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
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemcheck.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
31 #include <trace/events/kmem.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has noone operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache
*s
)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
135 * Mininum number of partial slabs. These will be left on the partial
136 * lists even if they are empty. kmem_cache_shrink may reclaim them.
138 #define MIN_PARTIAL 5
141 * Maximum number of desirable partial slabs.
142 * The existence of more partial slabs makes kmem_cache_shrink
143 * sort the partial list by the number of objects in the.
145 #define MAX_PARTIAL 10
147 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
148 SLAB_POISON | SLAB_STORE_USER)
151 * Debugging flags that require metadata to be stored in the slab. These get
152 * disabled when slub_debug=O is used and a cache's min order increases with
155 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
158 * Set of flags that will prevent slab merging
160 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
164 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
165 SLAB_CACHE_DMA | SLAB_NOTRACK)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000UL /* Poison object */
174 static int kmem_size
= sizeof(struct kmem_cache
);
177 static struct notifier_block slab_notifier
;
181 DOWN
, /* No slab functionality available */
182 PARTIAL
, /* Kmem_cache_node works */
183 UP
, /* Everything works but does not show up in sysfs */
187 /* A list of all slab caches on the system */
188 static DECLARE_RWSEM(slub_lock
);
189 static LIST_HEAD(slab_caches
);
192 * Tracking user of a slab.
195 unsigned long addr
; /* Called from address */
196 int cpu
; /* Was running on cpu */
197 int pid
; /* Pid context */
198 unsigned long when
; /* When did the operation occur */
201 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
204 static int sysfs_slab_add(struct kmem_cache
*);
205 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
206 static void sysfs_slab_remove(struct kmem_cache
*);
209 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
210 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
212 static inline void sysfs_slab_remove(struct kmem_cache
*s
)
220 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
222 #ifdef CONFIG_SLUB_STATS
223 __this_cpu_inc(s
->cpu_slab
->stat
[si
]);
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 int slab_is_available(void)
233 return slab_state
>= UP
;
236 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
238 return s
->node
[node
];
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache
*s
,
243 struct page
*page
, const void *object
)
250 base
= page_address(page
);
251 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
252 (object
- base
) % s
->size
) {
259 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
261 return *(void **)(object
+ s
->offset
);
264 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
266 *(void **)(object
+ s
->offset
) = fp
;
269 /* Loop over all objects in a slab */
270 #define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
274 /* Determine object index from a given position */
275 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
277 return (p
- addr
) / s
->size
;
280 static inline size_t slab_ksize(const struct kmem_cache
*s
)
282 #ifdef CONFIG_SLUB_DEBUG
284 * Debugging requires use of the padding between object
285 * and whatever may come after it.
287 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
292 * If we have the need to store the freelist pointer
293 * back there or track user information then we can
294 * only use the space before that information.
296 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
299 * Else we can use all the padding etc for the allocation
304 static inline int order_objects(int order
, unsigned long size
, int reserved
)
306 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
309 static inline struct kmem_cache_order_objects
oo_make(int order
,
310 unsigned long size
, int reserved
)
312 struct kmem_cache_order_objects x
= {
313 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
319 static inline int oo_order(struct kmem_cache_order_objects x
)
321 return x
.x
>> OO_SHIFT
;
324 static inline int oo_objects(struct kmem_cache_order_objects x
)
326 return x
.x
& OO_MASK
;
330 * Determine a map of object in use on a page.
332 * Slab lock or node listlock must be held to guarantee that the page does
333 * not vanish from under us.
335 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
338 void *addr
= page_address(page
);
340 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
341 set_bit(slab_index(p
, s
, addr
), map
);
344 #ifdef CONFIG_SLUB_DEBUG
348 #ifdef CONFIG_SLUB_DEBUG_ON
349 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
351 static int slub_debug
;
354 static char *slub_debug_slabs
;
355 static int disable_higher_order_debug
;
360 static void print_section(char *text
, u8
*addr
, unsigned int length
)
368 for (i
= 0; i
< length
; i
++) {
370 printk(KERN_ERR
"%8s 0x%p: ", text
, addr
+ i
);
373 printk(KERN_CONT
" %02x", addr
[i
]);
375 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
377 printk(KERN_CONT
" %s\n", ascii
);
384 printk(KERN_CONT
" ");
388 printk(KERN_CONT
" %s\n", ascii
);
392 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
393 enum track_item alloc
)
398 p
= object
+ s
->offset
+ sizeof(void *);
400 p
= object
+ s
->inuse
;
405 static void set_track(struct kmem_cache
*s
, void *object
,
406 enum track_item alloc
, unsigned long addr
)
408 struct track
*p
= get_track(s
, object
, alloc
);
412 p
->cpu
= smp_processor_id();
413 p
->pid
= current
->pid
;
416 memset(p
, 0, sizeof(struct track
));
419 static void init_tracking(struct kmem_cache
*s
, void *object
)
421 if (!(s
->flags
& SLAB_STORE_USER
))
424 set_track(s
, object
, TRACK_FREE
, 0UL);
425 set_track(s
, object
, TRACK_ALLOC
, 0UL);
428 static void print_track(const char *s
, struct track
*t
)
433 printk(KERN_ERR
"INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
434 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
437 static void print_tracking(struct kmem_cache
*s
, void *object
)
439 if (!(s
->flags
& SLAB_STORE_USER
))
442 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
443 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
446 static void print_page_info(struct page
*page
)
448 printk(KERN_ERR
"INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
449 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
453 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
459 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
461 printk(KERN_ERR
"========================================"
462 "=====================================\n");
463 printk(KERN_ERR
"BUG %s: %s\n", s
->name
, buf
);
464 printk(KERN_ERR
"----------------------------------------"
465 "-------------------------------------\n\n");
468 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
474 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
476 printk(KERN_ERR
"FIX %s: %s\n", s
->name
, buf
);
479 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
481 unsigned int off
; /* Offset of last byte */
482 u8
*addr
= page_address(page
);
484 print_tracking(s
, p
);
486 print_page_info(page
);
488 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
489 p
, p
- addr
, get_freepointer(s
, p
));
492 print_section("Bytes b4", p
- 16, 16);
494 print_section("Object", p
, min_t(unsigned long, s
->objsize
, PAGE_SIZE
));
496 if (s
->flags
& SLAB_RED_ZONE
)
497 print_section("Redzone", p
+ s
->objsize
,
498 s
->inuse
- s
->objsize
);
501 off
= s
->offset
+ sizeof(void *);
505 if (s
->flags
& SLAB_STORE_USER
)
506 off
+= 2 * sizeof(struct track
);
509 /* Beginning of the filler is the free pointer */
510 print_section("Padding", p
+ off
, s
->size
- off
);
515 static void object_err(struct kmem_cache
*s
, struct page
*page
,
516 u8
*object
, char *reason
)
518 slab_bug(s
, "%s", reason
);
519 print_trailer(s
, page
, object
);
522 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *fmt
, ...)
528 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
530 slab_bug(s
, "%s", buf
);
531 print_page_info(page
);
535 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
539 if (s
->flags
& __OBJECT_POISON
) {
540 memset(p
, POISON_FREE
, s
->objsize
- 1);
541 p
[s
->objsize
- 1] = POISON_END
;
544 if (s
->flags
& SLAB_RED_ZONE
)
545 memset(p
+ s
->objsize
, val
, s
->inuse
- s
->objsize
);
548 static u8
*check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
551 if (*start
!= (u8
)value
)
559 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
560 void *from
, void *to
)
562 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
563 memset(from
, data
, to
- from
);
566 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
567 u8
*object
, char *what
,
568 u8
*start
, unsigned int value
, unsigned int bytes
)
573 fault
= check_bytes(start
, value
, bytes
);
578 while (end
> fault
&& end
[-1] == value
)
581 slab_bug(s
, "%s overwritten", what
);
582 printk(KERN_ERR
"INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
583 fault
, end
- 1, fault
[0], value
);
584 print_trailer(s
, page
, object
);
586 restore_bytes(s
, what
, value
, fault
, end
);
594 * Bytes of the object to be managed.
595 * If the freepointer may overlay the object then the free
596 * pointer is the first word of the object.
598 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
601 * object + s->objsize
602 * Padding to reach word boundary. This is also used for Redzoning.
603 * Padding is extended by another word if Redzoning is enabled and
606 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
607 * 0xcc (RED_ACTIVE) for objects in use.
610 * Meta data starts here.
612 * A. Free pointer (if we cannot overwrite object on free)
613 * B. Tracking data for SLAB_STORE_USER
614 * C. Padding to reach required alignment boundary or at mininum
615 * one word if debugging is on to be able to detect writes
616 * before the word boundary.
618 * Padding is done using 0x5a (POISON_INUSE)
621 * Nothing is used beyond s->size.
623 * If slabcaches are merged then the objsize and inuse boundaries are mostly
624 * ignored. And therefore no slab options that rely on these boundaries
625 * may be used with merged slabcaches.
628 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
630 unsigned long off
= s
->inuse
; /* The end of info */
633 /* Freepointer is placed after the object. */
634 off
+= sizeof(void *);
636 if (s
->flags
& SLAB_STORE_USER
)
637 /* We also have user information there */
638 off
+= 2 * sizeof(struct track
);
643 return check_bytes_and_report(s
, page
, p
, "Object padding",
644 p
+ off
, POISON_INUSE
, s
->size
- off
);
647 /* Check the pad bytes at the end of a slab page */
648 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
656 if (!(s
->flags
& SLAB_POISON
))
659 start
= page_address(page
);
660 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
661 end
= start
+ length
;
662 remainder
= length
% s
->size
;
666 fault
= check_bytes(end
- remainder
, POISON_INUSE
, remainder
);
669 while (end
> fault
&& end
[-1] == POISON_INUSE
)
672 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
673 print_section("Padding", end
- remainder
, remainder
);
675 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
679 static int check_object(struct kmem_cache
*s
, struct page
*page
,
680 void *object
, u8 val
)
683 u8
*endobject
= object
+ s
->objsize
;
685 if (s
->flags
& SLAB_RED_ZONE
) {
686 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
687 endobject
, val
, s
->inuse
- s
->objsize
))
690 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
) {
691 check_bytes_and_report(s
, page
, p
, "Alignment padding",
692 endobject
, POISON_INUSE
, s
->inuse
- s
->objsize
);
696 if (s
->flags
& SLAB_POISON
) {
697 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
698 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
699 POISON_FREE
, s
->objsize
- 1) ||
700 !check_bytes_and_report(s
, page
, p
, "Poison",
701 p
+ s
->objsize
- 1, POISON_END
, 1)))
704 * check_pad_bytes cleans up on its own.
706 check_pad_bytes(s
, page
, p
);
709 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
711 * Object and freepointer overlap. Cannot check
712 * freepointer while object is allocated.
716 /* Check free pointer validity */
717 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
718 object_err(s
, page
, p
, "Freepointer corrupt");
720 * No choice but to zap it and thus lose the remainder
721 * of the free objects in this slab. May cause
722 * another error because the object count is now wrong.
724 set_freepointer(s
, p
, NULL
);
730 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
734 VM_BUG_ON(!irqs_disabled());
736 if (!PageSlab(page
)) {
737 slab_err(s
, page
, "Not a valid slab page");
741 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
742 if (page
->objects
> maxobj
) {
743 slab_err(s
, page
, "objects %u > max %u",
744 s
->name
, page
->objects
, maxobj
);
747 if (page
->inuse
> page
->objects
) {
748 slab_err(s
, page
, "inuse %u > max %u",
749 s
->name
, page
->inuse
, page
->objects
);
752 /* Slab_pad_check fixes things up after itself */
753 slab_pad_check(s
, page
);
758 * Determine if a certain object on a page is on the freelist. Must hold the
759 * slab lock to guarantee that the chains are in a consistent state.
761 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
764 void *fp
= page
->freelist
;
766 unsigned long max_objects
;
768 while (fp
&& nr
<= page
->objects
) {
771 if (!check_valid_pointer(s
, page
, fp
)) {
773 object_err(s
, page
, object
,
774 "Freechain corrupt");
775 set_freepointer(s
, object
, NULL
);
778 slab_err(s
, page
, "Freepointer corrupt");
779 page
->freelist
= NULL
;
780 page
->inuse
= page
->objects
;
781 slab_fix(s
, "Freelist cleared");
787 fp
= get_freepointer(s
, object
);
791 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
792 if (max_objects
> MAX_OBJS_PER_PAGE
)
793 max_objects
= MAX_OBJS_PER_PAGE
;
795 if (page
->objects
!= max_objects
) {
796 slab_err(s
, page
, "Wrong number of objects. Found %d but "
797 "should be %d", page
->objects
, max_objects
);
798 page
->objects
= max_objects
;
799 slab_fix(s
, "Number of objects adjusted.");
801 if (page
->inuse
!= page
->objects
- nr
) {
802 slab_err(s
, page
, "Wrong object count. Counter is %d but "
803 "counted were %d", page
->inuse
, page
->objects
- nr
);
804 page
->inuse
= page
->objects
- nr
;
805 slab_fix(s
, "Object count adjusted.");
807 return search
== NULL
;
810 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
813 if (s
->flags
& SLAB_TRACE
) {
814 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
816 alloc
? "alloc" : "free",
821 print_section("Object", (void *)object
, s
->objsize
);
828 * Hooks for other subsystems that check memory allocations. In a typical
829 * production configuration these hooks all should produce no code at all.
