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.
70 * There is no list for full slabs. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * Otherwise there is no need to track full slabs unless we have to
73 * track full slabs for debugging purposes.
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 used as a cpu cache. Allocations
82 * may be performed from the slab. The slab is not
83 * on any slab list and cannot be moved onto one.
85 * PageError Slab requires special handling due to debug
86 * options set. This moves slab handling out of
91 * Issues still to be resolved:
93 * - The per cpu array is updated for each new slab and and is a remote
94 * cacheline for most nodes. This could become a bouncing cacheline given
95 * enough frequent updates. There are 16 pointers in a cacheline.so at
96 * max 16 cpus could compete. Likely okay.
98 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
100 * - Variable sizing of the per node arrays
103 /* Enable to test recovery from slab corruption on boot */
104 #undef SLUB_RESILIENCY_TEST
109 * Small page size. Make sure that we do not fragment memory
111 #define DEFAULT_MAX_ORDER 1
112 #define DEFAULT_MIN_OBJECTS 4
117 * Large page machines are customarily able to handle larger
120 #define DEFAULT_MAX_ORDER 2
121 #define DEFAULT_MIN_OBJECTS 8
126 * Mininum number of partial slabs. These will be left on the partial
127 * lists even if they are empty. kmem_cache_shrink may reclaim them.
129 #define MIN_PARTIAL 2
132 * Maximum number of desirable partial slabs.
133 * The existence of more partial slabs makes kmem_cache_shrink
134 * sort the partial list by the number of objects in the.
136 #define MAX_PARTIAL 10
138 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_STORE_USER)
141 * Set of flags that will prevent slab merging
143 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
146 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
149 #ifndef ARCH_KMALLOC_MINALIGN
150 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
153 #ifndef ARCH_SLAB_MINALIGN
154 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
157 /* Internal SLUB flags */
158 #define __OBJECT_POISON 0x80000000 /* Poison object */
160 /* Not all arches define cache_line_size */
161 #ifndef cache_line_size
162 #define cache_line_size() L1_CACHE_BYTES
165 static int kmem_size
= sizeof(struct kmem_cache
);
168 static struct notifier_block slab_notifier
;
172 DOWN
, /* No slab functionality available */
173 PARTIAL
, /* kmem_cache_open() works but kmalloc does not */
174 UP
, /* Everything works */
178 /* A list of all slab caches on the system */
179 static DECLARE_RWSEM(slub_lock
);
180 LIST_HEAD(slab_caches
);
183 static int sysfs_slab_add(struct kmem_cache
*);
184 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
185 static void sysfs_slab_remove(struct kmem_cache
*);
187 static int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
188 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
) { return 0; }
189 static void sysfs_slab_remove(struct kmem_cache
*s
) {}
192 /********************************************************************
193 * Core slab cache functions
194 *******************************************************************/
196 int slab_is_available(void)
198 return slab_state
>= UP
;
201 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
204 return s
->node
[node
];
206 return &s
->local_node
;
213 static void print_section(char *text
, u8
*addr
, unsigned int length
)
221 for (i
= 0; i
< length
; i
++) {
223 printk(KERN_ERR
"%10s 0x%p: ", text
, addr
+ i
);
226 printk(" %02x", addr
[i
]);
228 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
230 printk(" %s\n",ascii
);
241 printk(" %s\n", ascii
);
246 * Slow version of get and set free pointer.
248 * This requires touching the cache lines of kmem_cache.
249 * The offset can also be obtained from the page. In that
250 * case it is in the cacheline that we already need to touch.
252 static void *get_freepointer(struct kmem_cache
*s
, void *object
)
254 return *(void **)(object
+ s
->offset
);
257 static void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
259 *(void **)(object
+ s
->offset
) = fp
;
263 * Tracking user of a slab.
266 void *addr
; /* Called from address */
267 int cpu
; /* Was running on cpu */
268 int pid
; /* Pid context */
269 unsigned long when
; /* When did the operation occur */
272 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
274 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
275 enum track_item alloc
)
280 p
= object
+ s
->offset
+ sizeof(void *);
282 p
= object
+ s
->inuse
;
287 static void set_track(struct kmem_cache
*s
, void *object
,
288 enum track_item alloc
, void *addr
)
293 p
= object
+ s
->offset
+ sizeof(void *);
295 p
= object
+ s
->inuse
;
300 p
->cpu
= smp_processor_id();
301 p
->pid
= current
? current
->pid
: -1;
304 memset(p
, 0, sizeof(struct track
));
307 static void init_tracking(struct kmem_cache
*s
, void *object
)
309 if (s
->flags
& SLAB_STORE_USER
) {
310 set_track(s
, object
, TRACK_FREE
, NULL
);
311 set_track(s
, object
, TRACK_ALLOC
, NULL
);
315 static void print_track(const char *s
, struct track
*t
)
320 printk(KERN_ERR
"%s: ", s
);
321 __print_symbol("%s", (unsigned long)t
->addr
);
322 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies
- t
->when
, t
->cpu
, t
->pid
);
325 static void print_trailer(struct kmem_cache
*s
, u8
*p
)
327 unsigned int off
; /* Offset of last byte */
329 if (s
->flags
& SLAB_RED_ZONE
)
330 print_section("Redzone", p
+ s
->objsize
,
331 s
->inuse
- s
->objsize
);
333 printk(KERN_ERR
"FreePointer 0x%p -> 0x%p\n",
335 get_freepointer(s
, p
));
338 off
= s
->offset
+ sizeof(void *);
342 if (s
->flags
& SLAB_STORE_USER
) {
343 print_track("Last alloc", get_track(s
, p
, TRACK_ALLOC
));
344 print_track("Last free ", get_track(s
, p
, TRACK_FREE
));
345 off
+= 2 * sizeof(struct track
);
349 /* Beginning of the filler is the free pointer */
350 print_section("Filler", p
+ off
, s
->size
- off
);
353 static void object_err(struct kmem_cache
*s
, struct page
*page
,
354 u8
*object
, char *reason
)
356 u8
*addr
= page_address(page
);
358 printk(KERN_ERR
"*** SLUB %s: %s@0x%p slab 0x%p\n",
359 s
->name
, reason
, object
, page
);
360 printk(KERN_ERR
" offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
361 object
- addr
, page
->flags
, page
->inuse
, page
->freelist
);
362 if (object
> addr
+ 16)
363 print_section("Bytes b4", object
- 16, 16);
364 print_section("Object", object
, min(s
->objsize
, 128));
365 print_trailer(s
, object
);
369 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *reason
, ...)
374 va_start(args
, reason
);
375 vsnprintf(buf
, sizeof(buf
), reason
, args
);
377 printk(KERN_ERR
"*** SLUB %s: %s in slab @0x%p\n", s
->name
, buf
,
382 static void init_object(struct kmem_cache
*s
, void *object
, int active
)
386 if (s
->flags
& __OBJECT_POISON
) {
387 memset(p
, POISON_FREE
, s
->objsize
- 1);
388 p
[s
->objsize
-1] = POISON_END
;
391 if (s
->flags
& SLAB_RED_ZONE
)
392 memset(p
+ s
->objsize
,
393 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
,
394 s
->inuse
- s
->objsize
);
397 static int check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
400 if (*start
!= (u8
)value
)
408 static inline int check_valid_pointer(struct kmem_cache
*s
,
409 struct page
*page
, const void *object
)
416 base
= page_address(page
);
417 if (object
< base
|| object
>= base
+ s
->objects
* s
->size
||
418 (object
- base
) % s
->size
) {
429 * Bytes of the object to be managed.
430 * If the freepointer may overlay the object then the free
431 * pointer is the first word of the object.
432 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
435 * object + s->objsize
436 * Padding to reach word boundary. This is also used for Redzoning.
437 * Padding is extended to word size if Redzoning is enabled
438 * and objsize == inuse.
439 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
440 * 0xcc (RED_ACTIVE) for objects in use.
443 * A. Free pointer (if we cannot overwrite object on free)
444 * B. Tracking data for SLAB_STORE_USER
445 * C. Padding to reach required alignment boundary
446 * Padding is done using 0x5a (POISON_INUSE)
450 * If slabcaches are merged then the objsize and inuse boundaries are to
451 * be ignored. And therefore no slab options that rely on these boundaries
452 * may be used with merged slabcaches.
