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 static int kmem_size
= sizeof(struct kmem_cache
);
163 static struct notifier_block slab_notifier
;
167 DOWN
, /* No slab functionality available */
168 PARTIAL
, /* kmem_cache_open() works but kmalloc does not */
169 UP
, /* Everything works */
173 /* A list of all slab caches on the system */
174 static DECLARE_RWSEM(slub_lock
);
175 LIST_HEAD(slab_caches
);
178 static int sysfs_slab_add(struct kmem_cache
*);
179 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
180 static void sysfs_slab_remove(struct kmem_cache
*);
182 static int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
183 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
) { return 0; }
184 static void sysfs_slab_remove(struct kmem_cache
*s
) {}
187 /********************************************************************
188 * Core slab cache functions
189 *******************************************************************/
191 int slab_is_available(void)
193 return slab_state
>= UP
;
196 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
199 return s
->node
[node
];
201 return &s
->local_node
;
208 static void print_section(char *text
, u8
*addr
, unsigned int length
)
216 for (i
= 0; i
< length
; i
++) {
218 printk(KERN_ERR
"%10s 0x%p: ", text
, addr
+ i
);
221 printk(" %02x", addr
[i
]);
223 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
225 printk(" %s\n",ascii
);
236 printk(" %s\n", ascii
);
241 * Slow version of get and set free pointer.
243 * This requires touching the cache lines of kmem_cache.
244 * The offset can also be obtained from the page. In that
245 * case it is in the cacheline that we already need to touch.
247 static void *get_freepointer(struct kmem_cache
*s
, void *object
)
249 return *(void **)(object
+ s
->offset
);
252 static void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
254 *(void **)(object
+ s
->offset
) = fp
;
258 * Tracking user of a slab.
261 void *addr
; /* Called from address */
262 int cpu
; /* Was running on cpu */
263 int pid
; /* Pid context */
264 unsigned long when
; /* When did the operation occur */
267 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
269 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
270 enum track_item alloc
)
275 p
= object
+ s
->offset
+ sizeof(void *);
277 p
= object
+ s
->inuse
;
282 static void set_track(struct kmem_cache
*s
, void *object
,
283 enum track_item alloc
, void *addr
)
288 p
= object
+ s
->offset
+ sizeof(void *);
290 p
= object
+ s
->inuse
;
295 p
->cpu
= smp_processor_id();
296 p
->pid
= current
? current
->pid
: -1;
299 memset(p
, 0, sizeof(struct track
));
302 static void init_tracking(struct kmem_cache
*s
, void *object
)
304 if (s
->flags
& SLAB_STORE_USER
) {
305 set_track(s
, object
, TRACK_FREE
, NULL
);
306 set_track(s
, object
, TRACK_ALLOC
, NULL
);
310 static void print_track(const char *s
, struct track
*t
)
315 printk(KERN_ERR
"%s: ", s
);
316 __print_symbol("%s", (unsigned long)t
->addr
);
317 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies
- t
->when
, t
->cpu
, t
->pid
);
320 static void print_trailer(struct kmem_cache
*s
, u8
*p
)
322 unsigned int off
; /* Offset of last byte */
324 if (s
->flags
& SLAB_RED_ZONE
)
325 print_section("Redzone", p
+ s
->objsize
,
326 s
->inuse
- s
->objsize
);
328 printk(KERN_ERR
"FreePointer 0x%p -> 0x%p\n",
330 get_freepointer(s
, p
));
333 off
= s
->offset
+ sizeof(void *);
337 if (s
->flags
& SLAB_STORE_USER
) {
338 print_track("Last alloc", get_track(s
, p
, TRACK_ALLOC
));
339 print_track("Last free ", get_track(s
, p
, TRACK_FREE
));
340 off
+= 2 * sizeof(struct track
);
344 /* Beginning of the filler is the free pointer */
345 print_section("Filler", p
+ off
, s
->size
- off
);
348 static void object_err(struct kmem_cache
*s
, struct page
*page
,
349 u8
*object
, char *reason
)
351 u8
*addr
= page_address(page
);
353 printk(KERN_ERR
"*** SLUB %s: %s@0x%p slab 0x%p\n",
354 s
->name
, reason
, object
, page
);
355 printk(KERN_ERR
" offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
356 object
- addr
, page
->flags
, page
->inuse
, page
->freelist
);
357 if (object
> addr
+ 16)
358 print_section("Bytes b4", object
- 16, 16);
359 print_section("Object", object
, min(s
->objsize
, 128));
360 print_trailer(s
, object
);
364 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *reason
, ...)
369 va_start(args
, reason
);
370 vsnprintf(buf
, sizeof(buf
), reason
, args
);
372 printk(KERN_ERR
"*** SLUB %s: %s in slab @0x%p\n", s
->name
, buf
,
377 static void init_object(struct kmem_cache
*s
, void *object
, int active
)
381 if (s
->flags
& __OBJECT_POISON
) {
382 memset(p
, POISON_FREE
, s
->objsize
- 1);
383 p
[s
->objsize
-1] = POISON_END
;
386 if (s
->flags
& SLAB_RED_ZONE
)
387 memset(p
+ s
->objsize
,
388 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
,
389 s
->inuse
- s
->objsize
);
392 static int check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
395 if (*start
!= (u8
)value
)
404 static int check_valid_pointer(struct kmem_cache
*s
, struct page
*page
,
412 base
= page_address(page
);
413 if (object
< base
|| object
>= base
+ s
->objects
* s
->size
||
414 (object
- base
) % s
->size
) {
425 * Bytes of the object to be managed.
426 * If the freepointer may overlay the object then the free
427 * pointer is the first word of the object.
428 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
431 * object + s->objsize
432 * Padding to reach word boundary. This is also used for Redzoning.
433 * Padding is extended to word size if Redzoning is enabled
434 * and objsize == inuse.
435 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
436 * 0xcc (RED_ACTIVE) for objects in use.
439 * A. Free pointer (if we cannot overwrite object on free)
440 * B. Tracking data for SLAB_STORE_USER
441 * C. Padding to reach required alignment boundary
442 * Padding is done using 0x5a (POISON_INUSE)
446 * If slabcaches are merged then the objsize and inuse boundaries are to
447 * be ignored. And therefore no slab options that rely on these boundaries
448 * may be used with merged slabcaches.
451 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
452 void *from
, void *to
)
454 printk(KERN_ERR
"@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
455 s
->name
, message
, data
, from
, to
- 1);
456 memset(from
, data
, to
- from
);
459 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
461 unsigned long off
= s
->inuse
; /* The end of info */
464 /* Freepointer is placed after the object. */
465 off
+= sizeof(void *);
467 if (s
->flags
& SLAB_STORE_USER
)
468 /* We also have user information there */
469 off
+= 2 * sizeof(struct track
);
474 if (check_bytes(p
+ off
, POISON_INUSE
, s
->size
- off
))
477 object_err(s
, page
, p
, "Object padding check fails");
482 restore_bytes(s
, "object padding", POISON_INUSE
, p
+ off
, p
+ s
->size
);
486 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
489 int length
, remainder
;
491 if (!(s
->flags
& SLAB_POISON
))
494 p
= page_address(page
);
495 length
= s
->objects
* s
->size
;
496 remainder
= (PAGE_SIZE
<< s
->order
) - length
;
500 if (!check_bytes(p
+ length
, POISON_INUSE
, remainder
)) {
501 slab_err(s
, page
, "Padding check failed");
502 restore_bytes(s
, "slab padding", POISON_INUSE
, p
+ length
,
503 p
+ length
+ remainder
);
509 static int check_object(struct kmem_cache
*s
, struct page
*page
,
510 void *object
, int active
)
513 u8
*endobject
= object
+ s
->objsize
;
515 if (s
->flags
& SLAB_RED_ZONE
) {
517 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
519 if (!check_bytes(endobject
, red
, s
->inuse
- s
->objsize
)) {
520 object_err(s
, page
, object
,
521 active
? "Redzone Active" : "Redzone Inactive");
522 restore_bytes(s
, "redzone", red
,
523 endobject
, object
+ s
->inuse
);
527 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
&&
528 !check_bytes(endobject
, POISON_INUSE
,
529 s
->inuse
- s
->objsize
)) {
530 object_err(s
, page
, p
, "Alignment padding check fails");
532 * Fix it so that there will not be another report.
534 * Hmmm... We may be corrupting an object that now expects
535 * to be longer than allowed.
537 restore_bytes(s
, "alignment padding", POISON_INUSE
,
538 endobject
, object
+ s
->inuse
);
542 if (s
->flags
& SLAB_POISON
) {
543 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
544 (!check_bytes(p
, POISON_FREE
, s
->objsize
- 1) ||
545 p
[s
->objsize
- 1] != POISON_END
)) {
547 object_err(s
, page
, p
, "Poison check failed");
548 restore_bytes(s
, "Poison", POISON_FREE
,
549 p
, p
+ s
->objsize
-1);
550 restore_bytes(s
, "Poison", POISON_END
,
551 p
+ s
->objsize
- 1, p
+ s
->objsize
);
555 * check_pad_bytes cleans up on its own.