831 static inline int slab_pre_alloc_hook(struct kmem_cache
*s
, gfp_t flags
)
833 flags
&= gfp_allowed_mask
;
834 lockdep_trace_alloc(flags
);
835 might_sleep_if(flags
& __GFP_WAIT
);
837 return should_failslab(s
->objsize
, flags
, s
->flags
);
840 static inline void slab_post_alloc_hook(struct kmem_cache
*s
, gfp_t flags
, void *object
)
842 flags
&= gfp_allowed_mask
;
843 kmemcheck_slab_alloc(s
, flags
, object
, slab_ksize(s
));
844 kmemleak_alloc_recursive(object
, s
->objsize
, 1, s
->flags
, flags
);
847 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
849 kmemleak_free_recursive(x
, s
->flags
);
852 * Trouble is that we may no longer disable interupts in the fast path
853 * So in order to make the debug calls that expect irqs to be
854 * disabled we need to disable interrupts temporarily.
856 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
860 local_irq_save(flags
);
861 kmemcheck_slab_free(s
, x
, s
->objsize
);
862 debug_check_no_locks_freed(x
, s
->objsize
);
863 local_irq_restore(flags
);
866 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
867 debug_check_no_obj_freed(x
, s
->objsize
);
871 * Tracking of fully allocated slabs for debugging purposes.
873 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
875 spin_lock(&n
->list_lock
);
876 list_add(&page
->lru
, &n
->full
);
877 spin_unlock(&n
->list_lock
);
880 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
882 struct kmem_cache_node
*n
;
884 if (!(s
->flags
& SLAB_STORE_USER
))
887 n
= get_node(s
, page_to_nid(page
));
889 spin_lock(&n
->list_lock
);
890 list_del(&page
->lru
);
891 spin_unlock(&n
->list_lock
);
894 /* Tracking of the number of slabs for debugging purposes */
895 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
897 struct kmem_cache_node
*n
= get_node(s
, node
);
899 return atomic_long_read(&n
->nr_slabs
);
902 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
904 return atomic_long_read(&n
->nr_slabs
);
907 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
909 struct kmem_cache_node
*n
= get_node(s
, node
);
912 * May be called early in order to allocate a slab for the
913 * kmem_cache_node structure. Solve the chicken-egg
914 * dilemma by deferring the increment of the count during
915 * bootstrap (see early_kmem_cache_node_alloc).
918 atomic_long_inc(&n
->nr_slabs
);
919 atomic_long_add(objects
, &n
->total_objects
);
922 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
924 struct kmem_cache_node
*n
= get_node(s
, node
);
926 atomic_long_dec(&n
->nr_slabs
);
927 atomic_long_sub(objects
, &n
->total_objects
);
930 /* Object debug checks for alloc/free paths */
931 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
934 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
937 init_object(s
, object
, SLUB_RED_INACTIVE
);
938 init_tracking(s
, object
);
941 static noinline
int alloc_debug_processing(struct kmem_cache
*s
, struct page
*page
,
942 void *object
, unsigned long addr
)
944 if (!check_slab(s
, page
))
947 if (!on_freelist(s
, page
, object
)) {
948 object_err(s
, page
, object
, "Object already allocated");
952 if (!check_valid_pointer(s
, page
, object
)) {
953 object_err(s
, page
, object
, "Freelist Pointer check fails");
957 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
960 /* Success perform special debug activities for allocs */
961 if (s
->flags
& SLAB_STORE_USER
)
962 set_track(s
, object
, TRACK_ALLOC
, addr
);
963 trace(s
, page
, object
, 1);
964 init_object(s
, object
, SLUB_RED_ACTIVE
);
968 if (PageSlab(page
)) {
970 * If this is a slab page then lets do the best we can
971 * to avoid issues in the future. Marking all objects
972 * as used avoids touching the remaining objects.
974 slab_fix(s
, "Marking all objects used");
975 page
->inuse
= page
->objects
;
976 page
->freelist
= NULL
;
981 static noinline
int free_debug_processing(struct kmem_cache
*s
,
982 struct page
*page
, void *object
, unsigned long addr
)
984 if (!check_slab(s
, page
))
987 if (!check_valid_pointer(s
, page
, object
)) {
988 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
992 if (on_freelist(s
, page
, object
)) {
993 object_err(s
, page
, object
, "Object already free");
997 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1000 if (unlikely(s
!= page
->slab
)) {
1001 if (!PageSlab(page
)) {
1002 slab_err(s
, page
, "Attempt to free object(0x%p) "
1003 "outside of slab", object
);
1004 } else if (!page
->slab
) {
1006 "SLUB <none>: no slab for object 0x%p.\n",
1010 object_err(s
, page
, object
,
1011 "page slab pointer corrupt.");
1015 /* Special debug activities for freeing objects */
1016 if (!PageSlubFrozen(page
) && !page
->freelist
)
1017 remove_full(s
, page
);
1018 if (s
->flags
& SLAB_STORE_USER
)
1019 set_track(s
, object
, TRACK_FREE
, addr
);
1020 trace(s
, page
, object
, 0);
1021 init_object(s
, object
, SLUB_RED_INACTIVE
);
1025 slab_fix(s
, "Object at 0x%p not freed", object
);
1029 static int __init
setup_slub_debug(char *str
)
1031 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1032 if (*str
++ != '=' || !*str
)
1034 * No options specified. Switch on full debugging.
1040 * No options but restriction on slabs. This means full
1041 * debugging for slabs matching a pattern.
1045 if (tolower(*str
) == 'o') {
1047 * Avoid enabling debugging on caches if its minimum order
1048 * would increase as a result.
1050 disable_higher_order_debug
= 1;
1057 * Switch off all debugging measures.
1062 * Determine which debug features should be switched on
1064 for (; *str
&& *str
!= ','; str
++) {
1065 switch (tolower(*str
)) {
1067 slub_debug
|= SLAB_DEBUG_FREE
;
1070 slub_debug
|= SLAB_RED_ZONE
;
1073 slub_debug
|= SLAB_POISON
;
1076 slub_debug
|= SLAB_STORE_USER
;
1079 slub_debug
|= SLAB_TRACE
;
1082 slub_debug
|= SLAB_FAILSLAB
;
1085 printk(KERN_ERR
"slub_debug option '%c' "
1086 "unknown. skipped\n", *str
);
1092 slub_debug_slabs
= str
+ 1;
1097 __setup("slub_debug", setup_slub_debug
);
1099 static unsigned long kmem_cache_flags(unsigned long objsize
,
1100 unsigned long flags
, const char *name
,
1101 void (*ctor
)(void *))
1104 * Enable debugging if selected on the kernel commandline.
1106 if (slub_debug
&& (!slub_debug_slabs
||
1107 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
))))
1108 flags
|= slub_debug
;
1113 static inline void setup_object_debug(struct kmem_cache
*s
,
1114 struct page
*page
, void *object
) {}
1116 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1117 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1119 static inline int free_debug_processing(struct kmem_cache
*s
,
1120 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1122 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1124 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1125 void *object
, u8 val
) { return 1; }
1126 static inline void add_full(struct kmem_cache_node
*n
, struct page
*page
) {}
1127 static inline unsigned long kmem_cache_flags(unsigned long objsize
,
1128 unsigned long flags
, const char *name
,
1129 void (*ctor
)(void *))
1133 #define slub_debug 0
1135 #define disable_higher_order_debug 0
1137 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1139 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1141 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1143 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1146 static inline int slab_pre_alloc_hook(struct kmem_cache
*s
, gfp_t flags
)
1149 static inline void slab_post_alloc_hook(struct kmem_cache
*s
, gfp_t flags
,
1152 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
) {}
1154 #endif /* CONFIG_SLUB_DEBUG */
1157 * Slab allocation and freeing
1159 static inline struct page
*alloc_slab_page(gfp_t flags
, int node
,
1160 struct kmem_cache_order_objects oo
)
1162 int order
= oo_order(oo
);
1164 flags
|= __GFP_NOTRACK
;
1166 if (node
== NUMA_NO_NODE
)
1167 return alloc_pages(flags
, order
);
1169 return alloc_pages_exact_node(node
, flags
, order
);
1172 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1175 struct kmem_cache_order_objects oo
= s
->oo
;
1178 flags
|= s
->allocflags
;
1181 * Let the initial higher-order allocation fail under memory pressure
1182 * so we fall-back to the minimum order allocation.
1184 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1186 page
= alloc_slab_page(alloc_gfp
, node
, oo
);
1187 if (unlikely(!page
)) {
1190 * Allocation may have failed due to fragmentation.
1191 * Try a lower order alloc if possible
1193 page
= alloc_slab_page(flags
, node
, oo
);
1197 stat(s
, ORDER_FALLBACK
);
1200 if (kmemcheck_enabled
1201 && !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1202 int pages
= 1 << oo_order(oo
);
1204 kmemcheck_alloc_shadow(page
, oo_order(oo
), flags
, node
);
1207 * Objects from caches that have a constructor don't get
1208 * cleared when they're allocated, so we need to do it here.
1211 kmemcheck_mark_uninitialized_pages(page
, pages
);
1213 kmemcheck_mark_unallocated_pages(page
, pages
);
1216 page
->objects
= oo_objects(oo
);
1217 mod_zone_page_state(page_zone(page
),
1218 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1219 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1225 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1228 setup_object_debug(s
, page
, object
);
1229 if (unlikely(s
->ctor
))
1233 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1240 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
1242 page
= allocate_slab(s
,
1243 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1247 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1249 page
->flags
|= 1 << PG_slab
;
1251 start
= page_address(page
);
1253 if (unlikely(s
->flags
& SLAB_POISON
))
1254 memset(start
, POISON_INUSE
, PAGE_SIZE
<< compound_order(page
));
1257 for_each_object(p
, s
, start
, page
->objects
) {
1258 setup_object(s
, page
, last
);
1259 set_freepointer(s
, last
, p
);
1262 setup_object(s
, page
, last
);
1263 set_freepointer(s
, last
, NULL
);
1265 page
->freelist
= start
;
1271 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1273 int order
= compound_order(page
);
1274 int pages
= 1 << order
;
1276 if (kmem_cache_debug(s
)) {
1279 slab_pad_check(s
, page
);
1280 for_each_object(p
, s
, page_address(page
),
1282 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1285 kmemcheck_free_shadow(page
, compound_order(page
));
1287 mod_zone_page_state(page_zone(page
),
1288 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1289 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1292 __ClearPageSlab(page
);
1293 reset_page_mapcount(page
);
1294 if (current
->reclaim_state
)
1295 current
->reclaim_state
->reclaimed_slab
+= pages
;
1296 __free_pages(page
, order
);
1299 #define need_reserve_slab_rcu \
1300 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1302 static void rcu_free_slab(struct rcu_head
*h
)
1306 if (need_reserve_slab_rcu
)
1307 page
= virt_to_head_page(h
);
1309 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1311 __free_slab(page
->slab
, page
);
1314 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1316 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1317 struct rcu_head
*head
;
1319 if (need_reserve_slab_rcu
) {
1320 int order
= compound_order(page
);
1321 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1323 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1324 head
= page_address(page
) + offset
;
1327 * RCU free overloads the RCU head over the LRU
1329 head
= (void *)&page
->lru
;
1332 call_rcu(head
, rcu_free_slab
);
1334 __free_slab(s
, page
);
1337 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1339 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1344 * Per slab locking using the pagelock
1346 static __always_inline
void slab_lock(struct page
*page
)
1348 bit_spin_lock(PG_locked
, &page
->flags
);
1351 static __always_inline
void slab_unlock(struct page
*page
)
1353 __bit_spin_unlock(PG_locked
, &page
->flags
);
1356 static __always_inline
int slab_trylock(struct page
*page
)
1360 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
1365 * Management of partially allocated slabs
1367 static void add_partial(struct kmem_cache_node
*n
,
1368 struct page
*page
, int tail
)
1370 spin_lock(&n
->list_lock
);
1373 list_add_tail(&page
->lru
, &n
->partial
);
1375 list_add(&page
->lru
, &n
->partial
);
1376 spin_unlock(&n
->list_lock
);
1379 static inline void __remove_partial(struct kmem_cache_node
*n
,
1382 list_del(&page
->lru
);
1386 static void remove_partial(struct kmem_cache
*s
, struct page
*page
)
1388 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1390 spin_lock(&n
->list_lock
);
1391 __remove_partial(n
, page
);
1392 spin_unlock(&n
->list_lock
);
1396 * Lock slab and remove from the partial list.
1398 * Must hold list_lock.
1400 static inline int lock_and_freeze_slab(struct kmem_cache_node
*n
,
1403 if (slab_trylock(page
)) {
1404 __remove_partial(n
, page
);
1405 __SetPageSlubFrozen(page
);
1412 * Try to allocate a partial slab from a specific node.
