEL
);
4261 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4262 validate_slab_cache(kmalloc_caches
[9]);
4266 static void resiliency_test(void) {};
4271 enum slab_stat_type
{
4272 SL_ALL
, /* All slabs */
4273 SL_PARTIAL
, /* Only partially allocated slabs */
4274 SL_CPU
, /* Only slabs used for cpu caches */
4275 SL_OBJECTS
, /* Determine allocated objects not slabs */
4276 SL_TOTAL
/* Determine object capacity not slabs */
4279 #define SO_ALL (1 << SL_ALL)
4280 #define SO_PARTIAL (1 << SL_PARTIAL)
4281 #define SO_CPU (1 << SL_CPU)
4282 #define SO_OBJECTS (1 << SL_OBJECTS)
4283 #define SO_TOTAL (1 << SL_TOTAL)
4285 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4286 char *buf
, unsigned long flags
)
4288 unsigned long total
= 0;
4291 unsigned long *nodes
;
4293 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4297 if (flags
& SO_CPU
) {
4300 for_each_possible_cpu(cpu
) {
4301 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4306 page
= ACCESS_ONCE(c
->page
);
4310 node
= page_to_nid(page
);
4311 if (flags
& SO_TOTAL
)
4313 else if (flags
& SO_OBJECTS
)
4321 page
= ACCESS_ONCE(c
->partial
);
4323 node
= page_to_nid(page
);
4324 if (flags
& SO_TOTAL
)
4326 else if (flags
& SO_OBJECTS
)
4336 lock_memory_hotplug();
4337 #ifdef CONFIG_SLUB_DEBUG
4338 if (flags
& SO_ALL
) {
4339 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4340 struct kmem_cache_node
*n
= get_node(s
, node
);
4342 if (flags
& SO_TOTAL
)
4343 x
= atomic_long_read(&n
->total_objects
);
4344 else if (flags
& SO_OBJECTS
)
4345 x
= atomic_long_read(&n
->total_objects
) -
4346 count_partial(n
, count_free
);
4348 x
= atomic_long_read(&n
->nr_slabs
);
4355 if (flags
& SO_PARTIAL
) {
4356 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4357 struct kmem_cache_node
*n
= get_node(s
, node
);
4359 if (flags
& SO_TOTAL
)
4360 x
= count_partial(n
, count_total
);
4361 else if (flags
& SO_OBJECTS
)
4362 x
= count_partial(n
, count_inuse
);
4369 x
= sprintf(buf
, "%lu", total
);
4371 for_each_node_state(node
, N_NORMAL_MEMORY
)
4373 x
+= sprintf(buf
+ x
, " N%d=%lu",
4376 unlock_memory_hotplug();
4378 return x
+ sprintf(buf
+ x
, "\n");
4381 #ifdef CONFIG_SLUB_DEBUG
4382 static int any_slab_objects(struct kmem_cache
*s
)
4386 for_each_online_node(node
) {
4387 struct kmem_cache_node
*n
= get_node(s
, node
);
4392 if (atomic_long_read(&n
->total_objects
))
4399 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4400 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4402 struct slab_attribute
{
4403 struct attribute attr
;
4404 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4405 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4408 #define SLAB_ATTR_RO(_name) \
4409 static struct slab_attribute _name##_attr = \
4410 __ATTR(_name, 0400, _name##_show, NULL)
4412 #define SLAB_ATTR(_name) \
4413 static struct slab_attribute _name##_attr = \
4414 __ATTR(_name, 0600, _name##_show, _name##_store)
4416 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4418 return sprintf(buf
, "%d\n", s
->size
);
4420 SLAB_ATTR_RO(slab_size
);
4422 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4424 return sprintf(buf
, "%d\n", s
->align
);
4426 SLAB_ATTR_RO(align
);
4428 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4430 return sprintf(buf
, "%d\n", s
->object_size
);
4432 SLAB_ATTR_RO(object_size
);
4434 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4436 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4438 SLAB_ATTR_RO(objs_per_slab
);
4440 static ssize_t
order_store(struct kmem_cache
*s
,
4441 const char *buf
, size_t length
)
4443 unsigned long order
;
4446 err
= kstrtoul(buf
, 10, &order
);
4450 if (order
> slub_max_order
|| order
< slub_min_order
)
4453 calculate_sizes(s
, order
);
4457 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4459 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4463 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4465 return sprintf(buf
, "%lu\n", s
->min_partial
);
4468 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4474 err
= kstrtoul(buf
, 10, &min
);
4478 set_min_partial(s
, min
);
4481 SLAB_ATTR(min_partial
);
4483 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4485 return sprintf(buf
, "%u\n", s
->cpu_partial
);
4488 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4491 unsigned long objects
;
4494 err
= kstrtoul(buf
, 10, &objects
);
