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>
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
93 * freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
99 * the fast path and disables lockless freelists.
102 #define FROZEN (1 << PG_active)
104 #ifdef CONFIG_SLUB_DEBUG
105 #define SLABDEBUG (1 << PG_error)
110 static inline int SlabFrozen(struct page
*page
)
112 return page
->flags
& FROZEN
;
115 static inline void SetSlabFrozen(struct page
*page
)
117 page
->flags
|= FROZEN
;
120 static inline void ClearSlabFrozen(struct page
*page
)
122 page
->flags
&= ~FROZEN
;
125 static inline int SlabDebug(struct page
*page
)
127 return page
->flags
& SLABDEBUG
;
130 static inline void SetSlabDebug(struct page
*page
)
132 page
->flags
|= SLABDEBUG
;
135 static inline void ClearSlabDebug(struct page
*page
)
137 page
->flags
&= ~SLABDEBUG
;
141 * Issues still to be resolved:
143 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
145 * - Variable sizing of the per node arrays
148 /* Enable to test recovery from slab corruption on boot */
149 #undef SLUB_RESILIENCY_TEST
154 * Small page size. Make sure that we do not fragment memory
156 #define DEFAULT_MAX_ORDER 1
157 #define DEFAULT_MIN_OBJECTS 4
162 * Large page machines are customarily able to handle larger
165 #define DEFAULT_MAX_ORDER 2
166 #define DEFAULT_MIN_OBJECTS 8
171 * Mininum number of partial slabs. These will be left on the partial
172 * lists even if they are empty. kmem_cache_shrink may reclaim them.
174 #define MIN_PARTIAL 2
177 * Maximum number of desirable partial slabs.
178 * The existence of more partial slabs makes kmem_cache_shrink
179 * sort the partial list by the number of objects in the.
181 #define MAX_PARTIAL 10
183 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
184 SLAB_POISON | SLAB_STORE_USER)
187 * Set of flags that will prevent slab merging
189 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
190 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
192 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
195 #ifndef ARCH_KMALLOC_MINALIGN
196 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
199 #ifndef ARCH_SLAB_MINALIGN
200 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 * The page->inuse field is 16 bit thus we have this limitation
206 #define MAX_OBJECTS_PER_SLAB 65535
208 /* Internal SLUB flags */
209 #define __OBJECT_POISON 0x80000000 /* Poison object */
210 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
212 /* Not all arches define cache_line_size */
213 #ifndef cache_line_size
214 #define cache_line_size() L1_CACHE_BYTES
217 static int kmem_size
= sizeof(struct kmem_cache
);
220 static struct notifier_block slab_notifier
;
224 DOWN
, /* No slab functionality available */
225 PARTIAL
, /* kmem_cache_open() works but kmalloc does not */
226 UP
, /* Everything works but does not show up in sysfs */
230 /* A list of all slab caches on the system */
231 static DECLARE_RWSEM(slub_lock
);
232 static LIST_HEAD(slab_caches
);
235 * Tracking user of a slab.
238 void *addr
; /* Called from address */
239 int cpu
; /* Was running on cpu */
240 int pid
; /* Pid context */
241 unsigned long when
; /* When did the operation occur */
244 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
246 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
247 static int sysfs_slab_add(struct kmem_cache
*);
248 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
249 static void sysfs_slab_remove(struct kmem_cache
*);
251 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
252 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
254 static inline void sysfs_slab_remove(struct kmem_cache
*s
) {}
257 /********************************************************************
258 * Core slab cache functions
259 *******************************************************************/
261 int slab_is_available(void)
263 return slab_state
>= UP
;
266 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
269 return s
->node
[node
];
271 return &s
->local_node
;
275 static inline struct kmem_cache_cpu
*get_cpu_slab(struct kmem_cache
*s
, int cpu
)
277 return &s
->cpu_slab
[cpu
];
280 static inline int check_valid_pointer(struct kmem_cache
*s
,
281 struct page
*page
, const void *object
)
288 base
= page_address(page
);
289 if (object
< base
|| object
>= base
+ s
->objects
* s
->size
||
290 (object
- base
) % s
->size
) {
298 * Slow version of get and set free pointer.
300 * This version requires touching the cache lines of kmem_cache which
301 * we avoid to do in the fast alloc free paths. There we obtain the offset
302 * from the page struct.
304 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
306 return *(void **)(object
+ s
->offset
);
309 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
311 *(void **)(object
+ s
->offset
) = fp
;
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr) \
316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
320 #define for_each_free_object(__p, __s, __free) \
321 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
323 /* Determine object index from a given position */
324 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
326 return (p
- addr
) / s
->size
;
329 #ifdef CONFIG_SLUB_DEBUG
333 #ifdef CONFIG_SLUB_DEBUG_ON
334 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
336 static int slub_debug
;
339 static char *slub_debug_slabs
;
344 static void print_section(char *text
, u8
*addr
, unsigned int length
)
352 for (i
= 0; i
< length
; i
++) {
354 printk(KERN_ERR
"%8s 0x%p: ", text
, addr
+ i
);
357 printk(" %02x", addr
[i
]);
359 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
361 printk(" %s\n",ascii
);
372 printk(" %s\n", ascii
);
376 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
377 enum track_item alloc
)
382 p
= object
+ s
->offset
+ sizeof(void *);
384 p
= object
+ s
->inuse
;
389 static void set_track(struct kmem_cache
*s
, void *object
,
390 enum track_item alloc
, void *addr
)
395 p
= object
+ s
->offset
+ sizeof(void *);
397 p
= object
+ s
->inuse
;
402 p
->cpu
= smp_processor_id();
403 p
->pid
= current
? current
->pid
: -1;
406 memset(p
, 0, sizeof(struct track
));
409 static void init_tracking(struct kmem_cache
*s
, void *object
)
411 if (!(s
->flags
& SLAB_STORE_USER
))
414 set_track(s
, object
, TRACK_FREE
, NULL
);
415 set_track(s
, object
, TRACK_ALLOC
, NULL
);
418 static void print_track(const char *s
, struct track
*t
)
423 printk(KERN_ERR
"INFO: %s in ", s
);
424 __print_symbol("%s", (unsigned long)t
->addr
);
425 printk(" age=%lu cpu=%u pid=%d\n", jiffies
- t
->when
, t
->cpu
, t
->pid
);
428 static void print_tracking(struct kmem_cache
*s
, void *object
)
430 if (!(s
->flags
& SLAB_STORE_USER
))
433 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
434 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
437 static void print_page_info(struct page
*page
)
439 printk(KERN_ERR
"INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
440 page
, page
->inuse
, page
->freelist
, page
->flags
);
444 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
450 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
452 printk(KERN_ERR
"========================================"
453 "=====================================\n");
454 printk(KERN_ERR
"BUG %s: %s\n", s
->name
, buf
);
455 printk(KERN_ERR
"----------------------------------------"
456 "-------------------------------------\n\n");
459 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
465 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
467 printk(KERN_ERR
"FIX %s: %s\n", s
->name
, buf
);
470 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
472 unsigned int off
; /* Offset of last byte */
473 u8
*addr
= page_address(page
);
475 print_tracking(s
, p
);
477 print_page_info(page
);
479 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
480 p
, p
- addr
, get_freepointer(s
, p
));
483 print_section("Bytes b4", p
- 16, 16);
485 print_section("Object", p
, min(s
->objsize
, 128));
487 if (s
->flags
& SLAB_RED_ZONE
)
488 print_section("Redzone", p
+ s
->objsize
,
489 s
->inuse
- s
->objsize
);
492 off
= s
->offset
+ sizeof(void *);
496 if (s
->flags
& SLAB_STORE_USER
)
497 off
+= 2 * sizeof(struct track
);
500 /* Beginning of the filler is the free pointer */
501 print_section("Padding", p
+ off
, s
->size
- off
);
506 static void object_err(struct kmem_cache
*s
, struct page
*page
,
507 u8
*object
, char *reason
)
510 print_trailer(s
, page
, object
);
513 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *fmt
, ...)
519 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
522 print_page_info(page
);
526 static void init_object(struct kmem_cache
*s
, void *object
, int active
)
530 if (s
->flags
& __OBJECT_POISON
) {
531 memset(p
, POISON_FREE
, s
->objsize
- 1);
532 p
[s
->objsize
-1] = POISON_END
;
535 if (s
->flags
& SLAB_RED_ZONE
)
536 memset(p
+ s
->objsize
,
537 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
,
538 s
->inuse
- s
->objsize
);
541 static u8
*check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
544 if (*start
!= (u8
)value
)
552 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
553 void *from
, void *to
)
555 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
556 memset(from
, data
, to
- from
);
559 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
560 u8
*object
, char *what
,
561 u8
* start
, unsigned int value
, unsigned int bytes
)
566 fault
= check_bytes(start
, value
, bytes
);
571 while (end
> fault
&& end
[-1] == value
)
574 slab_bug(s
, "%s overwritten", what
);
575 printk(KERN_ERR
"INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
576 fault
, end
- 1, fault
[0], value
);
577 print_trailer(s
, page
, object
);
579 restore_bytes(s
, what
, value
, fault
, end
);
587 * Bytes of the object to be managed.
588 * If the freepointer may overlay the object then the free
589 * pointer is the first word of the object.
591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
594 * object + s->objsize
595 * Padding to reach word boundary. This is also used for Redzoning.
596 * Padding is extended by another word if Redzoning is enabled and
599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
600 * 0xcc (RED_ACTIVE) for objects in use.
603 * Meta data starts here.
605 * A. Free pointer (if we cannot overwrite object on free)
606 * B. Tracking data for SLAB_STORE_USER
607 * C. Padding to reach required alignment boundary or at mininum
608 * one word if debuggin is on to be able to detect writes
609 * before the word boundary.
611 * Padding is done using 0x5a (POISON_INUSE)
614 * Nothing is used beyond s->size.
616 * If slabcaches are merged then the objsize and inuse boundaries are mostly
617 * ignored. And therefore no slab options that rely on these boundaries
618 * may be used with merged slabcaches.
