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
+ page
->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, __objects) \
331 for (__p = (__addr); __p < (__addr) + (__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 objects=%u used=%u fp=0x%p flags=0x%04lx\n",
455 page
, page
->objects
, 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
)
524 slab_bug(s
, "%s", 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
);
536 slab_bug(s
, "%s", buf
);
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 /* Check the pad bytes at the end of a slab page */
656 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
664 if (!(s
->flags
& SLAB_POISON
))
667 start
= page_address(page
);
668 length
= (PAGE_SIZE
<< s
->order
);
669 end
= start
+ length
;
670 remainder
= length
% s
->size
;
674 fault
= check_bytes(end
- remainder
, POISON_INUSE
, remainder
);
677 while (end
> fault
&& end
[-1] == POISON_INUSE
)
680 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
681 print_section("Padding", end
- remainder
, remainder
);
683 restore_bytes(s
, "slab padding", POISON_INUSE
, start
, end
);
687 static int check_object(struct kmem_cache
*s
, struct page
*page
,
688 void *object
, int active
)
691 u8
*endobject
= object
+ s
->objsize
;
693 if (s
->flags
& SLAB_RED_ZONE
) {
695 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
697 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
698 endobject
, red
, s
->inuse
- s
->objsize
))
701 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
) {
702 check_bytes_and_report(s
, page
, p
, "Alignment padding",
703 endobject
, POISON_INUSE
, s
->inuse
- s
->objsize
);
707 if (s
->flags
& SLAB_POISON
) {
708 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
709 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
710 POISON_FREE
, s
->objsize
- 1) ||
711 !check_bytes_and_report(s
, page
, p
, "Poison",
712 p
+ s
->objsize
- 1, POISON_END
, 1)))
715 * check_pad_bytes cleans up on its own.
717 check_pad_bytes(s
, page
, p
);
720 if (!s
->offset
&& active
)
722 * Object and freepointer overlap. Cannot check
723 * freepointer while object is allocated.
727 /* Check free pointer validity */
728 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
729 object_err(s
, page
, p
, "Freepointer corrupt");
731 * No choice but to zap it and thus loose the remainder
732 * of the free objects in this slab. May cause
733 * another error because the object count is now wrong.
735 set_freepointer(s
, p
, NULL
);
741 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
745 VM_BUG_ON(!irqs_disabled());
747 if (!PageSlab(page
)) {
748 slab_err(s
, page
, "Not a valid slab page");
752 maxobj
= (PAGE_SIZE
<< compound_order(page
)) / s
->size
;
753 if (page
->objects
> maxobj
) {
754 slab_err(s
, page
, "objects %u > max %u",
755 s
->name
, page
->objects
, maxobj
);
758 if (page
->inuse
> page
->objects
) {
759 slab_err(s
, page
, "inuse %u > max %u",
760 s
->name
, page
->inuse
, page
->objects
);
763 /* Slab_pad_check fixes things up after itself */
764 slab_pad_check(s
, page
);
769 * Determine if a certain object on a page is on the freelist. Must hold the
770 * slab lock to guarantee that the chains are in a consistent state.
772 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
775 void *fp
= page
->freelist
;
777 unsigned long max_objects
;
779 while (fp
&& nr
<= page
->objects
) {
782 if (!check_valid_pointer(s
, page
, fp
)) {
784 object_err(s
, page
, object
,
785 "Freechain corrupt");
786 set_freepointer(s
, object
, NULL
);
789 slab_err(s
, page
, "Freepointer corrupt");
790 page
->freelist
= NULL
;
791 page
->inuse
= page
->objects
;
792 slab_fix(s
, "Freelist cleared");
798 fp
= get_freepointer(s
, object
);
802 max_objects
= (PAGE_SIZE
<< compound_order(page
)) / s
->size
;
803 if (max_objects
> 65535)
806 if (page
->objects
!= max_objects
) {
807 slab_err(s
, page
, "Wrong number of objects. Found %d but "
808 "should be %d", page
->objects
, max_objects
);
809 page
->objects
= max_objects
;
810 slab_fix(s
, "Number of objects adjusted.");
812 if (page
->inuse
!= page
->objects
- nr
) {
813 slab_err(s
, page
, "Wrong object count. Counter is %d but "
814 "counted were %d", page
->inuse
, page
->objects
- nr
);
815 page
->inuse
= page
->objects
- nr
;
816 slab_fix(s
, "Object count adjusted.");
818 return search
== NULL
;
821 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
, int alloc
)
823 if (s
->flags
& SLAB_TRACE
) {
824 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
826 alloc
? "alloc" : "free",
831 print_section("Object", (void *)object
, s
->objsize
);
838 * Tracking of fully allocated slabs for debugging purposes.
840 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
842 spin_lock(&n
->list_lock
);
843 list_add(&page
->lru
, &n
->full
);
844 spin_unlock(&n
->list_lock
);
847 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
849 struct kmem_cache_node
*n
;
851 if (!(s
->flags
& SLAB_STORE_USER
))
854 n
= get_node(s
, page_to_nid(page
));
856 spin_lock(&n
->list_lock
);
857 list_del(&page
->lru
);
858 spin_unlock(&n
->list_lock
);
861 /* Tracking of the number of slabs for debugging purposes */
862 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
864 struct kmem_cache_node
*n
= get_node(s
, node
);
866 return atomic_long_read(&n
->nr_slabs
);
869 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
)
871 struct kmem_cache_node
*n
= get_node(s
, node
);
874 * May be called early in order to allocate a slab for the
875 * kmem_cache_node structure. Solve the chicken-egg
876 * dilemma by deferring the increment of the count during
877 * bootstrap (see early_kmem_cache_node_alloc).
879 if (!NUMA_BUILD
|| n
)
880 atomic_long_inc(&n
->nr_slabs
);
882 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
)
884 struct kmem_cache_node
*n
= get_node(s
, node
);
886 atomic_long_dec(&n
->nr_slabs
);
889 /* Object debug checks for alloc/free paths */
890 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
893 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
896 init_object(s
, object
, 0);
897 init_tracking(s
, object
);
900 static int alloc_debug_processing(struct kmem_cache
*s
, struct page
*page
,
901 void *object
, void *addr
)
903 if (!check_slab(s
, page
))
906 if (!on_freelist(s
, page
, object
)) {
907 object_err(s
, page
, object
, "Object already allocated");
911 if (!check_valid_pointer(s
, page
, object
)) {
912 object_err(s
, page
, object
, "Freelist Pointer check fails");
916 if (!check_object(s
, page
, object
, 0))
919 /* Success perform special debug activities for allocs */
920 if (s
->flags
& SLAB_STORE_USER
)
921 set_track(s
, object
, TRACK_ALLOC
, addr
);
922 trace(s
, page
, object
, 1);
923 init_object(s
, object
, 1);
927 if (PageSlab(page
)) {
929 * If this is a slab page then lets do the best we can
930 * to avoid issues in the future. Marking all objects
931 * as used avoids touching the remaining objects.
933 slab_fix(s
, "Marking all objects used");
934 page
->inuse
= page
->objects
;
935 page
->freelist
= NULL
;
940 static int free_debug_processing(struct kmem_cache
*s
, struct page
*page
,
941 void *object
, void *addr
)
943 if (!check_slab(s
, page
))
946 if (!check_valid_pointer(s
, page
, object
)) {
947 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
951 if (on_freelist(s
, page
, object
)) {
952 object_err(s
, page
, object
, "Object already free");
956 if (!check_object(s
, page
, object
, 1))
959 if (unlikely(s
!= page
->slab
)) {
960 if (!PageSlab(page
)) {
961 slab_err(s
, page
, "Attempt to free object(0x%p) "
962 "outside of slab", object
);
963 } else if (!page
->slab
) {
965 "SLUB <none>: no slab for object 0x%p.\n",
969 object_err(s
, page
, object
,
970 "page slab pointer corrupt.");
974 /* Special debug activities for freeing objects */
975 if (!SlabFrozen(page
) && !page
->freelist
)
976 remove_full(s
, page
);
977 if (s
->flags
& SLAB_STORE_USER
)
978 set_track(s
, object
, TRACK_FREE
, addr
);
979 trace(s
, page
, object
, 0);
980 init_object(s
, object
, 0);
984 slab_fix(s
, "Object at 0x%p not freed", object
);
988 static int __init
setup_slub_debug(char *str
)
990 slub_debug
= DEBUG_DEFAULT_FLAGS
;
991 if (*str
++ != '=' || !*str
)
993 * No options specified. Switch on full debugging.
999 * No options but restriction on slabs. This means full
1000 * debugging for slabs matching a pattern.
1007 * Switch off all debugging measures.
1012 * Determine which debug features should be switched on
1014 for (; *str
&& *str
!= ','; str
++) {
1015 switch (tolower(*str
)) {
1017 slub_debug
|= SLAB_DEBUG_FREE
;
1020 slub_debug
|= SLAB_RED_ZONE
;
1023 slub_debug
|= SLAB_POISON
;
1026 slub_debug
|= SLAB_STORE_USER
;
1029 slub_debug
|= SLAB_TRACE
;
1032 printk(KERN_ERR
"slub_debug option '%c' "
1033 "unknown. skipped\n", *str
);
1039 slub_debug_slabs
= str
+ 1;
1044 __setup("slub_debug", setup_slub_debug
);
1046 static unsigned long kmem_cache_flags(unsigned long objsize
,
1047 unsigned long flags
, const char *name
,
1048 void (*ctor
)(struct kmem_cache
*, void *))
1051 * Enable debugging if selected on the kernel commandline.
