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 or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache
*s
)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s
);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier
;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr
; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
194 int cpu
; /* Was running on cpu */
195 int pid
; /* Pid context */
196 unsigned long when
; /* When did the operation occur */
199 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
202 static int sysfs_slab_add(struct kmem_cache
*);
203 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
206 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
212 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache
*s
,
229 struct page
*page
, const void *object
)
236 base
= page_address(page
);
237 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
238 (object
- base
) % s
->size
) {
245 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
247 return *(void **)(object
+ s
->offset
);
250 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
252 prefetch(object
+ s
->offset
);
255 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p
, (void **)(object
+ s
->offset
), sizeof(p
));
262 p
= get_freepointer(s
, object
);
267 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
269 *(void **)(object
+ s
->offset
) = fp
;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
284 return (p
- addr
) / s
->size
;
287 static inline size_t slab_ksize(const struct kmem_cache
*s
)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s
->flags
& (SLAB_RED_ZONE
| SLAB_POISON
))
295 return s
->object_size
;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s
->flags
& (SLAB_DESTROY_BY_RCU
| SLAB_STORE_USER
))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order
, unsigned long size
, int reserved
)
313 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
316 static inline struct kmem_cache_order_objects
oo_make(int order
,
317 unsigned long size
, int reserved
)
319 struct kmem_cache_order_objects x
= {
320 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
326 static inline int oo_order(struct kmem_cache_order_objects x
)
328 return x
.x
>> OO_SHIFT
;
331 static inline int oo_objects(struct kmem_cache_order_objects x
)
333 return x
.x
& OO_MASK
;
337 * Per slab locking using the pagelock
339 static __always_inline
void slab_lock(struct page
*page
)
341 bit_spin_lock(PG_locked
, &page
->flags
);
344 static __always_inline
void slab_unlock(struct page
*page
)
346 __bit_spin_unlock(PG_locked
, &page
->flags
);
349 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
352 tmp
.counters
= counters_new
;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page
->frozen
= tmp
.frozen
;
360 page
->inuse
= tmp
.inuse
;
361 page
->objects
= tmp
.objects
;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
366 void *freelist_old
, unsigned long counters_old
,
367 void *freelist_new
, unsigned long counters_new
,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s
->flags
& __CMPXCHG_DOUBLE
) {
374 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
375 freelist_old
, counters_old
,
376 freelist_new
, counters_new
))
382 if (page
->freelist
== freelist_old
&&
383 page
->counters
== counters_old
) {
384 page
->freelist
= freelist_new
;
385 set_page_slub_counters(page
, counters_new
);
393 stat(s
, CMPXCHG_DOUBLE_FAIL
);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
402 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
403 void *freelist_old
, unsigned long counters_old
,
404 void *freelist_new
, unsigned long counters_new
,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s
->flags
& __CMPXCHG_DOUBLE
) {
410 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
411 freelist_old
, counters_old
,
412 freelist_new
, counters_new
))
419 local_irq_save(flags
);
421 if (page
->freelist
== freelist_old
&&
422 page
->counters
== counters_old
) {
423 page
->freelist
= freelist_new
;
424 set_page_slub_counters(page
, counters_new
);
426 local_irq_restore(flags
);
430 local_irq_restore(flags
);
434 stat(s
, CMPXCHG_DOUBLE_FAIL
);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
453 void *addr
= page_address(page
);
455 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
456 set_bit(slab_index(p
, s
, addr
), map
);
462 #if defined(CONFIG_SLUB_DEBUG_ON)
463 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
464 #elif defined(CONFIG_KASAN)
465 static int slub_debug
= SLAB_STORE_USER
;
467 static int slub_debug
;
470 static char *slub_debug_slabs
;
471 static int disable_higher_order_debug
;
474 * slub is about to manipulate internal object metadata. This memory lies
475 * outside the range of the allocated object, so accessing it would normally
476 * be reported by kasan as a bounds error. metadata_access_enable() is used
477 * to tell kasan that these accesses are OK.
479 static inline void metadata_access_enable(void)
481 kasan_disable_current();
484 static inline void metadata_access_disable(void)
486 kasan_enable_current();
492 static void print_section(char *text
, u8
*addr
, unsigned int length
)
494 metadata_access_enable();
495 print_hex_dump(KERN_ERR
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
497 metadata_access_disable();
500 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
501 enum track_item alloc
)
506 p
= object
+ s
->offset
+ sizeof(void *);
508 p
= object
+ s
->inuse
;
513 static void set_track(struct kmem_cache
*s
, void *object
,
514 enum track_item alloc
, unsigned long addr
)
516 struct track
*p
= get_track(s
, object
, alloc
);
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace
;
523 trace
.nr_entries
= 0;
524 trace
.max_entries
= TRACK_ADDRS_COUNT
;
525 trace
.entries
= p
->addrs
;
527 metadata_access_enable();
528 save_stack_trace(&trace
);
529 metadata_access_disable();
531 /* See rant in lockdep.c */
532 if (trace
.nr_entries
!= 0 &&
533 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
536 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
540 p
->cpu
= smp_processor_id();
541 p
->pid
= current
->pid
;
544 memset(p
, 0, sizeof(struct track
));
547 static void init_tracking(struct kmem_cache
*s
, void *object
)
549 if (!(s
->flags
& SLAB_STORE_USER
))
552 set_track(s
, object
, TRACK_FREE
, 0UL);
553 set_track(s
, object
, TRACK_ALLOC
, 0UL);
556 static void print_track(const char *s
, struct track
*t
)
561 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
562 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
563 #ifdef CONFIG_STACKTRACE
566 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
568 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
575 static void print_tracking(struct kmem_cache
*s
, void *object
)
577 if (!(s
->flags
& SLAB_STORE_USER
))
580 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
581 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
584 static void print_page_info(struct page
*page
)
586 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
587 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
591 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
593 struct va_format vaf
;
599 pr_err("=============================================================================\n");
600 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
601 pr_err("-----------------------------------------------------------------------------\n\n");
603 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
607 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
609 struct va_format vaf
;
615 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
619 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
621 unsigned int off
; /* Offset of last byte */
622 u8
*addr
= page_address(page
);
624 print_tracking(s
, p
);
626 print_page_info(page
);
628 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
629 p
, p
- addr
, get_freepointer(s
, p
));
632 print_section("Bytes b4 ", p
- 16, 16);
634 print_section("Object ", p
, min_t(unsigned long, s
->object_size
,
636 if (s
->flags
& SLAB_RED_ZONE
)
637 print_section("Redzone ", p
+ s
->object_size
,
638 s
->inuse
- s
->object_size
);
641 off
= s
->offset
+ sizeof(void *);
645 if (s
->flags
& SLAB_STORE_USER
)
646 off
+= 2 * sizeof(struct track
);
649 /* Beginning of the filler is the free pointer */
650 print_section("Padding ", p
+ off
, s
->size
- off
);
655 void object_err(struct kmem_cache
*s
, struct page
*page
,
656 u8
*object
, char *reason
)
658 slab_bug(s
, "%s", reason
);
659 print_trailer(s
, page
, object
);
662 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
663 const char *fmt
, ...)
669 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
671 slab_bug(s
, "%s", buf
);
672 print_page_info(page
);
676 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
680 if (s
->flags
& __OBJECT_POISON
) {
681 memset(p
, POISON_FREE
, s
->object_size
- 1);
682 p
[s
->object_size
- 1] = POISON_END
;
685 if (s
->flags
& SLAB_RED_ZONE
)
686 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
689 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
690 void *from
, void *to
)
692 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
693 memset(from
, data
, to
- from
);
696 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
697 u8
*object
, char *what
,
698 u8
*start
, unsigned int value
, unsigned int bytes
)
703 metadata_access_enable();
704 fault
= memchr_inv(start
, value
, bytes
);
705 metadata_access_disable();
710 while (end
> fault
&& end
[-1] == value
)
713 slab_bug(s
, "%s overwritten", what
);
714 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
715 fault
, end
- 1, fault
[0], value
);
716 print_trailer(s
, page
, object
);
718 restore_bytes(s
, what
, value
, fault
, end
);
726 * Bytes of the object to be managed.
727 * If the freepointer may overlay the object then the free
728 * pointer is the first word of the object.
730 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
733 * object + s->object_size
734 * Padding to reach word boundary. This is also used for Redzoning.
735 * Padding is extended by another word if Redzoning is enabled and
736 * object_size == inuse.
738 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
739 * 0xcc (RED_ACTIVE) for objects in use.
742 * Meta data starts here.
744 * A. Free pointer (if we cannot overwrite object on free)
745 * B. Tracking data for SLAB_STORE_USER
746 * C. Padding to reach required alignment boundary or at mininum
747 * one word if debugging is on to be able to detect writes
748 * before the word boundary.
750 * Padding is done using 0x5a (POISON_INUSE)
753 * Nothing is used beyond s->size.
755 * If slabcaches are merged then the object_size and inuse boundaries are mostly
756 * ignored. And therefore no slab options that rely on these boundaries
757 * may be used with merged slabcaches.
760 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
762 unsigned long off
= s
->inuse
; /* The end of info */
765 /* Freepointer is placed after the object. */
766 off
+= sizeof(void *);
768 if (s
->flags
& SLAB_STORE_USER
)
769 /* We also have user information there */
770 off
+= 2 * sizeof(struct track
);
775 return check_bytes_and_report(s
, page
, p
, "Object padding",
776 p
+ off
, POISON_INUSE
, s
->size
- off
);
779 /* Check the pad bytes at the end of a slab page */
780 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
788 if (!(s
->flags
& SLAB_POISON
))
791 start
= page_address(page
);
792 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
793 end
= start
+ length
;
794 remainder
= length
% s
->size
;
798 metadata_access_enable();
799 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
800 metadata_access_disable();
803 while (end
> fault
&& end
[-1] == POISON_INUSE
)
806 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
807 print_section("Padding ", end
- remainder
, remainder
);
809 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
813 static int check_object(struct kmem_cache
*s
, struct page
*page
,
814 void *object
, u8 val
)
817 u8
*endobject
= object
+ s
->object_size
;
819 if (s
->flags
& SLAB_RED_ZONE
) {
820 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
821 endobject
, val
, s
->inuse
- s
->object_size
))
824 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
825 check_bytes_and_report(s
, page
, p
, "Alignment padding",
826 endobject
, POISON_INUSE
,
827 s
->inuse
- s
->object_size
);
831 if (s
->flags
& SLAB_POISON
) {
832 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
833 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
834 POISON_FREE
, s
->object_size
- 1) ||
835 !check_bytes_and_report(s
, page
, p
, "Poison",
836 p
+ s
->object_size
- 1, POISON_END
, 1)))
839 * check_pad_bytes cleans up on its own.
841 check_pad_bytes(s
, page
, p
);
844 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
846 * Object and freepointer overlap. Cannot check
847 * freepointer while object is allocated.
851 /* Check free pointer validity */
852 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
853 object_err(s
, page
, p
, "Freepointer corrupt");
855 * No choice but to zap it and thus lose the remainder
856 * of the free objects in this slab. May cause
857 * another error because the object count is now wrong.
859 set_freepointer(s
, p
, NULL
);
865 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
869 VM_BUG_ON(!irqs_disabled());
871 if (!PageSlab(page
)) {
872 slab_err(s
, page
, "Not a valid slab page");
876 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
877 if (page
->objects
> maxobj
) {
878 slab_err(s
, page
, "objects %u > max %u",
879 page
->objects
, maxobj
);
882 if (page
->inuse
> page
->objects
) {
883 slab_err(s
, page
, "inuse %u > max %u",
884 page
->inuse
, page
->objects
);
887 /* Slab_pad_check fixes things up after itself */
888 slab_pad_check(s
, page
);
893 * Determine if a certain object on a page is on the freelist. Must hold the
894 * slab lock to guarantee that the chains are in a consistent state.
896 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
904 while (fp
&& nr
<= page
->objects
) {
907 if (!check_valid_pointer(s
, page
, fp
)) {
909 object_err(s
, page
, object
,
910 "Freechain corrupt");
911 set_freepointer(s
, object
, NULL
);
913 slab_err(s
, page
, "Freepointer corrupt");
914 page
->freelist
= NULL
;
915 page
->inuse
= page
->objects
;
916 slab_fix(s
, "Freelist cleared");
922 fp
= get_freepointer(s
, object
);
926 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
927 if (max_objects
> MAX_OBJS_PER_PAGE
)
928 max_objects
= MAX_OBJS_PER_PAGE
;
930 if (page
->objects
!= max_objects
) {
931 slab_err(s
, page
, "Wrong number of objects. Found %d but "
932 "should be %d", page
->objects
, max_objects
);
933 page
->objects
= max_objects
;
934 slab_fix(s
, "Number of objects adjusted.");
936 if (page
->inuse
!= page
->objects
- nr
) {
937 slab_err(s
, page
, "Wrong object count. Counter is %d but "
938 "counted were %d", page
->inuse
, page
->objects
- nr
);
939 page
->inuse
= page
->objects
- nr
;
940 slab_fix(s
, "Object count adjusted.");
942 return search
== NULL
;
945 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
948 if (s
->flags
& SLAB_TRACE
) {
949 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
951 alloc
? "alloc" : "free",
956 print_section("Object ", (void *)object
,
964 * Tracking of fully allocated slabs for debugging purposes.