455 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
456 void *from
, void *to
)
458 printk(KERN_ERR
"@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
459 s
->name
, message
, data
, from
, to
- 1);
460 memset(from
, data
, to
- from
);
463 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
465 unsigned long off
= s
->inuse
; /* The end of info */
468 /* Freepointer is placed after the object. */
469 off
+= sizeof(void *);
471 if (s
->flags
& SLAB_STORE_USER
)
472 /* We also have user information there */
473 off
+= 2 * sizeof(struct track
);
478 if (check_bytes(p
+ off
, POISON_INUSE
, s
->size
- off
))
481 object_err(s
, page
, p
, "Object padding check fails");
486 restore_bytes(s
, "object padding", POISON_INUSE
, p
+ off
, p
+ s
->size
);
490 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
493 int length
, remainder
;
495 if (!(s
->flags
& SLAB_POISON
))
498 p
= page_address(page
);
499 length
= s
->objects
* s
->size
;
500 remainder
= (PAGE_SIZE
<< s
->order
) - length
;
504 if (!check_bytes(p
+ length
, POISON_INUSE
, remainder
)) {
505 slab_err(s
, page
, "Padding check failed");
506 restore_bytes(s
, "slab padding", POISON_INUSE
, p
+ length
,
507 p
+ length
+ remainder
);
513 static int check_object(struct kmem_cache
*s
, struct page
*page
,
514 void *object
, int active
)
517 u8
*endobject
= object
+ s
->objsize
;
519 if (s
->flags
& SLAB_RED_ZONE
) {
521 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
523 if (!check_bytes(endobject
, red
, s
->inuse
- s
->objsize
)) {
524 object_err(s
, page
, object
,
525 active
? "Redzone Active" : "Redzone Inactive");
526 restore_bytes(s
, "redzone", red
,
527 endobject
, object
+ s
->inuse
);
531 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
&&
532 !check_bytes(endobject
, POISON_INUSE
,
533 s
->inuse
- s
->objsize
)) {
534 object_err(s
, page
, p
, "Alignment padding check fails");
536 * Fix it so that there will not be another report.
538 * Hmmm... We may be corrupting an object that now expects
539 * to be longer than allowed.
541 restore_bytes(s
, "alignment padding", POISON_INUSE
,
542 endobject
, object
+ s
->inuse
);
546 if (s
->flags
& SLAB_POISON
) {
547 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
548 (!check_bytes(p
, POISON_FREE
, s
->objsize
- 1) ||
549 p
[s
->objsize
- 1] != POISON_END
)) {
551 object_err(s
, page
, p
, "Poison check failed");
552 restore_bytes(s
, "Poison", POISON_FREE
,
553 p
, p
+ s
->objsize
-1);
554 restore_bytes(s
, "Poison", POISON_END
,
555 p
+ s
->objsize
- 1, p
+ s
->objsize
);
559 * check_pad_bytes cleans up on its own.
561 check_pad_bytes(s
, page
, p
);
564 if (!s
->offset
&& active
)
566 * Object and freepointer overlap. Cannot check
567 * freepointer while object is allocated.
571 /* Check free pointer validity */
572 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
573 object_err(s
, page
, p
, "Freepointer corrupt");
575 * No choice but to zap it and thus loose the remainder
576 * of the free objects in this slab. May cause
577 * another error because the object count maybe
580 set_freepointer(s
, p
, NULL
);
586 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
588 VM_BUG_ON(!irqs_disabled());
590 if (!PageSlab(page
)) {
591 slab_err(s
, page
, "Not a valid slab page flags=%lx "
592 "mapping=0x%p count=%d", page
->flags
, page
->mapping
,
596 if (page
->offset
* sizeof(void *) != s
->offset
) {
597 slab_err(s
, page
, "Corrupted offset %lu flags=0x%lx "
598 "mapping=0x%p count=%d",
599 (unsigned long)(page
->offset
* sizeof(void *)),
605 if (page
->inuse
> s
->objects
) {
606 slab_err(s
, page
, "inuse %u > max %u @0x%p flags=%lx "
607 "mapping=0x%p count=%d",
608 s
->name
, page
->inuse
, s
->objects
, page
->flags
,
609 page
->mapping
, page_count(page
));
612 /* Slab_pad_check fixes things up after itself */
613 slab_pad_check(s
, page
);
618 * Determine if a certain object on a page is on the freelist and
619 * therefore free. Must hold the slab lock for cpu slabs to
620 * guarantee that the chains are consistent.
622 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
625 void *fp
= page
->freelist
;
628 while (fp
&& nr
<= s
->objects
) {
631 if (!check_valid_pointer(s
, page
, fp
)) {
633 object_err(s
, page
, object
,
634 "Freechain corrupt");
635 set_freepointer(s
, object
, NULL
);
638 slab_err(s
, page
, "Freepointer 0x%p corrupt",
640 page
->freelist
= NULL
;
641 page
->inuse
= s
->objects
;
642 printk(KERN_ERR
"@@@ SLUB %s: Freelist "
643 "cleared. Slab 0x%p\n",
650 fp
= get_freepointer(s
, object
);
654 if (page
->inuse
!= s
->objects
- nr
) {
655 slab_err(s
, page
, "Wrong object count. Counter is %d but "
656 "counted were %d", s
, page
, page
->inuse
,
658 page
->inuse
= s
->objects
- nr
;
659 printk(KERN_ERR
"@@@ SLUB %s: Object count adjusted. "
660 "Slab @0x%p\n", s
->name
, page
);
662 return search
== NULL
;
666 * Tracking of fully allocated slabs for debugging
668 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
670 spin_lock(&n
->list_lock
);
671 list_add(&page
->lru
, &n
->full
);
672 spin_unlock(&n
->list_lock
);
675 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
677 struct kmem_cache_node
*n
;
679 if (!(s
->flags
& SLAB_STORE_USER
))
682 n
= get_node(s
, page_to_nid(page
));
684 spin_lock(&n
->list_lock
);
685 list_del(&page
->lru
);
686 spin_unlock(&n
->list_lock
);
689 static int alloc_object_checks(struct kmem_cache
*s
, struct page
*page
,
692 if (!check_slab(s
, page
))
695 if (object
&& !on_freelist(s
, page
, object
)) {
696 slab_err(s
, page
, "Object 0x%p already allocated", object
);
700 if (!check_valid_pointer(s
, page
, object
)) {
701 object_err(s
, page
, object
, "Freelist Pointer check fails");
708 if (!check_object(s
, page
, object
, 0))
713 if (PageSlab(page
)) {
715 * If this is a slab page then lets do the best we can
716 * to avoid issues in the future. Marking all objects
717 * as used avoids touching the remainder.
719 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
721 page
->inuse
= s
->objects
;
722 page
->freelist
= NULL
;
723 /* Fix up fields that may be corrupted */
724 page
->offset
= s
->offset
/ sizeof(void *);
729 static int free_object_checks(struct kmem_cache
*s
, struct page
*page
,
732 if (!check_slab(s
, page
))
735 if (!check_valid_pointer(s
, page
, object
)) {
736 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
740 if (on_freelist(s
, page
, object
)) {
741 slab_err(s
, page
, "Object 0x%p already free", object
);
745 if (!check_object(s
, page
, object
, 1))
748 if (unlikely(s
!= page
->slab
)) {
750 slab_err(s
, page
, "Attempt to free object(0x%p) "
751 "outside of slab", object
);
755 "SLUB <none>: no slab for object 0x%p.\n",
760 slab_err(s
, page
, "object at 0x%p belongs "
761 "to slab %s", object
, page
->slab
->name
);
766 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
767 s
->name
, page
, object
);
772 * Slab allocation and freeing
774 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
777 int pages
= 1 << s
->order
;
782 if (s
->flags
& SLAB_CACHE_DMA
)
786 page
= alloc_pages(flags
, s
->order
);
788 page
= alloc_pages_node(node
, flags
, s
->order
);
793 mod_zone_page_state(page_zone(page
),
794 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
795 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
801 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
804 if (PageError(page
)) {
805 init_object(s
, object
, 0);
806 init_tracking(s
, object
);
809 if (unlikely(s
->ctor
))
810 s
->ctor(object
, s
, SLAB_CTOR_CONSTRUCTOR
);
813 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
816 struct kmem_cache_node
*n
;
822 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
));
824 if (flags
& __GFP_WAIT
)
827 page
= allocate_slab(s
, flags
& GFP_LEVEL_MASK
, node
);
831 n
= get_node(s
, page_to_nid(page
));
833 atomic_long_inc(&n
->nr_slabs
);
834 page
->offset
= s
->offset
/ sizeof(void *);
836 page
->flags
|= 1 << PG_slab
;
837 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
838 SLAB_STORE_USER
| SLAB_TRACE
))
839 page
->flags
|= 1 << PG_error
;
841 start
= page_address(page
);
842 end
= start
+ s
->objects
* s
->size
;
844 if (unlikely(s
->flags
& SLAB_POISON
))
845 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
848 for (p
= start
+ s
->size
; p
< end
; p
+= s
->size
) {
849 setup_object(s
, page
, last
);
850 set_freepointer(s
, last
, p
);
853 setup_object(s
, page
, last
);
854 set_freepointer(s
, last
, NULL
);
856 page
->freelist
= start
;
859 if (flags
& __GFP_WAIT
)
864 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
866 int pages
= 1 << s
->order
;
868 if (unlikely(PageError(page
) || s
->dtor
)) {
869 void *start
= page_address(page
);
870 void *end
= start
+ (pages
<< PAGE_SHIFT
);
873 slab_pad_check(s
, page
);
874 for (p
= start
; p
<= end
- s
->size
; p
+= s
->size
) {
877 check_object(s
, page
, p
, 0);
881 mod_zone_page_state(page_zone(page
),
882 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
883 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
886 page
->mapping
= NULL
;
887 __free_pages(page
, s
->order
);
890 static void rcu_free_slab(struct rcu_head
*h
)
894 page
= container_of((struct list_head
*)h
, struct page
, lru
);
895 __free_slab(page
->slab
, page
);
898 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
900 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
902 * RCU free overloads the RCU head over the LRU
904 struct rcu_head
*head
= (void *)&page
->lru
;
906 call_rcu(head
, rcu_free_slab
);
908 __free_slab(s
, page
);
911 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
913 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
915 atomic_long_dec(&n
->nr_slabs
);
916 reset_page_mapcount(page
);
917 page
->flags
&= ~(1 << PG_slab
| 1 << PG_error
);
922 * Per slab locking using the pagelock
924 static __always_inline
void slab_lock(struct page
*page
)
926 bit_spin_lock(PG_locked
, &page
->flags
);
929 static __always_inline
void slab_unlock(struct page
*page
)
931 bit_spin_unlock(PG_locked
, &page
->flags
);
934 static __always_inline
int slab_trylock(struct page
*page
)
938 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
943 * Management of partially allocated slabs
945 static void add_partial_tail(struct kmem_cache_node
*n
, struct page
*page
)
947 spin_lock(&n
->list_lock
);
949 list_add_tail(&page
->lru
, &n
->partial
);
950 spin_unlock(&n
->list_lock
);
953 static void add_partial(struct kmem_cache_node
*n
, struct page
*page
)
955 spin_lock(&n
->list_lock
);
957 list_add(&page
->lru
, &n
->partial
);
958 spin_unlock(&n
->list_lock
);
961 static void remove_partial(struct kmem_cache
*s
,
964 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
966 spin_lock(&n
->list_lock
);
967 list_del(&page
->lru
);
969 spin_unlock(&n
->list_lock
);
973 * Lock page and remove it from the partial list
975 * Must hold list_lock
977 static int lock_and_del_slab(struct kmem_cache_node
*n
, struct page
*page
)
979 if (slab_trylock(page
)) {
980 list_del(&page
->lru
);
988 * Try to get a partial slab from a specific node
990 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
995 * Racy check. If we mistakenly see no partial slabs then we
996 * just allocate an empty slab. If we mistakenly try to get a
997 * partial slab then get_partials() will return NULL.
