2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is used as a cpu cache. Allocations
82 * may be performed from the slab. The slab is not
83 * on any slab list and cannot be moved onto one.
84 * The cpu slab may be equipped with an additioanl
85 * lockless_freelist that allows lockless access to
86 * free objects in addition to the regular freelist
87 * that requires the slab lock.
89 * PageError Slab requires special handling due to debug
90 * options set. This moves slab handling out of
91 * the fast path and disables lockless freelists.
94 static inline int SlabDebug(struct page
*page
)
96 #ifdef CONFIG_SLUB_DEBUG
97 return PageError(page
);
103 static inline void SetSlabDebug(struct page
*page
)
105 #ifdef CONFIG_SLUB_DEBUG
110 static inline void ClearSlabDebug(struct page
*page
)
112 #ifdef CONFIG_SLUB_DEBUG
113 ClearPageError(page
);
118 * Issues still to be resolved:
120 * - The per cpu array is updated for each new slab and and is a remote
121 * cacheline for most nodes. This could become a bouncing cacheline given
122 * enough frequent updates. There are 16 pointers in a cacheline, so at
123 * max 16 cpus could compete for the cacheline which may be okay.
125 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
127 * - Variable sizing of the per node arrays
130 /* Enable to test recovery from slab corruption on boot */
131 #undef SLUB_RESILIENCY_TEST
136 * Small page size. Make sure that we do not fragment memory
138 #define DEFAULT_MAX_ORDER 1
139 #define DEFAULT_MIN_OBJECTS 4
144 * Large page machines are customarily able to handle larger
147 #define DEFAULT_MAX_ORDER 2
148 #define DEFAULT_MIN_OBJECTS 8
153 * Mininum number of partial slabs. These will be left on the partial
154 * lists even if they are empty. kmem_cache_shrink may reclaim them.
156 #define MIN_PARTIAL 2
159 * Maximum number of desirable partial slabs.
160 * The existence of more partial slabs makes kmem_cache_shrink
161 * sort the partial list by the number of objects in the.
163 #define MAX_PARTIAL 10
165 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
166 SLAB_POISON | SLAB_STORE_USER)
169 * Set of flags that will prevent slab merging
171 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
172 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
174 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
177 #ifndef ARCH_KMALLOC_MINALIGN
178 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
181 #ifndef ARCH_SLAB_MINALIGN
182 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
185 /* Internal SLUB flags */
186 #define __OBJECT_POISON 0x80000000 /* Poison object */
188 /* Not all arches define cache_line_size */
189 #ifndef cache_line_size
190 #define cache_line_size() L1_CACHE_BYTES
193 static int kmem_size
= sizeof(struct kmem_cache
);
196 static struct notifier_block slab_notifier
;
200 DOWN
, /* No slab functionality available */
201 PARTIAL
, /* kmem_cache_open() works but kmalloc does not */
202 UP
, /* Everything works but does not show up in sysfs */
206 /* A list of all slab caches on the system */
207 static DECLARE_RWSEM(slub_lock
);
208 LIST_HEAD(slab_caches
);
211 * Tracking user of a slab.
214 void *addr
; /* Called from address */
215 int cpu
; /* Was running on cpu */
216 int pid
; /* Pid context */
217 unsigned long when
; /* When did the operation occur */
220 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
222 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
223 static int sysfs_slab_add(struct kmem_cache
*);
224 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
225 static void sysfs_slab_remove(struct kmem_cache
*);
227 static int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
228 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
) { return 0; }
229 static void sysfs_slab_remove(struct kmem_cache
*s
) {}
232 /********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
236 int slab_is_available(void)
238 return slab_state
>= UP
;
241 static inline struct kmem_cache_node
*get_node(struct kmem_cache
*s
, int node
)
244 return s
->node
[node
];
246 return &s
->local_node
;
250 static inline int check_valid_pointer(struct kmem_cache
*s
,
251 struct page
*page
, const void *object
)
258 base
= page_address(page
);
259 if (object
< base
|| object
>= base
+ s
->objects
* s
->size
||
260 (object
- base
) % s
->size
) {
268 * Slow version of get and set free pointer.
270 * This version requires touching the cache lines of kmem_cache which
271 * we avoid to do in the fast alloc free paths. There we obtain the offset
272 * from the page struct.
274 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
276 return *(void **)(object
+ s
->offset
);
279 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
281 *(void **)(object
+ s
->offset
) = fp
;
284 /* Loop over all objects in a slab */
285 #define for_each_object(__p, __s, __addr) \
286 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
290 #define for_each_free_object(__p, __s, __free) \
291 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
293 /* Determine object index from a given position */
294 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
296 return (p
- addr
) / s
->size
;
299 #ifdef CONFIG_SLUB_DEBUG
303 static int slub_debug
;
305 static char *slub_debug_slabs
;
310 static void print_section(char *text
, u8
*addr
, unsigned int length
)
318 for (i
= 0; i
< length
; i
++) {
320 printk(KERN_ERR
"%10s 0x%p: ", text
, addr
+ i
);
323 printk(" %02x", addr
[i
]);
325 ascii
[offset
] = isgraph(addr
[i
]) ? addr
[i
] : '.';
327 printk(" %s\n",ascii
);
338 printk(" %s\n", ascii
);
342 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
343 enum track_item alloc
)
348 p
= object
+ s
->offset
+ sizeof(void *);
350 p
= object
+ s
->inuse
;
355 static void set_track(struct kmem_cache
*s
, void *object
,
356 enum track_item alloc
, void *addr
)
361 p
= object
+ s
->offset
+ sizeof(void *);
363 p
= object
+ s
->inuse
;
368 p
->cpu
= smp_processor_id();
369 p
->pid
= current
? current
->pid
: -1;
372 memset(p
, 0, sizeof(struct track
));
375 static void init_tracking(struct kmem_cache
*s
, void *object
)
377 if (s
->flags
& SLAB_STORE_USER
) {
378 set_track(s
, object
, TRACK_FREE
, NULL
);
379 set_track(s
, object
, TRACK_ALLOC
, NULL
);
383 static void print_track(const char *s
, struct track
*t
)
388 printk(KERN_ERR
"%s: ", s
);
389 __print_symbol("%s", (unsigned long)t
->addr
);
390 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies
- t
->when
, t
->cpu
, t
->pid
);
393 static void print_trailer(struct kmem_cache
*s
, u8
*p
)
395 unsigned int off
; /* Offset of last byte */
397 if (s
->flags
& SLAB_RED_ZONE
)
398 print_section("Redzone", p
+ s
->objsize
,
399 s
->inuse
- s
->objsize
);
401 printk(KERN_ERR
"FreePointer 0x%p -> 0x%p\n",
403 get_freepointer(s
, p
));
406 off
= s
->offset
+ sizeof(void *);
410 if (s
->flags
& SLAB_STORE_USER
) {
411 print_track("Last alloc", get_track(s
, p
, TRACK_ALLOC
));
412 print_track("Last free ", get_track(s
, p
, TRACK_FREE
));
413 off
+= 2 * sizeof(struct track
);
417 /* Beginning of the filler is the free pointer */
418 print_section("Filler", p
+ off
, s
->size
- off
);
421 static void object_err(struct kmem_cache
*s
, struct page
*page
,
422 u8
*object
, char *reason
)
424 u8
*addr
= page_address(page
);
426 printk(KERN_ERR
"*** SLUB %s: %s@0x%p slab 0x%p\n",
427 s
->name
, reason
, object
, page
);
428 printk(KERN_ERR
" offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
429 object
- addr
, page
->flags
, page
->inuse
, page
->freelist
);
430 if (object
> addr
+ 16)
431 print_section("Bytes b4", object
- 16, 16);
432 print_section("Object", object
, min(s
->objsize
, 128));
433 print_trailer(s
, object
);
437 static void slab_err(struct kmem_cache
*s
, struct page
*page
, char *reason
, ...)
442 va_start(args
, reason
);
443 vsnprintf(buf
, sizeof(buf
), reason
, args
);
445 printk(KERN_ERR
"*** SLUB %s: %s in slab @0x%p\n", s
->name
, buf
,
450 static void init_object(struct kmem_cache
*s
, void *object
, int active
)
454 if (s
->flags
& __OBJECT_POISON
) {
455 memset(p
, POISON_FREE
, s
->objsize
- 1);
456 p
[s
->objsize
-1] = POISON_END
;
459 if (s
->flags
& SLAB_RED_ZONE
)
460 memset(p
+ s
->objsize
,
461 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
,
462 s
->inuse
- s
->objsize
);
465 static int check_bytes(u8
*start
, unsigned int value
, unsigned int bytes
)
468 if (*start
!= (u8
)value
)
480 * Bytes of the object to be managed.
481 * If the freepointer may overlay the object then the free
482 * pointer is the first word of the object.
484 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
487 * object + s->objsize
488 * Padding to reach word boundary. This is also used for Redzoning.
489 * Padding is extended by another word if Redzoning is enabled and
492 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
493 * 0xcc (RED_ACTIVE) for objects in use.
496 * Meta data starts here.
498 * A. Free pointer (if we cannot overwrite object on free)
499 * B. Tracking data for SLAB_STORE_USER
500 * C. Padding to reach required alignment boundary or at mininum
501 * one word if debuggin is on to be able to detect writes
502 * before the word boundary.
504 * Padding is done using 0x5a (POISON_INUSE)
507 * Nothing is used beyond s->size.
509 * If slabcaches are merged then the objsize and inuse boundaries are mostly
510 * ignored. And therefore no slab options that rely on these boundaries
511 * may be used with merged slabcaches.