557 check_pad_bytes(s
, page
, p
);
560 if (!s
->offset
&& active
)
562 * Object and freepointer overlap. Cannot check
563 * freepointer while object is allocated.
567 /* Check free pointer validity */
568 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
569 object_err(s
, page
, p
, "Freepointer corrupt");
571 * No choice but to zap it and thus loose the remainder
572 * of the free objects in this slab. May cause
573 * another error because the object count maybe
576 set_freepointer(s
, p
, NULL
);
582 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
584 VM_BUG_ON(!irqs_disabled());
586 if (!PageSlab(page
)) {
587 slab_err(s
, page
, "Not a valid slab page flags=%lx "
588 "mapping=0x%p count=%d", page
->flags
, page
->mapping
,
592 if (page
->offset
* sizeof(void *) != s
->offset
) {
593 slab_err(s
, page
, "Corrupted offset %lu flags=0x%lx "
594 "mapping=0x%p count=%d",
595 (unsigned long)(page
->offset
* sizeof(void *)),
601 if (page
->inuse
> s
->objects
) {
602 slab_err(s
, page
, "inuse %u > max %u @0x%p flags=%lx "
603 "mapping=0x%p count=%d",
604 s
->name
, page
->inuse
, s
->objects
, page
->flags
,
605 page
->mapping
, page_count(page
));
608 /* Slab_pad_check fixes things up after itself */
609 slab_pad_check(s
, page
);
614 * Determine if a certain object on a page is on the freelist and
615 * therefore free. Must hold the slab lock for cpu slabs to
616 * guarantee that the chains are consistent.
618 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
621 void *fp
= page
->freelist
;
624 while (fp
&& nr
<= s
->objects
) {
627 if (!check_valid_pointer(s
, page
, fp
)) {
629 object_err(s
, page
, object
,
630 "Freechain corrupt");
631 set_freepointer(s
, object
, NULL
);
634 slab_err(s
, page
, "Freepointer 0x%p corrupt",
636 page
->freelist
= NULL
;
637 page
->inuse
= s
->objects
;
638 printk(KERN_ERR
"@@@ SLUB %s: Freelist "
639 "cleared. Slab 0x%p\n",
646 fp
= get_freepointer(s
, object
);
650 if (page
->inuse
!= s
->objects
- nr
) {
651 slab_err(s
, page
, "Wrong object count. Counter is %d but "
652 "counted were %d", s
, page
, page
->inuse
,
654 page
->inuse
= s
->objects
- nr
;
655 printk(KERN_ERR
"@@@ SLUB %s: Object count adjusted. "
656 "Slab @0x%p\n", s
->name
, page
);
658 return search
== NULL
;
662 * Tracking of fully allocated slabs for debugging
664 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
666 spin_lock(&n
->list_lock
);
667 list_add(&page
->lru
, &n
->full
);
668 spin_unlock(&n
->list_lock
);
671 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
673 struct kmem_cache_node
*n
;
675 if (!(s
->flags
& SLAB_STORE_USER
))
678 n
= get_node(s
, page_to_nid(page
));
680 spin_lock(&n
->list_lock
);
681 list_del(&page
->lru
);
682 spin_unlock(&n
->list_lock
);
685 static int alloc_object_checks(struct kmem_cache
*s
, struct page
*page
,
688 if (!check_slab(s
, page
))
691 if (object
&& !on_freelist(s
, page
, object
)) {
692 slab_err(s
, page
, "Object 0x%p already allocated", object
);
696 if (!check_valid_pointer(s
, page
, object
)) {
697 object_err(s
, page
, object
, "Freelist Pointer check fails");
704 if (!check_object(s
, page
, object
, 0))
709 if (PageSlab(page
)) {
711 * If this is a slab page then lets do the best we can
712 * to avoid issues in the future. Marking all objects
713 * as used avoids touching the remainder.
715 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
717 page
->inuse
= s
->objects
;
718 page
->freelist
= NULL
;
719 /* Fix up fields that may be corrupted */
720 page
->offset
= s
->offset
/ sizeof(void *);
725 static int free_object_checks(struct kmem_cache
*s
, struct page
*page
,
728 if (!check_slab(s
, page
))
731 if (!check_valid_pointer(s
, page
, object
)) {
732 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
736 if (on_freelist(s
, page
, object
)) {
737 slab_err(s
, page
, "Object 0x%p already free", object
);
741 if (!check_object(s
, page
, object
, 1))
744 if (unlikely(s
!= page
->slab
)) {
746 slab_err(s
, page
, "Attempt to free object(0x%p) "
747 "outside of slab", object
);
751 "SLUB <none>: no slab for object 0x%p.\n",
756 slab_err(s
, page
, "object at 0x%p belongs "
757 "to slab %s", object
, page
->slab
->name
);
762 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
763 s
->name
, page
, object
);
768 * Slab allocation and freeing
770 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
773 int pages
= 1 << s
->order
;
778 if (s
->flags
& SLAB_CACHE_DMA
)
782 page
= alloc_pages(flags
, s
->order
);
784 page
= alloc_pages_node(node
, flags
, s
->order
);
789 mod_zone_page_state(page_zone(page
),
790 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
791 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
797 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
800 if (PageError(page
)) {
801 init_object(s
, object
, 0);
802 init_tracking(s
, object
);
805 if (unlikely(s
->ctor
))
806 s
->ctor(object
, s
, SLAB_CTOR_CONSTRUCTOR
);
809 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
812 struct kmem_cache_node
*n
;
818 if (flags
& __GFP_NO_GROW
)
821 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
));
823 if (flags
& __GFP_WAIT
)
826 page
= allocate_slab(s
, flags
& GFP_LEVEL_MASK
, node
);
830 n
= get_node(s
, page_to_nid(page
));
832 atomic_long_inc(&n
->nr_slabs
);
833 page
->offset
= s
->offset
/ sizeof(void *);
835 page
->flags
|= 1 << PG_slab
;
836 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
837 SLAB_STORE_USER
| SLAB_TRACE
))
838 page
->flags
|= 1 << PG_error
;
840 start
= page_address(page
);
841 end
= start
+ s
->objects
* s
->size
;
843 if (unlikely(s
->flags
& SLAB_POISON
))
844 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
847 for (p
= start
+ s
->size
; p
< end
; p
+= s
->size
) {
848 setup_object(s
, page
, last
);
849 set_freepointer(s
, last
, p
);
852 setup_object(s
, page
, last
);
853 set_freepointer(s
, last
, NULL
);
855 page
->freelist
= start
;
858 if (flags
& __GFP_WAIT
)
863 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
865 int pages
= 1 << s
->order
;
867 if (unlikely(PageError(page
) || s
->dtor
)) {
868 void *start
= page_address(page
);
869 void *end
= start
+ (pages
<< PAGE_SHIFT
);
872 slab_pad_check(s
, page
);
873 for (p
= start
; p
<= end
- s
->size
; p
+= s
->size
) {
876 check_object(s
, page
, p
, 0);
880 mod_zone_page_state(page_zone(page
),
881 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
882 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
885 page
->mapping
= NULL
;
886 __free_pages(page
, s
->order
);
889 static void rcu_free_slab(struct rcu_head
*h
)
893 page
= container_of((struct list_head
*)h
, struct page
, lru
);
894 __free_slab(page
->slab
, page
);
897 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
899 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
901 * RCU free overloads the RCU head over the LRU
903 struct rcu_head
*head
= (void *)&page
->lru
;
905 call_rcu(head
, rcu_free_slab
);
907 __free_slab(s
, page
);
910 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
912 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
914 atomic_long_dec(&n
->nr_slabs
);
915 reset_page_mapcount(page
);
916 page
->flags
&= ~(1 << PG_slab
| 1 << PG_error
);
921 * Per slab locking using the pagelock
923 static __always_inline
void slab_lock(struct page
*page
)
925 bit_spin_lock(PG_locked
, &page
->flags
);
928 static __always_inline
void slab_unlock(struct page
*page
)
930 bit_spin_unlock(PG_locked
, &page
->flags
);
933 static __always_inline
int slab_trylock(struct page
*page
)
937 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
942 * Management of partially allocated slabs
944 static void add_partial_tail(struct kmem_cache_node
*n
, struct page
*page
)
946 spin_lock(&n
->list_lock
);
948 list_add_tail(&page
->lru
, &n
->partial
);
949 spin_unlock(&n
->list_lock
);
952 static void add_partial(struct kmem_cache_node
*n
, struct page
*page
)
954 spin_lock(&n
->list_lock
);
956 list_add(&page
->lru
, &n
->partial
);
957 spin_unlock(&n
->list_lock
);
960 static void remove_partial(struct kmem_cache
*s
,
963 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
965 spin_lock(&n
->list_lock
);
966 list_del(&page
->lru
);
968 spin_unlock(&n
->list_lock
);
972 * Lock page and remove it from the partial list
974 * Must hold list_lock
976 static int lock_and_del_slab(struct kmem_cache_node
*n
, struct page
*page
)
978 if (slab_trylock(page
)) {
979 list_del(&page
->lru
);
987 * Try to get a partial slab from a specific node
989 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
994 * Racy check. If we mistakenly see no partial slabs then we
995 * just allocate an empty slab. If we mistakenly try to get a
996 * partial slab then get_partials() will return NULL.