966 static void add_full(struct kmem_cache
*s
,
967 struct kmem_cache_node
*n
, struct page
*page
)
969 if (!(s
->flags
& SLAB_STORE_USER
))
972 lockdep_assert_held(&n
->list_lock
);
973 list_add(&page
->lru
, &n
->full
);
976 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
978 if (!(s
->flags
& SLAB_STORE_USER
))
981 lockdep_assert_held(&n
->list_lock
);
982 list_del(&page
->lru
);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
988 struct kmem_cache_node
*n
= get_node(s
, node
);
990 return atomic_long_read(&n
->nr_slabs
);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
995 return atomic_long_read(&n
->nr_slabs
);
998 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1000 struct kmem_cache_node
*n
= get_node(s
, node
);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n
->nr_slabs
);
1010 atomic_long_add(objects
, &n
->total_objects
);
1013 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1015 struct kmem_cache_node
*n
= get_node(s
, node
);
1017 atomic_long_dec(&n
->nr_slabs
);
1018 atomic_long_sub(objects
, &n
->total_objects
);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1025 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1028 init_object(s
, object
, SLUB_RED_INACTIVE
);
1029 init_tracking(s
, object
);
1032 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1034 void *object
, unsigned long addr
)
1036 if (!check_slab(s
, page
))
1039 if (!check_valid_pointer(s
, page
, object
)) {
1040 object_err(s
, page
, object
, "Freelist Pointer check fails");
1044 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1047 /* Success perform special debug activities for allocs */
1048 if (s
->flags
& SLAB_STORE_USER
)
1049 set_track(s
, object
, TRACK_ALLOC
, addr
);
1050 trace(s
, page
, object
, 1);
1051 init_object(s
, object
, SLUB_RED_ACTIVE
);
1055 if (PageSlab(page
)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s
, "Marking all objects used");
1062 page
->inuse
= page
->objects
;
1063 page
->freelist
= NULL
;
1068 static noinline
struct kmem_cache_node
*free_debug_processing(
1069 struct kmem_cache
*s
, struct page
*page
, void *object
,
1070 unsigned long addr
, unsigned long *flags
)
1072 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1074 spin_lock_irqsave(&n
->list_lock
, *flags
);
1077 if (!check_slab(s
, page
))
1080 if (!check_valid_pointer(s
, page
, object
)) {
1081 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1085 if (on_freelist(s
, page
, object
)) {
1086 object_err(s
, page
, object
, "Object already free");
1090 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1093 if (unlikely(s
!= page
->slab_cache
)) {
1094 if (!PageSlab(page
)) {
1095 slab_err(s
, page
, "Attempt to free object(0x%p) "
1096 "outside of slab", object
);
1097 } else if (!page
->slab_cache
) {
1098 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1102 object_err(s
, page
, object
,
1103 "page slab pointer corrupt.");
1107 if (s
->flags
& SLAB_STORE_USER
)
1108 set_track(s
, object
, TRACK_FREE
, addr
);
1109 trace(s
, page
, object
, 0);
1110 init_object(s
, object
, SLUB_RED_INACTIVE
);
1114 * Keep node_lock to preserve integrity
1115 * until the object is actually freed
1121 spin_unlock_irqrestore(&n
->list_lock
, *flags
);
1122 slab_fix(s
, "Object at 0x%p not freed", object
);
1126 static int __init
setup_slub_debug(char *str
)
1128 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1129 if (*str
++ != '=' || !*str
)
1131 * No options specified. Switch on full debugging.
1137 * No options but restriction on slabs. This means full
1138 * debugging for slabs matching a pattern.
1145 * Switch off all debugging measures.
1150 * Determine which debug features should be switched on
1152 for (; *str
&& *str
!= ','; str
++) {
1153 switch (tolower(*str
)) {
1155 slub_debug
|= SLAB_DEBUG_FREE
;
1158 slub_debug
|= SLAB_RED_ZONE
;
1161 slub_debug
|= SLAB_POISON
;
1164 slub_debug
|= SLAB_STORE_USER
;
1167 slub_debug
|= SLAB_TRACE
;
1170 slub_debug
|= SLAB_FAILSLAB
;
1174 * Avoid enabling debugging on caches if its minimum
1175 * order would increase as a result.
1177 disable_higher_order_debug
= 1;
1180 pr_err("slub_debug option '%c' unknown. skipped\n",
1187 slub_debug_slabs
= str
+ 1;
1192 __setup("slub_debug", setup_slub_debug
);
1194 unsigned long kmem_cache_flags(unsigned long object_size
,
1195 unsigned long flags
, const char *name
,
1196 void (*ctor
)(void *))
1199 * Enable debugging if selected on the kernel commandline.
1201 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1202 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1203 flags
|= slub_debug
;
1208 static inline void setup_object_debug(struct kmem_cache
*s
,
1209 struct page
*page
, void *object
) {}
1211 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1212 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1214 static inline struct kmem_cache_node
*free_debug_processing(
1215 struct kmem_cache
*s
, struct page
*page
, void *object
,
1216 unsigned long addr
, unsigned long *flags
) { return NULL
; }
1218 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1220 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1221 void *object
, u8 val
) { return 1; }
1222 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1223 struct page
*page
) {}
1224 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1225 struct page
*page
) {}
1226 unsigned long kmem_cache_flags(unsigned long object_size
,
1227 unsigned long flags
, const char *name
,
1228 void (*ctor
)(void *))
1232 #define slub_debug 0
1234 #define disable_higher_order_debug 0
1236 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1238 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1240 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1242 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1245 #endif /* CONFIG_SLUB_DEBUG */
1248 * Hooks for other subsystems that check memory allocations. In a typical
1249 * production configuration these hooks all should produce no code at all.
1251 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1253 kmemleak_alloc(ptr
, size
, 1, flags
);
1254 kasan_kmalloc_large(ptr
, size
);
1257 static inline void kfree_hook(const void *x
)
1260 kasan_kfree_large(x
);
1263 static inline struct kmem_cache
*slab_pre_alloc_hook(struct kmem_cache
*s
,
1266 flags
&= gfp_allowed_mask
;
1267 lockdep_trace_alloc(flags
);
1268 might_sleep_if(gfpflags_allow_blocking(flags
));
1270 if (should_failslab(s
->object_size
, flags
, s
->flags
))
1273 return memcg_kmem_get_cache(s
, flags
);
1276 static inline void slab_post_alloc_hook(struct kmem_cache
*s
,
1277 gfp_t flags
, void *object
)
1279 flags
&= gfp_allowed_mask
;
1280 kmemcheck_slab_alloc(s
, flags
, object
, slab_ksize(s
));
1281 kmemleak_alloc_recursive(object
, s
->object_size
, 1, s
->flags
, flags
);
1282 memcg_kmem_put_cache(s
);
1283 kasan_slab_alloc(s
, object
);
1286 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
1288 kmemleak_free_recursive(x
, s
->flags
);
1291 * Trouble is that we may no longer disable interrupts in the fast path
1292 * So in order to make the debug calls that expect irqs to be
1293 * disabled we need to disable interrupts temporarily.
1295 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1297 unsigned long flags
;
1299 local_irq_save(flags
);
1300 kmemcheck_slab_free(s
, x
, s
->object_size
);
1301 debug_check_no_locks_freed(x
, s
->object_size
);
1302 local_irq_restore(flags
);
1305 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1306 debug_check_no_obj_freed(x
, s
->object_size
);
1308 kasan_slab_free(s
, x
);
1311 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1314 setup_object_debug(s
, page
, object
);
1315 if (unlikely(s
->ctor
)) {
1316 kasan_unpoison_object_data(s
, object
);
1318 kasan_poison_object_data(s
, object
);
1323 * Slab allocation and freeing
1325 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1326 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1329 int order
= oo_order(oo
);
1331 flags
|= __GFP_NOTRACK
;
1333 if (node
== NUMA_NO_NODE
)
1334 page
= alloc_pages(flags
, order
);
1336 page
= __alloc_pages_node(node
, flags
, order
);
1338 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1339 __free_pages(page
, order
);
1346 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1349 struct kmem_cache_order_objects oo
= s
->oo
;
1354 flags
&= gfp_allowed_mask
;
1356 if (gfpflags_allow_blocking(flags
))
1359 flags
|= s
->allocflags
;
1362 * Let the initial higher-order allocation fail under memory pressure
1363 * so we fall-back to the minimum order allocation.
1365 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1366 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1367 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~__GFP_DIRECT_RECLAIM
;
1369 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1370 if (unlikely(!page
)) {
1374 * Allocation may have failed due to fragmentation.
1375 * Try a lower order alloc if possible
1377 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1378 if (unlikely(!page
))
1380 stat(s
, ORDER_FALLBACK
);
1383 if (kmemcheck_enabled
&&
1384 !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1385 int pages
= 1 << oo_order(oo
);
1387 kmemcheck_alloc_shadow(page
, oo_order(oo
), alloc_gfp
, node
);
1390 * Objects from caches that have a constructor don't get
1391 * cleared when they're allocated, so we need to do it here.
1394 kmemcheck_mark_uninitialized_pages(page
, pages
);
1396 kmemcheck_mark_unallocated_pages(page
, pages
);
1399 page
->objects
= oo_objects(oo
);
1401 order
= compound_order(page
);
1402 page
->slab_cache
= s
;
1403 __SetPageSlab(page
);
1404 if (page_is_pfmemalloc(page
))
1405 SetPageSlabPfmemalloc(page
);
1407 start
= page_address(page
);
1409 if (unlikely(s
->flags
& SLAB_POISON
))
1410 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1412 kasan_poison_slab(page
);
1414 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1415 setup_object(s
, page
, p
);
1416 if (likely(idx
< page
->objects
))
1417 set_freepointer(s
, p
, p
+ s
->size
);
1419 set_freepointer(s
, p
, NULL
);
1422 page
->freelist
= start
;
1423 page
->inuse
= page
->objects
;
1427 if (gfpflags_allow_blocking(flags
))
1428 local_irq_disable();
1432 mod_zone_page_state(page_zone(page
),
1433 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1434 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1437 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1442 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1444 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1445 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
1449 return allocate_slab(s
,
1450 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1453 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1455 int order
= compound_order(page
);
1456 int pages
= 1 << order
;
1458 if (kmem_cache_debug(s
)) {
1461 slab_pad_check(s
, page
);
1462 for_each_object(p
, s
, page_address(page
),
1464 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1467 kmemcheck_free_shadow(page
, compound_order(page
));
1469 mod_zone_page_state(page_zone(page
),
1470 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1471 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1474 __ClearPageSlabPfmemalloc(page
);
1475 __ClearPageSlab(page
);
1477 page_mapcount_reset(page
);
1478 if (current
->reclaim_state
)
1479 current
->reclaim_state
->reclaimed_slab
+= pages
;
1480 __free_kmem_pages(page
, order
);
1483 #define need_reserve_slab_rcu \
1484 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486 static void rcu_free_slab(struct rcu_head
*h
)
1490 if (need_reserve_slab_rcu
)
1491 page
= virt_to_head_page(h
);
1493 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1495 __free_slab(page
->slab_cache
, page
);
1498 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1500 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1501 struct rcu_head
*head
;
1503 if (need_reserve_slab_rcu
) {
1504 int order
= compound_order(page
);
1505 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1507 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1508 head
= page_address(page
) + offset
;
1510 head
= &page
->rcu_head
;
1513 call_rcu(head
, rcu_free_slab
);
1515 __free_slab(s
, page
);
1518 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1520 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1525 * Management of partially allocated slabs.
1528 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1531 if (tail
== DEACTIVATE_TO_TAIL
)
1532 list_add_tail(&page
->lru
, &n
->partial
);
1534 list_add(&page
->lru
, &n
->partial
);
1537 static inline void add_partial(struct kmem_cache_node
*n
,
1538 struct page
*page
, int tail
)
1540 lockdep_assert_held(&n
->list_lock
);
1541 __add_partial(n
, page
, tail
);
1545 __remove_partial(struct kmem_cache_node
*n
, struct page
*page
)
1547 list_del(&page
->lru
);
1551 static inline void remove_partial(struct kmem_cache_node
*n
,
1554 lockdep_assert_held(&n
->list_lock
);
1555 __remove_partial(n
, page
);
1559 * Remove slab from the partial list, freeze it and
1560 * return the pointer to the freelist.
1562 * Returns a list of objects or NULL if it fails.
1564 static inline void *acquire_slab(struct kmem_cache
*s
,
1565 struct kmem_cache_node
*n
, struct page
*page
,
1566 int mode
, int *objects
)
1569 unsigned long counters
;
1572 lockdep_assert_held(&n
->list_lock
);
1575 * Zap the freelist and set the frozen bit.
1576 * The old freelist is the list of objects for the
1577 * per cpu allocation list.
1579 freelist
= page
->freelist
;
1580 counters
= page
->counters
;
1581 new.counters
= counters
;
1582 *objects
= new.objects
- new.inuse
;
1584 new.inuse
= page
->objects
;
1585 new.freelist
= NULL
;
1587 new.freelist
= freelist
;
1590 VM_BUG_ON(new.frozen
);
1593 if (!__cmpxchg_double_slab(s
, page
,
1595 new.freelist
, new.counters
,
1599 remove_partial(n
, page
);
1604 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1605 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1608 * Try to allocate a partial slab from a specific node.
