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 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
129 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
130 p
+= s
->red_left_pad
;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s
);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
194 #define __OBJECT_POISON 0x80000000UL /* Poison object */
195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
198 static struct notifier_block slab_notifier
;
202 * Tracking user of a slab.
204 #define TRACK_ADDRS_COUNT 16
206 unsigned long addr
; /* Called from address */
207 #ifdef CONFIG_STACKTRACE
208 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
210 int cpu
; /* Was running on cpu */
211 int pid
; /* Pid context */
212 unsigned long when
; /* When did the operation occur */
215 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
218 static int sysfs_slab_add(struct kmem_cache
*);
219 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
220 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
222 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
228 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
243 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
245 return *(void **)(object
+ s
->offset
);
248 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
250 prefetch(object
+ s
->offset
);
253 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
257 if (!debug_pagealloc_enabled())
258 return get_freepointer(s
, object
);
260 probe_kernel_read(&p
, (void **)(object
+ s
->offset
), sizeof(p
));
264 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
266 *(void **)(object
+ s
->offset
) = fp
;
269 /* Loop over all objects in a slab */
270 #define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = fixup_red_left(__s, __addr); \
272 __p < (__addr) + (__objects) * (__s)->size; \
275 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
276 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
277 __idx <= __objects; \
278 __p += (__s)->size, __idx++)
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
283 return (p
- addr
) / s
->size
;
286 static inline int order_objects(int order
, unsigned long size
, int reserved
)
288 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
291 static inline struct kmem_cache_order_objects
oo_make(int order
,
292 unsigned long size
, int reserved
)
294 struct kmem_cache_order_objects x
= {
295 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
301 static inline int oo_order(struct kmem_cache_order_objects x
)
303 return x
.x
>> OO_SHIFT
;
306 static inline int oo_objects(struct kmem_cache_order_objects x
)
308 return x
.x
& OO_MASK
;
312 * Per slab locking using the pagelock
314 static __always_inline
void slab_lock(struct page
*page
)
316 VM_BUG_ON_PAGE(PageTail(page
), page
);
317 bit_spin_lock(PG_locked
, &page
->flags
);
320 static __always_inline
void slab_unlock(struct page
*page
)
322 VM_BUG_ON_PAGE(PageTail(page
), page
);
323 __bit_spin_unlock(PG_locked
, &page
->flags
);
326 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
329 tmp
.counters
= counters_new
;
331 * page->counters can cover frozen/inuse/objects as well
332 * as page->_refcount. If we assign to ->counters directly
333 * we run the risk of losing updates to page->_refcount, so
334 * be careful and only assign to the fields we need.
336 page
->frozen
= tmp
.frozen
;
337 page
->inuse
= tmp
.inuse
;
338 page
->objects
= tmp
.objects
;
341 /* Interrupts must be disabled (for the fallback code to work right) */
342 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
343 void *freelist_old
, unsigned long counters_old
,
344 void *freelist_new
, unsigned long counters_new
,
347 VM_BUG_ON(!irqs_disabled());
348 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
349 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
350 if (s
->flags
& __CMPXCHG_DOUBLE
) {
351 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
352 freelist_old
, counters_old
,
353 freelist_new
, counters_new
))
359 if (page
->freelist
== freelist_old
&&
360 page
->counters
== counters_old
) {
361 page
->freelist
= freelist_new
;
362 set_page_slub_counters(page
, counters_new
);
370 stat(s
, CMPXCHG_DOUBLE_FAIL
);
372 #ifdef SLUB_DEBUG_CMPXCHG
373 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
379 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
380 void *freelist_old
, unsigned long counters_old
,
381 void *freelist_new
, unsigned long counters_new
,
384 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
385 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
386 if (s
->flags
& __CMPXCHG_DOUBLE
) {
387 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
388 freelist_old
, counters_old
,
389 freelist_new
, counters_new
))
396 local_irq_save(flags
);
398 if (page
->freelist
== freelist_old
&&
399 page
->counters
== counters_old
) {
400 page
->freelist
= freelist_new
;
401 set_page_slub_counters(page
, counters_new
);
403 local_irq_restore(flags
);
407 local_irq_restore(flags
);
411 stat(s
, CMPXCHG_DOUBLE_FAIL
);
413 #ifdef SLUB_DEBUG_CMPXCHG
414 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
420 #ifdef CONFIG_SLUB_DEBUG
422 * Determine a map of object in use on a page.
424 * Node listlock must be held to guarantee that the page does
425 * not vanish from under us.
427 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
430 void *addr
= page_address(page
);
432 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
433 set_bit(slab_index(p
, s
, addr
), map
);
436 static inline int size_from_object(struct kmem_cache
*s
)
438 if (s
->flags
& SLAB_RED_ZONE
)
439 return s
->size
- s
->red_left_pad
;
444 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
446 if (s
->flags
& SLAB_RED_ZONE
)
447 p
-= s
->red_left_pad
;
455 #if defined(CONFIG_SLUB_DEBUG_ON)
456 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
457 #elif defined(CONFIG_KASAN)
458 static int slub_debug
= SLAB_STORE_USER
;
460 static int slub_debug
;
463 static char *slub_debug_slabs
;
464 static int disable_higher_order_debug
;
467 * slub is about to manipulate internal object metadata. This memory lies
468 * outside the range of the allocated object, so accessing it would normally
469 * be reported by kasan as a bounds error. metadata_access_enable() is used
470 * to tell kasan that these accesses are OK.