514 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
515 void *from
, void *to
)
517 printk(KERN_ERR
"@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
518 s
->name
, message
, data
, from
, to
- 1);
519 memset(from
, data
, to
- from
);
522 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
524 unsigned long off
= s
->inuse
; /* The end of info */
527 /* Freepointer is placed after the object. */
528 off
+= sizeof(void *);
530 if (s
->flags
& SLAB_STORE_USER
)
531 /* We also have user information there */
532 off
+= 2 * sizeof(struct track
);
537 if (check_bytes(p
+ off
, POISON_INUSE
, s
->size
- off
))
540 object_err(s
, page
, p
, "Object padding check fails");
545 restore_bytes(s
, "object padding", POISON_INUSE
, p
+ off
, p
+ s
->size
);
549 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
552 int length
, remainder
;
554 if (!(s
->flags
& SLAB_POISON
))
557 p
= page_address(page
);
558 length
= s
->objects
* s
->size
;
559 remainder
= (PAGE_SIZE
<< s
->order
) - length
;
563 if (!check_bytes(p
+ length
, POISON_INUSE
, remainder
)) {
564 slab_err(s
, page
, "Padding check failed");
565 restore_bytes(s
, "slab padding", POISON_INUSE
, p
+ length
,
566 p
+ length
+ remainder
);
572 static int check_object(struct kmem_cache
*s
, struct page
*page
,
573 void *object
, int active
)
576 u8
*endobject
= object
+ s
->objsize
;
578 if (s
->flags
& SLAB_RED_ZONE
) {
580 active
? SLUB_RED_ACTIVE
: SLUB_RED_INACTIVE
;
582 if (!check_bytes(endobject
, red
, s
->inuse
- s
->objsize
)) {
583 object_err(s
, page
, object
,
584 active
? "Redzone Active" : "Redzone Inactive");
585 restore_bytes(s
, "redzone", red
,
586 endobject
, object
+ s
->inuse
);
590 if ((s
->flags
& SLAB_POISON
) && s
->objsize
< s
->inuse
&&
591 !check_bytes(endobject
, POISON_INUSE
,
592 s
->inuse
- s
->objsize
)) {
593 object_err(s
, page
, p
, "Alignment padding check fails");
595 * Fix it so that there will not be another report.
597 * Hmmm... We may be corrupting an object that now expects
598 * to be longer than allowed.
600 restore_bytes(s
, "alignment padding", POISON_INUSE
,
601 endobject
, object
+ s
->inuse
);
605 if (s
->flags
& SLAB_POISON
) {
606 if (!active
&& (s
->flags
& __OBJECT_POISON
) &&
607 (!check_bytes(p
, POISON_FREE
, s
->objsize
- 1) ||
608 p
[s
->objsize
- 1] != POISON_END
)) {
610 object_err(s
, page
, p
, "Poison check failed");
611 restore_bytes(s
, "Poison", POISON_FREE
,
612 p
, p
+ s
->objsize
-1);
613 restore_bytes(s
, "Poison", POISON_END
,
614 p
+ s
->objsize
- 1, p
+ s
->objsize
);
618 * check_pad_bytes cleans up on its own.
620 check_pad_bytes(s
, page
, p
);
623 if (!s
->offset
&& active
)
625 * Object and freepointer overlap. Cannot check
626 * freepointer while object is allocated.
630 /* Check free pointer validity */
631 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
632 object_err(s
, page
, p
, "Freepointer corrupt");
634 * No choice but to zap it and thus loose the remainder
635 * of the free objects in this slab. May cause
636 * another error because the object count is now wrong.
638 set_freepointer(s
, p
, NULL
);
644 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
646 VM_BUG_ON(!irqs_disabled());
648 if (!PageSlab(page
)) {
649 slab_err(s
, page
, "Not a valid slab page flags=%lx "
650 "mapping=0x%p count=%d", page
->flags
, page
->mapping
,
654 if (page
->offset
* sizeof(void *) != s
->offset
) {
655 slab_err(s
, page
, "Corrupted offset %lu flags=0x%lx "
656 "mapping=0x%p count=%d",
657 (unsigned long)(page
->offset
* sizeof(void *)),
663 if (page
->inuse
> s
->objects
) {
664 slab_err(s
, page
, "inuse %u > max %u @0x%p flags=%lx "
665 "mapping=0x%p count=%d",
666 s
->name
, page
->inuse
, s
->objects
, page
->flags
,
667 page
->mapping
, page_count(page
));
670 /* Slab_pad_check fixes things up after itself */
671 slab_pad_check(s
, page
);
676 * Determine if a certain object on a page is on the freelist. Must hold the
677 * slab lock to guarantee that the chains are in a consistent state.
679 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
682 void *fp
= page
->freelist
;
685 while (fp
&& nr
<= s
->objects
) {
688 if (!check_valid_pointer(s
, page
, fp
)) {
690 object_err(s
, page
, object
,
691 "Freechain corrupt");
692 set_freepointer(s
, object
, NULL
);
695 slab_err(s
, page
, "Freepointer 0x%p corrupt",
697 page
->freelist
= NULL
;
698 page
->inuse
= s
->objects
;
699 printk(KERN_ERR
"@@@ SLUB %s: Freelist "
700 "cleared. Slab 0x%p\n",
707 fp
= get_freepointer(s
, object
);
711 if (page
->inuse
!= s
->objects
- nr
) {
712 slab_err(s
, page
, "Wrong object count. Counter is %d but "
713 "counted were %d", s
, page
, page
->inuse
,
715 page
->inuse
= s
->objects
- nr
;
716 printk(KERN_ERR
"@@@ SLUB %s: Object count adjusted. "
717 "Slab @0x%p\n", s
->name
, page
);
719 return search
== NULL
;
723 * Tracking of fully allocated slabs for debugging purposes.
725 static void add_full(struct kmem_cache_node
*n
, struct page
*page
)
727 spin_lock(&n
->list_lock
);
728 list_add(&page
->lru
, &n
->full
);
729 spin_unlock(&n
->list_lock
);
732 static void remove_full(struct kmem_cache
*s
, struct page
*page
)
734 struct kmem_cache_node
*n
;
736 if (!(s
->flags
& SLAB_STORE_USER
))
739 n
= get_node(s
, page_to_nid(page
));
741 spin_lock(&n
->list_lock
);
742 list_del(&page
->lru
);
743 spin_unlock(&n
->list_lock
);
746 static int alloc_object_checks(struct kmem_cache
*s
, struct page
*page
,
749 if (!check_slab(s
, page
))
752 if (object
&& !on_freelist(s
, page
, object
)) {
753 slab_err(s
, page
, "Object 0x%p already allocated", object
);
757 if (!check_valid_pointer(s
, page
, object
)) {
758 object_err(s
, page
, object
, "Freelist Pointer check fails");
765 if (!check_object(s
, page
, object
, 0))
770 if (PageSlab(page
)) {
772 * If this is a slab page then lets do the best we can
773 * to avoid issues in the future. Marking all objects
774 * as used avoids touching the remaining objects.
776 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
778 page
->inuse
= s
->objects
;
779 page
->freelist
= NULL
;
780 /* Fix up fields that may be corrupted */
781 page
->offset
= s
->offset
/ sizeof(void *);
786 static int free_object_checks(struct kmem_cache
*s
, struct page
*page
,
789 if (!check_slab(s
, page
))
792 if (!check_valid_pointer(s
, page
, object
)) {
793 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
797 if (on_freelist(s
, page
, object
)) {
798 slab_err(s
, page
, "Object 0x%p already free", object
);
802 if (!check_object(s
, page
, object
, 1))
805 if (unlikely(s
!= page
->slab
)) {
807 slab_err(s
, page
, "Attempt to free object(0x%p) "
808 "outside of slab", object
);
812 "SLUB <none>: no slab for object 0x%p.\n",
817 slab_err(s
, page
, "object at 0x%p belongs "
818 "to slab %s", object
, page
->slab
->name
);
823 printk(KERN_ERR
"@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
824 s
->name
, page
, object
);
828 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
, int alloc
)
830 if (s
->flags
& SLAB_TRACE
) {
831 printk(KERN_INFO
"TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
833 alloc
? "alloc" : "free",
838 print_section("Object", (void *)object
, s
->objsize
);
844 static int __init
setup_slub_debug(char *str
)
846 if (!str
|| *str
!= '=')
847 slub_debug
= DEBUG_DEFAULT_FLAGS
;
850 if (*str
== 0 || *str
== ',')
851 slub_debug
= DEBUG_DEFAULT_FLAGS
;
853 for( ;*str
&& *str
!= ','; str
++)
855 case 'f' : case 'F' :
856 slub_debug
|= SLAB_DEBUG_FREE
;
858 case 'z' : case 'Z' :
859 slub_debug
|= SLAB_RED_ZONE
;
861 case 'p' : case 'P' :
862 slub_debug
|= SLAB_POISON
;
864 case 'u' : case 'U' :
865 slub_debug
|= SLAB_STORE_USER
;
867 case 't' : case 'T' :
868 slub_debug
|= SLAB_TRACE
;
871 printk(KERN_ERR
"slub_debug option '%c' "
872 "unknown. skipped\n",*str
);
877 slub_debug_slabs
= str
+ 1;
881 __setup("slub_debug", setup_slub_debug
);
883 static void kmem_cache_open_debug_check(struct kmem_cache
*s
)
886 * The page->offset field is only 16 bit wide. This is an offset
887 * in units of words from the beginning of an object. If the slab
888 * size is bigger then we cannot move the free pointer behind the
891 * On 32 bit platforms the limit is 256k. On 64bit platforms
894 * Debugging or ctor may create a need to move the free
895 * pointer. Fail if this happens.
897 if (s
->size
>= 65535 * sizeof(void *)) {
898 BUG_ON(s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
|
899 SLAB_STORE_USER
| SLAB_DESTROY_BY_RCU
));
904 * Enable debugging if selected on the kernel commandline.