1414 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
1419 * Racy check. If we mistakenly see no partial slabs then we
1420 * just allocate an empty slab. If we mistakenly try to get a
1421 * partial slab and there is none available then get_partials()
1424 if (!n
|| !n
->nr_partial
)
1427 spin_lock(&n
->list_lock
);
1428 list_for_each_entry(page
, &n
->partial
, lru
)
1429 if (lock_and_freeze_slab(n
, page
))
1433 spin_unlock(&n
->list_lock
);
1438 * Get a page from somewhere. Search in increasing NUMA distances.
1440 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1443 struct zonelist
*zonelist
;
1446 enum zone_type high_zoneidx
= gfp_zone(flags
);
1450 * The defrag ratio allows a configuration of the tradeoffs between
1451 * inter node defragmentation and node local allocations. A lower
1452 * defrag_ratio increases the tendency to do local allocations
1453 * instead of attempting to obtain partial slabs from other nodes.
1455 * If the defrag_ratio is set to 0 then kmalloc() always
1456 * returns node local objects. If the ratio is higher then kmalloc()
1457 * may return off node objects because partial slabs are obtained
1458 * from other nodes and filled up.
1460 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1461 * defrag_ratio = 1000) then every (well almost) allocation will
1462 * first attempt to defrag slab caches on other nodes. This means
1463 * scanning over all nodes to look for partial slabs which may be
1464 * expensive if we do it every time we are trying to find a slab
1465 * with available objects.
1467 if (!s
->remote_node_defrag_ratio
||
1468 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1472 zonelist
= node_zonelist(slab_node(current
->mempolicy
), flags
);
1473 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1474 struct kmem_cache_node
*n
;
1476 n
= get_node(s
, zone_to_nid(zone
));
1478 if (n
&& cpuset_zone_allowed_hardwall(zone
, flags
) &&
1479 n
->nr_partial
> s
->min_partial
) {
1480 page
= get_partial_node(n
);
1493 * Get a partial page, lock it and return it.
1495 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1498 int searchnode
= (node
== NUMA_NO_NODE
) ? numa_node_id() : node
;
1500 page
= get_partial_node(get_node(s
, searchnode
));
1501 if (page
|| node
!= NUMA_NO_NODE
)
1504 return get_any_partial(s
, flags
);
1508 * Move a page back to the lists.
1510 * Must be called with the slab lock held.
1512 * On exit the slab lock will have been dropped.
1514 static void unfreeze_slab(struct kmem_cache
*s
, struct page
*page
, int tail
)
1517 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1519 __ClearPageSlubFrozen(page
);
1522 if (page
->freelist
) {
1523 add_partial(n
, page
, tail
);
1524 stat(s
, tail
? DEACTIVATE_TO_TAIL
: DEACTIVATE_TO_HEAD
);
1526 stat(s
, DEACTIVATE_FULL
);
1527 if (kmem_cache_debug(s
) && (s
->flags
& SLAB_STORE_USER
))
1532 stat(s
, DEACTIVATE_EMPTY
);
1533 if (n
->nr_partial
< s
->min_partial
) {
1535 * Adding an empty slab to the partial slabs in order
1536 * to avoid page allocator overhead. This slab needs
1537 * to come after the other slabs with objects in
1538 * so that the others get filled first. That way the
1539 * size of the partial list stays small.
1541 * kmem_cache_shrink can reclaim any empty slabs from
1544 add_partial(n
, page
, 1);
1549 discard_slab(s
, page
);
1554 #ifdef CONFIG_CMPXCHG_LOCAL
1555 #ifdef CONFIG_PREEMPT
1557 * Calculate the next globally unique transaction for disambiguiation
1558 * during cmpxchg. The transactions start with the cpu number and are then
1559 * incremented by CONFIG_NR_CPUS.
1561 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1564 * No preemption supported therefore also no need to check for
1570 static inline unsigned long next_tid(unsigned long tid
)
1572 return tid
+ TID_STEP
;
1575 static inline unsigned int tid_to_cpu(unsigned long tid
)
1577 return tid
% TID_STEP
;
1580 static inline unsigned long tid_to_event(unsigned long tid
)
1582 return tid
/ TID_STEP
;
1585 static inline unsigned int init_tid(int cpu
)
1590 static inline void note_cmpxchg_failure(const char *n
,
1591 const struct kmem_cache
*s
, unsigned long tid
)
1593 #ifdef SLUB_DEBUG_CMPXCHG
1594 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1596 printk(KERN_INFO
"%s %s: cmpxchg redo ", n
, s
->name
);
1598 #ifdef CONFIG_PREEMPT
1599 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1600 printk("due to cpu change %d -> %d\n",
1601 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1604 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1605 printk("due to cpu running other code. Event %ld->%ld\n",
1606 tid_to_event(tid
), tid_to_event(actual_tid
));
1608 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1609 actual_tid
, tid
, next_tid(tid
));
1611 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1616 void init_kmem_cache_cpus(struct kmem_cache
*s
)
1618 #ifdef CONFIG_CMPXCHG_LOCAL
1621 for_each_possible_cpu(cpu
)
1622 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1627 * Remove the cpu slab
1629 static void deactivate_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1632 struct page
*page
= c
->page
;
1636 stat(s
, DEACTIVATE_REMOTE_FREES
);
1638 * Merge cpu freelist into slab freelist. Typically we get here
1639 * because both freelists are empty. So this is unlikely
1642 while (unlikely(c
->freelist
)) {
1645 tail
= 0; /* Hot objects. Put the slab first */
1647 /* Retrieve object from cpu_freelist */
1648 object
= c
->freelist
;
1649 c
->freelist
= get_freepointer(s
, c
->freelist
);
1651 /* And put onto the regular freelist */
1652 set_freepointer(s
, object
, page
->freelist
);
1653 page
->freelist
= object
;
1657 #ifdef CONFIG_CMPXCHG_LOCAL
1658 c
->tid
= next_tid(c
->tid
);
1660 unfreeze_slab(s
, page
, tail
);
1663 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
1665 stat(s
, CPUSLAB_FLUSH
);
1667 deactivate_slab(s
, c
);
1673 * Called from IPI handler with interrupts disabled.
1675 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1677 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
1679 if (likely(c
&& c
->page
))
1683 static void flush_cpu_slab(void *d
)
1685 struct kmem_cache
*s
= d
;
1687 __flush_cpu_slab(s
, smp_processor_id());
1690 static void flush_all(struct kmem_cache
*s
)
1692 on_each_cpu(flush_cpu_slab
, s
, 1);
1696 * Check if the objects in a per cpu structure fit numa
1697 * locality expectations.
1699 static inline int node_match(struct kmem_cache_cpu
*c
, int node
)
1702 if (node
!= NUMA_NO_NODE
&& c
->node
!= node
)
1708 static int count_free(struct page
*page
)
1710 return page
->objects
- page
->inuse
;
1713 static unsigned long count_partial(struct kmem_cache_node
*n
,
1714 int (*get_count
)(struct page
*))
1716 unsigned long flags
;
1717 unsigned long x
= 0;
1720 spin_lock_irqsave(&n
->list_lock
, flags
);
1721 list_for_each_entry(page
, &n
->partial
, lru
)
1722 x
+= get_count(page
);
1723 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1727 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
1729 #ifdef CONFIG_SLUB_DEBUG
1730 return atomic_long_read(&n
->total_objects
);
1736 static noinline
void
1737 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
1742 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1744 printk(KERN_WARNING
" cache: %s, object size: %d, buffer size: %d, "
1745 "default order: %d, min order: %d\n", s
->name
, s
->objsize
,
1746 s
->size
, oo_order(s
->oo
), oo_order(s
->min
));
1748 if (oo_order(s
->min
) > get_order(s
->objsize
))
1749 printk(KERN_WARNING
" %s debugging increased min order, use "
1750 "slub_debug=O to disable.\n", s
->name
);
1752 for_each_online_node(node
) {
1753 struct kmem_cache_node
*n
= get_node(s
, node
);
1754 unsigned long nr_slabs
;
1755 unsigned long nr_objs
;
1756 unsigned long nr_free
;
1761 nr_free
= count_partial(n
, count_free
);
1762 nr_slabs
= node_nr_slabs(n
);
1763 nr_objs
= node_nr_objs(n
);
1766 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1767 node
, nr_slabs
, nr_objs
, nr_free
);
1772 * Slow path. The lockless freelist is empty or we need to perform
1775 * Interrupts are disabled.
1777 * Processing is still very fast if new objects have been freed to the
1778 * regular freelist. In that case we simply take over the regular freelist
1779 * as the lockless freelist and zap the regular freelist.
1781 * If that is not working then we fall back to the partial lists. We take the
1782 * first element of the freelist as the object to allocate now and move the
1783 * rest of the freelist to the lockless freelist.
1785 * And if we were unable to get a new slab from the partial slab lists then
1786 * we need to allocate a new slab. This is the slowest path since it involves
1787 * a call to the page allocator and the setup of a new slab.
1789 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
1790 unsigned long addr
, struct kmem_cache_cpu
*c
)
1794 #ifdef CONFIG_CMPXCHG_LOCAL
1795 unsigned long flags
;
1797 local_irq_save(flags
);
1798 #ifdef CONFIG_PREEMPT
1800 * We may have been preempted and rescheduled on a different
1801 * cpu before disabling interrupts. Need to reload cpu area
1804 c
= this_cpu_ptr(s
->cpu_slab
);
1808 /* We handle __GFP_ZERO in the caller */
1809 gfpflags
&= ~__GFP_ZERO
;
1816 if (unlikely(!node_match(c
, node
)))
1819 stat(s
, ALLOC_REFILL
);
1822 object
= page
->freelist
;
1823 if (unlikely(!object
))
1825 if (kmem_cache_debug(s
))
1828 c
->freelist
= get_freepointer(s
, object
);
1829 page
->inuse
= page
->objects
;
1830 page
->freelist
= NULL
;
1834 #ifdef CONFIG_CMPXCHG_LOCAL
1835 c
->tid
= next_tid(c
->tid
);
1836 local_irq_restore(flags
);
1838 stat(s
, ALLOC_SLOWPATH
);
1842 deactivate_slab(s
, c
);
1845 page
= get_partial(s
, gfpflags
, node
);
1847 stat(s
, ALLOC_FROM_PARTIAL
);
1849 c
->node
= page_to_nid(page
);
1854 gfpflags
&= gfp_allowed_mask
;
1855 if (gfpflags
& __GFP_WAIT
)
1858 page
= new_slab(s
, gfpflags
, node
);
1860 if (gfpflags
& __GFP_WAIT
)
1861 local_irq_disable();
1864 c
= __this_cpu_ptr(s
->cpu_slab
);
1865 stat(s
, ALLOC_SLAB
);
1870 __SetPageSlubFrozen(page
);
1872 goto load_from_page
;
1874 if (!(gfpflags
& __GFP_NOWARN
) && printk_ratelimit())
1875 slab_out_of_memory(s
, gfpflags
, node
);
1876 #ifdef CONFIG_CMPXCHG_LOCAL
1877 local_irq_restore(flags
);
1881 if (!alloc_debug_processing(s
, page
, object
, addr
))
1885 page
->freelist
= get_freepointer(s
, object
);
1886 c
->node
= NUMA_NO_NODE
;
1891 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1892 * have the fastpath folded into their functions. So no function call
1893 * overhead for requests that can be satisfied on the fastpath.
1895 * The fastpath works by first checking if the lockless freelist can be used.
1896 * If not then __slab_alloc is called for slow processing.
1898 * Otherwise we can simply pick the next object from the lockless free list.
1900 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
1901 gfp_t gfpflags
, int node
, unsigned long addr
)
1904 struct kmem_cache_cpu
*c
;
1905 #ifdef CONFIG_CMPXCHG_LOCAL
1908 unsigned long flags
;
1911 if (slab_pre_alloc_hook(s
, gfpflags
))
1914 #ifndef CONFIG_CMPXCHG_LOCAL
1915 local_irq_save(flags
);
1921 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1922 * enabled. We may switch back and forth between cpus while
1923 * reading from one cpu area. That does not matter as long
1924 * as we end up on the original cpu again when doing the cmpxchg.
1926 c
= __this_cpu_ptr(s
->cpu_slab
);
1928 #ifdef CONFIG_CMPXCHG_LOCAL
1930 * The transaction ids are globally unique per cpu and per operation on
1931 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1932 * occurs on the right processor and that there was no operation on the
1933 * linked list in between.
1939 object
= c
->freelist
;
1940 if (unlikely(!object
|| !node_match(c
, node
)))
1942 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
1945 #ifdef CONFIG_CMPXCHG_LOCAL
1947 * The cmpxchg will only match if there was no additonal
1948 * operation and if we are on the right processor.
1950 * The cmpxchg does the following atomically (without lock semantics!)
1951 * 1. Relocate first pointer to the current per cpu area.
1952 * 2. Verify that tid and freelist have not been changed
1953 * 3. If they were not changed replace tid and freelist
1955 * Since this is without lock semantics the protection is only against
1956 * code executing on this cpu *not* from access by other cpus.