4497 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4500 s
->cpu_partial
= objects
;
4504 SLAB_ATTR(cpu_partial
);
4506 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4510 return sprintf(buf
, "%pS\n", s
->ctor
);
4514 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4516 return sprintf(buf
, "%d\n", s
->refcount
- 1);
4518 SLAB_ATTR_RO(aliases
);
4520 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4522 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4524 SLAB_ATTR_RO(partial
);
4526 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4528 return show_slab_objects(s
, buf
, SO_CPU
);
4530 SLAB_ATTR_RO(cpu_slabs
);
4532 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4534 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4536 SLAB_ATTR_RO(objects
);
4538 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4540 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4542 SLAB_ATTR_RO(objects_partial
);
4544 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4551 for_each_online_cpu(cpu
) {
4552 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
)->partial
;
4555 pages
+= page
->pages
;
4556 objects
+= page
->pobjects
;
4560 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
4563 for_each_online_cpu(cpu
) {
4564 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
) ->partial
;
4566 if (page
&& len
< PAGE_SIZE
- 20)
4567 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
4568 page
->pobjects
, page
->pages
);
4571 return len
+ sprintf(buf
+ len
, "\n");
4573 SLAB_ATTR_RO(slabs_cpu_partial
);
4575 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4577 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4580 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4581 const char *buf
, size_t length
)
4583 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4585 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4588 SLAB_ATTR(reclaim_account
);
4590 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4592 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4594 SLAB_ATTR_RO(hwcache_align
);
4596 #ifdef CONFIG_ZONE_DMA
4597 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4599 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4601 SLAB_ATTR_RO(cache_dma
);
4604 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4606 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4608 SLAB_ATTR_RO(destroy_by_rcu
);
4610 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4612 return sprintf(buf
, "%d\n", s
->reserved
);
4614 SLAB_ATTR_RO(reserved
);
4616 #ifdef CONFIG_SLUB_DEBUG
4617 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4619 return show_slab_objects(s
, buf
, SO_ALL
);
4621 SLAB_ATTR_RO(slabs
);
4623 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4625 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4627 SLAB_ATTR_RO(total_objects
);
4629 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4631 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
4634 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
4635 const char *buf
, size_t length
)
4637 s
->flags
&= ~SLAB_DEBUG_FREE
;
4638 if (buf
[0] == '1') {
4639 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4640 s
->flags
|= SLAB_DEBUG_FREE
;
4644 SLAB_ATTR(sanity_checks
);
4646 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
4648 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
4651 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
4654 s
->flags
&= ~SLAB_TRACE
;
4655 if (buf
[0] == '1') {
4656 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4657 s
->flags
|= SLAB_TRACE
;
4663 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
4665 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
4668 static ssize_t
red_zone_store(struct kmem_cache
*s
,
4669 const char *buf
, size_t length
)
4671 if (any_slab_objects(s
))
4674 s
->flags
&= ~SLAB_RED_ZONE
;
4675 if (buf
[0] == '1') {
4676 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4677 s
->flags
|= SLAB_RED_ZONE
;
4679 calculate_sizes(s
, -1);
4682 SLAB_ATTR(red_zone
);
4684 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
4686 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
4689 static ssize_t
poison_store(struct kmem_cache
*s
,
4690 const char *buf
, size_t length
)
4692 if (any_slab_objects(s
))
4695 s
->flags
&= ~SLAB_POISON
;
4696 if (buf
[0] == '1') {
4697 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4698 s
->flags
|= SLAB_POISON
;
