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 and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page
*page
)
113 return page
->flags
& FROZEN
;
116 static inline void SetSlabFrozen(struct page
*page
)
118 page
->flags
|= FROZEN
;
121 static inline void ClearSlabFrozen(struct page
*page
)
123 page
->flags
&= ~FROZEN
;
126 static inline int SlabDebug(struct page
*page
)
128 return page
->flags
& SLABDEBUG
;
131 static inline void SetSlabDebug(struct page
*page
)
133 page
->flags
|= SLABDEBUG
;
136 static inline void ClearSlabDebug(struct page
*page
)
138 page
->flags
&= ~SLABDEBUG
;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
207 #define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208 #define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
210 /* Not all arches define cache_line_size */
211 #ifndef cache_line_size
212 #define cache_line_size() L1_CACHE_BYTES
215 static int kmem_size
= sizeof(struct kmem_cache
);
218 static struct notifier_block slab_notifier
;
222 DOWN
, /* No slab functionality available */
223 PARTIAL
, /* kmem_cache_open() works but kmalloc does not */
224 UP
, /* Everything works but does not show up in sysfs */
228 /* A list of all slab caches on the system */
229 static DECLARE_RWSEM(slub_lock
);
230 static LIST_HEAD(slab_caches
);
233 * Tracking user of a slab.
236 void *addr
; /* Called from address */
237 int cpu
; /* Was running on cpu */
238 int pid
; /* Pid context */
239 unsigned long when
; /* When did the operation occur */
242 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
244 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
245 static int sysfs_slab_add(struct kmem_cache
*);
246 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
247 static void sysfs_slab_remove(struct kmem_cache
*);
250 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
251 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
253 static inline void sysfs_slab_remove(struct kmem_cache
*s
)
260 static inline void stat(struct kmem_cache_cpu
*c
, enum stat_item si
)
262 #ifdef CONFIG_SLUB_STATS
267 /********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
271 int slab_is_available(void)
273 return slab_state
>= UP
;
276 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
279 return s
->node
[node
];
281 return &s
->local_node
;
285 static inline struct kmem_cache_cpu
*get_cpu_slab(struct kmem_cache
*s
, int cpu
)
288 return s
->cpu_slab
[cpu
];
294 /* Verify that a pointer has an address that is valid within a slab page */
295 static inline int check_valid_pointer(struct kmem_cache
*s
,
296 struct page
*page
, const void *object
)
303 base
= page_address(page
);
304 if (object
< base
|| object
>= base
+ s
->objects
* s
->size
||
305 (object
- base
) % s
->size
) {
313 * Slow version of get and set free pointer.
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
319 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
321 return *(void **)(object
+ s
->offset
);
324 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
326 *(void **)(object
+ s
->offset
) = fp
;
329 /* Loop over all objects in a slab */
330 #define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
335 #define for_each_free_object(__p, __s, __free) \
336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
338 /* Determine object index from a given position */
339 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
341 return (p
- addr
) / s
->size
;
344 #ifdef CONFIG_SLUB_DEBUG
348 #ifdef CONFIG_SLUB_DEBUG_ON
349 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
351 static int slub_debug
;
354 static char *slub_debug_slabs
;
359 static void print_section(char *text
, u8
*addr
, unsigned int length
)
367 for (i
= 0; i
< length
; i
++) {
369 printk(KERN_ERR
"%8s 0x%p: ", text
, addr
+ i
);
372 printk(KERN_CONT
" %02x", addr
[i
]);
374 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
376 printk(KERN_CONT
" %s\n", ascii
);
383 printk(KERN_CONT
" ");
387 printk(KERN_CONT
" %s\n", ascii
);
391 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
392 enum track_item alloc
)
397 p
= object
+ s
->offset
+ sizeof(void *);
399 p
= object
+ s
->inuse
;
404 static void set_track(struct kmem_cache
*s
, void *object
,
405 enum track_item alloc
, void *addr
)
410 p
= object
+ s
->offset
+ sizeof(void *);
412 p
= object
+ s
->inuse
;
417 p
->cpu
= smp_processor_id();
418 p
->pid
= current
? current
->pid
: -1;
421 memset(p
, 0, sizeof(struct track
));
424 static void init_tracking(struct kmem_cache
*s
, void *object
)
426 if (!(s
->flags
& SLAB_STORE_USER
))
429 set_track(s
, object
, TRACK_FREE
, NULL
);
430 set_track(s
, object
, TRACK_ALLOC
, NULL
);
433 static void print_track(const char *s
, struct track
*t
)
438 printk(KERN_ERR
"INFO: %s in ", s
);
439 __print_symbol("%s", (unsigned long)t
->addr
);
440 printk(" age=%lu cpu=%u pid=%d\n", jiffies
- t
->when
, t
->cpu
, t
->pid
);
443 static void print_tracking(struct kmem_cache
*s
, void *object
)
445 if (!(s
->flags
& SLAB_STORE_USER
))
448 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
449 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
452 static void print_page_info(struct page
*page
)
454 printk(KERN_ERR
"INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
455 page
, page
->inuse
, page
->freelist
, page
->flags
);
459 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
465 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
467 printk(KERN_ERR
"========================================"
468 "=====================================\n");
469 printk(KERN_ERR
"BUG %s: %s\n", s
->name
, buf
);
470 printk(KERN_ERR
"----------------------------------------"
471 "-------------------------------------\n\n");
474 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
480 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
482 printk(KERN_ERR
"FIX %s: %s\n", s
->name
, buf
);
485 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
487 unsigned int off
; /* Offset of last byte */
488 u8
*addr
= page_address(page
);
490 print_tracking(s
, p
);
492 print_page_info(page
);
494 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p
, p
- addr
, get_freepointer(s
, p
));
498 print_section("Bytes b4", p
- 16, 16);
500 print_section("Object", p
, min(s
->objsize
, 128));
502 if (s
->flags
& SLAB_RED_ZONE
)
503 print_section("Redzone", p
+ s
->objsize
,
504 s
->inuse
- s
->objsize
);
507 off
= s
->offset
+ sizeof(void *);
511 if (s
->flags
& SLAB_STORE_USER
)
512 off
+= 2 * sizeof(struct track
);
515 /* Beginning of the filler is the free pointer */
516 print_section("Padding", p
+ off
, s
->size
- off
);
521 static void object_err(struct kmem_cache
*s
, struct page
*page
,
522 u8
*object
, char *reason
)
525 print_trailer(s
, page
, object
);
528 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *fmt
, ...)
534 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
537 print_page_info(page
);
541 static void init_object(struct kmem_cache
*s
, void *object
, int active
)
545 if (s
->flags
& __OBJECT_POISON
) {
546 memset(p
, POISON_FREE
, s
->objsize
- 1);
547 p
[s
->objsize
- 1] = POISON_END
;
550 if (s
->flags
& SLAB_RED_ZONE
)
551 memset(p
+ s
->objsize
,
552 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
,
553 s
->inuse
- s
->objsize
);
556 static u8
*check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
559 if (*start
!= (u8
)value
)
567 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
568 void *from
, void *to
)
570 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
571 memset(from
, data
, to
- from
);
574 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
575 u8
*object
, char *what
,
576 u8
*start
, unsigned int value
, unsigned int bytes
)
581 fault
= check_bytes(start
, value
, bytes
);
586 while (end
> fault
&& end
[-1] == value
)
589 slab_bug(s
, "%s overwritten", what
);
590 printk(KERN_ERR
"INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault
, end
- 1, fault
[0], value
);
592 print_trailer(s
, page
, object
);
594 restore_bytes(s
, what
, value
, fault
, end
);
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
611 * Padding is extended by another word if Redzoning is enabled and
614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
618 * Meta data starts here.
620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
622 * C. Padding to reach required alignment boundary or at mininum
623 * one word if debugging is on to be able to detect writes
624 * before the word boundary.
626 * Padding is done using 0x5a (POISON_INUSE)
629 * Nothing is used beyond s->size.
631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
633 * may be used with merged slabcaches.
636 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
638 unsigned long off
= s
->inuse
; /* The end of info */
641 /* Freepointer is placed after the object. */
642 off
+= sizeof(void *);
644 if (s
->flags
& SLAB_STORE_USER
)
645 /* We also have user information there */
646 off
+= 2 * sizeof(struct track
);
651 return check_bytes_and_report(s
, page
, p
, "Object padding",
652 p
+ off
, POISON_INUSE
, s
->size
- off
);
655 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
663 if (!(s
->flags
& SLAB_POISON
))
666 start
= page_address(page
);
667 end
= start
+ (PAGE_SIZE
<< s
->order
);
668 length
= s
->objects
* s
->size
;
669 remainder
= end
- (start
+ length
);
673 fault
= check_bytes(start
+ length
, POISON_INUSE
, remainder
);
676 while (end
> fault
&& end
[-1] == POISON_INUSE
)
679 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
680 print_section("Padding", start
, length
);
682 restore_bytes(s
, "slab padding", POISON_INUSE
, start
, end
);
686 static int check_object(struct kmem_cache
*s
, struct page
*page
,
687 void *object
, int active
)
690 u8
*endobject
= object
+ s
->objsize
;
692 if (s
->flags
& SLAB_RED_ZONE
) {
694 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
696 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
697 endobject
, red
, s
->inuse
- s
->objsize
))
700 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
) {
701 check_bytes_and_report(s
, page
, p
, "Alignment padding",
702 endobject
, POISON_INUSE
, s
->inuse
- s
->objsize
);
706 if (s
->flags
& SLAB_POISON
) {
707 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
708 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
709 POISON_FREE
, s
->objsize
- 1) ||
710 !check_bytes_and_report(s
, page
, p
, "Poison",
711 p
+ s
->objsize
- 1, POISON_END
, 1)))
714 * check_pad_bytes cleans up on its own.
716 check_pad_bytes(s
, page
, p
);
719 if (!s
->offset
&& active
)
721 * Object and freepointer overlap. Cannot check
722 * freepointer while object is allocated.
726 /* Check free pointer validity */
727 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
728 object_err(s
, page
, p
, "Freepointer corrupt");
730 * No choice but to zap it and thus loose the remainder
731 * of the free objects in this slab. May cause
732 * another error because the object count is now wrong.
734 set_freepointer(s
, p
, NULL
);
740 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
742 VM_BUG_ON(!irqs_disabled());
744 if (!PageSlab(page
)) {
745 slab_err(s
, page
, "Not a valid slab page");
748 if (page
->inuse
> s
->objects
) {
749 slab_err(s
, page
, "inuse %u > max %u",
750 s
->name
, page
->inuse
, s
->objects
);
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s
, page
);
759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
762 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
765 void *fp
= page
->freelist
;
768 while (fp
&& nr
<= s
->objects
) {
771 if (!check_valid_pointer(s
, page
, fp
)) {
773 object_err(s
, page
, object
,
774 "Freechain corrupt");
775 set_freepointer(s
, object
, NULL
);
778 slab_err(s
, page
, "Freepointer corrupt");
779 page
->freelist
= NULL
;
780 page
->inuse
= s
->objects
;
781 slab_fix(s
, "Freelist cleared");
787 fp
= get_freepointer(s
, object
);
791 if (page
->inuse
!= s
->objects
- nr
) {
792 slab_err(s
, page
, "Wrong object count. Counter is %d but "
793 "counted were %d", page
->inuse
, s
->objects
- nr
);
794 page
->inuse
= s
->objects
- nr
;
795 slab_fix(s
, "Object count adjusted.");
797 return search
== NULL
;
800 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
, int alloc
)
802 if (s
->flags
& SLAB_TRACE
) {
803 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
805 alloc
? "alloc" : "free",
810 print_section("Object", (void *)object
, s
->objsize
);
817 * Tracking of fully allocated slabs for debugging purposes.