906 if (slub_debug
&& (!slub_debug_slabs
||
907 strncmp(slub_debug_slabs
, s
->name
,
908 strlen(slub_debug_slabs
)) == 0))
909 s
->flags
|= slub_debug
;
913 static inline int alloc_object_checks(struct kmem_cache
*s
,
914 struct page
*page
, void *object
) { return 0; }
916 static inline int free_object_checks(struct kmem_cache
*s
,
917 struct page
*page
, void *object
) { return 0; }
919 static inline void add_full(struct kmem_cache_node
*n
, struct page
*page
) {}
920 static inline void remove_full(struct kmem_cache
*s
, struct page
*page
) {}
921 static inline void trace(struct kmem_cache
*s
, struct page
*page
,
922 void *object
, int alloc
) {}
923 static inline void init_object(struct kmem_cache
*s
,
924 void *object
, int active
) {}
925 static inline void init_tracking(struct kmem_cache
*s
, void *object
) {}
926 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
928 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
929 void *object
, int active
) { return 1; }
930 static inline void set_track(struct kmem_cache
*s
, void *object
,
931 enum track_item alloc
, void *addr
) {}
932 static inline void kmem_cache_open_debug_check(struct kmem_cache
*s
) {}
936 * Slab allocation and freeing
938 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
941 int pages
= 1 << s
->order
;
946 if (s
->flags
& SLAB_CACHE_DMA
)
950 page
= alloc_pages(flags
, s
->order
);
952 page
= alloc_pages_node(node
, flags
, s
->order
);
957 mod_zone_page_state(page_zone(page
),
958 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
959 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
965 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
968 if (SlabDebug(page
)) {
969 init_object(s
, object
, 0);
970 init_tracking(s
, object
);
973 if (unlikely(s
->ctor
))
974 s
->ctor(object
, s
, SLAB_CTOR_CONSTRUCTOR
);
977 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
980 struct kmem_cache_node
*n
;
986 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
));
988 if (flags
& __GFP_WAIT
)
991 page
= allocate_slab(s
, flags
& GFP_LEVEL_MASK
, node
);
995 n
= get_node(s
, page_to_nid(page
));
997 atomic_long_inc(&n
->nr_slabs
);
998 page
->offset
= s
->offset
/ sizeof(void *);
1000 page
->flags
|= 1 << PG_slab
;
1001 if (s
->flags
& (SLAB_DEBUG_FREE
| SLAB_RED_ZONE
| SLAB_POISON
|
1002 SLAB_STORE_USER
| SLAB_TRACE
))
1005 start
= page_address(page
);
1006 end
= start
+ s
->objects
* s
->size
;
1008 if (unlikely(s
->flags
& SLAB_POISON
))
1009 memset(start
, POISON_INUSE
, PAGE_SIZE
<< s
->order
);
1012 for_each_object(p
, s
, start
) {
1013 setup_object(s
, page
, last
);
1014 set_freepointer(s
, last
, p
);
1017 setup_object(s
, page
, last
);
1018 set_freepointer(s
, last
, NULL
);
1020 page
->freelist
= start
;
1021 page
->lockless_freelist
= NULL
;
1024 if (flags
& __GFP_WAIT
)
1025 local_irq_disable();
1029 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1031 int pages
= 1 << s
->order
;
1033 if (unlikely(SlabDebug(page
))) {
1036 slab_pad_check(s
, page
);
1037 for_each_object(p
, s
, page_address(page
))
1038 check_object(s
, page
, p
, 0);
1041 mod_zone_page_state(page_zone(page
),
1042 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1043 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1046 page
->mapping
= NULL
;
1047 __free_pages(page
, s
->order
);
1050 static void rcu_free_slab(struct rcu_head
*h
)
1054 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1055 __free_slab(page
->slab
, page
);
1058 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1060 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1062 * RCU free overloads the RCU head over the LRU
1064 struct rcu_head
*head
= (void *)&page
->lru
;
1066 call_rcu(head
, rcu_free_slab
);
1068 __free_slab(s
, page
);
1071 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1073 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1075 atomic_long_dec(&n
->nr_slabs
);
1076 reset_page_mapcount(page
);
1077 ClearSlabDebug(page
);
1078 __ClearPageSlab(page
);
1083 * Per slab locking using the pagelock
1085 static __always_inline
void slab_lock(struct page
*page
)
1087 bit_spin_lock(PG_locked
, &page
->flags
);
1090 static __always_inline
void slab_unlock(struct page
*page
)
1092 bit_spin_unlock(PG_locked
, &page
->flags
);
1095 static __always_inline
int slab_trylock(struct page
*page
)
1099 rc
= bit_spin_trylock(PG_locked
, &page
->flags
);
1104 * Management of partially allocated slabs
1106 static void add_partial_tail(struct kmem_cache_node
*n
, struct page
*page
)
1108 spin_lock(&n
->list_lock
);
1110 list_add_tail(&page
->lru
, &n
->partial
);
1111 spin_unlock(&n
->list_lock
);
1114 static void add_partial(struct kmem_cache_node
*n
, struct page
*page
)
1116 spin_lock(&n
->list_lock
);
1118 list_add(&page
->lru
, &n
->partial
);
1119 spin_unlock(&n
->list_lock
);
1122 static void remove_partial(struct kmem_cache
*s
,
1125 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1127 spin_lock(&n
->list_lock
);
1128 list_del(&page
->lru
);
1130 spin_unlock(&n
->list_lock
);
1134 * Lock slab and remove from the partial list.
1136 * Must hold list_lock.
1138 static int lock_and_del_slab(struct kmem_cache_node
*n
, struct page
*page
)
1140 if (slab_trylock(page
)) {
1141 list_del(&page
->lru
);
1149 * Try to allocate a partial slab from a specific node.
1151 static struct page
*get_partial_node(struct kmem_cache_node
*n
)
1156 * Racy check. If we mistakenly see no partial slabs then we
1157 * just allocate an empty slab. If we mistakenly try to get a
1158 * partial slab and there is none available then get_partials()
1161 if (!n
|| !n
->nr_partial
)
1164 spin_lock(&n
->list_lock
);
1165 list_for_each_entry(page
, &n
->partial
, lru
)
1166 if (lock_and_del_slab(n
, page
))
1170 spin_unlock(&n
->list_lock
);
1175 * Get a page from somewhere. Search in increasing NUMA distances.
1177 static struct page
*get_any_partial(struct kmem_cache
*s
, gfp_t flags
)
1180 struct zonelist
*zonelist
;
1185 * The defrag ratio allows a configuration of the tradeoffs between
1186 * inter node defragmentation and node local allocations. A lower
1187 * defrag_ratio increases the tendency to do local allocations
1188 * instead of attempting to obtain partial slabs from other nodes.
1190 * If the defrag_ratio is set to 0 then kmalloc() always
1191 * returns node local objects. If the ratio is higher then kmalloc()
1192 * may return off node objects because partial slabs are obtained
1193 * from other nodes and filled up.
1195 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1196 * defrag_ratio = 1000) then every (well almost) allocation will
1197 * first attempt to defrag slab caches on other nodes. This means
1198 * scanning over all nodes to look for partial slabs which may be
1199 * expensive if we do it every time we are trying to find a slab
1200 * with available objects.
1202 if (!s
->defrag_ratio
|| get_cycles() % 1024 > s
->defrag_ratio
)
1205 zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
1206 ->node_zonelists
[gfp_zone(flags
)];
1207 for (z
= zonelist
->zones
; *z
; z
++) {
1208 struct kmem_cache_node
*n
;
1210 n
= get_node(s
, zone_to_nid(*z
));
1212 if (n
&& cpuset_zone_allowed_hardwall(*z
, flags
) &&
1213 n
->nr_partial
> MIN_PARTIAL
) {
1214 page
= get_partial_node(n
);
1224 * Get a partial page, lock it and return it.
1226 static struct page
*get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
)
1229 int searchnode
= (node
== -1) ? numa_node_id() : node
;
1231 page
= get_partial_node(get_node(s
, searchnode
));
1232 if (page
|| (flags
& __GFP_THISNODE
))
1235 return get_any_partial(s
, flags
);
1239 * Move a page back to the lists.
1241 * Must be called with the slab lock held.
1243 * On exit the slab lock will have been dropped.
1245 static void putback_slab(struct kmem_cache
*s
, struct page
*page
)
1247 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1252 add_partial(n
, page
);
1253 else if (SlabDebug(page
) && (s
->flags
& SLAB_STORE_USER
))
1258 if (n
->nr_partial
< MIN_PARTIAL
) {
1260 * Adding an empty slab to the partial slabs in order
1261 * to avoid page allocator overhead. This slab needs
1262 * to come after the other slabs with objects in
1263 * order to fill them up. That way the size of the
1264 * partial list stays small. kmem_cache_shrink can
1265 * reclaim empty slabs from the partial list.
1267 add_partial_tail(n
, page
);
1271 discard_slab(s
, page
);
1277 * Remove the cpu slab
1279 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1282 * Merge cpu freelist into freelist. Typically we get here
1283 * because both freelists are empty. So this is unlikely
1286 while (unlikely(page
->lockless_freelist
)) {
1289 /* Retrieve object from cpu_freelist */
1290 object
= page
->lockless_freelist
;
1291 page
->lockless_freelist
= page
->lockless_freelist
[page
->offset
];
1293 /* And put onto the regular freelist */
1294 object
[page
->offset
] = page
->freelist
;
1295 page
->freelist
= object
;
1298 s
->cpu_slab
[cpu
] = NULL
;
1299 ClearPageActive(page
);
1301 putback_slab(s
, page
);
1304 static void flush_slab(struct kmem_cache
*s
, struct page
*page
, int cpu
)
1307 deactivate_slab(s
, page
, cpu
);
1312 * Called from IPI handler with interrupts disabled.
1314 static void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
1316 struct page
*page
= s
->cpu_slab
[cpu
];
1319 flush_slab(s
, page
, cpu
);
1322 static void flush_cpu_slab(void *d
)
1324 struct kmem_cache
*s
= d
;
1325 int cpu
= smp_processor_id();
1327 __flush_cpu_slab(s
, cpu
);
1330 static void flush_all(struct kmem_cache
*s
)
1333 on_each_cpu(flush_cpu_slab
, s
, 1, 1);
1335 unsigned long flags
;
1337 local_irq_save(flags
);
1339 local_irq_restore(flags
);
1344 * Slow path. The lockless freelist is empty or we need to perform
1347 * Interrupts are disabled.
1349 * Processing is still very fast if new objects have been freed to the
1350 * regular freelist. In that case we simply take over the regular freelist
1351 * as the lockless freelist and zap the regular freelist.
1353 * If that is not working then we fall back to the partial lists. We take the
1354 * first element of the freelist as the object to allocate now and move the
1355 * rest of the freelist to the lockless freelist.
1357 * And if we were unable to get a new slab from the partial slab lists then
1358 * we need to allocate a new slab. This is slowest path since we may sleep.
1360 static void *__slab_alloc(struct kmem_cache
*s
,
1361 gfp_t gfpflags
, int node
, void *addr
, struct page
*page
)
1364 int cpu
= smp_processor_id();
1370 if (unlikely(node
!= -1 && page_to_nid(page
) != node
))
1373 object
= page
->freelist
;
1374 if (unlikely(!object
))
1376 if (unlikely(SlabDebug(page
)))
1379 object
= page
->freelist
;
1380 page
->lockless_freelist
= object
[page
->offset
];
1381 page
->inuse
= s
->objects
;
1382 page
->freelist
= NULL
;
1387 deactivate_slab(s
, page
, cpu
);
1390 page
= get_partial(s
, gfpflags
, node
);
1393 s
->cpu_slab
[cpu
] = page
;
1394 SetPageActive(page
);
1398 page
= new_slab(s
, gfpflags
, node
);
1400 cpu
= smp_processor_id();
1401 if (s
->cpu_slab
[cpu
]) {
1403 * Someone else populated the cpu_slab while we
1404 * enabled interrupts, or we have gotten scheduled
1405 * on another cpu. The page may not be on the
1406 * requested node even if __GFP_THISNODE was
1407 * specified. So we need to recheck.