1053 if (slub_debug
&& (!slub_debug_slabs
||
1054 strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)) == 0))
1055 flags
|= slub_debug
;
1060 static inline void setup_object_debug(struct kmem_cache
*s
,
1061 struct page
*page
, void *object
) {}
1063 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1064 struct page
*page
, void *object
, void *addr
) { return 0; }
1066 static inline int free_debug_processing(struct kmem_cache
*s
,
1067 struct page
*page
, void *object
, void *addr
) { return 0; }
1069 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1071 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1072 void *object
, int active
) { return 1; }
1073 static inline void add_full(struct kmem_cache_node
*n
, struct page
*page
) {}
1074 static inline unsigned long kmem_cache_flags(unsigned long objsize
,
1075 unsigned long flags
, const char *name
,
1076 void (*ctor
)(struct kmem_cache
*, void *))
1080 #define slub_debug 0
1082 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1084 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
) {}
1085 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
) {}
1088 * Slab allocation and freeing
1090 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1093 int pages
= 1 << s
->order
;
1095 flags
|= s
->allocflags
;
1098 page
= alloc_pages(flags
, s
->order
);
1100 page
= alloc_pages_node(node
, flags
, s
->order
);
1105 page
->objects
= s
->objects
;
1106 mod_zone_page_state(page_zone(page
),
1107 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1108 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1114 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1117 setup_object_debug(s
, page
, object
);
1118 if (unlikely(s
->ctor
))
1122 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1129 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1131 page
= allocate_slab(s
,
1132 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1136 inc_slabs_node(s
, page_to_nid(page
));
1138 page
->flags
|= 1 << PG_slab
;
1139 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
1140 SLAB_STORE_USER
| SLAB_TRACE
))
1143 start
= page_address(page
);
1145 if (unlikely(s
->flags
& SLAB_POISON
))
1146 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
1149 for_each_object(p
, s
, start
, page
->objects
) {
1150 setup_object(s
, page
, last
);
1151 set_freepointer(s
, last
, p
);
1154 setup_object(s
, page
, last
);
1155 set_freepointer(s
, last
, NULL
);
1157 page
->freelist
= start
;
1163 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1165 int pages
= 1 << s
->order
;
1167 if (unlikely(SlabDebug(page
))) {
1170 slab_pad_check(s
, page
);
1171 for_each_object(p
, s
, page_address(page
),
1173 check_object(s
, page
, p
, 0);
1174 ClearSlabDebug(page
);
1177 mod_zone_page_state(page_zone(page
),
1178 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1179 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1182 __ClearPageSlab(page
);
1183 reset_page_mapcount(page
);
1184 __free_pages(page
, s
->order
);
1187 static void rcu_free_slab(struct rcu_head
*h
)
1191 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1192 __free_slab(page
->slab
, page
);
1195 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1197 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1199 * RCU free overloads the RCU head over the LRU
1201 struct rcu_head
*head
= (void *)&page
->lru
;
1203 call_rcu(head
, rcu_free_slab
);
1205 __free_slab(s
, page
);
1208 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1210 dec_slabs_node(s
, page_to_nid(page
));
1215 * Per slab locking using the pagelock
1217 static __always_inline
void slab_lock(struct page
*page
)
1219 bit_spin_lock(PG_locked
, &page
->flags
);
1222 static __always_inline
void slab_unlock(struct page
*page
)
1224 __bit_spin_unlock(PG_locked
, &page
->flags
);
1227 static __always_inline
int slab_trylock(struct page
*page
)
1231 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
1236 * Management of partially allocated slabs
1238 static void add_partial(struct kmem_cache_node
*n
,
1239 struct page
*page
, int tail
)
1241 spin_lock(&n
->list_lock
);
1244 list_add_tail(&page
->lru
, &n
->partial
);
1246 list_add(&page
->lru
, &n
->partial
);
1247 spin_unlock(&n
->list_lock
);
1250 static void remove_partial(struct kmem_cache
*s
,
1253 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1255 spin_lock(&n
->list_lock
);
1256 list_del(&page
->lru
);
1258 spin_unlock(&n
->list_lock
);
1262 * Lock slab and remove from the partial list.
1264 * Must hold list_lock.
1266 static inline int lock_and_freeze_slab(struct kmem_cache_node
*n
, struct page
*page
)
1268 if (slab_trylock(page
)) {
1269 list_del(&page
->lru
);
1271 SetSlabFrozen(page
);
1278 * Try to allocate a partial slab from a specific node.
1280 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
1285 * Racy check. If we mistakenly see no partial slabs then we
1286 * just allocate an empty slab. If we mistakenly try to get a
1287 * partial slab and there is none available then get_partials()
1290 if (!n
|| !n
->nr_partial
)
1293 spin_lock(&n
->list_lock
);
1294 list_for_each_entry(page
, &n
->partial
, lru
)
1295 if (lock_and_freeze_slab(n
, page
))
1299 spin_unlock(&n
->list_lock
);
1304 * Get a page from somewhere. Search in increasing NUMA distances.
1306 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1309 struct zonelist
*zonelist
;
1314 * The defrag ratio allows a configuration of the tradeoffs between
1315 * inter node defragmentation and node local allocations. A lower
1316 * defrag_ratio increases the tendency to do local allocations
1317 * instead of attempting to obtain partial slabs from other nodes.
1319 * If the defrag_ratio is set to 0 then kmalloc() always
1320 * returns node local objects. If the ratio is higher then kmalloc()
1321 * may return off node objects because partial slabs are obtained
1322 * from other nodes and filled up.
1324 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1325 * defrag_ratio = 1000) then every (well almost) allocation will
1326 * first attempt to defrag slab caches on other nodes. This means
1327 * scanning over all nodes to look for partial slabs which may be
1328 * expensive if we do it every time we are trying to find a slab
1329 * with available objects.
1331 if (!s
->remote_node_defrag_ratio
||
1332 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1335 zonelist
= &NODE_DATA(
1336 slab_node(current
->mempolicy
))->node_zonelists
[gfp_zone(flags
)];
1337 for (z
= zonelist
->zones
; *z
; z
++) {
1338 struct kmem_cache_node
*n
;
1340 n
= get_node(s
, zone_to_nid(*z
));
1342 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1343 n
->nr_partial
> MIN_PARTIAL
) {
1344 page
= get_partial_node(n
);
1354 * Get a partial page, lock it and return it.
1356 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1359 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1361 page
= get_partial_node(get_node(s
, searchnode
));
1362 if (page
|| (flags
& __GFP_THISNODE
))
1365 return get_any_partial(s
, flags
);
1369 * Move a page back to the lists.
1371 * Must be called with the slab lock held.
1373 * On exit the slab lock will have been dropped.
1375 static void unfreeze_slab(struct kmem_cache
*s
, struct page
*page
, int tail
)
1377 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1378 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, smp_processor_id());
1380 ClearSlabFrozen(page
);
1383 if (page
->freelist
) {
1384 add_partial(n
, page
, tail
);
1385 stat(c
, tail
? DEACTIVATE_TO_TAIL
: DEACTIVATE_TO_HEAD
);
1387 stat(c
, DEACTIVATE_FULL
);
1388 if (SlabDebug(page
) && (s
->flags
& SLAB_STORE_USER
))
1393 stat(c
, DEACTIVATE_EMPTY
);
1394 if (n
->nr_partial
< MIN_PARTIAL
) {
1396 * Adding an empty slab to the partial slabs in order
1397 * to avoid page allocator overhead. This slab needs
1398 * to come after the other slabs with objects in
1399 * so that the others get filled first. That way the
1400 * size of the partial list stays small.
1402 * kmem_cache_shrink can reclaim any empty slabs from the
1405 add_partial(n
, page
, 1);
1409 stat(get_cpu_slab(s
, raw_smp_processor_id()), FREE_SLAB
);
1410 discard_slab(s
, page
);
1416 * Remove the cpu slab
1418 static void deactivate_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1420 struct page
*page
= c
->page
;
1424 stat(c
, DEACTIVATE_REMOTE_FREES
);
1426 * Merge cpu freelist into slab freelist. Typically we get here
1427 * because both freelists are empty. So this is unlikely
1430 while (unlikely(c
->freelist
)) {
1433 tail
= 0; /* Hot objects. Put the slab first */
1435 /* Retrieve object from cpu_freelist */
1436 object
= c
->freelist
;
1437 c
->freelist
= c
->freelist
[c
->offset
];
1439 /* And put onto the regular freelist */
1440 object
[c
->offset
] = page
->freelist
;
1441 page
->freelist
= object
;
1445 unfreeze_slab(s
, page
, tail
);
1448 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1450 stat(c
, CPUSLAB_FLUSH
);
1452 deactivate_slab(s
, c
);
1458 * Called from IPI handler with interrupts disabled.
1460 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1462 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
1464 if (likely(c
&& c
->page
))
1468 static void flush_cpu_slab(void *d
)
1470 struct kmem_cache
*s
= d
;
1472 __flush_cpu_slab(s
, smp_processor_id());
1475 static void flush_all(struct kmem_cache
*s
)
1478 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1480 unsigned long flags
;
1482 local_irq_save(flags
);
1484 local_irq_restore(flags
);
1489 * Check if the objects in a per cpu structure fit numa
1490 * locality expectations.
1492 static inline int node_match(struct kmem_cache_cpu
*c
, int node
)
1495 if (node
!= -1 && c
->node
!= node
)
1502 * Slow path. The lockless freelist is empty or we need to perform
1505 * Interrupts are disabled.
1507 * Processing is still very fast if new objects have been freed to the
1508 * regular freelist. In that case we simply take over the regular freelist
1509 * as the lockless freelist and zap the regular freelist.
1511 * If that is not working then we fall back to the partial lists. We take the
1512 * first element of the freelist as the object to allocate now and move the
1513 * rest of the freelist to the lockless freelist.
1515 * And if we were unable to get a new slab from the partial slab lists then
1516 * we need to allocate a new slab. This is the slowest path since it involves
1517 * a call to the page allocator and the setup of a new slab.
1519 static void *__slab_alloc(struct kmem_cache
*s
,
1520 gfp_t gfpflags
, int node
, void *addr
, struct kmem_cache_cpu
*c
)
1525 /* We handle __GFP_ZERO in the caller */
1526 gfpflags
&= ~__GFP_ZERO
;
1532 if (unlikely(!node_match(c
, node
)))
1535 stat(c
, ALLOC_REFILL
);
1538 object
= c
->page
->freelist
;
1539 if (unlikely(!object
))
1541 if (unlikely(SlabDebug(c
->page
)))
1544 c
->freelist
= object
[c
->offset
];
1545 c
->page
->inuse
= c
->page
->objects
;
1546 c
->page
->freelist
= NULL
;
1547 c
->node
= page_to_nid(c
->page
);
1549 slab_unlock(c
->page
);
1550 stat(c
, ALLOC_SLOWPATH
);
1554 deactivate_slab(s
, c
);
1557 new = get_partial(s
, gfpflags
, node
);
1560 stat(c
, ALLOC_FROM_PARTIAL
);
1564 if (gfpflags
& __GFP_WAIT
)
1567 new = new_slab(s
, gfpflags
, node
);
1569 if (gfpflags
& __GFP_WAIT
)
1570 local_irq_disable();
1573 c
= get_cpu_slab(s
, smp_processor_id());
1574 stat(c
, ALLOC_SLAB
);
1584 * No memory available.
1586 * If the slab uses higher order allocs but the object is
1587 * smaller than a page size then we can fallback in emergencies
1588 * to the page allocator via kmalloc_large. The page allocator may
1589 * have failed to obtain a higher order page and we can try to
1590 * allocate a single page if the object fits into a single page.