819 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
821 spin_lock(&n
->list_lock
);
822 list_add(&page
->lru
, &n
->full
);
823 spin_unlock(&n
->list_lock
);
826 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
828 struct kmem_cache_node
*n
;
830 if (!(s
->flags
& SLAB_STORE_USER
))
833 n
= get_node(s
, page_to_nid(page
));
835 spin_lock(&n
->list_lock
);
836 list_del(&page
->lru
);
837 spin_unlock(&n
->list_lock
);
840 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
843 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
846 init_object(s
, object
, 0);
847 init_tracking(s
, object
);
850 static int alloc_debug_processing(struct kmem_cache
*s
, struct page
*page
,
851 void *object
, void *addr
)
853 if (!check_slab(s
, page
))
856 if (!on_freelist(s
, page
, object
)) {
857 object_err(s
, page
, object
, "Object already allocated");
861 if (!check_valid_pointer(s
, page
, object
)) {
862 object_err(s
, page
, object
, "Freelist Pointer check fails");
866 if (!check_object(s
, page
, object
, 0))
869 /* Success perform special debug activities for allocs */
870 if (s
->flags
& SLAB_STORE_USER
)
871 set_track(s
, object
, TRACK_ALLOC
, addr
);
872 trace(s
, page
, object
, 1);
873 init_object(s
, object
, 1);
877 if (PageSlab(page
)) {
879 * If this is a slab page then lets do the best we can
880 * to avoid issues in the future. Marking all objects
881 * as used avoids touching the remaining objects.
883 slab_fix(s
, "Marking all objects used");
884 page
->inuse
= s
->objects
;
885 page
->freelist
= NULL
;
890 static int free_debug_processing(struct kmem_cache
*s
, struct page
*page
,
891 void *object
, void *addr
)
893 if (!check_slab(s
, page
))
896 if (!check_valid_pointer(s
, page
, object
)) {
897 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
901 if (on_freelist(s
, page
, object
)) {
902 object_err(s
, page
, object
, "Object already free");
906 if (!check_object(s
, page
, object
, 1))
909 if (unlikely(s
!= page
->slab
)) {
910 if (!PageSlab(page
)) {
911 slab_err(s
, page
, "Attempt to free object(0x%p) "
912 "outside of slab", object
);
913 } else if (!page
->slab
) {
915 "SLUB <none>: no slab for object 0x%p.\n",
919 object_err(s
, page
, object
,
920 "page slab pointer corrupt.");
924 /* Special debug activities for freeing objects */
925 if (!SlabFrozen(page
) && !page
->freelist
)
926 remove_full(s
, page
);
927 if (s
->flags
& SLAB_STORE_USER
)
928 set_track(s
, object
, TRACK_FREE
, addr
);
929 trace(s
, page
, object
, 0);
930 init_object(s
, object
, 0);
934 slab_fix(s
, "Object at 0x%p not freed", object
);
938 static int __init
setup_slub_debug(char *str
)
940 slub_debug
= DEBUG_DEFAULT_FLAGS
;
941 if (*str
++ != '=' || !*str
)
943 * No options specified. Switch on full debugging.
949 * No options but restriction on slabs. This means full
950 * debugging for slabs matching a pattern.
957 * Switch off all debugging measures.
962 * Determine which debug features should be switched on
964 for (; *str
&& *str
!= ','; str
++) {
965 switch (tolower(*str
)) {
967 slub_debug
|= SLAB_DEBUG_FREE
;
970 slub_debug
|= SLAB_RED_ZONE
;
973 slub_debug
|= SLAB_POISON
;
976 slub_debug
|= SLAB_STORE_USER
;
979 slub_debug
|= SLAB_TRACE
;
982 printk(KERN_ERR
"slub_debug option '%c' "
983 "unknown. skipped\n", *str
);
989 slub_debug_slabs
= str
+ 1;
994 __setup("slub_debug", setup_slub_debug
);
996 static unsigned long kmem_cache_flags(unsigned long objsize
,
997 unsigned long flags
, const char *name
,
998 void (*ctor
)(struct kmem_cache
*, void *))
1001 * Enable debugging if selected on the kernel commandline.
1003 if (slub_debug
&& (!slub_debug_slabs
||
1004 strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)) == 0))
1005 flags
|= slub_debug
;
1010 static inline void setup_object_debug(struct kmem_cache
*s
,
1011 struct page
*page
, void *object
) {}
1013 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1014 struct page
*page
, void *object
, void *addr
) { return 0; }
1016 static inline int free_debug_processing(struct kmem_cache
*s
,
1017 struct page
*page
, void *object
, void *addr
) { return 0; }
1019 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1021 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1022 void *object
, int active
) { return 1; }
1023 static inline void add_full(struct kmem_cache_node
*n
, struct page
*page
) {}
1024 static inline unsigned long kmem_cache_flags(unsigned long objsize
,
1025 unsigned long flags
, const char *name
,
1026 void (*ctor
)(struct kmem_cache
*, void *))
1030 #define slub_debug 0
1033 * Slab allocation and freeing
1035 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1038 int pages
= 1 << s
->order
;
1040 flags
|= s
->allocflags
;
1043 page
= alloc_pages(flags
, s
->order
);
1045 page
= alloc_pages_node(node
, flags
, s
->order
);
1050 mod_zone_page_state(page_zone(page
),
1051 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1052 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1058 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1061 setup_object_debug(s
, page
, object
);
1062 if (unlikely(s
->ctor
))
1066 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1069 struct kmem_cache_node
*n
;
1074 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1076 page
= allocate_slab(s
,
1077 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1081 n
= get_node(s
, page_to_nid(page
));
1083 atomic_long_inc(&n
->nr_slabs
);
1085 page
->flags
|= 1 << PG_slab
;
1086 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
1087 SLAB_STORE_USER
| SLAB_TRACE
))
1090 start
= page_address(page
);
1092 if (unlikely(s
->flags
& SLAB_POISON
))
1093 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
1096 for_each_object(p
, s
, start
) {
1097 setup_object(s
, page
, last
);
1098 set_freepointer(s
, last
, p
);
1101 setup_object(s
, page
, last
);
1102 set_freepointer(s
, last
, NULL
);
1104 page
->freelist
= start
;
1110 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1112 int pages
= 1 << s
->order
;
1114 if (unlikely(SlabDebug(page
))) {
1117 slab_pad_check(s
, page
);
1118 for_each_object(p
, s
, page_address(page
))
1119 check_object(s
, page
, p
, 0);
1120 ClearSlabDebug(page
);
1123 mod_zone_page_state(page_zone(page
),
1124 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1125 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1128 __free_pages(page
, s
->order
);
1131 static void rcu_free_slab(struct rcu_head
*h
)
1135 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1136 __free_slab(page
->slab
, page
);
1139 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1141 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1143 * RCU free overloads the RCU head over the LRU
1145 struct rcu_head
*head
= (void *)&page
->lru
;
1147 call_rcu(head
, rcu_free_slab
);
1149 __free_slab(s
, page
);
1152 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1154 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1156 atomic_long_dec(&n
->nr_slabs
);
1157 reset_page_mapcount(page
);
1158 __ClearPageSlab(page
);
1163 * Per slab locking using the pagelock
1165 static __always_inline
void slab_lock(struct page
*page
)
1167 bit_spin_lock(PG_locked
, &page
->flags
);
1170 static __always_inline
void slab_unlock(struct page
*page
)
1172 __bit_spin_unlock(PG_locked
, &page
->flags
);
1175 static __always_inline
int slab_trylock(struct page
*page
)
1179 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
1184 * Management of partially allocated slabs
1186 static void add_partial(struct kmem_cache_node
*n
,
1187 struct page
*page
, int tail
)
1189 spin_lock(&n
->list_lock
);
1192 list_add_tail(&page
->lru
, &n
->partial
);
1194 list_add(&page
->lru
, &n
->partial
);
1195 spin_unlock(&n
->list_lock
);
1198 static void remove_partial(struct kmem_cache
*s
,
1201 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1203 spin_lock(&n
->list_lock
);
1204 list_del(&page
->lru
);
1206 spin_unlock(&n
->list_lock
);
1210 * Lock slab and remove from the partial list.
1212 * Must hold list_lock.
1214 static inline int lock_and_freeze_slab(struct kmem_cache_node
*n
, struct page
*page
)
1216 if (slab_trylock(page
)) {
1217 list_del(&page
->lru
);
1219 SetSlabFrozen(page
);
1226 * Try to allocate a partial slab from a specific node.
1228 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
1233 * Racy check. If we mistakenly see no partial slabs then we
1234 * just allocate an empty slab. If we mistakenly try to get a
1235 * partial slab and there is none available then get_partials()
1238 if (!n
|| !n
->nr_partial
)
1241 spin_lock(&n
->list_lock
);
1242 list_for_each_entry(page
, &n
->partial
, lru
)
1243 if (lock_and_freeze_slab(n
, page
))
1247 spin_unlock(&n
->list_lock
);
1252 * Get a page from somewhere. Search in increasing NUMA distances.
1254 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1257 struct zonelist
*zonelist
;
1262 * The defrag ratio allows a configuration of the tradeoffs between
1263 * inter node defragmentation and node local allocations. A lower
1264 * defrag_ratio increases the tendency to do local allocations
1265 * instead of attempting to obtain partial slabs from other nodes.
1267 * If the defrag_ratio is set to 0 then kmalloc() always
1268 * returns node local objects. If the ratio is higher then kmalloc()
1269 * may return off node objects because partial slabs are obtained
1270 * from other nodes and filled up.
1272 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1273 * defrag_ratio = 1000) then every (well almost) allocation will
1274 * first attempt to defrag slab caches on other nodes. This means
1275 * scanning over all nodes to look for partial slabs which may be
1276 * expensive if we do it every time we are trying to find a slab
1277 * with available objects.
1279 if (!s
->remote_node_defrag_ratio
||
1280 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1283 zonelist
= &NODE_DATA(
1284 slab_node(current
->mempolicy
))->node_zonelists
[gfp_zone(flags
)];
1285 for (z
= zonelist
->zones
; *z
; z
++) {
1286 struct kmem_cache_node
*n
;
1288 n
= get_node(s
, zone_to_nid(*z
));
1290 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1291 n
->nr_partial
> MIN_PARTIAL
) {
1292 page
= get_partial_node(n
);
1302 * Get a partial page, lock it and return it.
1304 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1307 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1309 page
= get_partial_node(get_node(s
, searchnode
));
1310 if (page
|| (flags
& __GFP_THISNODE
))
1313 return get_any_partial(s
, flags
);
1317 * Move a page back to the lists.
1319 * Must be called with the slab lock held.
1321 * On exit the slab lock will have been dropped.