999 if (!n
|| !n
->nr_partial
)
1002 spin_lock(&n
->list_lock
);
1003 list_for_each_entry(page
, &n
->partial
, lru
)
1004 if (lock_and_del_slab(n
, page
))
1008 spin_unlock(&n
->list_lock
);
1013 * Get a page from somewhere. Search in increasing NUMA
1016 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1019 struct zonelist
*zonelist
;
1024 * The defrag ratio allows to configure the tradeoffs between
1025 * inter node defragmentation and node local allocations.
1026 * A lower defrag_ratio increases the tendency to do local
1027 * allocations instead of scanning throught the partial
1028 * lists on other nodes.
1030 * If defrag_ratio is set to 0 then kmalloc() always
1031 * returns node local objects. If its higher then kmalloc()
1032 * may return off node objects in order to avoid fragmentation.
1034 * A higher ratio means slabs may be taken from other nodes
1035 * thus reducing the number of partial slabs on those nodes.
1037 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1038 * defrag_ratio = 1000) then every (well almost) allocation
1039 * will first attempt to defrag slab caches on other nodes. This
1040 * means scanning over all nodes to look for partial slabs which
1041 * may be a bit expensive to do on every slab allocation.
1043 if (!s
->defrag_ratio
|| get_cycles() % 1024 > s
->defrag_ratio
)
1046 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
1047 ->node_zonelists
[gfp_zone(flags
)];
1048 for (z
= zonelist
->zones
; *z
; z
++) {
1049 struct kmem_cache_node
*n
;
1051 n
= get_node(s
, zone_to_nid(*z
));
1053 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1054 n
->nr_partial
> MIN_PARTIAL
) {
1055 page
= get_partial_node(n
);
1065 * Get a partial page, lock it and return it.
1067 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1070 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1072 page
= get_partial_node(get_node(s
, searchnode
));
1073 if (page
|| (flags
& __GFP_THISNODE
))
1076 return get_any_partial(s
, flags
);
1080 * Move a page back to the lists.
1082 * Must be called with the slab lock held.
1084 * On exit the slab lock will have been dropped.
1086 static void putback_slab(struct kmem_cache
*s
, struct page
*page
)
1088 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1093 add_partial(n
, page
);
1094 else if (PageError(page
) && (s
->flags
& SLAB_STORE_USER
))
1099 if (n
->nr_partial
< MIN_PARTIAL
) {
1101 * Adding an empty page to the partial slabs in order
1102 * to avoid page allocator overhead. This page needs to
1103 * come after all the others that are not fully empty
1104 * in order to make sure that we do maximum
1107 add_partial_tail(n
, page
);
1111 discard_slab(s
, page
);
1117 * Remove the cpu slab
1119 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1121 s
->cpu_slab
[cpu
] = NULL
;
1122 ClearPageActive(page
);
1124 putback_slab(s
, page
);
1127 static void flush_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1130 deactivate_slab(s
, page
, cpu
);
1135 * Called from IPI handler with interrupts disabled.
1137 static void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1139 struct page
*page
= s
->cpu_slab
[cpu
];
1142 flush_slab(s
, page
, cpu
);
1145 static void flush_cpu_slab(void *d
)
1147 struct kmem_cache
*s
= d
;
1148 int cpu
= smp_processor_id();
1150 __flush_cpu_slab(s
, cpu
);
1153 static void flush_all(struct kmem_cache
*s
)
1156 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1158 unsigned long flags
;
1160 local_irq_save(flags
);
1162 local_irq_restore(flags
);
1167 * slab_alloc is optimized to only modify two cachelines on the fast path
1168 * (aside from the stack):
1170 * 1. The page struct
1171 * 2. The first cacheline of the object to be allocated.
1173 * The only cache lines that are read (apart from code) is the
1174 * per cpu array in the kmem_cache struct.
1176 * Fastpath is not possible if we need to get a new slab or have
1177 * debugging enabled (which means all slabs are marked with PageError)
1179 static void *slab_alloc(struct kmem_cache
*s
,
1180 gfp_t gfpflags
, int node
, void *addr
)
1184 unsigned long flags
;
1187 local_irq_save(flags
);
1188 cpu
= smp_processor_id();
1189 page
= s
->cpu_slab
[cpu
];
1194 if (unlikely(node
!= -1 && page_to_nid(page
) != node
))
1197 object
= page
->freelist
;
1198 if (unlikely(!object
))
1200 if (unlikely(PageError(page
)))
1205 page
->freelist
= object
[page
->offset
];
1207 local_irq_restore(flags
);
1211 deactivate_slab(s
, page
, cpu
);
1214 page
= get_partial(s
, gfpflags
, node
);
1217 s
->cpu_slab
[cpu
] = page
;
1218 SetPageActive(page
);
1222 page
= new_slab(s
, gfpflags
, node
);
1224 cpu
= smp_processor_id();
1225 if (s
->cpu_slab
[cpu
]) {
1227 * Someone else populated the cpu_slab while we enabled
1228 * interrupts, or we have got scheduled on another cpu.
1229 * The page may not be on the requested node.
1232 page_to_nid(s
->cpu_slab
[cpu
]) == node
) {
1234 * Current cpuslab is acceptable and we
1235 * want the current one since its cache hot
1237 discard_slab(s
, page
);
1238 page
= s
->cpu_slab
[cpu
];
1242 /* Dump the current slab */
1243 flush_slab(s
, s
->cpu_slab
[cpu
], cpu
);
1248 local_irq_restore(flags
);
1251 if (!alloc_object_checks(s
, page
, object
))
1253 if (s
->flags
& SLAB_STORE_USER
)
1254 set_track(s
, object
, TRACK_ALLOC
, addr
);
1255 if (s
->flags
& SLAB_TRACE
) {
1256 printk(KERN_INFO
"TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
1257 s
->name
, object
, page
->inuse
,
1261 init_object(s
, object
, 1);
1265 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1267 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1269 EXPORT_SYMBOL(kmem_cache_alloc
);
1272 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1274 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1276 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1280 * The fastpath only writes the cacheline of the page struct and the first
1281 * cacheline of the object.
1283 * No special cachelines need to be read
1285 static void slab_free(struct kmem_cache
*s
, struct page
*page
,
1286 void *x
, void *addr
)
1289 void **object
= (void *)x
;
1290 unsigned long flags
;
1292 local_irq_save(flags
);
1295 if (unlikely(PageError(page
)))
1298 prior
= object
[page
->offset
] = page
->freelist
;
1299 page
->freelist
= object
;
1302 if (unlikely(PageActive(page
)))
1304 * Cpu slabs are never on partial lists and are
1309 if (unlikely(!page
->inuse
))
1313 * Objects left in the slab. If it
1314 * was not on the partial list before
1317 if (unlikely(!prior
))
1318 add_partial(get_node(s
, page_to_nid(page
)), page
);
1322 local_irq_restore(flags
);
1328 * Slab on the partial list.