1958 if (unlikely(!this_cpu_cmpxchg_double(
1959 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
1961 get_freepointer(s
, object
), next_tid(tid
)))) {
1963 note_cmpxchg_failure("slab_alloc", s
, tid
);
1967 c
->freelist
= get_freepointer(s
, object
);
1969 stat(s
, ALLOC_FASTPATH
);
1972 #ifndef CONFIG_CMPXCHG_LOCAL
1973 local_irq_restore(flags
);
1976 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
1977 memset(object
, 0, s
->objsize
);
1979 slab_post_alloc_hook(s
, gfpflags
, object
);
1984 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1986 void *ret
= slab_alloc(s
, gfpflags
, NUMA_NO_NODE
, _RET_IP_
);
1988 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->objsize
, s
->size
, gfpflags
);
1992 EXPORT_SYMBOL(kmem_cache_alloc
);
1994 #ifdef CONFIG_TRACING
1995 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
1997 void *ret
= slab_alloc(s
, gfpflags
, NUMA_NO_NODE
, _RET_IP_
);
1998 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2001 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2003 void *kmalloc_order_trace(size_t size
, gfp_t flags
, unsigned int order
)
2005 void *ret
= kmalloc_order(size
, flags
, order
);
2006 trace_kmalloc(_RET_IP_
, ret
, size
, PAGE_SIZE
<< order
, flags
);
2009 EXPORT_SYMBOL(kmalloc_order_trace
);
2013 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2015 void *ret
= slab_alloc(s
, gfpflags
, node
, _RET_IP_
);
2017 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2018 s
->objsize
, s
->size
, gfpflags
, node
);
2022 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2024 #ifdef CONFIG_TRACING
2025 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2027 int node
, size_t size
)
2029 void *ret
= slab_alloc(s
, gfpflags
, node
, _RET_IP_
);
2031 trace_kmalloc_node(_RET_IP_
, ret
,
2032 size
, s
->size
, gfpflags
, node
);
2035 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2040 * Slow patch handling. This may still be called frequently since objects
2041 * have a longer lifetime than the cpu slabs in most processing loads.
2043 * So we still attempt to reduce cache line usage. Just take the slab
2044 * lock and free the item. If there is no additional partial page
2045 * handling required then we can return immediately.
2047 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2048 void *x
, unsigned long addr
)
2051 void **object
= (void *)x
;
2052 #ifdef CONFIG_CMPXCHG_LOCAL
2053 unsigned long flags
;
2055 local_irq_save(flags
);
2058 stat(s
, FREE_SLOWPATH
);
2060 if (kmem_cache_debug(s
))
2064 prior
= page
->freelist
;
2065 set_freepointer(s
, object
, prior
);
2066 page
->freelist
= object
;
2069 if (unlikely(PageSlubFrozen(page
))) {
2070 stat(s
, FREE_FROZEN
);
2074 if (unlikely(!page
->inuse
))
2078 * Objects left in the slab. If it was not on the partial list before
2081 if (unlikely(!prior
)) {
2082 add_partial(get_node(s
, page_to_nid(page
)), page
, 1);
2083 stat(s
, FREE_ADD_PARTIAL
);
2088 #ifdef CONFIG_CMPXCHG_LOCAL
2089 local_irq_restore(flags
);
2096 * Slab still on the partial list.
2098 remove_partial(s
, page
);
2099 stat(s
, FREE_REMOVE_PARTIAL
);
2102 #ifdef CONFIG_CMPXCHG_LOCAL
2103 local_irq_restore(flags
);
2106 discard_slab(s
, page
);
2110 if (!free_debug_processing(s
, page
, x
, addr
))
2116 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2117 * can perform fastpath freeing without additional function calls.
2119 * The fastpath is only possible if we are freeing to the current cpu slab
2120 * of this processor. This typically the case if we have just allocated
2123 * If fastpath is not possible then fall back to __slab_free where we deal
2124 * with all sorts of special processing.
2126 static __always_inline
void slab_free(struct kmem_cache
*s
,
2127 struct page
*page
, void *x
, unsigned long addr
)
2129 void **object
= (void *)x
;
2130 struct kmem_cache_cpu
*c
;
2131 #ifdef CONFIG_CMPXCHG_LOCAL
2134 unsigned long flags
;
2137 slab_free_hook(s
, x
);
2139 #ifndef CONFIG_CMPXCHG_LOCAL
2140 local_irq_save(flags
);
2147 * Determine the currently cpus per cpu slab.
2148 * The cpu may change afterward. However that does not matter since
2149 * data is retrieved via this pointer. If we are on the same cpu
2150 * during the cmpxchg then the free will succedd.
2152 c
= __this_cpu_ptr(s
->cpu_slab
);
2154 #ifdef CONFIG_CMPXCHG_LOCAL
2159 if (likely(page
== c
->page
&& c
->node
!= NUMA_NO_NODE
)) {
2160 set_freepointer(s
, object
, c
->freelist
);
2162 #ifdef CONFIG_CMPXCHG_LOCAL
2163 if (unlikely(!this_cpu_cmpxchg_double(
2164 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2166 object
, next_tid(tid
)))) {
2168 note_cmpxchg_failure("slab_free", s
, tid
);
2172 c
->freelist
= object
;
2174 stat(s
, FREE_FASTPATH
);
2176 __slab_free(s
, page
, x
, addr
);
2178 #ifndef CONFIG_CMPXCHG_LOCAL
2179 local_irq_restore(flags
);
2183 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2187 page
= virt_to_head_page(x
);
2189 slab_free(s
, page
, x
, _RET_IP_
);
2191 trace_kmem_cache_free(_RET_IP_
, x
);
2193 EXPORT_SYMBOL(kmem_cache_free
);
2196 * Object placement in a slab is made very easy because we always start at
2197 * offset 0. If we tune the size of the object to the alignment then we can
2198 * get the required alignment by putting one properly sized object after
2201 * Notice that the allocation order determines the sizes of the per cpu
2202 * caches. Each processor has always one slab available for allocations.
2203 * Increasing the allocation order reduces the number of times that slabs
2204 * must be moved on and off the partial lists and is therefore a factor in
2209 * Mininum / Maximum order of slab pages. This influences locking overhead
2210 * and slab fragmentation. A higher order reduces the number of partial slabs
2211 * and increases the number of allocations possible without having to
2212 * take the list_lock.
2214 static int slub_min_order
;
2215 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
2216 static int slub_min_objects
;
2219 * Merge control. If this is set then no merging of slab caches will occur.
2220 * (Could be removed. This was introduced to pacify the merge skeptics.)
2222 static int slub_nomerge
;
2225 * Calculate the order of allocation given an slab object size.
2227 * The order of allocation has significant impact on performance and other
2228 * system components. Generally order 0 allocations should be preferred since
2229 * order 0 does not cause fragmentation in the page allocator. Larger objects
2230 * be problematic to put into order 0 slabs because there may be too much
2231 * unused space left. We go to a higher order if more than 1/16th of the slab
2234 * In order to reach satisfactory performance we must ensure that a minimum
2235 * number of objects is in one slab. Otherwise we may generate too much
2236 * activity on the partial lists which requires taking the list_lock. This is
2237 * less a concern for large slabs though which are rarely used.
2239 * slub_max_order specifies the order where we begin to stop considering the
2240 * number of objects in a slab as critical. If we reach slub_max_order then
2241 * we try to keep the page order as low as possible. So we accept more waste
2242 * of space in favor of a small page order.
2244 * Higher order allocations also allow the placement of more objects in a
2245 * slab and thereby reduce object handling overhead. If the user has
2246 * requested a higher mininum order then we start with that one instead of
2247 * the smallest order which will fit the object.
2249 static inline int slab_order(int size
, int min_objects
,
2250 int max_order
, int fract_leftover
, int reserved
)
2254 int min_order
= slub_min_order
;
2256 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
2257 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
2259 for (order
= max(min_order
,
2260 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
2261 order
<= max_order
; order
++) {
2263 unsigned long slab_size
= PAGE_SIZE
<< order
;
2265 if (slab_size
< min_objects
* size
+ reserved
)
2268 rem
= (slab_size
- reserved
) % size
;
2270 if (rem
<= slab_size
/ fract_leftover
)
2278 static inline int calculate_order(int size
, int reserved
)
2286 * Attempt to find best configuration for a slab. This
2287 * works by first attempting to generate a layout with
2288 * the best configuration and backing off gradually.
2290 * First we reduce the acceptable waste in a slab. Then
2291 * we reduce the minimum objects required in a slab.
2293 min_objects
= slub_min_objects
;
2295 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
2296 max_objects
= order_objects(slub_max_order
, size
, reserved
);
2297 min_objects
= min(min_objects
, max_objects
);
2299 while (min_objects
> 1) {
2301 while (fraction
>= 4) {
2302 order
= slab_order(size
, min_objects
,
2303 slub_max_order
, fraction
, reserved
);
2304 if (order
<= slub_max_order
)
2312 * We were unable to place multiple objects in a slab. Now
2313 * lets see if we can place a single object there.
2315 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
2316 if (order
<= slub_max_order
)
2320 * Doh this slab cannot be placed using slub_max_order.
2322 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
2323 if (order
< MAX_ORDER
)
2329 * Figure out what the alignment of the objects will be.
2331 static unsigned long calculate_alignment(unsigned long flags
,
2332 unsigned long align
, unsigned long size
)
2335 * If the user wants hardware cache aligned objects then follow that
2336 * suggestion if the object is sufficiently large.
2338 * The hardware cache alignment cannot override the specified
2339 * alignment though. If that is greater then use it.
2341 if (flags
& SLAB_HWCACHE_ALIGN
) {
2342 unsigned long ralign
= cache_line_size();
2343 while (size
<= ralign
/ 2)
2345 align
= max(align
, ralign
);
2348 if (align
< ARCH_SLAB_MINALIGN
)
2349 align
= ARCH_SLAB_MINALIGN
;
2351 return ALIGN(align
, sizeof(void *));
2355 init_kmem_cache_node(struct kmem_cache_node
*n
, struct kmem_cache
*s
)
2358 spin_lock_init(&n
->list_lock
);
2359 INIT_LIST_HEAD(&n
->partial
);
2360 #ifdef CONFIG_SLUB_DEBUG
2361 atomic_long_set(&n
->nr_slabs
, 0);
2362 atomic_long_set(&n
->total_objects
, 0);
2363 INIT_LIST_HEAD(&n
->full
);
2367 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
2369 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
2370 SLUB_PAGE_SHIFT
* sizeof(struct kmem_cache_cpu
));
2372 #ifdef CONFIG_CMPXCHG_LOCAL
2374 * Must align to double word boundary for the double cmpxchg instructions
2377 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
), 2 * sizeof(void *));
2379 /* Regular alignment is sufficient */
2380 s
->cpu_slab
= alloc_percpu(struct kmem_cache_cpu
);
2386 init_kmem_cache_cpus(s
);
2391 static struct kmem_cache
*kmem_cache_node
;
2394 * No kmalloc_node yet so do it by hand. We know that this is the first
2395 * slab on the node for this slabcache. There are no concurrent accesses
2398 * Note that this function only works on the kmalloc_node_cache
2399 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2400 * memory on a fresh node that has no slab structures yet.
2402 static void early_kmem_cache_node_alloc(int node
)
2405 struct kmem_cache_node
*n
;
2406 unsigned long flags
;
2408 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
2410 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
2413 if (page_to_nid(page
) != node
) {
2414 printk(KERN_ERR
"SLUB: Unable to allocate memory from "
2416 printk(KERN_ERR
"SLUB: Allocating a useless per node structure "
2417 "in order to be able to continue\n");
2422 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
2424 kmem_cache_node
->node
[node
] = n
;
2425 #ifdef CONFIG_SLUB_DEBUG
2426 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
2427 init_tracking(kmem_cache_node
, n
);
2429 init_kmem_cache_node(n
, kmem_cache_node
);
2430 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
2433 * lockdep requires consistent irq usage for each lock
2434 * so even though there cannot be a race this early in
2435 * the boot sequence, we still disable irqs.
2437 local_irq_save(flags
);
2438 add_partial(n
, page
, 0);
2439 local_irq_restore(flags
);
2442 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
2446 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2447 struct kmem_cache_node
*n
= s
->node
[node
];
2450 kmem_cache_free(kmem_cache_node
, n
);
2452 s
->node
[node
] = NULL
;
2456 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
2460 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2461 struct kmem_cache_node
*n
;
2463 if (slab_state
== DOWN
) {
2464 early_kmem_cache_node_alloc(node
);
2467 n
= kmem_cache_alloc_node(kmem_cache_node
,
2471 free_kmem_cache_nodes(s
);
2476 init_kmem_cache_node(n
, s
);
2481 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
2483 if (min
< MIN_PARTIAL
)
2485 else if (min
> MAX_PARTIAL
)
2487 s
->min_partial
= min
;
2491 * calculate_sizes() determines the order and the distribution of data within
2494 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
2496 unsigned long flags
= s
->flags
;
2497 unsigned long size
= s
->objsize
;
2498 unsigned long align
= s
->align
;
2502 * Round up object size to the next word boundary. We can only
2503 * place the free pointer at word boundaries and this determines
2504 * the possible location of the free pointer.
2506 size
= ALIGN(size
, sizeof(void *));
2508 #ifdef CONFIG_SLUB_DEBUG
2510 * Determine if we can poison the object itself. If the user of
2511 * the slab may touch the object after free or before allocation
2512 * then we should never poison the object itself.