472 static inline void metadata_access_enable(void)
474 kasan_disable_current();
477 static inline void metadata_access_disable(void)
479 kasan_enable_current();
486 /* Verify that a pointer has an address that is valid within a slab page */
487 static inline int check_valid_pointer(struct kmem_cache
*s
,
488 struct page
*page
, void *object
)
495 base
= page_address(page
);
496 object
= restore_red_left(s
, object
);
497 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
498 (object
- base
) % s
->size
) {
505 static void print_section(char *text
, u8
*addr
, unsigned int length
)
507 metadata_access_enable();
508 print_hex_dump(KERN_ERR
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
510 metadata_access_disable();
513 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
514 enum track_item alloc
)
519 p
= object
+ s
->offset
+ sizeof(void *);
521 p
= object
+ s
->inuse
;
526 static void set_track(struct kmem_cache
*s
, void *object
,
527 enum track_item alloc
, unsigned long addr
)
529 struct track
*p
= get_track(s
, object
, alloc
);
532 #ifdef CONFIG_STACKTRACE
533 struct stack_trace trace
;
536 trace
.nr_entries
= 0;
537 trace
.max_entries
= TRACK_ADDRS_COUNT
;
538 trace
.entries
= p
->addrs
;
540 metadata_access_enable();
541 save_stack_trace(&trace
);
542 metadata_access_disable();
544 /* See rant in lockdep.c */
545 if (trace
.nr_entries
!= 0 &&
546 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
549 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
553 p
->cpu
= smp_processor_id();
554 p
->pid
= current
->pid
;
557 memset(p
, 0, sizeof(struct track
));
560 static void init_tracking(struct kmem_cache
*s
, void *object
)
562 if (!(s
->flags
& SLAB_STORE_USER
))
565 set_track(s
, object
, TRACK_FREE
, 0UL);
566 set_track(s
, object
, TRACK_ALLOC
, 0UL);
569 static void print_track(const char *s
, struct track
*t
)
574 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
575 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
576 #ifdef CONFIG_STACKTRACE
579 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
581 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
588 static void print_tracking(struct kmem_cache
*s
, void *object
)
590 if (!(s
->flags
& SLAB_STORE_USER
))
593 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
594 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
597 static void print_page_info(struct page
*page
)
599 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
600 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
604 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
606 struct va_format vaf
;
612 pr_err("=============================================================================\n");
613 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
614 pr_err("-----------------------------------------------------------------------------\n\n");
616 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
620 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
622 struct va_format vaf
;
628 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
632 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
634 unsigned int off
; /* Offset of last byte */
635 u8
*addr
= page_address(page
);
637 print_tracking(s
, p
);
639 print_page_info(page
);
641 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
642 p
, p
- addr
, get_freepointer(s
, p
));
644 if (s
->flags
& SLAB_RED_ZONE
)
645 print_section("Redzone ", p
- s
->red_left_pad
, s
->red_left_pad
);
646 else if (p
> addr
+ 16)
647 print_section("Bytes b4 ", p
- 16, 16);
649 print_section("Object ", p
, min_t(unsigned long, s
->object_size
,
651 if (s
->flags
& SLAB_RED_ZONE
)
652 print_section("Redzone ", p
+ s
->object_size
,
653 s
->inuse
- s
->object_size
);
656 off
= s
->offset
+ sizeof(void *);
660 if (s
->flags
& SLAB_STORE_USER
)
661 off
+= 2 * sizeof(struct track
);
663 if (off
!= size_from_object(s
))
664 /* Beginning of the filler is the free pointer */
665 print_section("Padding ", p
+ off
, size_from_object(s
) - off
);
670 void object_err(struct kmem_cache
*s
, struct page
*page
,
671 u8
*object
, char *reason
)
673 slab_bug(s
, "%s", reason
);
674 print_trailer(s
, page
, object
);
677 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
678 const char *fmt
, ...)
684 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
686 slab_bug(s
, "%s", buf
);
687 print_page_info(page
);
691 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
695 if (s
->flags
& SLAB_RED_ZONE
)
696 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
698 if (s
->flags
& __OBJECT_POISON
) {
699 memset(p
, POISON_FREE
, s
->object_size
- 1);
700 p
[s
->object_size
- 1] = POISON_END
;
703 if (s
->flags
& SLAB_RED_ZONE
)
704 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
707 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
708 void *from
, void *to
)
710 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
711 memset(from
, data
, to
- from
);
714 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
715 u8
*object
, char *what
,
716 u8
*start
, unsigned int value
, unsigned int bytes
)
721 metadata_access_enable();
722 fault
= memchr_inv(start
, value
, bytes
);
723 metadata_access_disable();
728 while (end
> fault
&& end
[-1] == value
)
731 slab_bug(s
, "%s overwritten", what
);
732 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
733 fault
, end
- 1, fault
[0], value
);
734 print_trailer(s
, page
, object
);
736 restore_bytes(s
, what
, value
, fault
, end
);
744 * Bytes of the object to be managed.
745 * If the freepointer may overlay the object then the free
746 * pointer is the first word of the object.
748 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
751 * object + s->object_size
752 * Padding to reach word boundary. This is also used for Redzoning.
753 * Padding is extended by another word if Redzoning is enabled and
754 * object_size == inuse.
756 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
757 * 0xcc (RED_ACTIVE) for objects in use.
760 * Meta data starts here.