998 if (!n
|| !n
->nr_partial
)
1001 spin_lock(&n
->list_lock
);
1002 list_for_each_entry(page
, &n
->partial
, lru
)
1003 if (lock_and_del_slab(n
, page
))
1007 spin_unlock(&n
->list_lock
);
1012 * Get a page from somewhere. Search in increasing NUMA
1015 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1018 struct zonelist
*zonelist
;
1023 * The defrag ratio allows to configure the tradeoffs between
1024 * inter node defragmentation and node local allocations.
1025 * A lower defrag_ratio increases the tendency to do local
1026 * allocations instead of scanning throught the partial
1027 * lists on other nodes.
1029 * If defrag_ratio is set to 0 then kmalloc() always
1030 * returns node local objects. If its higher then kmalloc()
1031 * may return off node objects in order to avoid fragmentation.
1033 * A higher ratio means slabs may be taken from other nodes
1034 * thus reducing the number of partial slabs on those nodes.
1036 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1037 * defrag_ratio = 1000) then every (well almost) allocation
1038 * will first attempt to defrag slab caches on other nodes. This
1039 * means scanning over all nodes to look for partial slabs which
1040 * may be a bit expensive to do on every slab allocation.
1042 if (!s
->defrag_ratio
|| get_cycles() % 1024 > s
->defrag_ratio
)
1045 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
1046 ->node_zonelists
[gfp_zone(flags
)];
1047 for (z
= zonelist
->zones
; *z
; z
++) {
1048 struct kmem_cache_node
*n
;
1050 n
= get_node(s
, zone_to_nid(*z
));
1052 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1053 n
->nr_partial
> MIN_PARTIAL
) {
1054 page
= get_partial_node(n
);
1064 * Get a partial page, lock it and return it.
1066 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1069 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1071 page
= get_partial_node(get_node(s
, searchnode
));
1072 if (page
|| (flags
& __GFP_THISNODE
))
1075 return get_any_partial(s
, flags
);
1079 * Move a page back to the lists.
1081 * Must be called with the slab lock held.
1083 * On exit the slab lock will have been dropped.
1085 static void putback_slab(struct kmem_cache
*s
, struct page
*page
)
1087 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1092 add_partial(n
, page
);
1093 else if (PageError(page
) && (s
->flags
& SLAB_STORE_USER
))
1098 if (n
->nr_partial
< MIN_PARTIAL
) {
1100 * Adding an empty page to the partial slabs in order
1101 * to avoid page allocator overhead. This page needs to
1102 * come after all the others that are not fully empty
1103 * in order to make sure that we do maximum
1106 add_partial_tail(n
, page
);
1110 discard_slab(s
, page
);
1116 * Remove the cpu slab
1118 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1120 s
->cpu_slab
[cpu
] = NULL
;
1121 ClearPageActive(page
);
1123 putback_slab(s
, page
);
1126 static void flush_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1129 deactivate_slab(s
, page
, cpu
);
1134 * Called from IPI handler with interrupts disabled.
1136 static void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1138 struct page
*page
= s
->cpu_slab
[cpu
];
1141 flush_slab(s
, page
, cpu
);
1144 static void flush_cpu_slab(void *d
)
1146 struct kmem_cache
*s
= d
;
1147 int cpu
= smp_processor_id();
1149 __flush_cpu_slab(s
, cpu
);
1152 static void flush_all(struct kmem_cache
*s
)
1155 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1157 unsigned long flags
;
1159 local_irq_save(flags
);
1161 local_irq_restore(flags
);
1166 * slab_alloc is optimized to only modify two cachelines on the fast path
1167 * (aside from the stack):
1169 * 1. The page struct
1170 * 2. The first cacheline of the object to be allocated.
1172 * The only cache lines that are read (apart from code) is the
1173 * per cpu array in the kmem_cache struct.
1175 * Fastpath is not possible if we need to get a new slab or have
1176 * debugging enabled (which means all slabs are marked with PageError)
1178 static void *slab_alloc(struct kmem_cache
*s
,
1179 gfp_t gfpflags
, int node
, void *addr
)
1183 unsigned long flags
;
1186 local_irq_save(flags
);
1187 cpu
= smp_processor_id();
1188 page
= s
->cpu_slab
[cpu
];
1193 if (unlikely(node
!= -1 && page_to_nid(page
) != node
))
1196 object
= page
->freelist
;
1197 if (unlikely(!object
))
1199 if (unlikely(PageError(page
)))
1204 page
->freelist
= object
[page
->offset
];
1206 local_irq_restore(flags
);
1210 deactivate_slab(s
, page
, cpu
);
1213 page
= get_partial(s
, gfpflags
, node
);
1216 s
->cpu_slab
[cpu
] = page
;
1217 SetPageActive(page
);
1221 page
= new_slab(s
, gfpflags
, node
);
1223 cpu
= smp_processor_id();
1224 if (s
->cpu_slab
[cpu
]) {
1226 * Someone else populated the cpu_slab while we enabled
1227 * interrupts, or we have got scheduled on another cpu.
1228 * The page may not be on the requested node.
1231 page_to_nid(s
->cpu_slab
[cpu
]) == node
) {
1233 * Current cpuslab is acceptable and we
1234 * want the current one since its cache hot
1236 discard_slab(s
, page
);
1237 page
= s
->cpu_slab
[cpu
];
1241 /* Dump the current slab */
1242 flush_slab(s
, s
->cpu_slab
[cpu
], cpu
);
1247 local_irq_restore(flags
);
1250 if (!alloc_object_checks(s
, page
, object
))
1252 if (s
->flags
& SLAB_STORE_USER
)
1253 set_track(s
, object
, TRACK_ALLOC
, addr
);
1254 if (s
->flags
& SLAB_TRACE
) {
1255 printk(KERN_INFO
"TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
1256 s
->name
, object
, page
->inuse
,
1260 init_object(s
, object
, 1);
1264 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1266 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1268 EXPORT_SYMBOL(kmem_cache_alloc
);
1271 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1273 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1275 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1279 * The fastpath only writes the cacheline of the page struct and the first
1280 * cacheline of the object.
1282 * No special cachelines need to be read
1284 static void slab_free(struct kmem_cache
*s
, struct page
*page
,
1285 void *x
, void *addr
)
1288 void **object
= (void *)x
;
1289 unsigned long flags
;
1291 local_irq_save(flags
);
1294 if (unlikely(PageError(page
)))
1297 prior
= object
[page
->offset
] = page
->freelist
;
1298 page
->freelist
= object
;
1301 if (unlikely(PageActive(page
)))
1303 * Cpu slabs are never on partial lists and are
1308 if (unlikely(!page
->inuse
))
1312 * Objects left in the slab. If it
1313 * was not on the partial list before
1316 if (unlikely(!prior
))
1317 add_partial(get_node(s
, page_to_nid(page
)), page
);
1321 local_irq_restore(flags
);
1327 * Slab on the partial list.
1329 remove_partial(s
, page
);
1332 discard_slab(s
, page
);
1333 local_irq_restore(flags
);
1337 if (!free_object_checks(s
, page
, x
))
1339 if (!PageActive(page
) && !page
->freelist
)
1340 remove_full(s
, page
);
1341 if (s
->flags
& SLAB_STORE_USER
)
1342 set_track(s
, x
, TRACK_FREE
, addr
);
1343 if (s
->flags
& SLAB_TRACE
) {
1344 printk(KERN_INFO
"TRACE %s free 0x%p inuse=%d fp=0x%p\n",
1345 s
->name
, object
, page
->inuse
,
1347 print_section("Object", (void *)object
, s
->objsize
);
1350 init_object(s
, object
, 0);
1354 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1358 page
= virt_to_head_page(x
);
1360 slab_free(s
, page
, x
, __builtin_return_address(0));
1362 EXPORT_SYMBOL(kmem_cache_free
);
1364 /* Figure out on which slab object the object resides */
1365 static struct page
*get_object_page(const void *x
)
1367 struct page
*page
= virt_to_head_page(x
);
1369 if (!PageSlab(page
))
1376 * kmem_cache_open produces objects aligned at "size" and the first object
1377 * is placed at offset 0 in the slab (We have no metainformation on the
1378 * slab, all slabs are in essence "off slab").
1380 * In order to get the desired alignment one just needs to align the
1383 * Notice that the allocation order determines the sizes of the per cpu
1384 * caches. Each processor has always one slab available for allocations.
1385 * Increasing the allocation order reduces the number of times that slabs
1386 * must be moved on and off the partial lists and therefore may influence
1389 * The offset is used to relocate the free list link in each object. It is
1390 * therefore possible to move the free list link behind the object. This
1391 * is necessary for RCU to work properly and also useful for debugging.