1410 page_to_nid(s
->cpu_slab
[cpu
]) == node
) {
1412 * Current cpuslab is acceptable and we
1413 * want the current one since its cache hot
1415 discard_slab(s
, page
);
1416 page
= s
->cpu_slab
[cpu
];
1420 /* New slab does not fit our expectations */
1421 flush_slab(s
, s
->cpu_slab
[cpu
], cpu
);
1428 object
= page
->freelist
;
1429 if (!alloc_object_checks(s
, page
, object
))
1431 if (s
->flags
& SLAB_STORE_USER
)
1432 set_track(s
, object
, TRACK_ALLOC
, addr
);
1433 trace(s
, page
, object
, 1);
1434 init_object(s
, object
, 1);
1437 page
->freelist
= object
[page
->offset
];
1443 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1444 * have the fastpath folded into their functions. So no function call
1445 * overhead for requests that can be satisfied on the fastpath.
1447 * The fastpath works by first checking if the lockless freelist can be used.
1448 * If not then __slab_alloc is called for slow processing.
1450 * Otherwise we can simply pick the next object from the lockless free list.
1452 static void __always_inline
*slab_alloc(struct kmem_cache
*s
,
1453 gfp_t gfpflags
, int node
, void *addr
)
1457 unsigned long flags
;
1459 local_irq_save(flags
);
1460 page
= s
->cpu_slab
[smp_processor_id()];
1461 if (unlikely(!page
|| !page
->lockless_freelist
||
1462 (node
!= -1 && page_to_nid(page
) != node
)))
1464 object
= __slab_alloc(s
, gfpflags
, node
, addr
, page
);
1467 object
= page
->lockless_freelist
;
1468 page
->lockless_freelist
= object
[page
->offset
];
1470 local_irq_restore(flags
);
1474 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
1476 return slab_alloc(s
, gfpflags
, -1, __builtin_return_address(0));
1478 EXPORT_SYMBOL(kmem_cache_alloc
);
1481 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
1483 return slab_alloc(s
, gfpflags
, node
, __builtin_return_address(0));
1485 EXPORT_SYMBOL(kmem_cache_alloc_node
);
1489 * Slow patch handling. This may still be called frequently since objects
1490 * have a longer lifetime than the cpu slabs in most processing loads.
1492 * So we still attempt to reduce cache line usage. Just take the slab
1493 * lock and free the item. If there is no additional partial page
1494 * handling required then we can return immediately.
1496 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
1497 void *x
, void *addr
)
1500 void **object
= (void *)x
;
1504 if (unlikely(SlabDebug(page
)))
1507 prior
= object
[page
->offset
] = page
->freelist
;
1508 page
->freelist
= object
;
1511 if (unlikely(PageActive(page
)))
1513 * Cpu slabs are never on partial lists and are
1518 if (unlikely(!page
->inuse
))
1522 * Objects left in the slab. If it
1523 * was not on the partial list before
1526 if (unlikely(!prior
))
1527 add_partial(get_node(s
, page_to_nid(page
)), page
);
1536 * Slab still on the partial list.
1538 remove_partial(s
, page
);
1541 discard_slab(s
, page
);
1545 if (!free_object_checks(s
, page
, x
))
1547 if (!PageActive(page
) && !page
->freelist
)
1548 remove_full(s
, page
);
1549 if (s
->flags
& SLAB_STORE_USER
)
1550 set_track(s
, x
, TRACK_FREE
, addr
);
1551 trace(s
, page
, object
, 0);
1552 init_object(s
, object
, 0);
1557 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1558 * can perform fastpath freeing without additional function calls.
1560 * The fastpath is only possible if we are freeing to the current cpu slab
1561 * of this processor. This typically the case if we have just allocated
1564 * If fastpath is not possible then fall back to __slab_free where we deal
1565 * with all sorts of special processing.
1567 static void __always_inline
slab_free(struct kmem_cache
*s
,
1568 struct page
*page
, void *x
, void *addr
)
1570 void **object
= (void *)x
;
1571 unsigned long flags
;
1573 local_irq_save(flags
);
1574 if (likely(page
== s
->cpu_slab
[smp_processor_id()] &&
1575 !SlabDebug(page
))) {
1576 object
[page
->offset
] = page
->lockless_freelist
;
1577 page
->lockless_freelist
= object
;
1579 __slab_free(s
, page
, x
, addr
);
1581 local_irq_restore(flags
);
1584 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
1588 page
= virt_to_head_page(x
);
1590 slab_free(s
, page
, x
, __builtin_return_address(0));
1592 EXPORT_SYMBOL(kmem_cache_free
);
1594 /* Figure out on which slab object the object resides */
1595 static struct page
*get_object_page(const void *x
)
1597 struct page
*page
= virt_to_head_page(x
);
1599 if (!PageSlab(page
))
1606 * Object placement in a slab is made very easy because we always start at
1607 * offset 0. If we tune the size of the object to the alignment then we can
1608 * get the required alignment by putting one properly sized object after
1611 * Notice that the allocation order determines the sizes of the per cpu
1612 * caches. Each processor has always one slab available for allocations.
1613 * Increasing the allocation order reduces the number of times that slabs
1614 * must be moved on and off the partial lists and is therefore a factor in
1619 * Mininum / Maximum order of slab pages. This influences locking overhead
1620 * and slab fragmentation. A higher order reduces the number of partial slabs
1621 * and increases the number of allocations possible without having to
1622 * take the list_lock.
1624 static int slub_min_order
;
1625 static int slub_max_order
= DEFAULT_MAX_ORDER
;
1626 static int slub_min_objects
= DEFAULT_MIN_OBJECTS
;
1629 * Merge control. If this is set then no merging of slab caches will occur.
1630 * (Could be removed. This was introduced to pacify the merge skeptics.)
1632 static int slub_nomerge
;
1635 * Calculate the order of allocation given an slab object size.
1637 * The order of allocation has significant impact on performance and other
1638 * system components. Generally order 0 allocations should be preferred since
1639 * order 0 does not cause fragmentation in the page allocator. Larger objects
1640 * be problematic to put into order 0 slabs because there may be too much
1641 * unused space left. We go to a higher order if more than 1/8th of the slab
1644 * In order to reach satisfactory performance we must ensure that a minimum
1645 * number of objects is in one slab. Otherwise we may generate too much
1646 * activity on the partial lists which requires taking the list_lock. This is
1647 * less a concern for large slabs though which are rarely used.
1649 * slub_max_order specifies the order where we begin to stop considering the
1650 * number of objects in a slab as critical. If we reach slub_max_order then
1651 * we try to keep the page order as low as possible. So we accept more waste
1652 * of space in favor of a small page order.
1654 * Higher order allocations also allow the placement of more objects in a
1655 * slab and thereby reduce object handling overhead. If the user has
1656 * requested a higher mininum order then we start with that one instead of
1657 * the smallest order which will fit the object.
1659 static inline int slab_order(int size
, int min_objects
,
1660 int max_order
, int fract_leftover
)
1665 for (order
= max(slub_min_order
,
1666 fls(min_objects
* size
- 1) - PAGE_SHIFT
);
1667 order
<= max_order
; order
++) {
1669 unsigned long slab_size
= PAGE_SIZE
<< order
;
1671 if (slab_size
< min_objects
* size
)
1674 rem
= slab_size
% size
;
1676 if (rem
<= slab_size
/ fract_leftover
)
1684 static inline int calculate_order(int size
)
1691 * Attempt to find best configuration for a slab. This
1692 * works by first attempting to generate a layout with
1693 * the best configuration and backing off gradually.
1695 * First we reduce the acceptable waste in a slab. Then
1696 * we reduce the minimum objects required in a slab.
1698 min_objects
= slub_min_objects
;
1699 while (min_objects
> 1) {
1701 while (fraction
>= 4) {
1702 order
= slab_order(size
, min_objects
,
1703 slub_max_order
, fraction
);
1704 if (order
<= slub_max_order
)
1712 * We were unable to place multiple objects in a slab. Now
1713 * lets see if we can place a single object there.
1715 order
= slab_order(size
, 1, slub_max_order
, 1);
1716 if (order
<= slub_max_order
)
1720 * Doh this slab cannot be placed using slub_max_order.
1722 order
= slab_order(size
, 1, MAX_ORDER
, 1);
1723 if (order
<= MAX_ORDER
)
1729 * Figure out what the alignment of the objects will be.
1731 static unsigned long calculate_alignment(unsigned long flags
,
1732 unsigned long align
, unsigned long size
)
1735 * If the user wants hardware cache aligned objects then
1736 * follow that suggestion if the object is sufficiently
1739 * The hardware cache alignment cannot override the
1740 * specified alignment though. If that is greater
1743 if ((flags
& SLAB_HWCACHE_ALIGN
) &&
1744 size
> cache_line_size() / 2)
1745 return max_t(unsigned long, align
, cache_line_size());
1747 if (align
< ARCH_SLAB_MINALIGN
)
1748 return ARCH_SLAB_MINALIGN
;
1750 return ALIGN(align
, sizeof(void *));
1753 static void init_kmem_cache_node(struct kmem_cache_node
*n
)
1756 atomic_long_set(&n
->nr_slabs
, 0);
1757 spin_lock_init(&n
->list_lock
);
1758 INIT_LIST_HEAD(&n
->partial
);
1759 INIT_LIST_HEAD(&n
->full
);
1764 * No kmalloc_node yet so do it by hand. We know that this is the first
1765 * slab on the node for this slabcache. There are no concurrent accesses
1768 * Note that this function only works on the kmalloc_node_cache
1769 * when allocating for the kmalloc_node_cache.