621 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
623 unsigned long off
= s
->inuse
; /* The end of info */
626 /* Freepointer is placed after the object. */
627 off
+= sizeof(void *);
629 if (s
->flags
& SLAB_STORE_USER
)
630 /* We also have user information there */
631 off
+= 2 * sizeof(struct track
);
636 return check_bytes_and_report(s
, page
, p
, "Object padding",
637 p
+ off
, POISON_INUSE
, s
->size
- off
);
640 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
648 if (!(s
->flags
& SLAB_POISON
))
651 start
= page_address(page
);
652 end
= start
+ (PAGE_SIZE
<< s
->order
);
653 length
= s
->objects
* s
->size
;
654 remainder
= end
- (start
+ length
);
658 fault
= check_bytes(start
+ length
, POISON_INUSE
, remainder
);
661 while (end
> fault
&& end
[-1] == POISON_INUSE
)
664 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
665 print_section("Padding", start
, length
);
667 restore_bytes(s
, "slab padding", POISON_INUSE
, start
, end
);
671 static int check_object(struct kmem_cache
*s
, struct page
*page
,
672 void *object
, int active
)
675 u8
*endobject
= object
+ s
->objsize
;
677 if (s
->flags
& SLAB_RED_ZONE
) {
679 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
681 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
682 endobject
, red
, s
->inuse
- s
->objsize
))
685 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
)
686 check_bytes_and_report(s
, page
, p
, "Alignment padding", endobject
,
687 POISON_INUSE
, s
->inuse
- s
->objsize
);
690 if (s
->flags
& SLAB_POISON
) {
691 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
692 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
693 POISON_FREE
, s
->objsize
- 1) ||
694 !check_bytes_and_report(s
, page
, p
, "Poison",
695 p
+ s
->objsize
-1, POISON_END
, 1)))
698 * check_pad_bytes cleans up on its own.
700 check_pad_bytes(s
, page
, p
);
703 if (!s
->offset
&& active
)
705 * Object and freepointer overlap. Cannot check
706 * freepointer while object is allocated.
710 /* Check free pointer validity */
711 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
712 object_err(s
, page
, p
, "Freepointer corrupt");
714 * No choice but to zap it and thus loose the remainder
715 * of the free objects in this slab. May cause
716 * another error because the object count is now wrong.
718 set_freepointer(s
, p
, NULL
);
724 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
726 VM_BUG_ON(!irqs_disabled());
728 if (!PageSlab(page
)) {
729 slab_err(s
, page
, "Not a valid slab page");
732 if (page
->offset
* sizeof(void *) != s
->offset
) {
733 slab_err(s
, page
, "Corrupted offset %lu",
734 (unsigned long)(page
->offset
* sizeof(void *)));
737 if (page
->inuse
> s
->objects
) {
738 slab_err(s
, page
, "inuse %u > max %u",
739 s
->name
, page
->inuse
, s
->objects
);
742 /* Slab_pad_check fixes things up after itself */
743 slab_pad_check(s
, page
);
748 * Determine if a certain object on a page is on the freelist. Must hold the
749 * slab lock to guarantee that the chains are in a consistent state.
751 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
754 void *fp
= page
->freelist
;
757 while (fp
&& nr
<= s
->objects
) {
760 if (!check_valid_pointer(s
, page
, fp
)) {
762 object_err(s
, page
, object
,
763 "Freechain corrupt");
764 set_freepointer(s
, object
, NULL
);
767 slab_err(s
, page
, "Freepointer corrupt");
768 page
->freelist
= NULL
;
769 page
->inuse
= s
->objects
;
770 slab_fix(s
, "Freelist cleared");
776 fp
= get_freepointer(s
, object
);
780 if (page
->inuse
!= s
->objects
- nr
) {
781 slab_err(s
, page
, "Wrong object count. Counter is %d but "
782 "counted were %d", page
->inuse
, s
->objects
- nr
);
783 page
->inuse
= s
->objects
- nr
;
784 slab_fix(s
, "Object count adjusted.");
786 return search
== NULL
;
789 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
, int alloc
)
791 if (s
->flags
& SLAB_TRACE
) {
792 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
794 alloc
? "alloc" : "free",
799 print_section("Object", (void *)object
, s
->objsize
);
806 * Tracking of fully allocated slabs for debugging purposes.
808 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
810 spin_lock(&n
->list_lock
);
811 list_add(&page
->lru
, &n
->full
);
812 spin_unlock(&n
->list_lock
);
815 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
817 struct kmem_cache_node
*n
;
819 if (!(s
->flags
& SLAB_STORE_USER
))
822 n
= get_node(s
, page_to_nid(page
));
824 spin_lock(&n
->list_lock
);
825 list_del(&page
->lru
);
826 spin_unlock(&n
->list_lock
);
829 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
832 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
835 init_object(s
, object
, 0);
836 init_tracking(s
, object
);
839 static int alloc_debug_processing(struct kmem_cache
*s
, struct page
*page
,
840 void *object
, void *addr
)
842 if (!check_slab(s
, page
))
845 if (object
&& !on_freelist(s
, page
, object
)) {
846 object_err(s
, page
, object
, "Object already allocated");
850 if (!check_valid_pointer(s
, page
, object
)) {
851 object_err(s
, page
, object
, "Freelist Pointer check fails");
855 if (object
&& !check_object(s
, page
, object
, 0))
858 /* Success perform special debug activities for allocs */
859 if (s
->flags
& SLAB_STORE_USER
)
860 set_track(s
, object
, TRACK_ALLOC
, addr
);
861 trace(s
, page
, object
, 1);
862 init_object(s
, object
, 1);
866 if (PageSlab(page
)) {
868 * If this is a slab page then lets do the best we can
869 * to avoid issues in the future. Marking all objects
870 * as used avoids touching the remaining objects.
872 slab_fix(s
, "Marking all objects used");
873 page
->inuse
= s
->objects
;
874 page
->freelist
= NULL
;
875 /* Fix up fields that may be corrupted */
876 page
->offset
= s
->offset
/ sizeof(void *);
881 static int free_debug_processing(struct kmem_cache
*s
, struct page
*page
,
882 void *object
, void *addr
)
884 if (!check_slab(s
, page
))
887 if (!check_valid_pointer(s
, page
, object
)) {
888 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
892 if (on_freelist(s
, page
, object
)) {
893 object_err(s
, page
, object
, "Object already free");
897 if (!check_object(s
, page
, object
, 1))
900 if (unlikely(s
!= page
->slab
)) {
902 slab_err(s
, page
, "Attempt to free object(0x%p) "
903 "outside of slab", object
);
907 "SLUB <none>: no slab for object 0x%p.\n",
912 object_err(s
, page
, object
,
913 "page slab pointer corrupt.");
917 /* Special debug activities for freeing objects */
918 if (!SlabFrozen(page
) && !page
->freelist
)
919 remove_full(s
, page
);
920 if (s
->flags
& SLAB_STORE_USER
)
921 set_track(s
, object
, TRACK_FREE
, addr
);
922 trace(s
, page
, object
, 0);
923 init_object(s
, object
, 0);
927 slab_fix(s
, "Object at 0x%p not freed", object
);
931 static int __init
setup_slub_debug(char *str
)
933 slub_debug
= DEBUG_DEFAULT_FLAGS
;
934 if (*str
++ != '=' || !*str
)
936 * No options specified. Switch on full debugging.
942 * No options but restriction on slabs. This means full
943 * debugging for slabs matching a pattern.
950 * Switch off all debugging measures.
955 * Determine which debug features should be switched on
957 for ( ;*str
&& *str
!= ','; str
++) {
958 switch (tolower(*str
)) {
960 slub_debug
|= SLAB_DEBUG_FREE
;
963 slub_debug
|= SLAB_RED_ZONE
;
966 slub_debug
|= SLAB_POISON
;
969 slub_debug
|= SLAB_STORE_USER
;
972 slub_debug
|= SLAB_TRACE
;
975 printk(KERN_ERR
"slub_debug option '%c' "
976 "unknown. skipped\n",*str
);
982 slub_debug_slabs
= str
+ 1;
987 __setup("slub_debug", setup_slub_debug
);
989 static unsigned long kmem_cache_flags(unsigned long objsize
,
990 unsigned long flags
, const char *name
,
991 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
994 * The page->offset field is only 16 bit wide. This is an offset
995 * in units of words from the beginning of an object. If the slab
996 * size is bigger then we cannot move the free pointer behind the
999 * On 32 bit platforms the limit is 256k. On 64bit platforms
1000 * the limit is 512k.
1002 * Debugging or ctor may create a need to move the free
1003 * pointer. Fail if this happens.
1005 if (objsize
>= 65535 * sizeof(void *)) {
1006 BUG_ON(flags
& (SLAB_RED_ZONE
| SLAB_POISON
|
1007 SLAB_STORE_USER
| SLAB_DESTROY_BY_RCU
));
1011 * Enable debugging if selected on the kernel commandline.