1330 remove_partial(s
, page
);
1333 discard_slab(s
, page
);
1334 local_irq_restore(flags
);
1338 if (!free_object_checks(s
, page
, x
))
1340 if (!PageActive(page
) && !page
->freelist
)
1341 remove_full(s
, page
);
1342 if (s
->flags
& SLAB_STORE_USER
)
1343 set_track(s
, x
, TRACK_FREE
, addr
);
1344 if (s
->flags
& SLAB_TRACE
) {
1345 printk(KERN_INFO
"TRACE %s free 0x%p inuse=%d fp=0x%p\n",
1346 s
->name
, object
, page
->inuse
,
1348 print_section("Object", (void *)object
, s
->objsize
);
1351 init_object(s
, object
, 0);
1355 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1359 page
= virt_to_head_page(x
);
1361 slab_free(s
, page
, x
, __builtin_return_address(0));
1363 EXPORT_SYMBOL(kmem_cache_free
);
1365 /* Figure out on which slab object the object resides */
1366 static struct page
*get_object_page(const void *x
)
1368 struct page
*page
= virt_to_head_page(x
);
1370 if (!PageSlab(page
))
1377 * kmem_cache_open produces objects aligned at "size" and the first object
1378 * is placed at offset 0 in the slab (We have no metainformation on the
1379 * slab, all slabs are in essence "off slab").
1381 * In order to get the desired alignment one just needs to align the
1384 * Notice that the allocation order determines the sizes of the per cpu
1385 * caches. Each processor has always one slab available for allocations.
1386 * Increasing the allocation order reduces the number of times that slabs
1387 * must be moved on and off the partial lists and therefore may influence
1390 * The offset is used to relocate the free list link in each object. It is
1391 * therefore possible to move the free list link behind the object. This
1392 * is necessary for RCU to work properly and also useful for debugging.
1396 * Mininum / Maximum order of slab pages. This influences locking overhead
1397 * and slab fragmentation. A higher order reduces the number of partial slabs
1398 * and increases the number of allocations possible without having to
1399 * take the list_lock.
1401 static int slub_min_order
;
1402 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1405 * Minimum number of objects per slab. This is necessary in order to
1406 * reduce locking overhead. Similar to the queue size in SLAB.
1408 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1411 * Merge control. If this is set then no merging of slab caches will occur.
1413 static int slub_nomerge
;
1418 static int slub_debug
;
1420 static char *slub_debug_slabs
;
1423 * Calculate the order of allocation given an slab object size.
1425 * The order of allocation has significant impact on other elements
1426 * of the system. Generally order 0 allocations should be preferred
1427 * since they do not cause fragmentation in the page allocator. Larger
1428 * objects may have problems with order 0 because there may be too much
1429 * space left unused in a slab. We go to a higher order if more than 1/8th
1430 * of the slab would be wasted.
1432 * In order to reach satisfactory performance we must ensure that
1433 * a minimum number of objects is in one slab. Otherwise we may
1434 * generate too much activity on the partial lists. This is less a
1435 * concern for large slabs though. slub_max_order specifies the order
1436 * where we begin to stop considering the number of objects in a slab.
1438 * Higher order allocations also allow the placement of more objects
1439 * in a slab and thereby reduce object handling overhead. If the user
1440 * has requested a higher mininum order then we start with that one
1443 static int calculate_order(int size
)
1448 for (order
= max(slub_min_order
, fls(size
- 1) - PAGE_SHIFT
);
1449 order
< MAX_ORDER
; order
++) {
1450 unsigned long slab_size
= PAGE_SIZE
<< order
;
1452 if (slub_max_order
> order
&&
1453 slab_size
< slub_min_objects
* size
)
1456 if (slab_size
< size
)
1459 rem
= slab_size
% size
;
1461 if (rem
<= (PAGE_SIZE
<< order
) / 8)
1465 if (order
>= MAX_ORDER
)
1471 * Function to figure out which alignment to use from the
1472 * various ways of specifying it.
1474 static unsigned long calculate_alignment(unsigned long flags
,
1475 unsigned long align
, unsigned long size
)
1478 * If the user wants hardware cache aligned objects then
1479 * follow that suggestion if the object is sufficiently
1482 * The hardware cache alignment cannot override the
1483 * specified alignment though. If that is greater
1486 if ((flags
& SLAB_HWCACHE_ALIGN
) &&
1487 size
> cache_line_size() / 2)
1488 return max_t(unsigned long, align
, cache_line_size());
1490 if (align
< ARCH_SLAB_MINALIGN
)
1491 return ARCH_SLAB_MINALIGN
;
1493 return ALIGN(align
, sizeof(void *));
1496 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1499 atomic_long_set(&n
->nr_slabs
, 0);
1500 spin_lock_init(&n
->list_lock
);
1501 INIT_LIST_HEAD(&n
->partial
);
1502 INIT_LIST_HEAD(&n
->full
);
1507 * No kmalloc_node yet so do it by hand. We know that this is the first
1508 * slab on the node for this slabcache. There are no concurrent accesses
1511 * Note that this function only works on the kmalloc_node_cache
1512 * when allocating for the kmalloc_node_cache.
1514 static struct kmem_cache_node
* __init
early_kmem_cache_node_alloc(gfp_t gfpflags
,
1518 struct kmem_cache_node
*n
;
1520 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
1522 page
= new_slab(kmalloc_caches
, gfpflags
| GFP_THISNODE
, node
);
1523 /* new_slab() disables interupts */
1529 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
1531 kmalloc_caches
->node
[node
] = n
;
1532 init_object(kmalloc_caches
, n
, 1);
1533 init_kmem_cache_node(n
);
1534 atomic_long_inc(&n
->nr_slabs
);
1535 add_partial(n
, page
);
1539 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1543 for_each_online_node(node
) {
1544 struct kmem_cache_node
*n
= s
->node
[node
];
1545 if (n
&& n
!= &s
->local_node
)
1546 kmem_cache_free(kmalloc_caches
, n
);
1547 s
->node
[node
] = NULL
;
1551 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1556 if (slab_state
>= UP
)
1557 local_node
= page_to_nid(virt_to_page(s
));
1561 for_each_online_node(node
) {
1562 struct kmem_cache_node
*n
;
1564 if (local_node
== node
)
1567 if (slab_state
== DOWN
) {
1568 n
= early_kmem_cache_node_alloc(gfpflags
,
1572 n
= kmem_cache_alloc_node(kmalloc_caches
,
1576 free_kmem_cache_nodes(s
);
1582 init_kmem_cache_node(n
);
1587 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1591 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1593 init_kmem_cache_node(&s
->local_node
);
1599 * calculate_sizes() determines the order and the distribution of data within
1602 static int calculate_sizes(struct kmem_cache
*s
)
1604 unsigned long flags
= s
->flags
;
1605 unsigned long size
= s
->objsize
;
1606 unsigned long align
= s
->align
;
1609 * Determine if we can poison the object itself. If the user of
1610 * the slab may touch the object after free or before allocation
1611 * then we should never poison the object itself.
1613 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
1614 !s
->ctor
&& !s
->dtor
)
1615 s
->flags
|= __OBJECT_POISON
;
1617 s
->flags
&= ~__OBJECT_POISON
;
1620 * Round up object size to the next word boundary. We can only
1621 * place the free pointer at word boundaries and this determines
1622 * the possible location of the free pointer.
1624 size
= ALIGN(size
, sizeof(void *));
1627 * If we are redzoning then check if there is some space between the
1628 * end of the object and the free pointer. If not then add an
1629 * additional word, so that we can establish a redzone between
1630 * the object and the freepointer to be able to check for overwrites.
1632 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
1633 size
+= sizeof(void *);
1636 * With that we have determined how much of the slab is in actual
1637 * use by the object. This is the potential offset to the free
1642 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
1643 s
->ctor
|| s
->dtor
)) {
1645 * Relocate free pointer after the object if it is not
1646 * permitted to overwrite the first word of the object on
1649 * This is the case if we do RCU, have a constructor or
1650 * destructor or are poisoning the objects.
1653 size
+= sizeof(void *);
1656 if (flags
& SLAB_STORE_USER
)
1658 * Need to store information about allocs and frees after
1661 size
+= 2 * sizeof(struct track
);
1663 if (flags
& SLAB_RED_ZONE
)
1665 * Add some empty padding so that we can catch
1666 * overwrites from earlier objects rather than let
1667 * tracking information or the free pointer be
1668 * corrupted if an user writes before the start
1671 size
+= sizeof(void *);
1673 * Determine the alignment based on various parameters that the
1674 * user specified and the dynamic determination of cache line size
1677 align
= calculate_alignment(flags
, align
, s
->objsize
);
1680 * SLUB stores one object immediately after another beginning from
1681 * offset 0. In order to align the objects we have to simply size
1682 * each object to conform to the alignment.
1684 size
= ALIGN(size
, align
);
1687 s
->order
= calculate_order(size
);
1692 * Determine the number of objects per slab
1694 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
1697 * Verify that the number of objects is within permitted limits.
1698 * The page->inuse field is only 16 bit wide! So we cannot have
1699 * more than 64k objects per slab.
1701 if (!s
->objects
|| s
->objects
> 65535)
1707 static int __init
finish_bootstrap(void)
1709 struct list_head
*h
;
1714 list_for_each(h
, &slab_caches
) {
1715 struct kmem_cache
*s
=
1716 container_of(h
, struct kmem_cache
, list
);
1718 err
= sysfs_slab_add(s
);
1724 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
1725 const char *name
, size_t size
,
1726 size_t align
, unsigned long flags
,
1727 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
1728 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
1730 memset(s
, 0, kmem_size
);
1739 * The page->offset field is only 16 bit wide. This is an offset
1740 * in units of words from the beginning of an object. If the slab
1741 * size is bigger then we cannot move the free pointer behind the
1744 * On 32 bit platforms the limit is 256k. On 64bit platforms
1745 * the limit is 512k.