1591 * That is only possible if certain conditions are met that are being
1592 * checked when a slab is created.
1594 if (!(gfpflags
& __GFP_NORETRY
) &&
1595 (s
->flags
& __PAGE_ALLOC_FALLBACK
)) {
1596 if (gfpflags
& __GFP_WAIT
)
1598 object
= kmalloc_large(s
->objsize
, gfpflags
);
1599 if (gfpflags
& __GFP_WAIT
)
1600 local_irq_disable();
1605 if (!alloc_debug_processing(s
, c
->page
, object
, addr
))
1609 c
->page
->freelist
= object
[c
->offset
];
1615 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1616 * have the fastpath folded into their functions. So no function call
1617 * overhead for requests that can be satisfied on the fastpath.
1619 * The fastpath works by first checking if the lockless freelist can be used.
1620 * If not then __slab_alloc is called for slow processing.
1622 * Otherwise we can simply pick the next object from the lockless free list.
1624 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
1625 gfp_t gfpflags
, int node
, void *addr
)
1628 struct kmem_cache_cpu
*c
;
1629 unsigned long flags
;
1631 local_irq_save(flags
);
1632 c
= get_cpu_slab(s
, smp_processor_id());
1633 if (unlikely(!c
->freelist
|| !node_match(c
, node
)))
1635 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
1638 object
= c
->freelist
;
1639 c
->freelist
= object
[c
->offset
];
1640 stat(c
, ALLOC_FASTPATH
);
1642 local_irq_restore(flags
);
1644 if (unlikely((gfpflags
& __GFP_ZERO
) && object
))
1645 memset(object
, 0, c
->objsize
);
1650 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1652 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1654 EXPORT_SYMBOL(kmem_cache_alloc
);
1657 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1659 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1661 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1665 * Slow patch handling. This may still be called frequently since objects
1666 * have a longer lifetime than the cpu slabs in most processing loads.
1668 * So we still attempt to reduce cache line usage. Just take the slab
1669 * lock and free the item. If there is no additional partial page
1670 * handling required then we can return immediately.
1672 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
1673 void *x
, void *addr
, unsigned int offset
)
1676 void **object
= (void *)x
;
1677 struct kmem_cache_cpu
*c
;
1679 c
= get_cpu_slab(s
, raw_smp_processor_id());
1680 stat(c
, FREE_SLOWPATH
);
1683 if (unlikely(SlabDebug(page
)))
1687 prior
= object
[offset
] = page
->freelist
;
1688 page
->freelist
= object
;
1691 if (unlikely(SlabFrozen(page
))) {
1692 stat(c
, FREE_FROZEN
);
1696 if (unlikely(!page
->inuse
))
1700 * Objects left in the slab. If it was not on the partial list before
1703 if (unlikely(!prior
)) {
1704 add_partial(get_node(s
, page_to_nid(page
)), page
, 1);
1705 stat(c
, FREE_ADD_PARTIAL
);
1715 * Slab still on the partial list.
1717 remove_partial(s
, page
);
1718 stat(c
, FREE_REMOVE_PARTIAL
);
1722 discard_slab(s
, page
);
1726 if (!free_debug_processing(s
, page
, x
, addr
))
1732 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1733 * can perform fastpath freeing without additional function calls.
1735 * The fastpath is only possible if we are freeing to the current cpu slab
1736 * of this processor. This typically the case if we have just allocated
1739 * If fastpath is not possible then fall back to __slab_free where we deal
1740 * with all sorts of special processing.
1742 static __always_inline
void slab_free(struct kmem_cache
*s
,
1743 struct page
*page
, void *x
, void *addr
)
1745 void **object
= (void *)x
;
1746 struct kmem_cache_cpu
*c
;
1747 unsigned long flags
;
1749 local_irq_save(flags
);
1750 c
= get_cpu_slab(s
, smp_processor_id());
1751 debug_check_no_locks_freed(object
, c
->objsize
);
1752 if (likely(page
== c
->page
&& c
->node
>= 0)) {
1753 object
[c
->offset
] = c
->freelist
;
1754 c
->freelist
= object
;
1755 stat(c
, FREE_FASTPATH
);
1757 __slab_free(s
, page
, x
, addr
, c
->offset
);
1759 local_irq_restore(flags
);
1762 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1766 page
= virt_to_head_page(x
);
1768 slab_free(s
, page
, x
, __builtin_return_address(0));
1770 EXPORT_SYMBOL(kmem_cache_free
);
1772 /* Figure out on which slab object the object resides */
1773 static struct page
*get_object_page(const void *x
)
1775 struct page
*page
= virt_to_head_page(x
);
1777 if (!PageSlab(page
))
1784 * Object placement in a slab is made very easy because we always start at
1785 * offset 0. If we tune the size of the object to the alignment then we can
1786 * get the required alignment by putting one properly sized object after
1789 * Notice that the allocation order determines the sizes of the per cpu
1790 * caches. Each processor has always one slab available for allocations.
1791 * Increasing the allocation order reduces the number of times that slabs
1792 * must be moved on and off the partial lists and is therefore a factor in
1797 * Mininum / Maximum order of slab pages. This influences locking overhead
1798 * and slab fragmentation. A higher order reduces the number of partial slabs
1799 * and increases the number of allocations possible without having to
1800 * take the list_lock.
1802 static int slub_min_order
;
1803 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1804 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1807 * Merge control. If this is set then no merging of slab caches will occur.
1808 * (Could be removed. This was introduced to pacify the merge skeptics.)
1810 static int slub_nomerge
;
1813 * Calculate the order of allocation given an slab object size.
1815 * The order of allocation has significant impact on performance and other
1816 * system components. Generally order 0 allocations should be preferred since
1817 * order 0 does not cause fragmentation in the page allocator. Larger objects
1818 * be problematic to put into order 0 slabs because there may be too much
1819 * unused space left. We go to a higher order if more than 1/8th of the slab
1822 * In order to reach satisfactory performance we must ensure that a minimum
1823 * number of objects is in one slab. Otherwise we may generate too much
1824 * activity on the partial lists which requires taking the list_lock. This is
1825 * less a concern for large slabs though which are rarely used.
1827 * slub_max_order specifies the order where we begin to stop considering the
1828 * number of objects in a slab as critical. If we reach slub_max_order then
1829 * we try to keep the page order as low as possible. So we accept more waste
1830 * of space in favor of a small page order.
1832 * Higher order allocations also allow the placement of more objects in a
1833 * slab and thereby reduce object handling overhead. If the user has
1834 * requested a higher mininum order then we start with that one instead of
1835 * the smallest order which will fit the object.
1837 static inline int slab_order(int size
, int min_objects
,
1838 int max_order
, int fract_leftover
)
1842 int min_order
= slub_min_order
;
1844 if ((PAGE_SIZE
<< min_order
) / size
> 65535)
1845 return get_order(size
* 65535) - 1;
1847 for (order
= max(min_order
,
1848 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
1849 order
<= max_order
; order
++) {
1851 unsigned long slab_size
= PAGE_SIZE
<< order
;
1853 if (slab_size
< min_objects
* size
)
1856 rem
= slab_size
% size
;
1858 if (rem
<= slab_size
/ fract_leftover
)
1866 static inline int calculate_order(int size
)
1873 * Attempt to find best configuration for a slab. This
1874 * works by first attempting to generate a layout with
1875 * the best configuration and backing off gradually.
1877 * First we reduce the acceptable waste in a slab. Then
1878 * we reduce the minimum objects required in a slab.
1880 min_objects
= slub_min_objects
;
1881 while (min_objects
> 1) {
1883 while (fraction
>= 4) {
1884 order
= slab_order(size
, min_objects
,
1885 slub_max_order
, fraction
);
1886 if (order
<= slub_max_order
)
1894 * We were unable to place multiple objects in a slab. Now
1895 * lets see if we can place a single object there.
1897 order
= slab_order(size
, 1, slub_max_order
, 1);
1898 if (order
<= slub_max_order
)
1902 * Doh this slab cannot be placed using slub_max_order.
1904 order
= slab_order(size
, 1, MAX_ORDER
, 1);
1905 if (order
<= MAX_ORDER
)
1911 * Figure out what the alignment of the objects will be.
1913 static unsigned long calculate_alignment(unsigned long flags
,
1914 unsigned long align
, unsigned long size
)
1917 * If the user wants hardware cache aligned objects then follow that
1918 * suggestion if the object is sufficiently large.
1920 * The hardware cache alignment cannot override the specified
1921 * alignment though. If that is greater then use it.
1923 if (flags
& SLAB_HWCACHE_ALIGN
) {
1924 unsigned long ralign
= cache_line_size();
1925 while (size
<= ralign
/ 2)
1927 align
= max(align
, ralign
);
1930 if (align
< ARCH_SLAB_MINALIGN
)
1931 align
= ARCH_SLAB_MINALIGN
;
1933 return ALIGN(align
, sizeof(void *));
1936 static void init_kmem_cache_cpu(struct kmem_cache
*s
,
1937 struct kmem_cache_cpu
*c
)
1942 c
->offset
= s
->offset
/ sizeof(void *);
1943 c
->objsize
= s
->objsize
;
1944 #ifdef CONFIG_SLUB_STATS
1945 memset(c
->stat
, 0, NR_SLUB_STAT_ITEMS
* sizeof(unsigned));
1949 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1952 spin_lock_init(&n
->list_lock
);
1953 INIT_LIST_HEAD(&n
->partial
);
1954 #ifdef CONFIG_SLUB_DEBUG
1955 atomic_long_set(&n
->nr_slabs
, 0);
1956 INIT_LIST_HEAD(&n
->full
);
1962 * Per cpu array for per cpu structures.
1964 * The per cpu array places all kmem_cache_cpu structures from one processor
1965 * close together meaning that it becomes possible that multiple per cpu
1966 * structures are contained in one cacheline. This may be particularly
1967 * beneficial for the kmalloc caches.
1969 * A desktop system typically has around 60-80 slabs. With 100 here we are
1970 * likely able to get per cpu structures for all caches from the array defined
1971 * here. We must be able to cover all kmalloc caches during bootstrap.
1973 * If the per cpu array is exhausted then fall back to kmalloc
1974 * of individual cachelines. No sharing is possible then.