1323 static void unfreeze_slab(struct kmem_cache
*s
, struct page
*page
, int tail
)
1325 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1326 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, smp_processor_id());
1328 ClearSlabFrozen(page
);
1331 if (page
->freelist
) {
1332 add_partial(n
, page
, tail
);
1333 stat(c
, tail
? DEACTIVATE_TO_TAIL
: DEACTIVATE_TO_HEAD
);
1335 stat(c
, DEACTIVATE_FULL
);
1336 if (SlabDebug(page
) && (s
->flags
& SLAB_STORE_USER
))
1341 stat(c
, DEACTIVATE_EMPTY
);
1342 if (n
->nr_partial
< MIN_PARTIAL
) {
1344 * Adding an empty slab to the partial slabs in order
1345 * to avoid page allocator overhead. This slab needs
1346 * to come after the other slabs with objects in
1347 * so that the others get filled first. That way the
1348 * size of the partial list stays small.
1350 * kmem_cache_shrink can reclaim any empty slabs from the
1353 add_partial(n
, page
, 1);
1357 stat(get_cpu_slab(s
, raw_smp_processor_id()), FREE_SLAB
);
1358 discard_slab(s
, page
);
1364 * Remove the cpu slab
1366 static void deactivate_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1368 struct page
*page
= c
->page
;
1372 stat(c
, DEACTIVATE_REMOTE_FREES
);
1374 * Merge cpu freelist into slab freelist. Typically we get here
1375 * because both freelists are empty. So this is unlikely
1378 while (unlikely(c
->freelist
)) {
1381 tail
= 0; /* Hot objects. Put the slab first */
1383 /* Retrieve object from cpu_freelist */
1384 object
= c
->freelist
;
1385 c
->freelist
= c
->freelist
[c
->offset
];
1387 /* And put onto the regular freelist */
1388 object
[c
->offset
] = page
->freelist
;
1389 page
->freelist
= object
;
1393 unfreeze_slab(s
, page
, tail
);
1396 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1398 stat(c
, CPUSLAB_FLUSH
);
1400 deactivate_slab(s
, c
);
1406 * Called from IPI handler with interrupts disabled.
1408 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1410 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
1412 if (likely(c
&& c
->page
))
1416 static void flush_cpu_slab(void *d
)
1418 struct kmem_cache
*s
= d
;
1420 __flush_cpu_slab(s
, smp_processor_id());
1423 static void flush_all(struct kmem_cache
*s
)
1426 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1428 unsigned long flags
;
1430 local_irq_save(flags
);
1432 local_irq_restore(flags
);
1437 * Check if the objects in a per cpu structure fit numa
1438 * locality expectations.
1440 static inline int node_match(struct kmem_cache_cpu
*c
, int node
)
1443 if (node
!= -1 && c
->node
!= node
)
1450 * Slow path. The lockless freelist is empty or we need to perform
1453 * Interrupts are disabled.
1455 * Processing is still very fast if new objects have been freed to the
1456 * regular freelist. In that case we simply take over the regular freelist
1457 * as the lockless freelist and zap the regular freelist.
1459 * If that is not working then we fall back to the partial lists. We take the
1460 * first element of the freelist as the object to allocate now and move the
1461 * rest of the freelist to the lockless freelist.
1463 * And if we were unable to get a new slab from the partial slab lists then
1464 * we need to allocate a new slab. This is the slowest path since it involves
1465 * a call to the page allocator and the setup of a new slab.
1467 static void *__slab_alloc(struct kmem_cache
*s
,
1468 gfp_t gfpflags
, int node
, void *addr
, struct kmem_cache_cpu
*c
)
1473 /* We handle __GFP_ZERO in the caller */
1474 gfpflags
&= ~__GFP_ZERO
;
1480 if (unlikely(!node_match(c
, node
)))
1483 stat(c
, ALLOC_REFILL
);
1486 object
= c
->page
->freelist
;
1487 if (unlikely(!object
))
1489 if (unlikely(SlabDebug(c
->page
)))
1492 c
->freelist
= object
[c
->offset
];
1493 c
->page
->inuse
= s
->objects
;
1494 c
->page
->freelist
= NULL
;
1495 c
->node
= page_to_nid(c
->page
);
1497 slab_unlock(c
->page
);
1498 stat(c
, ALLOC_SLOWPATH
);
1502 deactivate_slab(s
, c
);
1505 new = get_partial(s
, gfpflags
, node
);
1508 stat(c
, ALLOC_FROM_PARTIAL
);
1512 if (gfpflags
& __GFP_WAIT
)
1515 new = new_slab(s
, gfpflags
, node
);
1517 if (gfpflags
& __GFP_WAIT
)
1518 local_irq_disable();
1521 c
= get_cpu_slab(s
, smp_processor_id());
1522 stat(c
, ALLOC_SLAB
);
1532 * No memory available.
1534 * If the slab uses higher order allocs but the object is
1535 * smaller than a page size then we can fallback in emergencies
1536 * to the page allocator via kmalloc_large. The page allocator may
1537 * have failed to obtain a higher order page and we can try to
1538 * allocate a single page if the object fits into a single page.
1539 * That is only possible if certain conditions are met that are being
1540 * checked when a slab is created.
1542 if (!(gfpflags
& __GFP_NORETRY
) &&
1543 (s
->flags
& __PAGE_ALLOC_FALLBACK
)) {
1544 if (gfpflags
& __GFP_WAIT
)
1546 object
= kmalloc_large(s
->objsize
, gfpflags
);
1547 if (gfpflags
& __GFP_WAIT
)
1548 local_irq_disable();
1553 if (!alloc_debug_processing(s
, c
->page
, object
, addr
))
1557 c
->page
->freelist
= object
[c
->offset
];
1563 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1564 * have the fastpath folded into their functions. So no function call
1565 * overhead for requests that can be satisfied on the fastpath.
1567 * The fastpath works by first checking if the lockless freelist can be used.
1568 * If not then __slab_alloc is called for slow processing.
1570 * Otherwise we can simply pick the next object from the lockless free list.
1572 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
1573 gfp_t gfpflags
, int node
, void *addr
)
1576 struct kmem_cache_cpu
*c
;
1577 unsigned long flags
;
1578 unsigned int objsize
;
1580 local_irq_save(flags
);
1581 c
= get_cpu_slab(s
, smp_processor_id());
1582 objsize
= c
->objsize
;
1583 if (unlikely(!c
->freelist
|| !node_match(c
, node
)))
1585 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
1588 object
= c
->freelist
;
1589 c
->freelist
= object
[c
->offset
];
1590 stat(c
, ALLOC_FASTPATH
);
1592 local_irq_restore(flags
);
1594 if (unlikely((gfpflags
& __GFP_ZERO
) && object
))
1595 memset(object
, 0, objsize
);
1600 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1602 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1604 EXPORT_SYMBOL(kmem_cache_alloc
);
1607 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1609 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1611 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1615 * Slow patch handling. This may still be called frequently since objects
1616 * have a longer lifetime than the cpu slabs in most processing loads.
1618 * So we still attempt to reduce cache line usage. Just take the slab
1619 * lock and free the item. If there is no additional partial page
1620 * handling required then we can return immediately.
1622 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
1623 void *x
, void *addr
, unsigned int offset
)
1626 void **object
= (void *)x
;
1627 struct kmem_cache_cpu
*c
;
1629 c
= get_cpu_slab(s
, raw_smp_processor_id());
1630 stat(c
, FREE_SLOWPATH
);
1633 if (unlikely(SlabDebug(page
)))
1637 prior
= object
[offset
] = page
->freelist
;
1638 page
->freelist
= object
;
1641 if (unlikely(SlabFrozen(page
))) {
1642 stat(c
, FREE_FROZEN
);
1646 if (unlikely(!page
->inuse
))
1650 * Objects left in the slab. If it was not on the partial list before
1653 if (unlikely(!prior
)) {
1654 add_partial(get_node(s
, page_to_nid(page
)), page
, 1);
1655 stat(c
, FREE_ADD_PARTIAL
);
1665 * Slab still on the partial list.
1667 remove_partial(s
, page
);
1668 stat(c
, FREE_REMOVE_PARTIAL
);
1672 discard_slab(s
, page
);
1676 if (!free_debug_processing(s
, page
, x
, addr
))
1682 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1683 * can perform fastpath freeing without additional function calls.
1685 * The fastpath is only possible if we are freeing to the current cpu slab
1686 * of this processor. This typically the case if we have just allocated
1689 * If fastpath is not possible then fall back to __slab_free where we deal
1690 * with all sorts of special processing.
1692 static __always_inline
void slab_free(struct kmem_cache
*s
,
1693 struct page
*page
, void *x
, void *addr
)
1695 void **object
= (void *)x
;
1696 struct kmem_cache_cpu
*c
;
1697 unsigned long flags
;
1699 local_irq_save(flags
);
1700 c
= get_cpu_slab(s
, smp_processor_id());
1701 debug_check_no_locks_freed(object
, c
->objsize
);
1702 if (likely(page
== c
->page
&& c
->node
>= 0)) {
1703 object
[c
->offset
] = c
->freelist
;
1704 c
->freelist
= object
;
1705 stat(c
, FREE_FASTPATH
);
1707 __slab_free(s
, page
, x
, addr
, c
->offset
);
1709 local_irq_restore(flags
);
1712 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1716 page
= virt_to_head_page(x
);
1718 slab_free(s
, page
, x
, __builtin_return_address(0));
1720 EXPORT_SYMBOL(kmem_cache_free
);
1722 /* Figure out on which slab object the object resides */
1723 static struct page
*get_object_page(const void *x
)
1725 struct page
*page
= virt_to_head_page(x
);
1727 if (!PageSlab(page
))
1734 * Object placement in a slab is made very easy because we always start at
1735 * offset 0. If we tune the size of the object to the alignment then we can
1736 * get the required alignment by putting one properly sized object after
1739 * Notice that the allocation order determines the sizes of the per cpu
1740 * caches. Each processor has always one slab available for allocations.
1741 * Increasing the allocation order reduces the number of times that slabs
1742 * must be moved on and off the partial lists and is therefore a factor in
1747 * Mininum / Maximum order of slab pages. This influences locking overhead
1748 * and slab fragmentation. A higher order reduces the number of partial slabs
1749 * and increases the number of allocations possible without having to
1750 * take the list_lock.
1752 static int slub_min_order
;
1753 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1754 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1757 * Merge control. If this is set then no merging of slab caches will occur.
1758 * (Could be removed. This was introduced to pacify the merge skeptics.)
1760 static int slub_nomerge
;
1763 * Calculate the order of allocation given an slab object size.
1765 * The order of allocation has significant impact on performance and other
1766 * system components. Generally order 0 allocations should be preferred since
1767 * order 0 does not cause fragmentation in the page allocator. Larger objects
1768 * be problematic to put into order 0 slabs because there may be too much
1769 * unused space left. We go to a higher order if more than 1/8th of the slab
1772 * In order to reach satisfactory performance we must ensure that a minimum
1773 * number of objects is in one slab. Otherwise we may generate too much
1774 * activity on the partial lists which requires taking the list_lock. This is
1775 * less a concern for large slabs though which are rarely used.
1777 * slub_max_order specifies the order where we begin to stop considering the
1778 * number of objects in a slab as critical. If we reach slub_max_order then
1779 * we try to keep the page order as low as possible. So we accept more waste
1780 * of space in favor of a small page order.