762 * A. Free pointer (if we cannot overwrite object on free)
763 * B. Tracking data for SLAB_STORE_USER
764 * C. Padding to reach required alignment boundary or at mininum
765 * one word if debugging is on to be able to detect writes
766 * before the word boundary.
768 * Padding is done using 0x5a (POISON_INUSE)
771 * Nothing is used beyond s->size.
773 * If slabcaches are merged then the object_size and inuse boundaries are mostly
774 * ignored. And therefore no slab options that rely on these boundaries
775 * may be used with merged slabcaches.
778 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
780 unsigned long off
= s
->inuse
; /* The end of info */
783 /* Freepointer is placed after the object. */
784 off
+= sizeof(void *);
786 if (s
->flags
& SLAB_STORE_USER
)
787 /* We also have user information there */
788 off
+= 2 * sizeof(struct track
);
790 if (size_from_object(s
) == off
)
793 return check_bytes_and_report(s
, page
, p
, "Object padding",
794 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
797 /* Check the pad bytes at the end of a slab page */
798 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
806 if (!(s
->flags
& SLAB_POISON
))
809 start
= page_address(page
);
810 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
811 end
= start
+ length
;
812 remainder
= length
% s
->size
;
816 metadata_access_enable();
817 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
818 metadata_access_disable();
821 while (end
> fault
&& end
[-1] == POISON_INUSE
)
824 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
825 print_section("Padding ", end
- remainder
, remainder
);
827 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
831 static int check_object(struct kmem_cache
*s
, struct page
*page
,
832 void *object
, u8 val
)
835 u8
*endobject
= object
+ s
->object_size
;
837 if (s
->flags
& SLAB_RED_ZONE
) {
838 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
839 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
842 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
843 endobject
, val
, s
->inuse
- s
->object_size
))
846 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
847 check_bytes_and_report(s
, page
, p
, "Alignment padding",
848 endobject
, POISON_INUSE
,
849 s
->inuse
- s
->object_size
);
853 if (s
->flags
& SLAB_POISON
) {
854 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
855 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
856 POISON_FREE
, s
->object_size
- 1) ||
857 !check_bytes_and_report(s
, page
, p
, "Poison",
858 p
+ s
->object_size
- 1, POISON_END
, 1)))
861 * check_pad_bytes cleans up on its own.
863 check_pad_bytes(s
, page
, p
);
866 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
868 * Object and freepointer overlap. Cannot check
869 * freepointer while object is allocated.
873 /* Check free pointer validity */
874 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
875 object_err(s
, page
, p
, "Freepointer corrupt");
877 * No choice but to zap it and thus lose the remainder
878 * of the free objects in this slab. May cause
879 * another error because the object count is now wrong.
881 set_freepointer(s
, p
, NULL
);
887 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
891 VM_BUG_ON(!irqs_disabled());
893 if (!PageSlab(page
)) {
894 slab_err(s
, page
, "Not a valid slab page");
898 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
899 if (page
->objects
> maxobj
) {
900 slab_err(s
, page
, "objects %u > max %u",
901 page
->objects
, maxobj
);
904 if (page
->inuse
> page
->objects
) {
905 slab_err(s
, page
, "inuse %u > max %u",
906 page
->inuse
, page
->objects
);
909 /* Slab_pad_check fixes things up after itself */
910 slab_pad_check(s
, page
);
915 * Determine if a certain object on a page is on the freelist. Must hold the
916 * slab lock to guarantee that the chains are in a consistent state.
918 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
926 while (fp
&& nr
<= page
->objects
) {
929 if (!check_valid_pointer(s
, page
, fp
)) {
931 object_err(s
, page
, object
,
932 "Freechain corrupt");
933 set_freepointer(s
, object
, NULL
);
935 slab_err(s
, page
, "Freepointer corrupt");
936 page
->freelist
= NULL
;
937 page
->inuse
= page
->objects
;
938 slab_fix(s
, "Freelist cleared");
944 fp
= get_freepointer(s
, object
);
948 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
949 if (max_objects
> MAX_OBJS_PER_PAGE
)
950 max_objects
= MAX_OBJS_PER_PAGE
;
952 if (page
->objects
!= max_objects
) {
953 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
954 page
->objects
, max_objects
);
955 page
->objects
= max_objects
;
956 slab_fix(s
, "Number of objects adjusted.");
958 if (page
->inuse
!= page
->objects
- nr
) {
959 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
960 page
->inuse
, page
->objects
- nr
);
961 page
->inuse
= page
->objects
- nr
;
962 slab_fix(s
, "Object count adjusted.");
964 return search
== NULL
;
967 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
970 if (s
->flags
& SLAB_TRACE
) {
971 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
973 alloc
? "alloc" : "free",
978 print_section("Object ", (void *)object
,
986 * Tracking of fully allocated slabs for debugging purposes.
988 static void add_full(struct kmem_cache
*s
,
989 struct kmem_cache_node
*n
, struct page
*page
)
991 if (!(s
->flags
& SLAB_STORE_USER
))
994 lockdep_assert_held(&n
->list_lock
);
995 list_add(&page
->lru
, &n
->full
);
998 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1000 if (!(s
->flags
& SLAB_STORE_USER
))
1003 lockdep_assert_held(&n
->list_lock
);
1004 list_del(&page
->lru
);
1007 /* Tracking of the number of slabs for debugging purposes */
1008 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1010 struct kmem_cache_node
*n
= get_node(s
, node
);
1012 return atomic_long_read(&n
->nr_slabs
);
1015 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1017 return atomic_long_read(&n
->nr_slabs
);
1020 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1022 struct kmem_cache_node
*n
= get_node(s
, node
);
1025 * May be called early in order to allocate a slab for the
1026 * kmem_cache_node structure. Solve the chicken-egg
1027 * dilemma by deferring the increment of the count during
1028 * bootstrap (see early_kmem_cache_node_alloc).