1747 * Debugging or ctor/dtors may create a need to move the free
1748 * pointer. Fail if this happens.
1750 if (s
->size
>= 65535 * sizeof(void *)) {
1751 BUG_ON(flags
& (SLAB_RED_ZONE
| SLAB_POISON
|
1752 SLAB_STORE_USER
| SLAB_DESTROY_BY_RCU
));
1753 BUG_ON(ctor
|| dtor
);
1757 * Enable debugging if selected on the kernel commandline.
1759 if (slub_debug
&& (!slub_debug_slabs
||
1760 strncmp(slub_debug_slabs
, name
,
1761 strlen(slub_debug_slabs
)) == 0))
1762 s
->flags
|= slub_debug
;
1764 if (!calculate_sizes(s
))
1769 s
->defrag_ratio
= 100;
1772 if (init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
1775 if (flags
& SLAB_PANIC
)
1776 panic("Cannot create slab %s size=%lu realsize=%u "
1777 "order=%u offset=%u flags=%lx\n",
1778 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
1782 EXPORT_SYMBOL(kmem_cache_open
);
1785 * Check if a given pointer is valid
1787 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
1792 page
= get_object_page(object
);
1794 if (!page
|| s
!= page
->slab
)
1795 /* No slab or wrong slab */
1798 if (!check_valid_pointer(s
, page
, object
))
1802 * We could also check if the object is on the slabs freelist.
1803 * But this would be too expensive and it seems that the main
1804 * purpose of kmem_ptr_valid is to check if the object belongs
1805 * to a certain slab.
1809 EXPORT_SYMBOL(kmem_ptr_validate
);
1812 * Determine the size of a slab object
1814 unsigned int kmem_cache_size(struct kmem_cache
*s
)
1818 EXPORT_SYMBOL(kmem_cache_size
);
1820 const char *kmem_cache_name(struct kmem_cache
*s
)
1824 EXPORT_SYMBOL(kmem_cache_name
);
1827 * Attempt to free all slabs on a node
1829 static int free_list(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1830 struct list_head
*list
)
1832 int slabs_inuse
= 0;
1833 unsigned long flags
;
1834 struct page
*page
, *h
;
1836 spin_lock_irqsave(&n
->list_lock
, flags
);
1837 list_for_each_entry_safe(page
, h
, list
, lru
)
1839 list_del(&page
->lru
);
1840 discard_slab(s
, page
);
1843 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1848 * Release all resources used by slab cache
1850 static int kmem_cache_close(struct kmem_cache
*s
)
1856 /* Attempt to free all objects */
1857 for_each_online_node(node
) {
1858 struct kmem_cache_node
*n
= get_node(s
, node
);
1860 n
->nr_partial
-= free_list(s
, n
, &n
->partial
);
1861 if (atomic_long_read(&n
->nr_slabs
))
1864 free_kmem_cache_nodes(s
);
1869 * Close a cache and release the kmem_cache structure
1870 * (must be used for caches created using kmem_cache_create)
1872 void kmem_cache_destroy(struct kmem_cache
*s
)
1874 down_write(&slub_lock
);
1878 if (kmem_cache_close(s
))
1880 sysfs_slab_remove(s
);
1883 up_write(&slub_lock
);
1885 EXPORT_SYMBOL(kmem_cache_destroy
);
1887 /********************************************************************
1889 *******************************************************************/
1891 struct kmem_cache kmalloc_caches
[KMALLOC_SHIFT_HIGH
+ 1] __cacheline_aligned
;
1892 EXPORT_SYMBOL(kmalloc_caches
);
1894 #ifdef CONFIG_ZONE_DMA
1895 static struct kmem_cache
*kmalloc_caches_dma
[KMALLOC_SHIFT_HIGH
+ 1];
1898 static int __init
setup_slub_min_order(char *str
)
1900 get_option (&str
, &slub_min_order
);
1905 __setup("slub_min_order=", setup_slub_min_order
);
1907 static int __init
setup_slub_max_order(char *str
)
1909 get_option (&str
, &slub_max_order
);
1914 __setup("slub_max_order=", setup_slub_max_order
);
1916 static int __init
setup_slub_min_objects(char *str
)
1918 get_option (&str
, &slub_min_objects
);
1923 __setup("slub_min_objects=", setup_slub_min_objects
);
1925 static int __init
setup_slub_nomerge(char *str
)
1931 __setup("slub_nomerge", setup_slub_nomerge
);
1933 static int __init
setup_slub_debug(char *str
)
1935 if (!str
|| *str
!= '=')
1936 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1939 if (*str
== 0 || *str
== ',')
1940 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1942 for( ;*str
&& *str
!= ','; str
++)
1944 case 'f' : case 'F' :
1945 slub_debug
|= SLAB_DEBUG_FREE
;
1947 case 'z' : case 'Z' :
1948 slub_debug
|= SLAB_RED_ZONE
;
1950 case 'p' : case 'P' :
1951 slub_debug
|= SLAB_POISON
;
1953 case 'u' : case 'U' :
1954 slub_debug
|= SLAB_STORE_USER
;
1956 case 't' : case 'T' :
1957 slub_debug
|= SLAB_TRACE
;
1960 printk(KERN_ERR
"slub_debug option '%c' "
1961 "unknown. skipped\n",*str
);
1966 slub_debug_slabs
= str
+ 1;
1970 __setup("slub_debug", setup_slub_debug
);
1972 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
1973 const char *name
, int size
, gfp_t gfp_flags
)
1975 unsigned int flags
= 0;
1977 if (gfp_flags
& SLUB_DMA
)
1978 flags
= SLAB_CACHE_DMA
;
1980 down_write(&slub_lock
);
1981 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
1985 list_add(&s
->list
, &slab_caches
);
1986 up_write(&slub_lock
);
1987 if (sysfs_slab_add(s
))
1992 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
1995 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
1997 int index
= kmalloc_index(size
);
2002 /* Allocation too large? */
2005 #ifdef CONFIG_ZONE_DMA
2006 if ((flags
& SLUB_DMA
)) {
2007 struct kmem_cache
*s
;
2008 struct kmem_cache
*x
;
2012 s
= kmalloc_caches_dma
[index
];
2016 /* Dynamically create dma cache */
2017 x
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2019 panic("Unable to allocate memory for dma cache\n");
2021 if (index
<= KMALLOC_SHIFT_HIGH
)
2022 realsize
= 1 << index
;
2030 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d",
2031 (unsigned int)realsize
);
2032 s
= create_kmalloc_cache(x
, text
, realsize
, flags
);
2033 kmalloc_caches_dma
[index
] = s
;
2037 return &kmalloc_caches
[index
];
2040 void *__kmalloc(size_t size
, gfp_t flags
)
2042 struct kmem_cache
*s
= get_slab(size
, flags
);
2045 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2048 EXPORT_SYMBOL(__kmalloc
);
2051 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2053 struct kmem_cache
*s
= get_slab(size
, flags
);
2056 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2059 EXPORT_SYMBOL(__kmalloc_node
);
2062 size_t ksize(const void *object
)
2064 struct page
*page
= get_object_page(object
);
2065 struct kmem_cache
*s
;
2072 * Debugging requires use of the padding between object
2073 * and whatever may come after it.
2075 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2079 * If we have the need to store the freelist pointer
2080 * back there or track user information then we can
2081 * only use the space before that information.
2083 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2087 * Else we can use all the padding etc for the allocation
2091 EXPORT_SYMBOL(ksize
);
2093 void kfree(const void *x
)
2095 struct kmem_cache
*s
;
2101 page
= virt_to_head_page(x
);
2104 slab_free(s
, page
, (void *)x
, __builtin_return_address(0));
2106 EXPORT_SYMBOL(kfree
);
2109 * kmem_cache_shrink removes empty slabs from the partial lists
2110 * and then sorts the partially allocated slabs by the number
2111 * of items in use. The slabs with the most items in use
2112 * come first. New allocations will remove these from the
2113 * partial list because they are full. The slabs with the
2114 * least items are placed last. If it happens that the objects
2115 * are freed then the page can be returned to the page allocator.
2117 int kmem_cache_shrink(struct kmem_cache
*s
)
2121 struct kmem_cache_node
*n
;
2124 struct list_head
*slabs_by_inuse
=
2125 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2126 unsigned long flags
;
2128 if (!slabs_by_inuse
)
2132 for_each_online_node(node
) {
2133 n
= get_node(s
, node
);
2138 for (i
= 0; i
< s
->objects
; i
++)
2139 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2141 spin_lock_irqsave(&n
->list_lock
, flags
);
2144 * Build lists indexed by the items in use in
2145 * each slab or free slabs if empty.
2147 * Note that concurrent frees may occur while
2148 * we hold the list_lock. page->inuse here is
2151 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2152 if (!page
->inuse
&& slab_trylock(page
)) {
2154 * Must hold slab lock here because slab_free
2155 * may have freed the last object and be
2156 * waiting to release the slab.
2158 list_del(&page
->lru
);
2161 discard_slab(s
, page
);
2163 if (n
->nr_partial
> MAX_PARTIAL
)
2164 list_move(&page
->lru
,
2165 slabs_by_inuse
+ page
->inuse
);
2169 if (n
->nr_partial
<= MAX_PARTIAL
)
2173 * Rebuild the partial list with the slabs filled up
2174 * most first and the least used slabs at the end.