1395 * Mininum / Maximum order of slab pages. This influences locking overhead
1396 * and slab fragmentation. A higher order reduces the number of partial slabs
1397 * and increases the number of allocations possible without having to
1398 * take the list_lock.
1400 static int slub_min_order
;
1401 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1404 * Minimum number of objects per slab. This is necessary in order to
1405 * reduce locking overhead. Similar to the queue size in SLAB.
1407 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1410 * Merge control. If this is set then no merging of slab caches will occur.
1412 static int slub_nomerge
;
1417 static int slub_debug
;
1419 static char *slub_debug_slabs
;
1422 * Calculate the order of allocation given an slab object size.
1424 * The order of allocation has significant impact on other elements
1425 * of the system. Generally order 0 allocations should be preferred
1426 * since they do not cause fragmentation in the page allocator. Larger
1427 * objects may have problems with order 0 because there may be too much
1428 * space left unused in a slab. We go to a higher order if more than 1/8th
1429 * of the slab would be wasted.
1431 * In order to reach satisfactory performance we must ensure that
1432 * a minimum number of objects is in one slab. Otherwise we may
1433 * generate too much activity on the partial lists. This is less a
1434 * concern for large slabs though. slub_max_order specifies the order
1435 * where we begin to stop considering the number of objects in a slab.
1437 * Higher order allocations also allow the placement of more objects
1438 * in a slab and thereby reduce object handling overhead. If the user
1439 * has requested a higher mininum order then we start with that one
1442 static int calculate_order(int size
)
1447 for (order
= max(slub_min_order
, fls(size
- 1) - PAGE_SHIFT
);
1448 order
< MAX_ORDER
; order
++) {
1449 unsigned long slab_size
= PAGE_SIZE
<< order
;
1451 if (slub_max_order
> order
&&
1452 slab_size
< slub_min_objects
* size
)
1455 if (slab_size
< size
)
1458 rem
= slab_size
% size
;
1460 if (rem
<= (PAGE_SIZE
<< order
) / 8)
1464 if (order
>= MAX_ORDER
)
1470 * Function to figure out which alignment to use from the
1471 * various ways of specifying it.
1473 static unsigned long calculate_alignment(unsigned long flags
,
1474 unsigned long align
, unsigned long size
)
1477 * If the user wants hardware cache aligned objects then
1478 * follow that suggestion if the object is sufficiently
1481 * The hardware cache alignment cannot override the
1482 * specified alignment though. If that is greater
1485 if ((flags
& SLAB_HWCACHE_ALIGN
) &&
1486 size
> L1_CACHE_BYTES
/ 2)
1487 return max_t(unsigned long, align
, L1_CACHE_BYTES
);
1489 if (align
< ARCH_SLAB_MINALIGN
)
1490 return ARCH_SLAB_MINALIGN
;
1492 return ALIGN(align
, sizeof(void *));
1495 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1498 atomic_long_set(&n
->nr_slabs
, 0);
1499 spin_lock_init(&n
->list_lock
);
1500 INIT_LIST_HEAD(&n
->partial
);
1501 INIT_LIST_HEAD(&n
->full
);
1506 * No kmalloc_node yet so do it by hand. We know that this is the first
1507 * slab on the node for this slabcache. There are no concurrent accesses
1510 * Note that this function only works on the kmalloc_node_cache
1511 * when allocating for the kmalloc_node_cache.
1513 static struct kmem_cache_node
* __init
early_kmem_cache_node_alloc(gfp_t gfpflags
,
1517 struct kmem_cache_node
*n
;
1519 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
1521 page
= new_slab(kmalloc_caches
, gfpflags
| GFP_THISNODE
, node
);
1522 /* new_slab() disables interupts */
1528 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
1530 kmalloc_caches
->node
[node
] = n
;
1531 init_object(kmalloc_caches
, n
, 1);
1532 init_kmem_cache_node(n
);
1533 atomic_long_inc(&n
->nr_slabs
);
1534 add_partial(n
, page
);
1538 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1542 for_each_online_node(node
) {
1543 struct kmem_cache_node
*n
= s
->node
[node
];
1544 if (n
&& n
!= &s
->local_node
)
1545 kmem_cache_free(kmalloc_caches
, n
);
1546 s
->node
[node
] = NULL
;
1550 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1555 if (slab_state
>= UP
)
1556 local_node
= page_to_nid(virt_to_page(s
));
1560 for_each_online_node(node
) {
1561 struct kmem_cache_node
*n
;
1563 if (local_node
== node
)
1566 if (slab_state
== DOWN
) {
1567 n
= early_kmem_cache_node_alloc(gfpflags
,
1571 n
= kmem_cache_alloc_node(kmalloc_caches
,
1575 free_kmem_cache_nodes(s
);
1581 init_kmem_cache_node(n
);
1586 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1590 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1592 init_kmem_cache_node(&s
->local_node
);
1598 * calculate_sizes() determines the order and the distribution of data within
1601 static int calculate_sizes(struct kmem_cache
*s
)
1603 unsigned long flags
= s
->flags
;
1604 unsigned long size
= s
->objsize
;
1605 unsigned long align
= s
->align
;
1608 * Determine if we can poison the object itself. If the user of
1609 * the slab may touch the object after free or before allocation
1610 * then we should never poison the object itself.
1612 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
1613 !s
->ctor
&& !s
->dtor
)
1614 s
->flags
|= __OBJECT_POISON
;
1616 s
->flags
&= ~__OBJECT_POISON
;
1619 * Round up object size to the next word boundary. We can only
1620 * place the free pointer at word boundaries and this determines
1621 * the possible location of the free pointer.
1623 size
= ALIGN(size
, sizeof(void *));
1626 * If we are redzoning then check if there is some space between the
1627 * end of the object and the free pointer. If not then add an
1628 * additional word, so that we can establish a redzone between
1629 * the object and the freepointer to be able to check for overwrites.
1631 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
1632 size
+= sizeof(void *);
1635 * With that we have determined how much of the slab is in actual
1636 * use by the object. This is the potential offset to the free
1641 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
1642 s
->ctor
|| s
->dtor
)) {
1644 * Relocate free pointer after the object if it is not
1645 * permitted to overwrite the first word of the object on
1648 * This is the case if we do RCU, have a constructor or
1649 * destructor or are poisoning the objects.
1652 size
+= sizeof(void *);
1655 if (flags
& SLAB_STORE_USER
)
1657 * Need to store information about allocs and frees after
1660 size
+= 2 * sizeof(struct track
);
1662 if (flags
& DEBUG_DEFAULT_FLAGS
)
1664 * Add some empty padding so that we can catch
1665 * overwrites from earlier objects rather than let
1666 * tracking information or the free pointer be
1667 * corrupted if an user writes before the start
1670 size
+= sizeof(void *);
1672 * Determine the alignment based on various parameters that the
1673 * user specified (this is unecessarily complex due to the attempt
1674 * to be compatible with SLAB. Should be cleaned up some day).
1676 align
= calculate_alignment(flags
, align
, s
->objsize
);
1679 * SLUB stores one object immediately after another beginning from
1680 * offset 0. In order to align the objects we have to simply size
1681 * each object to conform to the alignment.
1683 size
= ALIGN(size
, align
);
1686 s
->order
= calculate_order(size
);
1691 * Determine the number of objects per slab
1693 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
1696 * Verify that the number of objects is within permitted limits.
1697 * The page->inuse field is only 16 bit wide! So we cannot have
1698 * more than 64k objects per slab.
1700 if (!s
->objects
|| s
->objects
> 65535)
1706 static int __init
finish_bootstrap(void)
1708 struct list_head
*h
;
1713 list_for_each(h
, &slab_caches
) {
1714 struct kmem_cache
*s
=
1715 container_of(h
, struct kmem_cache
, list
);
1717 err
= sysfs_slab_add(s
);
1723 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
1724 const char *name
, size_t size
,
1725 size_t align
, unsigned long flags
,
1726 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
1727 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
1729 memset(s
, 0, kmem_size
);
1738 * The page->offset field is only 16 bit wide. This is an offset
1739 * in units of words from the beginning of an object. If the slab
1740 * size is bigger then we cannot move the free pointer behind the
1743 * On 32 bit platforms the limit is 256k. On 64bit platforms
1744 * the limit is 512k.
1746 * Debugging or ctor/dtors may create a need to move the free
1747 * pointer. Fail if this happens.
1749 if (s
->size
>= 65535 * sizeof(void *)) {
1750 BUG_ON(flags
& (SLAB_RED_ZONE
| SLAB_POISON
|
1751 SLAB_STORE_USER
| SLAB_DESTROY_BY_RCU
));
1752 BUG_ON(ctor
|| dtor
);
1756 * Enable debugging if selected on the kernel commandline.