1031 atomic_long_inc(&n
->nr_slabs
);
1032 atomic_long_add(objects
, &n
->total_objects
);
1035 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1037 struct kmem_cache_node
*n
= get_node(s
, node
);
1039 atomic_long_dec(&n
->nr_slabs
);
1040 atomic_long_sub(objects
, &n
->total_objects
);
1043 /* Object debug checks for alloc/free paths */
1044 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1047 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1050 init_object(s
, object
, SLUB_RED_INACTIVE
);
1051 init_tracking(s
, object
);
1054 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1056 void *object
, unsigned long addr
)
1058 if (!check_slab(s
, page
))
1061 if (!check_valid_pointer(s
, page
, object
)) {
1062 object_err(s
, page
, object
, "Freelist Pointer check fails");
1066 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1072 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1074 void *object
, unsigned long addr
)
1076 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1077 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1081 /* Success perform special debug activities for allocs */
1082 if (s
->flags
& SLAB_STORE_USER
)
1083 set_track(s
, object
, TRACK_ALLOC
, addr
);
1084 trace(s
, page
, object
, 1);
1085 init_object(s
, object
, SLUB_RED_ACTIVE
);
1089 if (PageSlab(page
)) {
1091 * If this is a slab page then lets do the best we can
1092 * to avoid issues in the future. Marking all objects
1093 * as used avoids touching the remaining objects.
1095 slab_fix(s
, "Marking all objects used");
1096 page
->inuse
= page
->objects
;
1097 page
->freelist
= NULL
;
1102 static inline int free_consistency_checks(struct kmem_cache
*s
,
1103 struct page
*page
, void *object
, unsigned long addr
)
1105 if (!check_valid_pointer(s
, page
, object
)) {
1106 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1110 if (on_freelist(s
, page
, object
)) {
1111 object_err(s
, page
, object
, "Object already free");
1115 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1118 if (unlikely(s
!= page
->slab_cache
)) {
1119 if (!PageSlab(page
)) {
1120 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1122 } else if (!page
->slab_cache
) {
1123 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1127 object_err(s
, page
, object
,
1128 "page slab pointer corrupt.");
1134 /* Supports checking bulk free of a constructed freelist */
1135 static noinline
int free_debug_processing(
1136 struct kmem_cache
*s
, struct page
*page
,
1137 void *head
, void *tail
, int bulk_cnt
,
1140 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1141 void *object
= head
;
1143 unsigned long uninitialized_var(flags
);
1146 spin_lock_irqsave(&n
->list_lock
, flags
);
1149 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1150 if (!check_slab(s
, page
))
1157 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1158 if (!free_consistency_checks(s
, page
, object
, addr
))
1162 if (s
->flags
& SLAB_STORE_USER
)
1163 set_track(s
, object
, TRACK_FREE
, addr
);
1164 trace(s
, page
, object
, 0);
1165 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1166 init_object(s
, object
, SLUB_RED_INACTIVE
);
1168 /* Reached end of constructed freelist yet? */
1169 if (object
!= tail
) {
1170 object
= get_freepointer(s
, object
);
1176 if (cnt
!= bulk_cnt
)
1177 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1181 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1183 slab_fix(s
, "Object at 0x%p not freed", object
);
1187 static int __init
setup_slub_debug(char *str
)
1189 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1190 if (*str
++ != '=' || !*str
)
1192 * No options specified. Switch on full debugging.
1198 * No options but restriction on slabs. This means full
1199 * debugging for slabs matching a pattern.
1206 * Switch off all debugging measures.
1211 * Determine which debug features should be switched on
1213 for (; *str
&& *str
!= ','; str
++) {
1214 switch (tolower(*str
)) {
1216 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1219 slub_debug
|= SLAB_RED_ZONE
;
1222 slub_debug
|= SLAB_POISON
;
1225 slub_debug
|= SLAB_STORE_USER
;
1228 slub_debug
|= SLAB_TRACE
;
1231 slub_debug
|= SLAB_FAILSLAB
;
1235 * Avoid enabling debugging on caches if its minimum
1236 * order would increase as a result.
1238 disable_higher_order_debug
= 1;
1241 pr_err("slub_debug option '%c' unknown. skipped\n",
1248 slub_debug_slabs
= str
+ 1;
1253 __setup("slub_debug", setup_slub_debug
);
1255 unsigned long kmem_cache_flags(unsigned long object_size
,
1256 unsigned long flags
, const char *name
,
1257 void (*ctor
)(void *))
1260 * Enable debugging if selected on the kernel commandline.
1262 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1263 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1264 flags
|= slub_debug
;
1268 #else /* !CONFIG_SLUB_DEBUG */
1269 static inline void setup_object_debug(struct kmem_cache
*s
,
1270 struct page
*page
, void *object
) {}
1272 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1273 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1275 static inline int free_debug_processing(
1276 struct kmem_cache
*s
, struct page
*page
,
1277 void *head
, void *tail
, int bulk_cnt
,
1278 unsigned long addr
) { return 0; }
1280 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1282 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1283 void *object
, u8 val
) { return 1; }
1284 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1285 struct page
*page
) {}
1286 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1287 struct page
*page
) {}
1288 unsigned long kmem_cache_flags(unsigned long object_size
,
1289 unsigned long flags
, const char *name
,
1290 void (*ctor
)(void *))
1294 #define slub_debug 0
1296 #define disable_higher_order_debug 0
1298 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1300 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1302 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1304 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1307 #endif /* CONFIG_SLUB_DEBUG */
1310 * Hooks for other subsystems that check memory allocations. In a typical
1311 * production configuration these hooks all should produce no code at all.