2176 for (i
= s
->objects
- 1; i
>= 0; i
--)
2177 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2180 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2183 kfree(slabs_by_inuse
);
2186 EXPORT_SYMBOL(kmem_cache_shrink
);
2189 * krealloc - reallocate memory. The contents will remain unchanged.
2191 * @p: object to reallocate memory for.
2192 * @new_size: how many bytes of memory are required.
2193 * @flags: the type of memory to allocate.
2195 * The contents of the object pointed to are preserved up to the
2196 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2197 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2198 * %NULL pointer, the object pointed to is freed.
2200 void *krealloc(const void *p
, size_t new_size
, gfp_t flags
)
2202 struct kmem_cache
*new_cache
;
2207 return kmalloc(new_size
, flags
);
2209 if (unlikely(!new_size
)) {
2214 page
= virt_to_head_page(p
);
2216 new_cache
= get_slab(new_size
, flags
);
2219 * If new size fits in the current cache, bail out.
2221 if (likely(page
->slab
== new_cache
))
2224 ret
= kmalloc(new_size
, flags
);
2226 memcpy(ret
, p
, min(new_size
, ksize(p
)));
2231 EXPORT_SYMBOL(krealloc
);
2233 /********************************************************************
2234 * Basic setup of slabs
2235 *******************************************************************/
2237 void __init
kmem_cache_init(void)
2243 * Must first have the slab cache available for the allocations of the
2244 * struct kmalloc_cache_node's. There is special bootstrap code in
2245 * kmem_cache_open for slab_state == DOWN.
2247 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2248 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2251 /* Able to allocate the per node structures */
2252 slab_state
= PARTIAL
;
2254 /* Caches that are not of the two-to-the-power-of size */
2255 create_kmalloc_cache(&kmalloc_caches
[1],
2256 "kmalloc-96", 96, GFP_KERNEL
);
2257 create_kmalloc_cache(&kmalloc_caches
[2],
2258 "kmalloc-192", 192, GFP_KERNEL
);
2260 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2261 create_kmalloc_cache(&kmalloc_caches
[i
],
2262 "kmalloc", 1 << i
, GFP_KERNEL
);
2266 /* Provide the correct kmalloc names now that the caches are up */
2267 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2268 kmalloc_caches
[i
]. name
=
2269 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
2272 register_cpu_notifier(&slab_notifier
);
2275 if (nr_cpu_ids
) /* Remove when nr_cpu_ids is fixed upstream ! */
2276 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
)
2277 + nr_cpu_ids
* sizeof(struct page
*);
2279 printk(KERN_INFO
"SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2280 " Processors=%d, Nodes=%d\n",
2281 KMALLOC_SHIFT_HIGH
, cache_line_size(),
2282 slub_min_order
, slub_max_order
, slub_min_objects
,
2283 nr_cpu_ids
, nr_node_ids
);
2287 * Find a mergeable slab cache
2289 static int slab_unmergeable(struct kmem_cache
*s
)
2291 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
2294 if (s
->ctor
|| s
->dtor
)
2300 static struct kmem_cache
*find_mergeable(size_t size
,
2301 size_t align
, unsigned long flags
,
2302 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
2303 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
2305 struct list_head
*h
;
2307 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
2313 size
= ALIGN(size
, sizeof(void *));
2314 align
= calculate_alignment(flags
, align
, size
);
2315 size
= ALIGN(size
, align
);
2317 list_for_each(h
, &slab_caches
) {
2318 struct kmem_cache
*s
=
2319 container_of(h
, struct kmem_cache
, list
);
2321 if (slab_unmergeable(s
))
2327 if (((flags
| slub_debug
) & SLUB_MERGE_SAME
) !=
2328 (s
->flags
& SLUB_MERGE_SAME
))
2331 * Check if alignment is compatible.
2332 * Courtesy of Adrian Drzewiecki
2334 if ((s
->size
& ~(align
-1)) != s
->size
)
2337 if (s
->size
- size
>= sizeof(void *))
2345 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
2346 size_t align
, unsigned long flags
,
2347 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
2348 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
2350 struct kmem_cache
*s
;
2352 down_write(&slub_lock
);
2353 s
= find_mergeable(size
, align
, flags
, dtor
, ctor
);
2357 * Adjust the object sizes so that we clear
2358 * the complete object on kzalloc.
2360 s
->objsize
= max(s
->objsize
, (int)size
);
2361 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
2362 if (sysfs_slab_alias(s
, name
))
2365 s
= kmalloc(kmem_size
, GFP_KERNEL
);
2366 if (s
&& kmem_cache_open(s
, GFP_KERNEL
, name
,
2367 size
, align
, flags
, ctor
, dtor
)) {
2368 if (sysfs_slab_add(s
)) {
2372 list_add(&s
->list
, &slab_caches
);
2376 up_write(&slub_lock
);
2380 up_write(&slub_lock
);
2381 if (flags
& SLAB_PANIC
)
2382 panic("Cannot create slabcache %s\n", name
);
2387 EXPORT_SYMBOL(kmem_cache_create
);
2389 void *kmem_cache_zalloc(struct kmem_cache
*s
, gfp_t flags
)
2393 x
= slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2395 memset(x
, 0, s
->objsize
);
2398 EXPORT_SYMBOL(kmem_cache_zalloc
);
2401 static void for_all_slabs(void (*func
)(struct kmem_cache
*, int), int cpu
)
2403 struct list_head
*h
;
2405 down_read(&slub_lock
);
2406 list_for_each(h
, &slab_caches
) {
2407 struct kmem_cache
*s
=
2408 container_of(h
, struct kmem_cache
, list
);
2412 up_read(&slub_lock
);
2416 * Use the cpu notifier to insure that the slab are flushed
2419 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
2420 unsigned long action
, void *hcpu
)
2422 long cpu
= (long)hcpu
;
2425 case CPU_UP_CANCELED
:
2427 for_all_slabs(__flush_cpu_slab
, cpu
);
2435 static struct notifier_block __cpuinitdata slab_notifier
=
2436 { &slab_cpuup_callback
, NULL
, 0 };
2442 /*****************************************************************
2443 * Generic reaper used to support the page allocator
2444 * (the cpu slabs are reaped by a per slab workqueue).
2446 * Maybe move this to the page allocator?
2447 ****************************************************************/
2449 static DEFINE_PER_CPU(unsigned long, reap_node
);
2451 static void init_reap_node(int cpu
)
2455 node
= next_node(cpu_to_node(cpu
), node_online_map
);
2456 if (node
== MAX_NUMNODES
)
2457 node
= first_node(node_online_map
);
2459 __get_cpu_var(reap_node
) = node
;
2462 static void next_reap_node(void)
2464 int node
= __get_cpu_var(reap_node
);
2467 * Also drain per cpu pages on remote zones
2469 if (node
!= numa_node_id())
2470 drain_node_pages(node
);
2472 node
= next_node(node
, node_online_map
);
2473 if (unlikely(node
>= MAX_NUMNODES
))
2474 node
= first_node(node_online_map
);
2475 __get_cpu_var(reap_node
) = node
;
2478 #define init_reap_node(cpu) do { } while (0)
2479 #define next_reap_node(void) do { } while (0)
2482 #define REAPTIMEOUT_CPUC (2*HZ)
2485 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
2487 static void cache_reap(struct work_struct
*unused
)
2490 refresh_cpu_vm_stats(smp_processor_id());
2491 schedule_delayed_work(&__get_cpu_var(reap_work
),
2495 static void __devinit
start_cpu_timer(int cpu
)
2497 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
2500 * When this gets called from do_initcalls via cpucache_init(),
2501 * init_workqueues() has already run, so keventd will be setup
2504 if (keventd_up() && reap_work
->work
.func
== NULL
) {
2505 init_reap_node(cpu
);
2506 INIT_DELAYED_WORK(reap_work
, cache_reap
);
2507 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
2511 static int __init
cpucache_init(void)
2516 * Register the timers that drain pcp pages and update vm statistics
2518 for_each_online_cpu(cpu
)
2519 start_cpu_timer(cpu
);
2522 __initcall(cpucache_init
);
2525 #ifdef SLUB_RESILIENCY_TEST
2526 static unsigned long validate_slab_cache(struct kmem_cache
*s
);
2528 static void resiliency_test(void)
2532 printk(KERN_ERR
"SLUB resiliency testing\n");
2533 printk(KERN_ERR
"-----------------------\n");
2534 printk(KERN_ERR
"A. Corruption after allocation\n");
2536 p
= kzalloc(16, GFP_KERNEL
);
2538 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
2539 " 0x12->0x%p\n\n", p
+ 16);
2541 validate_slab_cache(kmalloc_caches
+ 4);
2543 /* Hmmm... The next two are dangerous */
2544 p
= kzalloc(32, GFP_KERNEL
);
2545 p
[32 + sizeof(void *)] = 0x34;
2546 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
2547 " 0x34 -> -0x%p\n", p
);
2548 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2550 validate_slab_cache(kmalloc_caches
+ 5);
2551 p
= kzalloc(64, GFP_KERNEL
);
2552 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
2554 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2556 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2557 validate_slab_cache(kmalloc_caches
+ 6);
2559 printk(KERN_ERR
"\nB. Corruption after free\n");
2560 p
= kzalloc(128, GFP_KERNEL
);
2563 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
2564 validate_slab_cache(kmalloc_caches
+ 7);
2566 p
= kzalloc(256, GFP_KERNEL
);
2569 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
2570 validate_slab_cache(kmalloc_caches
+ 8);
2572 p
= kzalloc(512, GFP_KERNEL
);
2575 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
2576 validate_slab_cache(kmalloc_caches
+ 9);
2579 static void resiliency_test(void) {};
2583 * These are not as efficient as kmalloc for the non debug case.