2514 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
2516 s
->flags
|= __OBJECT_POISON
;
2518 s
->flags
&= ~__OBJECT_POISON
;
2522 * If we are Redzoning then check if there is some space between the
2523 * end of the object and the free pointer. If not then add an
2524 * additional word to have some bytes to store Redzone information.
2526 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
2527 size
+= sizeof(void *);
2531 * With that we have determined the number of bytes in actual use
2532 * by the object. This is the potential offset to the free pointer.
2536 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
2539 * Relocate free pointer after the object if it is not
2540 * permitted to overwrite the first word of the object on
2543 * This is the case if we do RCU, have a constructor or
2544 * destructor or are poisoning the objects.
2547 size
+= sizeof(void *);
2550 #ifdef CONFIG_SLUB_DEBUG
2551 if (flags
& SLAB_STORE_USER
)
2553 * Need to store information about allocs and frees after
2556 size
+= 2 * sizeof(struct track
);
2558 if (flags
& SLAB_RED_ZONE
)
2560 * Add some empty padding so that we can catch
2561 * overwrites from earlier objects rather than let
2562 * tracking information or the free pointer be
2563 * corrupted if a user writes before the start
2566 size
+= sizeof(void *);
2570 * Determine the alignment based on various parameters that the
2571 * user specified and the dynamic determination of cache line size
2574 align
= calculate_alignment(flags
, align
, s
->objsize
);
2578 * SLUB stores one object immediately after another beginning from
2579 * offset 0. In order to align the objects we have to simply size
2580 * each object to conform to the alignment.
2582 size
= ALIGN(size
, align
);
2584 if (forced_order
>= 0)
2585 order
= forced_order
;
2587 order
= calculate_order(size
, s
->reserved
);
2594 s
->allocflags
|= __GFP_COMP
;
2596 if (s
->flags
& SLAB_CACHE_DMA
)
2597 s
->allocflags
|= SLUB_DMA
;
2599 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
2600 s
->allocflags
|= __GFP_RECLAIMABLE
;
2603 * Determine the number of objects per slab
2605 s
->oo
= oo_make(order
, size
, s
->reserved
);
2606 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
2607 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
2610 return !!oo_objects(s
->oo
);
2614 static int kmem_cache_open(struct kmem_cache
*s
,
2615 const char *name
, size_t size
,
2616 size_t align
, unsigned long flags
,
2617 void (*ctor
)(void *))
2619 memset(s
, 0, kmem_size
);
2624 s
->flags
= kmem_cache_flags(size
, flags
, name
, ctor
);
2627 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
2628 s
->reserved
= sizeof(struct rcu_head
);
2630 if (!calculate_sizes(s
, -1))
2632 if (disable_higher_order_debug
) {
2634 * Disable debugging flags that store metadata if the min slab
2637 if (get_order(s
->size
) > get_order(s
->objsize
)) {
2638 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
2640 if (!calculate_sizes(s
, -1))
2646 * The larger the object size is, the more pages we want on the partial
2647 * list to avoid pounding the page allocator excessively.
2649 set_min_partial(s
, ilog2(s
->size
));
2652 s
->remote_node_defrag_ratio
= 1000;
2654 if (!init_kmem_cache_nodes(s
))
2657 if (alloc_kmem_cache_cpus(s
))
2660 free_kmem_cache_nodes(s
);
2662 if (flags
& SLAB_PANIC
)
2663 panic("Cannot create slab %s size=%lu realsize=%u "
2664 "order=%u offset=%u flags=%lx\n",
2665 s
->name
, (unsigned long)size
, s
->size
, oo_order(s
->oo
),
2671 * Determine the size of a slab object
2673 unsigned int kmem_cache_size(struct kmem_cache
*s
)
2677 EXPORT_SYMBOL(kmem_cache_size
);
2679 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
2682 #ifdef CONFIG_SLUB_DEBUG
2683 void *addr
= page_address(page
);
2685 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
2686 sizeof(long), GFP_ATOMIC
);
2689 slab_err(s
, page
, "%s", text
);
2692 get_map(s
, page
, map
);
2693 for_each_object(p
, s
, addr
, page
->objects
) {
2695 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
2696 printk(KERN_ERR
"INFO: Object 0x%p @offset=%tu\n",
2698 print_tracking(s
, p
);
2707 * Attempt to free all partial slabs on a node.
2709 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
2711 unsigned long flags
;
2712 struct page
*page
, *h
;
2714 spin_lock_irqsave(&n
->list_lock
, flags
);
2715 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
2717 __remove_partial(n
, page
);
2718 discard_slab(s
, page
);
2720 list_slab_objects(s
, page
,
2721 "Objects remaining on kmem_cache_close()");
2724 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2728 * Release all resources used by a slab cache.
2730 static inline int kmem_cache_close(struct kmem_cache
*s
)
2735 free_percpu(s
->cpu_slab
);
2736 /* Attempt to free all objects */
2737 for_each_node_state(node
, N_NORMAL_MEMORY
) {
2738 struct kmem_cache_node
*n
= get_node(s
, node
);
2741 if (n
->nr_partial
|| slabs_node(s
, node
))
2744 free_kmem_cache_nodes(s
);
2749 * Close a cache and release the kmem_cache structure
2750 * (must be used for caches created using kmem_cache_create)
2752 void kmem_cache_destroy(struct kmem_cache
*s
)
2754 down_write(&slub_lock
);
2758 if (kmem_cache_close(s
)) {
2759 printk(KERN_ERR
"SLUB %s: %s called for cache that "
2760 "still has objects.\n", s
->name
, __func__
);
2763 if (s
->flags
& SLAB_DESTROY_BY_RCU
)
2765 sysfs_slab_remove(s
);
2767 up_write(&slub_lock
);
2769 EXPORT_SYMBOL(kmem_cache_destroy
);
2771 /********************************************************************
2773 *******************************************************************/
2775 struct kmem_cache
*kmalloc_caches
[SLUB_PAGE_SHIFT
];
2776 EXPORT_SYMBOL(kmalloc_caches
);
2778 static struct kmem_cache
*kmem_cache
;
2780 #ifdef CONFIG_ZONE_DMA
2781 static struct kmem_cache
*kmalloc_dma_caches
[SLUB_PAGE_SHIFT
];
2784 static int __init
setup_slub_min_order(char *str
)
2786 get_option(&str
, &slub_min_order
);
2791 __setup("slub_min_order=", setup_slub_min_order
);
2793 static int __init
setup_slub_max_order(char *str
)
2795 get_option(&str
, &slub_max_order
);
2796 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
2801 __setup("slub_max_order=", setup_slub_max_order
);
2803 static int __init
setup_slub_min_objects(char *str
)
2805 get_option(&str
, &slub_min_objects
);
2810 __setup("slub_min_objects=", setup_slub_min_objects
);
2812 static int __init
setup_slub_nomerge(char *str
)
2818 __setup("slub_nomerge", setup_slub_nomerge
);
2820 static struct kmem_cache
*__init
create_kmalloc_cache(const char *name
,
2821 int size
, unsigned int flags
)
2823 struct kmem_cache
*s
;
2825 s
= kmem_cache_alloc(kmem_cache
, GFP_NOWAIT
);
2828 * This function is called with IRQs disabled during early-boot on
2829 * single CPU so there's no need to take slub_lock here.
2831 if (!kmem_cache_open(s
, name
, size
, ARCH_KMALLOC_MINALIGN
,
2835 list_add(&s
->list
, &slab_caches
);
2839 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2844 * Conversion table for small slabs sizes / 8 to the index in the
2845 * kmalloc array. This is necessary for slabs < 192 since we have non power
2846 * of two cache sizes there. The size of larger slabs can be determined using
2849 static s8 size_index
[24] = {
2876 static inline int size_index_elem(size_t bytes
)
2878 return (bytes
- 1) / 8;
2881 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2887 return ZERO_SIZE_PTR
;
2889 index
= size_index
[size_index_elem(size
)];
2891 index
= fls(size
- 1);
2893 #ifdef CONFIG_ZONE_DMA
2894 if (unlikely((flags
& SLUB_DMA
)))
2895 return kmalloc_dma_caches
[index
];
2898 return kmalloc_caches
[index
];
2901 void *__kmalloc(size_t size
, gfp_t flags
)
2903 struct kmem_cache
*s
;
2906 if (unlikely(size
> SLUB_MAX_SIZE
))
2907 return kmalloc_large(size
, flags
);
2909 s
= get_slab(size
, flags
);
2911 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2914 ret
= slab_alloc(s
, flags
, NUMA_NO_NODE
, _RET_IP_
);
2916 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
2920 EXPORT_SYMBOL(__kmalloc
);
2923 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
2928 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
2929 page
= alloc_pages_node(node
, flags
, get_order(size
));
2931 ptr
= page_address(page
);
2933 kmemleak_alloc(ptr
, size
, 1, flags
);
2937 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2939 struct kmem_cache
*s
;
2942 if (unlikely(size
> SLUB_MAX_SIZE
)) {
2943 ret
= kmalloc_large_node(size
, flags
, node
);
2945 trace_kmalloc_node(_RET_IP_
, ret
,
2946 size
, PAGE_SIZE
<< get_order(size
),
2952 s
= get_slab(size
, flags
);
2954 if (unlikely(ZERO_OR_NULL_PTR(s
)))
2957 ret
= slab_alloc(s
, flags
, node
, _RET_IP_
);
2959 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
2963 EXPORT_SYMBOL(__kmalloc_node
);
2966 size_t ksize(const void *object
)
2970 if (unlikely(object
== ZERO_SIZE_PTR
))
2973 page
= virt_to_head_page(object
);
2975 if (unlikely(!PageSlab(page
))) {
2976 WARN_ON(!PageCompound(page
));
2977 return PAGE_SIZE
<< compound_order(page
);
2980 return slab_ksize(page
->slab
);
2982 EXPORT_SYMBOL(ksize
);
2984 void kfree(const void *x
)
2987 void *object
= (void *)x
;
2989 trace_kfree(_RET_IP_
, x
);
2991 if (unlikely(ZERO_OR_NULL_PTR(x
)))
2994 page
= virt_to_head_page(x
);
2995 if (unlikely(!PageSlab(page
))) {
2996 BUG_ON(!PageCompound(page
));
3001 slab_free(page
->slab
, page
, object
, _RET_IP_
);
3003 EXPORT_SYMBOL(kfree
);
3006 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3007 * the remaining slabs by the number of items in use. The slabs with the
3008 * most items in use come first. New allocations will then fill those up
3009 * and thus they can be removed from the partial lists.
3011 * The slabs with the least items are placed last. This results in them
3012 * being allocated from last increasing the chance that the last objects
3013 * are freed in them.
3015 int kmem_cache_shrink(struct kmem_cache
*s
)
3019 struct kmem_cache_node
*n
;
3022 int objects
= oo_objects(s
->max
);
3023 struct list_head
*slabs_by_inuse
=
3024 kmalloc(sizeof(struct list_head
) * objects
, GFP_KERNEL
);
3025 unsigned long flags
;
3027 if (!slabs_by_inuse
)
3031 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3032 n
= get_node(s
, node
);
3037 for (i
= 0; i
< objects
; i
++)
3038 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
3040 spin_lock_irqsave(&n
->list_lock
, flags
);
3043 * Build lists indexed by the items in use in each slab.
3045 * Note that concurrent frees may occur while we hold the
3046 * list_lock. page->inuse here is the upper limit.
3048 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3049 if (!page
->inuse
&& slab_trylock(page
)) {
3051 * Must hold slab lock here because slab_free
3052 * may have freed the last object and be
3053 * waiting to release the slab.
3055 __remove_partial(n
, page
);
3057 discard_slab(s
, page
);
3059 list_move(&page
->lru
,
3060 slabs_by_inuse
+ page
->inuse
);
3065 * Rebuild the partial list with the slabs filled up most
3066 * first and the least used slabs at the end.
3068 for (i
= objects
- 1; i
>= 0; i
--)
3069 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
3071 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3074 kfree(slabs_by_inuse
);
3077 EXPORT_SYMBOL(kmem_cache_shrink
);
3079 #if defined(CONFIG_MEMORY_HOTPLUG)
3080 static int slab_mem_going_offline_callback(void *arg
)
3082 struct kmem_cache
*s
;
3084 down_read(&slub_lock
);
3085 list_for_each_entry(s
, &slab_caches
, list
)
3086 kmem_cache_shrink(s
);
3087 up_read(&slub_lock
);
3092 static void slab_mem_offline_callback(void *arg
)
3094 struct kmem_cache_node
*n
;
3095 struct kmem_cache
*s
;
3096 struct memory_notify
*marg
= arg
;
3099 offline_node
= marg
->status_change_nid
;
3102 * If the node still has available memory. we need kmem_cache_node
3105 if (offline_node
< 0)
3108 down_read(&slub_lock
);
3109 list_for_each_entry(s
, &slab_caches
, list
) {
3110 n
= get_node(s
, offline_node
);
3113 * if n->nr_slabs > 0, slabs still exist on the node
3114 * that is going down. We were unable to free them,
3115 * and offline_pages() function shouldn't call this
3116 * callback. So, we must fail.