1610 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1611 struct kmem_cache_cpu
*c
, gfp_t flags
)
1613 struct page
*page
, *page2
;
1614 void *object
= NULL
;
1619 * Racy check. If we mistakenly see no partial slabs then we
1620 * just allocate an empty slab. If we mistakenly try to get a
1621 * partial slab and there is none available then get_partials()
1624 if (!n
|| !n
->nr_partial
)
1627 spin_lock(&n
->list_lock
);
1628 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1631 if (!pfmemalloc_match(page
, flags
))
1634 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1638 available
+= objects
;
1641 stat(s
, ALLOC_FROM_PARTIAL
);
1644 put_cpu_partial(s
, page
, 0);
1645 stat(s
, CPU_PARTIAL_NODE
);
1647 if (!kmem_cache_has_cpu_partial(s
)
1648 || available
> s
->cpu_partial
/ 2)
1652 spin_unlock(&n
->list_lock
);
1657 * Get a page from somewhere. Search in increasing NUMA distances.
1659 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1660 struct kmem_cache_cpu
*c
)
1663 struct zonelist
*zonelist
;
1666 enum zone_type high_zoneidx
= gfp_zone(flags
);
1668 unsigned int cpuset_mems_cookie
;
1671 * The defrag ratio allows a configuration of the tradeoffs between
1672 * inter node defragmentation and node local allocations. A lower
1673 * defrag_ratio increases the tendency to do local allocations
1674 * instead of attempting to obtain partial slabs from other nodes.
1676 * If the defrag_ratio is set to 0 then kmalloc() always
1677 * returns node local objects. If the ratio is higher then kmalloc()
1678 * may return off node objects because partial slabs are obtained
1679 * from other nodes and filled up.
1681 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1682 * defrag_ratio = 1000) then every (well almost) allocation will
1683 * first attempt to defrag slab caches on other nodes. This means
1684 * scanning over all nodes to look for partial slabs which may be
1685 * expensive if we do it every time we are trying to find a slab
1686 * with available objects.
1688 if (!s
->remote_node_defrag_ratio
||
1689 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1693 cpuset_mems_cookie
= read_mems_allowed_begin();
1694 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1695 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1696 struct kmem_cache_node
*n
;
1698 n
= get_node(s
, zone_to_nid(zone
));
1700 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1701 n
->nr_partial
> s
->min_partial
) {
1702 object
= get_partial_node(s
, n
, c
, flags
);
1705 * Don't check read_mems_allowed_retry()
1706 * here - if mems_allowed was updated in
1707 * parallel, that was a harmless race
1708 * between allocation and the cpuset
1715 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1721 * Get a partial page, lock it and return it.
1723 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1724 struct kmem_cache_cpu
*c
)
1727 int searchnode
= node
;
1729 if (node
== NUMA_NO_NODE
)
1730 searchnode
= numa_mem_id();
1731 else if (!node_present_pages(node
))
1732 searchnode
= node_to_mem_node(node
);
1734 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1735 if (object
|| node
!= NUMA_NO_NODE
)
1738 return get_any_partial(s
, flags
, c
);
1741 #ifdef CONFIG_PREEMPT
1743 * Calculate the next globally unique transaction for disambiguiation
1744 * during cmpxchg. The transactions start with the cpu number and are then
1745 * incremented by CONFIG_NR_CPUS.
1747 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1750 * No preemption supported therefore also no need to check for
1756 static inline unsigned long next_tid(unsigned long tid
)
1758 return tid
+ TID_STEP
;
1761 static inline unsigned int tid_to_cpu(unsigned long tid
)
1763 return tid
% TID_STEP
;
1766 static inline unsigned long tid_to_event(unsigned long tid
)
1768 return tid
/ TID_STEP
;
1771 static inline unsigned int init_tid(int cpu
)
1776 static inline void note_cmpxchg_failure(const char *n
,
1777 const struct kmem_cache
*s
, unsigned long tid
)
1779 #ifdef SLUB_DEBUG_CMPXCHG
1780 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1782 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1784 #ifdef CONFIG_PREEMPT
1785 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1786 pr_warn("due to cpu change %d -> %d\n",
1787 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1790 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1791 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1792 tid_to_event(tid
), tid_to_event(actual_tid
));
1794 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1795 actual_tid
, tid
, next_tid(tid
));
1797 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1800 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1804 for_each_possible_cpu(cpu
)
1805 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1809 * Remove the cpu slab
1811 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1814 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1815 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1817 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1819 int tail
= DEACTIVATE_TO_HEAD
;
1823 if (page
->freelist
) {
1824 stat(s
, DEACTIVATE_REMOTE_FREES
);
1825 tail
= DEACTIVATE_TO_TAIL
;
1829 * Stage one: Free all available per cpu objects back
1830 * to the page freelist while it is still frozen. Leave the
1833 * There is no need to take the list->lock because the page
1836 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
1838 unsigned long counters
;
1841 prior
= page
->freelist
;
1842 counters
= page
->counters
;
1843 set_freepointer(s
, freelist
, prior
);
1844 new.counters
= counters
;
1846 VM_BUG_ON(!new.frozen
);
1848 } while (!__cmpxchg_double_slab(s
, page
,
1850 freelist
, new.counters
,
1851 "drain percpu freelist"));
1853 freelist
= nextfree
;
1857 * Stage two: Ensure that the page is unfrozen while the
1858 * list presence reflects the actual number of objects
1861 * We setup the list membership and then perform a cmpxchg
1862 * with the count. If there is a mismatch then the page
1863 * is not unfrozen but the page is on the wrong list.
1865 * Then we restart the process which may have to remove
1866 * the page from the list that we just put it on again
1867 * because the number of objects in the slab may have
1872 old
.freelist
= page
->freelist
;
1873 old
.counters
= page
->counters
;
1874 VM_BUG_ON(!old
.frozen
);
1876 /* Determine target state of the slab */
1877 new.counters
= old
.counters
;
1880 set_freepointer(s
, freelist
, old
.freelist
);
1881 new.freelist
= freelist
;
1883 new.freelist
= old
.freelist
;
1887 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
1889 else if (new.freelist
) {
1894 * Taking the spinlock removes the possiblity
1895 * that acquire_slab() will see a slab page that
1898 spin_lock(&n
->list_lock
);
1902 if (kmem_cache_debug(s
) && !lock
) {
1905 * This also ensures that the scanning of full
1906 * slabs from diagnostic functions will not see
1909 spin_lock(&n
->list_lock
);
1917 remove_partial(n
, page
);
1919 else if (l
== M_FULL
)
1921 remove_full(s
, n
, page
);
1923 if (m
== M_PARTIAL
) {
1925 add_partial(n
, page
, tail
);
1928 } else if (m
== M_FULL
) {
1930 stat(s
, DEACTIVATE_FULL
);
1931 add_full(s
, n
, page
);
1937 if (!__cmpxchg_double_slab(s
, page
,
1938 old
.freelist
, old
.counters
,
1939 new.freelist
, new.counters
,
1944 spin_unlock(&n
->list_lock
);
1947 stat(s
, DEACTIVATE_EMPTY
);
1948 discard_slab(s
, page
);
1954 * Unfreeze all the cpu partial slabs.
1956 * This function must be called with interrupts disabled
1957 * for the cpu using c (or some other guarantee must be there
1958 * to guarantee no concurrent accesses).
1960 static void unfreeze_partials(struct kmem_cache
*s
,
1961 struct kmem_cache_cpu
*c
)
1963 #ifdef CONFIG_SLUB_CPU_PARTIAL
1964 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
1965 struct page
*page
, *discard_page
= NULL
;
1967 while ((page
= c
->partial
)) {
1971 c
->partial
= page
->next
;
1973 n2
= get_node(s
, page_to_nid(page
));
1976 spin_unlock(&n
->list_lock
);
1979 spin_lock(&n
->list_lock
);
1984 old
.freelist
= page
->freelist
;
1985 old
.counters
= page
->counters
;
1986 VM_BUG_ON(!old
.frozen
);
1988 new.counters
= old
.counters
;
1989 new.freelist
= old
.freelist
;
1993 } while (!__cmpxchg_double_slab(s
, page
,
1994 old
.freelist
, old
.counters
,
1995 new.freelist
, new.counters
,
1996 "unfreezing slab"));
1998 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
1999 page
->next
= discard_page
;
2000 discard_page
= page
;
2002 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2003 stat(s
, FREE_ADD_PARTIAL
);
2008 spin_unlock(&n
->list_lock
);
2010 while (discard_page
) {
2011 page
= discard_page
;
2012 discard_page
= discard_page
->next
;
2014 stat(s
, DEACTIVATE_EMPTY
);
2015 discard_slab(s
, page
);
2022 * Put a page that was just frozen (in __slab_free) into a partial page
2023 * slot if available. This is done without interrupts disabled and without
2024 * preemption disabled. The cmpxchg is racy and may put the partial page
2025 * onto a random cpus partial slot.
2027 * If we did not find a slot then simply move all the partials to the
2028 * per node partial list.
2030 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2032 #ifdef CONFIG_SLUB_CPU_PARTIAL
2033 struct page
*oldpage
;
2041 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2044 pobjects
= oldpage
->pobjects
;
2045 pages
= oldpage
->pages
;
2046 if (drain
&& pobjects
> s
->cpu_partial
) {
2047 unsigned long flags
;
2049 * partial array is full. Move the existing
2050 * set to the per node partial list.
2052 local_irq_save(flags
);
2053 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2054 local_irq_restore(flags
);
2058 stat(s
, CPU_PARTIAL_DRAIN
);
2063 pobjects
+= page
->objects
- page
->inuse
;
2065 page
->pages
= pages
;
2066 page
->pobjects
= pobjects
;
2067 page
->next
= oldpage
;
2069 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2071 if (unlikely(!s
->cpu_partial
)) {
2072 unsigned long flags
;
2074 local_irq_save(flags
);
2075 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2076 local_irq_restore(flags
);
2082 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2084 stat(s
, CPUSLAB_FLUSH
);
2085 deactivate_slab(s
, c
->page
, c
->freelist
);
2087 c
->tid
= next_tid(c
->tid
);
2095 * Called from IPI handler with interrupts disabled.
2097 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2099 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2105 unfreeze_partials(s
, c
);
2109 static void flush_cpu_slab(void *d
)
2111 struct kmem_cache
*s
= d
;
2113 __flush_cpu_slab(s
, smp_processor_id());
2116 static bool has_cpu_slab(int cpu
, void *info
)
2118 struct kmem_cache
*s
= info
;
2119 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2121 return c
->page
|| c
->partial
;
2124 static void flush_all(struct kmem_cache
*s
)
2126 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2130 * Check if the objects in a per cpu structure fit numa
2131 * locality expectations.
2133 static inline int node_match(struct page
*page
, int node
)
2136 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2142 #ifdef CONFIG_SLUB_DEBUG
2143 static int count_free(struct page
*page
)
2145 return page
->objects
- page
->inuse
;
2148 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2150 return atomic_long_read(&n
->total_objects
);
2152 #endif /* CONFIG_SLUB_DEBUG */
2154 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2155 static unsigned long count_partial(struct kmem_cache_node
*n
,
2156 int (*get_count
)(struct page
*))
2158 unsigned long flags
;
2159 unsigned long x
= 0;
2162 spin_lock_irqsave(&n
->list_lock
, flags
);
2163 list_for_each_entry(page
, &n
->partial
, lru
)
2164 x
+= get_count(page
);
2165 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2168 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2170 static noinline
void
2171 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2173 #ifdef CONFIG_SLUB_DEBUG
2174 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2175 DEFAULT_RATELIMIT_BURST
);
2177 struct kmem_cache_node
*n
;
2179 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2182 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2184 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2185 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2188 if (oo_order(s
->min
) > get_order(s
->object_size
))
2189 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2192 for_each_kmem_cache_node(s
, node
, n
) {
2193 unsigned long nr_slabs
;
2194 unsigned long nr_objs
;
2195 unsigned long nr_free
;
2197 nr_free
= count_partial(n
, count_free
);
2198 nr_slabs
= node_nr_slabs(n
);
2199 nr_objs
= node_nr_objs(n
);
2201 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2202 node
, nr_slabs
, nr_objs
, nr_free
);
2207 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2208 int node
, struct kmem_cache_cpu
**pc
)
2211 struct kmem_cache_cpu
*c
= *pc
;
2214 freelist
= get_partial(s
, flags
, node
, c
);
2219 page
= new_slab(s
, flags
, node
);
2221 c
= raw_cpu_ptr(s
->cpu_slab
);
2226 * No other reference to the page yet so we can
2227 * muck around with it freely without cmpxchg
2229 freelist
= page
->freelist
;
2230 page
->freelist
= NULL
;
2232 stat(s
, ALLOC_SLAB
);
2241 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2243 if (unlikely(PageSlabPfmemalloc(page
)))
2244 return gfp_pfmemalloc_allowed(gfpflags
);
2250 * Check the page->freelist of a page and either transfer the freelist to the
2251 * per cpu freelist or deactivate the page.
2253 * The page is still frozen if the return value is not NULL.
2255 * If this function returns NULL then the page has been unfrozen.
2257 * This function must be called with interrupt disabled.