1013 if (slub_debug
&& (!slub_debug_slabs
||
1014 strncmp(slub_debug_slabs
, name
,
1015 strlen(slub_debug_slabs
)) == 0))
1016 flags
|= slub_debug
;
1022 static inline void setup_object_debug(struct kmem_cache
*s
,
1023 struct page
*page
, void *object
) {}
1025 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1026 struct page
*page
, void *object
, void *addr
) { return 0; }
1028 static inline int free_debug_processing(struct kmem_cache
*s
,
1029 struct page
*page
, void *object
, void *addr
) { return 0; }
1031 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1033 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1034 void *object
, int active
) { return 1; }
1035 static inline void add_full(struct kmem_cache_node
*n
, struct page
*page
) {}
1036 static inline unsigned long kmem_cache_flags(unsigned long objsize
,
1037 unsigned long flags
, const char *name
,
1038 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
1042 #define slub_debug 0
1045 * Slab allocation and freeing
1047 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1050 int pages
= 1 << s
->order
;
1053 flags
|= __GFP_COMP
;
1055 if (s
->flags
& SLAB_CACHE_DMA
)
1058 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
1059 flags
|= __GFP_RECLAIMABLE
;
1062 page
= alloc_pages(flags
, s
->order
);
1064 page
= alloc_pages_node(node
, flags
, s
->order
);
1069 mod_zone_page_state(page_zone(page
),
1070 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1071 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1077 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1080 setup_object_debug(s
, page
, object
);
1081 if (unlikely(s
->ctor
))
1082 s
->ctor(object
, s
, 0);
1085 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1088 struct kmem_cache_node
*n
;
1094 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1096 if (flags
& __GFP_WAIT
)
1099 page
= allocate_slab(s
,
1100 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1104 n
= get_node(s
, page_to_nid(page
));
1106 atomic_long_inc(&n
->nr_slabs
);
1107 page
->offset
= s
->offset
/ sizeof(void *);
1109 page
->flags
|= 1 << PG_slab
;
1110 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
1111 SLAB_STORE_USER
| SLAB_TRACE
))
1114 start
= page_address(page
);
1115 end
= start
+ s
->objects
* s
->size
;
1117 if (unlikely(s
->flags
& SLAB_POISON
))
1118 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
1121 for_each_object(p
, s
, start
) {
1122 setup_object(s
, page
, last
);
1123 set_freepointer(s
, last
, p
);
1126 setup_object(s
, page
, last
);
1127 set_freepointer(s
, last
, NULL
);
1129 page
->freelist
= start
;
1132 if (flags
& __GFP_WAIT
)
1133 local_irq_disable();
1137 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1139 int pages
= 1 << s
->order
;
1141 if (unlikely(SlabDebug(page
))) {
1144 slab_pad_check(s
, page
);
1145 for_each_object(p
, s
, page_address(page
))
1146 check_object(s
, page
, p
, 0);
1147 ClearSlabDebug(page
);
1150 mod_zone_page_state(page_zone(page
),
1151 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1152 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1155 __free_pages(page
, s
->order
);
1158 static void rcu_free_slab(struct rcu_head
*h
)
1162 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1163 __free_slab(page
->slab
, page
);
1166 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1168 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1170 * RCU free overloads the RCU head over the LRU
1172 struct rcu_head
*head
= (void *)&page
->lru
;
1174 call_rcu(head
, rcu_free_slab
);
1176 __free_slab(s
, page
);
1179 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1181 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1183 atomic_long_dec(&n
->nr_slabs
);
1184 reset_page_mapcount(page
);
1185 __ClearPageSlab(page
);
1190 * Per slab locking using the pagelock
1192 static __always_inline
void slab_lock(struct page
*page
)
1194 bit_spin_lock(PG_locked
, &page
->flags
);
1197 static __always_inline
void slab_unlock(struct page
*page
)
1199 bit_spin_unlock(PG_locked
, &page
->flags
);
1202 static __always_inline
int slab_trylock(struct page
*page
)
1206 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
1211 * Management of partially allocated slabs
1213 static void add_partial_tail(struct kmem_cache_node
*n
, struct page
*page
)
1215 spin_lock(&n
->list_lock
);
1217 list_add_tail(&page
->lru
, &n
->partial
);
1218 spin_unlock(&n
->list_lock
);
1221 static void add_partial(struct kmem_cache_node
*n
, struct page
*page
)
1223 spin_lock(&n
->list_lock
);
1225 list_add(&page
->lru
, &n
->partial
);
1226 spin_unlock(&n
->list_lock
);
1229 static void remove_partial(struct kmem_cache
*s
,
1232 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1234 spin_lock(&n
->list_lock
);
1235 list_del(&page
->lru
);
1237 spin_unlock(&n
->list_lock
);
1241 * Lock slab and remove from the partial list.
1243 * Must hold list_lock.
1245 static inline int lock_and_freeze_slab(struct kmem_cache_node
*n
, struct page
*page
)
1247 if (slab_trylock(page
)) {
1248 list_del(&page
->lru
);
1250 SetSlabFrozen(page
);
1257 * Try to allocate a partial slab from a specific node.
1259 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
1264 * Racy check. If we mistakenly see no partial slabs then we
1265 * just allocate an empty slab. If we mistakenly try to get a
1266 * partial slab and there is none available then get_partials()
1269 if (!n
|| !n
->nr_partial
)
1272 spin_lock(&n
->list_lock
);
1273 list_for_each_entry(page
, &n
->partial
, lru
)
1274 if (lock_and_freeze_slab(n
, page
))
1278 spin_unlock(&n
->list_lock
);
1283 * Get a page from somewhere. Search in increasing NUMA distances.
1285 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1288 struct zonelist
*zonelist
;
1293 * The defrag ratio allows a configuration of the tradeoffs between
1294 * inter node defragmentation and node local allocations. A lower
1295 * defrag_ratio increases the tendency to do local allocations
1296 * instead of attempting to obtain partial slabs from other nodes.
1298 * If the defrag_ratio is set to 0 then kmalloc() always
1299 * returns node local objects. If the ratio is higher then kmalloc()
1300 * may return off node objects because partial slabs are obtained
1301 * from other nodes and filled up.
1303 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1304 * defrag_ratio = 1000) then every (well almost) allocation will
1305 * first attempt to defrag slab caches on other nodes. This means
1306 * scanning over all nodes to look for partial slabs which may be
1307 * expensive if we do it every time we are trying to find a slab
1308 * with available objects.
1310 if (!s
->defrag_ratio
|| get_cycles() % 1024 > s
->defrag_ratio
)
1313 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
1314 ->node_zonelists
[gfp_zone(flags
)];
1315 for (z
= zonelist
->zones
; *z
; z
++) {
1316 struct kmem_cache_node
*n
;
1318 n
= get_node(s
, zone_to_nid(*z
));
1320 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1321 n
->nr_partial
> MIN_PARTIAL
) {
1322 page
= get_partial_node(n
);
1332 * Get a partial page, lock it and return it.
1334 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1337 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1339 page
= get_partial_node(get_node(s
, searchnode
));
1340 if (page
|| (flags
& __GFP_THISNODE
))
1343 return get_any_partial(s
, flags
);
1347 * Move a page back to the lists.
1349 * Must be called with the slab lock held.
1351 * On exit the slab lock will have been dropped.
1353 static void unfreeze_slab(struct kmem_cache
*s
, struct page
*page
)
1355 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1357 ClearSlabFrozen(page
);
1361 add_partial(n
, page
);
1362 else if (SlabDebug(page
) && (s
->flags
& SLAB_STORE_USER
))
1367 if (n
->nr_partial
< MIN_PARTIAL
) {
1369 * Adding an empty slab to the partial slabs in order
1370 * to avoid page allocator overhead. This slab needs
1371 * to come after the other slabs with objects in
1372 * order to fill them up. That way the size of the
1373 * partial list stays small. kmem_cache_shrink can
1374 * reclaim empty slabs from the partial list.
1376 add_partial_tail(n
, page
);
1380 discard_slab(s
, page
);
1386 * Remove the cpu slab
1388 static void deactivate_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1390 struct page
*page
= c
->page
;
1392 * Merge cpu freelist into freelist. Typically we get here
1393 * because both freelists are empty. So this is unlikely
1396 while (unlikely(c
->freelist
)) {
1399 /* Retrieve object from cpu_freelist */
1400 object
= c
->freelist
;
1401 c
->freelist
= c
->freelist
[page
->offset
];
1403 /* And put onto the regular freelist */
1404 object
[page
->offset
] = page
->freelist
;
1405 page
->freelist
= object
;
1409 unfreeze_slab(s
, page
);
1412 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1415 deactivate_slab(s
, c
);
1420 * Called from IPI handler with interrupts disabled.
1422 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1424 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
1426 if (likely(c
&& c
->page
))
1430 static void flush_cpu_slab(void *d
)
1432 struct kmem_cache
*s
= d
;
1434 __flush_cpu_slab(s
, smp_processor_id());
1437 static void flush_all(struct kmem_cache
*s
)
1440 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1442 unsigned long flags
;
1444 local_irq_save(flags
);
1446 local_irq_restore(flags
);
1451 * Check if the objects in a per cpu structure fit numa
1452 * locality expectations.
1454 static inline int node_match(struct kmem_cache_cpu
*c
, int node
)
1457 if (node
!= -1 && c
->node
!= node
)
1464 * Slow path. The lockless freelist is empty or we need to perform
1467 * Interrupts are disabled.
1469 * Processing is still very fast if new objects have been freed to the
1470 * regular freelist. In that case we simply take over the regular freelist
1471 * as the lockless freelist and zap the regular freelist.
1473 * If that is not working then we fall back to the partial lists. We take the
1474 * first element of the freelist as the object to allocate now and move the
1475 * rest of the freelist to the lockless freelist.
1477 * And if we were unable to get a new slab from the partial slab lists then
1478 * we need to allocate a new slab. This is slowest path since we may sleep.
1480 static void *__slab_alloc(struct kmem_cache
*s
,
1481 gfp_t gfpflags
, int node
, void *addr
, struct kmem_cache_cpu
*c
)
1490 if (unlikely(!node_match(c
, node
)))
1493 object
= c
->page
->freelist
;
1494 if (unlikely(!object
))
1496 if (unlikely(SlabDebug(c
->page
)))
1499 object
= c
->page
->freelist
;
1500 c
->freelist
= object
[c
->page
->offset
];
1501 c
->page
->inuse
= s
->objects
;
1502 c
->page
->freelist
= NULL
;
1503 c
->node
= page_to_nid(c
->page
);
1504 slab_unlock(c
->page
);
1508 deactivate_slab(s
, c
);
1511 new = get_partial(s
, gfpflags
, node
);
1517 new = new_slab(s
, gfpflags
, node
);
1519 c
= get_cpu_slab(s
, smp_processor_id());
1522 * Someone else populated the cpu_slab while we
1523 * enabled interrupts, or we have gotten scheduled
1524 * on another cpu. The page may not be on the
1525 * requested node even if __GFP_THISNODE was
1526 * specified. So we need to recheck.
1528 if (node_match(c
, node
)) {
1530 * Current cpuslab is acceptable and we
1531 * want the current one since its cache hot
1533 discard_slab(s
, new);
1537 /* New slab does not fit our expectations */
1547 object
= c
->page
->freelist
;
1548 if (!alloc_debug_processing(s
, c
->page
, object
, addr
))
1552 c
->page
->freelist
= object
[c
->page
->offset
];
1553 slab_unlock(c
->page
);
1558 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1559 * have the fastpath folded into their functions. So no function call
1560 * overhead for requests that can be satisfied on the fastpath.