1782 * Higher order allocations also allow the placement of more objects in a
1783 * slab and thereby reduce object handling overhead. If the user has
1784 * requested a higher mininum order then we start with that one instead of
1785 * the smallest order which will fit the object.
1787 static inline int slab_order(int size
, int min_objects
,
1788 int max_order
, int fract_leftover
)
1792 int min_order
= slub_min_order
;
1794 for (order
= max(min_order
,
1795 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
1796 order
<= max_order
; order
++) {
1798 unsigned long slab_size
= PAGE_SIZE
<< order
;
1800 if (slab_size
< min_objects
* size
)
1803 rem
= slab_size
% size
;
1805 if (rem
<= slab_size
/ fract_leftover
)
1813 static inline int calculate_order(int size
)
1820 * Attempt to find best configuration for a slab. This
1821 * works by first attempting to generate a layout with
1822 * the best configuration and backing off gradually.
1824 * First we reduce the acceptable waste in a slab. Then
1825 * we reduce the minimum objects required in a slab.
1827 min_objects
= slub_min_objects
;
1828 while (min_objects
> 1) {
1830 while (fraction
>= 4) {
1831 order
= slab_order(size
, min_objects
,
1832 slub_max_order
, fraction
);
1833 if (order
<= slub_max_order
)
1841 * We were unable to place multiple objects in a slab. Now
1842 * lets see if we can place a single object there.
1844 order
= slab_order(size
, 1, slub_max_order
, 1);
1845 if (order
<= slub_max_order
)
1849 * Doh this slab cannot be placed using slub_max_order.
1851 order
= slab_order(size
, 1, MAX_ORDER
, 1);
1852 if (order
<= MAX_ORDER
)
1858 * Figure out what the alignment of the objects will be.
1860 static unsigned long calculate_alignment(unsigned long flags
,
1861 unsigned long align
, unsigned long size
)
1864 * If the user wants hardware cache aligned objects then follow that
1865 * suggestion if the object is sufficiently large.
1867 * The hardware cache alignment cannot override the specified
1868 * alignment though. If that is greater then use it.
1870 if (flags
& SLAB_HWCACHE_ALIGN
) {
1871 unsigned long ralign
= cache_line_size();
1872 while (size
<= ralign
/ 2)
1874 align
= max(align
, ralign
);
1877 if (align
< ARCH_SLAB_MINALIGN
)
1878 align
= ARCH_SLAB_MINALIGN
;
1880 return ALIGN(align
, sizeof(void *));
1883 static void init_kmem_cache_cpu(struct kmem_cache
*s
,
1884 struct kmem_cache_cpu
*c
)
1889 c
->offset
= s
->offset
/ sizeof(void *);
1890 c
->objsize
= s
->objsize
;
1893 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1896 atomic_long_set(&n
->nr_slabs
, 0);
1897 spin_lock_init(&n
->list_lock
);
1898 INIT_LIST_HEAD(&n
->partial
);
1899 #ifdef CONFIG_SLUB_DEBUG
1900 INIT_LIST_HEAD(&n
->full
);
1906 * Per cpu array for per cpu structures.
1908 * The per cpu array places all kmem_cache_cpu structures from one processor
1909 * close together meaning that it becomes possible that multiple per cpu
1910 * structures are contained in one cacheline. This may be particularly
1911 * beneficial for the kmalloc caches.
1913 * A desktop system typically has around 60-80 slabs. With 100 here we are
1914 * likely able to get per cpu structures for all caches from the array defined
1915 * here. We must be able to cover all kmalloc caches during bootstrap.
1917 * If the per cpu array is exhausted then fall back to kmalloc
1918 * of individual cachelines. No sharing is possible then.
1920 #define NR_KMEM_CACHE_CPU 100
1922 static DEFINE_PER_CPU(struct kmem_cache_cpu
,
1923 kmem_cache_cpu
)[NR_KMEM_CACHE_CPU
];
1925 static DEFINE_PER_CPU(struct kmem_cache_cpu
*, kmem_cache_cpu_free
);
1926 static cpumask_t kmem_cach_cpu_free_init_once
= CPU_MASK_NONE
;
1928 static struct kmem_cache_cpu
*alloc_kmem_cache_cpu(struct kmem_cache
*s
,
1929 int cpu
, gfp_t flags
)
1931 struct kmem_cache_cpu
*c
= per_cpu(kmem_cache_cpu_free
, cpu
);
1934 per_cpu(kmem_cache_cpu_free
, cpu
) =
1935 (void *)c
->freelist
;
1937 /* Table overflow: So allocate ourselves */
1939 ALIGN(sizeof(struct kmem_cache_cpu
), cache_line_size()),
1940 flags
, cpu_to_node(cpu
));
1945 init_kmem_cache_cpu(s
, c
);
1949 static void free_kmem_cache_cpu(struct kmem_cache_cpu
*c
, int cpu
)
1951 if (c
< per_cpu(kmem_cache_cpu
, cpu
) ||
1952 c
> per_cpu(kmem_cache_cpu
, cpu
) + NR_KMEM_CACHE_CPU
) {
1956 c
->freelist
= (void *)per_cpu(kmem_cache_cpu_free
, cpu
);
1957 per_cpu(kmem_cache_cpu_free
, cpu
) = c
;
1960 static void free_kmem_cache_cpus(struct kmem_cache
*s
)
1964 for_each_online_cpu(cpu
) {
1965 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
1968 s
->cpu_slab
[cpu
] = NULL
;
1969 free_kmem_cache_cpu(c
, cpu
);
1974 static int alloc_kmem_cache_cpus(struct kmem_cache
*s
, gfp_t flags
)
1978 for_each_online_cpu(cpu
) {
1979 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
1984 c
= alloc_kmem_cache_cpu(s
, cpu
, flags
);
1986 free_kmem_cache_cpus(s
);
1989 s
->cpu_slab
[cpu
] = c
;
1995 * Initialize the per cpu array.
1997 static void init_alloc_cpu_cpu(int cpu
)
2001 if (cpu_isset(cpu
, kmem_cach_cpu_free_init_once
))
2004 for (i
= NR_KMEM_CACHE_CPU
- 1; i
>= 0; i
--)
2005 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu
, cpu
)[i
], cpu
);
2007 cpu_set(cpu
, kmem_cach_cpu_free_init_once
);
2010 static void __init
init_alloc_cpu(void)
2014 for_each_online_cpu(cpu
)
2015 init_alloc_cpu_cpu(cpu
);
2019 static inline void free_kmem_cache_cpus(struct kmem_cache
*s
) {}
2020 static inline void init_alloc_cpu(void) {}
2022 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
, gfp_t flags
)
2024 init_kmem_cache_cpu(s
, &s
->cpu_slab
);
2031 * No kmalloc_node yet so do it by hand. We know that this is the first
2032 * slab on the node for this slabcache. There are no concurrent accesses
2035 * Note that this function only works on the kmalloc_node_cache
2036 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2037 * memory on a fresh node that has no slab structures yet.
2039 static struct kmem_cache_node
*early_kmem_cache_node_alloc(gfp_t gfpflags
,
2043 struct kmem_cache_node
*n
;
2044 unsigned long flags
;
2046 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
2048 page
= new_slab(kmalloc_caches
, gfpflags
, node
);
2051 if (page_to_nid(page
) != node
) {
2052 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
2054 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
2055 "in order to be able to continue\n");
2060 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
2062 kmalloc_caches
->node
[node
] = n
;
2063 #ifdef CONFIG_SLUB_DEBUG
2064 init_object(kmalloc_caches
, n
, 1);
2065 init_tracking(kmalloc_caches
, n
);
2067 init_kmem_cache_node(n
);
2068 atomic_long_inc(&n
->nr_slabs
);
2071 * lockdep requires consistent irq usage for each lock
2072 * so even though there cannot be a race this early in
2073 * the boot sequence, we still disable irqs.
2075 local_irq_save(flags
);
2076 add_partial(n
, page
, 0);
2077 local_irq_restore(flags
);
2081 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2085 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2086 struct kmem_cache_node
*n
= s
->node
[node
];
2087 if (n
&& n
!= &s
->local_node
)
2088 kmem_cache_free(kmalloc_caches
, n
);
2089 s
->node
[node
] = NULL
;
2093 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
2098 if (slab_state
>= UP
)
2099 local_node
= page_to_nid(virt_to_page(s
));
2103 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2104 struct kmem_cache_node
*n
;
2106 if (local_node
== node
)
2109 if (slab_state
== DOWN
) {
2110 n
= early_kmem_cache_node_alloc(gfpflags
,
2114 n
= kmem_cache_alloc_node(kmalloc_caches
,
2118 free_kmem_cache_nodes(s
);
2124 init_kmem_cache_node(n
);
2129 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2133 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
2135 init_kmem_cache_node(&s
->local_node
);
2141 * calculate_sizes() determines the order and the distribution of data within
2144 static int calculate_sizes(struct kmem_cache
*s
)
2146 unsigned long flags
= s
->flags
;
2147 unsigned long size
= s
->objsize
;
2148 unsigned long align
= s
->align
;
2151 * Round up object size to the next word boundary. We can only
2152 * place the free pointer at word boundaries and this determines
2153 * the possible location of the free pointer.
2155 size
= ALIGN(size
, sizeof(void *));
2157 #ifdef CONFIG_SLUB_DEBUG
2159 * Determine if we can poison the object itself. If the user of
2160 * the slab may touch the object after free or before allocation
2161 * then we should never poison the object itself.
2163 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2165 s
->flags
|= __OBJECT_POISON
;
2167 s
->flags
&= ~__OBJECT_POISON
;
2171 * If we are Redzoning then check if there is some space between the
2172 * end of the object and the free pointer. If not then add an
2173 * additional word to have some bytes to store Redzone information.
2175 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
2176 size
+= sizeof(void *);
2180 * With that we have determined the number of bytes in actual use
2181 * by the object. This is the potential offset to the free pointer.
2185 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
2188 * Relocate free pointer after the object if it is not
2189 * permitted to overwrite the first word of the object on
2192 * This is the case if we do RCU, have a constructor or
2193 * destructor or are poisoning the objects.
2196 size
+= sizeof(void *);
2199 #ifdef CONFIG_SLUB_DEBUG
2200 if (flags
& SLAB_STORE_USER
)
2202 * Need to store information about allocs and frees after
2205 size
+= 2 * sizeof(struct track
);
2207 if (flags
& SLAB_RED_ZONE
)
2209 * Add some empty padding so that we can catch
2210 * overwrites from earlier objects rather than let
2211 * tracking information or the free pointer be
2212 * corrupted if an user writes before the start
2215 size
+= sizeof(void *);
2219 * Determine the alignment based on various parameters that the
2220 * user specified and the dynamic determination of cache line size
2223 align
= calculate_alignment(flags
, align
, s
->objsize
);
2226 * SLUB stores one object immediately after another beginning from
2227 * offset 0. In order to align the objects we have to simply size
2228 * each object to conform to the alignment.
2230 size
= ALIGN(size
, align
);
2233 if ((flags
& __KMALLOC_CACHE
) &&
2234 PAGE_SIZE
/ size
< slub_min_objects
) {
2236 * Kmalloc cache that would not have enough objects in
2237 * an order 0 page. Kmalloc slabs can fallback to
2238 * page allocator order 0 allocs so take a reasonably large
2239 * order that will allows us a good number of objects.