1976 #define NR_KMEM_CACHE_CPU 100
1978 static DEFINE_PER_CPU(struct kmem_cache_cpu
,
1979 kmem_cache_cpu
)[NR_KMEM_CACHE_CPU
];
1981 static DEFINE_PER_CPU(struct kmem_cache_cpu
*, kmem_cache_cpu_free
);
1982 static cpumask_t kmem_cach_cpu_free_init_once
= CPU_MASK_NONE
;
1984 static struct kmem_cache_cpu
*alloc_kmem_cache_cpu(struct kmem_cache
*s
,
1985 int cpu
, gfp_t flags
)
1987 struct kmem_cache_cpu
*c
= per_cpu(kmem_cache_cpu_free
, cpu
);
1990 per_cpu(kmem_cache_cpu_free
, cpu
) =
1991 (void *)c
->freelist
;
1993 /* Table overflow: So allocate ourselves */
1995 ALIGN(sizeof(struct kmem_cache_cpu
), cache_line_size()),
1996 flags
, cpu_to_node(cpu
));
2001 init_kmem_cache_cpu(s
, c
);
2005 static void free_kmem_cache_cpu(struct kmem_cache_cpu
*c
, int cpu
)
2007 if (c
< per_cpu(kmem_cache_cpu
, cpu
) ||
2008 c
> per_cpu(kmem_cache_cpu
, cpu
) + NR_KMEM_CACHE_CPU
) {
2012 c
->freelist
= (void *)per_cpu(kmem_cache_cpu_free
, cpu
);
2013 per_cpu(kmem_cache_cpu_free
, cpu
) = c
;
2016 static void free_kmem_cache_cpus(struct kmem_cache
*s
)
2020 for_each_online_cpu(cpu
) {
2021 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
2024 s
->cpu_slab
[cpu
] = NULL
;
2025 free_kmem_cache_cpu(c
, cpu
);
2030 static int alloc_kmem_cache_cpus(struct kmem_cache
*s
, gfp_t flags
)
2034 for_each_online_cpu(cpu
) {
2035 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
2040 c
= alloc_kmem_cache_cpu(s
, cpu
, flags
);
2042 free_kmem_cache_cpus(s
);
2045 s
->cpu_slab
[cpu
] = c
;
2051 * Initialize the per cpu array.
2053 static void init_alloc_cpu_cpu(int cpu
)
2057 if (cpu_isset(cpu
, kmem_cach_cpu_free_init_once
))
2060 for (i
= NR_KMEM_CACHE_CPU
- 1; i
>= 0; i
--)
2061 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu
, cpu
)[i
], cpu
);
2063 cpu_set(cpu
, kmem_cach_cpu_free_init_once
);
2066 static void __init
init_alloc_cpu(void)
2070 for_each_online_cpu(cpu
)
2071 init_alloc_cpu_cpu(cpu
);
2075 static inline void free_kmem_cache_cpus(struct kmem_cache
*s
) {}
2076 static inline void init_alloc_cpu(void) {}
2078 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
, gfp_t flags
)
2080 init_kmem_cache_cpu(s
, &s
->cpu_slab
);
2087 * No kmalloc_node yet so do it by hand. We know that this is the first
2088 * slab on the node for this slabcache. There are no concurrent accesses
2091 * Note that this function only works on the kmalloc_node_cache
2092 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2093 * memory on a fresh node that has no slab structures yet.
2095 static struct kmem_cache_node
*early_kmem_cache_node_alloc(gfp_t gfpflags
,
2099 struct kmem_cache_node
*n
;
2100 unsigned long flags
;
2102 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
2104 page
= new_slab(kmalloc_caches
, gfpflags
, node
);
2107 if (page_to_nid(page
) != node
) {
2108 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
2110 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
2111 "in order to be able to continue\n");
2116 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
2118 kmalloc_caches
->node
[node
] = n
;
2119 #ifdef CONFIG_SLUB_DEBUG
2120 init_object(kmalloc_caches
, n
, 1);
2121 init_tracking(kmalloc_caches
, n
);
2123 init_kmem_cache_node(n
);
2124 inc_slabs_node(kmalloc_caches
, node
);
2127 * lockdep requires consistent irq usage for each lock
2128 * so even though there cannot be a race this early in
2129 * the boot sequence, we still disable irqs.
2131 local_irq_save(flags
);
2132 add_partial(n
, page
, 0);
2133 local_irq_restore(flags
);
2137 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2141 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2142 struct kmem_cache_node
*n
= s
->node
[node
];
2143 if (n
&& n
!= &s
->local_node
)
2144 kmem_cache_free(kmalloc_caches
, n
);
2145 s
->node
[node
] = NULL
;
2149 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
2154 if (slab_state
>= UP
)
2155 local_node
= page_to_nid(virt_to_page(s
));
2159 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2160 struct kmem_cache_node
*n
;
2162 if (local_node
== node
)
2165 if (slab_state
== DOWN
) {
2166 n
= early_kmem_cache_node_alloc(gfpflags
,
2170 n
= kmem_cache_alloc_node(kmalloc_caches
,
2174 free_kmem_cache_nodes(s
);
2180 init_kmem_cache_node(n
);
2185 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2189 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
2191 init_kmem_cache_node(&s
->local_node
);
2197 * calculate_sizes() determines the order and the distribution of data within
2200 static int calculate_sizes(struct kmem_cache
*s
)
2202 unsigned long flags
= s
->flags
;
2203 unsigned long size
= s
->objsize
;
2204 unsigned long align
= s
->align
;
2207 * Round up object size to the next word boundary. We can only
2208 * place the free pointer at word boundaries and this determines
2209 * the possible location of the free pointer.
2211 size
= ALIGN(size
, sizeof(void *));
2213 #ifdef CONFIG_SLUB_DEBUG
2215 * Determine if we can poison the object itself. If the user of
2216 * the slab may touch the object after free or before allocation
2217 * then we should never poison the object itself.
2219 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2221 s
->flags
|= __OBJECT_POISON
;
2223 s
->flags
&= ~__OBJECT_POISON
;
2227 * If we are Redzoning then check if there is some space between the
2228 * end of the object and the free pointer. If not then add an
2229 * additional word to have some bytes to store Redzone information.
2231 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
2232 size
+= sizeof(void *);
2236 * With that we have determined the number of bytes in actual use
2237 * by the object. This is the potential offset to the free pointer.
2241 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
2244 * Relocate free pointer after the object if it is not
2245 * permitted to overwrite the first word of the object on
2248 * This is the case if we do RCU, have a constructor or
2249 * destructor or are poisoning the objects.
2252 size
+= sizeof(void *);
2255 #ifdef CONFIG_SLUB_DEBUG
2256 if (flags
& SLAB_STORE_USER
)
2258 * Need to store information about allocs and frees after
2261 size
+= 2 * sizeof(struct track
);
2263 if (flags
& SLAB_RED_ZONE
)
2265 * Add some empty padding so that we can catch
2266 * overwrites from earlier objects rather than let
2267 * tracking information or the free pointer be
2268 * corrupted if an user writes before the start
2271 size
+= sizeof(void *);
2275 * Determine the alignment based on various parameters that the
2276 * user specified and the dynamic determination of cache line size
2279 align
= calculate_alignment(flags
, align
, s
->objsize
);
2282 * SLUB stores one object immediately after another beginning from
2283 * offset 0. In order to align the objects we have to simply size
2284 * each object to conform to the alignment.
2286 size
= ALIGN(size
, align
);
2289 if ((flags
& __KMALLOC_CACHE
) &&
2290 PAGE_SIZE
/ size
< slub_min_objects
) {
2292 * Kmalloc cache that would not have enough objects in
2293 * an order 0 page. Kmalloc slabs can fallback to
2294 * page allocator order 0 allocs so take a reasonably large
2295 * order that will allows us a good number of objects.
2297 s
->order
= max(slub_max_order
, PAGE_ALLOC_COSTLY_ORDER
);
2298 s
->flags
|= __PAGE_ALLOC_FALLBACK
;
2299 s
->allocflags
|= __GFP_NOWARN
;
2301 s
->order
= calculate_order(size
);
2308 s
->allocflags
|= __GFP_COMP
;
2310 if (s
->flags
& SLAB_CACHE_DMA
)
2311 s
->allocflags
|= SLUB_DMA
;
2313 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
2314 s
->allocflags
|= __GFP_RECLAIMABLE
;
2317 * Determine the number of objects per slab
2319 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
2321 return !!s
->objects
;
2325 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
2326 const char *name
, size_t size
,
2327 size_t align
, unsigned long flags
,
2328 void (*ctor
)(struct kmem_cache
*, void *))
2330 memset(s
, 0, kmem_size
);
2335 s
->flags
= kmem_cache_flags(size
, flags
, name
, ctor
);
2337 if (!calculate_sizes(s
))
2342 s
->remote_node_defrag_ratio
= 100;
2344 if (!init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
2347 if (alloc_kmem_cache_cpus(s
, gfpflags
& ~SLUB_DMA
))
2349 free_kmem_cache_nodes(s
);
2351 if (flags
& SLAB_PANIC
)
2352 panic("Cannot create slab %s size=%lu realsize=%u "
2353 "order=%u offset=%u flags=%lx\n",
2354 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
2360 * Check if a given pointer is valid
2362 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
2366 page
= get_object_page(object
);
2368 if (!page
|| s
!= page
->slab
)
2369 /* No slab or wrong slab */
2372 if (!check_valid_pointer(s
, page
, object
))
2376 * We could also check if the object is on the slabs freelist.
2377 * But this would be too expensive and it seems that the main
2378 * purpose of kmem_ptr_valid() is to check if the object belongs
2379 * to a certain slab.
2383 EXPORT_SYMBOL(kmem_ptr_validate
);
2386 * Determine the size of a slab object
2388 unsigned int kmem_cache_size(struct kmem_cache
*s
)
2392 EXPORT_SYMBOL(kmem_cache_size
);
2394 const char *kmem_cache_name(struct kmem_cache
*s
)
2398 EXPORT_SYMBOL(kmem_cache_name
);
2400 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
2403 #ifdef CONFIG_SLUB_DEBUG
2404 void *addr
= page_address(page
);
2406 DECLARE_BITMAP(map
, page
->objects
);
2408 bitmap_zero(map
, page
->objects
);
2409 slab_err(s
, page
, "%s", text
);
2411 for_each_free_object(p
, s
, page
->freelist
)
2412 set_bit(slab_index(p
, s
, addr
), map
);
2414 for_each_object(p
, s
, addr
, page
->objects
) {
2416 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
2417 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu\n",
2419 print_tracking(s
, p
);
2427 * Attempt to free all partial slabs on a node.
2429 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
2431 unsigned long flags
;
2432 struct page
*page
, *h
;
2434 spin_lock_irqsave(&n
->list_lock
, flags
);
2435 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
2437 list_del(&page
->lru
);
2438 discard_slab(s
, page
);
2441 list_slab_objects(s
, page
,
2442 "Objects remaining on kmem_cache_close()");
2445 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2449 * Release all resources used by a slab cache.