1313 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1315 kmemleak_alloc(ptr
, size
, 1, flags
);
1316 kasan_kmalloc_large(ptr
, size
, flags
);
1319 static inline void kfree_hook(const void *x
)
1322 kasan_kfree_large(x
);
1325 static inline void slab_free_hook(struct kmem_cache
*s
, void *x
)
1327 kmemleak_free_recursive(x
, s
->flags
);
1330 * Trouble is that we may no longer disable interrupts in the fast path
1331 * So in order to make the debug calls that expect irqs to be
1332 * disabled we need to disable interrupts temporarily.
1334 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1336 unsigned long flags
;
1338 local_irq_save(flags
);
1339 kmemcheck_slab_free(s
, x
, s
->object_size
);
1340 debug_check_no_locks_freed(x
, s
->object_size
);
1341 local_irq_restore(flags
);
1344 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1345 debug_check_no_obj_freed(x
, s
->object_size
);
1347 kasan_slab_free(s
, x
);
1350 static inline void slab_free_freelist_hook(struct kmem_cache
*s
,
1351 void *head
, void *tail
)
1354 * Compiler cannot detect this function can be removed if slab_free_hook()
1355 * evaluates to nothing. Thus, catch all relevant config debug options here.
1357 #if defined(CONFIG_KMEMCHECK) || \
1358 defined(CONFIG_LOCKDEP) || \
1359 defined(CONFIG_DEBUG_KMEMLEAK) || \
1360 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1361 defined(CONFIG_KASAN)
1363 void *object
= head
;
1364 void *tail_obj
= tail
? : head
;
1367 slab_free_hook(s
, object
);
1368 } while ((object
!= tail_obj
) &&
1369 (object
= get_freepointer(s
, object
)));
1373 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1376 setup_object_debug(s
, page
, object
);
1377 if (unlikely(s
->ctor
)) {
1378 kasan_unpoison_object_data(s
, object
);
1380 kasan_poison_object_data(s
, object
);
1385 * Slab allocation and freeing
1387 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1388 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1391 int order
= oo_order(oo
);
1393 flags
|= __GFP_NOTRACK
;
1395 if (node
== NUMA_NO_NODE
)
1396 page
= alloc_pages(flags
, order
);
1398 page
= __alloc_pages_node(node
, flags
, order
);
1400 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1401 __free_pages(page
, order
);
1408 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1409 /* Pre-initialize the random sequence cache */
1410 static int init_cache_random_seq(struct kmem_cache
*s
)
1413 unsigned long i
, count
= oo_objects(s
->oo
);
1415 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1417 pr_err("SLUB: Unable to initialize free list for %s\n",
1422 /* Transform to an offset on the set of pages */
1423 if (s
->random_seq
) {
1424 for (i
= 0; i
< count
; i
++)
1425 s
->random_seq
[i
] *= s
->size
;
1430 /* Initialize each random sequence freelist per cache */
1431 static void __init
init_freelist_randomization(void)
1433 struct kmem_cache
*s
;
1435 mutex_lock(&slab_mutex
);
1437 list_for_each_entry(s
, &slab_caches
, list
)
1438 init_cache_random_seq(s
);
1440 mutex_unlock(&slab_mutex
);
1443 /* Get the next entry on the pre-computed freelist randomized */
1444 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1445 unsigned long *pos
, void *start
,
1446 unsigned long page_limit
,
1447 unsigned long freelist_count
)
1452 * If the target page allocation failed, the number of objects on the
1453 * page might be smaller than the usual size defined by the cache.
1456 idx
= s
->random_seq
[*pos
];
1458 if (*pos
>= freelist_count
)
1460 } while (unlikely(idx
>= page_limit
));
1462 return (char *)start
+ idx
;
1465 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1466 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1471 unsigned long idx
, pos
, page_limit
, freelist_count
;
1473 if (page
->objects
< 2 || !s
->random_seq
)
1476 freelist_count
= oo_objects(s
->oo
);
1477 pos
= get_random_int() % freelist_count
;
1479 page_limit
= page
->objects
* s
->size
;
1480 start
= fixup_red_left(s
, page_address(page
));
1482 /* First entry is used as the base of the freelist */
1483 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1485 page
->freelist
= cur
;
1487 for (idx
= 1; idx
< page
->objects
; idx
++) {
1488 setup_object(s
, page
, cur
);
1489 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1491 set_freepointer(s
, cur
, next
);
1494 setup_object(s
, page
, cur
);
1495 set_freepointer(s
, cur
, NULL
);
1500 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1504 static inline void init_freelist_randomization(void) { }
1505 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1509 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1511 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1514 struct kmem_cache_order_objects oo
= s
->oo
;
1520 flags
&= gfp_allowed_mask
;
1522 if (gfpflags_allow_blocking(flags
))
1525 flags
|= s
->allocflags
;
1528 * Let the initial higher-order allocation fail under memory pressure
1529 * so we fall-back to the minimum order allocation.