1562 * The fastpath works by first checking if the lockless freelist can be used.
1563 * If not then __slab_alloc is called for slow processing.
1565 * Otherwise we can simply pick the next object from the lockless free list.
1567 static void __always_inline
*slab_alloc(struct kmem_cache
*s
,
1568 gfp_t gfpflags
, int node
, void *addr
)
1571 unsigned long flags
;
1572 struct kmem_cache_cpu
*c
;
1574 local_irq_save(flags
);
1575 c
= get_cpu_slab(s
, smp_processor_id());
1576 if (unlikely(!c
->page
|| !c
->freelist
||
1577 !node_match(c
, node
)))
1579 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
1582 object
= c
->freelist
;
1583 c
->freelist
= object
[c
->page
->offset
];
1585 local_irq_restore(flags
);
1587 if (unlikely((gfpflags
& __GFP_ZERO
) && object
))
1588 memset(object
, 0, s
->objsize
);
1593 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1595 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1597 EXPORT_SYMBOL(kmem_cache_alloc
);
1600 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1602 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1604 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1608 * Slow patch handling. This may still be called frequently since objects
1609 * have a longer lifetime than the cpu slabs in most processing loads.
1611 * So we still attempt to reduce cache line usage. Just take the slab
1612 * lock and free the item. If there is no additional partial page
1613 * handling required then we can return immediately.
1615 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
1616 void *x
, void *addr
)
1619 void **object
= (void *)x
;
1623 if (unlikely(SlabDebug(page
)))
1626 prior
= object
[page
->offset
] = page
->freelist
;
1627 page
->freelist
= object
;
1630 if (unlikely(SlabFrozen(page
)))
1633 if (unlikely(!page
->inuse
))
1637 * Objects left in the slab. If it
1638 * was not on the partial list before
1641 if (unlikely(!prior
))
1642 add_partial(get_node(s
, page_to_nid(page
)), page
);
1651 * Slab still on the partial list.
1653 remove_partial(s
, page
);
1656 discard_slab(s
, page
);
1660 if (!free_debug_processing(s
, page
, x
, addr
))
1666 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1667 * can perform fastpath freeing without additional function calls.
1669 * The fastpath is only possible if we are freeing to the current cpu slab
1670 * of this processor. This typically the case if we have just allocated
1673 * If fastpath is not possible then fall back to __slab_free where we deal
1674 * with all sorts of special processing.
1676 static void __always_inline
slab_free(struct kmem_cache
*s
,
1677 struct page
*page
, void *x
, void *addr
)
1679 void **object
= (void *)x
;
1680 unsigned long flags
;
1681 struct kmem_cache_cpu
*c
;
1683 local_irq_save(flags
);
1684 debug_check_no_locks_freed(object
, s
->objsize
);
1685 c
= get_cpu_slab(s
, smp_processor_id());
1686 if (likely(page
== c
->page
&& !SlabDebug(page
))) {
1687 object
[page
->offset
] = c
->freelist
;
1688 c
->freelist
= object
;
1690 __slab_free(s
, page
, x
, addr
);
1692 local_irq_restore(flags
);
1695 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1699 page
= virt_to_head_page(x
);
1701 slab_free(s
, page
, x
, __builtin_return_address(0));
1703 EXPORT_SYMBOL(kmem_cache_free
);
1705 /* Figure out on which slab object the object resides */
1706 static struct page
*get_object_page(const void *x
)
1708 struct page
*page
= virt_to_head_page(x
);
1710 if (!PageSlab(page
))
1717 * Object placement in a slab is made very easy because we always start at
1718 * offset 0. If we tune the size of the object to the alignment then we can
1719 * get the required alignment by putting one properly sized object after
1722 * Notice that the allocation order determines the sizes of the per cpu
1723 * caches. Each processor has always one slab available for allocations.
1724 * Increasing the allocation order reduces the number of times that slabs
1725 * must be moved on and off the partial lists and is therefore a factor in
1730 * Mininum / Maximum order of slab pages. This influences locking overhead
1731 * and slab fragmentation. A higher order reduces the number of partial slabs
1732 * and increases the number of allocations possible without having to
1733 * take the list_lock.
1735 static int slub_min_order
;
1736 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1737 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1740 * Merge control. If this is set then no merging of slab caches will occur.
1741 * (Could be removed. This was introduced to pacify the merge skeptics.)
1743 static int slub_nomerge
;
1746 * Calculate the order of allocation given an slab object size.
1748 * The order of allocation has significant impact on performance and other
1749 * system components. Generally order 0 allocations should be preferred since
1750 * order 0 does not cause fragmentation in the page allocator. Larger objects
1751 * be problematic to put into order 0 slabs because there may be too much
1752 * unused space left. We go to a higher order if more than 1/8th of the slab
1755 * In order to reach satisfactory performance we must ensure that a minimum
1756 * number of objects is in one slab. Otherwise we may generate too much
1757 * activity on the partial lists which requires taking the list_lock. This is
1758 * less a concern for large slabs though which are rarely used.
1760 * slub_max_order specifies the order where we begin to stop considering the
1761 * number of objects in a slab as critical. If we reach slub_max_order then
1762 * we try to keep the page order as low as possible. So we accept more waste
1763 * of space in favor of a small page order.
1765 * Higher order allocations also allow the placement of more objects in a
1766 * slab and thereby reduce object handling overhead. If the user has
1767 * requested a higher mininum order then we start with that one instead of
1768 * the smallest order which will fit the object.
1770 static inline int slab_order(int size
, int min_objects
,
1771 int max_order
, int fract_leftover
)
1775 int min_order
= slub_min_order
;
1778 * If we would create too many object per slab then reduce
1779 * the slab order even if it goes below slub_min_order.
1781 while (min_order
> 0 &&
1782 (PAGE_SIZE
<< min_order
) >= MAX_OBJECTS_PER_SLAB
* size
)
1785 for (order
= max(min_order
,
1786 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
1787 order
<= max_order
; order
++) {
1789 unsigned long slab_size
= PAGE_SIZE
<< order
;
1791 if (slab_size
< min_objects
* size
)
1794 rem
= slab_size
% size
;
1796 if (rem
<= slab_size
/ fract_leftover
)
1799 /* If the next size is too high then exit now */
1800 if (slab_size
* 2 >= MAX_OBJECTS_PER_SLAB
* size
)
1807 static inline int calculate_order(int size
)
1814 * Attempt to find best configuration for a slab. This
1815 * works by first attempting to generate a layout with
1816 * the best configuration and backing off gradually.
1818 * First we reduce the acceptable waste in a slab. Then
1819 * we reduce the minimum objects required in a slab.
1821 min_objects
= slub_min_objects
;
1822 while (min_objects
> 1) {
1824 while (fraction
>= 4) {
1825 order
= slab_order(size
, min_objects
,
1826 slub_max_order
, fraction
);
1827 if (order
<= slub_max_order
)
1835 * We were unable to place multiple objects in a slab. Now
1836 * lets see if we can place a single object there.
1838 order
= slab_order(size
, 1, slub_max_order
, 1);
1839 if (order
<= slub_max_order
)
1843 * Doh this slab cannot be placed using slub_max_order.
1845 order
= slab_order(size
, 1, MAX_ORDER
, 1);
1846 if (order
<= MAX_ORDER
)
1852 * Figure out what the alignment of the objects will be.
1854 static unsigned long calculate_alignment(unsigned long flags
,
1855 unsigned long align
, unsigned long size
)
1858 * If the user wants hardware cache aligned objects then
1859 * follow that suggestion if the object is sufficiently
1862 * The hardware cache alignment cannot override the
1863 * specified alignment though. If that is greater
1866 if ((flags
& SLAB_HWCACHE_ALIGN
) &&
1867 size
> cache_line_size() / 2)
1868 return max_t(unsigned long, align
, cache_line_size());
1870 if (align
< ARCH_SLAB_MINALIGN
)
1871 return ARCH_SLAB_MINALIGN
;
1873 return ALIGN(align
, sizeof(void *));
1876 static void init_kmem_cache_cpu(struct kmem_cache
*s
,
1877 struct kmem_cache_cpu
*c
)
1884 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
, gfp_t flags
)
1888 for_each_possible_cpu(cpu
)
1889 init_kmem_cache_cpu(s
, get_cpu_slab(s
, cpu
));
1894 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1897 atomic_long_set(&n
->nr_slabs
, 0);
1898 spin_lock_init(&n
->list_lock
);
1899 INIT_LIST_HEAD(&n
->partial
);
1900 #ifdef CONFIG_SLUB_DEBUG
1901 INIT_LIST_HEAD(&n
->full
);
1907 * No kmalloc_node yet so do it by hand. We know that this is the first
1908 * slab on the node for this slabcache. There are no concurrent accesses
1911 * Note that this function only works on the kmalloc_node_cache
1912 * when allocating for the kmalloc_node_cache.
1914 static struct kmem_cache_node
*early_kmem_cache_node_alloc(gfp_t gfpflags
,
1918 struct kmem_cache_node
*n
;
1920 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
1922 page
= new_slab(kmalloc_caches
, gfpflags
, node
);
1925 if (page_to_nid(page
) != node
) {
1926 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
1928 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
1929 "in order to be able to continue\n");
1934 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
1936 kmalloc_caches
->node
[node
] = n
;
1937 #ifdef CONFIG_SLUB_DEBUG
1938 init_object(kmalloc_caches
, n
, 1);
1939 init_tracking(kmalloc_caches
, n
);
1941 init_kmem_cache_node(n
);
1942 atomic_long_inc(&n
->nr_slabs
);
1943 add_partial(n
, page
);
1946 * new_slab() disables interupts. If we do not reenable interrupts here
1947 * then bootup would continue with interrupts disabled.