1758 if (slub_debug
&& (!slub_debug_slabs
||
1759 strncmp(slub_debug_slabs
, name
,
1760 strlen(slub_debug_slabs
)) == 0))
1761 s
->flags
|= slub_debug
;
1763 if (!calculate_sizes(s
))
1768 s
->defrag_ratio
= 100;
1771 if (init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
1774 if (flags
& SLAB_PANIC
)
1775 panic("Cannot create slab %s size=%lu realsize=%u "
1776 "order=%u offset=%u flags=%lx\n",
1777 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
1781 EXPORT_SYMBOL(kmem_cache_open
);
1784 * Check if a given pointer is valid
1786 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
1791 page
= get_object_page(object
);
1793 if (!page
|| s
!= page
->slab
)
1794 /* No slab or wrong slab */
1797 addr
= page_address(page
);
1798 if (object
< addr
|| object
>= addr
+ s
->objects
* s
->size
)
1802 if ((object
- addr
) % s
->size
)
1803 /* Improperly aligned */
1807 * We could also check if the object is on the slabs freelist.
1808 * But this would be too expensive and it seems that the main
1809 * purpose of kmem_ptr_valid is to check if the object belongs
1810 * to a certain slab.
1814 EXPORT_SYMBOL(kmem_ptr_validate
);
1817 * Determine the size of a slab object
1819 unsigned int kmem_cache_size(struct kmem_cache
*s
)
1823 EXPORT_SYMBOL(kmem_cache_size
);
1825 const char *kmem_cache_name(struct kmem_cache
*s
)
1829 EXPORT_SYMBOL(kmem_cache_name
);
1832 * Attempt to free all slabs on a node
1834 static int free_list(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1835 struct list_head
*list
)
1837 int slabs_inuse
= 0;
1838 unsigned long flags
;
1839 struct page
*page
, *h
;
1841 spin_lock_irqsave(&n
->list_lock
, flags
);
1842 list_for_each_entry_safe(page
, h
, list
, lru
)
1844 list_del(&page
->lru
);
1845 discard_slab(s
, page
);
1848 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1853 * Release all resources used by slab cache
1855 static int kmem_cache_close(struct kmem_cache
*s
)
1861 /* Attempt to free all objects */
1862 for_each_online_node(node
) {
1863 struct kmem_cache_node
*n
= get_node(s
, node
);
1865 n
->nr_partial
-= free_list(s
, n
, &n
->partial
);
1866 if (atomic_long_read(&n
->nr_slabs
))
1869 free_kmem_cache_nodes(s
);
1874 * Close a cache and release the kmem_cache structure
1875 * (must be used for caches created using kmem_cache_create)
1877 void kmem_cache_destroy(struct kmem_cache
*s
)
1879 down_write(&slub_lock
);
1883 if (kmem_cache_close(s
))
1885 sysfs_slab_remove(s
);
1888 up_write(&slub_lock
);
1890 EXPORT_SYMBOL(kmem_cache_destroy
);
1892 /********************************************************************
1894 *******************************************************************/
1896 struct kmem_cache kmalloc_caches
[KMALLOC_SHIFT_HIGH
+ 1] __cacheline_aligned
;
1897 EXPORT_SYMBOL(kmalloc_caches
);
1899 #ifdef CONFIG_ZONE_DMA
1900 static struct kmem_cache
*kmalloc_caches_dma
[KMALLOC_SHIFT_HIGH
+ 1];
1903 static int __init
setup_slub_min_order(char *str
)
1905 get_option (&str
, &slub_min_order
);
1910 __setup("slub_min_order=", setup_slub_min_order
);
1912 static int __init
setup_slub_max_order(char *str
)
1914 get_option (&str
, &slub_max_order
);
1919 __setup("slub_max_order=", setup_slub_max_order
);
1921 static int __init
setup_slub_min_objects(char *str
)
1923 get_option (&str
, &slub_min_objects
);
1928 __setup("slub_min_objects=", setup_slub_min_objects
);
1930 static int __init
setup_slub_nomerge(char *str
)
1936 __setup("slub_nomerge", setup_slub_nomerge
);
1938 static int __init
setup_slub_debug(char *str
)
1940 if (!str
|| *str
!= '=')
1941 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1944 if (*str
== 0 || *str
== ',')
1945 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1947 for( ;*str
&& *str
!= ','; str
++)
1949 case 'f' : case 'F' :
1950 slub_debug
|= SLAB_DEBUG_FREE
;
1952 case 'z' : case 'Z' :
1953 slub_debug
|= SLAB_RED_ZONE
;
1955 case 'p' : case 'P' :
1956 slub_debug
|= SLAB_POISON
;
1958 case 'u' : case 'U' :
1959 slub_debug
|= SLAB_STORE_USER
;
1961 case 't' : case 'T' :
1962 slub_debug
|= SLAB_TRACE
;
1965 printk(KERN_ERR
"slub_debug option '%c' "
1966 "unknown. skipped\n",*str
);
1971 slub_debug_slabs
= str
+ 1;
1975 __setup("slub_debug", setup_slub_debug
);
1977 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
1978 const char *name
, int size
, gfp_t gfp_flags
)
1980 unsigned int flags
= 0;
1982 if (gfp_flags
& SLUB_DMA
)
1983 flags
= SLAB_CACHE_DMA
;
1985 down_write(&slub_lock
);
1986 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
1990 list_add(&s
->list
, &slab_caches
);
1991 up_write(&slub_lock
);
1992 if (sysfs_slab_add(s
))
1997 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2000 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2002 int index
= kmalloc_index(size
);
2007 /* Allocation too large? */
2010 #ifdef CONFIG_ZONE_DMA
2011 if ((flags
& SLUB_DMA
)) {
2012 struct kmem_cache
*s
;
2013 struct kmem_cache
*x
;
2017 s
= kmalloc_caches_dma
[index
];
2021 /* Dynamically create dma cache */
2022 x
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2024 panic("Unable to allocate memory for dma cache\n");
2026 if (index
<= KMALLOC_SHIFT_HIGH
)
2027 realsize
= 1 << index
;
2035 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d",
2036 (unsigned int)realsize
);
2037 s
= create_kmalloc_cache(x
, text
, realsize
, flags
);
2038 kmalloc_caches_dma
[index
] = s
;
2042 return &kmalloc_caches
[index
];
2045 void *__kmalloc(size_t size
, gfp_t flags
)
2047 struct kmem_cache
*s
= get_slab(size
, flags
);
2050 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2053 EXPORT_SYMBOL(__kmalloc
);
2056 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2058 struct kmem_cache
*s
= get_slab(size
, flags
);
2061 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2064 EXPORT_SYMBOL(__kmalloc_node
);
2067 size_t ksize(const void *object
)
2069 struct page
*page
= get_object_page(object
);
2070 struct kmem_cache
*s
;
2077 * Debugging requires use of the padding between object
2078 * and whatever may come after it.
2080 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2084 * If we have the need to store the freelist pointer
2085 * back there or track user information then we can
2086 * only use the space before that information.
2088 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2092 * Else we can use all the padding etc for the allocation
2096 EXPORT_SYMBOL(ksize
);
2098 void kfree(const void *x
)
2100 struct kmem_cache
*s
;
2106 page
= virt_to_head_page(x
);
2109 slab_free(s
, page
, (void *)x
, __builtin_return_address(0));
2111 EXPORT_SYMBOL(kfree
);
2114 * kmem_cache_shrink removes empty slabs from the partial lists
2115 * and then sorts the partially allocated slabs by the number
2116 * of items in use. The slabs with the most items in use
2117 * come first. New allocations will remove these from the
2118 * partial list because they are full. The slabs with the
2119 * least items are placed last. If it happens that the objects
2120 * are freed then the page can be returned to the page allocator.
2122 int kmem_cache_shrink(struct kmem_cache
*s
)
2126 struct kmem_cache_node
*n
;
2129 struct list_head
*slabs_by_inuse
=
2130 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2131 unsigned long flags
;
2133 if (!slabs_by_inuse
)
2137 for_each_online_node(node
) {
2138 n
= get_node(s
, node
);
2143 for (i
= 0; i
< s
->objects
; i
++)
2144 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2146 spin_lock_irqsave(&n
->list_lock
, flags
);
2149 * Build lists indexed by the items in use in
2150 * each slab or free slabs if empty.
2152 * Note that concurrent frees may occur while
2153 * we hold the list_lock. page->inuse here is
2156 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2157 if (!page
->inuse
&& slab_trylock(page
)) {
2159 * Must hold slab lock here because slab_free
2160 * may have freed the last object and be
2161 * waiting to release the slab.
2163 list_del(&page
->lru
);
2166 discard_slab(s
, page
);
2168 if (n
->nr_partial
> MAX_PARTIAL
)
2169 list_move(&page
->lru
,
2170 slabs_by_inuse
+ page
->inuse
);
2174 if (n
->nr_partial
<= MAX_PARTIAL
)
2178 * Rebuild the partial list with the slabs filled up
2179 * most first and the least used slabs at the end.
2181 for (i
= s
->objects
- 1; i
>= 0; i
--)
2182 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2185 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2188 kfree(slabs_by_inuse
);
2191 EXPORT_SYMBOL(kmem_cache_shrink
);
2194 * krealloc - reallocate memory. The contents will remain unchanged.
2196 * @p: object to reallocate memory for.
2197 * @new_size: how many bytes of memory are required.
2198 * @flags: the type of memory to allocate.