4700 calculate_sizes(s
, -1);
4705 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
4707 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
4710 static ssize_t
store_user_store(struct kmem_cache
*s
,
4711 const char *buf
, size_t length
)
4713 if (any_slab_objects(s
))
4716 s
->flags
&= ~SLAB_STORE_USER
;
4717 if (buf
[0] == '1') {
4718 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4719 s
->flags
|= SLAB_STORE_USER
;
4721 calculate_sizes(s
, -1);
4724 SLAB_ATTR(store_user
);
4726 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
4731 static ssize_t
validate_store(struct kmem_cache
*s
,
4732 const char *buf
, size_t length
)
4736 if (buf
[0] == '1') {
4737 ret
= validate_slab_cache(s
);
4743 SLAB_ATTR(validate
);
4745 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
4747 if (!(s
->flags
& SLAB_STORE_USER
))
4749 return list_locations(s
, buf
, TRACK_ALLOC
);
4751 SLAB_ATTR_RO(alloc_calls
);
4753 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
4755 if (!(s
->flags
& SLAB_STORE_USER
))
4757 return list_locations(s
, buf
, TRACK_FREE
);
4759 SLAB_ATTR_RO(free_calls
);
4760 #endif /* CONFIG_SLUB_DEBUG */
4762 #ifdef CONFIG_FAILSLAB
4763 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
4765 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
4768 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
4771 s
->flags
&= ~SLAB_FAILSLAB
;
4773 s
->flags
|= SLAB_FAILSLAB
;
4776 SLAB_ATTR(failslab
);
4779 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
4784 static ssize_t
shrink_store(struct kmem_cache
*s
,
4785 const char *buf
, size_t length
)
4787 if (buf
[0] == '1') {
4788 int rc
= kmem_cache_shrink(s
);
4799 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
4801 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
4804 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
4805 const char *buf
, size_t length
)
4807 unsigned long ratio
;
4810 err
= kstrtoul(buf
, 10, &ratio
);
4815 s
->remote_node_defrag_ratio
= ratio
* 10;
4819 SLAB_ATTR(remote_node_defrag_ratio
);
4822 #ifdef CONFIG_SLUB_STATS
4823 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
4825 unsigned long sum
= 0;
4828 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
4833 for_each_online_cpu(cpu
) {
4834 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
4840 len
= sprintf(buf
, "%lu", sum
);
4843 for_each_online_cpu(cpu
) {
4844 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
4845 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
4849 return len
+ sprintf(buf
+ len
, "\n");
4852 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
4856 for_each_online_cpu(cpu
)
4857 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
4860 #define STAT_ATTR(si, text) \
4861 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4863 return show_stat(s, buf, si); \
4865 static ssize_t text##_store(struct kmem_cache *s, \
4866 const char *buf, size_t length) \
4868 if (buf[0] != '0') \
4870 clear_stat(s, si); \
4875 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4876 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4877 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4878 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4879 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4880 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4881 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4882 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4883 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4884 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4885 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
4886 STAT_ATTR(FREE_SLAB
, free_slab
);
4887 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4888 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4889 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4890 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4891 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4892 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4893 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
4894 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
4895 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
4896 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
4897 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
4898 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
4899 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