2259 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2262 unsigned long counters
;
2266 freelist
= page
->freelist
;
2267 counters
= page
->counters
;
2269 new.counters
= counters
;
2270 VM_BUG_ON(!new.frozen
);
2272 new.inuse
= page
->objects
;
2273 new.frozen
= freelist
!= NULL
;
2275 } while (!__cmpxchg_double_slab(s
, page
,
2284 * Slow path. The lockless freelist is empty or we need to perform
2287 * Processing is still very fast if new objects have been freed to the
2288 * regular freelist. In that case we simply take over the regular freelist
2289 * as the lockless freelist and zap the regular freelist.
2291 * If that is not working then we fall back to the partial lists. We take the
2292 * first element of the freelist as the object to allocate now and move the
2293 * rest of the freelist to the lockless freelist.
2295 * And if we were unable to get a new slab from the partial slab lists then
2296 * we need to allocate a new slab. This is the slowest path since it involves
2297 * a call to the page allocator and the setup of a new slab.
2299 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2300 unsigned long addr
, struct kmem_cache_cpu
*c
)
2304 unsigned long flags
;
2306 local_irq_save(flags
);
2307 #ifdef CONFIG_PREEMPT
2309 * We may have been preempted and rescheduled on a different
2310 * cpu before disabling interrupts. Need to reload cpu area
2313 c
= this_cpu_ptr(s
->cpu_slab
);
2321 if (unlikely(!node_match(page
, node
))) {
2322 int searchnode
= node
;
2324 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2325 searchnode
= node_to_mem_node(node
);
2327 if (unlikely(!node_match(page
, searchnode
))) {
2328 stat(s
, ALLOC_NODE_MISMATCH
);
2329 deactivate_slab(s
, page
, c
->freelist
);
2337 * By rights, we should be searching for a slab page that was
2338 * PFMEMALLOC but right now, we are losing the pfmemalloc
2339 * information when the page leaves the per-cpu allocator
2341 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2342 deactivate_slab(s
, page
, c
->freelist
);
2348 /* must check again c->freelist in case of cpu migration or IRQ */
2349 freelist
= c
->freelist
;
2353 freelist
= get_freelist(s
, page
);
2357 stat(s
, DEACTIVATE_BYPASS
);
2361 stat(s
, ALLOC_REFILL
);
2365 * freelist is pointing to the list of objects to be used.
2366 * page is pointing to the page from which the objects are obtained.
2367 * That page must be frozen for per cpu allocations to work.
2369 VM_BUG_ON(!c
->page
->frozen
);
2370 c
->freelist
= get_freepointer(s
, freelist
);
2371 c
->tid
= next_tid(c
->tid
);
2372 local_irq_restore(flags
);
2378 page
= c
->page
= c
->partial
;
2379 c
->partial
= page
->next
;
2380 stat(s
, CPU_PARTIAL_ALLOC
);
2385 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2387 if (unlikely(!freelist
)) {
2388 slab_out_of_memory(s
, gfpflags
, node
);
2389 local_irq_restore(flags
);
2394 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2397 /* Only entered in the debug case */
2398 if (kmem_cache_debug(s
) &&
2399 !alloc_debug_processing(s
, page
, freelist
, addr
))
2400 goto new_slab
; /* Slab failed checks. Next slab needed */
2402 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
2405 local_irq_restore(flags
);
2410 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2411 * have the fastpath folded into their functions. So no function call
2412 * overhead for requests that can be satisfied on the fastpath.
2414 * The fastpath works by first checking if the lockless freelist can be used.
2415 * If not then __slab_alloc is called for slow processing.
2417 * Otherwise we can simply pick the next object from the lockless free list.
2419 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2420 gfp_t gfpflags
, int node
, unsigned long addr
)
2423 struct kmem_cache_cpu
*c
;
2427 s
= slab_pre_alloc_hook(s
, gfpflags
);
2432 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2433 * enabled. We may switch back and forth between cpus while
2434 * reading from one cpu area. That does not matter as long
2435 * as we end up on the original cpu again when doing the cmpxchg.
2437 * We should guarantee that tid and kmem_cache are retrieved on
2438 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2439 * to check if it is matched or not.
2442 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2443 c
= raw_cpu_ptr(s
->cpu_slab
);
2444 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2445 unlikely(tid
!= READ_ONCE(c
->tid
)));
2448 * Irqless object alloc/free algorithm used here depends on sequence
2449 * of fetching cpu_slab's data. tid should be fetched before anything
2450 * on c to guarantee that object and page associated with previous tid
2451 * won't be used with current tid. If we fetch tid first, object and
2452 * page could be one associated with next tid and our alloc/free
2453 * request will be failed. In this case, we will retry. So, no problem.
2458 * The transaction ids are globally unique per cpu and per operation on
2459 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2460 * occurs on the right processor and that there was no operation on the
2461 * linked list in between.
2464 object
= c
->freelist
;
2466 if (unlikely(!object
|| !node_match(page
, node
))) {
2467 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2468 stat(s
, ALLOC_SLOWPATH
);
2470 void *next_object
= get_freepointer_safe(s
, object
);
2473 * The cmpxchg will only match if there was no additional
2474 * operation and if we are on the right processor.
2476 * The cmpxchg does the following atomically (without lock
2478 * 1. Relocate first pointer to the current per cpu area.
2479 * 2. Verify that tid and freelist have not been changed
2480 * 3. If they were not changed replace tid and freelist
2482 * Since this is without lock semantics the protection is only
2483 * against code executing on this cpu *not* from access by
2486 if (unlikely(!this_cpu_cmpxchg_double(
2487 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2489 next_object
, next_tid(tid
)))) {
2491 note_cmpxchg_failure("slab_alloc", s
, tid
);
2494 prefetch_freepointer(s
, next_object
);
2495 stat(s
, ALLOC_FASTPATH
);
2498 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2499 memset(object
, 0, s
->object_size
);
2501 slab_post_alloc_hook(s
, gfpflags
, object
);
2506 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2507 gfp_t gfpflags
, unsigned long addr
)
2509 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2512 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2514 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2516 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2521 EXPORT_SYMBOL(kmem_cache_alloc
);
2523 #ifdef CONFIG_TRACING
2524 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2526 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2527 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2528 kasan_kmalloc(s
, ret
, size
);
2531 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2535 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2537 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2539 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2540 s
->object_size
, s
->size
, gfpflags
, node
);
2544 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2546 #ifdef CONFIG_TRACING
2547 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2549 int node
, size_t size
)
2551 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2553 trace_kmalloc_node(_RET_IP_
, ret
,
2554 size
, s
->size
, gfpflags
, node
);
2556 kasan_kmalloc(s
, ret
, size
);
2559 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2564 * Slow path handling. This may still be called frequently since objects
2565 * have a longer lifetime than the cpu slabs in most processing loads.
2567 * So we still attempt to reduce cache line usage. Just take the slab
2568 * lock and free the item. If there is no additional partial page
2569 * handling required then we can return immediately.
2571 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2572 void *x
, unsigned long addr
)
2575 void **object
= (void *)x
;
2578 unsigned long counters
;
2579 struct kmem_cache_node
*n
= NULL
;
2580 unsigned long uninitialized_var(flags
);
2582 stat(s
, FREE_SLOWPATH
);
2584 if (kmem_cache_debug(s
) &&
2585 !(n
= free_debug_processing(s
, page
, x
, addr
, &flags
)))
2590 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2593 prior
= page
->freelist
;
2594 counters
= page
->counters
;
2595 set_freepointer(s
, object
, prior
);
2596 new.counters
= counters
;
2597 was_frozen
= new.frozen
;
2599 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2601 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2604 * Slab was on no list before and will be
2606 * We can defer the list move and instead
2611 } else { /* Needs to be taken off a list */
2613 n
= get_node(s
, page_to_nid(page
));
2615 * Speculatively acquire the list_lock.
2616 * If the cmpxchg does not succeed then we may
2617 * drop the list_lock without any processing.
2619 * Otherwise the list_lock will synchronize with
2620 * other processors updating the list of slabs.
2622 spin_lock_irqsave(&n
->list_lock
, flags
);
2627 } while (!cmpxchg_double_slab(s
, page
,
2629 object
, new.counters
,
2635 * If we just froze the page then put it onto the
2636 * per cpu partial list.
2638 if (new.frozen
&& !was_frozen
) {
2639 put_cpu_partial(s
, page
, 1);
2640 stat(s
, CPU_PARTIAL_FREE
);
2643 * The list lock was not taken therefore no list
2644 * activity can be necessary.
2647 stat(s
, FREE_FROZEN
);
2651 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2655 * Objects left in the slab. If it was not on the partial list before
2658 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2659 if (kmem_cache_debug(s
))
2660 remove_full(s
, n
, page
);
2661 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2662 stat(s
, FREE_ADD_PARTIAL
);
2664 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2670 * Slab on the partial list.
2672 remove_partial(n
, page
);
2673 stat(s
, FREE_REMOVE_PARTIAL
);
2675 /* Slab must be on the full list */
2676 remove_full(s
, n
, page
);
2679 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2681 discard_slab(s
, page
);
2685 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2686 * can perform fastpath freeing without additional function calls.
2688 * The fastpath is only possible if we are freeing to the current cpu slab
2689 * of this processor. This typically the case if we have just allocated
2692 * If fastpath is not possible then fall back to __slab_free where we deal
2693 * with all sorts of special processing.
2695 static __always_inline
void slab_free(struct kmem_cache
*s
,
2696 struct page
*page
, void *x
, unsigned long addr
)
2698 void **object
= (void *)x
;
2699 struct kmem_cache_cpu
*c
;
2702 slab_free_hook(s
, x
);
2706 * Determine the currently cpus per cpu slab.
2707 * The cpu may change afterward. However that does not matter since
2708 * data is retrieved via this pointer. If we are on the same cpu
2709 * during the cmpxchg then the free will succeed.
2712 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2713 c
= raw_cpu_ptr(s
->cpu_slab
);
2714 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2715 unlikely(tid
!= READ_ONCE(c
->tid
)));
2717 /* Same with comment on barrier() in slab_alloc_node() */
2720 if (likely(page
== c
->page
)) {
2721 set_freepointer(s
, object
, c
->freelist
);
2723 if (unlikely(!this_cpu_cmpxchg_double(
2724 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2726 object
, next_tid(tid
)))) {
2728 note_cmpxchg_failure("slab_free", s
, tid
);
2731 stat(s
, FREE_FASTPATH
);
2733 __slab_free(s
, page
, x
, addr
);
2737 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2739 s
= cache_from_obj(s
, x
);
2742 slab_free(s
, virt_to_head_page(x
), x
, _RET_IP_
);
2743 trace_kmem_cache_free(_RET_IP_
, x
);
2745 EXPORT_SYMBOL(kmem_cache_free
);
2747 /* Note that interrupts must be enabled when calling this function. */
2748 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
2750 struct kmem_cache_cpu
*c
;
2754 local_irq_disable();
2755 c
= this_cpu_ptr(s
->cpu_slab
);
2757 for (i
= 0; i
< size
; i
++) {
2758 void *object
= p
[i
];
2761 /* kmem cache debug support */
2762 s
= cache_from_obj(s
, object
);
2765 slab_free_hook(s
, object
);
2767 page
= virt_to_head_page(object
);
2769 if (c
->page
== page
) {
2770 /* Fastpath: local CPU free */
2771 set_freepointer(s
, object
, c
->freelist
);
2772 c
->freelist
= object
;
2774 c
->tid
= next_tid(c
->tid
);
2776 /* Slowpath: overhead locked cmpxchg_double_slab */
2777 __slab_free(s
, page
, object
, _RET_IP_
);
2778 local_irq_disable();
2779 c
= this_cpu_ptr(s
->cpu_slab
);
2783 c
->tid
= next_tid(c
->tid
);
2786 EXPORT_SYMBOL(kmem_cache_free_bulk
);
2788 /* Note that interrupts must be enabled when calling this function. */
2789 bool kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
2792 struct kmem_cache_cpu
*c
;
2796 * Drain objects in the per cpu slab, while disabling local
2797 * IRQs, which protects against PREEMPT and interrupts
2798 * handlers invoking normal fastpath.
2800 local_irq_disable();
2801 c
= this_cpu_ptr(s
->cpu_slab
);
2803 for (i
= 0; i
< size
; i
++) {
2804 void *object
= c
->freelist
;
2806 if (unlikely(!object
)) {
2809 * Invoking slow path likely have side-effect
2810 * of re-populating per CPU c->freelist
2812 p
[i
] = __slab_alloc(s
, flags
, NUMA_NO_NODE
,
2814 if (unlikely(!p
[i
])) {
2815 __kmem_cache_free_bulk(s
, i
, p
);
2818 local_irq_disable();
2819 c
= this_cpu_ptr(s
->cpu_slab
);
2820 continue; /* goto for-loop */
2823 /* kmem_cache debug support */
2824 s
= slab_pre_alloc_hook(s
, flags
);
2826 __kmem_cache_free_bulk(s
, i
, p
);
2827 c
->tid
= next_tid(c
->tid
);
2832 c
->freelist
= get_freepointer(s
, object
);
2835 /* kmem_cache debug support */
2836 slab_post_alloc_hook(s
, flags
, object
);
2838 c
->tid
= next_tid(c
->tid
);
2841 /* Clear memory outside IRQ disabled fastpath loop */
2842 if (unlikely(flags
& __GFP_ZERO
)) {
2845 for (j
= 0; j
< i
; j
++)
2846 memset(p
[j
], 0, s
->object_size
);
2851 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
2855 * Object placement in a slab is made very easy because we always start at
2856 * offset 0. If we tune the size of the object to the alignment then we can
2857 * get the required alignment by putting one properly sized object after
2860 * Notice that the allocation order determines the sizes of the per cpu
2861 * caches. Each processor has always one slab available for allocations.