3118 BUG_ON(slabs_node(s
, offline_node
));
3120 s
->node
[offline_node
] = NULL
;
3121 kmem_cache_free(kmem_cache_node
, n
);
3124 up_read(&slub_lock
);
3127 static int slab_mem_going_online_callback(void *arg
)
3129 struct kmem_cache_node
*n
;
3130 struct kmem_cache
*s
;
3131 struct memory_notify
*marg
= arg
;
3132 int nid
= marg
->status_change_nid
;
3136 * If the node's memory is already available, then kmem_cache_node is
3137 * already created. Nothing to do.
3143 * We are bringing a node online. No memory is available yet. We must
3144 * allocate a kmem_cache_node structure in order to bring the node
3147 down_read(&slub_lock
);
3148 list_for_each_entry(s
, &slab_caches
, list
) {
3150 * XXX: kmem_cache_alloc_node will fallback to other nodes
3151 * since memory is not yet available from the node that
3154 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3159 init_kmem_cache_node(n
, s
);
3163 up_read(&slub_lock
);
3167 static int slab_memory_callback(struct notifier_block
*self
,
3168 unsigned long action
, void *arg
)
3173 case MEM_GOING_ONLINE
:
3174 ret
= slab_mem_going_online_callback(arg
);
3176 case MEM_GOING_OFFLINE
:
3177 ret
= slab_mem_going_offline_callback(arg
);
3180 case MEM_CANCEL_ONLINE
:
3181 slab_mem_offline_callback(arg
);
3184 case MEM_CANCEL_OFFLINE
:
3188 ret
= notifier_from_errno(ret
);
3194 #endif /* CONFIG_MEMORY_HOTPLUG */
3196 /********************************************************************
3197 * Basic setup of slabs
3198 *******************************************************************/
3201 * Used for early kmem_cache structures that were allocated using
3202 * the page allocator
3205 static void __init
kmem_cache_bootstrap_fixup(struct kmem_cache
*s
)
3209 list_add(&s
->list
, &slab_caches
);
3212 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3213 struct kmem_cache_node
*n
= get_node(s
, node
);
3217 list_for_each_entry(p
, &n
->partial
, lru
)
3220 #ifdef CONFIG_SLUB_DEBUG
3221 list_for_each_entry(p
, &n
->full
, lru
)
3228 void __init
kmem_cache_init(void)
3232 struct kmem_cache
*temp_kmem_cache
;
3234 struct kmem_cache
*temp_kmem_cache_node
;
3235 unsigned long kmalloc_size
;
3237 kmem_size
= offsetof(struct kmem_cache
, node
) +
3238 nr_node_ids
* sizeof(struct kmem_cache_node
*);
3240 /* Allocate two kmem_caches from the page allocator */
3241 kmalloc_size
= ALIGN(kmem_size
, cache_line_size());
3242 order
= get_order(2 * kmalloc_size
);
3243 kmem_cache
= (void *)__get_free_pages(GFP_NOWAIT
, order
);
3246 * Must first have the slab cache available for the allocations of the
3247 * struct kmem_cache_node's. There is special bootstrap code in
3248 * kmem_cache_open for slab_state == DOWN.
3250 kmem_cache_node
= (void *)kmem_cache
+ kmalloc_size
;
3252 kmem_cache_open(kmem_cache_node
, "kmem_cache_node",
3253 sizeof(struct kmem_cache_node
),
3254 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
3256 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
3258 /* Able to allocate the per node structures */
3259 slab_state
= PARTIAL
;
3261 temp_kmem_cache
= kmem_cache
;
3262 kmem_cache_open(kmem_cache
, "kmem_cache", kmem_size
,
3263 0, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
, NULL
);
3264 kmem_cache
= kmem_cache_alloc(kmem_cache
, GFP_NOWAIT
);
3265 memcpy(kmem_cache
, temp_kmem_cache
, kmem_size
);
3268 * Allocate kmem_cache_node properly from the kmem_cache slab.
3269 * kmem_cache_node is separately allocated so no need to
3270 * update any list pointers.
3272 temp_kmem_cache_node
= kmem_cache_node
;
3274 kmem_cache_node
= kmem_cache_alloc(kmem_cache
, GFP_NOWAIT
);
3275 memcpy(kmem_cache_node
, temp_kmem_cache_node
, kmem_size
);
3277 kmem_cache_bootstrap_fixup(kmem_cache_node
);
3280 kmem_cache_bootstrap_fixup(kmem_cache
);
3282 /* Free temporary boot structure */
3283 free_pages((unsigned long)temp_kmem_cache
, order
);
3285 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3288 * Patch up the size_index table if we have strange large alignment
3289 * requirements for the kmalloc array. This is only the case for
3290 * MIPS it seems. The standard arches will not generate any code here.
3292 * Largest permitted alignment is 256 bytes due to the way we
3293 * handle the index determination for the smaller caches.
3295 * Make sure that nothing crazy happens if someone starts tinkering
3296 * around with ARCH_KMALLOC_MINALIGN
3298 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 256 ||
3299 (KMALLOC_MIN_SIZE
& (KMALLOC_MIN_SIZE
- 1)));
3301 for (i
= 8; i
< KMALLOC_MIN_SIZE
; i
+= 8) {
3302 int elem
= size_index_elem(i
);
3303 if (elem
>= ARRAY_SIZE(size_index
))
3305 size_index
[elem
] = KMALLOC_SHIFT_LOW
;
3308 if (KMALLOC_MIN_SIZE
== 64) {
3310 * The 96 byte size cache is not used if the alignment
3313 for (i
= 64 + 8; i
<= 96; i
+= 8)
3314 size_index
[size_index_elem(i
)] = 7;
3315 } else if (KMALLOC_MIN_SIZE
== 128) {
3317 * The 192 byte sized cache is not used if the alignment
3318 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3321 for (i
= 128 + 8; i
<= 192; i
+= 8)
3322 size_index
[size_index_elem(i
)] = 8;
3325 /* Caches that are not of the two-to-the-power-of size */
3326 if (KMALLOC_MIN_SIZE
<= 32) {
3327 kmalloc_caches
[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3331 if (KMALLOC_MIN_SIZE
<= 64) {
3332 kmalloc_caches
[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3336 for (i
= KMALLOC_SHIFT_LOW
; i
< SLUB_PAGE_SHIFT
; i
++) {
3337 kmalloc_caches
[i
] = create_kmalloc_cache("kmalloc", 1 << i
, 0);
3343 /* Provide the correct kmalloc names now that the caches are up */
3344 if (KMALLOC_MIN_SIZE
<= 32) {
3345 kmalloc_caches
[1]->name
= kstrdup(kmalloc_caches
[1]->name
, GFP_NOWAIT
);
3346 BUG_ON(!kmalloc_caches
[1]->name
);
3349 if (KMALLOC_MIN_SIZE
<= 64) {
3350 kmalloc_caches
[2]->name
= kstrdup(kmalloc_caches
[2]->name
, GFP_NOWAIT
);
3351 BUG_ON(!kmalloc_caches
[2]->name
);
3354 for (i
= KMALLOC_SHIFT_LOW
; i
< SLUB_PAGE_SHIFT
; i
++) {
3355 char *s
= kasprintf(GFP_NOWAIT
, "kmalloc-%d", 1 << i
);
3358 kmalloc_caches
[i
]->name
= s
;
3362 register_cpu_notifier(&slab_notifier
);
3365 #ifdef CONFIG_ZONE_DMA
3366 for (i
= 0; i
< SLUB_PAGE_SHIFT
; i
++) {
3367 struct kmem_cache
*s
= kmalloc_caches
[i
];
3370 char *name
= kasprintf(GFP_NOWAIT
,
3371 "dma-kmalloc-%d", s
->objsize
);
3374 kmalloc_dma_caches
[i
] = create_kmalloc_cache(name
,
3375 s
->objsize
, SLAB_CACHE_DMA
);
3380 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3381 " CPUs=%d, Nodes=%d\n",
3382 caches
, cache_line_size(),
3383 slub_min_order
, slub_max_order
, slub_min_objects
,
3384 nr_cpu_ids
, nr_node_ids
);
3387 void __init
kmem_cache_init_late(void)
3392 * Find a mergeable slab cache
3394 static int slab_unmergeable(struct kmem_cache
*s
)
3396 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
3403 * We may have set a slab to be unmergeable during bootstrap.
3405 if (s
->refcount
< 0)
3411 static struct kmem_cache
*find_mergeable(size_t size
,
3412 size_t align
, unsigned long flags
, const char *name
,
3413 void (*ctor
)(void *))
3415 struct kmem_cache
*s
;
3417 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
3423 size
= ALIGN(size
, sizeof(void *));
3424 align
= calculate_alignment(flags
, align
, size
);
3425 size
= ALIGN(size
, align
);
3426 flags
= kmem_cache_flags(size
, flags
, name
, NULL
);
3428 list_for_each_entry(s
, &slab_caches
, list
) {
3429 if (slab_unmergeable(s
))
3435 if ((flags
& SLUB_MERGE_SAME
) != (s
->flags
& SLUB_MERGE_SAME
))
3438 * Check if alignment is compatible.
3439 * Courtesy of Adrian Drzewiecki
3441 if ((s
->size
& ~(align
- 1)) != s
->size
)
3444 if (s
->size
- size
>= sizeof(void *))
3452 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
3453 size_t align
, unsigned long flags
, void (*ctor
)(void *))
3455 struct kmem_cache
*s
;
3461 down_write(&slub_lock
);
3462 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
3466 * Adjust the object sizes so that we clear
3467 * the complete object on kzalloc.
3469 s
->objsize
= max(s
->objsize
, (int)size
);
3470 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3472 if (sysfs_slab_alias(s
, name
)) {
3476 up_write(&slub_lock
);
3480 n
= kstrdup(name
, GFP_KERNEL
);
3484 s
= kmalloc(kmem_size
, GFP_KERNEL
);
3486 if (kmem_cache_open(s
, n
,
3487 size
, align
, flags
, ctor
)) {
3488 list_add(&s
->list
, &slab_caches
);
3489 if (sysfs_slab_add(s
)) {
3495 up_write(&slub_lock
);
3502 up_write(&slub_lock
);
3504 if (flags
& SLAB_PANIC
)
3505 panic("Cannot create slabcache %s\n", name
);
3510 EXPORT_SYMBOL(kmem_cache_create
);
3514 * Use the cpu notifier to insure that the cpu slabs are flushed when
3517 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
3518 unsigned long action
, void *hcpu
)
3520 long cpu
= (long)hcpu
;
3521 struct kmem_cache
*s
;
3522 unsigned long flags
;
3525 case CPU_UP_CANCELED
:
3526 case CPU_UP_CANCELED_FROZEN
:
3528 case CPU_DEAD_FROZEN
:
3529 down_read(&slub_lock
);
3530 list_for_each_entry(s
, &slab_caches
, list
) {
3531 local_irq_save(flags
);
3532 __flush_cpu_slab(s
, cpu
);
3533 local_irq_restore(flags
);
3535 up_read(&slub_lock
);
3543 static struct notifier_block __cpuinitdata slab_notifier
= {
3544 .notifier_call
= slab_cpuup_callback
3549 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
3551 struct kmem_cache
*s
;
3554 if (unlikely(size
> SLUB_MAX_SIZE
))
3555 return kmalloc_large(size
, gfpflags
);
3557 s
= get_slab(size
, gfpflags
);
3559 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3562 ret
= slab_alloc(s
, gfpflags
, NUMA_NO_NODE
, caller
);
3564 /* Honor the call site pointer we recieved. */
3565 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
3571 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3572 int node
, unsigned long caller
)
3574 struct kmem_cache
*s
;
3577 if (unlikely(size
> SLUB_MAX_SIZE
)) {
3578 ret
= kmalloc_large_node(size
, gfpflags
, node
);
3580 trace_kmalloc_node(caller
, ret
,
3581 size
, PAGE_SIZE
<< get_order(size
),
3587 s
= get_slab(size
, gfpflags
);
3589 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3592 ret
= slab_alloc(s
, gfpflags
, node
, caller
);
3594 /* Honor the call site pointer we recieved. */
3595 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
3602 static int count_inuse(struct page
*page
)
3607 static int count_total(struct page
*page
)
3609 return page
->objects
;
3613 #ifdef CONFIG_SLUB_DEBUG
3614 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3618 void *addr
= page_address(page
);
3620 if (!check_slab(s
, page
) ||
3621 !on_freelist(s
, page
, NULL
))
3624 /* Now we know that a valid freelist exists */
3625 bitmap_zero(map
, page
->objects
);
3627 get_map(s
, page
, map
);
3628 for_each_object(p
, s
, addr
, page
->objects
) {
3629 if (test_bit(slab_index(p
, s
, addr
), map
))
3630 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
3634 for_each_object(p
, s
, addr
, page
->objects
)
3635 if (!test_bit(slab_index(p
, s
, addr
), map
))
3636 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
3641 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
3644 if (slab_trylock(page
)) {
3645 validate_slab(s
, page
, map
);
3648 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
3652 static int validate_slab_node(struct kmem_cache
*s
,
3653 struct kmem_cache_node
*n
, unsigned long *map
)
3655 unsigned long count
= 0;
3657 unsigned long flags
;
3659 spin_lock_irqsave(&n
->list_lock
, flags
);
3661 list_for_each_entry(page
, &n
->partial
, lru
) {
3662 validate_slab_slab(s
, page
, map
);
3665 if (count
!= n
->nr_partial
)
3666 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
3667 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
3669 if (!(s
->flags
& SLAB_STORE_USER
))
3672 list_for_each_entry(page
, &n
->full
, lru
) {
3673 validate_slab_slab(s
, page
, map
);
3676 if (count
!= atomic_long_read(&n
->nr_slabs
))
3677 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
3678 "counter=%ld\n", s
->name
, count
,
3679 atomic_long_read(&n
->nr_slabs
));
3682 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3686 static long validate_slab_cache(struct kmem_cache
*s
)
3689 unsigned long count
= 0;
3690 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
3691 sizeof(unsigned long), GFP_KERNEL
);
3697 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3698 struct kmem_cache_node
*n
= get_node(s
, node
);
3700 count
+= validate_slab_node(s
, n
, map
);
3706 * Generate lists of code addresses where slabcache objects are allocated
3711 unsigned long count
;
3718 DECLARE_BITMAP(cpus
, NR_CPUS
);
3724 unsigned long count
;
3725 struct location
*loc
;
3728 static void free_loc_track(struct loc_track
*t
)
3731 free_pages((unsigned long)t
->loc
,
3732 get_order(sizeof(struct location
) * t
->max
));
3735 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
3740 order
= get_order(sizeof(struct location
) * max
);
3742 l
= (void *)__get_free_pages(flags
, order
);
3747 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
3755 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
3756 const struct track
*track
)
3758 long start
, end
, pos
;
3760 unsigned long caddr
;
3761 unsigned long age
= jiffies
- track
->when
;
3767 pos
= start
+ (end
- start
+ 1) / 2;
3770 * There is nothing at "end". If we end up there
3771 * we need to add something to before end.