1953 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1957 for_each_node_state(node
, N_NORMAL_MEMORY
) {
1958 struct kmem_cache_node
*n
= s
->node
[node
];
1959 if (n
&& n
!= &s
->local_node
)
1960 kmem_cache_free(kmalloc_caches
, n
);
1961 s
->node
[node
] = NULL
;
1965 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1970 if (slab_state
>= UP
)
1971 local_node
= page_to_nid(virt_to_page(s
));
1975 for_each_node_state(node
, N_NORMAL_MEMORY
) {
1976 struct kmem_cache_node
*n
;
1978 if (local_node
== node
)
1981 if (slab_state
== DOWN
) {
1982 n
= early_kmem_cache_node_alloc(gfpflags
,
1986 n
= kmem_cache_alloc_node(kmalloc_caches
,
1990 free_kmem_cache_nodes(s
);
1996 init_kmem_cache_node(n
);
2001 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2005 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
2007 init_kmem_cache_node(&s
->local_node
);
2013 * calculate_sizes() determines the order and the distribution of data within
2016 static int calculate_sizes(struct kmem_cache
*s
)
2018 unsigned long flags
= s
->flags
;
2019 unsigned long size
= s
->objsize
;
2020 unsigned long align
= s
->align
;
2023 * Determine if we can poison the object itself. If the user of
2024 * the slab may touch the object after free or before allocation
2025 * then we should never poison the object itself.
2027 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2029 s
->flags
|= __OBJECT_POISON
;
2031 s
->flags
&= ~__OBJECT_POISON
;
2034 * Round up object size to the next word boundary. We can only
2035 * place the free pointer at word boundaries and this determines
2036 * the possible location of the free pointer.
2038 size
= ALIGN(size
, sizeof(void *));
2040 #ifdef CONFIG_SLUB_DEBUG
2042 * If we are Redzoning then check if there is some space between the
2043 * end of the object and the free pointer. If not then add an
2044 * additional word to have some bytes to store Redzone information.
2046 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
2047 size
+= sizeof(void *);
2051 * With that we have determined the number of bytes in actual use
2052 * by the object. This is the potential offset to the free pointer.
2056 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
2059 * Relocate free pointer after the object if it is not
2060 * permitted to overwrite the first word of the object on
2063 * This is the case if we do RCU, have a constructor or
2064 * destructor or are poisoning the objects.
2067 size
+= sizeof(void *);
2070 #ifdef CONFIG_SLUB_DEBUG
2071 if (flags
& SLAB_STORE_USER
)
2073 * Need to store information about allocs and frees after
2076 size
+= 2 * sizeof(struct track
);
2078 if (flags
& SLAB_RED_ZONE
)
2080 * Add some empty padding so that we can catch
2081 * overwrites from earlier objects rather than let
2082 * tracking information or the free pointer be
2083 * corrupted if an user writes before the start
2086 size
+= sizeof(void *);
2090 * Determine the alignment based on various parameters that the
2091 * user specified and the dynamic determination of cache line size
2094 align
= calculate_alignment(flags
, align
, s
->objsize
);
2097 * SLUB stores one object immediately after another beginning from
2098 * offset 0. In order to align the objects we have to simply size
2099 * each object to conform to the alignment.
2101 size
= ALIGN(size
, align
);
2104 s
->order
= calculate_order(size
);
2109 * Determine the number of objects per slab
2111 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
2114 * Verify that the number of objects is within permitted limits.
2115 * The page->inuse field is only 16 bit wide! So we cannot have
2116 * more than 64k objects per slab.
2118 if (!s
->objects
|| s
->objects
> MAX_OBJECTS_PER_SLAB
)
2124 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
2125 const char *name
, size_t size
,
2126 size_t align
, unsigned long flags
,
2127 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
2129 memset(s
, 0, kmem_size
);
2134 s
->flags
= kmem_cache_flags(size
, flags
, name
, ctor
);
2136 if (!calculate_sizes(s
))
2141 s
->defrag_ratio
= 100;
2143 if (!init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
2146 if (alloc_kmem_cache_cpus(s
, gfpflags
& ~SLUB_DMA
))
2149 if (flags
& SLAB_PANIC
)
2150 panic("Cannot create slab %s size=%lu realsize=%u "
2151 "order=%u offset=%u flags=%lx\n",
2152 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
2158 * Check if a given pointer is valid
2160 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
2164 page
= get_object_page(object
);
2166 if (!page
|| s
!= page
->slab
)
2167 /* No slab or wrong slab */
2170 if (!check_valid_pointer(s
, page
, object
))
2174 * We could also check if the object is on the slabs freelist.
2175 * But this would be too expensive and it seems that the main
2176 * purpose of kmem_ptr_valid is to check if the object belongs
2177 * to a certain slab.
2181 EXPORT_SYMBOL(kmem_ptr_validate
);
2184 * Determine the size of a slab object
2186 unsigned int kmem_cache_size(struct kmem_cache
*s
)
2190 EXPORT_SYMBOL(kmem_cache_size
);
2192 const char *kmem_cache_name(struct kmem_cache
*s
)
2196 EXPORT_SYMBOL(kmem_cache_name
);
2199 * Attempt to free all slabs on a node. Return the number of slabs we
2200 * were unable to free.
2202 static int free_list(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
2203 struct list_head
*list
)
2205 int slabs_inuse
= 0;
2206 unsigned long flags
;
2207 struct page
*page
, *h
;
2209 spin_lock_irqsave(&n
->list_lock
, flags
);
2210 list_for_each_entry_safe(page
, h
, list
, lru
)
2212 list_del(&page
->lru
);
2213 discard_slab(s
, page
);
2216 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2221 * Release all resources used by a slab cache.
2223 static inline int kmem_cache_close(struct kmem_cache
*s
)
2229 /* Attempt to free all objects */
2230 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2231 struct kmem_cache_node
*n
= get_node(s
, node
);
2233 n
->nr_partial
-= free_list(s
, n
, &n
->partial
);
2234 if (atomic_long_read(&n
->nr_slabs
))
2237 free_kmem_cache_nodes(s
);
2242 * Close a cache and release the kmem_cache structure
2243 * (must be used for caches created using kmem_cache_create)
2245 void kmem_cache_destroy(struct kmem_cache
*s
)
2247 down_write(&slub_lock
);
2251 up_write(&slub_lock
);
2252 if (kmem_cache_close(s
))
2254 sysfs_slab_remove(s
);
2257 up_write(&slub_lock
);
2259 EXPORT_SYMBOL(kmem_cache_destroy
);
2261 /********************************************************************
2263 *******************************************************************/
2265 struct kmem_cache kmalloc_caches
[PAGE_SHIFT
] __cacheline_aligned
;
2266 EXPORT_SYMBOL(kmalloc_caches
);
2268 #ifdef CONFIG_ZONE_DMA
2269 static struct kmem_cache
*kmalloc_caches_dma
[PAGE_SHIFT
];
2272 static int __init
setup_slub_min_order(char *str
)
2274 get_option (&str
, &slub_min_order
);
2279 __setup("slub_min_order=", setup_slub_min_order
);
2281 static int __init
setup_slub_max_order(char *str
)
2283 get_option (&str
, &slub_max_order
);
2288 __setup("slub_max_order=", setup_slub_max_order
);
2290 static int __init
setup_slub_min_objects(char *str
)
2292 get_option (&str
, &slub_min_objects
);
2297 __setup("slub_min_objects=", setup_slub_min_objects
);
2299 static int __init
setup_slub_nomerge(char *str
)
2305 __setup("slub_nomerge", setup_slub_nomerge
);
2307 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
2308 const char *name
, int size
, gfp_t gfp_flags
)
2310 unsigned int flags
= 0;
2312 if (gfp_flags
& SLUB_DMA
)
2313 flags
= SLAB_CACHE_DMA
;
2315 down_write(&slub_lock
);
2316 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
2320 list_add(&s
->list
, &slab_caches
);
2321 up_write(&slub_lock
);
2322 if (sysfs_slab_add(s
))
2327 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2330 #ifdef CONFIG_ZONE_DMA
2332 static void sysfs_add_func(struct work_struct
*w
)
2334 struct kmem_cache
*s
;
2336 down_write(&slub_lock
);
2337 list_for_each_entry(s
, &slab_caches
, list
) {
2338 if (s
->flags
& __SYSFS_ADD_DEFERRED
) {
2339 s
->flags
&= ~__SYSFS_ADD_DEFERRED
;
2343 up_write(&slub_lock
);
2346 static DECLARE_WORK(sysfs_add_work
, sysfs_add_func
);
2348 static noinline
struct kmem_cache
*dma_kmalloc_cache(int index
, gfp_t flags
)
2350 struct kmem_cache
*s
;
2354 s
= kmalloc_caches_dma
[index
];
2358 /* Dynamically create dma cache */
2359 if (flags
& __GFP_WAIT
)
2360 down_write(&slub_lock
);
2362 if (!down_write_trylock(&slub_lock
))
2366 if (kmalloc_caches_dma
[index
])
2369 realsize
= kmalloc_caches
[index
].objsize
;
2370 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d", (unsigned int)realsize
),
2371 s
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2373 if (!s
|| !text
|| !kmem_cache_open(s
, flags
, text
,
2374 realsize
, ARCH_KMALLOC_MINALIGN
,
2375 SLAB_CACHE_DMA
|__SYSFS_ADD_DEFERRED
, NULL
)) {
2381 list_add(&s
->list
, &slab_caches
);
2382 kmalloc_caches_dma
[index
] = s
;
2384 schedule_work(&sysfs_add_work
);
2387 up_write(&slub_lock
);
2389 return kmalloc_caches_dma
[index
];
2394 * Conversion table for small slabs sizes / 8 to the index in the
2395 * kmalloc array. This is necessary for slabs < 192 since we have non power
2396 * of two cache sizes there. The size of larger slabs can be determined using
2399 static s8 size_index
[24] = {
2426 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2432 return ZERO_SIZE_PTR
;
2434 index
= size_index
[(size
- 1) / 8];
2436 index
= fls(size
- 1);
2438 #ifdef CONFIG_ZONE_DMA
2439 if (unlikely((flags
& SLUB_DMA
)))
2440 return dma_kmalloc_cache(index
, flags
);
2443 return &kmalloc_caches
[index
];
2446 void *__kmalloc(size_t size
, gfp_t flags
)
2448 struct kmem_cache
*s
;
2450 if (unlikely(size
> PAGE_SIZE
/ 2))
2451 return (void *)__get_free_pages(flags
| __GFP_COMP
,
2454 s
= get_slab(size
, flags
);
2456 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2459 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2461 EXPORT_SYMBOL(__kmalloc
);
2464 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2466 struct kmem_cache
*s
;
2468 if (unlikely(size
> PAGE_SIZE
/ 2))
2469 return (void *)__get_free_pages(flags
| __GFP_COMP
,
2472 s
= get_slab(size
, flags
);
2474 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2477 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2479 EXPORT_SYMBOL(__kmalloc_node
);
2482 size_t ksize(const void *object
)
2485 struct kmem_cache
*s
;
2488 if (unlikely(object
== ZERO_SIZE_PTR
))
2491 page
= get_object_page(object
);
2497 * Debugging requires use of the padding between object
2498 * and whatever may come after it.