2862 * Increasing the allocation order reduces the number of times that slabs
2863 * must be moved on and off the partial lists and is therefore a factor in
2868 * Mininum / Maximum order of slab pages. This influences locking overhead
2869 * and slab fragmentation. A higher order reduces the number of partial slabs
2870 * and increases the number of allocations possible without having to
2871 * take the list_lock.
2873 static int slub_min_order
;
2874 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
2875 static int slub_min_objects
;
2878 * Calculate the order of allocation given an slab object size.
2880 * The order of allocation has significant impact on performance and other
2881 * system components. Generally order 0 allocations should be preferred since
2882 * order 0 does not cause fragmentation in the page allocator. Larger objects
2883 * be problematic to put into order 0 slabs because there may be too much
2884 * unused space left. We go to a higher order if more than 1/16th of the slab
2887 * In order to reach satisfactory performance we must ensure that a minimum
2888 * number of objects is in one slab. Otherwise we may generate too much
2889 * activity on the partial lists which requires taking the list_lock. This is
2890 * less a concern for large slabs though which are rarely used.
2892 * slub_max_order specifies the order where we begin to stop considering the
2893 * number of objects in a slab as critical. If we reach slub_max_order then
2894 * we try to keep the page order as low as possible. So we accept more waste
2895 * of space in favor of a small page order.
2897 * Higher order allocations also allow the placement of more objects in a
2898 * slab and thereby reduce object handling overhead. If the user has
2899 * requested a higher mininum order then we start with that one instead of
2900 * the smallest order which will fit the object.
2902 static inline int slab_order(int size
, int min_objects
,
2903 int max_order
, int fract_leftover
, int reserved
)
2907 int min_order
= slub_min_order
;
2909 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
2910 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
2912 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
2913 order
<= max_order
; order
++) {
2915 unsigned long slab_size
= PAGE_SIZE
<< order
;
2917 rem
= (slab_size
- reserved
) % size
;
2919 if (rem
<= slab_size
/ fract_leftover
)
2926 static inline int calculate_order(int size
, int reserved
)
2934 * Attempt to find best configuration for a slab. This
2935 * works by first attempting to generate a layout with
2936 * the best configuration and backing off gradually.
2938 * First we increase the acceptable waste in a slab. Then
2939 * we reduce the minimum objects required in a slab.
2941 min_objects
= slub_min_objects
;
2943 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
2944 max_objects
= order_objects(slub_max_order
, size
, reserved
);
2945 min_objects
= min(min_objects
, max_objects
);
2947 while (min_objects
> 1) {
2949 while (fraction
>= 4) {
2950 order
= slab_order(size
, min_objects
,
2951 slub_max_order
, fraction
, reserved
);
2952 if (order
<= slub_max_order
)
2960 * We were unable to place multiple objects in a slab. Now
2961 * lets see if we can place a single object there.
2963 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
2964 if (order
<= slub_max_order
)
2968 * Doh this slab cannot be placed using slub_max_order.
2970 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
2971 if (order
< MAX_ORDER
)
2977 init_kmem_cache_node(struct kmem_cache_node
*n
)
2980 spin_lock_init(&n
->list_lock
);
2981 INIT_LIST_HEAD(&n
->partial
);
2982 #ifdef CONFIG_SLUB_DEBUG
2983 atomic_long_set(&n
->nr_slabs
, 0);
2984 atomic_long_set(&n
->total_objects
, 0);
2985 INIT_LIST_HEAD(&n
->full
);
2989 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
2991 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
2992 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
2995 * Must align to double word boundary for the double cmpxchg
2996 * instructions to work; see __pcpu_double_call_return_bool().
2998 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
2999 2 * sizeof(void *));
3004 init_kmem_cache_cpus(s
);
3009 static struct kmem_cache
*kmem_cache_node
;
3012 * No kmalloc_node yet so do it by hand. We know that this is the first
3013 * slab on the node for this slabcache. There are no concurrent accesses
3016 * Note that this function only works on the kmem_cache_node
3017 * when allocating for the kmem_cache_node. This is used for bootstrapping
3018 * memory on a fresh node that has no slab structures yet.
3020 static void early_kmem_cache_node_alloc(int node
)
3023 struct kmem_cache_node
*n
;
3025 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3027 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3030 if (page_to_nid(page
) != node
) {
3031 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3032 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3037 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3040 kmem_cache_node
->node
[node
] = n
;
3041 #ifdef CONFIG_SLUB_DEBUG
3042 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3043 init_tracking(kmem_cache_node
, n
);
3045 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
));
3046 init_kmem_cache_node(n
);
3047 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3050 * No locks need to be taken here as it has just been
3051 * initialized and there is no concurrent access.
3053 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3056 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3059 struct kmem_cache_node
*n
;
3061 for_each_kmem_cache_node(s
, node
, n
) {
3062 kmem_cache_free(kmem_cache_node
, n
);
3063 s
->node
[node
] = NULL
;
3067 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3071 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3072 struct kmem_cache_node
*n
;
3074 if (slab_state
== DOWN
) {
3075 early_kmem_cache_node_alloc(node
);
3078 n
= kmem_cache_alloc_node(kmem_cache_node
,
3082 free_kmem_cache_nodes(s
);
3087 init_kmem_cache_node(n
);
3092 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3094 if (min
< MIN_PARTIAL
)
3096 else if (min
> MAX_PARTIAL
)
3098 s
->min_partial
= min
;
3102 * calculate_sizes() determines the order and the distribution of data within
3105 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3107 unsigned long flags
= s
->flags
;
3108 unsigned long size
= s
->object_size
;
3112 * Round up object size to the next word boundary. We can only
3113 * place the free pointer at word boundaries and this determines
3114 * the possible location of the free pointer.
3116 size
= ALIGN(size
, sizeof(void *));
3118 #ifdef CONFIG_SLUB_DEBUG
3120 * Determine if we can poison the object itself. If the user of
3121 * the slab may touch the object after free or before allocation
3122 * then we should never poison the object itself.
3124 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
3126 s
->flags
|= __OBJECT_POISON
;
3128 s
->flags
&= ~__OBJECT_POISON
;
3132 * If we are Redzoning then check if there is some space between the
3133 * end of the object and the free pointer. If not then add an
3134 * additional word to have some bytes to store Redzone information.
3136 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3137 size
+= sizeof(void *);
3141 * With that we have determined the number of bytes in actual use
3142 * by the object. This is the potential offset to the free pointer.
3146 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
3149 * Relocate free pointer after the object if it is not
3150 * permitted to overwrite the first word of the object on
3153 * This is the case if we do RCU, have a constructor or
3154 * destructor or are poisoning the objects.
3157 size
+= sizeof(void *);
3160 #ifdef CONFIG_SLUB_DEBUG
3161 if (flags
& SLAB_STORE_USER
)
3163 * Need to store information about allocs and frees after
3166 size
+= 2 * sizeof(struct track
);
3168 if (flags
& SLAB_RED_ZONE
)
3170 * Add some empty padding so that we can catch
3171 * overwrites from earlier objects rather than let
3172 * tracking information or the free pointer be
3173 * corrupted if a user writes before the start
3176 size
+= sizeof(void *);
3180 * SLUB stores one object immediately after another beginning from
3181 * offset 0. In order to align the objects we have to simply size
3182 * each object to conform to the alignment.
3184 size
= ALIGN(size
, s
->align
);
3186 if (forced_order
>= 0)
3187 order
= forced_order
;
3189 order
= calculate_order(size
, s
->reserved
);
3196 s
->allocflags
|= __GFP_COMP
;
3198 if (s
->flags
& SLAB_CACHE_DMA
)
3199 s
->allocflags
|= GFP_DMA
;
3201 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3202 s
->allocflags
|= __GFP_RECLAIMABLE
;
3205 * Determine the number of objects per slab
3207 s
->oo
= oo_make(order
, size
, s
->reserved
);
3208 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3209 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3212 return !!oo_objects(s
->oo
);
3215 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3217 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3220 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
3221 s
->reserved
= sizeof(struct rcu_head
);
3223 if (!calculate_sizes(s
, -1))
3225 if (disable_higher_order_debug
) {
3227 * Disable debugging flags that store metadata if the min slab
3230 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3231 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3233 if (!calculate_sizes(s
, -1))
3238 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3239 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3240 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_DEBUG_FLAGS
) == 0)
3241 /* Enable fast mode */
3242 s
->flags
|= __CMPXCHG_DOUBLE
;
3246 * The larger the object size is, the more pages we want on the partial
3247 * list to avoid pounding the page allocator excessively.
3249 set_min_partial(s
, ilog2(s
->size
) / 2);
3252 * cpu_partial determined the maximum number of objects kept in the
3253 * per cpu partial lists of a processor.
3255 * Per cpu partial lists mainly contain slabs that just have one
3256 * object freed. If they are used for allocation then they can be
3257 * filled up again with minimal effort. The slab will never hit the
3258 * per node partial lists and therefore no locking will be required.
3260 * This setting also determines
3262 * A) The number of objects from per cpu partial slabs dumped to the
3263 * per node list when we reach the limit.
3264 * B) The number of objects in cpu partial slabs to extract from the
3265 * per node list when we run out of per cpu objects. We only fetch
3266 * 50% to keep some capacity around for frees.
3268 if (!kmem_cache_has_cpu_partial(s
))
3270 else if (s
->size
>= PAGE_SIZE
)
3272 else if (s
->size
>= 1024)
3274 else if (s
->size
>= 256)
3275 s
->cpu_partial
= 13;
3277 s
->cpu_partial
= 30;
3280 s
->remote_node_defrag_ratio
= 1000;
3282 if (!init_kmem_cache_nodes(s
))
3285 if (alloc_kmem_cache_cpus(s
))
3288 free_kmem_cache_nodes(s
);
3290 if (flags
& SLAB_PANIC
)
3291 panic("Cannot create slab %s size=%lu realsize=%u "
3292 "order=%u offset=%u flags=%lx\n",
3293 s
->name
, (unsigned long)s
->size
, s
->size
,
3294 oo_order(s
->oo
), s
->offset
, flags
);
3298 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3301 #ifdef CONFIG_SLUB_DEBUG
3302 void *addr
= page_address(page
);
3304 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3305 sizeof(long), GFP_ATOMIC
);
3308 slab_err(s
, page
, text
, s
->name
);
3311 get_map(s
, page
, map
);
3312 for_each_object(p
, s
, addr
, page
->objects
) {
3314 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3315 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3316 print_tracking(s
, p
);
3325 * Attempt to free all partial slabs on a node.
3326 * This is called from kmem_cache_close(). We must be the last thread
3327 * using the cache and therefore we do not need to lock anymore.
3329 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3331 struct page
*page
, *h
;
3333 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3335 __remove_partial(n
, page
);
3336 discard_slab(s
, page
);
3338 list_slab_objects(s
, page
,
3339 "Objects remaining in %s on kmem_cache_close()");
3345 * Release all resources used by a slab cache.
3347 static inline int kmem_cache_close(struct kmem_cache
*s
)
3350 struct kmem_cache_node
*n
;
3353 /* Attempt to free all objects */
3354 for_each_kmem_cache_node(s
, node
, n
) {
3356 if (n
->nr_partial
|| slabs_node(s
, node
))
3359 free_percpu(s
->cpu_slab
);
3360 free_kmem_cache_nodes(s
);
3364 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3366 return kmem_cache_close(s
);
3369 /********************************************************************
3371 *******************************************************************/
3373 static int __init
setup_slub_min_order(char *str
)
3375 get_option(&str
, &slub_min_order
);
3380 __setup("slub_min_order=", setup_slub_min_order
);
3382 static int __init
setup_slub_max_order(char *str
)
3384 get_option(&str
, &slub_max_order
);
3385 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3390 __setup("slub_max_order=", setup_slub_max_order
);
3392 static int __init
setup_slub_min_objects(char *str
)
3394 get_option(&str
, &slub_min_objects
);
3399 __setup("slub_min_objects=", setup_slub_min_objects
);
3401 void *__kmalloc(size_t size
, gfp_t flags
)
3403 struct kmem_cache
*s
;
3406 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3407 return kmalloc_large(size
, flags
);
3409 s
= kmalloc_slab(size
, flags
);
3411 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3414 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3416 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3418 kasan_kmalloc(s
, ret
, size
);
3422 EXPORT_SYMBOL(__kmalloc
);
3425 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3430 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
3431 page
= alloc_kmem_pages_node(node
, flags
, get_order(size
));
3433 ptr
= page_address(page
);
3435 kmalloc_large_node_hook(ptr
, size
, flags
);
3439 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3441 struct kmem_cache
*s
;
3444 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3445 ret
= kmalloc_large_node(size
, flags
, node
);
3447 trace_kmalloc_node(_RET_IP_
, ret
,
3448 size
, PAGE_SIZE
<< get_order(size
),
3454 s
= kmalloc_slab(size
, flags
);
3456 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3459 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3461 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3463 kasan_kmalloc(s
, ret
, size
);
3467 EXPORT_SYMBOL(__kmalloc_node
);
3470 static size_t __ksize(const void *object
)
3474 if (unlikely(object
== ZERO_SIZE_PTR
))
3477 page
= virt_to_head_page(object
);
3479 if (unlikely(!PageSlab(page
))) {
3480 WARN_ON(!PageCompound(page
));
3481 return PAGE_SIZE
<< compound_order(page
);
3484 return slab_ksize(page
->slab_cache
);
3487 size_t ksize(const void *object
)
3489 size_t size
= __ksize(object
);
3490 /* We assume that ksize callers could use whole allocated area,
3491 so we need unpoison this area. */
3492 kasan_krealloc(object
, size
);
3495 EXPORT_SYMBOL(ksize
);
3497 void kfree(const void *x
)
3500 void *object
= (void *)x
;
3502 trace_kfree(_RET_IP_
, x
);
3504 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3507 page
= virt_to_head_page(x
);
3508 if (unlikely(!PageSlab(page
))) {
3509 BUG_ON(!PageCompound(page
));
3511 __free_kmem_pages(page
, compound_order(page
));
3514 slab_free(page
->slab_cache
, page
, object
, _RET_IP_
);
3516 EXPORT_SYMBOL(kfree
);
3518 #define SHRINK_PROMOTE_MAX 32
3521 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3522 * up most to the head of the partial lists. New allocations will then
3523 * fill those up and thus they can be removed from the partial lists.