1771 static struct kmem_cache_node
* __init
early_kmem_cache_node_alloc(gfp_t gfpflags
,
1775 struct kmem_cache_node
*n
;
1777 BUG_ON(kmalloc_caches
->size
< sizeof(struct kmem_cache_node
));
1779 page
= new_slab(kmalloc_caches
, gfpflags
| GFP_THISNODE
, node
);
1780 /* new_slab() disables interupts */
1786 page
->freelist
= get_freepointer(kmalloc_caches
, n
);
1788 kmalloc_caches
->node
[node
] = n
;
1789 init_object(kmalloc_caches
, n
, 1);
1790 init_kmem_cache_node(n
);
1791 atomic_long_inc(&n
->nr_slabs
);
1792 add_partial(n
, page
);
1796 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1800 for_each_online_node(node
) {
1801 struct kmem_cache_node
*n
= s
->node
[node
];
1802 if (n
&& n
!= &s
->local_node
)
1803 kmem_cache_free(kmalloc_caches
, n
);
1804 s
->node
[node
] = NULL
;
1808 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1813 if (slab_state
>= UP
)
1814 local_node
= page_to_nid(virt_to_page(s
));
1818 for_each_online_node(node
) {
1819 struct kmem_cache_node
*n
;
1821 if (local_node
== node
)
1824 if (slab_state
== DOWN
) {
1825 n
= early_kmem_cache_node_alloc(gfpflags
,
1829 n
= kmem_cache_alloc_node(kmalloc_caches
,
1833 free_kmem_cache_nodes(s
);
1839 init_kmem_cache_node(n
);
1844 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
1848 static int init_kmem_cache_nodes(struct kmem_cache
*s
, gfp_t gfpflags
)
1850 init_kmem_cache_node(&s
->local_node
);
1856 * calculate_sizes() determines the order and the distribution of data within
1859 static int calculate_sizes(struct kmem_cache
*s
)
1861 unsigned long flags
= s
->flags
;
1862 unsigned long size
= s
->objsize
;
1863 unsigned long align
= s
->align
;
1866 * Determine if we can poison the object itself. If the user of
1867 * the slab may touch the object after free or before allocation
1868 * then we should never poison the object itself.
1870 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
1872 s
->flags
|= __OBJECT_POISON
;
1874 s
->flags
&= ~__OBJECT_POISON
;
1877 * Round up object size to the next word boundary. We can only
1878 * place the free pointer at word boundaries and this determines
1879 * the possible location of the free pointer.
1881 size
= ALIGN(size
, sizeof(void *));
1883 #ifdef CONFIG_SLUB_DEBUG
1885 * If we are Redzoning then check if there is some space between the
1886 * end of the object and the free pointer. If not then add an
1887 * additional word to have some bytes to store Redzone information.
1889 if ((flags
& SLAB_RED_ZONE
) && size
== s
->objsize
)
1890 size
+= sizeof(void *);
1894 * With that we have determined the number of bytes in actual use
1895 * by the object. This is the potential offset to the free pointer.
1899 #ifdef CONFIG_SLUB_DEBUG
1900 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
1903 * Relocate free pointer after the object if it is not
1904 * permitted to overwrite the first word of the object on
1907 * This is the case if we do RCU, have a constructor or
1908 * destructor or are poisoning the objects.
1911 size
+= sizeof(void *);
1914 if (flags
& SLAB_STORE_USER
)
1916 * Need to store information about allocs and frees after
1919 size
+= 2 * sizeof(struct track
);
1921 if (flags
& SLAB_RED_ZONE
)
1923 * Add some empty padding so that we can catch
1924 * overwrites from earlier objects rather than let
1925 * tracking information or the free pointer be
1926 * corrupted if an user writes before the start
1929 size
+= sizeof(void *);
1933 * Determine the alignment based on various parameters that the
1934 * user specified and the dynamic determination of cache line size
1937 align
= calculate_alignment(flags
, align
, s
->objsize
);
1940 * SLUB stores one object immediately after another beginning from
1941 * offset 0. In order to align the objects we have to simply size
1942 * each object to conform to the alignment.
1944 size
= ALIGN(size
, align
);
1947 s
->order
= calculate_order(size
);
1952 * Determine the number of objects per slab
1954 s
->objects
= (PAGE_SIZE
<< s
->order
) / size
;
1957 * Verify that the number of objects is within permitted limits.
1958 * The page->inuse field is only 16 bit wide! So we cannot have
1959 * more than 64k objects per slab.
1961 if (!s
->objects
|| s
->objects
> 65535)
1967 static int kmem_cache_open(struct kmem_cache
*s
, gfp_t gfpflags
,
1968 const char *name
, size_t size
,
1969 size_t align
, unsigned long flags
,
1970 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
1972 memset(s
, 0, kmem_size
);
1978 kmem_cache_open_debug_check(s
);
1980 if (!calculate_sizes(s
))
1985 s
->defrag_ratio
= 100;
1988 if (init_kmem_cache_nodes(s
, gfpflags
& ~SLUB_DMA
))
1991 if (flags
& SLAB_PANIC
)
1992 panic("Cannot create slab %s size=%lu realsize=%u "
1993 "order=%u offset=%u flags=%lx\n",
1994 s
->name
, (unsigned long)size
, s
->size
, s
->order
,
1998 EXPORT_SYMBOL(kmem_cache_open
);
2001 * Check if a given pointer is valid
2003 int kmem_ptr_validate(struct kmem_cache
*s
, const void *object
)
2007 page
= get_object_page(object
);
2009 if (!page
|| s
!= page
->slab
)
2010 /* No slab or wrong slab */
2013 if (!check_valid_pointer(s
, page
, object
))
2017 * We could also check if the object is on the slabs freelist.
2018 * But this would be too expensive and it seems that the main
2019 * purpose of kmem_ptr_valid is to check if the object belongs
2020 * to a certain slab.
2024 EXPORT_SYMBOL(kmem_ptr_validate
);
2027 * Determine the size of a slab object
2029 unsigned int kmem_cache_size(struct kmem_cache
*s
)
2033 EXPORT_SYMBOL(kmem_cache_size
);
2035 const char *kmem_cache_name(struct kmem_cache
*s
)
2039 EXPORT_SYMBOL(kmem_cache_name
);
2042 * Attempt to free all slabs on a node. Return the number of slabs we
2043 * were unable to free.
2045 static int free_list(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
2046 struct list_head
*list
)
2048 int slabs_inuse
= 0;
2049 unsigned long flags
;
2050 struct page
*page
, *h
;
2052 spin_lock_irqsave(&n
->list_lock
, flags
);
2053 list_for_each_entry_safe(page
, h
, list
, lru
)
2055 list_del(&page
->lru
);
2056 discard_slab(s
, page
);
2059 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2064 * Release all resources used by a slab cache.
2066 static int kmem_cache_close(struct kmem_cache
*s
)
2072 /* Attempt to free all objects */
2073 for_each_online_node(node
) {
2074 struct kmem_cache_node
*n
= get_node(s
, node
);
2076 n
->nr_partial
-= free_list(s
, n
, &n
->partial
);
2077 if (atomic_long_read(&n
->nr_slabs
))
2080 free_kmem_cache_nodes(s
);
2085 * Close a cache and release the kmem_cache structure
2086 * (must be used for caches created using kmem_cache_create)
2088 void kmem_cache_destroy(struct kmem_cache
*s
)
2090 down_write(&slub_lock
);
2094 if (kmem_cache_close(s
))
2096 sysfs_slab_remove(s
);
2099 up_write(&slub_lock
);
2101 EXPORT_SYMBOL(kmem_cache_destroy
);
2103 /********************************************************************
2105 *******************************************************************/
2107 struct kmem_cache kmalloc_caches
[KMALLOC_SHIFT_HIGH
+ 1] __cacheline_aligned
;
2108 EXPORT_SYMBOL(kmalloc_caches
);
2110 #ifdef CONFIG_ZONE_DMA
2111 static struct kmem_cache
*kmalloc_caches_dma
[KMALLOC_SHIFT_HIGH
+ 1];
2114 static int __init
setup_slub_min_order(char *str
)
2116 get_option (&str
, &slub_min_order
);
2121 __setup("slub_min_order=", setup_slub_min_order
);
2123 static int __init
setup_slub_max_order(char *str
)
2125 get_option (&str
, &slub_max_order
);
2130 __setup("slub_max_order=", setup_slub_max_order
);
2132 static int __init
setup_slub_min_objects(char *str
)
2134 get_option (&str
, &slub_min_objects
);
2139 __setup("slub_min_objects=", setup_slub_min_objects
);
2141 static int __init
setup_slub_nomerge(char *str
)
2147 __setup("slub_nomerge", setup_slub_nomerge
);
2149 static struct kmem_cache
*create_kmalloc_cache(struct kmem_cache
*s
,
2150 const char *name
, int size
, gfp_t gfp_flags
)
2152 unsigned int flags
= 0;
2154 if (gfp_flags
& SLUB_DMA
)
2155 flags
= SLAB_CACHE_DMA
;
2157 down_write(&slub_lock
);
2158 if (!kmem_cache_open(s
, gfp_flags
, name
, size
, ARCH_KMALLOC_MINALIGN
,
2162 list_add(&s
->list
, &slab_caches
);
2163 up_write(&slub_lock
);
2164 if (sysfs_slab_add(s
))
2169 panic("Creation of kmalloc slab %s size=%d failed.\n", name
, size
);
2172 static struct kmem_cache
*get_slab(size_t size
, gfp_t flags
)
2174 int index
= kmalloc_index(size
);
2179 /* Allocation too large? */
2182 #ifdef CONFIG_ZONE_DMA
2183 if ((flags
& SLUB_DMA
)) {
2184 struct kmem_cache
*s
;
2185 struct kmem_cache
*x
;
2189 s
= kmalloc_caches_dma
[index
];
2193 /* Dynamically create dma cache */
2194 x
= kmalloc(kmem_size
, flags
& ~SLUB_DMA
);
2196 panic("Unable to allocate memory for dma cache\n");
2198 if (index
<= KMALLOC_SHIFT_HIGH
)
2199 realsize
= 1 << index
;
2207 text
= kasprintf(flags
& ~SLUB_DMA
, "kmalloc_dma-%d",
2208 (unsigned int)realsize
);
2209 s
= create_kmalloc_cache(x
, text
, realsize
, flags
);
2210 kmalloc_caches_dma
[index
] = s
;
2214 return &kmalloc_caches
[index
];
2217 void *__kmalloc(size_t size
, gfp_t flags
)
2219 struct kmem_cache
*s
= get_slab(size
, flags
);
2222 return slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2225 EXPORT_SYMBOL(__kmalloc
);
2228 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
2230 struct kmem_cache
*s
= get_slab(size
, flags
);
2233 return slab_alloc(s
, flags
, node
, __builtin_return_address(0));
2236 EXPORT_SYMBOL(__kmalloc_node
);
2239 size_t ksize(const void *object
)
2241 struct page
*page
= get_object_page(object
);
2242 struct kmem_cache
*s
;
2249 * Debugging requires use of the padding between object
2250 * and whatever may come after it.