1531 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1532 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1533 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1535 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1536 if (unlikely(!page
)) {
1540 * Allocation may have failed due to fragmentation.
1541 * Try a lower order alloc if possible
1543 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1544 if (unlikely(!page
))
1546 stat(s
, ORDER_FALLBACK
);
1549 if (kmemcheck_enabled
&&
1550 !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1551 int pages
= 1 << oo_order(oo
);
1553 kmemcheck_alloc_shadow(page
, oo_order(oo
), alloc_gfp
, node
);
1556 * Objects from caches that have a constructor don't get
1557 * cleared when they're allocated, so we need to do it here.
1560 kmemcheck_mark_uninitialized_pages(page
, pages
);
1562 kmemcheck_mark_unallocated_pages(page
, pages
);
1565 page
->objects
= oo_objects(oo
);
1567 order
= compound_order(page
);
1568 page
->slab_cache
= s
;
1569 __SetPageSlab(page
);
1570 if (page_is_pfmemalloc(page
))
1571 SetPageSlabPfmemalloc(page
);
1573 start
= page_address(page
);
1575 if (unlikely(s
->flags
& SLAB_POISON
))
1576 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1578 kasan_poison_slab(page
);
1580 shuffle
= shuffle_freelist(s
, page
);
1583 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1584 setup_object(s
, page
, p
);
1585 if (likely(idx
< page
->objects
))
1586 set_freepointer(s
, p
, p
+ s
->size
);
1588 set_freepointer(s
, p
, NULL
);
1590 page
->freelist
= fixup_red_left(s
, start
);
1593 page
->inuse
= page
->objects
;
1597 if (gfpflags_allow_blocking(flags
))
1598 local_irq_disable();
1602 mod_zone_page_state(page_zone(page
),
1603 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1604 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1607 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1612 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1614 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1615 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
1619 return allocate_slab(s
,
1620 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1623 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1625 int order
= compound_order(page
);
1626 int pages
= 1 << order
;
1628 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1631 slab_pad_check(s
, page
);
1632 for_each_object(p
, s
, page_address(page
),
1634 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1637 kmemcheck_free_shadow(page
, compound_order(page
));
1639 mod_zone_page_state(page_zone(page
),
1640 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1641 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1644 __ClearPageSlabPfmemalloc(page
);
1645 __ClearPageSlab(page
);
1647 page_mapcount_reset(page
);
1648 if (current
->reclaim_state
)
1649 current
->reclaim_state
->reclaimed_slab
+= pages
;
1650 memcg_uncharge_slab(page
, order
, s
);
1651 __free_pages(page
, order
);
1654 #define need_reserve_slab_rcu \
1655 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1657 static void rcu_free_slab(struct rcu_head
*h
)
1661 if (need_reserve_slab_rcu
)
1662 page
= virt_to_head_page(h
);
1664 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1666 __free_slab(page
->slab_cache
, page
);
1669 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1671 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1672 struct rcu_head
*head
;
1674 if (need_reserve_slab_rcu
) {
1675 int order
= compound_order(page
);
1676 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1678 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1679 head
= page_address(page
) + offset
;
1681 head
= &page
->rcu_head
;
1684 call_rcu(head
, rcu_free_slab
);
1686 __free_slab(s
, page
);
1689 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1691 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1696 * Management of partially allocated slabs.
1699 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1702 if (tail
== DEACTIVATE_TO_TAIL
)
1703 list_add_tail(&page
->lru
, &n
->partial
);
1705 list_add(&page
->lru
, &n
->partial
);
1708 static inline void add_partial(struct kmem_cache_node
*n
,
1709 struct page
*page
, int tail
)
1711 lockdep_assert_held(&n
->list_lock
);
1712 __add_partial(n
, page
, tail
);
1715 static inline void remove_partial(struct kmem_cache_node
*n
,
1718 lockdep_assert_held(&n
->list_lock
);
1719 list_del(&page
->lru
);
1724 * Remove slab from the partial list, freeze it and
1725 * return the pointer to the freelist.
1727 * Returns a list of objects or NULL if it fails.
1729 static inline void *acquire_slab(struct kmem_cache
*s
,
1730 struct kmem_cache_node
*n
, struct page
*page
,
1731 int mode
, int *objects
)
1734 unsigned long counters
;
1737 lockdep_assert_held(&n
->list_lock
);
1740 * Zap the freelist and set the frozen bit.
1741 * The old freelist is the list of objects for the
1742 * per cpu allocation list.
1744 freelist
= page
->freelist
;
1745 counters
= page
->counters
;
1746 new.counters
= counters
;
1747 *objects
= new.objects
- new.inuse
;
1749 new.inuse
= page
->objects
;
1750 new.freelist
= NULL
;
1752 new.freelist
= freelist
;
1755 VM_BUG_ON(new.frozen
);
1758 if (!__cmpxchg_double_slab(s
, page
,
1760 new.freelist
, new.counters
,
1764 remove_partial(n
, page
);
1769 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1770 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1773 * Try to allocate a partial slab from a specific node.