3525 * The slabs with the least items are placed last. This results in them
3526 * being allocated from last increasing the chance that the last objects
3527 * are freed in them.
3529 int __kmem_cache_shrink(struct kmem_cache
*s
, bool deactivate
)
3533 struct kmem_cache_node
*n
;
3536 struct list_head discard
;
3537 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3538 unsigned long flags
;
3543 * Disable empty slabs caching. Used to avoid pinning offline
3544 * memory cgroups by kmem pages that can be freed.
3550 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3551 * so we have to make sure the change is visible.
3553 kick_all_cpus_sync();
3557 for_each_kmem_cache_node(s
, node
, n
) {
3558 INIT_LIST_HEAD(&discard
);
3559 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3560 INIT_LIST_HEAD(promote
+ i
);
3562 spin_lock_irqsave(&n
->list_lock
, flags
);
3565 * Build lists of slabs to discard or promote.
3567 * Note that concurrent frees may occur while we hold the
3568 * list_lock. page->inuse here is the upper limit.
3570 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3571 int free
= page
->objects
- page
->inuse
;
3573 /* Do not reread page->inuse */
3576 /* We do not keep full slabs on the list */
3579 if (free
== page
->objects
) {
3580 list_move(&page
->lru
, &discard
);
3582 } else if (free
<= SHRINK_PROMOTE_MAX
)
3583 list_move(&page
->lru
, promote
+ free
- 1);
3587 * Promote the slabs filled up most to the head of the
3590 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3591 list_splice(promote
+ i
, &n
->partial
);
3593 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3595 /* Release empty slabs */
3596 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3597 discard_slab(s
, page
);
3599 if (slabs_node(s
, node
))
3606 static int slab_mem_going_offline_callback(void *arg
)
3608 struct kmem_cache
*s
;
3610 mutex_lock(&slab_mutex
);
3611 list_for_each_entry(s
, &slab_caches
, list
)
3612 __kmem_cache_shrink(s
, false);
3613 mutex_unlock(&slab_mutex
);
3618 static void slab_mem_offline_callback(void *arg
)
3620 struct kmem_cache_node
*n
;
3621 struct kmem_cache
*s
;
3622 struct memory_notify
*marg
= arg
;
3625 offline_node
= marg
->status_change_nid_normal
;
3628 * If the node still has available memory. we need kmem_cache_node
3631 if (offline_node
< 0)
3634 mutex_lock(&slab_mutex
);
3635 list_for_each_entry(s
, &slab_caches
, list
) {
3636 n
= get_node(s
, offline_node
);
3639 * if n->nr_slabs > 0, slabs still exist on the node
3640 * that is going down. We were unable to free them,
3641 * and offline_pages() function shouldn't call this
3642 * callback. So, we must fail.
3644 BUG_ON(slabs_node(s
, offline_node
));
3646 s
->node
[offline_node
] = NULL
;
3647 kmem_cache_free(kmem_cache_node
, n
);
3650 mutex_unlock(&slab_mutex
);
3653 static int slab_mem_going_online_callback(void *arg
)
3655 struct kmem_cache_node
*n
;
3656 struct kmem_cache
*s
;
3657 struct memory_notify
*marg
= arg
;
3658 int nid
= marg
->status_change_nid_normal
;
3662 * If the node's memory is already available, then kmem_cache_node is
3663 * already created. Nothing to do.
3669 * We are bringing a node online. No memory is available yet. We must
3670 * allocate a kmem_cache_node structure in order to bring the node
3673 mutex_lock(&slab_mutex
);
3674 list_for_each_entry(s
, &slab_caches
, list
) {
3676 * XXX: kmem_cache_alloc_node will fallback to other nodes
3677 * since memory is not yet available from the node that
3680 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3685 init_kmem_cache_node(n
);
3689 mutex_unlock(&slab_mutex
);
3693 static int slab_memory_callback(struct notifier_block
*self
,
3694 unsigned long action
, void *arg
)
3699 case MEM_GOING_ONLINE
:
3700 ret
= slab_mem_going_online_callback(arg
);
3702 case MEM_GOING_OFFLINE
:
3703 ret
= slab_mem_going_offline_callback(arg
);
3706 case MEM_CANCEL_ONLINE
:
3707 slab_mem_offline_callback(arg
);
3710 case MEM_CANCEL_OFFLINE
:
3714 ret
= notifier_from_errno(ret
);
3720 static struct notifier_block slab_memory_callback_nb
= {
3721 .notifier_call
= slab_memory_callback
,
3722 .priority
= SLAB_CALLBACK_PRI
,
3725 /********************************************************************
3726 * Basic setup of slabs
3727 *******************************************************************/
3730 * Used for early kmem_cache structures that were allocated using
3731 * the page allocator. Allocate them properly then fix up the pointers
3732 * that may be pointing to the wrong kmem_cache structure.
3735 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
3738 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
3739 struct kmem_cache_node
*n
;
3741 memcpy(s
, static_cache
, kmem_cache
->object_size
);
3744 * This runs very early, and only the boot processor is supposed to be
3745 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3748 __flush_cpu_slab(s
, smp_processor_id());
3749 for_each_kmem_cache_node(s
, node
, n
) {
3752 list_for_each_entry(p
, &n
->partial
, lru
)
3755 #ifdef CONFIG_SLUB_DEBUG
3756 list_for_each_entry(p
, &n
->full
, lru
)
3760 slab_init_memcg_params(s
);
3761 list_add(&s
->list
, &slab_caches
);
3765 void __init
kmem_cache_init(void)
3767 static __initdata
struct kmem_cache boot_kmem_cache
,
3768 boot_kmem_cache_node
;
3770 if (debug_guardpage_minorder())
3773 kmem_cache_node
= &boot_kmem_cache_node
;
3774 kmem_cache
= &boot_kmem_cache
;
3776 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
3777 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
3779 register_hotmemory_notifier(&slab_memory_callback_nb
);
3781 /* Able to allocate the per node structures */
3782 slab_state
= PARTIAL
;
3784 create_boot_cache(kmem_cache
, "kmem_cache",
3785 offsetof(struct kmem_cache
, node
) +
3786 nr_node_ids
* sizeof(struct kmem_cache_node
*),
3787 SLAB_HWCACHE_ALIGN
);
3789 kmem_cache
= bootstrap(&boot_kmem_cache
);
3792 * Allocate kmem_cache_node properly from the kmem_cache slab.
3793 * kmem_cache_node is separately allocated so no need to
3794 * update any list pointers.
3796 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
3798 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3799 setup_kmalloc_cache_index_table();
3800 create_kmalloc_caches(0);
3803 register_cpu_notifier(&slab_notifier
);
3806 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3808 slub_min_order
, slub_max_order
, slub_min_objects
,
3809 nr_cpu_ids
, nr_node_ids
);
3812 void __init
kmem_cache_init_late(void)
3817 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
3818 unsigned long flags
, void (*ctor
)(void *))
3820 struct kmem_cache
*s
, *c
;
3822 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
3827 * Adjust the object sizes so that we clear
3828 * the complete object on kzalloc.
3830 s
->object_size
= max(s
->object_size
, (int)size
);
3831 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
3833 for_each_memcg_cache(c
, s
) {
3834 c
->object_size
= s
->object_size
;
3835 c
->inuse
= max_t(int, c
->inuse
,
3836 ALIGN(size
, sizeof(void *)));
3839 if (sysfs_slab_alias(s
, name
)) {
3848 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
3852 err
= kmem_cache_open(s
, flags
);
3856 /* Mutex is not taken during early boot */
3857 if (slab_state
<= UP
)
3860 memcg_propagate_slab_attrs(s
);
3861 err
= sysfs_slab_add(s
);
3863 kmem_cache_close(s
);
3870 * Use the cpu notifier to insure that the cpu slabs are flushed when
3873 static int slab_cpuup_callback(struct notifier_block
*nfb
,
3874 unsigned long action
, void *hcpu
)
3876 long cpu
= (long)hcpu
;
3877 struct kmem_cache
*s
;
3878 unsigned long flags
;
3881 case CPU_UP_CANCELED
:
3882 case CPU_UP_CANCELED_FROZEN
:
3884 case CPU_DEAD_FROZEN
:
3885 mutex_lock(&slab_mutex
);
3886 list_for_each_entry(s
, &slab_caches
, list
) {
3887 local_irq_save(flags
);
3888 __flush_cpu_slab(s
, cpu
);
3889 local_irq_restore(flags
);
3891 mutex_unlock(&slab_mutex
);
3899 static struct notifier_block slab_notifier
= {
3900 .notifier_call
= slab_cpuup_callback
3905 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
3907 struct kmem_cache
*s
;
3910 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3911 return kmalloc_large(size
, gfpflags
);
3913 s
= kmalloc_slab(size
, gfpflags
);
3915 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3918 ret
= slab_alloc(s
, gfpflags
, caller
);
3920 /* Honor the call site pointer we received. */
3921 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
3927 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
3928 int node
, unsigned long caller
)
3930 struct kmem_cache
*s
;
3933 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3934 ret
= kmalloc_large_node(size
, gfpflags
, node
);
3936 trace_kmalloc_node(caller
, ret
,
3937 size
, PAGE_SIZE
<< get_order(size
),
3943 s
= kmalloc_slab(size
, gfpflags
);
3945 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3948 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
3950 /* Honor the call site pointer we received. */
3951 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
3958 static int count_inuse(struct page
*page
)
3963 static int count_total(struct page
*page
)
3965 return page
->objects
;
3969 #ifdef CONFIG_SLUB_DEBUG
3970 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
3974 void *addr
= page_address(page
);
3976 if (!check_slab(s
, page
) ||
3977 !on_freelist(s
, page
, NULL
))
3980 /* Now we know that a valid freelist exists */
3981 bitmap_zero(map
, page
->objects
);
3983 get_map(s
, page
, map
);
3984 for_each_object(p
, s
, addr
, page
->objects
) {
3985 if (test_bit(slab_index(p
, s
, addr
), map
))
3986 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
3990 for_each_object(p
, s
, addr
, page
->objects
)
3991 if (!test_bit(slab_index(p
, s
, addr
), map
))
3992 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
3997 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4001 validate_slab(s
, page
, map
);
4005 static int validate_slab_node(struct kmem_cache
*s
,
4006 struct kmem_cache_node
*n
, unsigned long *map
)
4008 unsigned long count
= 0;
4010 unsigned long flags
;
4012 spin_lock_irqsave(&n
->list_lock
, flags
);
4014 list_for_each_entry(page
, &n
->partial
, lru
) {
4015 validate_slab_slab(s
, page
, map
);
4018 if (count
!= n
->nr_partial
)
4019 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4020 s
->name
, count
, n
->nr_partial
);
4022 if (!(s
->flags
& SLAB_STORE_USER
))
4025 list_for_each_entry(page
, &n
->full
, lru
) {
4026 validate_slab_slab(s
, page
, map
);
4029 if (count
!= atomic_long_read(&n
->nr_slabs
))
4030 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4031 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4034 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4038 static long validate_slab_cache(struct kmem_cache
*s
)
4041 unsigned long count
= 0;
4042 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4043 sizeof(unsigned long), GFP_KERNEL
);
4044 struct kmem_cache_node
*n
;
4050 for_each_kmem_cache_node(s
, node
, n
)
4051 count
+= validate_slab_node(s
, n
, map
);
4056 * Generate lists of code addresses where slabcache objects are allocated
4061 unsigned long count
;
4068 DECLARE_BITMAP(cpus
, NR_CPUS
);
4074 unsigned long count
;
4075 struct location
*loc
;
4078 static void free_loc_track(struct loc_track
*t
)
4081 free_pages((unsigned long)t
->loc
,
4082 get_order(sizeof(struct location
) * t
->max
));
4085 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4090 order
= get_order(sizeof(struct location
) * max
);
4092 l
= (void *)__get_free_pages(flags
, order
);
4097 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4105 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4106 const struct track
*track
)
4108 long start
, end
, pos
;
4110 unsigned long caddr
;
4111 unsigned long age
= jiffies
- track
->when
;
4117 pos
= start
+ (end
- start
+ 1) / 2;
4120 * There is nothing at "end". If we end up there
4121 * we need to add something to before end.