2252 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
2256 * If we have the need to store the freelist pointer
2257 * back there or track user information then we can
2258 * only use the space before that information.
2260 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
2264 * Else we can use all the padding etc for the allocation
2268 EXPORT_SYMBOL(ksize
);
2270 void kfree(const void *x
)
2272 struct kmem_cache
*s
;
2278 page
= virt_to_head_page(x
);
2281 slab_free(s
, page
, (void *)x
, __builtin_return_address(0));
2283 EXPORT_SYMBOL(kfree
);
2286 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2287 * the remaining slabs by the number of items in use. The slabs with the
2288 * most items in use come first. New allocations will then fill those up
2289 * and thus they can be removed from the partial lists.
2291 * The slabs with the least items are placed last. This results in them
2292 * being allocated from last increasing the chance that the last objects
2293 * are freed in them.
2295 int kmem_cache_shrink(struct kmem_cache
*s
)
2299 struct kmem_cache_node
*n
;
2302 struct list_head
*slabs_by_inuse
=
2303 kmalloc(sizeof(struct list_head
) * s
->objects
, GFP_KERNEL
);
2304 unsigned long flags
;
2306 if (!slabs_by_inuse
)
2310 for_each_online_node(node
) {
2311 n
= get_node(s
, node
);
2316 for (i
= 0; i
< s
->objects
; i
++)
2317 INIT_LIST_HEAD(slabs_by_inuse
+ i
);
2319 spin_lock_irqsave(&n
->list_lock
, flags
);
2322 * Build lists indexed by the items in use in each slab.
2324 * Note that concurrent frees may occur while we hold the
2325 * list_lock. page->inuse here is the upper limit.
2327 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
2328 if (!page
->inuse
&& slab_trylock(page
)) {
2330 * Must hold slab lock here because slab_free
2331 * may have freed the last object and be
2332 * waiting to release the slab.
2334 list_del(&page
->lru
);
2337 discard_slab(s
, page
);
2339 if (n
->nr_partial
> MAX_PARTIAL
)
2340 list_move(&page
->lru
,
2341 slabs_by_inuse
+ page
->inuse
);
2345 if (n
->nr_partial
<= MAX_PARTIAL
)
2349 * Rebuild the partial list with the slabs filled up most
2350 * first and the least used slabs at the end.
2352 for (i
= s
->objects
- 1; i
>= 0; i
--)
2353 list_splice(slabs_by_inuse
+ i
, n
->partial
.prev
);
2356 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2359 kfree(slabs_by_inuse
);
2362 EXPORT_SYMBOL(kmem_cache_shrink
);
2365 * krealloc - reallocate memory. The contents will remain unchanged.
2366 * @p: object to reallocate memory for.
2367 * @new_size: how many bytes of memory are required.
2368 * @flags: the type of memory to allocate.
2370 * The contents of the object pointed to are preserved up to the
2371 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2372 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2373 * %NULL pointer, the object pointed to is freed.
2375 void *krealloc(const void *p
, size_t new_size
, gfp_t flags
)
2381 return kmalloc(new_size
, flags
);
2383 if (unlikely(!new_size
)) {
2392 ret
= kmalloc(new_size
, flags
);
2394 memcpy(ret
, p
, min(new_size
, ks
));
2399 EXPORT_SYMBOL(krealloc
);
2401 /********************************************************************
2402 * Basic setup of slabs
2403 *******************************************************************/
2405 void __init
kmem_cache_init(void)
2411 * Must first have the slab cache available for the allocations of the
2412 * struct kmem_cache_node's. There is special bootstrap code in
2413 * kmem_cache_open for slab_state == DOWN.
2415 create_kmalloc_cache(&kmalloc_caches
[0], "kmem_cache_node",
2416 sizeof(struct kmem_cache_node
), GFP_KERNEL
);
2419 /* Able to allocate the per node structures */
2420 slab_state
= PARTIAL
;
2422 /* Caches that are not of the two-to-the-power-of size */
2423 create_kmalloc_cache(&kmalloc_caches
[1],
2424 "kmalloc-96", 96, GFP_KERNEL
);
2425 create_kmalloc_cache(&kmalloc_caches
[2],
2426 "kmalloc-192", 192, GFP_KERNEL
);
2428 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2429 create_kmalloc_cache(&kmalloc_caches
[i
],
2430 "kmalloc", 1 << i
, GFP_KERNEL
);
2434 /* Provide the correct kmalloc names now that the caches are up */
2435 for (i
= KMALLOC_SHIFT_LOW
; i
<= KMALLOC_SHIFT_HIGH
; i
++)
2436 kmalloc_caches
[i
]. name
=
2437 kasprintf(GFP_KERNEL
, "kmalloc-%d", 1 << i
);
2440 register_cpu_notifier(&slab_notifier
);
2443 kmem_size
= offsetof(struct kmem_cache
, cpu_slab
) +
2444 nr_cpu_ids
* sizeof(struct page
*);
2446 printk(KERN_INFO
"SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2447 " Processors=%d, Nodes=%d\n",
2448 KMALLOC_SHIFT_HIGH
, cache_line_size(),
2449 slub_min_order
, slub_max_order
, slub_min_objects
,
2450 nr_cpu_ids
, nr_node_ids
);
2454 * Find a mergeable slab cache
2456 static int slab_unmergeable(struct kmem_cache
*s
)
2458 if (slub_nomerge
|| (s
->flags
& SLUB_NEVER_MERGE
))
2467 static struct kmem_cache
*find_mergeable(size_t size
,
2468 size_t align
, unsigned long flags
,
2469 void (*ctor
)(void *, struct kmem_cache
*, unsigned long))
2471 struct list_head
*h
;
2473 if (slub_nomerge
|| (flags
& SLUB_NEVER_MERGE
))
2479 size
= ALIGN(size
, sizeof(void *));
2480 align
= calculate_alignment(flags
, align
, size
);
2481 size
= ALIGN(size
, align
);
2483 list_for_each(h
, &slab_caches
) {
2484 struct kmem_cache
*s
=
2485 container_of(h
, struct kmem_cache
, list
);
2487 if (slab_unmergeable(s
))
2493 if (((flags
| slub_debug
) & SLUB_MERGE_SAME
) !=
2494 (s
->flags
& SLUB_MERGE_SAME
))
2497 * Check if alignment is compatible.
2498 * Courtesy of Adrian Drzewiecki
2500 if ((s
->size
& ~(align
-1)) != s
->size
)
2503 if (s
->size
- size
>= sizeof(void *))
2511 struct kmem_cache
*kmem_cache_create(const char *name
, size_t size
,
2512 size_t align
, unsigned long flags
,
2513 void (*ctor
)(void *, struct kmem_cache
*, unsigned long),
2514 void (*dtor
)(void *, struct kmem_cache
*, unsigned long))
2516 struct kmem_cache
*s
;
2519 down_write(&slub_lock
);
2520 s
= find_mergeable(size
, align
, flags
, ctor
);
2524 * Adjust the object sizes so that we clear
2525 * the complete object on kzalloc.
2527 s
->objsize
= max(s
->objsize
, (int)size
);
2528 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
2529 if (sysfs_slab_alias(s
, name
))
2532 s
= kmalloc(kmem_size
, GFP_KERNEL
);
2533 if (s
&& kmem_cache_open(s
, GFP_KERNEL
, name
,
2534 size
, align
, flags
, ctor
)) {
2535 if (sysfs_slab_add(s
)) {
2539 list_add(&s
->list
, &slab_caches
);
2543 up_write(&slub_lock
);
2547 up_write(&slub_lock
);
2548 if (flags
& SLAB_PANIC
)
2549 panic("Cannot create slabcache %s\n", name
);
2554 EXPORT_SYMBOL(kmem_cache_create
);
2556 void *kmem_cache_zalloc(struct kmem_cache
*s
, gfp_t flags
)
2560 x
= slab_alloc(s
, flags
, -1, __builtin_return_address(0));
2562 memset(x
, 0, s
->objsize
);
2565 EXPORT_SYMBOL(kmem_cache_zalloc
);
2568 static void for_all_slabs(void (*func
)(struct kmem_cache
*, int), int cpu
)
2570 struct list_head
*h
;
2572 down_read(&slub_lock
);
2573 list_for_each(h
, &slab_caches
) {
2574 struct kmem_cache
*s
=
2575 container_of(h
, struct kmem_cache
, list
);
2579 up_read(&slub_lock
);
2583 * Use the cpu notifier to insure that the cpu slabs are flushed when
2586 static int __cpuinit
slab_cpuup_callback(struct notifier_block
*nfb
,
2587 unsigned long action
, void *hcpu
)
2589 long cpu
= (long)hcpu
;
2592 case CPU_UP_CANCELED
:
2593 case CPU_UP_CANCELED_FROZEN
:
2595 case CPU_DEAD_FROZEN
:
2596 for_all_slabs(__flush_cpu_slab
, cpu
);
2604 static struct notifier_block __cpuinitdata slab_notifier
=
2605 { &slab_cpuup_callback
, NULL
, 0 };
2609 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, void *caller
)
2611 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2616 return slab_alloc(s
, gfpflags
, -1, caller
);
2619 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
2620 int node
, void *caller
)
2622 struct kmem_cache
*s
= get_slab(size
, gfpflags
);
2627 return slab_alloc(s
, gfpflags
, node
, caller
);
2630 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2631 static int validate_slab(struct kmem_cache
*s
, struct page
*page
)
2634 void *addr
= page_address(page
);
2635 DECLARE_BITMAP(map
, s
->objects
);
2637 if (!check_slab(s
, page
) ||
2638 !on_freelist(s
, page
, NULL
))
2641 /* Now we know that a valid freelist exists */
2642 bitmap_zero(map
, s
->objects
);
2644 for_each_free_object(p
, s
, page
->freelist
) {
2645 set_bit(slab_index(p
, s
, addr
), map
);
2646 if (!check_object(s
, page
, p
, 0))
2650 for_each_object(p
, s
, addr
)
2651 if (!test_bit(slab_index(p
, s
, addr
), map
))
2652 if (!check_object(s
, page
, p
, 1))
2657 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
)
2659 if (slab_trylock(page
)) {
2660 validate_slab(s
, page
);
2663 printk(KERN_INFO
"SLUB %s: Skipped busy slab 0x%p\n",
2666 if (s
->flags
& DEBUG_DEFAULT_FLAGS
) {
2667 if (!SlabDebug(page
))
2668 printk(KERN_ERR
"SLUB %s: SlabDebug not set "
2669 "on slab 0x%p\n", s
->name
, page
);
2671 if (SlabDebug(page
))
2672 printk(KERN_ERR
"SLUB %s: SlabDebug set on "
2673 "slab 0x%p\n", s
->name
, page
);