2451 static inline int kmem_cache_close(struct kmem_cache
*s
)
2457 /* Attempt to free all objects */
2458 free_kmem_cache_cpus(s
);
2459 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2460 struct kmem_cache_node
*n
= get_node(s
, node
);
2463 if (n
->nr_partial
|| slabs_node(s
, node
))
2466 free_kmem_cache_nodes(s
);
2471 * Close a cache and release the kmem_cache structure
2472 * (must be used for caches created using kmem_cache_create)
2474 void kmem_cache_destroy(struct kmem_cache
*s
)
2476 down_write(&slub_lock
);
2480 up_write(&slub_lock
);
2481 if (kmem_cache_close(s
)) {
2482 printk(KERN_ERR
"SLUB %s: %s called for cache that "
2483 "still has objects.\n", s
->name
, __func__
);
2486 sysfs_slab_remove(s
);
2488 up_write(&slub_lock
);
2490 EXPORT_SYMBOL(kmem_cache_destroy
);
2492 /********************************************************************
2494 *******************************************************************/
2496 struct kmem_cache kmalloc_caches
[PAGE_SHIFT
+ 1] __cacheline_aligned
;
2497 EXPORT_SYMBOL(kmalloc_caches
);
2499 static int __init
setup_slub_min_order(char *str
)
2501 get_option(&str
, &slub_min_order
);
2506 __setup("slub_min_order=", setup_slub_min_order
);
2508 static int __init
setup_slub_max_order(char *str
)
2510 get_option(&str
, &slub_max_order
);
2515 __setup("slub_max_order=", setup_slub_max_order
);
2517 static int __init
setup_slub_min_objects(char *str
)
2519 get_option(&str
, &slub_min_objects
);
2524 __setup("slub_min_objects=", setup_slub_min_objects
);
2526 static int __init
setup_slub_nomerge(char *str
)
2532 __setup("slub_nomerge", setup_slub_nomerge
);
2534 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
2535 const char *name
, int size
, gfp_t gfp_flags
)
2537 unsigned int flags
= 0;
2539 if (gfp_flags
& SLUB_DMA
)
2540 flags
= SLAB_CACHE_DMA
;
2542 down_write(&slub_lock
);
2543 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
2544 flags
| __KMALLOC_CACHE
, NULL
))
2547 list_add(&s
->list
, &slab_caches
);
2548 up_write(&slub_lock
);
2549 if (sysfs_slab_add(s
))
2554 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2557 #ifdef CONFIG_ZONE_DMA
2558 static struct kmem_cache
*kmalloc_caches_dma
[PAGE_SHIFT
+ 1];
2560 static void sysfs_add_func(struct work_struct
*w
)
2562 struct kmem_cache
*s
;
2564 down_write(&slub_lock
);
2565 list_for_each_entry(s
, &slab_caches
, list
) {
2566 if (s
->flags
& __SYSFS_ADD_DEFERRED
) {
2567 s
->flags
&= ~__SYSFS_ADD_DEFERRED
;
2571 up_write(&slub_lock
);
2574 static DECLARE_WORK(sysfs_add_work
, sysfs_add_func
);
2576 static noinline
struct kmem_cache
*dma_kmalloc_cache(int index
, gfp_t flags
)
2578 struct kmem_cache
*s
;
2582 s
= kmalloc_caches_dma
[index
];
2586 /* Dynamically create dma cache */
2587 if (flags
& __GFP_WAIT
)
2588 down_write(&slub_lock
);
2590 if (!down_write_trylock(&slub_lock
))
2594 if (kmalloc_caches_dma
[index
])
2597 realsize
= kmalloc_caches
[index
].objsize
;
2598 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d",
2599 (unsigned int)realsize
);
2600 s
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2602 if (!s
|| !text
|| !kmem_cache_open(s
, flags
, text
,
2603 realsize
, ARCH_KMALLOC_MINALIGN
,
2604 SLAB_CACHE_DMA
|__SYSFS_ADD_DEFERRED
, NULL
)) {
2610 list_add(&s
->list
, &slab_caches
);
2611 kmalloc_caches_dma
[index
] = s
;
2613 schedule_work(&sysfs_add_work
);
2616 up_write(&slub_lock
);
2618 return kmalloc_caches_dma
[index
];
2623 * Conversion table for small slabs sizes / 8 to the index in the
2624 * kmalloc array. This is necessary for slabs < 192 since we have non power
2625 * of two cache sizes there. The size of larger slabs can be determined using
2628 static s8 size_index
[24] = {
2655 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2661 return ZERO_SIZE_PTR
;
2663 index
= size_index
[(size
- 1) / 8];
2665 index
= fls(size
- 1);
2667 #ifdef CONFIG_ZONE_DMA
2668 if (unlikely((flags
& SLUB_DMA
)))
2669 return dma_kmalloc_cache(index
, flags
);
2672 return &kmalloc_caches
[index
];
2675 void *__kmalloc(size_t size
, gfp_t flags
)
2677 struct kmem_cache
*s
;
2679 if (unlikely(size
> PAGE_SIZE
))
2680 return kmalloc_large(size
, flags
);
2682 s
= get_slab(size
, flags
);
2684 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2687 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2689 EXPORT_SYMBOL(__kmalloc
);
2691 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
2693 struct page
*page
= alloc_pages_node(node
, flags
| __GFP_COMP
,
2697 return page_address(page
);
2703 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2705 struct kmem_cache
*s
;
2707 if (unlikely(size
> PAGE_SIZE
))
2708 return kmalloc_large_node(size
, flags
, node
);
2710 s
= get_slab(size
, flags
);
2712 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2715 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2717 EXPORT_SYMBOL(__kmalloc_node
);
2720 size_t ksize(const void *object
)
2723 struct kmem_cache
*s
;
2725 if (unlikely(object
== ZERO_SIZE_PTR
))
2728 page
= virt_to_head_page(object
);
2730 if (unlikely(!PageSlab(page
)))
2731 return PAGE_SIZE
<< compound_order(page
);
2735 #ifdef CONFIG_SLUB_DEBUG
2737 * Debugging requires use of the padding between object
2738 * and whatever may come after it.
2740 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2745 * If we have the need to store the freelist pointer
2746 * back there or track user information then we can
2747 * only use the space before that information.
2749 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2752 * Else we can use all the padding etc for the allocation
2756 EXPORT_SYMBOL(ksize
);
2758 void kfree(const void *x
)
2761 void *object
= (void *)x
;
2763 if (unlikely(ZERO_OR_NULL_PTR(x
)))
2766 page
= virt_to_head_page(x
);
2767 if (unlikely(!PageSlab(page
))) {
2771 slab_free(page
->slab
, page
, object
, __builtin_return_address(0));
2773 EXPORT_SYMBOL(kfree
);
2776 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2777 * the remaining slabs by the number of items in use. The slabs with the
2778 * most items in use come first. New allocations will then fill those up
2779 * and thus they can be removed from the partial lists.
2781 * The slabs with the least items are placed last. This results in them
2782 * being allocated from last increasing the chance that the last objects
2783 * are freed in them.
2785 int kmem_cache_shrink(struct kmem_cache
*s
)
2789 struct kmem_cache_node
*n
;
2792 struct list_head
*slabs_by_inuse
=
2793 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2794 unsigned long flags
;
2796 if (!slabs_by_inuse
)
2800 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2801 n
= get_node(s
, node
);
2806 for (i
= 0; i
< s
->objects
; i
++)
2807 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2809 spin_lock_irqsave(&n
->list_lock
, flags
);
2812 * Build lists indexed by the items in use in each slab.
2814 * Note that concurrent frees may occur while we hold the
2815 * list_lock. page->inuse here is the upper limit.
2817 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2818 if (!page
->inuse
&& slab_trylock(page
)) {
2820 * Must hold slab lock here because slab_free
2821 * may have freed the last object and be
2822 * waiting to release the slab.
2824 list_del(&page
->lru
);
2827 discard_slab(s
, page
);
2829 list_move(&page
->lru
,
2830 slabs_by_inuse
+ page
->inuse
);
2835 * Rebuild the partial list with the slabs filled up most
2836 * first and the least used slabs at the end.
2838 for (i
= s
->objects
- 1; i
>= 0; i
--)
2839 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2841 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2844 kfree(slabs_by_inuse
);
2847 EXPORT_SYMBOL(kmem_cache_shrink
);
2849 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2850 static int slab_mem_going_offline_callback(void *arg
)
2852 struct kmem_cache
*s
;
2854 down_read(&slub_lock
);
2855 list_for_each_entry(s
, &slab_caches
, list
)
2856 kmem_cache_shrink(s
);
2857 up_read(&slub_lock
);
2862 static void slab_mem_offline_callback(void *arg
)
2864 struct kmem_cache_node
*n
;
2865 struct kmem_cache
*s
;
2866 struct memory_notify
*marg
= arg
;
2869 offline_node
= marg
->status_change_nid
;
2872 * If the node still has available memory. we need kmem_cache_node
2875 if (offline_node
< 0)
2878 down_read(&slub_lock
);
2879 list_for_each_entry(s
, &slab_caches
, list
) {
2880 n
= get_node(s
, offline_node
);
2883 * if n->nr_slabs > 0, slabs still exist on the node
2884 * that is going down. We were unable to free them,
2885 * and offline_pages() function shoudn't call this
2886 * callback. So, we must fail.
2888 BUG_ON(slabs_node(s
, offline_node
));
2890 s
->node
[offline_node
] = NULL
;
2891 kmem_cache_free(kmalloc_caches
, n
);
2894 up_read(&slub_lock
);
2897 static int slab_mem_going_online_callback(void *arg
)
2899 struct kmem_cache_node
*n
;
2900 struct kmem_cache
*s
;
2901 struct memory_notify
*marg
= arg
;
2902 int nid
= marg
->status_change_nid
;
2906 * If the node's memory is already available, then kmem_cache_node is
2907 * already created. Nothing to do.