1775 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1776 struct kmem_cache_cpu
*c
, gfp_t flags
)
1778 struct page
*page
, *page2
;
1779 void *object
= NULL
;
1784 * Racy check. If we mistakenly see no partial slabs then we
1785 * just allocate an empty slab. If we mistakenly try to get a
1786 * partial slab and there is none available then get_partials()
1789 if (!n
|| !n
->nr_partial
)
1792 spin_lock(&n
->list_lock
);
1793 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1796 if (!pfmemalloc_match(page
, flags
))
1799 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1803 available
+= objects
;
1806 stat(s
, ALLOC_FROM_PARTIAL
);
1809 put_cpu_partial(s
, page
, 0);
1810 stat(s
, CPU_PARTIAL_NODE
);
1812 if (!kmem_cache_has_cpu_partial(s
)
1813 || available
> s
->cpu_partial
/ 2)
1817 spin_unlock(&n
->list_lock
);
1822 * Get a page from somewhere. Search in increasing NUMA distances.
1824 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1825 struct kmem_cache_cpu
*c
)
1828 struct zonelist
*zonelist
;
1831 enum zone_type high_zoneidx
= gfp_zone(flags
);
1833 unsigned int cpuset_mems_cookie
;
1836 * The defrag ratio allows a configuration of the tradeoffs between
1837 * inter node defragmentation and node local allocations. A lower
1838 * defrag_ratio increases the tendency to do local allocations
1839 * instead of attempting to obtain partial slabs from other nodes.
1841 * If the defrag_ratio is set to 0 then kmalloc() always
1842 * returns node local objects. If the ratio is higher then kmalloc()
1843 * may return off node objects because partial slabs are obtained
1844 * from other nodes and filled up.
1846 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1847 * (which makes defrag_ratio = 1000) then every (well almost)
1848 * allocation will first attempt to defrag slab caches on other nodes.
1849 * This means scanning over all nodes to look for partial slabs which
1850 * may be expensive if we do it every time we are trying to find a slab
1851 * with available objects.
1853 if (!s
->remote_node_defrag_ratio
||
1854 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1858 cpuset_mems_cookie
= read_mems_allowed_begin();
1859 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1860 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1861 struct kmem_cache_node
*n
;
1863 n
= get_node(s
, zone_to_nid(zone
));
1865 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1866 n
->nr_partial
> s
->min_partial
) {
1867 object
= get_partial_node(s
, n
, c
, flags
);
1870 * Don't check read_mems_allowed_retry()
1871 * here - if mems_allowed was updated in
1872 * parallel, that was a harmless race
1873 * between allocation and the cpuset
1880 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1886 * Get a partial page, lock it and return it.
1888 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1889 struct kmem_cache_cpu
*c
)
1892 int searchnode
= node
;
1894 if (node
== NUMA_NO_NODE
)
1895 searchnode
= numa_mem_id();
1896 else if (!node_present_pages(node
))
1897 searchnode
= node_to_mem_node(node
);
1899 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1900 if (object
|| node
!= NUMA_NO_NODE
)
1903 return get_any_partial(s
, flags
, c
);
1906 #ifdef CONFIG_PREEMPT
1908 * Calculate the next globally unique transaction for disambiguiation
1909 * during cmpxchg. The transactions start with the cpu number and are then
1910 * incremented by CONFIG_NR_CPUS.
1912 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1915 * No preemption supported therefore also no need to check for
1921 static inline unsigned long next_tid(unsigned long tid
)
1923 return tid
+ TID_STEP
;
1926 static inline unsigned int tid_to_cpu(unsigned long tid
)
1928 return tid
% TID_STEP
;
1931 static inline unsigned long tid_to_event(unsigned long tid
)
1933 return tid
/ TID_STEP
;
1936 static inline unsigned int init_tid(int cpu
)
1941 static inline void note_cmpxchg_failure(const char *n
,
1942 const struct kmem_cache
*s
, unsigned long tid
)
1944 #ifdef SLUB_DEBUG_CMPXCHG
1945 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1947 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1949 #ifdef CONFIG_PREEMPT
1950 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1951 pr_warn("due to cpu change %d -> %d\n",
1952 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1955 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1956 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1957 tid_to_event(tid
), tid_to_event(actual_tid
));
1959 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1960 actual_tid
, tid
, next_tid(tid
));
1962 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1965 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1969 for_each_possible_cpu(cpu
)
1970 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1974 * Remove the cpu slab
1976 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1979 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1980 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1982 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1984 int tail
= DEACTIVATE_TO_HEAD
;
1988 if (page
->freelist
) {
1989 stat(s
, DEACTIVATE_REMOTE_FREES
);
1990 tail
= DEACTIVATE_TO_TAIL
;
1994 * Stage one: Free all available per cpu objects back
1995 * to the page freelist while it is still frozen. Leave the
1998 * There is no need to take the list->lock because the page
2001 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2003 unsigned long counters
;
2006 prior
= page
->freelist
;
2007 counters
= page
->counters
;
2008 set_freepointer(s
, freelist
, prior
);
2009 new.counters
= counters
;
2011 VM_BUG_ON(!new.frozen
);
2013 } while (!__cmpxchg_double_slab(s
, page
,
2015 freelist
, new.counters
,
2016 "drain percpu freelist"));
2018 freelist
= nextfree
;
2022 * Stage two: Ensure that the page is unfrozen while the
2023 * list presence reflects the actual number of objects
2026 * We setup the list membership and then perform a cmpxchg
2027 * with the count. If there is a mismatch then the page
2028 * is not unfrozen but the page is on the wrong list.