2677 static int validate_slab_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
2679 unsigned long count
= 0;
2681 unsigned long flags
;
2683 spin_lock_irqsave(&n
->list_lock
, flags
);
2685 list_for_each_entry(page
, &n
->partial
, lru
) {
2686 validate_slab_slab(s
, page
);
2689 if (count
!= n
->nr_partial
)
2690 printk(KERN_ERR
"SLUB %s: %ld partial slabs counted but "
2691 "counter=%ld\n", s
->name
, count
, n
->nr_partial
);
2693 if (!(s
->flags
& SLAB_STORE_USER
))
2696 list_for_each_entry(page
, &n
->full
, lru
) {
2697 validate_slab_slab(s
, page
);
2700 if (count
!= atomic_long_read(&n
->nr_slabs
))
2701 printk(KERN_ERR
"SLUB: %s %ld slabs counted but "
2702 "counter=%ld\n", s
->name
, count
,
2703 atomic_long_read(&n
->nr_slabs
));
2706 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2710 static unsigned long validate_slab_cache(struct kmem_cache
*s
)
2713 unsigned long count
= 0;
2716 for_each_online_node(node
) {
2717 struct kmem_cache_node
*n
= get_node(s
, node
);
2719 count
+= validate_slab_node(s
, n
);
2724 #ifdef SLUB_RESILIENCY_TEST
2725 static void resiliency_test(void)
2729 printk(KERN_ERR
"SLUB resiliency testing\n");
2730 printk(KERN_ERR
"-----------------------\n");
2731 printk(KERN_ERR
"A. Corruption after allocation\n");
2733 p
= kzalloc(16, GFP_KERNEL
);
2735 printk(KERN_ERR
"\n1. kmalloc-16: Clobber Redzone/next pointer"
2736 " 0x12->0x%p\n\n", p
+ 16);
2738 validate_slab_cache(kmalloc_caches
+ 4);
2740 /* Hmmm... The next two are dangerous */
2741 p
= kzalloc(32, GFP_KERNEL
);
2742 p
[32 + sizeof(void *)] = 0x34;
2743 printk(KERN_ERR
"\n2. kmalloc-32: Clobber next pointer/next slab"
2744 " 0x34 -> -0x%p\n", p
);
2745 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2747 validate_slab_cache(kmalloc_caches
+ 5);
2748 p
= kzalloc(64, GFP_KERNEL
);
2749 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
2751 printk(KERN_ERR
"\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2753 printk(KERN_ERR
"If allocated object is overwritten then not detectable\n\n");
2754 validate_slab_cache(kmalloc_caches
+ 6);
2756 printk(KERN_ERR
"\nB. Corruption after free\n");
2757 p
= kzalloc(128, GFP_KERNEL
);
2760 printk(KERN_ERR
"1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
2761 validate_slab_cache(kmalloc_caches
+ 7);
2763 p
= kzalloc(256, GFP_KERNEL
);
2766 printk(KERN_ERR
"\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
2767 validate_slab_cache(kmalloc_caches
+ 8);
2769 p
= kzalloc(512, GFP_KERNEL
);
2772 printk(KERN_ERR
"\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
2773 validate_slab_cache(kmalloc_caches
+ 9);
2776 static void resiliency_test(void) {};
2780 * Generate lists of code addresses where slabcache objects are allocated
2785 unsigned long count
;
2798 unsigned long count
;
2799 struct location
*loc
;
2802 static void free_loc_track(struct loc_track
*t
)
2805 free_pages((unsigned long)t
->loc
,
2806 get_order(sizeof(struct location
) * t
->max
));
2809 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
)
2815 max
= PAGE_SIZE
/ sizeof(struct location
);
2817 order
= get_order(sizeof(struct location
) * max
);
2819 l
= (void *)__get_free_pages(GFP_KERNEL
, order
);
2825 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
2833 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
2834 const struct track
*track
)
2836 long start
, end
, pos
;
2839 unsigned long age
= jiffies
- track
->when
;
2845 pos
= start
+ (end
- start
+ 1) / 2;
2848 * There is nothing at "end". If we end up there
2849 * we need to add something to before end.
2854 caddr
= t
->loc
[pos
].addr
;
2855 if (track
->addr
== caddr
) {
2861 if (age
< l
->min_time
)
2863 if (age
> l
->max_time
)
2866 if (track
->pid
< l
->min_pid
)
2867 l
->min_pid
= track
->pid
;
2868 if (track
->pid
> l
->max_pid
)
2869 l
->max_pid
= track
->pid
;
2871 cpu_set(track
->cpu
, l
->cpus
);
2873 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
2877 if (track
->addr
< caddr
)
2884 * Not found. Insert new tracking element.
2886 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
))
2892 (t
->count
- pos
) * sizeof(struct location
));
2895 l
->addr
= track
->addr
;
2899 l
->min_pid
= track
->pid
;
2900 l
->max_pid
= track
->pid
;
2901 cpus_clear(l
->cpus
);
2902 cpu_set(track
->cpu
, l
->cpus
);
2903 nodes_clear(l
->nodes
);
2904 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
2908 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
2909 struct page
*page
, enum track_item alloc
)
2911 void *addr
= page_address(page
);
2912 DECLARE_BITMAP(map
, s
->objects
);
2915 bitmap_zero(map
, s
->objects
);
2916 for_each_free_object(p
, s
, page
->freelist
)
2917 set_bit(slab_index(p
, s
, addr
), map
);
2919 for_each_object(p
, s
, addr
)
2920 if (!test_bit(slab_index(p
, s
, addr
), map
))
2921 add_location(t
, s
, get_track(s
, p
, alloc
));
2924 static int list_locations(struct kmem_cache
*s
, char *buf
,
2925 enum track_item alloc
)
2935 /* Push back cpu slabs */
2938 for_each_online_node(node
) {
2939 struct kmem_cache_node
*n
= get_node(s
, node
);
2940 unsigned long flags
;
2943 if (!atomic_read(&n
->nr_slabs
))
2946 spin_lock_irqsave(&n
->list_lock
, flags
);
2947 list_for_each_entry(page
, &n
->partial
, lru
)
2948 process_slab(&t
, s
, page
, alloc
);
2949 list_for_each_entry(page
, &n
->full
, lru
)
2950 process_slab(&t
, s
, page
, alloc
);
2951 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2954 for (i
= 0; i
< t
.count
; i
++) {
2955 struct location
*l
= &t
.loc
[i
];
2957 if (n
> PAGE_SIZE
- 100)
2959 n
+= sprintf(buf
+ n
, "%7ld ", l
->count
);
2962 n
+= sprint_symbol(buf
+ n
, (unsigned long)l
->addr
);
2964 n
+= sprintf(buf
+ n
, "<not-available>");
2966 if (l
->sum_time
!= l
->min_time
) {
2967 unsigned long remainder
;
2969 n
+= sprintf(buf
+ n
, " age=%ld/%ld/%ld",
2971 div_long_long_rem(l
->sum_time
, l
->count
, &remainder
),
2974 n
+= sprintf(buf
+ n
, " age=%ld",
2977 if (l
->min_pid
!= l
->max_pid
)
2978 n
+= sprintf(buf
+ n
, " pid=%ld-%ld",
2979 l
->min_pid
, l
->max_pid
);
2981 n
+= sprintf(buf
+ n
, " pid=%ld",
2984 if (num_online_cpus() > 1 && !cpus_empty(l
->cpus
)) {
2985 n
+= sprintf(buf
+ n
, " cpus=");
2986 n
+= cpulist_scnprintf(buf
+ n
, PAGE_SIZE
- n
- 50,
2990 if (num_online_nodes() > 1 && !nodes_empty(l
->nodes
)) {
2991 n
+= sprintf(buf
+ n
, " nodes=");
2992 n
+= nodelist_scnprintf(buf
+ n
, PAGE_SIZE
- n
- 50,
2996 n
+= sprintf(buf
+ n
, "\n");
3001 n
+= sprintf(buf
, "No data\n");
3005 static unsigned long count_partial(struct kmem_cache_node
*n
)
3007 unsigned long flags
;
3008 unsigned long x
= 0;
3011 spin_lock_irqsave(&n
->list_lock
, flags
);
3012 list_for_each_entry(page
, &n
->partial
, lru
)
3014 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3018 enum slab_stat_type
{
3025 #define SO_FULL (1 << SL_FULL)
3026 #define SO_PARTIAL (1 << SL_PARTIAL)
3027 #define SO_CPU (1 << SL_CPU)
3028 #define SO_OBJECTS (1 << SL_OBJECTS)
3030 static unsigned long slab_objects(struct kmem_cache
*s
,
3031 char *buf
, unsigned long flags
)
3033 unsigned long total
= 0;
3037 unsigned long *nodes
;
3038 unsigned long *per_cpu
;
3040 nodes
= kzalloc(2 * sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
3041 per_cpu
= nodes
+ nr_node_ids
;
3043 for_each_possible_cpu(cpu
) {
3044 struct page
*page
= s
->cpu_slab
[cpu
];
3048 node
= page_to_nid(page
);
3049 if (flags
& SO_CPU
) {
3052 if (flags
& SO_OBJECTS
)
3063 for_each_online_node(node
) {
3064 struct kmem_cache_node
*n
= get_node(s
, node
);
3066 if (flags
& SO_PARTIAL
) {
3067 if (flags
& SO_OBJECTS
)
3068 x
= count_partial(n
);
3075 if (flags
& SO_FULL
) {
3076 int full_slabs
= atomic_read(&n
->nr_slabs
)
3080 if (flags
& SO_OBJECTS
)
3081 x
= full_slabs
* s
->objects
;
3089 x
= sprintf(buf
, "%lu", total
);
3091 for_each_online_node(node
)
3093 x
+= sprintf(buf
+ x
, " N%d=%lu",
3097 return x
+ sprintf(buf
+ x
, "\n");
3100 static int any_slab_objects(struct kmem_cache
*s
)
3105 for_each_possible_cpu(cpu
)
3106 if (s
->cpu_slab
[cpu
])
3109 for_each_node(node
) {
3110 struct kmem_cache_node
*n
= get_node(s
, node
);
3112 if (n
->nr_partial
|| atomic_read(&n
->nr_slabs
))
3118 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3119 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3121 struct slab_attribute
{
3122 struct attribute attr
;
3123 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
3124 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
3127 #define SLAB_ATTR_RO(_name) \
3128 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3130 #define SLAB_ATTR(_name) \
3131 static struct slab_attribute _name##_attr = \
3132 __ATTR(_name, 0644, _name##_show, _name##_store)
3134 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
3136 return sprintf(buf
, "%d\n", s
->size
);
3138 SLAB_ATTR_RO(slab_size
);
3140 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
3142 return sprintf(buf
, "%d\n", s
->align
);
3144 SLAB_ATTR_RO(align
);
3146 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
3148 return sprintf(buf
, "%d\n", s
->objsize
);
3150 SLAB_ATTR_RO(object_size
);
3152 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
3154 return sprintf(buf