4126 caddr
= t
->loc
[pos
].addr
;
4127 if (track
->addr
== caddr
) {
4133 if (age
< l
->min_time
)
4135 if (age
> l
->max_time
)
4138 if (track
->pid
< l
->min_pid
)
4139 l
->min_pid
= track
->pid
;
4140 if (track
->pid
> l
->max_pid
)
4141 l
->max_pid
= track
->pid
;
4143 cpumask_set_cpu(track
->cpu
,
4144 to_cpumask(l
->cpus
));
4146 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4150 if (track
->addr
< caddr
)
4157 * Not found. Insert new tracking element.
4159 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4165 (t
->count
- pos
) * sizeof(struct location
));
4168 l
->addr
= track
->addr
;
4172 l
->min_pid
= track
->pid
;
4173 l
->max_pid
= track
->pid
;
4174 cpumask_clear(to_cpumask(l
->cpus
));
4175 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4176 nodes_clear(l
->nodes
);
4177 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4181 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4182 struct page
*page
, enum track_item alloc
,
4185 void *addr
= page_address(page
);
4188 bitmap_zero(map
, page
->objects
);
4189 get_map(s
, page
, map
);
4191 for_each_object(p
, s
, addr
, page
->objects
)
4192 if (!test_bit(slab_index(p
, s
, addr
), map
))
4193 add_location(t
, s
, get_track(s
, p
, alloc
));
4196 static int list_locations(struct kmem_cache
*s
, char *buf
,
4197 enum track_item alloc
)
4201 struct loc_track t
= { 0, 0, NULL
};
4203 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4204 sizeof(unsigned long), GFP_KERNEL
);
4205 struct kmem_cache_node
*n
;
4207 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4210 return sprintf(buf
, "Out of memory\n");
4212 /* Push back cpu slabs */
4215 for_each_kmem_cache_node(s
, node
, n
) {
4216 unsigned long flags
;
4219 if (!atomic_long_read(&n
->nr_slabs
))
4222 spin_lock_irqsave(&n
->list_lock
, flags
);
4223 list_for_each_entry(page
, &n
->partial
, lru
)
4224 process_slab(&t
, s
, page
, alloc
, map
);
4225 list_for_each_entry(page
, &n
->full
, lru
)
4226 process_slab(&t
, s
, page
, alloc
, map
);
4227 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4230 for (i
= 0; i
< t
.count
; i
++) {
4231 struct location
*l
= &t
.loc
[i
];
4233 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4235 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4238 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4240 len
+= sprintf(buf
+ len
, "<not-available>");
4242 if (l
->sum_time
!= l
->min_time
) {
4243 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4245 (long)div_u64(l
->sum_time
, l
->count
),
4248 len
+= sprintf(buf
+ len
, " age=%ld",
4251 if (l
->min_pid
!= l
->max_pid
)
4252 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4253 l
->min_pid
, l
->max_pid
);
4255 len
+= sprintf(buf
+ len
, " pid=%ld",
4258 if (num_online_cpus() > 1 &&
4259 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4260 len
< PAGE_SIZE
- 60)
4261 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4263 cpumask_pr_args(to_cpumask(l
->cpus
)));
4265 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4266 len
< PAGE_SIZE
- 60)
4267 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4269 nodemask_pr_args(&l
->nodes
));
4271 len
+= sprintf(buf
+ len
, "\n");
4277 len
+= sprintf(buf
, "No data\n");
4282 #ifdef SLUB_RESILIENCY_TEST
4283 static void __init
resiliency_test(void)
4287 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4289 pr_err("SLUB resiliency testing\n");
4290 pr_err("-----------------------\n");
4291 pr_err("A. Corruption after allocation\n");
4293 p
= kzalloc(16, GFP_KERNEL
);
4295 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4298 validate_slab_cache(kmalloc_caches
[4]);
4300 /* Hmmm... The next two are dangerous */
4301 p
= kzalloc(32, GFP_KERNEL
);
4302 p
[32 + sizeof(void *)] = 0x34;
4303 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4305 pr_err("If allocated object is overwritten then not detectable\n\n");
4307 validate_slab_cache(kmalloc_caches
[5]);
4308 p
= kzalloc(64, GFP_KERNEL
);
4309 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4311 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4313 pr_err("If allocated object is overwritten then not detectable\n\n");
4314 validate_slab_cache(kmalloc_caches
[6]);
4316 pr_err("\nB. Corruption after free\n");
4317 p
= kzalloc(128, GFP_KERNEL
);
4320 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4321 validate_slab_cache(kmalloc_caches
[7]);
4323 p
= kzalloc(256, GFP_KERNEL
);
4326 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4327 validate_slab_cache(kmalloc_caches
[8]);
4329 p
= kzalloc(512, GFP_KERNEL
);
4332 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4333 validate_slab_cache(kmalloc_caches
[9]);
4337 static void resiliency_test(void) {};
4342 enum slab_stat_type
{
4343 SL_ALL
, /* All slabs */
4344 SL_PARTIAL
, /* Only partially allocated slabs */
4345 SL_CPU
, /* Only slabs used for cpu caches */
4346 SL_OBJECTS
, /* Determine allocated objects not slabs */
4347 SL_TOTAL
/* Determine object capacity not slabs */
4350 #define SO_ALL (1 << SL_ALL)
4351 #define SO_PARTIAL (1 << SL_PARTIAL)
4352 #define SO_CPU (1 << SL_CPU)
4353 #define SO_OBJECTS (1 << SL_OBJECTS)
4354 #define SO_TOTAL (1 << SL_TOTAL)
4356 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4357 char *buf
, unsigned long flags
)
4359 unsigned long total
= 0;
4362 unsigned long *nodes
;
4364 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4368 if (flags
& SO_CPU
) {
4371 for_each_possible_cpu(cpu
) {
4372 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4377 page
= READ_ONCE(c
->page
);
4381 node
= page_to_nid(page
);
4382 if (flags
& SO_TOTAL
)
4384 else if (flags
& SO_OBJECTS
)
4392 page
= READ_ONCE(c
->partial
);
4394 node
= page_to_nid(page
);
4395 if (flags
& SO_TOTAL
)
4397 else if (flags
& SO_OBJECTS
)
4408 #ifdef CONFIG_SLUB_DEBUG
4409 if (flags
& SO_ALL
) {
4410 struct kmem_cache_node
*n
;
4412 for_each_kmem_cache_node(s
, node
, n
) {
4414 if (flags
& SO_TOTAL
)
4415 x
= atomic_long_read(&n
->total_objects
);
4416 else if (flags
& SO_OBJECTS
)
4417 x
= atomic_long_read(&n
->total_objects
) -
4418 count_partial(n
, count_free
);
4420 x
= atomic_long_read(&n
->nr_slabs
);
4427 if (flags
& SO_PARTIAL
) {
4428 struct kmem_cache_node
*n
;
4430 for_each_kmem_cache_node(s
, node
, n
) {
4431 if (flags
& SO_TOTAL
)
4432 x
= count_partial(n
, count_total
);
4433 else if (flags
& SO_OBJECTS
)
4434 x
= count_partial(n
, count_inuse
);
4441 x
= sprintf(buf
, "%lu", total
);
4443 for (node
= 0; node
< nr_node_ids
; node
++)
4445 x
+= sprintf(buf
+ x
, " N%d=%lu",
4450 return x
+ sprintf(buf
+ x
, "\n");
4453 #ifdef CONFIG_SLUB_DEBUG
4454 static int any_slab_objects(struct kmem_cache
*s
)
4457 struct kmem_cache_node
*n
;
4459 for_each_kmem_cache_node(s
, node
, n
)
4460 if (atomic_long_read(&n
->total_objects
))
4467 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4468 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4470 struct slab_attribute
{
4471 struct attribute attr
;
4472 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4473 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4476 #define SLAB_ATTR_RO(_name) \
4477 static struct slab_attribute _name##_attr = \
4478 __ATTR(_name, 0400, _name##_show, NULL)
4480 #define SLAB_ATTR(_name) \
4481 static struct slab_attribute _name##_attr = \
4482 __ATTR(_name, 0600, _name##_show, _name##_store)
4484 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4486 return sprintf(buf
, "%d\n", s
->size
);
4488 SLAB_ATTR_RO(slab_size
);
4490 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4492 return sprintf(buf
, "%d\n", s
->align
);
4494 SLAB_ATTR_RO(align
);
4496 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4498 return sprintf(buf
, "%d\n", s
->object_size
);
4500 SLAB_ATTR_RO(object_size
);
4502 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4504 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4506 SLAB_ATTR_RO(objs_per_slab
);
4508 static ssize_t
order_store(struct kmem_cache
*s
,
4509 const char *buf
, size_t length
)
4511 unsigned long order
;
4514 err
= kstrtoul(buf
, 10, &order
);
4518 if (order
> slub_max_order
|| order
< slub_min_order
)
4521 calculate_sizes(s
, order
);
4525 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4527 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4531 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4533 return sprintf(buf
, "%lu\n", s
->min_partial
);
4536 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4542 err
= kstrtoul(buf
, 10, &min
);
4546 set_min_partial(s
, min
);
4549 SLAB_ATTR(min_partial
);
4551 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4553 return sprintf(buf
, "%u\n", s
->cpu_partial
);
4556 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4559 unsigned long objects
;
4562 err
= kstrtoul(buf
, 10, &objects
);
4565 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4568 s
->cpu_partial
= objects
;
4572 SLAB_ATTR(cpu_partial
);
4574 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4578 return sprintf(buf
, "%pS\n", s
->ctor
);
4582 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4584 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4586 SLAB_ATTR_RO(aliases
);
4588 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4590 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4592 SLAB_ATTR_RO(partial
);
4594 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4596 return show_slab_objects(s
, buf
, SO_CPU
);
4598 SLAB_ATTR_RO(cpu_slabs
);
4600 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4602 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4604 SLAB_ATTR_RO(objects
);
4606 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4608 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4610 SLAB_ATTR_RO(objects_partial
);
4612 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4619 for_each_online_cpu(cpu
) {
4620 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
)->partial
;
4623 pages
+= page
->pages
;
4624 objects
+= page
->pobjects
;
4628 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
4631 for_each_online_cpu(cpu
) {
4632 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
) ->partial
;
4634 if (page
&& len
< PAGE_SIZE
- 20)
4635 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
4636 page
->pobjects
, page
->pages
);
4639 return len
+ sprintf(buf
+ len
, "\n");
4641 SLAB_ATTR_RO(slabs_cpu_partial
);
4643 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4645 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4648 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4649 const char *buf
, size_t length
)
4651 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4653 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4656 SLAB_ATTR(reclaim_account
);
4658 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4660 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4662 SLAB_ATTR_RO(hwcache_align
);
4664 #ifdef CONFIG_ZONE_DMA
4665 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4667 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4669 SLAB_ATTR_RO(cache_dma
);
4672 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4674 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4676 SLAB_ATTR_RO(destroy_by_rcu
);
4678 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4680 return sprintf(buf
, "%d\n", s
->reserved
);
4682 SLAB_ATTR_RO(reserved
);
4684 #ifdef CONFIG_SLUB_DEBUG
4685 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4687 return show_slab_objects(s
, buf
, SO_ALL
);
4689 SLAB_ATTR_RO(slabs
);
4691 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4693 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4695 SLAB_ATTR_RO(total_objects
);
4697 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4699 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DEBUG_FREE
));
4702 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
4703 const char *buf
, size_t length
)
4705 s
->flags
&= ~SLAB_DEBUG_FREE
;
4706 if (buf
[0] == '1') {
4707 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4708 s
->flags
|= SLAB_DEBUG_FREE
;
4712 SLAB_ATTR(sanity_checks
);
4714 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
4716 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
4719 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
4723 * Tracing a merged cache is going to give confusing results
4724 * as well as cause other issues like converting a mergeable
4725 * cache into an umergeable one.