2913 * We are bringing a node online. No memory is availabe yet. We must
2914 * allocate a kmem_cache_node structure in order to bring the node
2917 down_read(&slub_lock
);
2918 list_for_each_entry(s
, &slab_caches
, list
) {
2920 * XXX: kmem_cache_alloc_node will fallback to other nodes
2921 * since memory is not yet available from the node that
2924 n
= kmem_cache_alloc(kmalloc_caches
, GFP_KERNEL
);
2929 init_kmem_cache_node(n
);
2933 up_read(&slub_lock
);
2937 static int slab_memory_callback(struct notifier_block
*self
,
2938 unsigned long action
, void *arg
)
2943 case MEM_GOING_ONLINE
:
2944 ret
= slab_mem_going_online_callback(arg
);
2946 case MEM_GOING_OFFLINE
:
2947 ret
= slab_mem_going_offline_callback(arg
);
2950 case MEM_CANCEL_ONLINE
:
2951 slab_mem_offline_callback(arg
);
2954 case MEM_CANCEL_OFFLINE
:
2958 ret
= notifier_from_errno(ret
);
2962 #endif /* CONFIG_MEMORY_HOTPLUG */
2964 /********************************************************************
2965 * Basic setup of slabs
2966 *******************************************************************/
2968 void __init
kmem_cache_init(void)
2977 * Must first have the slab cache available for the allocations of the
2978 * struct kmem_cache_node's. There is special bootstrap code in
2979 * kmem_cache_open for slab_state == DOWN.
2981 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2982 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2983 kmalloc_caches
[0].refcount
= -1;
2986 hotplug_memory_notifier(slab_memory_callback
, 1);
2989 /* Able to allocate the per node structures */
2990 slab_state
= PARTIAL
;
2992 /* Caches that are not of the two-to-the-power-of size */
2993 if (KMALLOC_MIN_SIZE
<= 64) {
2994 create_kmalloc_cache(&kmalloc_caches
[1],
2995 "kmalloc-96", 96, GFP_KERNEL
);
2998 if (KMALLOC_MIN_SIZE
<= 128) {
2999 create_kmalloc_cache(&kmalloc_caches
[2],
3000 "kmalloc-192", 192, GFP_KERNEL
);
3004 for (i
= KMALLOC_SHIFT_LOW
; i
<= PAGE_SHIFT
; i
++) {
3005 create_kmalloc_cache(&kmalloc_caches
[i
],
3006 "kmalloc", 1 << i
, GFP_KERNEL
);
3012 * Patch up the size_index table if we have strange large alignment
3013 * requirements for the kmalloc array. This is only the case for
3014 * MIPS it seems. The standard arches will not generate any code here.
3016 * Largest permitted alignment is 256 bytes due to the way we
3017 * handle the index determination for the smaller caches.
3019 * Make sure that nothing crazy happens if someone starts tinkering
3020 * around with ARCH_KMALLOC_MINALIGN
3022 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 256 ||
3023 (KMALLOC_MIN_SIZE
& (KMALLOC_MIN_SIZE
- 1)));
3025 for (i
= 8; i
< KMALLOC_MIN_SIZE
; i
+= 8)
3026 size_index
[(i
- 1) / 8] = KMALLOC_SHIFT_LOW
;
3030 /* Provide the correct kmalloc names now that the caches are up */
3031 for (i
= KMALLOC_SHIFT_LOW
; i
<= PAGE_SHIFT
; i
++)
3032 kmalloc_caches
[i
]. name
=
3033 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
3036 register_cpu_notifier(&slab_notifier
);
3037 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
) +
3038 nr_cpu_ids
* sizeof(struct kmem_cache_cpu
*);
3040 kmem_size
= sizeof(struct kmem_cache
);
3044 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3045 " CPUs=%d, Nodes=%d\n",
3046 caches
, cache_line_size(),
3047 slub_min_order
, slub_max_order
, slub_min_objects
,
3048 nr_cpu_ids
, nr_node_ids
);
3052 * Find a mergeable slab cache
3054 static int slab_unmergeable(struct kmem_cache
*s
)
3056 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
3059 if ((s
->flags
& __PAGE_ALLOC_FALLBACK
))
3066 * We may have set a slab to be unmergeable during bootstrap.
3068 if (s
->refcount
< 0)
3074 static struct kmem_cache
*find_mergeable(size_t size
,
3075 size_t align
, unsigned long flags
, const char *name
,
3076 void (*ctor
)(struct kmem_cache
*, void *))
3078 struct kmem_cache
*s
;
3080 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
3086 size
= ALIGN(size
, sizeof(void *));
3087 align
= calculate_alignment(flags
, align
, size
);
3088 size
= ALIGN(size
, align
);
3089 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
3091 list_for_each_entry(s
, &slab_caches
, list
) {
3092 if (slab_unmergeable(s
))
3098 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
3101 * Check if alignment is compatible.
3102 * Courtesy of Adrian Drzewiecki
3104 if ((s
->size
& ~(align
- 1)) != s
->size
)
3107 if (s
->size
- size
>= sizeof(void *))
3115 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
3116 size_t align
, unsigned long flags
,
3117 void (*ctor
)(struct kmem_cache
*, void *))
3119 struct kmem_cache
*s
;
3121 down_write(&slub_lock
);
3122 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
3128 * Adjust the object sizes so that we clear
3129 * the complete object on kzalloc.
3131 s
->objsize
= max(s
->objsize
, (int)size
);
3134 * And then we need to update the object size in the
3135 * per cpu structures
3137 for_each_online_cpu(cpu
)
3138 get_cpu_slab(s
, cpu
)->objsize
= s
->objsize
;
3140 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3141 up_write(&slub_lock
);
3143 if (sysfs_slab_alias(s
, name
))
3148 s
= kmalloc(kmem_size
, GFP_KERNEL
);
3150 if (kmem_cache_open(s
, GFP_KERNEL
, name
,
3151 size
, align
, flags
, ctor
)) {
3152 list_add(&s
->list
, &slab_caches
);
3153 up_write(&slub_lock
);
3154 if (sysfs_slab_add(s
))
3160 up_write(&slub_lock
);
3163 if (flags
& SLAB_PANIC
)
3164 panic("Cannot create slabcache %s\n", name
);
3169 EXPORT_SYMBOL(kmem_cache_create
);
3173 * Use the cpu notifier to insure that the cpu slabs are flushed when
3176 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
3177 unsigned long action
, void *hcpu
)
3179 long cpu
= (long)hcpu
;
3180 struct kmem_cache
*s
;
3181 unsigned long flags
;
3184 case CPU_UP_PREPARE
:
3185 case CPU_UP_PREPARE_FROZEN
:
3186 init_alloc_cpu_cpu(cpu
);
3187 down_read(&slub_lock
);
3188 list_for_each_entry(s
, &slab_caches
, list
)
3189 s
->cpu_slab
[cpu
] = alloc_kmem_cache_cpu(s
, cpu
,
3191 up_read(&slub_lock
);
3194 case CPU_UP_CANCELED
:
3195 case CPU_UP_CANCELED_FROZEN
:
3197 case CPU_DEAD_FROZEN
:
3198 down_read(&slub_lock
);
3199 list_for_each_entry(s
, &slab_caches
, list
) {
3200 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3202 local_irq_save(flags
);
3203 __flush_cpu_slab(s
, cpu
);
3204 local_irq_restore(flags
);
3205 free_kmem_cache_cpu(c
, cpu
);
3206 s
->cpu_slab
[cpu
] = NULL
;
3208 up_read(&slub_lock
);
3216 static struct notifier_block __cpuinitdata slab_notifier
= {
3217 .notifier_call
= slab_cpuup_callback
3222 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
3224 struct kmem_cache
*s
;
3226 if (unlikely(size
> PAGE_SIZE
))
3227 return kmalloc_large(size
, gfpflags
);
3229 s
= get_slab(size
, gfpflags
);
3231 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3234 return slab_alloc(s
, gfpflags
, -1, caller
);
3237 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3238 int node
, void *caller
)
3240 struct kmem_cache
*s
;
3242 if (unlikely(size
> PAGE_SIZE
))
3243 return kmalloc_large_node(size
, gfpflags
, node
);
3245 s
= get_slab(size
, gfpflags
);
3247 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3250 return slab_alloc(s
, gfpflags
, node
, caller
);
3253 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3254 static unsigned long count_partial(struct kmem_cache_node
*n
)
3256 unsigned long flags
;
3257 unsigned long x
= 0;
3260 spin_lock_irqsave(&n
->list_lock
, flags
);
3261 list_for_each_entry(page
, &n
->partial
, lru
)
3263 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3268 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3269 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3273 void *addr
= page_address(page
);
3275 if (!check_slab(s
, page
) ||
3276 !on_freelist(s
, page
, NULL
))
3279 /* Now we know that a valid freelist exists */
3280 bitmap_zero(map
, page
->objects
);
3282 for_each_free_object(p
, s
, page
->freelist
) {
3283 set_bit(slab_index(p
, s
, addr
), map
);
3284 if (!check_object(s
, page
, p
, 0))
3288 for_each_object(p
, s
, addr
, page
->objects
)
3289 if (!test_bit(slab_index(p
, s
, addr
), map
))
3290 if (!check_object(s
, page
, p
, 1))
3295 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
3298 if (slab_trylock(page
)) {
3299 validate_slab(s
, page
, map
);
3302 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
3305 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
3306 if (!SlabDebug(page
))
3307 printk(KERN_ERR
"SLUB %s: SlabDebug not set "
3308 "on slab 0x%p\n", s
->name
, page
);
3310 if (SlabDebug(page
))
3311 printk(KERN_ERR
"SLUB %s: SlabDebug set on "
3312 "slab 0x%p\n", s
->name
, page
);
3316 static int validate_slab_node(struct kmem_cache
*s
,
3317 struct kmem_cache_node
*n
, unsigned long *map
)
3319 unsigned long count
= 0;
3321 unsigned long flags
;
3323 spin_lock_irqsave(&n
->list_lock
, flags
);
3325 list_for_each_entry(page
, &n
->partial
, lru
) {
3326 validate_slab_slab(s
, page
, map
);
3329 if (count
!= n
->nr_partial
)
3330 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
3331 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
3333 if (!(s
->flags
& SLAB_STORE_USER
))
3336 list_for_each_entry(page
, &n
->full
, lru
) {
3337 validate_slab_slab(s
, page
, map
);
3340 if (count
!= atomic_long_read(&n
->nr_slabs
))
3341 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
3342 "counter=%ld\n", s
->name
, count
,
3343 atomic_long_read(&n
->nr_slabs
));
3346 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3350 static long validate_slab_cache(struct kmem_cache
*s
)
3353 unsigned long count
= 0;
3354 unsigned long *map
= kmalloc(BITS_TO_LONGS(s
->objects
) *
3355 sizeof(unsigned long), GFP_KERNEL
);
3361 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3362 struct kmem_cache_node
*n
= get_node(s
, node
);
3364 count
+= validate_slab_node(s
, n
, map
);
3370 #ifdef SLUB_RESILIENCY_TEST
3371 static void resiliency_test(void)
3375 printk(KERN_ERR
"SLUB resiliency testing\n");
3376 printk(KERN_ERR
"-----------------------\n");
3377 printk(KERN_ERR
"A. Corruption after allocation\n");
3379 p
= kzalloc(16, GFP_KERNEL
);
3381 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
3382 " 0x12->0x%p\n\n", p
+ 16);
3384 validate_slab_cache(kmalloc_caches
+ 4);
3386 /* Hmmm... The next two are dangerous */
3387 p
= kzalloc(32, GFP_KERNEL
);
3388 p
[32 + sizeof(void *)] = 0x34;
3389 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
3390 " 0x34 -> -0x%p\n", p
);
3392 "If allocated object is overwritten then not detectable\n\n");
3394 validate_slab_cache(kmalloc_caches
+ 5);
3395 p
= kzalloc(64, GFP_KERNEL
);
3396 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
3398 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3401 "If allocated object is overwritten then not detectable\n\n");
3402 validate_slab_cache(kmalloc_caches
+ 6);
3404 printk(KERN_ERR
"\nB. Corruption after free\n");
3405 p
= kzalloc(128, GFP_KERNEL
);
3408 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
3409 validate_slab_cache(kmalloc_caches
+ 7);
3411 p
= kzalloc(256, GFP_KERNEL
);
3414 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3416 validate_slab_cache(kmalloc_caches
+ 8);
3418 p
= kzalloc(512, GFP_KERNEL
);
3421 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
3422 validate_slab_cache(kmalloc_caches
+ 9);
3425 static void resiliency_test(void) {};
3429 * Generate lists of code addresses where slabcache objects are allocated
3434 unsigned long count
;
3447 unsigned long count
;
3448 struct location
*loc
;
3451 static void free_loc_track(struct loc_track
*t
)
3454 free_pages((unsigned long)t
->loc
,
3455 get_order(sizeof(struct location
) * t
->max
));
3458 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
3463 order
= get_order(sizeof(struct location
) * max
);
3465 l
= (void *)__get_free_pages(flags
, order
);
3470 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
3478 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
3479 const struct track
*track
)
3481 long start
, end
, pos
;
3484 unsigned long age
= jiffies
- track
->when
;
3490 pos
= start
+ (end
- start
+ 1) / 2;
3493 * There is nothing at "end". If we end up there
3494 * we need to add something to before end.