, "%d\n", s
->objects
);
3156 SLAB_ATTR_RO(objs_per_slab
);
3158 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
3160 return sprintf(buf
, "%d\n", s
->order
);
3162 SLAB_ATTR_RO(order
);
3164 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
3167 int n
= sprint_symbol(buf
, (unsigned long)s
->ctor
);
3169 return n
+ sprintf(buf
+ n
, "\n");
3175 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
3177 return sprintf(buf
, "%d\n", s
->refcount
- 1);
3179 SLAB_ATTR_RO(aliases
);
3181 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
3183 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
);
3185 SLAB_ATTR_RO(slabs
);
3187 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
3189 return slab_objects(s
, buf
, SO_PARTIAL
);
3191 SLAB_ATTR_RO(partial
);
3193 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
3195 return slab_objects(s
, buf
, SO_CPU
);
3197 SLAB_ATTR_RO(cpu_slabs
);
3199 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
3201 return slab_objects(s
, buf
, SO_FULL
|SO_PARTIAL
|SO_CPU
|SO_OBJECTS
);
3203 SLAB_ATTR_RO(objects
);
3205 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
3207 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
3210 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
3211 const char *buf
, size_t length
)
3213 s
->flags
&= ~SLAB_DEBUG_FREE
;
3215 s
->flags
|= SLAB_DEBUG_FREE
;
3218 SLAB_ATTR(sanity_checks
);
3220 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
3222 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
3225 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
3228 s
->flags
&= ~SLAB_TRACE
;
3230 s
->flags
|= SLAB_TRACE
;
3235 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
3237 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
3240 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
3241 const char *buf
, size_t length
)
3243 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
3245 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
3248 SLAB_ATTR(reclaim_account
);
3250 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
3252 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
3254 SLAB_ATTR_RO(hwcache_align
);
3256 #ifdef CONFIG_ZONE_DMA
3257 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
3259 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
3261 SLAB_ATTR_RO(cache_dma
);
3264 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
3266 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
3268 SLAB_ATTR_RO(destroy_by_rcu
);
3270 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
3272 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
3275 static ssize_t
red_zone_store(struct kmem_cache
*s
,
3276 const char *buf
, size_t length
)
3278 if (any_slab_objects(s
))
3281 s
->flags
&= ~SLAB_RED_ZONE
;
3283 s
->flags
|= SLAB_RED_ZONE
;
3287 SLAB_ATTR(red_zone
);
3289 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
3291 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
3294 static ssize_t
poison_store(struct kmem_cache
*s
,
3295 const char *buf
, size_t length
)
3297 if (any_slab_objects(s
))
3300 s
->flags
&= ~SLAB_POISON
;
3302 s
->flags
|= SLAB_POISON
;
3308 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
3310 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
3313 static ssize_t
store_user_store(struct kmem_cache
*s
,
3314 const char *buf
, size_t length
)
3316 if (any_slab_objects(s
))
3319 s
->flags
&= ~SLAB_STORE_USER
;
3321 s
->flags
|= SLAB_STORE_USER
;
3325 SLAB_ATTR(store_user
);
3327 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
3332 static ssize_t
validate_store(struct kmem_cache
*s
,
3333 const char *buf
, size_t length
)
3336 validate_slab_cache(s
);
3341 SLAB_ATTR(validate
);
3343 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
3348 static ssize_t
shrink_store(struct kmem_cache
*s
,
3349 const char *buf
, size_t length
)
3351 if (buf
[0] == '1') {
3352 int rc
= kmem_cache_shrink(s
);
3362 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
3364 if (!(s
->flags
& SLAB_STORE_USER
))
3366 return list_locations(s
, buf
, TRACK_ALLOC
);
3368 SLAB_ATTR_RO(alloc_calls
);
3370 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
3372 if (!(s
->flags
& SLAB_STORE_USER
))
3374 return list_locations(s
, buf
, TRACK_FREE
);
3376 SLAB_ATTR_RO(free_calls
);
3379 static ssize_t
defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
3381 return sprintf(buf
, "%d\n", s
->defrag_ratio
/ 10);
3384 static ssize_t
defrag_ratio_store(struct kmem_cache
*s
,
3385 const char *buf
, size_t length
)
3387 int n
= simple_strtoul(buf
, NULL
, 10);
3390 s
->defrag_ratio
= n
* 10;
3393 SLAB_ATTR(defrag_ratio
);
3396 static struct attribute
* slab_attrs
[] = {
3397 &slab_size_attr
.attr
,
3398 &object_size_attr
.attr
,
3399 &objs_per_slab_attr
.attr
,
3404 &cpu_slabs_attr
.attr
,
3408 &sanity_checks_attr
.attr
,
3410 &hwcache_align_attr
.attr
,
3411 &reclaim_account_attr
.attr
,
3412 &destroy_by_rcu_attr
.attr
,
3413 &red_zone_attr
.attr
,
3415 &store_user_attr
.attr
,
3416 &validate_attr
.attr
,
3418 &alloc_calls_attr
.attr
,
3419 &free_calls_attr
.attr
,
3420 #ifdef CONFIG_ZONE_DMA
3421 &cache_dma_attr
.attr
,
3424 &defrag_ratio_attr
.attr
,
3429 static struct attribute_group slab_attr_group
= {
3430 .attrs
= slab_attrs
,
3433 static ssize_t
slab_attr_show(struct kobject
*kobj
,
3434 struct attribute
*attr
,
3437 struct slab_attribute
*attribute
;
3438 struct kmem_cache
*s
;
3441 attribute
= to_slab_attr(attr
);
3444 if (!attribute
->show
)
3447 err
= attribute
->show(s
, buf
);
3452 static ssize_t
slab_attr_store(struct kobject
*kobj
,
3453 struct attribute
*attr
,
3454 const char *buf
, size_t len
)
3456 struct slab_attribute
*attribute
;
3457 struct kmem_cache
*s
;
3460 attribute
= to_slab_attr(attr
);
3463 if (!attribute
->store
)
3466 err
= attribute
->store(s
, buf
, len
);
3471 static struct sysfs_ops slab_sysfs_ops
= {
3472 .show
= slab_attr_show
,
3473 .store
= slab_attr_store
,
3476 static struct kobj_type slab_ktype
= {
3477 .sysfs_ops
= &slab_sysfs_ops
,
3480 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
3482 struct kobj_type
*ktype
= get_ktype(kobj
);
3484 if (ktype
== &slab_ktype
)
3489 static struct kset_uevent_ops slab_uevent_ops
= {
3490 .filter
= uevent_filter
,
3493 decl_subsys(slab
, &slab_ktype
, &slab_uevent_ops
);
3495 #define ID_STR_LENGTH 64
3497 /* Create a unique string id for a slab cache:
3499 * :[flags-]size:[memory address of kmemcache]
3501 static char *create_unique_id(struct kmem_cache
*s
)
3503 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
3510 * First flags affecting slabcache operations. We will only
3511 * get here for aliasable slabs so we do not need to support
3512 * too many flags. The flags here must cover all flags that
3513 * are matched during merging to guarantee that the id is
3516 if (s
->flags
& SLAB_CACHE_DMA
)
3518 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3520 if (s
->flags
& SLAB_DEBUG_FREE
)
3524 p
+= sprintf(p
, "%07d", s
->size
);
3525 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
3529 static int sysfs_slab_add(struct kmem_cache
*s
)
3535 if (slab_state
< SYSFS
)
3536 /* Defer until later */
3539 unmergeable
= slab_unmergeable(s
);
3542 * Slabcache can never be merged so we can use the name proper.
3543 * This is typically the case for debug situations. In that
3544 * case we can catch duplicate names easily.
3546 sysfs_remove_link(&slab_subsys
.kobj
, s
->name
);
3550 * Create a unique name for the slab as a target
3553 name
= create_unique_id(s
);
3556 kobj_set_kset_s(s
, slab_subsys
);
3557 kobject_set_name(&s
->kobj
, name
);
3558 kobject_init(&s
->kobj
);
3559 err
= kobject_add(&s
->kobj
);
3563 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
3566 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
3568 /* Setup first alias */
3569 sysfs_slab_alias(s
, s
->name
);
3575 static void sysfs_slab_remove(struct kmem_cache
*s
)
3577 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
3578 kobject_del(&s
->kobj
);
3582 * Need to buffer aliases during bootup until sysfs becomes
3583 * available lest we loose that information.
3585 struct saved_alias
{
3586 struct kmem_cache
*s
;
3588 struct saved_alias
*next
;
3591 struct saved_alias
*alias_list
;
3593 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
3595 struct saved_alias
*al
;
3597 if (slab_state
== SYSFS
) {
3599 * If we have a leftover link then remove it.
3601 sysfs_remove_link(&slab_subsys
.kobj
, name
);
3602 return sysfs_create_link(&slab_subsys
.kobj
,
3606 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
3612 al
->next
= alias_list
;
3617 static int __init
slab_sysfs_init(void)
3619 struct list_head
*h
;
3622 err
= subsystem_register(&slab_subsys
);
3624 printk(KERN_ERR
"Cannot register slab subsystem.\n");
3630 list_for_each(h
, &slab_caches
) {
3631 struct kmem_cache
*s
=
3632 container_of(h
, struct kmem_cache
, list
);
3634 err
= sysfs_slab_add(s
);
3638 while (alias_list
) {
3639 struct saved_alias
*al
= alias_list
;
3641 alias_list
= alias_list
->next
;
3642 err
= sysfs_slab_alias(al
->s
, al
->name
);
3651 __initcall(slab_sysfs_init
);