2241 s
->order
= max(slub_max_order
, PAGE_ALLOC_COSTLY_ORDER
);
2242 s
->flags
|= __PAGE_ALLOC_FALLBACK
;
2243 s
->allocflags
|= __GFP_NOWARN
;
2245 s
->order
= calculate_order(size
);
2252 s
->allocflags
|= __GFP_COMP
;
2254 if (s
->flags
& SLAB_CACHE_DMA
)
2255 s
->allocflags
|= SLUB_DMA
;
2257 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
2258 s
->allocflags
|= __GFP_RECLAIMABLE
;
2261 * Determine the number of objects per slab
2263 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
2265 return !!s
->objects
;
2269 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
2270 const char *name
, size_t size
,
2271 size_t align
, unsigned long flags
,
2272 void (*ctor
)(struct kmem_cache
*, void *))
2274 memset(s
, 0, kmem_size
);
2279 s
->flags
= kmem_cache_flags(size
, flags
, name
, ctor
);
2281 if (!calculate_sizes(s
))
2286 s
->remote_node_defrag_ratio
= 100;
2288 if (!init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
2291 if (alloc_kmem_cache_cpus(s
, gfpflags
& ~SLUB_DMA
))
2293 free_kmem_cache_nodes(s
);
2295 if (flags
& SLAB_PANIC
)
2296 panic("Cannot create slab %s size=%lu realsize=%u "
2297 "order=%u offset=%u flags=%lx\n",
2298 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
2304 * Check if a given pointer is valid
2306 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
2310 page
= get_object_page(object
);
2312 if (!page
|| s
!= page
->slab
)
2313 /* No slab or wrong slab */
2316 if (!check_valid_pointer(s
, page
, object
))
2320 * We could also check if the object is on the slabs freelist.
2321 * But this would be too expensive and it seems that the main
2322 * purpose of kmem_ptr_valid() is to check if the object belongs
2323 * to a certain slab.
2327 EXPORT_SYMBOL(kmem_ptr_validate
);
2330 * Determine the size of a slab object
2332 unsigned int kmem_cache_size(struct kmem_cache
*s
)
2336 EXPORT_SYMBOL(kmem_cache_size
);
2338 const char *kmem_cache_name(struct kmem_cache
*s
)
2342 EXPORT_SYMBOL(kmem_cache_name
);
2345 * Attempt to free all slabs on a node. Return the number of slabs we
2346 * were unable to free.
2348 static int free_list(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
2349 struct list_head
*list
)
2351 int slabs_inuse
= 0;
2352 unsigned long flags
;
2353 struct page
*page
, *h
;
2355 spin_lock_irqsave(&n
->list_lock
, flags
);
2356 list_for_each_entry_safe(page
, h
, list
, lru
)
2358 list_del(&page
->lru
);
2359 discard_slab(s
, page
);
2362 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2367 * Release all resources used by a slab cache.
2369 static inline int kmem_cache_close(struct kmem_cache
*s
)
2375 /* Attempt to free all objects */
2376 free_kmem_cache_cpus(s
);
2377 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2378 struct kmem_cache_node
*n
= get_node(s
, node
);
2380 n
->nr_partial
-= free_list(s
, n
, &n
->partial
);
2381 if (atomic_long_read(&n
->nr_slabs
))
2384 free_kmem_cache_nodes(s
);
2389 * Close a cache and release the kmem_cache structure
2390 * (must be used for caches created using kmem_cache_create)
2392 void kmem_cache_destroy(struct kmem_cache
*s
)
2394 down_write(&slub_lock
);
2398 up_write(&slub_lock
);
2399 if (kmem_cache_close(s
))
2401 sysfs_slab_remove(s
);
2403 up_write(&slub_lock
);
2405 EXPORT_SYMBOL(kmem_cache_destroy
);
2407 /********************************************************************
2409 *******************************************************************/
2411 struct kmem_cache kmalloc_caches
[PAGE_SHIFT
+ 1] __cacheline_aligned
;
2412 EXPORT_SYMBOL(kmalloc_caches
);
2414 #ifdef CONFIG_ZONE_DMA
2415 static struct kmem_cache
*kmalloc_caches_dma
[PAGE_SHIFT
+ 1];
2418 static int __init
setup_slub_min_order(char *str
)
2420 get_option(&str
, &slub_min_order
);
2425 __setup("slub_min_order=", setup_slub_min_order
);
2427 static int __init
setup_slub_max_order(char *str
)
2429 get_option(&str
, &slub_max_order
);
2434 __setup("slub_max_order=", setup_slub_max_order
);
2436 static int __init
setup_slub_min_objects(char *str
)
2438 get_option(&str
, &slub_min_objects
);
2443 __setup("slub_min_objects=", setup_slub_min_objects
);
2445 static int __init
setup_slub_nomerge(char *str
)
2451 __setup("slub_nomerge", setup_slub_nomerge
);
2453 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
2454 const char *name
, int size
, gfp_t gfp_flags
)
2456 unsigned int flags
= 0;
2458 if (gfp_flags
& SLUB_DMA
)
2459 flags
= SLAB_CACHE_DMA
;
2461 down_write(&slub_lock
);
2462 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
2463 flags
| __KMALLOC_CACHE
, NULL
))
2466 list_add(&s
->list
, &slab_caches
);
2467 up_write(&slub_lock
);
2468 if (sysfs_slab_add(s
))
2473 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2476 #ifdef CONFIG_ZONE_DMA
2478 static void sysfs_add_func(struct work_struct
*w
)
2480 struct kmem_cache
*s
;
2482 down_write(&slub_lock
);
2483 list_for_each_entry(s
, &slab_caches
, list
) {
2484 if (s
->flags
& __SYSFS_ADD_DEFERRED
) {
2485 s
->flags
&= ~__SYSFS_ADD_DEFERRED
;
2489 up_write(&slub_lock
);
2492 static DECLARE_WORK(sysfs_add_work
, sysfs_add_func
);
2494 static noinline
struct kmem_cache
*dma_kmalloc_cache(int index
, gfp_t flags
)
2496 struct kmem_cache
*s
;
2500 s
= kmalloc_caches_dma
[index
];
2504 /* Dynamically create dma cache */
2505 if (flags
& __GFP_WAIT
)
2506 down_write(&slub_lock
);
2508 if (!down_write_trylock(&slub_lock
))
2512 if (kmalloc_caches_dma
[index
])
2515 realsize
= kmalloc_caches
[index
].objsize
;
2516 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d",
2517 (unsigned int)realsize
);
2518 s
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2520 if (!s
|| !text
|| !kmem_cache_open(s
, flags
, text
,
2521 realsize
, ARCH_KMALLOC_MINALIGN
,
2522 SLAB_CACHE_DMA
|__SYSFS_ADD_DEFERRED
, NULL
)) {
2528 list_add(&s
->list
, &slab_caches
);
2529 kmalloc_caches_dma
[index
] = s
;
2531 schedule_work(&sysfs_add_work
);
2534 up_write(&slub_lock
);
2536 return kmalloc_caches_dma
[index
];
2541 * Conversion table for small slabs sizes / 8 to the index in the
2542 * kmalloc array. This is necessary for slabs < 192 since we have non power
2543 * of two cache sizes there. The size of larger slabs can be determined using
2546 static s8 size_index
[24] = {
2573 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2579 return ZERO_SIZE_PTR
;
2581 index
= size_index
[(size
- 1) / 8];
2583 index
= fls(size
- 1);
2585 #ifdef CONFIG_ZONE_DMA
2586 if (unlikely((flags
& SLUB_DMA
)))
2587 return dma_kmalloc_cache(index
, flags
);
2590 return &kmalloc_caches
[index
];
2593 void *__kmalloc(size_t size
, gfp_t flags
)
2595 struct kmem_cache
*s
;
2597 if (unlikely(size
> PAGE_SIZE
))
2598 return kmalloc_large(size
, flags
);
2600 s
= get_slab(size
, flags
);
2602 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2605 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2607 EXPORT_SYMBOL(__kmalloc
);
2609 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
2611 struct page
*page
= alloc_pages_node(node
, flags
| __GFP_COMP
,
2615 return page_address(page
);
2621 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2623 struct kmem_cache
*s
;
2625 if (unlikely(size
> PAGE_SIZE
))
2626 return kmalloc_large_node(size
, flags
, node
);
2628 s
= get_slab(size
, flags
);
2630 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2633 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2635 EXPORT_SYMBOL(__kmalloc_node
);
2638 size_t ksize(const void *object
)
2641 struct kmem_cache
*s
;
2643 if (unlikely(object
== ZERO_SIZE_PTR
))
2646 page
= virt_to_head_page(object
);
2648 if (unlikely(!PageSlab(page
)))
2649 return PAGE_SIZE
<< compound_order(page
);
2653 #ifdef CONFIG_SLUB_DEBUG
2655 * Debugging requires use of the padding between object
2656 * and whatever may come after it.
2658 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2663 * If we have the need to store the freelist pointer
2664 * back there or track user information then we can
2665 * only use the space before that information.
2667 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2670 * Else we can use all the padding etc for the allocation
2674 EXPORT_SYMBOL(ksize
);
2676 void kfree(const void *x
)
2679 void *object
= (void *)x
;
2681 if (unlikely(ZERO_OR_NULL_PTR(x
)))
2684 page
= virt_to_head_page(x
);
2685 if (unlikely(!PageSlab(page
))) {
2689 slab_free(page
->slab
, page
, object
, __builtin_return_address(0));
2691 EXPORT_SYMBOL(kfree
);
2693 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SLABINFO)
2694 static unsigned long count_partial(struct kmem_cache_node
*n
)
2696 unsigned long flags
;
2697 unsigned long x
= 0;
2700 spin_lock_irqsave(&n
->list_lock
, flags
);
2701 list_for_each_entry(page
, &n
->partial
, lru
)
2703 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2709 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2710 * the remaining slabs by the number of items in use. The slabs with the
2711 * most items in use come first. New allocations will then fill those up
2712 * and thus they can be removed from the partial lists.
2714 * The slabs with the least items are placed last. This results in them
2715 * being allocated from last increasing the chance that the last objects
2716 * are freed in them.
2718 int kmem_cache_shrink(struct kmem_cache
*s
)
2722 struct kmem_cache_node
*n
;
2725 struct list_head
*slabs_by_inuse
=
2726 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2727 unsigned long flags
;
2729 if (!slabs_by_inuse
)
2733 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2734 n
= get_node(s
, node
);
2739 for (i
= 0; i
< s
->objects
; i
++)
2740 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2742 spin_lock_irqsave(&n
->list_lock
, flags
);
2745 * Build lists indexed by the items in use in each slab.
2747 * Note that concurrent frees may occur while we hold the
2748 * list_lock. page->inuse here is the upper limit.
2750 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2751 if (!page
->inuse
&& slab_trylock(page
)) {
2753 * Must hold slab lock here because slab_free
2754 * may have freed the last object and be
2755 * waiting to release the slab.
2757 list_del(&page
->lru
);
2760 discard_slab(s
, page
);
2762 list_move(&page
->lru
,
2763 slabs_by_inuse
+ page
->inuse
);
2768 * Rebuild the partial list with the slabs filled up most
2769 * first and the least used slabs at the end.