3776 caddr
= t
->loc
[pos
].addr
;
3777 if (track
->addr
== caddr
) {
3783 if (age
< l
->min_time
)
3785 if (age
> l
->max_time
)
3788 if (track
->pid
< l
->min_pid
)
3789 l
->min_pid
= track
->pid
;
3790 if (track
->pid
> l
->max_pid
)
3791 l
->max_pid
= track
->pid
;
3793 cpumask_set_cpu(track
->cpu
,
3794 to_cpumask(l
->cpus
));
3796 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3800 if (track
->addr
< caddr
)
3807 * Not found. Insert new tracking element.
3809 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
3815 (t
->count
- pos
) * sizeof(struct location
));
3818 l
->addr
= track
->addr
;
3822 l
->min_pid
= track
->pid
;
3823 l
->max_pid
= track
->pid
;
3824 cpumask_clear(to_cpumask(l
->cpus
));
3825 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
3826 nodes_clear(l
->nodes
);
3827 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
3831 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
3832 struct page
*page
, enum track_item alloc
,
3835 void *addr
= page_address(page
);
3838 bitmap_zero(map
, page
->objects
);
3839 get_map(s
, page
, map
);
3841 for_each_object(p
, s
, addr
, page
->objects
)
3842 if (!test_bit(slab_index(p
, s
, addr
), map
))
3843 add_location(t
, s
, get_track(s
, p
, alloc
));
3846 static int list_locations(struct kmem_cache
*s
, char *buf
,
3847 enum track_item alloc
)
3851 struct loc_track t
= { 0, 0, NULL
};
3853 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
3854 sizeof(unsigned long), GFP_KERNEL
);
3856 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
3859 return sprintf(buf
, "Out of memory\n");
3861 /* Push back cpu slabs */
3864 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3865 struct kmem_cache_node
*n
= get_node(s
, node
);
3866 unsigned long flags
;
3869 if (!atomic_long_read(&n
->nr_slabs
))
3872 spin_lock_irqsave(&n
->list_lock
, flags
);
3873 list_for_each_entry(page
, &n
->partial
, lru
)
3874 process_slab(&t
, s
, page
, alloc
, map
);
3875 list_for_each_entry(page
, &n
->full
, lru
)
3876 process_slab(&t
, s
, page
, alloc
, map
);
3877 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3880 for (i
= 0; i
< t
.count
; i
++) {
3881 struct location
*l
= &t
.loc
[i
];
3883 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
3885 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
3888 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
3890 len
+= sprintf(buf
+ len
, "<not-available>");
3892 if (l
->sum_time
!= l
->min_time
) {
3893 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
3895 (long)div_u64(l
->sum_time
, l
->count
),
3898 len
+= sprintf(buf
+ len
, " age=%ld",
3901 if (l
->min_pid
!= l
->max_pid
)
3902 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
3903 l
->min_pid
, l
->max_pid
);
3905 len
+= sprintf(buf
+ len
, " pid=%ld",
3908 if (num_online_cpus() > 1 &&
3909 !cpumask_empty(to_cpumask(l
->cpus
)) &&
3910 len
< PAGE_SIZE
- 60) {
3911 len
+= sprintf(buf
+ len
, " cpus=");
3912 len
+= cpulist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3913 to_cpumask(l
->cpus
));
3916 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
3917 len
< PAGE_SIZE
- 60) {
3918 len
+= sprintf(buf
+ len
, " nodes=");
3919 len
+= nodelist_scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
3923 len
+= sprintf(buf
+ len
, "\n");
3929 len
+= sprintf(buf
, "No data\n");
3934 #ifdef SLUB_RESILIENCY_TEST
3935 static void resiliency_test(void)
3939 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || SLUB_PAGE_SHIFT
< 10);
3941 printk(KERN_ERR
"SLUB resiliency testing\n");
3942 printk(KERN_ERR
"-----------------------\n");
3943 printk(KERN_ERR
"A. Corruption after allocation\n");
3945 p
= kzalloc(16, GFP_KERNEL
);
3947 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
3948 " 0x12->0x%p\n\n", p
+ 16);
3950 validate_slab_cache(kmalloc_caches
[4]);
3952 /* Hmmm... The next two are dangerous */
3953 p
= kzalloc(32, GFP_KERNEL
);
3954 p
[32 + sizeof(void *)] = 0x34;
3955 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
3956 " 0x34 -> -0x%p\n", p
);
3958 "If allocated object is overwritten then not detectable\n\n");
3960 validate_slab_cache(kmalloc_caches
[5]);
3961 p
= kzalloc(64, GFP_KERNEL
);
3962 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
3964 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3967 "If allocated object is overwritten then not detectable\n\n");
3968 validate_slab_cache(kmalloc_caches
[6]);
3970 printk(KERN_ERR
"\nB. Corruption after free\n");
3971 p
= kzalloc(128, GFP_KERNEL
);
3974 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
3975 validate_slab_cache(kmalloc_caches
[7]);
3977 p
= kzalloc(256, GFP_KERNEL
);
3980 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3982 validate_slab_cache(kmalloc_caches
[8]);
3984 p
= kzalloc(512, GFP_KERNEL
);
3987 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
3988 validate_slab_cache(kmalloc_caches
[9]);
3992 static void resiliency_test(void) {};
3997 enum slab_stat_type
{
3998 SL_ALL
, /* All slabs */
3999 SL_PARTIAL
, /* Only partially allocated slabs */
4000 SL_CPU
, /* Only slabs used for cpu caches */
4001 SL_OBJECTS
, /* Determine allocated objects not slabs */
4002 SL_TOTAL
/* Determine object capacity not slabs */
4005 #define SO_ALL (1 << SL_ALL)
4006 #define SO_PARTIAL (1 << SL_PARTIAL)
4007 #define SO_CPU (1 << SL_CPU)
4008 #define SO_OBJECTS (1 << SL_OBJECTS)
4009 #define SO_TOTAL (1 << SL_TOTAL)
4011 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4012 char *buf
, unsigned long flags
)
4014 unsigned long total
= 0;
4017 unsigned long *nodes
;
4018 unsigned long *per_cpu
;
4020 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4023 per_cpu
= nodes
+ nr_node_ids
;
4025 if (flags
& SO_CPU
) {
4028 for_each_possible_cpu(cpu
) {
4029 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
4031 if (!c
|| c
->node
< 0)
4035 if (flags
& SO_TOTAL
)
4036 x
= c
->page
->objects
;
4037 else if (flags
& SO_OBJECTS
)
4043 nodes
[c
->node
] += x
;
4049 lock_memory_hotplug();
4050 #ifdef CONFIG_SLUB_DEBUG
4051 if (flags
& SO_ALL
) {
4052 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4053 struct kmem_cache_node
*n
= get_node(s
, node
);
4055 if (flags
& SO_TOTAL
)
4056 x
= atomic_long_read(&n
->total_objects
);
4057 else if (flags
& SO_OBJECTS
)
4058 x
= atomic_long_read(&n
->total_objects
) -
4059 count_partial(n
, count_free
);
4062 x
= atomic_long_read(&n
->nr_slabs
);
4069 if (flags
& SO_PARTIAL
) {
4070 for_each_node_state(node
, N_NORMAL_MEMORY
) {
4071 struct kmem_cache_node
*n
= get_node(s
, node
);
4073 if (flags
& SO_TOTAL
)
4074 x
= count_partial(n
, count_total
);
4075 else if (flags
& SO_OBJECTS
)
4076 x
= count_partial(n
, count_inuse
);
4083 x
= sprintf(buf
, "%lu", total
);
4085 for_each_node_state(node
, N_NORMAL_MEMORY
)
4087 x
+= sprintf(buf
+ x
, " N%d=%lu",
4090 unlock_memory_hotplug();
4092 return x
+ sprintf(buf
+ x
, "\n");
4095 #ifdef CONFIG_SLUB_DEBUG
4096 static int any_slab_objects(struct kmem_cache
*s
)
4100 for_each_online_node(node
) {
4101 struct kmem_cache_node
*n
= get_node(s
, node
);
4106 if (atomic_long_read(&n
->total_objects
))
4113 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4114 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4116 struct slab_attribute
{
4117 struct attribute attr
;
4118 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4119 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4122 #define SLAB_ATTR_RO(_name) \
4123 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4125 #define SLAB_ATTR(_name) \
4126 static struct slab_attribute _name##_attr = \
4127 __ATTR(_name, 0644, _name##_show, _name##_store)
4129 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4131 return sprintf(buf
, "%d\n", s
->size
);
4133 SLAB_ATTR_RO(slab_size
);
4135 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4137 return sprintf(buf
, "%d\n", s
->align
);
4139 SLAB_ATTR_RO(align
);
4141 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4143 return sprintf(buf
, "%d\n", s
->objsize
);
4145 SLAB_ATTR_RO(object_size
);
4147 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4149 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4151 SLAB_ATTR_RO(objs_per_slab
);
4153 static ssize_t
order_store(struct kmem_cache
*s
,
4154 const char *buf
, size_t length
)
4156 unsigned long order
;
4159 err
= strict_strtoul(buf
, 10, &order
);
4163 if (order
> slub_max_order
|| order
< slub_min_order
)
4166 calculate_sizes(s
, order
);
4170 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4172 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4176 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4178 return sprintf(buf
, "%lu\n", s
->min_partial
);
4181 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4187 err
= strict_strtoul(buf
, 10, &min
);
4191 set_min_partial(s
, min
);
4194 SLAB_ATTR(min_partial
);
4196 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4200 return sprintf(buf
, "%pS\n", s
->ctor
);
4204 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4206 return sprintf(buf
, "%d\n", s
->refcount
- 1);
4208 SLAB_ATTR_RO(aliases
);
4210 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4212 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4214 SLAB_ATTR_RO(partial
);
4216 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4218 return show_slab_objects(s
, buf
, SO_CPU
);
4220 SLAB_ATTR_RO(cpu_slabs
);
4222 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4224 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4226 SLAB_ATTR_RO(objects
);
4228 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4230 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4232 SLAB_ATTR_RO(objects_partial
);
4234 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4236 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4239 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4240 const char *buf
, size_t length
)
4242 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4244 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4247 SLAB_ATTR(reclaim_account
);
4249 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4251 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4253 SLAB_ATTR_RO(hwcache_align
);
4255 #ifdef CONFIG_ZONE_DMA
4256 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4258 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4260 SLAB_ATTR_RO(cache_dma
);
4263 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4265 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4267 SLAB_ATTR_RO(destroy_by_rcu
);
4269 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4271 return sprintf(buf
, "%d\n", s
->reserved
);
4273 SLAB_ATTR_RO(reserved
);
4275 #ifdef CONFIG_SLUB_DEBUG
4276 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4278 return show_slab_objects(s
, buf
, SO_ALL
);
4280 SLAB_ATTR_RO(slabs
);
4282 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4284 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4286 SLAB_ATTR_RO(total_objects
);
4288 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4290 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
4293 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
4294 const char *buf
, size_t length
)
4296 s
->flags
&= ~SLAB_DEBUG_FREE
;
4298 s
->flags
|= SLAB_DEBUG_FREE
;
4301 SLAB_ATTR(sanity_checks
);
4303 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
4305 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
4308 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
4311 s
->flags
&= ~SLAB_TRACE
;
4313 s
->flags
|= SLAB_TRACE
;
4318 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
4320 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
4323 static ssize_t
red_zone_store(struct kmem_cache
*s
,
4324 const char *buf
, size_t length
)
4326 if (any_slab_objects(s
))
4329 s
->flags
&= ~SLAB_RED_ZONE
;
4331 s
->flags
|= SLAB_RED_ZONE
;
4332 calculate_sizes(s
, -1);
4335 SLAB_ATTR(red_zone
);
4337 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
4339 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
4342 static ssize_t
poison_store(struct kmem_cache
*s
,
4343 const char *buf
, size_t length
)