2584 * We do not have the page struct available so we have to touch one
2585 * cacheline in struct kmem_cache to check slab flags.
2587 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
2589 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2594 return slab_alloc(s
, gfpflags
, -1, caller
);
2597 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
2598 int node
, void *caller
)
2600 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2605 return slab_alloc(s
, gfpflags
, node
, caller
);
2610 static int validate_slab(struct kmem_cache
*s
, struct page
*page
)
2613 void *addr
= page_address(page
);
2614 unsigned long map
[BITS_TO_LONGS(s
->objects
)];
2616 if (!check_slab(s
, page
) ||
2617 !on_freelist(s
, page
, NULL
))
2620 /* Now we know that a valid freelist exists */
2621 bitmap_zero(map
, s
->objects
);
2623 for(p
= page
->freelist
; p
; p
= get_freepointer(s
, p
)) {
2624 set_bit((p
- addr
) / s
->size
, map
);
2625 if (!check_object(s
, page
, p
, 0))
2629 for(p
= addr
; p
< addr
+ s
->objects
* s
->size
; p
+= s
->size
)
2630 if (!test_bit((p
- addr
) / s
->size
, map
))
2631 if (!check_object(s
, page
, p
, 1))
2636 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
)
2638 if (slab_trylock(page
)) {
2639 validate_slab(s
, page
);
2642 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
2645 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
2646 if (!PageError(page
))
2647 printk(KERN_ERR
"SLUB %s: PageError not set "
2648 "on slab 0x%p\n", s
->name
, page
);
2650 if (PageError(page
))
2651 printk(KERN_ERR
"SLUB %s: PageError set on "
2652 "slab 0x%p\n", s
->name
, page
);
2656 static int validate_slab_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
2658 unsigned long count
= 0;
2660 unsigned long flags
;
2662 spin_lock_irqsave(&n
->list_lock
, flags
);
2664 list_for_each_entry(page
, &n
->partial
, lru
) {
2665 validate_slab_slab(s
, page
);
2668 if (count
!= n
->nr_partial
)
2669 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
2670 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
2672 if (!(s
->flags
& SLAB_STORE_USER
))
2675 list_for_each_entry(page
, &n
->full
, lru
) {
2676 validate_slab_slab(s
, page
);
2679 if (count
!= atomic_long_read(&n
->nr_slabs
))
2680 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
2681 "counter=%ld\n", s
->name
, count
,
2682 atomic_long_read(&n
->nr_slabs
));
2685 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2689 static unsigned long validate_slab_cache(struct kmem_cache
*s
)
2692 unsigned long count
= 0;
2695 for_each_online_node(node
) {
2696 struct kmem_cache_node
*n
= get_node(s
, node
);
2698 count
+= validate_slab_node(s
, n
);
2704 * Generate lists of locations where slabcache objects are allocated
2709 unsigned long count
;
2715 unsigned long count
;
2716 struct location
*loc
;
2719 static void free_loc_track(struct loc_track
*t
)
2722 free_pages((unsigned long)t
->loc
,
2723 get_order(sizeof(struct location
) * t
->max
));
2726 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
)
2732 max
= PAGE_SIZE
/ sizeof(struct location
);
2734 order
= get_order(sizeof(struct location
) * max
);
2736 l
= (void *)__get_free_pages(GFP_KERNEL
, order
);
2742 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
2750 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
2753 long start
, end
, pos
;
2761 pos
= start
+ (end
- start
+ 1) / 2;
2764 * There is nothing at "end". If we end up there
2765 * we need to add something to before end.
2770 caddr
= t
->loc
[pos
].addr
;
2771 if (addr
== caddr
) {
2772 t
->loc
[pos
].count
++;
2783 * Not found. Insert new tracking element
2785 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
))
2791 (t
->count
- pos
) * sizeof(struct location
));
2798 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
2799 struct page
*page
, enum track_item alloc
)
2801 void *addr
= page_address(page
);
2802 unsigned long map
[BITS_TO_LONGS(s
->objects
)];
2805 bitmap_zero(map
, s
->objects
);
2806 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
2807 set_bit((p
- addr
) / s
->size
, map
);
2809 for (p
= addr
; p
< addr
+ s
->objects
* s
->size
; p
+= s
->size
)
2810 if (!test_bit((p
- addr
) / s
->size
, map
)) {
2811 void *addr
= get_track(s
, p
, alloc
)->addr
;
2813 add_location(t
, s
, addr
);
2817 static int list_locations(struct kmem_cache
*s
, char *buf
,
2818 enum track_item alloc
)
2828 /* Push back cpu slabs */
2831 for_each_online_node(node
) {
2832 struct kmem_cache_node
*n
= get_node(s
, node
);
2833 unsigned long flags
;
2836 if (!atomic_read(&n
->nr_slabs
))
2839 spin_lock_irqsave(&n
->list_lock
, flags
);
2840 list_for_each_entry(page
, &n
->partial
, lru
)
2841 process_slab(&t
, s
, page
, alloc
);
2842 list_for_each_entry(page
, &n
->full
, lru
)
2843 process_slab(&t
, s
, page
, alloc
);
2844 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2847 for (i
= 0; i
< t
.count
; i
++) {
2848 void *addr
= t
.loc
[i
].addr
;
2850 if (n
> PAGE_SIZE
- 100)
2852 n
+= sprintf(buf
+ n
, "%7ld ", t
.loc
[i
].count
);
2854 n
+= sprint_symbol(buf
+ n
, (unsigned long)t
.loc
[i
].addr
);
2856 n
+= sprintf(buf
+ n
, "<not-available>");
2857 n
+= sprintf(buf
+ n
, "\n");
2862 n
+= sprintf(buf
, "No data\n");
2866 static unsigned long count_partial(struct kmem_cache_node
*n
)
2868 unsigned long flags
;
2869 unsigned long x
= 0;
2872 spin_lock_irqsave(&n
->list_lock
, flags
);
2873 list_for_each_entry(page
, &n
->partial
, lru
)
2875 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2879 enum slab_stat_type
{
2886 #define SO_FULL (1 << SL_FULL)
2887 #define SO_PARTIAL (1 << SL_PARTIAL)
2888 #define SO_CPU (1 << SL_CPU)
2889 #define SO_OBJECTS (1 << SL_OBJECTS)
2891 static unsigned long slab_objects(struct kmem_cache
*s
,
2892 char *buf
, unsigned long flags
)
2894 unsigned long total
= 0;
2898 unsigned long *nodes
;
2899 unsigned long *per_cpu
;
2901 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
2902 per_cpu
= nodes
+ nr_node_ids
;
2904 for_each_possible_cpu(cpu
) {
2905 struct page
*page
= s
->cpu_slab
[cpu
];
2909 node
= page_to_nid(page
);
2910 if (flags
& SO_CPU
) {
2913 if (flags
& SO_OBJECTS
)
2924 for_each_online_node(node
) {
2925 struct kmem_cache_node
*n
= get_node(s
, node
);
2927 if (flags
& SO_PARTIAL
) {
2928 if (flags
& SO_OBJECTS
)
2929 x
= count_partial(n
);
2936 if (flags
& SO_FULL
) {
2937 int full_slabs
= atomic_read(&n
->nr_slabs
)
2941 if (flags
& SO_OBJECTS
)
2942 x
= full_slabs
* s
->objects
;
2950 x
= sprintf(buf
, "%lu", total
);
2952 for_each_online_node(node
)
2954 x
+= sprintf(buf
+ x
, " N%d=%lu",
2958 return x
+ sprintf(buf
+ x
, "\n");
2961 static int any_slab_objects(struct kmem_cache
*s
)
2966 for_each_possible_cpu(cpu
)
2967 if (s
->cpu_slab
[cpu
])
2970 for_each_node(node
) {
2971 struct kmem_cache_node
*n
= get_node(s
, node
);
2973 if (n
->nr_partial
|| atomic_read(&n
->nr_slabs
))
2979 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2980 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
2982 struct slab_attribute
{
2983 struct attribute attr
;
2984 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
2985 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
2988 #define SLAB_ATTR_RO(_name) \
2989 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2991 #define SLAB_ATTR(_name) \
2992 static struct slab_attribute _name##_attr = \
2993 __ATTR(_name, 0644, _name##_show, _name##_store)
2995 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
2997 return sprintf(buf
, "%d\n", s
->size
);
2999 SLAB_ATTR_RO(slab_size
);
3001 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3003 return sprintf(buf
, "%d\n", s
->align
);
3005 SLAB_ATTR_RO(align
);
3007 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3009 return sprintf(buf
, "%d\n", s
->objsize
);
3011 SLAB_ATTR_RO(object_size
);
3013 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3015 return sprintf(buf
, "%d\n", s
->objects
);
3017 SLAB_ATTR_RO(objs_per_slab
);
3019 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3021 return sprintf(buf
, "%d\n", s
->order
);
3023 SLAB_ATTR_RO(order
);
3025 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3028 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3030 return n
+ sprintf(buf
+ n
, "\n");
3036 static ssize_t
dtor_show(struct kmem_cache
*s
, char *buf
)
3039 int n
= sprint_symbol(buf
, (unsigned long)s
->dtor
);
3041 return n
+ sprintf(buf
+ n
, "\n");
3047 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3049 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3051 SLAB_ATTR_RO(aliases