4900 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
4903 static struct attribute
*slab_attrs
[] = {
4904 &slab_size_attr
.attr
,
4905 &object_size_attr
.attr
,
4906 &objs_per_slab_attr
.attr
,
4908 &min_partial_attr
.attr
,
4909 &cpu_partial_attr
.attr
,
4911 &objects_partial_attr
.attr
,
4913 &cpu_slabs_attr
.attr
,
4917 &hwcache_align_attr
.attr
,
4918 &reclaim_account_attr
.attr
,
4919 &destroy_by_rcu_attr
.attr
,
4921 &reserved_attr
.attr
,
4922 &slabs_cpu_partial_attr
.attr
,
4923 #ifdef CONFIG_SLUB_DEBUG
4924 &total_objects_attr
.attr
,
4926 &sanity_checks_attr
.attr
,
4928 &red_zone_attr
.attr
,
4930 &store_user_attr
.attr
,
4931 &validate_attr
.attr
,
4932 &alloc_calls_attr
.attr
,
4933 &free_calls_attr
.attr
,
4935 #ifdef CONFIG_ZONE_DMA
4936 &cache_dma_attr
.attr
,
4939 &remote_node_defrag_ratio_attr
.attr
,
4941 #ifdef CONFIG_SLUB_STATS
4942 &alloc_fastpath_attr
.attr
,
4943 &alloc_slowpath_attr
.attr
,
4944 &free_fastpath_attr
.attr
,
4945 &free_slowpath_attr
.attr
,
4946 &free_frozen_attr
.attr
,
4947 &free_add_partial_attr
.attr
,
4948 &free_remove_partial_attr
.attr
,
4949 &alloc_from_partial_attr
.attr
,
4950 &alloc_slab_attr
.attr
,
4951 &alloc_refill_attr
.attr
,
4952 &alloc_node_mismatch_attr
.attr
,
4953 &free_slab_attr
.attr
,
4954 &cpuslab_flush_attr
.attr
,
4955 &deactivate_full_attr
.attr
,
4956 &deactivate_empty_attr
.attr
,
4957 &deactivate_to_head_attr
.attr
,
4958 &deactivate_to_tail_attr
.attr
,
4959 &deactivate_remote_frees_attr
.attr
,
4960 &deactivate_bypass_attr
.attr
,
4961 &order_fallback_attr
.attr
,
4962 &cmpxchg_double_fail_attr
.attr
,
4963 &cmpxchg_double_cpu_fail_attr
.attr
,
4964 &cpu_partial_alloc_attr
.attr
,
4965 &cpu_partial_free_attr
.attr
,
4966 &cpu_partial_node_attr
.attr
,
4967 &cpu_partial_drain_attr
.attr
,
4969 #ifdef CONFIG_FAILSLAB
4970 &failslab_attr
.attr
,
4976 static struct attribute_group slab_attr_group
= {
4977 .attrs
= slab_attrs
,
4980 static ssize_t
slab_attr_show(struct kobject
*kobj
,
4981 struct attribute
*attr
,
4984 struct slab_attribute
*attribute
;
4985 struct kmem_cache
*s
;
4988 attribute
= to_slab_attr(attr
);
4991 if (!attribute
->show
)
4994 err
= attribute
->show(s
, buf
);
4999 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5000 struct attribute
*attr
,
5001 const char *buf
, size_t len
)
5003 struct slab_attribute
*attribute
;
5004 struct kmem_cache
*s
;
5007 attribute
= to_slab_attr(attr
);
5010 if (!attribute
->store
)
5013 err
= attribute
->store(s
, buf
, len
);
5014 #ifdef CONFIG_MEMCG_KMEM
5015 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5018 mutex_lock(&slab_mutex
);
5019 if (s
->max_attr_size
< len
)
5020 s
->max_attr_size
= len
;
5023 * This is a best effort propagation, so this function's return
5024 * value will be determined by the parent cache only. This is
5025 * basically because not all attributes will have a well
5026 * defined semantics for rollbacks - most of the actions will
5027 * have permanent effects.
5029 * Returning the error value of any of the children that fail
5030 * is not 100 % defined, in the sense that users seeing the
5031 * error code won't be able to know anything about the state of
5034 * Only returning the error code for the parent cache at least
5035 * has well defined semantics. The cache being written to
5036 * directly either failed or succeeded, in which case we loop
5037 * through the descendants with best-effort propagation.
5039 for_each_memcg_cache_index(i
) {
5040 struct kmem_cache
*c
= cache_from_memcg_idx(s
, i
);
5042 attribute
->store(c
, buf
, len
);
5044 mutex_unlock(&slab_mutex
);
5050 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5052 #ifdef CONFIG_MEMCG_KMEM
5054 char *buffer
= NULL
;
5056 if (!is_root_cache(s
))
5060 * This mean this cache had no attribute written. Therefore, no point
5061 * in copying default values around
5063 if (!s
->max_attr_size
)
5066 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5069 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5071 if (!attr
|| !attr
->store
|| !attr
->show
)
5075 * It is really bad that we have to allocate here, so we will
5076 * do it only as a fallback. If we actually allocate, though,
5077 * we can just use the allocated buffer until the end.
5079 * Most of the slub attributes will tend to be very small in
5080 * size, but sysfs allows buffers up to a page, so they can
5081 * theoretically happen.