2771 for (i
= s
->objects
- 1; i
>= 0; i
--)
2772 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2774 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2777 kfree(slabs_by_inuse
);
2780 EXPORT_SYMBOL(kmem_cache_shrink
);
2782 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2783 static int slab_mem_going_offline_callback(void *arg
)
2785 struct kmem_cache
*s
;
2787 down_read(&slub_lock
);
2788 list_for_each_entry(s
, &slab_caches
, list
)
2789 kmem_cache_shrink(s
);
2790 up_read(&slub_lock
);
2795 static void slab_mem_offline_callback(void *arg
)
2797 struct kmem_cache_node
*n
;
2798 struct kmem_cache
*s
;
2799 struct memory_notify
*marg
= arg
;
2802 offline_node
= marg
->status_change_nid
;
2805 * If the node still has available memory. we need kmem_cache_node
2808 if (offline_node
< 0)
2811 down_read(&slub_lock
);
2812 list_for_each_entry(s
, &slab_caches
, list
) {
2813 n
= get_node(s
, offline_node
);
2816 * if n->nr_slabs > 0, slabs still exist on the node
2817 * that is going down. We were unable to free them,
2818 * and offline_pages() function shoudn't call this
2819 * callback. So, we must fail.
2821 BUG_ON(atomic_long_read(&n
->nr_slabs
));
2823 s
->node
[offline_node
] = NULL
;
2824 kmem_cache_free(kmalloc_caches
, n
);
2827 up_read(&slub_lock
);
2830 static int slab_mem_going_online_callback(void *arg
)
2832 struct kmem_cache_node
*n
;
2833 struct kmem_cache
*s
;
2834 struct memory_notify
*marg
= arg
;
2835 int nid
= marg
->status_change_nid
;
2839 * If the node's memory is already available, then kmem_cache_node is
2840 * already created. Nothing to do.
2846 * We are bringing a node online. No memory is availabe yet. We must
2847 * allocate a kmem_cache_node structure in order to bring the node
2850 down_read(&slub_lock
);
2851 list_for_each_entry(s
, &slab_caches
, list
) {
2853 * XXX: kmem_cache_alloc_node will fallback to other nodes
2854 * since memory is not yet available from the node that
2857 n
= kmem_cache_alloc(kmalloc_caches
, GFP_KERNEL
);
2862 init_kmem_cache_node(n
);
2866 up_read(&slub_lock
);
2870 static int slab_memory_callback(struct notifier_block
*self
,
2871 unsigned long action
, void *arg
)
2876 case MEM_GOING_ONLINE
:
2877 ret
= slab_mem_going_online_callback(arg
);
2879 case MEM_GOING_OFFLINE
:
2880 ret
= slab_mem_going_offline_callback(arg
);
2883 case MEM_CANCEL_ONLINE
:
2884 slab_mem_offline_callback(arg
);
2887 case MEM_CANCEL_OFFLINE
:
2891 ret
= notifier_from_errno(ret
);
2895 #endif /* CONFIG_MEMORY_HOTPLUG */
2897 /********************************************************************
2898 * Basic setup of slabs
2899 *******************************************************************/
2901 void __init
kmem_cache_init(void)
2910 * Must first have the slab cache available for the allocations of the
2911 * struct kmem_cache_node's. There is special bootstrap code in
2912 * kmem_cache_open for slab_state == DOWN.
2914 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2915 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2916 kmalloc_caches
[0].refcount
= -1;
2919 hotplug_memory_notifier(slab_memory_callback
, 1);
2922 /* Able to allocate the per node structures */
2923 slab_state
= PARTIAL
;
2925 /* Caches that are not of the two-to-the-power-of size */
2926 if (KMALLOC_MIN_SIZE
<= 64) {
2927 create_kmalloc_cache(&kmalloc_caches
[1],
2928 "kmalloc-96", 96, GFP_KERNEL
);
2931 if (KMALLOC_MIN_SIZE
<= 128) {
2932 create_kmalloc_cache(&kmalloc_caches
[2],
2933 "kmalloc-192", 192, GFP_KERNEL
);
2937 for (i
= KMALLOC_SHIFT_LOW
; i
<= PAGE_SHIFT
; i
++) {
2938 create_kmalloc_cache(&kmalloc_caches
[i
],
2939 "kmalloc", 1 << i
, GFP_KERNEL
);
2945 * Patch up the size_index table if we have strange large alignment
2946 * requirements for the kmalloc array. This is only the case for
2947 * MIPS it seems. The standard arches will not generate any code here.
2949 * Largest permitted alignment is 256 bytes due to the way we
2950 * handle the index determination for the smaller caches.
2952 * Make sure that nothing crazy happens if someone starts tinkering
2953 * around with ARCH_KMALLOC_MINALIGN
2955 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 256 ||
2956 (KMALLOC_MIN_SIZE
& (KMALLOC_MIN_SIZE
- 1)));
2958 for (i
= 8; i
< KMALLOC_MIN_SIZE
; i
+= 8)
2959 size_index
[(i
- 1) / 8] = KMALLOC_SHIFT_LOW
;
2963 /* Provide the correct kmalloc names now that the caches are up */
2964 for (i
= KMALLOC_SHIFT_LOW
; i
<= PAGE_SHIFT
; i
++)
2965 kmalloc_caches
[i
]. name
=
2966 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
2969 register_cpu_notifier(&slab_notifier
);
2970 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
) +
2971 nr_cpu_ids
* sizeof(struct kmem_cache_cpu
*);
2973 kmem_size
= sizeof(struct kmem_cache
);
2977 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2978 " CPUs=%d, Nodes=%d\n",
2979 caches
, cache_line_size(),
2980 slub_min_order
, slub_max_order
, slub_min_objects
,
2981 nr_cpu_ids
, nr_node_ids
);
2985 * Find a mergeable slab cache
2987 static int slab_unmergeable(struct kmem_cache
*s
)
2989 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
2992 if ((s
->flags
& __PAGE_ALLOC_FALLBACK
))
2999 * We may have set a slab to be unmergeable during bootstrap.
3001 if (s
->refcount
< 0)
3007 static struct kmem_cache
*find_mergeable(size_t size
,
3008 size_t align
, unsigned long flags
, const char *name
,
3009 void (*ctor
)(struct kmem_cache
*, void *))
3011 struct kmem_cache
*s
;
3013 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
3019 size
= ALIGN(size
, sizeof(void *));
3020 align
= calculate_alignment(flags
, align
, size
);
3021 size
= ALIGN(size
, align
);
3022 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
3024 list_for_each_entry(s
, &slab_caches
, list
) {
3025 if (slab_unmergeable(s
))
3031 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
3034 * Check if alignment is compatible.
3035 * Courtesy of Adrian Drzewiecki
3037 if ((s
->size
& ~(align
- 1)) != s
->size
)
3040 if (s
->size
- size
>= sizeof(void *))
3048 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
3049 size_t align
, unsigned long flags
,
3050 void (*ctor
)(struct kmem_cache
*, void *))
3052 struct kmem_cache
*s
;
3054 down_write(&slub_lock
);
3055 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
3061 * Adjust the object sizes so that we clear
3062 * the complete object on kzalloc.
3064 s
->objsize
= max(s
->objsize
, (int)size
);
3067 * And then we need to update the object size in the
3068 * per cpu structures
3070 for_each_online_cpu(cpu
)
3071 get_cpu_slab(s
, cpu
)->objsize
= s
->objsize
;
3073 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3074 up_write(&slub_lock
);
3076 if (sysfs_slab_alias(s
, name
))
3081 s
= kmalloc(kmem_size
, GFP_KERNEL
);
3083 if (kmem_cache_open(s
, GFP_KERNEL
, name
,
3084 size
, align
, flags
, ctor
)) {
3085 list_add(&s
->list
, &slab_caches
);
3086 up_write(&slub_lock
);
3087 if (sysfs_slab_add(s
))
3093 up_write(&slub_lock
);
3096 if (flags
& SLAB_PANIC
)
3097 panic("Cannot create slabcache %s\n", name
);
3102 EXPORT_SYMBOL(kmem_cache_create
);
3106 * Use the cpu notifier to insure that the cpu slabs are flushed when
3109 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
3110 unsigned long action
, void *hcpu
)
3112 long cpu
= (long)hcpu
;
3113 struct kmem_cache
*s
;
3114 unsigned long flags
;
3117 case CPU_UP_PREPARE
:
3118 case CPU_UP_PREPARE_FROZEN
:
3119 init_alloc_cpu_cpu(cpu
);
3120 down_read(&slub_lock
);
3121 list_for_each_entry(s
, &slab_caches
, list
)
3122 s
->cpu_slab
[cpu
] = alloc_kmem_cache_cpu(s
, cpu
,
3124 up_read(&slub_lock
);
3127 case CPU_UP_CANCELED
:
3128 case CPU_UP_CANCELED_FROZEN
:
3130 case CPU_DEAD_FROZEN
:
3131 down_read(&slub_lock
);
3132 list_for_each_entry(s
, &slab_caches
, list
) {
3133 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3135 local_irq_save(flags
);
3136 __flush_cpu_slab(s
, cpu
);
3137 local_irq_restore(flags
);
3138 free_kmem_cache_cpu(c
, cpu
);
3139 s
->cpu_slab
[cpu
] = NULL
;
3141 up_read(&slub_lock
);
3149 static struct notifier_block __cpuinitdata slab_notifier
= {
3150 .notifier_call
= slab_cpuup_callback
3155 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
3157 struct kmem_cache
*s
;
3159 if (unlikely(size
> PAGE_SIZE
))
3160 return kmalloc_large(size
, gfpflags
);
3162 s
= get_slab(size
, gfpflags
);
3164 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3167 return slab_alloc(s
, gfpflags
, -1, caller
);
3170 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3171 int node
, void *caller
)
3173 struct kmem_cache
*s
;
3175 if (unlikely(size
> PAGE_SIZE
))
3176 return kmalloc_large_node(size
, gfpflags
, node
);
3178 s
= get_slab(size
, gfpflags
);
3180 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3183 return slab_alloc(s
, gfpflags
, node
, caller
);
3186 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3187 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3191 void *addr
= page_address(page
);
3193 if (!check_slab(s
, page
) ||
3194 !on_freelist(s
, page
, NULL
))
3197 /* Now we know that a valid freelist exists */
3198 bitmap_zero(map
, s
->objects
);
3200 for_each_free_object(p
, s
, page
->freelist
) {
3201 set_bit(slab_index(p
, s
, addr
), map
);
3202 if (!check_object(s
, page
, p
, 0))
3206 for_each_object(p
, s
, addr
)
3207 if (!test_bit(slab_index(p
, s
, addr
), map
))
3208 if (!check_object(s
, page
, p
, 1))
3213 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
3216 if (slab_trylock(page
)) {
3217 validate_slab(s
, page
, map
);
3220 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
3223 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
3224 if (!SlabDebug(page
))
3225 printk(KERN_ERR
"SLUB %s: SlabDebug not set "
3226 "on slab 0x%p\n", s
->name
, page
);
3228 if (SlabDebug(page
))
3229 printk(KERN_ERR
"SLUB %s: SlabDebug set on "
3230 "slab 0x%p\n", s
->name
, page
);
3234 static int validate_slab_node(struct kmem_cache
*s
,
3235 struct kmem_cache_node
*n
, unsigned long *map
)
3237 unsigned long count
= 0;
3239 unsigned long flags
;
3241 spin_lock_irqsave(&n
->list_lock
, flags
);
3243 list_for_each_entry(page
, &n
->partial
, lru
) {
3244 validate_slab_slab(s
, page
, map
);
3247 if (count
!= n
->nr_partial
)
3248 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
3249 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
3251 if (!(s
->flags
& SLAB_STORE_USER
))
3254 list_for_each_entry(page
, &n
->full
, lru
) {
3255 validate_slab_slab(s
, page
, map
);
3258 if (count
!= atomic_long_read(&n
->nr_slabs
))
3259 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
3260 "counter=%ld\n", s
->name
, count
,
3261 atomic_long_read(&n
->nr_slabs
));
3264 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3268 static long validate_slab_cache(struct kmem_cache
*s
)
3271 unsigned long count
= 0;
3272 unsigned long *map