2200 * The contents of the object pointed to are preserved up to the
2201 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2202 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2203 * %NULL pointer, the object pointed to is freed.
2205 void *krealloc(const void *p
, size_t new_size
, gfp_t flags
)
2207 struct kmem_cache
*new_cache
;
2212 return kmalloc(new_size
, flags
);
2214 if (unlikely(!new_size
)) {
2219 page
= virt_to_head_page(p
);
2221 new_cache
= get_slab(new_size
, flags
);
2224 * If new size fits in the current cache, bail out.
2226 if (likely(page
->slab
== new_cache
))
2229 ret
= kmalloc(new_size
, flags
);
2231 memcpy(ret
, p
, min(new_size
, ksize(p
)));
2236 EXPORT_SYMBOL(krealloc
);
2238 /********************************************************************
2239 * Basic setup of slabs
2240 *******************************************************************/
2242 void __init
kmem_cache_init(void)
2248 * Must first have the slab cache available for the allocations of the
2249 * struct kmalloc_cache_node's. There is special bootstrap code in
2250 * kmem_cache_open for slab_state == DOWN.
2252 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2253 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2256 /* Able to allocate the per node structures */
2257 slab_state
= PARTIAL
;
2259 /* Caches that are not of the two-to-the-power-of size */
2260 create_kmalloc_cache(&kmalloc_caches
[1],
2261 "kmalloc-96", 96, GFP_KERNEL
);
2262 create_kmalloc_cache(&kmalloc_caches
[2],
2263 "kmalloc-192", 192, GFP_KERNEL
);
2265 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2266 create_kmalloc_cache(&kmalloc_caches
[i
],
2267 "kmalloc", 1 << i
, GFP_KERNEL
);
2271 /* Provide the correct kmalloc names now that the caches are up */
2272 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2273 kmalloc_caches
[i
]. name
=
2274 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
2277 register_cpu_notifier(&slab_notifier
);
2280 if (nr_cpu_ids
) /* Remove when nr_cpu_ids is fixed upstream ! */
2281 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
)
2282 + nr_cpu_ids
* sizeof(struct page
*);
2284 printk(KERN_INFO
"SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2285 " Processors=%d, Nodes=%d\n",
2286 KMALLOC_SHIFT_HIGH
, L1_CACHE_BYTES
,
2287 slub_min_order
, slub_max_order
, slub_min_objects
,
2288 nr_cpu_ids
, nr_node_ids
);
2292 * Find a mergeable slab cache
2294 static int slab_unmergeable(struct kmem_cache
*s
)
2296 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
2299 if (s
->ctor
|| s
->dtor
)
2305 static struct kmem_cache
*find_mergeable(size_t size
,
2306 size_t align
, unsigned long flags
,
2307 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
2308 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
2310 struct list_head
*h
;
2312 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
2318 size
= ALIGN(size
, sizeof(void *));
2319 align
= calculate_alignment(flags
, align
, size
);
2320 size
= ALIGN(size
, align
);
2322 list_for_each(h
, &slab_caches
) {
2323 struct kmem_cache
*s
=
2324 container_of(h
, struct kmem_cache
, list
);
2326 if (slab_unmergeable(s
))
2332 if (((flags
| slub_debug
) & SLUB_MERGE_SAME
) !=
2333 (s
->flags
& SLUB_MERGE_SAME
))
2336 * Check if alignment is compatible.
2337 * Courtesy of Adrian Drzewiecki
2339 if ((s
->size
& ~(align
-1)) != s
->size
)
2342 if (s
->size
- size
>= sizeof(void *))
2350 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
2351 size_t align
, unsigned long flags
,
2352 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
2353 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
2355 struct kmem_cache
*s
;
2357 down_write(&slub_lock
);
2358 s
= find_mergeable(size
, align
, flags
, dtor
, ctor
);
2362 * Adjust the object sizes so that we clear
2363 * the complete object on kzalloc.
2365 s
->objsize
= max(s
->objsize
, (int)size
);
2366 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
2367 if (sysfs_slab_alias(s
, name
))
2370 s
= kmalloc(kmem_size
, GFP_KERNEL
);
2371 if (s
&& kmem_cache_open(s
, GFP_KERNEL
, name
,
2372 size
, align
, flags
, ctor
, dtor
)) {
2373 if (sysfs_slab_add(s
)) {
2377 list_add(&s
->list
, &slab_caches
);
2381 up_write(&slub_lock
);
2385 up_write(&slub_lock
);
2386 if (flags
& SLAB_PANIC
)
2387 panic("Cannot create slabcache %s\n", name
);
2392 EXPORT_SYMBOL(kmem_cache_create
);
2394 void *kmem_cache_zalloc(struct kmem_cache
*s
, gfp_t flags
)
2398 x
= slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2400 memset(x
, 0, s
->objsize
);
2403 EXPORT_SYMBOL(kmem_cache_zalloc
);
2406 static void for_all_slabs(void (*func
)(struct kmem_cache
*, int), int cpu
)
2408 struct list_head
*h
;
2410 down_read(&slub_lock
);
2411 list_for_each(h
, &slab_caches
) {
2412 struct kmem_cache
*s
=
2413 container_of(h
, struct kmem_cache
, list
);
2417 up_read(&slub_lock
);
2421 * Use the cpu notifier to insure that the slab are flushed
2424 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
2425 unsigned long action
, void *hcpu
)
2427 long cpu
= (long)hcpu
;
2430 case CPU_UP_CANCELED
:
2432 for_all_slabs(__flush_cpu_slab
, cpu
);
2440 static struct notifier_block __cpuinitdata slab_notifier
=
2441 { &slab_cpuup_callback
, NULL
, 0 };
2447 /*****************************************************************
2448 * Generic reaper used to support the page allocator
2449 * (the cpu slabs are reaped by a per slab workqueue).
2451 * Maybe move this to the page allocator?
2452 ****************************************************************/
2454 static DEFINE_PER_CPU(unsigned long, reap_node
);
2456 static void init_reap_node(int cpu
)
2460 node
= next_node(cpu_to_node(cpu
), node_online_map
);
2461 if (node
== MAX_NUMNODES
)
2462 node
= first_node(node_online_map
);
2464 __get_cpu_var(reap_node
) = node
;
2467 static void next_reap_node(void)
2469 int node
= __get_cpu_var(reap_node
);
2472 * Also drain per cpu pages on remote zones
2474 if (node
!= numa_node_id())
2475 drain_node_pages(node
);
2477 node
= next_node(node
, node_online_map
);
2478 if (unlikely(node
>= MAX_NUMNODES
))
2479 node
= first_node(node_online_map
);
2480 __get_cpu_var(reap_node
) = node
;
2483 #define init_reap_node(cpu) do { } while (0)
2484 #define next_reap_node(void) do { } while (0)
2487 #define REAPTIMEOUT_CPUC (2*HZ)
2490 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
2492 static void cache_reap(struct work_struct
*unused
)
2495 refresh_cpu_vm_stats(smp_processor_id());
2496 schedule_delayed_work(&__get_cpu_var(reap_work
),
2500 static void __devinit
start_cpu_timer(int cpu
)
2502 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
2505 * When this gets called from do_initcalls via cpucache_init(),
2506 * init_workqueues() has already run, so keventd will be setup
2509 if (keventd_up() && reap_work
->work
.func
== NULL
) {
2510 init_reap_node(cpu
);
2511 INIT_DELAYED_WORK(reap_work
, cache_reap
);
2512 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
2516 static int __init
cpucache_init(void)
2521 * Register the timers that drain pcp pages and update vm statistics
2523 for_each_online_cpu(cpu
)
2524 start_cpu_timer(cpu
);
2527 __initcall(cpucache_init
);
2530 #ifdef SLUB_RESILIENCY_TEST
2531 static unsigned long validate_slab_cache(struct kmem_cache
*s
);
2533 static void resiliency_test(void)
2537 printk(KERN_ERR
"SLUB resiliency testing\n");
2538 printk(KERN_ERR
"-----------------------\n");
2539 printk(KERN_ERR
"A. Corruption after allocation\n");
2541 p
= kzalloc(16, GFP_KERNEL
);
2543 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
2544 " 0x12->0x%p\n\n", p
+ 16);
2546 validate_slab_cache(kmalloc_caches
+ 4);
2548 /* Hmmm... The next two are dangerous */
2549 p
= kzalloc(32, GFP_KERNEL
);
2550 p
[32 + sizeof(void *)] = 0x34;
2551 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
2552 " 0x34 -> -0x%p\n", p
);
2553 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2555 validate_slab_cache(kmalloc_caches
+ 5);
2556 p
= kzalloc(64, GFP_KERNEL
);
2557 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
2559 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2561 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2562 validate_slab_cache(kmalloc_caches
+ 6);
2564 printk(KERN_ERR
"\nB. Corruption after free\n");
2565 p
= kzalloc(128, GFP_KERNEL
);
2568 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
2569 validate_slab_cache(kmalloc_caches
+ 7);
2571 p
= kzalloc(256, GFP_KERNEL
);
2574 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
2575 validate_slab_cache(kmalloc_caches
+ 8);
2577 p
= kzalloc(512, GFP_KERNEL
);
2580 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
2581 validate_slab_cache(kmalloc_caches
+ 9);
2584 static void resiliency_test(void) {};
2588 * These are not as efficient as kmalloc for the non debug case.