5085 else if (s
->max_attr_size
< ARRAY_SIZE(mbuf
))
5088 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5089 if (WARN_ON(!buffer
))
5094 attr
->show(s
->memcg_params
->root_cache
, buf
);
5095 attr
->store(s
, buf
, strlen(buf
));
5099 free_page((unsigned long)buffer
);
5103 static const struct sysfs_ops slab_sysfs_ops
= {
5104 .show
= slab_attr_show
,
5105 .store
= slab_attr_store
,
5108 static struct kobj_type slab_ktype
= {
5109 .sysfs_ops
= &slab_sysfs_ops
,
5112 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5114 struct kobj_type
*ktype
= get_ktype(kobj
);
5116 if (ktype
== &slab_ktype
)
5121 static const struct kset_uevent_ops slab_uevent_ops
= {
5122 .filter
= uevent_filter
,
5125 static struct kset
*slab_kset
;
5127 #define ID_STR_LENGTH 64
5129 /* Create a unique string id for a slab cache:
5131 * Format :[flags-]size
5133 static char *create_unique_id(struct kmem_cache
*s
)
5135 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5142 * First flags affecting slabcache operations. We will only
5143 * get here for aliasable slabs so we do not need to support
5144 * too many flags. The flags here must cover all flags that
5145 * are matched during merging to guarantee that the id is
5148 if (s
->flags
& SLAB_CACHE_DMA
)
5150 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5152 if (s
->flags
& SLAB_DEBUG_FREE
)
5154 if (!(s
->flags
& SLAB_NOTRACK
))
5158 p
+= sprintf(p
, "%07d", s
->size
);
5160 #ifdef CONFIG_MEMCG_KMEM
5161 if (!is_root_cache(s
))
5162 p
+= sprintf(p
, "-%08d",
5163 memcg_cache_id(s
->memcg_params
->memcg
));
5166 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5170 static int sysfs_slab_add(struct kmem_cache
*s
)
5174 int unmergeable
= slab_unmergeable(s
);
5178 * Slabcache can never be merged so we can use the name proper.
5179 * This is typically the case for debug situations. In that
5180 * case we can catch duplicate names easily.
5182 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5186 * Create a unique name for the slab as a target
5189 name
= create_unique_id(s
);
5192 s
->kobj
.kset
= slab_kset
;
5193 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5195 kobject_put(&s
->kobj
);
5199 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5201 kobject_del(&s
->kobj
);
5202 kobject_put(&s
->kobj
);
5205 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5207 /* Setup first alias */
5208 sysfs_slab_alias(s
, s
->name
);
5214 static void sysfs_slab_remove(struct kmem_cache
*s
)
5216 if (slab_state
< FULL
)
5218 * Sysfs has not been setup yet so no need to remove the
5223 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5224 kobject_del(&s
->kobj
);
5225 kobject_put(&s
->kobj
);
5229 * Need to buffer aliases during bootup until sysfs becomes
5230 * available lest we lose that information.
5232 struct saved_alias
{
5233 struct kmem_cache
*s
;
5235 struct saved_alias
*next
;
5238 static struct saved_alias
*alias_list
;
5240 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5242 struct saved_alias
*al
;
5244 if (slab_state
== FULL
) {
5246 * If we have a leftover link then remove it.
5248 sysfs_remove_link(&slab_kset
->kobj
, name
);
5249 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5252 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5258 al
->next
= alias_list
;
5263 static int __init
slab_sysfs_init(void)
5265 struct kmem_cache
*s
;
5268 mutex_lock(&slab_mutex
);
5270 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5272 mutex_unlock(&slab_mutex
);
5273 printk(KERN_ERR
"Cannot register slab subsystem.\n");
5279 list_for_each_entry(s
, &slab_caches
, list
) {
5280 err
= sysfs_slab_add(s
);
5282 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
5283 " to sysfs\n", s
->name
);
5286 while (alias_list
) {
5287 struct saved_alias
*al
= alias_list
;
5289 alias_list
= alias_list
->next
;
5290 err
= sysfs_slab_alias(al
->s
, al
->name
);
5292 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
5293 " %s to sysfs\n", al
->name
);
5297 mutex_unlock(&slab_mutex
);
5302 __initcall(slab_sysfs_init
);
5303 #endif /* CONFIG_SYSFS */
5306 * The /proc/slabinfo ABI
5308 #ifdef CONFIG_SLABINFO
5309 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5311 unsigned long nr_slabs
= 0;
5312 unsigned long nr_objs
= 0;
5313 unsigned long nr_free
= 0;
5316 for_each_online_node(node
) {
5317 struct kmem_cache_node
*n
= get_node(s
, node
);
5322 nr_slabs
+= node_nr_slabs(n
);
5323 nr_objs
+= node_nr_objs(n
);
5324 nr_free
+= count_partial(n
, count_free
);
5327 sinfo
->active_objs
= nr_objs
- nr_free
;
5328 sinfo
->num_objs
= nr_objs
;
5329 sinfo
->active_slabs
= nr_slabs
;
5330 sinfo
->num_slabs
= nr_slabs
;
5331 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5332 sinfo
->cache_order
= oo_order(s
->oo
);
5335 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5339 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5340 size_t count
, loff_t
*ppos
)
5344 #endif /* CONFIG_SLABINFO */