= kmalloc(BITS_TO_LONGS(s
->objects
) *
3273 sizeof(unsigned long), GFP_KERNEL
);
3279 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3280 struct kmem_cache_node
*n
= get_node(s
, node
);
3282 count
+= validate_slab_node(s
, n
, map
);
3288 #ifdef SLUB_RESILIENCY_TEST
3289 static void resiliency_test(void)
3293 printk(KERN_ERR
"SLUB resiliency testing\n");
3294 printk(KERN_ERR
"-----------------------\n");
3295 printk(KERN_ERR
"A. Corruption after allocation\n");
3297 p
= kzalloc(16, GFP_KERNEL
);
3299 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
3300 " 0x12->0x%p\n\n", p
+ 16);
3302 validate_slab_cache(kmalloc_caches
+ 4);
3304 /* Hmmm... The next two are dangerous */
3305 p
= kzalloc(32, GFP_KERNEL
);
3306 p
[32 + sizeof(void *)] = 0x34;
3307 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
3308 " 0x34 -> -0x%p\n", p
);
3310 "If allocated object is overwritten then not detectable\n\n");
3312 validate_slab_cache(kmalloc_caches
+ 5);
3313 p
= kzalloc(64, GFP_KERNEL
);
3314 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
3316 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3319 "If allocated object is overwritten then not detectable\n\n");
3320 validate_slab_cache(kmalloc_caches
+ 6);
3322 printk(KERN_ERR
"\nB. Corruption after free\n");
3323 p
= kzalloc(128, GFP_KERNEL
);
3326 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
3327 validate_slab_cache(kmalloc_caches
+ 7);
3329 p
= kzalloc(256, GFP_KERNEL
);
3332 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3334 validate_slab_cache(kmalloc_caches
+ 8);
3336 p
= kzalloc(512, GFP_KERNEL
);
3339 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
3340 validate_slab_cache(kmalloc_caches
+ 9);
3343 static void resiliency_test(void) {};
3347 * Generate lists of code addresses where slabcache objects are allocated
3352 unsigned long count
;
3365 unsigned long count
;
3366 struct location
*loc
;
3369 static void free_loc_track(struct loc_track
*t
)
3372 free_pages((unsigned long)t
->loc
,
3373 get_order(sizeof(struct location
) * t
->max
));
3376 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
3381 order
= get_order(sizeof(struct location
) * max
);
3383 l
= (void *)__get_free_pages(flags
, order
);
3388 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
3396 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
3397 const struct track
*track
)
3399 long start
, end
, pos
;
3402 unsigned long age
= jiffies
- track
->when
;
3408 pos
= start
+ (end
- start
+ 1) / 2;
3411 * There is nothing at "end". If we end up there
3412 * we need to add something to before end.
3417 caddr
= t
->loc
[pos
].addr
;
3418 if (track
->addr
== caddr
) {
3424 if (age
< l
->min_time
)
3426 if (age
> l
->max_time
)
3429 if (track
->pid
< l
->min_pid
)
3430 l
->min_pid
= track
->pid
;
3431 if (track
->pid
> l
->max_pid
)
3432 l
->max_pid
= track
->pid
;
3434 cpu_set(track
->cpu
, l
->cpus
);
3436 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3440 if (track
->addr
< caddr
)
3447 * Not found. Insert new tracking element.
3449 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
3455 (t
->count
- pos
) * sizeof(struct location
));
3458 l
->addr
= track
->addr
;
3462 l
->min_pid
= track
->pid
;
3463 l
->max_pid
= track
->pid
;
3464 cpus_clear(l
->cpus
);
3465 cpu_set(track
->cpu
, l
->cpus
);
3466 nodes_clear(l
->nodes
);
3467 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3471 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
3472 struct page
*page
, enum track_item alloc
)
3474 void *addr
= page_address(page
);
3475 DECLARE_BITMAP(map
, s
->objects
);
3478 bitmap_zero(map
, s
->objects
);
3479 for_each_free_object(p
, s
, page
->freelist
)
3480 set_bit(slab_index(p
, s
, addr
), map
);
3482 for_each_object(p
, s
, addr
)
3483 if (!test_bit(slab_index(p
, s
, addr
), map
))
3484 add_location(t
, s
, get_track(s
, p
, alloc
));
3487 static int list_locations(struct kmem_cache
*s
, char *buf
,
3488 enum track_item alloc
)
3492 struct loc_track t
= { 0, 0, NULL
};
3495 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
3497 return sprintf(buf
, "Out of memory\n");
3499 /* Push back cpu slabs */
3502 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3503 struct kmem_cache_node
*n
= get_node(s
, node
);
3504 unsigned long flags
;
3507 if (!atomic_long_read(&n
->nr_slabs
))
3510 spin_lock_irqsave(&n
->list_lock
, flags
);
3511 list_for_each_entry(page
, &n
->partial
, lru
)
3512 process_slab(&t
, s
, page
, alloc
);
3513 list_for_each_entry(page
, &n
->full
, lru
)
3514 process_slab(&t
, s
, page
, alloc
);
3515 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3518 for (i
= 0; i
< t
.count
; i
++) {
3519 struct location
*l
= &t
.loc
[i
];
3521 if (len
> PAGE_SIZE
- 100)
3523 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
3526 len
+= sprint_symbol(buf
+ len
, (unsigned long)l
->addr
);
3528 len
+= sprintf(buf
+ len
, "<not-available>");
3530 if (l
->sum_time
!= l
->min_time
) {
3531 unsigned long remainder
;
3533 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
3535 div_long_long_rem(l
->sum_time
, l
->count
, &remainder
),
3538 len
+= sprintf(buf
+ len
, " age=%ld",
3541 if (l
->min_pid
!= l
->max_pid
)
3542 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
3543 l
->min_pid
, l
->max_pid
);
3545 len
+= sprintf(buf
+ len
, " pid=%ld",
3548 if (num_online_cpus() > 1 && !cpus_empty(l
->cpus
) &&
3549 len
< PAGE_SIZE
- 60) {
3550 len
+= sprintf(buf
+ len
, " cpus=");
3551 len
+= cpulist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3555 if (num_online_nodes() > 1 && !nodes_empty(l
->nodes
) &&
3556 len
< PAGE_SIZE
- 60) {
3557 len
+= sprintf(buf
+ len
, " nodes=");
3558 len
+= nodelist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3562 len
+= sprintf(buf
+ len
, "\n");
3567 len
+= sprintf(buf
, "No data\n");
3571 enum slab_stat_type
{
3578 #define SO_FULL (1 << SL_FULL)
3579 #define SO_PARTIAL (1 << SL_PARTIAL)
3580 #define SO_CPU (1 << SL_CPU)
3581 #define SO_OBJECTS (1 << SL_OBJECTS)
3583 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
3584 char *buf
, unsigned long flags
)
3586 unsigned long total
= 0;
3590 unsigned long *nodes
;
3591 unsigned long *per_cpu
;
3593 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
3596 per_cpu
= nodes
+ nr_node_ids
;
3598 for_each_possible_cpu(cpu
) {
3600 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3610 if (flags
& SO_CPU
) {
3611 if (flags
& SO_OBJECTS
)
3622 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3623 struct kmem_cache_node
*n
= get_node(s
, node
);
3625 if (flags
& SO_PARTIAL
) {
3626 if (flags
& SO_OBJECTS
)
3627 x
= count_partial(n
);
3634 if (flags
& SO_FULL
) {
3635 int full_slabs
= atomic_long_read(&n
->nr_slabs
)
3639 if (flags
& SO_OBJECTS
)
3640 x
= full_slabs
* s
->objects
;
3648 x
= sprintf(buf
, "%lu", total
);
3650 for_each_node_state(node
, N_NORMAL_MEMORY
)
3652 x
+= sprintf(buf
+ x
, " N%d=%lu",
3656 return x
+ sprintf(buf
+ x
, "\n");
3659 static int any_slab_objects(struct kmem_cache
*s
)
3664 for_each_possible_cpu(cpu
) {
3665 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3671 for_each_online_node(node
) {
3672 struct kmem_cache_node
*n
= get_node(s
, node
);
3677 if (n
->nr_partial
|| atomic_long_read(&n
->nr_slabs
))
3683 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3684 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3686 struct slab_attribute
{
3687 struct attribute attr
;
3688 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
3689 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
3692 #define SLAB_ATTR_RO(_name) \
3693 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3695 #define SLAB_ATTR(_name) \
3696 static struct slab_attribute _name##_attr = \
3697 __ATTR(_name, 0644, _name##_show, _name##_store)
3699 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
3701 return sprintf(buf
, "%d\n", s
->size
);
3703 SLAB_ATTR_RO(slab_size
);
3705 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3707 return sprintf(buf
, "%d\n", s
->align
);
3709 SLAB_ATTR_RO(align
);
3711 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3713 return sprintf(buf
, "%d\n", s
->objsize
);
3715 SLAB_ATTR_RO(object_size
);
3717 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3719 return sprintf(buf
, "%d\n", s
->objects
);
3721 SLAB_ATTR_RO(objs_per_slab
);
3723 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3725 return sprintf(buf
, "%d\n", s
->order
);
3727 SLAB_ATTR_RO(order
);
3729 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3732 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3734 return n
+ sprintf(buf
+ n
, "\n");
3740 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3742 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3744 SLAB_ATTR_RO(aliases
);
3746 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3748 return show_slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3750 SLAB_ATTR_RO(slabs
);
3752 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3754 return show_slab_objects(s
, buf
, SO_PARTIAL
);
3756 SLAB_ATTR_RO(partial
);
3758 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3760 return show_slab_objects(s
, buf
, SO_CPU
);
3762 SLAB_ATTR_RO(cpu_slabs
);
3764 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3766 return show_slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3768 SLAB_ATTR_RO(objects
);
3770 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3772 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3775 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3776 const char *buf
, size_t length
)
3778 s
->flags
&= ~SLAB_DEBUG_FREE
;
3780 s
->flags
|= SLAB_DEBUG_FREE
;
3783 SLAB_ATTR(sanity_checks
);
3785 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3787 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3790 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3793 s
->flags
&= ~SLAB_TRACE
;
3795 s
->flags
|= SLAB_TRACE
;
3800 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3802 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3805 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3806 const char *buf
, size_t length
)
3808 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3810 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3813 SLAB_ATTR(reclaim_account
);
3815 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3817 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3819 SLAB_ATTR_RO(hwcache_align
);
3821 #ifdef CONFIG_ZONE_DMA
3822 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3824 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3826 SLAB_ATTR_RO(cache_dma
);
3829 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3831 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3833 SLAB_ATTR_RO(destroy_by_rcu
);