2500 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2504 * If we have the need to store the freelist pointer
2505 * back there or track user information then we can
2506 * only use the space before that information.
2508 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2512 * Else we can use all the padding etc for the allocation
2516 EXPORT_SYMBOL(ksize
);
2518 void kfree(const void *x
)
2522 if (unlikely(ZERO_OR_NULL_PTR(x
)))
2525 page
= virt_to_head_page(x
);
2526 if (unlikely(!PageSlab(page
))) {
2530 slab_free(page
->slab
, page
, (void *)x
, __builtin_return_address(0));
2532 EXPORT_SYMBOL(kfree
);
2535 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2536 * the remaining slabs by the number of items in use. The slabs with the
2537 * most items in use come first. New allocations will then fill those up
2538 * and thus they can be removed from the partial lists.
2540 * The slabs with the least items are placed last. This results in them
2541 * being allocated from last increasing the chance that the last objects
2542 * are freed in them.
2544 int kmem_cache_shrink(struct kmem_cache
*s
)
2548 struct kmem_cache_node
*n
;
2551 struct list_head
*slabs_by_inuse
=
2552 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2553 unsigned long flags
;
2555 if (!slabs_by_inuse
)
2559 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2560 n
= get_node(s
, node
);
2565 for (i
= 0; i
< s
->objects
; i
++)
2566 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2568 spin_lock_irqsave(&n
->list_lock
, flags
);
2571 * Build lists indexed by the items in use in each slab.
2573 * Note that concurrent frees may occur while we hold the
2574 * list_lock. page->inuse here is the upper limit.
2576 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2577 if (!page
->inuse
&& slab_trylock(page
)) {
2579 * Must hold slab lock here because slab_free
2580 * may have freed the last object and be
2581 * waiting to release the slab.
2583 list_del(&page
->lru
);
2586 discard_slab(s
, page
);
2588 list_move(&page
->lru
,
2589 slabs_by_inuse
+ page
->inuse
);
2594 * Rebuild the partial list with the slabs filled up most
2595 * first and the least used slabs at the end.
2597 for (i
= s
->objects
- 1; i
>= 0; i
--)
2598 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2600 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2603 kfree(slabs_by_inuse
);
2606 EXPORT_SYMBOL(kmem_cache_shrink
);
2608 /********************************************************************
2609 * Basic setup of slabs
2610 *******************************************************************/
2612 void __init
kmem_cache_init(void)
2619 * Must first have the slab cache available for the allocations of the
2620 * struct kmem_cache_node's. There is special bootstrap code in
2621 * kmem_cache_open for slab_state == DOWN.
2623 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2624 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2625 kmalloc_caches
[0].refcount
= -1;
2629 /* Able to allocate the per node structures */
2630 slab_state
= PARTIAL
;
2632 /* Caches that are not of the two-to-the-power-of size */
2633 if (KMALLOC_MIN_SIZE
<= 64) {
2634 create_kmalloc_cache(&kmalloc_caches
[1],
2635 "kmalloc-96", 96, GFP_KERNEL
);
2638 if (KMALLOC_MIN_SIZE
<= 128) {
2639 create_kmalloc_cache(&kmalloc_caches
[2],
2640 "kmalloc-192", 192, GFP_KERNEL
);
2644 for (i
= KMALLOC_SHIFT_LOW
; i
< PAGE_SHIFT
; i
++) {
2645 create_kmalloc_cache(&kmalloc_caches
[i
],
2646 "kmalloc", 1 << i
, GFP_KERNEL
);
2652 * Patch up the size_index table if we have strange large alignment
2653 * requirements for the kmalloc array. This is only the case for
2654 * mips it seems. The standard arches will not generate any code here.
2656 * Largest permitted alignment is 256 bytes due to the way we
2657 * handle the index determination for the smaller caches.
2659 * Make sure that nothing crazy happens if someone starts tinkering
2660 * around with ARCH_KMALLOC_MINALIGN
2662 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 256 ||
2663 (KMALLOC_MIN_SIZE
& (KMALLOC_MIN_SIZE
- 1)));
2665 for (i
= 8; i
< KMALLOC_MIN_SIZE
; i
+= 8)
2666 size_index
[(i
- 1) / 8] = KMALLOC_SHIFT_LOW
;
2670 /* Provide the correct kmalloc names now that the caches are up */
2671 for (i
= KMALLOC_SHIFT_LOW
; i
< PAGE_SHIFT
; i
++)
2672 kmalloc_caches
[i
]. name
=
2673 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
2676 register_cpu_notifier(&slab_notifier
);
2679 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
) +
2680 nr_cpu_ids
* sizeof(struct kmem_cache_cpu
);
2682 printk(KERN_INFO
"SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2683 " CPUs=%d, Nodes=%d\n",
2684 caches
, cache_line_size(),
2685 slub_min_order
, slub_max_order
, slub_min_objects
,
2686 nr_cpu_ids
, nr_node_ids
);
2690 * Find a mergeable slab cache
2692 static int slab_unmergeable(struct kmem_cache
*s
)
2694 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
2701 * We may have set a slab to be unmergeable during bootstrap.
2703 if (s
->refcount
< 0)
2709 static struct kmem_cache
*find_mergeable(size_t size
,
2710 size_t align
, unsigned long flags
, const char *name
,
2711 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
2713 struct kmem_cache
*s
;
2715 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
2721 size
= ALIGN(size
, sizeof(void *));
2722 align
= calculate_alignment(flags
, align
, size
);
2723 size
= ALIGN(size
, align
);
2724 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
2726 list_for_each_entry(s
, &slab_caches
, list
) {
2727 if (slab_unmergeable(s
))
2733 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
2736 * Check if alignment is compatible.
2737 * Courtesy of Adrian Drzewiecki
2739 if ((s
->size
& ~(align
-1)) != s
->size
)
2742 if (s
->size
- size
>= sizeof(void *))
2750 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
2751 size_t align
, unsigned long flags
,
2752 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
2754 struct kmem_cache
*s
;
2756 down_write(&slub_lock
);
2757 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
2761 * Adjust the object sizes so that we clear
2762 * the complete object on kzalloc.
2764 s
->objsize
= max(s
->objsize
, (int)size
);
2765 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
2766 up_write(&slub_lock
);
2767 if (sysfs_slab_alias(s
, name
))
2771 s
= kmalloc(kmem_size
, GFP_KERNEL
);
2773 if (kmem_cache_open(s
, GFP_KERNEL
, name
,
2774 size
, align
, flags
, ctor
)) {
2775 list_add(&s
->list
, &slab_caches
);
2776 up_write(&slub_lock
);
2777 if (sysfs_slab_add(s
))
2783 up_write(&slub_lock
);
2786 if (flags
& SLAB_PANIC
)
2787 panic("Cannot create slabcache %s\n", name
);
2792 EXPORT_SYMBOL(kmem_cache_create
);
2796 * Use the cpu notifier to insure that the cpu slabs are flushed when
2799 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
2800 unsigned long action
, void *hcpu
)
2802 long cpu
= (long)hcpu
;
2803 struct kmem_cache
*s
;
2804 unsigned long flags
;
2807 case CPU_UP_CANCELED
:
2808 case CPU_UP_CANCELED_FROZEN
:
2810 case CPU_DEAD_FROZEN
:
2811 down_read(&slub_lock
);
2812 list_for_each_entry(s
, &slab_caches
, list
) {
2813 local_irq_save(flags
);
2814 __flush_cpu_slab(s
, cpu
);
2815 local_irq_restore(flags
);
2817 up_read(&slub_lock
);
2825 static struct notifier_block __cpuinitdata slab_notifier
=
2826 { &slab_cpuup_callback
, NULL
, 0 };
2830 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
2832 struct kmem_cache
*s
;
2834 if (unlikely(size
> PAGE_SIZE
/ 2))
2835 return (void *)__get_free_pages(gfpflags
| __GFP_COMP
,
2837 s
= get_slab(size
, gfpflags
);
2839 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2842 return slab_alloc(s
, gfpflags
, -1, caller
);
2845 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
2846 int node
, void *caller
)
2848 struct kmem_cache
*s
;
2850 if (unlikely(size
> PAGE_SIZE
/ 2))
2851 return (void *)__get_free_pages(gfpflags
| __GFP_COMP
,
2853 s
= get_slab(size
, gfpflags
);
2855 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2858 return slab_alloc(s
, gfpflags
, node
, caller
);
2861 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2862 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
2866 void *addr
= page_address(page
);
2868 if (!check_slab(s
, page
) ||
2869 !on_freelist(s
, page
, NULL
))
2872 /* Now we know that a valid freelist exists */
2873 bitmap_zero(map
, s
->objects
);
2875 for_each_free_object(p
, s
, page
->freelist
) {
2876 set_bit(slab_index(p
, s
, addr
), map
);
2877 if (!check_object(s
, page
, p
, 0))
2881 for_each_object(p
, s
, addr
)
2882 if (!test_bit(slab_index(p
, s
, addr
), map
))
2883 if (!check_object(s
, page
, p
, 1))
2888 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
2891 if (slab_trylock(page
)) {
2892 validate_slab(s
, page
, map
);
2895 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
2898 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
2899 if (!SlabDebug(page
))
2900 printk(KERN_ERR
"SLUB %s: SlabDebug not set "
2901 "on slab 0x%p\n", s
->name
, page
);
2903 if (SlabDebug(page
))
2904 printk(KERN_ERR
"SLUB %s: SlabDebug set on "
2905 "slab 0x%p\n", s
->name
, page
);
2909 static int validate_slab_node(struct kmem_cache
*s
,
2910 struct kmem_cache_node
*n
, unsigned long *map