2589 * We do not have the page struct available so we have to touch one
2590 * cacheline in struct kmem_cache to check slab flags.
2592 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
2594 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2599 return slab_alloc(s
, gfpflags
, -1, caller
);
2602 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
2603 int node
, void *caller
)
2605 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2610 return slab_alloc(s
, gfpflags
, node
, caller
);
2615 static int validate_slab(struct kmem_cache
*s
, struct page
*page
)
2618 void *addr
= page_address(page
);
2619 unsigned long map
[BITS_TO_LONGS(s
->objects
)];
2621 if (!check_slab(s
, page
) ||
2622 !on_freelist(s
, page
, NULL
))
2625 /* Now we know that a valid freelist exists */
2626 bitmap_zero(map
, s
->objects
);
2628 for(p
= page
->freelist
; p
; p
= get_freepointer(s
, p
)) {
2629 set_bit((p
- addr
) / s
->size
, map
);
2630 if (!check_object(s
, page
, p
, 0))
2634 for(p
= addr
; p
< addr
+ s
->objects
* s
->size
; p
+= s
->size
)
2635 if (!test_bit((p
- addr
) / s
->size
, map
))
2636 if (!check_object(s
, page
, p
, 1))
2641 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
)
2643 if (slab_trylock(page
)) {
2644 validate_slab(s
, page
);
2647 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
2650 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
2651 if (!PageError(page
))
2652 printk(KERN_ERR
"SLUB %s: PageError not set "
2653 "on slab 0x%p\n", s
->name
, page
);
2655 if (PageError(page
))
2656 printk(KERN_ERR
"SLUB %s: PageError set on "
2657 "slab 0x%p\n", s
->name
, page
);
2661 static int validate_slab_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
2663 unsigned long count
= 0;
2665 unsigned long flags
;
2667 spin_lock_irqsave(&n
->list_lock
, flags
);
2669 list_for_each_entry(page
, &n
->partial
, lru
) {
2670 validate_slab_slab(s
, page
);
2673 if (count
!= n
->nr_partial
)
2674 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
2675 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
2677 if (!(s
->flags
& SLAB_STORE_USER
))
2680 list_for_each_entry(page
, &n
->full
, lru
) {
2681 validate_slab_slab(s
, page
);
2684 if (count
!= atomic_long_read(&n
->nr_slabs
))
2685 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
2686 "counter=%ld\n", s
->name
, count
,
2687 atomic_long_read(&n
->nr_slabs
));
2690 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2694 static unsigned long validate_slab_cache(struct kmem_cache
*s
)
2697 unsigned long count
= 0;
2700 for_each_online_node(node
) {
2701 struct kmem_cache_node
*n
= get_node(s
, node
);
2703 count
+= validate_slab_node(s
, n
);
2709 * Generate lists of locations where slabcache objects are allocated
2714 unsigned long count
;
2720 unsigned long count
;
2721 struct location
*loc
;
2724 static void free_loc_track(struct loc_track
*t
)
2727 free_pages((unsigned long)t
->loc
,
2728 get_order(sizeof(struct location
) * t
->max
));
2731 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
)
2737 max
= PAGE_SIZE
/ sizeof(struct location
);
2739 order
= get_order(sizeof(struct location
) * max
);
2741 l
= (void *)__get_free_pages(GFP_KERNEL
, order
);
2747 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
2755 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
2758 long start
, end
, pos
;
2766 pos
= start
+ (end
- start
+ 1) / 2;
2769 * There is nothing at "end". If we end up there
2770 * we need to add something to before end.
2775 caddr
= t
->loc
[pos
].addr
;
2776 if (addr
== caddr
) {
2777 t
->loc
[pos
].count
++;
2788 * Not found. Insert new tracking element
2790 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
))
2796 (t
->count
- pos
) * sizeof(struct location
));
2803 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
2804 struct page
*page
, enum track_item alloc
)
2806 void *addr
= page_address(page
);
2807 unsigned long map
[BITS_TO_LONGS(s
->objects
)];
2810 bitmap_zero(map
, s
->objects
);
2811 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
2812 set_bit((p
- addr
) / s
->size
, map
);
2814 for (p
= addr
; p
< addr
+ s
->objects
* s
->size
; p
+= s
->size
)
2815 if (!test_bit((p
- addr
) / s
->size
, map
)) {
2816 void *addr
= get_track(s
, p
, alloc
)->addr
;
2818 add_location(t
, s
, addr
);
2822 static int list_locations(struct kmem_cache
*s
, char *buf
,
2823 enum track_item alloc
)
2833 /* Push back cpu slabs */
2836 for_each_online_node(node
) {
2837 struct kmem_cache_node
*n
= get_node(s
, node
);
2838 unsigned long flags
;
2841 if (!atomic_read(&n
->nr_slabs
))
2844 spin_lock_irqsave(&n
->list_lock
, flags
);
2845 list_for_each_entry(page
, &n
->partial
, lru
)
2846 process_slab(&t
, s
, page
, alloc
);
2847 list_for_each_entry(page
, &n
->full
, lru
)
2848 process_slab(&t
, s
, page
, alloc
);
2849 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2852 for (i
= 0; i
< t
.count
; i
++) {
2853 void *addr
= t
.loc
[i
].addr
;
2855 if (n
> PAGE_SIZE
- 100)
2857 n
+= sprintf(buf
+ n
, "%7ld ", t
.loc
[i
].count
);
2859 n
+= sprint_symbol(buf
+ n
, (unsigned long)t
.loc
[i
].addr
);
2861 n
+= sprintf(buf
+ n
, "<not-available>");
2862 n
+= sprintf(buf
+ n
, "\n");
2867 n
+= sprintf(buf
, "No data\n");
2871 static unsigned long count_partial(struct kmem_cache_node
*n
)
2873 unsigned long flags
;
2874 unsigned long x
= 0;
2877 spin_lock_irqsave(&n
->list_lock
, flags
);
2878 list_for_each_entry(page
, &n
->partial
, lru
)
2880 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2884 enum slab_stat_type
{
2891 #define SO_FULL (1 << SL_FULL)
2892 #define SO_PARTIAL (1 << SL_PARTIAL)
2893 #define SO_CPU (1 << SL_CPU)
2894 #define SO_OBJECTS (1 << SL_OBJECTS)
2896 static unsigned long slab_objects(struct kmem_cache
*s
,
2897 char *buf
, unsigned long flags
)
2899 unsigned long total
= 0;
2903 unsigned long *nodes
;
2904 unsigned long *per_cpu
;
2906 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
2907 per_cpu
= nodes
+ nr_node_ids
;
2909 for_each_possible_cpu(cpu
) {
2910 struct page
*page
= s
->cpu_slab
[cpu
];
2914 node
= page_to_nid(page
);
2915 if (flags
& SO_CPU
) {
2918 if (flags
& SO_OBJECTS
)
2929 for_each_online_node(node
) {
2930 struct kmem_cache_node
*n
= get_node(s
, node
);
2932 if (flags
& SO_PARTIAL
) {
2933 if (flags
& SO_OBJECTS
)
2934 x
= count_partial(n
);
2941 if (flags
& SO_FULL
) {
2942 int full_slabs
= atomic_read(&n
->nr_slabs
)
2946 if (flags
& SO_OBJECTS
)
2947 x
= full_slabs
* s
->objects
;
2955 x
= sprintf(buf
, "%lu", total
);
2957 for_each_online_node(node
)
2959 x
+= sprintf(buf
+ x
, " N%d=%lu",
2963 return x
+ sprintf(buf
+ x
, "\n");
2966 static int any_slab_objects(struct kmem_cache
*s
)
2971 for_each_possible_cpu(cpu
)
2972 if (s
->cpu_slab
[cpu
])
2975 for_each_node(node
) {
2976 struct kmem_cache_node
*n
= get_node(s
, node
);
2978 if (n
->nr_partial
|| atomic_read(&n
->nr_slabs
))
2984 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2985 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
2987 struct slab_attribute
{
2988 struct attribute attr
;
2989 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
2990 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
2993 #define SLAB_ATTR_RO(_name) \
2994 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2996 #define SLAB_ATTR(_name) \
2997 static struct slab_attribute _name##_attr = \
2998 __ATTR(_name, 0644, _name##_show, _name##_store)
3000 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
3002 return sprintf(buf
, "%d\n", s
->size
);
3004 SLAB_ATTR_RO(slab_size
);
3006 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3008 return sprintf(buf
, "%d\n", s
->align
);
3010 SLAB_ATTR_RO(align
);
3012 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3014 return sprintf(buf
, "%d\n", s
->objsize
);
3016 SLAB_ATTR_RO(object_size
);
3018 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3020 return sprintf(buf
, "%d\n", s
->objects
);
3022 SLAB_ATTR_RO(objs_per_slab
);
3024 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3026 return sprintf(buf
, "%d\n", s
->order
);
3028 SLAB_ATTR_RO(order
);
3030 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3033 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3035 return n
+ sprintf(buf
+ n
, "\n");
3041 static ssize_t
dtor_show(struct kmem_cache
*s
, char *buf
)
3044 int n
= sprint_symbol(buf
, (unsigned long)s
->dtor
);
3046 return n
+ sprintf(buf
+ n
, "\n");
3052 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3054 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3056 SLAB_ATTR_RO(aliases