3835 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3837 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3840 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3841 const char *buf
, size_t length
)
3843 if (any_slab_objects(s
))
3846 s
->flags
&= ~SLAB_RED_ZONE
;
3848 s
->flags
|= SLAB_RED_ZONE
;
3852 SLAB_ATTR(red_zone
);
3854 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3856 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3859 static ssize_t
poison_store(struct kmem_cache
*s
,
3860 const char *buf
, size_t length
)
3862 if (any_slab_objects(s
))
3865 s
->flags
&= ~SLAB_POISON
;
3867 s
->flags
|= SLAB_POISON
;
3873 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3875 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3878 static ssize_t
store_user_store(struct kmem_cache
*s
,
3879 const char *buf
, size_t length
)
3881 if (any_slab_objects(s
))
3884 s
->flags
&= ~SLAB_STORE_USER
;
3886 s
->flags
|= SLAB_STORE_USER
;
3890 SLAB_ATTR(store_user
);
3892 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3897 static ssize_t
validate_store(struct kmem_cache
*s
,
3898 const char *buf
, size_t length
)
3902 if (buf
[0] == '1') {
3903 ret
= validate_slab_cache(s
);
3909 SLAB_ATTR(validate
);
3911 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3916 static ssize_t
shrink_store(struct kmem_cache
*s
,
3917 const char *buf
, size_t length
)
3919 if (buf
[0] == '1') {
3920 int rc
= kmem_cache_shrink(s
);
3930 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
3932 if (!(s
->flags
& SLAB_STORE_USER
))
3934 return list_locations(s
, buf
, TRACK_ALLOC
);
3936 SLAB_ATTR_RO(alloc_calls
);
3938 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
3940 if (!(s
->flags
& SLAB_STORE_USER
))
3942 return list_locations(s
, buf
, TRACK_FREE
);
3944 SLAB_ATTR_RO(free_calls
);
3947 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
3949 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
3952 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
3953 const char *buf
, size_t length
)
3955 int n
= simple_strtoul(buf
, NULL
, 10);
3958 s
->remote_node_defrag_ratio
= n
* 10;
3961 SLAB_ATTR(remote_node_defrag_ratio
);
3964 #ifdef CONFIG_SLUB_STATS
3965 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
3967 unsigned long sum
= 0;
3970 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
3975 for_each_online_cpu(cpu
) {
3976 unsigned x
= get_cpu_slab(s
, cpu
)->stat
[si
];
3982 len
= sprintf(buf
, "%lu", sum
);
3984 for_each_online_cpu(cpu
) {
3985 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
3986 len
+= sprintf(buf
+ len
, " c%d=%u", cpu
, data
[cpu
]);
3989 return len
+ sprintf(buf
+ len
, "\n");
3992 #define STAT_ATTR(si, text) \
3993 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3995 return show_stat(s, buf, si); \
3997 SLAB_ATTR_RO(text); \
3999 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4000 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4001 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4002 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4003 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4004 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4005 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4006 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4007 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4008 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4009 STAT_ATTR(FREE_SLAB
, free_slab
);
4010 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4011 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4012 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4013 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4014 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4015 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4019 static struct attribute
*slab_attrs
[] = {
4020 &slab_size_attr
.attr
,
4021 &object_size_attr
.attr
,
4022 &objs_per_slab_attr
.attr
,
4027 &cpu_slabs_attr
.attr
,
4031 &sanity_checks_attr
.attr
,
4033 &hwcache_align_attr
.attr
,
4034 &reclaim_account_attr
.attr
,
4035 &destroy_by_rcu_attr
.attr
,
4036 &red_zone_attr
.attr
,
4038 &store_user_attr
.attr
,
4039 &validate_attr
.attr
,
4041 &alloc_calls_attr
.attr
,
4042 &free_calls_attr
.attr
,
4043 #ifdef CONFIG_ZONE_DMA
4044 &cache_dma_attr
.attr
,
4047 &remote_node_defrag_ratio_attr
.attr
,
4049 #ifdef CONFIG_SLUB_STATS
4050 &alloc_fastpath_attr
.attr
,
4051 &alloc_slowpath_attr
.attr
,
4052 &free_fastpath_attr
.attr
,
4053 &free_slowpath_attr
.attr
,
4054 &free_frozen_attr
.attr
,
4055 &free_add_partial_attr
.attr
,
4056 &free_remove_partial_attr
.attr
,
4057 &alloc_from_partial_attr
.attr
,
4058 &alloc_slab_attr
.attr
,
4059 &alloc_refill_attr
.attr
,
4060 &free_slab_attr
.attr
,
4061 &cpuslab_flush_attr
.attr
,
4062 &deactivate_full_attr
.attr
,
4063 &deactivate_empty_attr
.attr
,
4064 &deactivate_to_head_attr
.attr
,
4065 &deactivate_to_tail_attr
.attr
,
4066 &deactivate_remote_frees_attr
.attr
,
4071 static struct attribute_group slab_attr_group
= {
4072 .attrs
= slab_attrs
,
4075 static ssize_t
slab_attr_show(struct kobject
*kobj
,
4076 struct attribute
*attr
,
4079 struct slab_attribute
*attribute
;
4080 struct kmem_cache
*s
;
4083 attribute
= to_slab_attr(attr
);
4086 if (!attribute
->show
)
4089 err
= attribute
->show(s
, buf
);
4094 static ssize_t
slab_attr_store(struct kobject
*kobj
,
4095 struct attribute
*attr
,
4096 const char *buf
, size_t len
)
4098 struct slab_attribute
*attribute
;
4099 struct kmem_cache
*s
;
4102 attribute
= to_slab_attr(attr
);
4105 if (!attribute
->store
)
4108 err
= attribute
->store(s
, buf
, len
);
4113 static void kmem_cache_release(struct kobject
*kobj
)
4115 struct kmem_cache
*s
= to_slab(kobj
);
4120 static struct sysfs_ops slab_sysfs_ops
= {
4121 .show
= slab_attr_show
,
4122 .store
= slab_attr_store
,
4125 static struct kobj_type slab_ktype
= {
4126 .sysfs_ops
= &slab_sysfs_ops
,
4127 .release
= kmem_cache_release
4130 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
4132 struct kobj_type
*ktype
= get_ktype(kobj
);
4134 if (ktype
== &slab_ktype
)
4139 static struct kset_uevent_ops slab_uevent_ops
= {
4140 .filter
= uevent_filter
,
4143 static struct kset
*slab_kset
;
4145 #define ID_STR_LENGTH 64
4147 /* Create a unique string id for a slab cache:
4149 * Format :[flags-]size
4151 static char *create_unique_id(struct kmem_cache
*s
)
4153 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
4160 * First flags affecting slabcache operations. We will only
4161 * get here for aliasable slabs so we do not need to support
4162 * too many flags. The flags here must cover all flags that
4163 * are matched during merging to guarantee that the id is
4166 if (s
->flags
& SLAB_CACHE_DMA
)
4168 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
4170 if (s
->flags
& SLAB_DEBUG_FREE
)
4174 p
+= sprintf(p
, "%07d", s
->size
);
4175 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
4179 static int sysfs_slab_add(struct kmem_cache
*s
)
4185 if (slab_state
< SYSFS
)
4186 /* Defer until later */
4189 unmergeable
= slab_unmergeable(s
);
4192 * Slabcache can never be merged so we can use the name proper.
4193 * This is typically the case for debug situations. In that
4194 * case we can catch duplicate names easily.
4196 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
4200 * Create a unique name for the slab as a target
4203 name
= create_unique_id(s
);
4206 s
->kobj
.kset
= slab_kset
;
4207 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, name
);
4209 kobject_put(&s
->kobj
);
4213 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
4216 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
4218 /* Setup first alias */
4219 sysfs_slab_alias(s
, s
->name
);
4225 static void sysfs_slab_remove(struct kmem_cache
*s
)
4227 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
4228 kobject_del(&s
->kobj
);
4229 kobject_put(&s
->kobj
);
4233 * Need to buffer aliases during bootup until sysfs becomes
4234 * available lest we loose that information.
4236 struct saved_alias
{
4237 struct kmem_cache
*s
;
4239 struct saved_alias
*next
;
4242 static struct saved_alias
*alias_list
;
4244 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
4246 struct saved_alias
*al
;
4248 if (slab_state
== SYSFS
) {
4250 * If we have a leftover link then remove it.
4252 sysfs_remove_link(&slab_kset
->kobj
, name
);
4253 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
4256 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
4262 al
->next
= alias_list
;
4267 static int __init
slab_sysfs_init(void)
4269 struct kmem_cache
*s
;
4272 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
4274 printk(KERN_ERR
"Cannot register slab subsystem.\n");
4280 list_for_each_entry(s
, &slab_caches
, list
) {
4281 err
= sysfs_slab_add(s
);
4283 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
4284 " to sysfs\n", s
->name
);
4287 while (alias_list
) {
4288 struct saved_alias
*al
= alias_list
;
4290 alias_list
= alias_list
->next
;
4291 err
= sysfs_slab_alias(al
->s
, al
->name
);
4293 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
4294 " %s to sysfs\n", s
->name
);
4302 __initcall(slab_sysfs_init
);
4306 * The /proc/slabinfo ABI
4308 #ifdef CONFIG_SLABINFO
4310 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4311 size_t count
, loff_t
*ppos
)
4317 static void print_slabinfo_header(struct seq_file
*m
)
4319 seq_puts(m
, "slabinfo - version: 2.1\n");
4320 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4321 "<objperslab> <pagesperslab>");
4322 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4323 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4327 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4331 down_read(&slub_lock
);
4333 print_slabinfo_header(m
);
4335 return seq_list_start(&slab_caches
, *pos
);
4338 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4340 return seq_list_next(p
, &slab_caches
, pos
);
4343 static void s_stop(struct seq_file
*m
, void *p
)
4345 up_read(&slub_lock
);
4348 static int s_show(struct seq_file
*m
, void *p
)
4350 unsigned long nr_partials
= 0;
4351 unsigned long nr_slabs
= 0;
4352 unsigned long nr_inuse
= 0;
4353 unsigned long nr_objs
;
4354 struct kmem_cache
*s
;
4357 s
= list_entry(p
, struct kmem_cache
, list
);
4359 for_each_online_node(node
) {
4360 struct kmem_cache_node
*n
= get_node(s
, node
);
4365 nr_partials
+= n
->nr_partial
;
4366 nr_slabs
+= atomic_long_read(&n
->nr_slabs
);
4367 nr_inuse
+= count_partial(n
);
4370 nr_objs
= nr_slabs
* s
->objects
;
4371 nr_inuse
+= (nr_slabs
- nr_partials
) * s
->objects
;
4373 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d", s
->name
, nr_inuse
,
4374 nr_objs
, s
->size
, s
->objects
, (1 << s
->order
));
4375 seq_printf(m
, " : tunables %4u %4u %4u", 0, 0, 0);
4376 seq_printf(m
, " : slabdata %6lu %6lu %6lu", nr_slabs
, nr_slabs
,
4382 const struct seq_operations slabinfo_op
= {
4389 #endif /* CONFIG_SLABINFO */