);
3053 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3055 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3057 SLAB_ATTR_RO(slabs
);
3059 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3061 return slab_objects(s
, buf
, SO_PARTIAL
);
3063 SLAB_ATTR_RO(partial
);
3065 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3067 return slab_objects(s
, buf
, SO_CPU
);
3069 SLAB_ATTR_RO(cpu_slabs
);
3071 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3073 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3075 SLAB_ATTR_RO(objects
);
3077 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3079 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3082 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3083 const char *buf
, size_t length
)
3085 s
->flags
&= ~SLAB_DEBUG_FREE
;
3087 s
->flags
|= SLAB_DEBUG_FREE
;
3090 SLAB_ATTR(sanity_checks
);
3092 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3094 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3097 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3100 s
->flags
&= ~SLAB_TRACE
;
3102 s
->flags
|= SLAB_TRACE
;
3107 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3109 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3112 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3113 const char *buf
, size_t length
)
3115 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3117 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3120 SLAB_ATTR(reclaim_account
);
3122 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3124 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3126 SLAB_ATTR_RO(hwcache_align
);
3128 #ifdef CONFIG_ZONE_DMA
3129 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3131 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3133 SLAB_ATTR_RO(cache_dma
);
3136 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3138 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3140 SLAB_ATTR_RO(destroy_by_rcu
);
3142 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3144 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3147 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3148 const char *buf
, size_t length
)
3150 if (any_slab_objects(s
))
3153 s
->flags
&= ~SLAB_RED_ZONE
;
3155 s
->flags
|= SLAB_RED_ZONE
;
3159 SLAB_ATTR(red_zone
);
3161 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3163 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3166 static ssize_t
poison_store(struct kmem_cache
*s
,
3167 const char *buf
, size_t length
)
3169 if (any_slab_objects(s
))
3172 s
->flags
&= ~SLAB_POISON
;
3174 s
->flags
|= SLAB_POISON
;
3180 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3182 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3185 static ssize_t
store_user_store(struct kmem_cache
*s
,
3186 const char *buf
, size_t length
)
3188 if (any_slab_objects(s
))
3191 s
->flags
&= ~SLAB_STORE_USER
;
3193 s
->flags
|= SLAB_STORE_USER
;
3197 SLAB_ATTR(store_user
);
3199 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3204 static ssize_t
validate_store(struct kmem_cache
*s
,
3205 const char *buf
, size_t length
)
3208 validate_slab_cache(s
);
3213 SLAB_ATTR(validate
);
3215 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3220 static ssize_t
shrink_store(struct kmem_cache
*s
,
3221 const char *buf
, size_t length
)
3223 if (buf
[0] == '1') {
3224 int rc
= kmem_cache_shrink(s
);
3234 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
3236 if (!(s
->flags
& SLAB_STORE_USER
))
3238 return list_locations(s
, buf
, TRACK_ALLOC
);
3240 SLAB_ATTR_RO(alloc_calls
);
3242 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
3244 if (!(s
->flags
& SLAB_STORE_USER
))
3246 return list_locations(s
, buf
, TRACK_FREE
);
3248 SLAB_ATTR_RO(free_calls
);
3251 static ssize_t
defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
3253 return sprintf(buf
, "%d\n", s
->defrag_ratio
/ 10);
3256 static ssize_t
defrag_ratio_store(struct kmem_cache
*s
,
3257 const char *buf
, size_t length
)
3259 int n
= simple_strtoul(buf
, NULL
, 10);
3262 s
->defrag_ratio
= n
* 10;
3265 SLAB_ATTR(defrag_ratio
);
3268 static struct attribute
* slab_attrs
[] = {
3269 &slab_size_attr
.attr
,
3270 &object_size_attr
.attr
,
3271 &objs_per_slab_attr
.attr
,
3276 &cpu_slabs_attr
.attr
,
3281 &sanity_checks_attr
.attr
,
3283 &hwcache_align_attr
.attr
,
3284 &reclaim_account_attr
.attr
,
3285 &destroy_by_rcu_attr
.attr
,
3286 &red_zone_attr
.attr
,
3288 &store_user_attr
.attr
,
3289 &validate_attr
.attr
,
3291 &alloc_calls_attr
.attr
,
3292 &free_calls_attr
.attr
,
3293 #ifdef CONFIG_ZONE_DMA
3294 &cache_dma_attr
.attr
,
3297 &defrag_ratio_attr
.attr
,
3302 static struct attribute_group slab_attr_group
= {
3303 .attrs
= slab_attrs
,
3306 static ssize_t
slab_attr_show(struct kobject
*kobj
,
3307 struct attribute
*attr
,
3310 struct slab_attribute
*attribute
;
3311 struct kmem_cache
*s
;
3314 attribute
= to_slab_attr(attr
);
3317 if (!attribute
->show
)
3320 err
= attribute
->show(s
, buf
);
3325 static ssize_t
slab_attr_store(struct kobject
*kobj
,
3326 struct attribute
*attr
,
3327 const char *buf
, size_t len
)
3329 struct slab_attribute
*attribute
;
3330 struct kmem_cache
*s
;
3333 attribute
= to_slab_attr(attr
);
3336 if (!attribute
->store
)
3339 err
= attribute
->store(s
, buf
, len
);
3344 static struct sysfs_ops slab_sysfs_ops
= {
3345 .show
= slab_attr_show
,
3346 .store
= slab_attr_store
,
3349 static struct kobj_type slab_ktype
= {
3350 .sysfs_ops
= &slab_sysfs_ops
,
3353 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
3355 struct kobj_type
*ktype
= get_ktype(kobj
);
3357 if (ktype
== &slab_ktype
)
3362 static struct kset_uevent_ops slab_uevent_ops
= {
3363 .filter
= uevent_filter
,
3366 decl_subsys(slab
, &slab_ktype
, &slab_uevent_ops
);
3368 #define ID_STR_LENGTH 64
3370 /* Create a unique string id for a slab cache:
3372 * :[flags-]size:[memory address of kmemcache]
3374 static char *create_unique_id(struct kmem_cache
*s
)
3376 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
3383 * First flags affecting slabcache operations. We will only
3384 * get here for aliasable slabs so we do not need to support
3385 * too many flags. The flags here must cover all flags that
3386 * are matched during merging to guarantee that the id is
3389 if (s
->flags
& SLAB_CACHE_DMA
)
3391 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3393 if (s
->flags
& SLAB_DEBUG_FREE
)
3397 p
+= sprintf(p
, "%07d", s
->size
);
3398 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
3402 static int sysfs_slab_add(struct kmem_cache
*s
)
3408 if (slab_state
< SYSFS
)
3409 /* Defer until later */
3412 unmergeable
= slab_unmergeable(s
);
3415 * Slabcache can never be merged so we can use the name proper.
3416 * This is typically the case for debug situations. In that
3417 * case we can catch duplicate names easily.
3419 sysfs_remove_link(&slab_subsys
.kobj
, s
->name
);
3423 * Create a unique name for the slab as a target
3426 name
= create_unique_id(s
);
3429 kobj_set_kset_s(s
, slab_subsys
);
3430 kobject_set_name(&s
->kobj
, name
);
3431 kobject_init(&s
->kobj
);
3432 err
= kobject_add(&s
->kobj
);
3436 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
3439 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
3441 /* Setup first alias */
3442 sysfs_slab_alias(s
, s
->name
);
3448 static void sysfs_slab_remove(struct kmem_cache
*s
)
3450 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
3451 kobject_del(&s
->kobj
);
3455 * Need to buffer aliases during bootup until sysfs becomes
3456 * available lest we loose that information.
3458 struct saved_alias
{
3459 struct kmem_cache
*s
;
3461 struct saved_alias
*next
;
3464 struct saved_alias
*alias_list
;
3466 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
3468 struct saved_alias
*al
;
3470 if (slab_state
== SYSFS
) {
3472 * If we have a leftover link then remove it.
3474 sysfs_remove_link(&slab_subsys
.kobj
, name
);
3475 return sysfs_create_link(&slab_subsys
.kobj
,
3479 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
3485 al
->next
= alias_list
;
3490 static int __init
slab_sysfs_init(void)
3494 err
= subsystem_register(&slab_subsys
);
3496 printk(KERN_ERR
"Cannot register slab subsystem.\n");
3502 while (alias_list
) {
3503 struct saved_alias
*al
= alias_list
;
3505 alias_list
= alias_list
->next
;
3506 err
= sysfs_slab_alias(al
->s
, al
->name
);
3515 __initcall(slab_sysfs_init
);
3517 __initcall(finish_bootstrap
);