3499 caddr
= t
->loc
[pos
].addr
;
3500 if (track
->addr
== caddr
) {
3506 if (age
< l
->min_time
)
3508 if (age
> l
->max_time
)
3511 if (track
->pid
< l
->min_pid
)
3512 l
->min_pid
= track
->pid
;
3513 if (track
->pid
> l
->max_pid
)
3514 l
->max_pid
= track
->pid
;
3516 cpu_set(track
->cpu
, l
->cpus
);
3518 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3522 if (track
->addr
< caddr
)
3529 * Not found. Insert new tracking element.
3531 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
3537 (t
->count
- pos
) * sizeof(struct location
));
3540 l
->addr
= track
->addr
;
3544 l
->min_pid
= track
->pid
;
3545 l
->max_pid
= track
->pid
;
3546 cpus_clear(l
->cpus
);
3547 cpu_set(track
->cpu
, l
->cpus
);
3548 nodes_clear(l
->nodes
);
3549 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3553 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
3554 struct page
*page
, enum track_item alloc
)
3556 void *addr
= page_address(page
);
3557 DECLARE_BITMAP(map
, page
->objects
);
3560 bitmap_zero(map
, page
->objects
);
3561 for_each_free_object(p
, s
, page
->freelist
)
3562 set_bit(slab_index(p
, s
, addr
), map
);
3564 for_each_object(p
, s
, addr
, page
->objects
)
3565 if (!test_bit(slab_index(p
, s
, addr
), map
))
3566 add_location(t
, s
, get_track(s
, p
, alloc
));
3569 static int list_locations(struct kmem_cache
*s
, char *buf
,
3570 enum track_item alloc
)
3574 struct loc_track t
= { 0, 0, NULL
};
3577 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
3579 return sprintf(buf
, "Out of memory\n");
3581 /* Push back cpu slabs */
3584 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3585 struct kmem_cache_node
*n
= get_node(s
, node
);
3586 unsigned long flags
;
3589 if (!atomic_long_read(&n
->nr_slabs
))
3592 spin_lock_irqsave(&n
->list_lock
, flags
);
3593 list_for_each_entry(page
, &n
->partial
, lru
)
3594 process_slab(&t
, s
, page
, alloc
);
3595 list_for_each_entry(page
, &n
->full
, lru
)
3596 process_slab(&t
, s
, page
, alloc
);
3597 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3600 for (i
= 0; i
< t
.count
; i
++) {
3601 struct location
*l
= &t
.loc
[i
];
3603 if (len
> PAGE_SIZE
- 100)
3605 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
3608 len
+= sprint_symbol(buf
+ len
, (unsigned long)l
->addr
);
3610 len
+= sprintf(buf
+ len
, "<not-available>");
3612 if (l
->sum_time
!= l
->min_time
) {
3613 unsigned long remainder
;
3615 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
3617 div_long_long_rem(l
->sum_time
, l
->count
, &remainder
),
3620 len
+= sprintf(buf
+ len
, " age=%ld",
3623 if (l
->min_pid
!= l
->max_pid
)
3624 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
3625 l
->min_pid
, l
->max_pid
);
3627 len
+= sprintf(buf
+ len
, " pid=%ld",
3630 if (num_online_cpus() > 1 && !cpus_empty(l
->cpus
) &&
3631 len
< PAGE_SIZE
- 60) {
3632 len
+= sprintf(buf
+ len
, " cpus=");
3633 len
+= cpulist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3637 if (num_online_nodes() > 1 && !nodes_empty(l
->nodes
) &&
3638 len
< PAGE_SIZE
- 60) {
3639 len
+= sprintf(buf
+ len
, " nodes=");
3640 len
+= nodelist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3644 len
+= sprintf(buf
+ len
, "\n");
3649 len
+= sprintf(buf
, "No data\n");
3653 enum slab_stat_type
{
3660 #define SO_FULL (1 << SL_FULL)
3661 #define SO_PARTIAL (1 << SL_PARTIAL)
3662 #define SO_CPU (1 << SL_CPU)
3663 #define SO_OBJECTS (1 << SL_OBJECTS)
3665 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
3666 char *buf
, unsigned long flags
)
3668 unsigned long total
= 0;
3672 unsigned long *nodes
;
3673 unsigned long *per_cpu
;
3675 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
3678 per_cpu
= nodes
+ nr_node_ids
;
3680 for_each_possible_cpu(cpu
) {
3682 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3692 if (flags
& SO_CPU
) {
3693 if (flags
& SO_OBJECTS
)
3704 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3705 struct kmem_cache_node
*n
= get_node(s
, node
);
3707 if (flags
& SO_PARTIAL
) {
3708 if (flags
& SO_OBJECTS
)
3709 x
= count_partial(n
);
3716 if (flags
& SO_FULL
) {
3717 int full_slabs
= atomic_long_read(&n
->nr_slabs
)
3721 if (flags
& SO_OBJECTS
)
3722 x
= full_slabs
* s
->objects
;
3730 x
= sprintf(buf
, "%lu", total
);
3732 for_each_node_state(node
, N_NORMAL_MEMORY
)
3734 x
+= sprintf(buf
+ x
, " N%d=%lu",
3738 return x
+ sprintf(buf
+ x
, "\n");
3741 static int any_slab_objects(struct kmem_cache
*s
)
3746 for_each_possible_cpu(cpu
) {
3747 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3753 for_each_online_node(node
) {
3754 struct kmem_cache_node
*n
= get_node(s
, node
);
3759 if (n
->nr_partial
|| atomic_long_read(&n
->nr_slabs
))
3765 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3766 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3768 struct slab_attribute
{
3769 struct attribute attr
;
3770 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
3771 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
3774 #define SLAB_ATTR_RO(_name) \
3775 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3777 #define SLAB_ATTR(_name) \
3778 static struct slab_attribute _name##_attr = \
3779 __ATTR(_name, 0644, _name##_show, _name##_store)
3781 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
3783 return sprintf(buf
, "%d\n", s
->size
);
3785 SLAB_ATTR_RO(slab_size
);
3787 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3789 return sprintf(buf
, "%d\n", s
->align
);
3791 SLAB_ATTR_RO(align
);
3793 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3795 return sprintf(buf
, "%d\n", s
->objsize
);
3797 SLAB_ATTR_RO(object_size
);
3799 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3801 return sprintf(buf
, "%d\n", s
->objects
);
3803 SLAB_ATTR_RO(objs_per_slab
);
3805 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3807 return sprintf(buf
, "%d\n", s
->order
);
3809 SLAB_ATTR_RO(order
);
3811 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3814 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3816 return n
+ sprintf(buf
+ n
, "\n");
3822 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3824 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3826 SLAB_ATTR_RO(aliases
);
3828 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3830 return show_slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3832 SLAB_ATTR_RO(slabs
);
3834 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3836 return show_slab_objects(s
, buf
, SO_PARTIAL
);
3838 SLAB_ATTR_RO(partial
);
3840 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3842 return show_slab_objects(s
, buf
, SO_CPU
);
3844 SLAB_ATTR_RO(cpu_slabs
);
3846 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3848 return show_slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3850 SLAB_ATTR_RO(objects
);
3852 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3854 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3857 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3858 const char *buf
, size_t length
)
3860 s
->flags
&= ~SLAB_DEBUG_FREE
;
3862 s
->flags
|= SLAB_DEBUG_FREE
;
3865 SLAB_ATTR(sanity_checks
);
3867 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3869 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3872 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3875 s
->flags
&= ~SLAB_TRACE
;
3877 s
->flags
|= SLAB_TRACE
;
3882 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3884 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3887 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3888 const char *buf
, size_t length
)
3890 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3892 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3895 SLAB_ATTR(reclaim_account
);
3897 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3899 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3901 SLAB_ATTR_RO(hwcache_align
);
3903 #ifdef CONFIG_ZONE_DMA
3904 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3906 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3908 SLAB_ATTR_RO(cache_dma
);
3911 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3913 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3915 SLAB_ATTR_RO(destroy_by_rcu
);
3917 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3919 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3922 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3923 const char *buf
, size_t length
)
3925 if (any_slab_objects(s
))
3928 s
->flags
&= ~SLAB_RED_ZONE
;
3930 s
->flags
|= SLAB_RED_ZONE
;
3934 SLAB_ATTR(red_zone
);
3936 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3938 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3941 static ssize_t
poison_store(struct kmem_cache
*s
,
3942 const char *buf
, size_t length
)
3944 if (any_slab_objects(s
))
3947 s
->flags
&= ~SLAB_POISON
;
3949 s
->flags
|= SLAB_POISON
;
3955 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3957 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3960 static ssize_t
store_user_store(struct kmem_cache
*s
,
3961 const char *buf
, size_t length
)
3963 if (any_slab_objects(s
))
3966 s
->flags
&= ~SLAB_STORE_USER
;
3968 s
->flags
|= SLAB_STORE_USER
;
3972 SLAB_ATTR(store_user
);
3974 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3979 static ssize_t
validate_store(struct kmem_cache
*s
,
3980 const char *buf
, size_t length
)
3984 if (buf
[0] == '1') {
3985 ret
= validate_slab_cache(s
);
3991 SLAB_ATTR(validate
);
3993 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3998 static ssize_t
shrink_store(struct kmem_cache
*s
,
3999 const char *buf