4345 if (any_slab_objects(s
))
4348 s
->flags
&= ~SLAB_POISON
;
4350 s
->flags
|= SLAB_POISON
;
4351 calculate_sizes(s
, -1);
4356 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
4358 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
4361 static ssize_t
store_user_store(struct kmem_cache
*s
,
4362 const char *buf
, size_t length
)
4364 if (any_slab_objects(s
))
4367 s
->flags
&= ~SLAB_STORE_USER
;
4369 s
->flags
|= SLAB_STORE_USER
;
4370 calculate_sizes(s
, -1);
4373 SLAB_ATTR(store_user
);
4375 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
4380 static ssize_t
validate_store(struct kmem_cache
*s
,
4381 const char *buf
, size_t length
)
4385 if (buf
[0] == '1') {
4386 ret
= validate_slab_cache(s
);
4392 SLAB_ATTR(validate
);
4394 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
4396 if (!(s
->flags
& SLAB_STORE_USER
))
4398 return list_locations(s
, buf
, TRACK_ALLOC
);
4400 SLAB_ATTR_RO(alloc_calls
);
4402 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
4404 if (!(s
->flags
& SLAB_STORE_USER
))
4406 return list_locations(s
, buf
, TRACK_FREE
);
4408 SLAB_ATTR_RO(free_calls
);
4409 #endif /* CONFIG_SLUB_DEBUG */
4411 #ifdef CONFIG_FAILSLAB
4412 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
4414 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
4417 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
4420 s
->flags
&= ~SLAB_FAILSLAB
;
4422 s
->flags
|= SLAB_FAILSLAB
;
4425 SLAB_ATTR(failslab
);
4428 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
4433 static ssize_t
shrink_store(struct kmem_cache
*s
,
4434 const char *buf
, size_t length
)
4436 if (buf
[0] == '1') {
4437 int rc
= kmem_cache_shrink(s
);
4448 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
4450 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
4453 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
4454 const char *buf
, size_t length
)
4456 unsigned long ratio
;
4459 err
= strict_strtoul(buf
, 10, &ratio
);
4464 s
->remote_node_defrag_ratio
= ratio
* 10;
4468 SLAB_ATTR(remote_node_defrag_ratio
);
4471 #ifdef CONFIG_SLUB_STATS
4472 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
4474 unsigned long sum
= 0;
4477 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
4482 for_each_online_cpu(cpu
) {
4483 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
4489 len
= sprintf(buf
, "%lu", sum
);
4492 for_each_online_cpu(cpu
) {
4493 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
4494 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
4498 return len
+ sprintf(buf
+ len
, "\n");
4501 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
4505 for_each_online_cpu(cpu
)
4506 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
4509 #define STAT_ATTR(si, text) \
4510 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4512 return show_stat(s, buf, si); \
4514 static ssize_t text##_store(struct kmem_cache *s, \
4515 const char *buf, size_t length) \
4517 if (buf[0] != '0') \
4519 clear_stat(s, si); \
4524 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4525 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4526 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4527 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4528 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4529 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4530 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4531 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4532 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4533 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4534 STAT_ATTR(FREE_SLAB
, free_slab
);
4535 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4536 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4537 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4538 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4539 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4540 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4541 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
4544 static struct attribute
*slab_attrs
[] = {
4545 &slab_size_attr
.attr
,
4546 &object_size_attr
.attr
,
4547 &objs_per_slab_attr
.attr
,
4549 &min_partial_attr
.attr
,
4551 &objects_partial_attr
.attr
,
4553 &cpu_slabs_attr
.attr
,
4557 &hwcache_align_attr
.attr
,
4558 &reclaim_account_attr
.attr
,
4559 &destroy_by_rcu_attr
.attr
,
4561 &reserved_attr
.attr
,
4562 #ifdef CONFIG_SLUB_DEBUG
4563 &total_objects_attr
.attr
,
4565 &sanity_checks_attr
.attr
,
4567 &red_zone_attr
.attr
,
4569 &store_user_attr
.attr
,
4570 &validate_attr
.attr
,
4571 &alloc_calls_attr
.attr
,
4572 &free_calls_attr
.attr
,
4574 #ifdef CONFIG_ZONE_DMA
4575 &cache_dma_attr
.attr
,
4578 &remote_node_defrag_ratio_attr
.attr
,
4580 #ifdef CONFIG_SLUB_STATS
4581 &alloc_fastpath_attr
.attr
,
4582 &alloc_slowpath_attr
.attr
,
4583 &free_fastpath_attr
.attr
,
4584 &free_slowpath_attr
.attr
,
4585 &free_frozen_attr
.attr
,
4586 &free_add_partial_attr
.attr
,
4587 &free_remove_partial_attr
.attr
,
4588 &alloc_from_partial_attr
.attr
,
4589 &alloc_slab_attr
.attr
,
4590 &alloc_refill_attr
.attr
,
4591 &free_slab_attr
.attr
,
4592 &cpuslab_flush_attr
.attr
,
4593 &deactivate_full_attr
.attr
,
4594 &deactivate_empty_attr
.attr
,
4595 &deactivate_to_head_attr
.attr
,
4596 &deactivate_to_tail_attr
.attr
,
4597 &deactivate_remote_frees_attr
.attr
,
4598 &order_fallback_attr
.attr
,
4600 #ifdef CONFIG_FAILSLAB
4601 &failslab_attr
.attr
,
4607 static struct attribute_group slab_attr_group
= {
4608 .attrs
= slab_attrs
,
4611 static ssize_t
slab_attr_show(struct kobject
*kobj
,
4612 struct attribute
*attr
,
4615 struct slab_attribute
*attribute
;
4616 struct kmem_cache
*s
;
4619 attribute
= to_slab_attr(attr
);
4622 if (!attribute
->show
)
4625 err
= attribute
->show(s
, buf
);
4630 static ssize_t
slab_attr_store(struct kobject
*kobj
,
4631 struct attribute
*attr
,
4632 const char *buf
, size_t len
)
4634 struct slab_attribute
*attribute
;
4635 struct kmem_cache
*s
;
4638 attribute
= to_slab_attr(attr
);
4641 if (!attribute
->store
)
4644 err
= attribute
->store(s
, buf
, len
);
4649 static void kmem_cache_release(struct kobject
*kobj
)
4651 struct kmem_cache
*s
= to_slab(kobj
);
4657 static const struct sysfs_ops slab_sysfs_ops
= {
4658 .show
= slab_attr_show
,
4659 .store
= slab_attr_store
,
4662 static struct kobj_type slab_ktype
= {
4663 .sysfs_ops
= &slab_sysfs_ops
,
4664 .release
= kmem_cache_release
4667 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
4669 struct kobj_type
*ktype
= get_ktype(kobj
);
4671 if (ktype
== &slab_ktype
)
4676 static const struct kset_uevent_ops slab_uevent_ops
= {
4677 .filter
= uevent_filter
,
4680 static struct kset
*slab_kset
;
4682 #define ID_STR_LENGTH 64
4684 /* Create a unique string id for a slab cache:
4686 * Format :[flags-]size
4688 static char *create_unique_id(struct kmem_cache
*s
)
4690 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
4697 * First flags affecting slabcache operations. We will only
4698 * get here for aliasable slabs so we do not need to support
4699 * too many flags. The flags here must cover all flags that
4700 * are matched during merging to guarantee that the id is
4703 if (s
->flags
& SLAB_CACHE_DMA
)
4705 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
4707 if (s
->flags
& SLAB_DEBUG_FREE
)
4709 if (!(s
->flags
& SLAB_NOTRACK
))
4713 p
+= sprintf(p
, "%07d", s
->size
);
4714 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
4718 static int sysfs_slab_add(struct kmem_cache
*s
)
4724 if (slab_state
< SYSFS
)
4725 /* Defer until later */
4728 unmergeable
= slab_unmergeable(s
);
4731 * Slabcache can never be merged so we can use the name proper.
4732 * This is typically the case for debug situations. In that
4733 * case we can catch duplicate names easily.
4735 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
4739 * Create a unique name for the slab as a target
4742 name
= create_unique_id(s
);
4745 s
->kobj
.kset
= slab_kset
;
4746 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, name
);
4748 kobject_put(&s
->kobj
);
4752 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
4754 kobject_del(&s
->kobj
);
4755 kobject_put(&s
->kobj
);
4758 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
4760 /* Setup first alias */
4761 sysfs_slab_alias(s
, s
->name
);
4767 static void sysfs_slab_remove(struct kmem_cache
*s
)
4769 if (slab_state
< SYSFS
)
4771 * Sysfs has not been setup yet so no need to remove the
4776 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
4777 kobject_del(&s
->kobj
);
4778 kobject_put(&s
->kobj
);
4782 * Need to buffer aliases during bootup until sysfs becomes
4783 * available lest we lose that information.
4785 struct saved_alias
{
4786 struct kmem_cache
*s
;
4788 struct saved_alias
*next
;
4791 static struct saved_alias
*alias_list
;
4793 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
4795 struct saved_alias
*al
;
4797 if (slab_state
== SYSFS
) {
4799 * If we have a leftover link then remove it.
4801 sysfs_remove_link(&slab_kset
->kobj
, name
);
4802 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
4805 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
4811 al
->next
= alias_list
;
4816 static int __init
slab_sysfs_init(void)
4818 struct kmem_cache
*s
;
4821 down_write(&slub_lock
);
4823 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
4825 up_write(&slub_lock
);
4826 printk(KERN_ERR
"Cannot register slab subsystem.\n");
4832 list_for_each_entry(s
, &slab_caches
, list
) {
4833 err
= sysfs_slab_add(s
);
4835 printk(KERN_ERR
"SLUB: Unable to add boot slab %s"
4836 " to sysfs\n", s
->name
);
4839 while (alias_list
) {
4840 struct saved_alias
*al
= alias_list
;
4842 alias_list
= alias_list
->next
;
4843 err
= sysfs_slab_alias(al
->s
, al
->name
);
4845 printk(KERN_ERR
"SLUB: Unable to add boot slab alias"
4846 " %s to sysfs\n", s
->name
);
4850 up_write(&slub_lock
);
4855 __initcall(slab_sysfs_init
);
4856 #endif /* CONFIG_SYSFS */
4859 * The /proc/slabinfo ABI
4861 #ifdef CONFIG_SLABINFO
4862 static void print_slabinfo_header(struct seq_file
*m
)
4864 seq_puts(m
, "slabinfo - version: 2.1\n");
4865 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
4866 "<objperslab> <pagesperslab>");
4867 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
4868 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4872 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4876 down_read(&slub_lock
);
4878 print_slabinfo_header(m
);
4880 return seq_list_start(&slab_caches
, *pos
);
4883 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4885 return seq_list_next(p
, &slab_caches
, pos
);
4888 static void s_stop(struct seq_file
*m
, void *p
)
4890 up_read(&slub_lock
);
4893 static int s_show(struct seq_file
*m
, void *p
)
4895 unsigned long nr_partials
= 0;
4896 unsigned long nr_slabs
= 0;
4897 unsigned long nr_inuse
= 0;
4898 unsigned long nr_objs
= 0;
4899 unsigned long nr_free
= 0;
4900 struct kmem_cache
*s
;
4903 s
= list_entry(p
, struct kmem_cache
, list
);
4905 for_each_online_node(node
) {
4906 struct kmem_cache_node
*n
= get_node(s
, node
);
4911 nr_partials
+= n
->nr_partial
;
4912 nr_slabs
+= atomic_long_read(&n
->nr_slabs
);
4913 nr_objs
+= atomic_long_read(&n
->total_objects
);
4914 nr_free
+= count_partial(n
, count_free
);
4917 nr_inuse
= nr_objs
- nr_free
;
4919 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d", s
->name
, nr_inuse
,
4920 nr_objs
, s
->size
, oo_objects(s
->oo
),
4921 (1 << oo_order(s
->oo
)));
4922 seq_printf(m
, " : tunables %4u %4u %4u", 0, 0, 0);
4923 seq_printf(m
, " : slabdata %6lu %6lu %6lu", nr_slabs
, nr_slabs
,
4929 static const struct seq_operations slabinfo_op
= {
4936 static int slabinfo_open(struct inode
*inode
, struct file
*file
)
4938 return seq_open(file
, &slabinfo_op
);
4941 static const struct file_operations proc_slabinfo_operations
= {
4942 .open
= slabinfo_open
,
4944 .llseek
= seq_lseek
,
4945 .release
= seq_release
,
4948 static int __init
slab_proc_init(void)
4950 proc_create("slabinfo", S_IRUGO
, NULL
, &proc_slabinfo_operations
);
4953 module_init(slab_proc_init
);
4954 #endif /* CONFIG_SLABINFO */