)
2912 unsigned long count
= 0;
2914 unsigned long flags
;
2916 spin_lock_irqsave(&n
->list_lock
, flags
);
2918 list_for_each_entry(page
, &n
->partial
, lru
) {
2919 validate_slab_slab(s
, page
, map
);
2922 if (count
!= n
->nr_partial
)
2923 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
2924 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
2926 if (!(s
->flags
& SLAB_STORE_USER
))
2929 list_for_each_entry(page
, &n
->full
, lru
) {
2930 validate_slab_slab(s
, page
, map
);
2933 if (count
!= atomic_long_read(&n
->nr_slabs
))
2934 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
2935 "counter=%ld\n", s
->name
, count
,
2936 atomic_long_read(&n
->nr_slabs
));
2939 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2943 static long validate_slab_cache(struct kmem_cache
*s
)
2946 unsigned long count
= 0;
2947 unsigned long *map
= kmalloc(BITS_TO_LONGS(s
->objects
) *
2948 sizeof(unsigned long), GFP_KERNEL
);
2954 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2955 struct kmem_cache_node
*n
= get_node(s
, node
);
2957 count
+= validate_slab_node(s
, n
, map
);
2963 #ifdef SLUB_RESILIENCY_TEST
2964 static void resiliency_test(void)
2968 printk(KERN_ERR
"SLUB resiliency testing\n");
2969 printk(KERN_ERR
"-----------------------\n");
2970 printk(KERN_ERR
"A. Corruption after allocation\n");
2972 p
= kzalloc(16, GFP_KERNEL
);
2974 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
2975 " 0x12->0x%p\n\n", p
+ 16);
2977 validate_slab_cache(kmalloc_caches
+ 4);
2979 /* Hmmm... The next two are dangerous */
2980 p
= kzalloc(32, GFP_KERNEL
);
2981 p
[32 + sizeof(void *)] = 0x34;
2982 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
2983 " 0x34 -> -0x%p\n", p
);
2984 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2986 validate_slab_cache(kmalloc_caches
+ 5);
2987 p
= kzalloc(64, GFP_KERNEL
);
2988 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
2990 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2992 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2993 validate_slab_cache(kmalloc_caches
+ 6);
2995 printk(KERN_ERR
"\nB. Corruption after free\n");
2996 p
= kzalloc(128, GFP_KERNEL
);
2999 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
3000 validate_slab_cache(kmalloc_caches
+ 7);
3002 p
= kzalloc(256, GFP_KERNEL
);
3005 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
3006 validate_slab_cache(kmalloc_caches
+ 8);
3008 p
= kzalloc(512, GFP_KERNEL
);
3011 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
3012 validate_slab_cache(kmalloc_caches
+ 9);
3015 static void resiliency_test(void) {};
3019 * Generate lists of code addresses where slabcache objects are allocated
3024 unsigned long count
;
3037 unsigned long count
;
3038 struct location
*loc
;
3041 static void free_loc_track(struct loc_track
*t
)
3044 free_pages((unsigned long)t
->loc
,
3045 get_order(sizeof(struct location
) * t
->max
));
3048 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
3053 order
= get_order(sizeof(struct location
) * max
);
3055 l
= (void *)__get_free_pages(flags
, order
);
3060 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
3068 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
3069 const struct track
*track
)
3071 long start
, end
, pos
;
3074 unsigned long age
= jiffies
- track
->when
;
3080 pos
= start
+ (end
- start
+ 1) / 2;
3083 * There is nothing at "end". If we end up there
3084 * we need to add something to before end.
3089 caddr
= t
->loc
[pos
].addr
;
3090 if (track
->addr
== caddr
) {
3096 if (age
< l
->min_time
)
3098 if (age
> l
->max_time
)
3101 if (track
->pid
< l
->min_pid
)
3102 l
->min_pid
= track
->pid
;
3103 if (track
->pid
> l
->max_pid
)
3104 l
->max_pid
= track
->pid
;
3106 cpu_set(track
->cpu
, l
->cpus
);
3108 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3112 if (track
->addr
< caddr
)
3119 * Not found. Insert new tracking element.
3121 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
3127 (t
->count
- pos
) * sizeof(struct location
));
3130 l
->addr
= track
->addr
;
3134 l
->min_pid
= track
->pid
;
3135 l
->max_pid
= track
->pid
;
3136 cpus_clear(l
->cpus
);
3137 cpu_set(track
->cpu
, l
->cpus
);
3138 nodes_clear(l
->nodes
);
3139 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3143 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
3144 struct page
*page
, enum track_item alloc
)
3146 void *addr
= page_address(page
);
3147 DECLARE_BITMAP(map
, s
->objects
);
3150 bitmap_zero(map
, s
->objects
);
3151 for_each_free_object(p
, s
, page
->freelist
)
3152 set_bit(slab_index(p
, s
, addr
), map
);
3154 for_each_object(p
, s
, addr
)
3155 if (!test_bit(slab_index(p
, s
, addr
), map
))
3156 add_location(t
, s
, get_track(s
, p
, alloc
));
3159 static int list_locations(struct kmem_cache
*s
, char *buf
,
3160 enum track_item alloc
)
3164 struct loc_track t
= { 0, 0, NULL
};
3167 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
3169 return sprintf(buf
, "Out of memory\n");
3171 /* Push back cpu slabs */
3174 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3175 struct kmem_cache_node
*n
= get_node(s
, node
);
3176 unsigned long flags
;
3179 if (!atomic_long_read(&n
->nr_slabs
))
3182 spin_lock_irqsave(&n
->list_lock
, flags
);
3183 list_for_each_entry(page
, &n
->partial
, lru
)
3184 process_slab(&t
, s
, page
, alloc
);
3185 list_for_each_entry(page
, &n
->full
, lru
)
3186 process_slab(&t
, s
, page
, alloc
);
3187 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3190 for (i
= 0; i
< t
.count
; i
++) {
3191 struct location
*l
= &t
.loc
[i
];
3193 if (n
> PAGE_SIZE
- 100)
3195 n
+= sprintf(buf
+ n
, "%7ld ", l
->count
);
3198 n
+= sprint_symbol(buf
+ n
, (unsigned long)l
->addr
);
3200 n
+= sprintf(buf
+ n
, "<not-available>");
3202 if (l
->sum_time
!= l
->min_time
) {
3203 unsigned long remainder
;
3205 n
+= sprintf(buf
+ n
, " age=%ld/%ld/%ld",
3207 div_long_long_rem(l
->sum_time
, l
->count
, &remainder
),
3210 n
+= sprintf(buf
+ n
, " age=%ld",
3213 if (l
->min_pid
!= l
->max_pid
)
3214 n
+= sprintf(buf
+ n
, " pid=%ld-%ld",
3215 l
->min_pid
, l
->max_pid
);
3217 n
+= sprintf(buf
+ n
, " pid=%ld",
3220 if (num_online_cpus() > 1 && !cpus_empty(l
->cpus
) &&
3221 n
< PAGE_SIZE
- 60) {
3222 n
+= sprintf(buf
+ n
, " cpus=");
3223 n
+= cpulist_scnprintf(buf
+ n
, PAGE_SIZE
- n
- 50,
3227 if (num_online_nodes() > 1 && !nodes_empty(l
->nodes
) &&
3228 n
< PAGE_SIZE
- 60) {
3229 n
+= sprintf(buf
+ n
, " nodes=");
3230 n
+= nodelist_scnprintf(buf
+ n
, PAGE_SIZE
- n
- 50,
3234 n
+= sprintf(buf
+ n
, "\n");
3239 n
+= sprintf(buf
, "No data\n");
3243 static unsigned long count_partial(struct kmem_cache_node
*n
)
3245 unsigned long flags
;
3246 unsigned long x
= 0;
3249 spin_lock_irqsave(&n
->list_lock
, flags
);
3250 list_for_each_entry(page
, &n
->partial
, lru
)
3252 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3256 enum slab_stat_type
{
3263 #define SO_FULL (1 << SL_FULL)
3264 #define SO_PARTIAL (1 << SL_PARTIAL)
3265 #define SO_CPU (1 << SL_CPU)
3266 #define SO_OBJECTS (1 << SL_OBJECTS)
3268 static unsigned long slab_objects(struct kmem_cache
*s
,
3269 char *buf
, unsigned long flags
)
3271 unsigned long total
= 0;
3275 unsigned long *nodes
;
3276 unsigned long *per_cpu
;
3278 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
3279 per_cpu
= nodes
+ nr_node_ids
;
3281 for_each_possible_cpu(cpu
) {
3283 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3290 if (flags
& SO_CPU
) {
3293 if (flags
& SO_OBJECTS
)
3298 nodes
[c
->node
] += x
;
3304 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3305 struct kmem_cache_node
*n
= get_node(s
, node
);
3307 if (flags
& SO_PARTIAL
) {
3308 if (flags
& SO_OBJECTS
)
3309 x
= count_partial(n
);
3316 if (flags
& SO_FULL
) {
3317 int full_slabs
= atomic_long_read(&n
->nr_slabs
)
3321 if (flags
& SO_OBJECTS
)
3322 x
= full_slabs
* s
->objects
;
3330 x
= sprintf(buf
, "%lu", total
);
3332 for_each_node_state(node
, N_NORMAL_MEMORY
)
3334 x
+= sprintf(buf
+ x
, " N%d=%lu",
3338 return x
+ sprintf(buf
+ x
, "\n");
3341 static int any_slab_objects(struct kmem_cache
*s
)
3346 for_each_possible_cpu(cpu
) {
3347 struct kmem_cache_cpu
*c
= get_cpu_slab(s
, cpu
);
3353 for_each_online_node(node
) {
3354 struct kmem_cache_node
*n
= get_node(s
, node
);
3359 if (n
->nr_partial
|| atomic_long_read(&n
->nr_slabs
))
3365 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3366 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3368 struct slab_attribute
{
3369 struct attribute attr
;
3370 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
3371 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
3374 #define SLAB_ATTR_RO(_name) \
3375 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3377 #define SLAB_ATTR(_name) \
3378 static struct slab_attribute _name##_attr = \
3379 __ATTR(_name, 0644, _name##_show, _name##_store)
3381 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
3383 return sprintf(buf
, "%d\n", s
->size
);
3385 SLAB_ATTR_RO(slab_size
);
3387 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3389 return sprintf(buf
, "%d\n", s
->align
);
3391 SLAB_ATTR_RO(align
);
3393 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3395 return sprintf(buf
, "%d\n", s
->objsize
);
3397 SLAB_ATTR_RO(object_size
);