);
3058 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3060 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3062 SLAB_ATTR_RO(slabs
);
3064 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3066 return slab_objects(s
, buf
, SO_PARTIAL
);
3068 SLAB_ATTR_RO(partial
);
3070 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3072 return slab_objects(s
, buf
, SO_CPU
);
3074 SLAB_ATTR_RO(cpu_slabs
);
3076 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3078 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3080 SLAB_ATTR_RO(objects
);
3082 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3084 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3087 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3088 const char *buf
, size_t length
)
3090 s
->flags
&= ~SLAB_DEBUG_FREE
;
3092 s
->flags
|= SLAB_DEBUG_FREE
;
3095 SLAB_ATTR(sanity_checks
);
3097 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3099 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3102 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3105 s
->flags
&= ~SLAB_TRACE
;
3107 s
->flags
|= SLAB_TRACE
;
3112 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3114 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3117 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3118 const char *buf
, size_t length
)
3120 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3122 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3125 SLAB_ATTR(reclaim_account
);
3127 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3129 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3131 SLAB_ATTR_RO(hwcache_align
);
3133 #ifdef CONFIG_ZONE_DMA
3134 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3136 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3138 SLAB_ATTR_RO(cache_dma
);
3141 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3143 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3145 SLAB_ATTR_RO(destroy_by_rcu
);
3147 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3149 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3152 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3153 const char *buf
, size_t length
)
3155 if (any_slab_objects(s
))
3158 s
->flags
&= ~SLAB_RED_ZONE
;
3160 s
->flags
|= SLAB_RED_ZONE
;
3164 SLAB_ATTR(red_zone
);
3166 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3168 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3171 static ssize_t
poison_store(struct kmem_cache
*s
,
3172 const char *buf
, size_t length
)
3174 if (any_slab_objects(s
))
3177 s
->flags
&= ~SLAB_POISON
;
3179 s
->flags
|= SLAB_POISON
;
3185 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3187 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3190 static ssize_t
store_user_store(struct kmem_cache
*s
,
3191 const char *buf
, size_t length
)
3193 if (any_slab_objects(s
))
3196 s
->flags
&= ~SLAB_STORE_USER
;
3198 s
->flags
|= SLAB_STORE_USER
;
3202 SLAB_ATTR(store_user
);
3204 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3209 static ssize_t
validate_store(struct kmem_cache
*s
,
3210 const char *buf
, size_t length
)
3213 validate_slab_cache(s
);
3218 SLAB_ATTR(validate
);
3220 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3225 static ssize_t
shrink_store(struct kmem_cache
*s
,
3226 const char *buf
, size_t length
)
3228 if (buf
[0] == '1') {
3229 int rc
= kmem_cache_shrink(s
);
3239 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
3241 if (!(s
->flags
& SLAB_STORE_USER
))
3243 return list_locations(s
, buf
, TRACK_ALLOC
);
3245 SLAB_ATTR_RO(alloc_calls
);
3247 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
3249 if (!(s
->flags
& SLAB_STORE_USER
))
3251 return list_locations(s
, buf
, TRACK_FREE
);
3253 SLAB_ATTR_RO(free_calls
);
3256 static ssize_t
defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
3258 return sprintf(buf
, "%d\n", s
->defrag_ratio
/ 10);
3261 static ssize_t
defrag_ratio_store(struct kmem_cache
*s
,
3262 const char *buf
, size_t length
)
3264 int n
= simple_strtoul(buf
, NULL
, 10);
3267 s
->defrag_ratio
= n
* 10;
3270 SLAB_ATTR(defrag_ratio
);
3273 static struct attribute
* slab_attrs
[] = {
3274 &slab_size_attr
.attr
,
3275 &object_size_attr
.attr
,
3276 &objs_per_slab_attr
.attr
,
3281 &cpu_slabs_attr
.attr
,
3286 &sanity_checks_attr
.attr
,
3288 &hwcache_align_attr
.attr
,
3289 &reclaim_account_attr
.attr
,
3290 &destroy_by_rcu_attr
.attr
,
3291 &red_zone_attr
.attr
,
3293 &store_user_attr
.attr
,
3294 &validate_attr
.attr
,
3296 &alloc_calls_attr
.attr
,
3297 &free_calls_attr
.attr
,
3298 #ifdef CONFIG_ZONE_DMA
3299 &cache_dma_attr
.attr
,
3302 &defrag_ratio_attr
.attr
,
3307 static struct attribute_group slab_attr_group
= {
3308 .attrs
= slab_attrs
,
3311 static ssize_t
slab_attr_show(struct kobject
*kobj
,
3312 struct attribute
*attr
,
3315 struct slab_attribute
*attribute
;
3316 struct kmem_cache
*s
;
3319 attribute
= to_slab_attr(attr
);
3322 if (!attribute
->show
)
3325 err
= attribute
->show(s
, buf
);
3330 static ssize_t
slab_attr_store(struct kobject
*kobj
,
3331 struct attribute
*attr
,
3332 const char *buf
, size_t len
)
3334 struct slab_attribute
*attribute
;
3335 struct kmem_cache
*s
;
3338 attribute
= to_slab_attr(attr
);
3341 if (!attribute
->store
)
3344 err
= attribute
->store(s
, buf
, len
);
3349 static struct sysfs_ops slab_sysfs_ops
= {
3350 .show
= slab_attr_show
,
3351 .store
= slab_attr_store
,
3354 static struct kobj_type slab_ktype
= {
3355 .sysfs_ops
= &slab_sysfs_ops
,
3358 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
3360 struct kobj_type
*ktype
= get_ktype(kobj
);
3362 if (ktype
== &slab_ktype
)
3367 static struct kset_uevent_ops slab_uevent_ops
= {
3368 .filter
= uevent_filter
,
3371 decl_subsys(slab
, &slab_ktype
, &slab_uevent_ops
);
3373 #define ID_STR_LENGTH 64
3375 /* Create a unique string id for a slab cache:
3377 * :[flags-]size:[memory address of kmemcache]
3379 static char *create_unique_id(struct kmem_cache
*s
)
3381 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
3388 * First flags affecting slabcache operations. We will only
3389 * get here for aliasable slabs so we do not need to support
3390 * too many flags. The flags here must cover all flags that
3391 * are matched during merging to guarantee that the id is
3394 if (s
->flags
& SLAB_CACHE_DMA
)
3396 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3398 if (s
->flags
& SLAB_DEBUG_FREE
)
3402 p
+= sprintf(p
, "%07d", s
->size
);
3403 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
3407 static int sysfs_slab_add(struct kmem_cache
*s
)
3413 if (slab_state
< SYSFS
)
3414 /* Defer until later */
3417 unmergeable
= slab_unmergeable(s
);
3420 * Slabcache can never be merged so we can use the name proper.
3421 * This is typically the case for debug situations. In that
3422 * case we can catch duplicate names easily.
3424 sysfs_remove_link(&slab_subsys
.kset
.kobj
, s
->name
);
3428 * Create a unique name for the slab as a target
3431 name
= create_unique_id(s
);
3434 kobj_set_kset_s(s
, slab_subsys
);
3435 kobject_set_name(&s
->kobj
, name
);
3436 kobject_init(&s
->kobj
);
3437 err
= kobject_add(&s
->kobj
);
3441 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
3444 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
3446 /* Setup first alias */
3447 sysfs_slab_alias(s
, s
->name
);
3453 static void sysfs_slab_remove(struct kmem_cache
*s
)
3455 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
3456 kobject_del(&s
->kobj
);
3460 * Need to buffer aliases during bootup until sysfs becomes
3461 * available lest we loose that information.
3463 struct saved_alias
{
3464 struct kmem_cache
*s
;
3466 struct saved_alias
*next
;
3469 struct saved_alias
*alias_list
;
3471 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
3473 struct saved_alias
*al
;
3475 if (slab_state
== SYSFS
) {
3477 * If we have a leftover link then remove it.
3479 sysfs_remove_link(&slab_subsys
.kset
.kobj
, name
);
3480 return sysfs_create_link(&slab_subsys
.kset
.kobj
,
3484 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
3490 al
->next
= alias_list
;
3495 static int __init
slab_sysfs_init(void)
3499 err
= subsystem_register(&slab_subsys
);
3501 printk(KERN_ERR
"Cannot register slab subsystem.\n");
3507 while (alias_list
) {
3508 struct saved_alias
*al
= alias_list
;
3510 alias_list
= alias_list
->next
;
3511 err
= sysfs_slab_alias(al
->s
, al
->name
);
3520 __initcall(slab_sysfs_init
);
3522 __initcall(finish_bootstrap
);