4727 if (s
->refcount
> 1)
4730 s
->flags
&= ~SLAB_TRACE
;
4731 if (buf
[0] == '1') {
4732 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4733 s
->flags
|= SLAB_TRACE
;
4739 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
4741 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
4744 static ssize_t
red_zone_store(struct kmem_cache
*s
,
4745 const char *buf
, size_t length
)
4747 if (any_slab_objects(s
))
4750 s
->flags
&= ~SLAB_RED_ZONE
;
4751 if (buf
[0] == '1') {
4752 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4753 s
->flags
|= SLAB_RED_ZONE
;
4755 calculate_sizes(s
, -1);
4758 SLAB_ATTR(red_zone
);
4760 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
4762 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
4765 static ssize_t
poison_store(struct kmem_cache
*s
,
4766 const char *buf
, size_t length
)
4768 if (any_slab_objects(s
))
4771 s
->flags
&= ~SLAB_POISON
;
4772 if (buf
[0] == '1') {
4773 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4774 s
->flags
|= SLAB_POISON
;
4776 calculate_sizes(s
, -1);
4781 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
4783 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
4786 static ssize_t
store_user_store(struct kmem_cache
*s
,
4787 const char *buf
, size_t length
)
4789 if (any_slab_objects(s
))
4792 s
->flags
&= ~SLAB_STORE_USER
;
4793 if (buf
[0] == '1') {
4794 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4795 s
->flags
|= SLAB_STORE_USER
;
4797 calculate_sizes(s
, -1);
4800 SLAB_ATTR(store_user
);
4802 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
4807 static ssize_t
validate_store(struct kmem_cache
*s
,
4808 const char *buf
, size_t length
)
4812 if (buf
[0] == '1') {
4813 ret
= validate_slab_cache(s
);
4819 SLAB_ATTR(validate
);
4821 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
4823 if (!(s
->flags
& SLAB_STORE_USER
))
4825 return list_locations(s
, buf
, TRACK_ALLOC
);
4827 SLAB_ATTR_RO(alloc_calls
);
4829 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
4831 if (!(s
->flags
& SLAB_STORE_USER
))
4833 return list_locations(s
, buf
, TRACK_FREE
);
4835 SLAB_ATTR_RO(free_calls
);
4836 #endif /* CONFIG_SLUB_DEBUG */
4838 #ifdef CONFIG_FAILSLAB
4839 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
4841 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
4844 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
4847 if (s
->refcount
> 1)
4850 s
->flags
&= ~SLAB_FAILSLAB
;
4852 s
->flags
|= SLAB_FAILSLAB
;
4855 SLAB_ATTR(failslab
);
4858 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
4863 static ssize_t
shrink_store(struct kmem_cache
*s
,
4864 const char *buf
, size_t length
)
4867 kmem_cache_shrink(s
);
4875 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
4877 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
4880 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
4881 const char *buf
, size_t length
)
4883 unsigned long ratio
;
4886 err
= kstrtoul(buf
, 10, &ratio
);
4891 s
->remote_node_defrag_ratio
= ratio
* 10;
4895 SLAB_ATTR(remote_node_defrag_ratio
);
4898 #ifdef CONFIG_SLUB_STATS
4899 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
4901 unsigned long sum
= 0;
4904 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
4909 for_each_online_cpu(cpu
) {
4910 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
4916 len
= sprintf(buf
, "%lu", sum
);
4919 for_each_online_cpu(cpu
) {
4920 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
4921 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
4925 return len
+ sprintf(buf
+ len
, "\n");
4928 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
4932 for_each_online_cpu(cpu
)
4933 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
4936 #define STAT_ATTR(si, text) \
4937 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4939 return show_stat(s, buf, si); \
4941 static ssize_t text##_store(struct kmem_cache *s, \
4942 const char *buf, size_t length) \
4944 if (buf[0] != '0') \
4946 clear_stat(s, si); \
4951 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4952 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
4953 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
4954 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
4955 STAT_ATTR(FREE_FROZEN
, free_frozen
);
4956 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
4957 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
4958 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
4959 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
4960 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
4961 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
4962 STAT_ATTR(FREE_SLAB
, free_slab
);
4963 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
4964 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
4965 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
4966 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
4967 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
4968 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
4969 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
4970 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
4971 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
4972 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
4973 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
4974 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
4975 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
4976 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
4979 static struct attribute
*slab_attrs
[] = {
4980 &slab_size_attr
.attr
,
4981 &object_size_attr
.attr
,
4982 &objs_per_slab_attr
.attr
,
4984 &min_partial_attr
.attr
,
4985 &cpu_partial_attr
.attr
,
4987 &objects_partial_attr
.attr
,
4989 &cpu_slabs_attr
.attr
,
4993 &hwcache_align_attr
.attr
,
4994 &reclaim_account_attr
.attr
,
4995 &destroy_by_rcu_attr
.attr
,
4997 &reserved_attr
.attr
,
4998 &slabs_cpu_partial_attr
.attr
,
4999 #ifdef CONFIG_SLUB_DEBUG
5000 &total_objects_attr
.attr
,
5002 &sanity_checks_attr
.attr
,
5004 &red_zone_attr
.attr
,
5006 &store_user_attr
.attr
,
5007 &validate_attr
.attr
,
5008 &alloc_calls_attr
.attr
,
5009 &free_calls_attr
.attr
,
5011 #ifdef CONFIG_ZONE_DMA
5012 &cache_dma_attr
.attr
,
5015 &remote_node_defrag_ratio_attr
.attr
,
5017 #ifdef CONFIG_SLUB_STATS
5018 &alloc_fastpath_attr
.attr
,
5019 &alloc_slowpath_attr
.attr
,
5020 &free_fastpath_attr
.attr
,
5021 &free_slowpath_attr
.attr
,
5022 &free_frozen_attr
.attr
,
5023 &free_add_partial_attr
.attr
,
5024 &free_remove_partial_attr
.attr
,
5025 &alloc_from_partial_attr
.attr
,
5026 &alloc_slab_attr
.attr
,
5027 &alloc_refill_attr
.attr
,
5028 &alloc_node_mismatch_attr
.attr
,
5029 &free_slab_attr
.attr
,
5030 &cpuslab_flush_attr
.attr
,
5031 &deactivate_full_attr
.attr
,
5032 &deactivate_empty_attr
.attr
,
5033 &deactivate_to_head_attr
.attr
,
5034 &deactivate_to_tail_attr
.attr
,
5035 &deactivate_remote_frees_attr
.attr
,
5036 &deactivate_bypass_attr
.attr
,
5037 &order_fallback_attr
.attr
,
5038 &cmpxchg_double_fail_attr
.attr
,
5039 &cmpxchg_double_cpu_fail_attr
.attr
,
5040 &cpu_partial_alloc_attr
.attr
,
5041 &cpu_partial_free_attr
.attr
,
5042 &cpu_partial_node_attr
.attr
,
5043 &cpu_partial_drain_attr
.attr
,
5045 #ifdef CONFIG_FAILSLAB
5046 &failslab_attr
.attr
,
5052 static struct attribute_group slab_attr_group
= {
5053 .attrs
= slab_attrs
,
5056 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5057 struct attribute
*attr
,
5060 struct slab_attribute
*attribute
;
5061 struct kmem_cache
*s
;
5064 attribute
= to_slab_attr(attr
);
5067 if (!attribute
->show
)
5070 err
= attribute
->show(s
, buf
);
5075 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5076 struct attribute
*attr
,
5077 const char *buf
, size_t len
)
5079 struct slab_attribute
*attribute
;
5080 struct kmem_cache
*s
;
5083 attribute
= to_slab_attr(attr
);
5086 if (!attribute
->store
)
5089 err
= attribute
->store(s
, buf
, len
);
5090 #ifdef CONFIG_MEMCG_KMEM
5091 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5092 struct kmem_cache
*c
;
5094 mutex_lock(&slab_mutex
);
5095 if (s
->max_attr_size
< len
)
5096 s
->max_attr_size
= len
;
5099 * This is a best effort propagation, so this function's return
5100 * value will be determined by the parent cache only. This is
5101 * basically because not all attributes will have a well
5102 * defined semantics for rollbacks - most of the actions will
5103 * have permanent effects.
5105 * Returning the error value of any of the children that fail
5106 * is not 100 % defined, in the sense that users seeing the
5107 * error code won't be able to know anything about the state of
5110 * Only returning the error code for the parent cache at least
5111 * has well defined semantics. The cache being written to
5112 * directly either failed or succeeded, in which case we loop
5113 * through the descendants with best-effort propagation.
5115 for_each_memcg_cache(c
, s
)
5116 attribute
->store(c
, buf
, len
);
5117 mutex_unlock(&slab_mutex
);
5123 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5125 #ifdef CONFIG_MEMCG_KMEM
5127 char *buffer
= NULL
;
5128 struct kmem_cache
*root_cache
;
5130 if (is_root_cache(s
))
5133 root_cache
= s
->memcg_params
.root_cache
;
5136 * This mean this cache had no attribute written. Therefore, no point
5137 * in copying default values around
5139 if (!root_cache
->max_attr_size
)
5142 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5145 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5147 if (!attr
|| !attr
->store
|| !attr
->show
)
5151 * It is really bad that we have to allocate here, so we will
5152 * do it only as a fallback. If we actually allocate, though,
5153 * we can just use the allocated buffer until the end.
5155 * Most of the slub attributes will tend to be very small in
5156 * size, but sysfs allows buffers up to a page, so they can
5157 * theoretically happen.
5161 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5164 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5165 if (WARN_ON(!buffer
))
5170 attr
->show(root_cache
, buf
);
5171 attr
->store(s
, buf
, strlen(buf
));
5175 free_page((unsigned long)buffer
);
5179 static void kmem_cache_release(struct kobject
*k
)
5181 slab_kmem_cache_release(to_slab(k
));
5184 static const struct sysfs_ops slab_sysfs_ops
= {
5185 .show
= slab_attr_show
,
5186 .store
= slab_attr_store
,
5189 static struct kobj_type slab_ktype
= {
5190 .sysfs_ops
= &slab_sysfs_ops
,
5191 .release
= kmem_cache_release
,
5194 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5196 struct kobj_type
*ktype
= get_ktype(kobj
);
5198 if (ktype
== &slab_ktype
)
5203 static const struct kset_uevent_ops slab_uevent_ops
= {
5204 .filter
= uevent_filter
,
5207 static struct kset
*slab_kset
;
5209 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5211 #ifdef CONFIG_MEMCG_KMEM
5212 if (!is_root_cache(s
))
5213 return s
->memcg_params
.root_cache
->memcg_kset
;
5218 #define ID_STR_LENGTH 64
5220 /* Create a unique string id for a slab cache:
5222 * Format :[flags-]size
5224 static char *create_unique_id(struct kmem_cache
*s
)
5226 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5233 * First flags affecting slabcache operations. We will only
5234 * get here for aliasable slabs so we do not need to support
5235 * too many flags. The flags here must cover all flags that
5236 * are matched during merging to guarantee that the id is
5239 if (s
->flags
& SLAB_CACHE_DMA
)
5241 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5243 if (s
->flags
& SLAB_DEBUG_FREE
)
5245 if (!(s
->flags
& SLAB_NOTRACK
))
5249 p
+= sprintf(p
, "%07d", s
->size
);
5251 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5255 static int sysfs_slab_add(struct kmem_cache
*s
)
5259 int unmergeable
= slab_unmergeable(s
);
5263 * Slabcache can never be merged so we can use the name proper.
5264 * This is typically the case for debug situations. In that
5265 * case we can catch duplicate names easily.
5267 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5271 * Create a unique name for the slab as a target
5274 name
= create_unique_id(s
);
5277 s
->kobj
.kset
= cache_kset(s
);
5278 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5282 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5286 #ifdef CONFIG_MEMCG_KMEM
5287 if (is_root_cache(s
)) {
5288 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5289 if (!s
->memcg_kset
) {
5296 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5298 /* Setup first alias */
5299 sysfs_slab_alias(s
, s
->name
);
5306 kobject_del(&s
->kobj
);
5310 void sysfs_slab_remove(struct kmem_cache
*s
)
5312 if (slab_state
< FULL
)
5314 * Sysfs has not been setup yet so no need to remove the
5319 #ifdef CONFIG_MEMCG_KMEM
5320 kset_unregister(s
->memcg_kset
);
5322 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5323 kobject_del(&s
->kobj
);
5324 kobject_put(&s
->kobj
);
5328 * Need to buffer aliases during bootup until sysfs becomes
5329 * available lest we lose that information.
5331 struct saved_alias
{
5332 struct kmem_cache
*s
;
5334 struct saved_alias
*next
;
5337 static struct saved_alias
*alias_list
;
5339 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5341 struct saved_alias
*al
;
5343 if (slab_state
== FULL
) {
5345 * If we have a leftover link then remove it.
5347 sysfs_remove_link(&slab_kset
->kobj
, name
);
5348 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5351 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5357 al
->next
= alias_list
;
5362 static int __init
slab_sysfs_init(void)
5364 struct kmem_cache
*s
;
5367 mutex_lock(&slab_mutex
);
5369 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5371 mutex_unlock(&slab_mutex
);
5372 pr_err("Cannot register slab subsystem.\n");
5378 list_for_each_entry(s
, &slab_caches
, list
) {
5379 err
= sysfs_slab_add(s
);
5381 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5385 while (alias_list
) {
5386 struct saved_alias
*al
= alias_list
;
5388 alias_list
= alias_list
->next
;
5389 err
= sysfs_slab_alias(al
->s
, al
->name
);
5391 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5396 mutex_unlock(&slab_mutex
);
5401 __initcall(slab_sysfs_init
);
5402 #endif /* CONFIG_SYSFS */
5405 * The /proc/slabinfo ABI
5407 #ifdef CONFIG_SLABINFO
5408 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5410 unsigned long nr_slabs
= 0;
5411 unsigned long nr_objs
= 0;
5412 unsigned long nr_free
= 0;
5414 struct kmem_cache_node
*n
;
5416 for_each_kmem_cache_node(s
, node
, n
) {
5417 nr_slabs
+= node_nr_slabs(n
);
5418 nr_objs
+= node_nr_objs(n
);
5419 nr_free
+= count_partial(n
, count_free
);
5422 sinfo
->active_objs
= nr_objs
- nr_free
;
5423 sinfo
->num_objs
= nr_objs
;
5424 sinfo
->active_slabs
= nr_slabs
;
5425 sinfo
->num_slabs
= nr_slabs
;
5426 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5427 sinfo
->cache_order
= oo_order(s
->oo
);
5430 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5434 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5435 size_t count
, loff_t
*ppos
)
5439 #endif /* CONFIG_SLABINFO */