3399 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3401 return sprintf(buf
, "%d\n", s
->objects
);
3403 SLAB_ATTR_RO(objs_per_slab
);
3405 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3407 return sprintf(buf
, "%d\n", s
->order
);
3409 SLAB_ATTR_RO(order
);
3411 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3414 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3416 return n
+ sprintf(buf
+ n
, "\n");
3422 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3424 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3426 SLAB_ATTR_RO(aliases
);
3428 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3430 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3432 SLAB_ATTR_RO(slabs
);
3434 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3436 return slab_objects(s
, buf
, SO_PARTIAL
);
3438 SLAB_ATTR_RO(partial
);
3440 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3442 return slab_objects(s
, buf
, SO_CPU
);
3444 SLAB_ATTR_RO(cpu_slabs
);
3446 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3448 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3450 SLAB_ATTR_RO(objects
);
3452 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3454 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3457 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3458 const char *buf
, size_t length
)
3460 s
->flags
&= ~SLAB_DEBUG_FREE
;
3462 s
->flags
|= SLAB_DEBUG_FREE
;
3465 SLAB_ATTR(sanity_checks
);
3467 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3469 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3472 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3475 s
->flags
&= ~SLAB_TRACE
;
3477 s
->flags
|= SLAB_TRACE
;
3482 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3484 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3487 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3488 const char *buf
, size_t length
)
3490 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3492 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3495 SLAB_ATTR(reclaim_account
);
3497 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3499 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3501 SLAB_ATTR_RO(hwcache_align
);
3503 #ifdef CONFIG_ZONE_DMA
3504 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3506 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3508 SLAB_ATTR_RO(cache_dma
);
3511 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3513 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3515 SLAB_ATTR_RO(destroy_by_rcu
);
3517 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3519 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3522 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3523 const char *buf
, size_t length
)
3525 if (any_slab_objects(s
))
3528 s
->flags
&= ~SLAB_RED_ZONE
;
3530 s
->flags
|= SLAB_RED_ZONE
;
3534 SLAB_ATTR(red_zone
);
3536 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3538 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3541 static ssize_t
poison_store(struct kmem_cache
*s
,
3542 const char *buf
, size_t length
)
3544 if (any_slab_objects(s
))
3547 s
->flags
&= ~SLAB_POISON
;
3549 s
->flags
|= SLAB_POISON
;
3555 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3557 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3560 static ssize_t
store_user_store(struct kmem_cache
*s
,
3561 const char *buf
, size_t length
)
3563 if (any_slab_objects(s
))
3566 s
->flags
&= ~SLAB_STORE_USER
;
3568 s
->flags
|= SLAB_STORE_USER
;
3572 SLAB_ATTR(store_user
);
3574 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3579 static ssize_t
validate_store(struct kmem_cache
*s
,
3580 const char *buf
, size_t length
)
3584 if (buf
[0] == '1') {
3585 ret
= validate_slab_cache(s
);
3591 SLAB_ATTR(validate
);
3593 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3598 static ssize_t
shrink_store(struct kmem_cache
*s
,
3599 const char *buf
, size_t length
)
3601 if (buf
[0] == '1') {
3602 int rc
= kmem_cache_shrink(s
);
3612 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
3614 if (!(s
->flags
& SLAB_STORE_USER
))
3616 return list_locations(s
, buf
, TRACK_ALLOC
);
3618 SLAB_ATTR_RO(alloc_calls
);
3620 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
3622 if (!(s
->flags
& SLAB_STORE_USER
))
3624 return list_locations(s
, buf
, TRACK_FREE
);
3626 SLAB_ATTR_RO(free_calls
);
3629 static ssize_t
defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
3631 return sprintf(buf
, "%d\n", s
->defrag_ratio
/ 10);
3634 static ssize_t
defrag_ratio_store(struct kmem_cache
*s
,
3635 const char *buf
, size_t length
)
3637 int n
= simple_strtoul(buf
, NULL
, 10);
3640 s
->defrag_ratio
= n
* 10;
3643 SLAB_ATTR(defrag_ratio
);
3646 static struct attribute
* slab_attrs
[] = {
3647 &slab_size_attr
.attr
,
3648 &object_size_attr
.attr
,
3649 &objs_per_slab_attr
.attr
,
3654 &cpu_slabs_attr
.attr
,
3658 &sanity_checks_attr
.attr
,
3660 &hwcache_align_attr
.attr
,
3661 &reclaim_account_attr
.attr
,
3662 &destroy_by_rcu_attr
.attr
,
3663 &red_zone_attr
.attr
,
3665 &store_user_attr
.attr
,
3666 &validate_attr
.attr
,
3668 &alloc_calls_attr
.attr
,
3669 &free_calls_attr
.attr
,
3670 #ifdef CONFIG_ZONE_DMA
3671 &cache_dma_attr
.attr
,
3674 &defrag_ratio_attr
.attr
,
3679 static struct attribute_group slab_attr_group
= {
3680 .attrs
= slab_attrs
,
3683 static ssize_t
slab_attr_show(struct kobject
*kobj
,
3684 struct attribute
*attr
,
3687 struct slab_attribute
*attribute
;
3688 struct kmem_cache
*s
;
3691 attribute
= to_slab_attr(attr
);
3694 if (!attribute
->show
)
3697 err
= attribute
->show(s
, buf
);
3702 static ssize_t
slab_attr_store(struct kobject
*kobj
,
3703 struct attribute
*attr
,
3704 const char *buf
, size_t len
)
3706 struct slab_attribute
*attribute
;
3707 struct kmem_cache
*s
;
3710 attribute
= to_slab_attr(attr
);
3713 if (!attribute
->store
)
3716 err
= attribute
->store(s
, buf
, len
);
3721 static struct sysfs_ops slab_sysfs_ops
= {
3722 .show
= slab_attr_show
,
3723 .store
= slab_attr_store
,
3726 static struct kobj_type slab_ktype
= {
3727 .sysfs_ops
= &slab_sysfs_ops
,
3730 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
3732 struct kobj_type
*ktype
= get_ktype(kobj
);
3734 if (ktype
== &slab_ktype
)
3739 static struct kset_uevent_ops slab_uevent_ops
= {
3740 .filter
= uevent_filter
,
3743 static decl_subsys(slab
, &slab_ktype
, &slab_uevent_ops
);
3745 #define ID_STR_LENGTH 64
3747 /* Create a unique string id for a slab cache:
3749 * :[flags-]size:[memory address of kmemcache]
3751 static char *create_unique_id(struct kmem_cache
*s
)
3753 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
3760 * First flags affecting slabcache operations. We will only
3761 * get here for aliasable slabs so we do not need to support
3762 * too many flags. The flags here must cover all flags that
3763 * are matched during merging to guarantee that the id is
3766 if (s
->flags
& SLAB_CACHE_DMA
)
3768 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3770 if (s
->flags
& SLAB_DEBUG_FREE
)
3774 p
+= sprintf(p
, "%07d", s
->size
);
3775 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
3779 static int sysfs_slab_add(struct kmem_cache
*s
)
3785 if (slab_state
< SYSFS
)
3786 /* Defer until later */
3789 unmergeable
= slab_unmergeable(s
);
3792 * Slabcache can never be merged so we can use the name proper.
3793 * This is typically the case for debug situations. In that
3794 * case we can catch duplicate names easily.
3796 sysfs_remove_link(&slab_subsys
.kobj
, s
->name
);
3800 * Create a unique name for the slab as a target
3803 name
= create_unique_id(s
);
3806 kobj_set_kset_s(s
, slab_subsys
);
3807 kobject_set_name(&s
->kobj
, name
);
3808 kobject_init(&s
->kobj
);
3809 err
= kobject_add(&s
->kobj
);
3813 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
3816 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
3818 /* Setup first alias */
3819 sysfs_slab_alias(s
, s
->name
);
3825 static void sysfs_slab_remove(struct kmem_cache
*s
)
3827 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
3828 kobject_del(&s
->kobj
);
3832 * Need to buffer aliases during bootup until sysfs becomes
3833 * available lest we loose that information.
3835 struct saved_alias
{
3836 struct kmem_cache
*s
;
3838 struct saved_alias
*next
;
3841 static struct saved_alias
*alias_list
;
3843 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
3845 struct saved_alias
*al
;
3847 if (slab_state
== SYSFS
) {
3849 * If we have a leftover link then remove it.
3851 sysfs_remove_link(&slab_subsys
.kobj
, name
);
3852 return sysfs_create_link(&slab_subsys
.kobj
,
3856 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
3862 al
->next
= alias_list
;
3867 static int __init
slab_sysfs_init(void)
3869 struct kmem_cache
*s
;
3872 err
= subsystem_register(&slab_subsys
);
3874 printk(KERN_ERR
"Cannot register slab subsystem.\n");
3880 list_for_each_entry(s
, &slab_caches
, list
) {
3881 err
= sysfs_slab_add(s
);
3883 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
3884 " to sysfs\n", s
->name
);
3887 while (alias_list
) {
3888 struct saved_alias
*al
= alias_list
;
3890 alias_list
= alias_list
->next
;
3891 err
= sysfs_slab_alias(al
->s
, al
->name
);
3893 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
3894 " %s to sysfs\n", s
->name
);
3902 __initcall(slab_sysfs_init
);