, size_t length
)
4001 if (buf
[0] == '1') {
4002 int rc
= kmem_cache_shrink(s
);
4012 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
4014 if (!(s
->flags
& SLAB_STORE_USER
))
4016 return list_locations(s
, buf
, TRACK_ALLOC
);
4018 SLAB_ATTR_RO(alloc_calls
);
4020 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
4022 if (!(s
->flags
& SLAB_STORE_USER
))
4024 return list_locations(s
, buf
, TRACK_FREE
);
4026 SLAB_ATTR_RO(free_calls
);
4029 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
4031 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
4034 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
4035 const char *buf
, size_t length
)
4037 int n
= simple_strtoul(buf
, NULL
, 10);
4040 s
->remote_node_defrag_ratio
= n
* 10;
4043 SLAB_ATTR(remote_node_defrag_ratio
);
4046 #ifdef CONFIG_SLUB_STATS
4047 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
4049 unsigned long sum
= 0;
4052 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
4057 for_each_online_cpu(cpu
) {
4058 unsigned x
= get_cpu_slab(s
, cpu
)->stat
[si
];
4064 len
= sprintf(buf
, "%lu", sum
);
4067 for_each_online_cpu(cpu
) {
4068 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
4069 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
4073 return len
+ sprintf(buf
+ len
, "\n");
4076 #define STAT_ATTR(si, text) \
4077 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4079 return show_stat(s, buf, si); \
4081 SLAB_ATTR_RO(text); \
4083 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4084 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4085 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4086 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4087 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4088 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4089 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4090 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4091 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4092 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4093 STAT_ATTR(FREE_SLAB
, free_slab
);
4094 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4095 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4096 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4097 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4098 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4099 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4103 static struct attribute
*slab_attrs
[] = {
4104 &slab_size_attr
.attr
,
4105 &object_size_attr
.attr
,
4106 &objs_per_slab_attr
.attr
,
4111 &cpu_slabs_attr
.attr
,
4115 &sanity_checks_attr
.attr
,
4117 &hwcache_align_attr
.attr
,
4118 &reclaim_account_attr
.attr
,
4119 &destroy_by_rcu_attr
.attr
,
4120 &red_zone_attr
.attr
,
4122 &store_user_attr
.attr
,
4123 &validate_attr
.attr
,
4125 &alloc_calls_attr
.attr
,
4126 &free_calls_attr
.attr
,
4127 #ifdef CONFIG_ZONE_DMA
4128 &cache_dma_attr
.attr
,
4131 &remote_node_defrag_ratio_attr
.attr
,
4133 #ifdef CONFIG_SLUB_STATS
4134 &alloc_fastpath_attr
.attr
,
4135 &alloc_slowpath_attr
.attr
,
4136 &free_fastpath_attr
.attr
,
4137 &free_slowpath_attr
.attr
,
4138 &free_frozen_attr
.attr
,
4139 &free_add_partial_attr
.attr
,
4140 &free_remove_partial_attr
.attr
,
4141 &alloc_from_partial_attr
.attr
,
4142 &alloc_slab_attr
.attr
,
4143 &alloc_refill_attr
.attr
,
4144 &free_slab_attr
.attr
,
4145 &cpuslab_flush_attr
.attr
,
4146 &deactivate_full_attr
.attr
,
4147 &deactivate_empty_attr
.attr
,
4148 &deactivate_to_head_attr
.attr
,
4149 &deactivate_to_tail_attr
.attr
,
4150 &deactivate_remote_frees_attr
.attr
,
4155 static struct attribute_group slab_attr_group
= {
4156 .attrs
= slab_attrs
,
4159 static ssize_t
slab_attr_show(struct kobject
*kobj
,
4160 struct attribute
*attr
,
4163 struct slab_attribute
*attribute
;
4164 struct kmem_cache
*s
;
4167 attribute
= to_slab_attr(attr
);
4170 if (!attribute
->show
)
4173 err
= attribute
->show(s
, buf
);
4178 static ssize_t
slab_attr_store(struct kobject
*kobj
,
4179 struct attribute
*attr
,
4180 const char *buf
, size_t len
)
4182 struct slab_attribute
*attribute
;
4183 struct kmem_cache
*s
;
4186 attribute
= to_slab_attr(attr
);
4189 if (!attribute
->store
)
4192 err
= attribute
->store(s
, buf
, len
);
4197 static void kmem_cache_release(struct kobject
*kobj
)
4199 struct kmem_cache
*s
= to_slab(kobj
);
4204 static struct sysfs_ops slab_sysfs_ops
= {
4205 .show
= slab_attr_show
,
4206 .store
= slab_attr_store
,
4209 static struct kobj_type slab_ktype
= {
4210 .sysfs_ops
= &slab_sysfs_ops
,
4211 .release
= kmem_cache_release
4214 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
4216 struct kobj_type
*ktype
= get_ktype(kobj
);
4218 if (ktype
== &slab_ktype
)
4223 static struct kset_uevent_ops slab_uevent_ops
= {
4224 .filter
= uevent_filter
,
4227 static struct kset
*slab_kset
;
4229 #define ID_STR_LENGTH 64
4231 /* Create a unique string id for a slab cache:
4233 * Format :[flags-]size
4235 static char *create_unique_id(struct kmem_cache
*s
)
4237 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
4244 * First flags affecting slabcache operations. We will only
4245 * get here for aliasable slabs so we do not need to support
4246 * too many flags. The flags here must cover all flags that
4247 * are matched during merging to guarantee that the id is
4250 if (s
->flags
& SLAB_CACHE_DMA
)
4252 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
4254 if (s
->flags
& SLAB_DEBUG_FREE
)
4258 p
+= sprintf(p
, "%07d", s
->size
);
4259 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
4263 static int sysfs_slab_add(struct kmem_cache
*s
)
4269 if (slab_state
< SYSFS
)
4270 /* Defer until later */
4273 unmergeable
= slab_unmergeable(s
);
4276 * Slabcache can never be merged so we can use the name proper.
4277 * This is typically the case for debug situations. In that
4278 * case we can catch duplicate names easily.
4280 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
4284 * Create a unique name for the slab as a target
4287 name
= create_unique_id(s
);
4290 s
->kobj
.kset
= slab_kset
;
4291 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, name
);
4293 kobject_put(&s
->kobj
);
4297 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
4300 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
4302 /* Setup first alias */
4303 sysfs_slab_alias(s
, s
->name
);
4309 static void sysfs_slab_remove(struct kmem_cache
*s
)
4311 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
4312 kobject_del(&s
->kobj
);
4313 kobject_put(&s
->kobj
);
4317 * Need to buffer aliases during bootup until sysfs becomes
4318 * available lest we loose that information.
4320 struct saved_alias
{
4321 struct kmem_cache
*s
;
4323 struct saved_alias
*next
;
4326 static struct saved_alias
*alias_list
;
4328 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
4330 struct saved_alias
*al
;
4332 if (slab_state
== SYSFS
) {
4334 * If we have a leftover link then remove it.
4336 sysfs_remove_link(&slab_kset
->kobj
, name
);
4337 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
4340 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
4346 al
->next
= alias_list
;
4351 static int __init
slab_sysfs_init(void)
4353 struct kmem_cache
*s
;
4356 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
4358 printk(KERN_ERR
"Cannot register slab subsystem.\n");
4364 list_for_each_entry(s
, &slab_caches
, list
) {
4365 err
= sysfs_slab_add(s
);
4367 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
4368 " to sysfs\n", s
->name
);
4371 while (alias_list
) {
4372 struct saved_alias
*al
= alias_list
;
4374 alias_list
= alias_list
->next
;
4375 err
= sysfs_slab_alias(al
->s
, al
->name
);
4377 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
4378 " %s to sysfs\n", s
->name
);
4386 __initcall(slab_sysfs_init
);
4390 * The /proc/slabinfo ABI
4392 #ifdef CONFIG_SLABINFO
4394 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4395 size_t count
, loff_t
*ppos
)
4401 static void print_slabinfo_header(struct seq_file
*m
)
4403 seq_puts(m
, "slabinfo - version: 2.1\n");
4404 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4405 "<objperslab> <pagesperslab>");
4406 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4407 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4411 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4415 down_read(&slub_lock
);
4417 print_slabinfo_header(m
);
4419 return seq_list_start(&slab_caches
, *pos
);
4422 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4424 return seq_list_next(p
, &slab_caches
, pos
);
4427 static void s_stop(struct seq_file
*m
, void *p
)
4429 up_read(&slub_lock
);
4432 static int s_show(struct seq_file
*m
, void *p
)
4434 unsigned long nr_partials
= 0;
4435 unsigned long nr_slabs
= 0;
4436 unsigned long nr_inuse
= 0;
4437 unsigned long nr_objs
;
4438 struct kmem_cache
*s
;
4441 s
= list_entry(p
, struct kmem_cache
, list
);
4443 for_each_online_node(node
) {
4444 struct kmem_cache_node
*n
= get_node(s
, node
);
4449 nr_partials
+= n
->nr_partial
;
4450 nr_slabs
+= atomic_long_read(&n
->nr_slabs
);
4451 nr_inuse
+= count_partial(n
);
4454 nr_objs
= nr_slabs
* s
->objects
;
4455 nr_inuse
+= (nr_slabs
- nr_partials
) * s
->objects
;
4457 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d", s
->name
, nr_inuse
,
4458 nr_objs
, s
->size
, s
->objects
, (1 << s
->order
));
4459 seq_printf(m
, " : tunables %4u %4u %4u", 0, 0, 0);
4460 seq_printf(m
, " : slabdata %6lu %6lu %6lu", nr_slabs
, nr_slabs
,
4466 const struct seq_operations slabinfo_op
= {
4473 #endif /* CONFIG_SLABINFO */