2030 * Then we restart the process which may have to remove
2031 * the page from the list that we just put it on again
2032 * because the number of objects in the slab may have
2037 old
.freelist
= page
->freelist
;
2038 old
.counters
= page
->counters
;
2039 VM_BUG_ON(!old
.frozen
);
2041 /* Determine target state of the slab */
2042 new.counters
= old
.counters
;
2045 set_freepointer(s
, freelist
, old
.freelist
);
2046 new.freelist
= freelist
;
2048 new.freelist
= old
.freelist
;
2052 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2054 else if (new.freelist
) {
2059 * Taking the spinlock removes the possiblity
2060 * that acquire_slab() will see a slab page that
2063 spin_lock(&n
->list_lock
);
2067 if (kmem_cache_debug(s
) && !lock
) {
2070 * This also ensures that the scanning of full
2071 * slabs from diagnostic functions will not see
2074 spin_lock(&n
->list_lock
);
2082 remove_partial(n
, page
);
2084 else if (l
== M_FULL
)
2086 remove_full(s
, n
, page
);
2088 if (m
== M_PARTIAL
) {
2090 add_partial(n
, page
, tail
);
2093 } else if (m
== M_FULL
) {
2095 stat(s
, DEACTIVATE_FULL
);
2096 add_full(s
, n
, page
);
2102 if (!__cmpxchg_double_slab(s
, page
,
2103 old
.freelist
, old
.counters
,
2104 new.freelist
, new.counters
,
2109 spin_unlock(&n
->list_lock
);
2112 stat(s
, DEACTIVATE_EMPTY
);
2113 discard_slab(s
, page
);
2119 * Unfreeze all the cpu partial slabs.
2121 * This function must be called with interrupts disabled
2122 * for the cpu using c (or some other guarantee must be there
2123 * to guarantee no concurrent accesses).
2125 static void unfreeze_partials(struct kmem_cache
*s
,
2126 struct kmem_cache_cpu
*c
)
2128 #ifdef CONFIG_SLUB_CPU_PARTIAL
2129 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2130 struct page
*page
, *discard_page
= NULL
;
2132 while ((page
= c
->partial
)) {
2136 c
->partial
= page
->next
;
2138 n2
= get_node(s
, page_to_nid(page
));
2141 spin_unlock(&n
->list_lock
);
2144 spin_lock(&n
->list_lock
);
2149 old
.freelist
= page
->freelist
;
2150 old
.counters
= page
->counters
;
2151 VM_BUG_ON(!old
.frozen
);
2153 new.counters
= old
.counters
;
2154 new.freelist
= old
.freelist
;
2158 } while (!__cmpxchg_double_slab(s
, page
,
2159 old
.freelist
, old
.counters
,
2160 new.freelist
, new.counters
,
2161 "unfreezing slab"));
2163 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2164 page
->next
= discard_page
;
2165 discard_page
= page
;
2167 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2168 stat(s
, FREE_ADD_PARTIAL
);
2173 spin_unlock(&n
->list_lock
);
2175 while (discard_page
) {
2176 page
= discard_page
;
2177 discard_page
= discard_page
->next
;
2179 stat(s
, DEACTIVATE_EMPTY
);
2180 discard_slab(s
, page
);
2187 * Put a page that was just frozen (in __slab_free) into a partial page
2188 * slot if available. This is done without interrupts disabled and without
2189 * preemption disabled. The cmpxchg is racy and may put the partial page
2190 * onto a random cpus partial slot.
2192 * If we did not find a slot then simply move all the partials to the
2193 * per node partial list.
2195 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2197 #ifdef CONFIG_SLUB_CPU_PARTIAL
2198 struct page
*oldpage
;
2206 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2209 pobjects
= oldpage
->pobjects
;
2210 pages
= oldpage
->pages
;
2211 if (drain
&& pobjects
> s
->cpu_partial
) {
2212 unsigned long flags
;
2214 * partial array is full. Move the existing
2215 * set to the per node partial list.
2217 local_irq_save(flags
);
2218 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2219 local_irq_restore(flags
);
2223 stat(s
, CPU_PARTIAL_DRAIN
);
2228 pobjects
+= page
->objects
- page
->inuse
;
2230 page
->pages
= pages
;
2231 page
->pobjects
= pobjects
;
2232 page
->next
= oldpage
;
2234 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2236 if (unlikely(!s
->cpu_partial
)) {
2237 unsigned long flags
;
2239 local_irq_save(flags
);
2240 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2241 local_irq_restore(flags
);
2247 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2249 stat(s
, CPUSLAB_FLUSH
);
2250 deactivate_slab(s
, c
->page
, c
->freelist
);
2252 c
->tid
= next_tid(c
->tid
);
2260 * Called from IPI handler with interrupts disabled.
2262 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2264 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2270 unfreeze_partials(s
, c
);
2274 static void flush_cpu_slab(void *d
)
2276 struct kmem_cache
*s
= d
;
2278 __flush_cpu_slab(s
, smp_processor_id());
2281 static bool has_cpu_slab(int cpu
, void *info
)
2283 struct kmem_cache
*s
= info
;
2284 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2286 return c
->page
|| c
->partial
;
2289 static void flush_all(struct kmem_cache
*s
)
2291 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2295 * Check if the objects in a per cpu structure fit numa
2296 * locality expectations.
2298 static inline int node_match(struct page
*page
, int node
)
2301 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2307 #ifdef CONFIG_SLUB_DEBUG
2308 static int count_free(struct page
*page
)
2310 return page
->objects
- page
->inuse
;
2313 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2315 return atomic_long_read(&n
->total_objects
);
2317 #endif /* CONFIG_SLUB_DEBUG */
2319 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2320 static unsigned long count_partial(struct kmem_cache_node
*n
,
2321 int (*get_count
)(struct page
*))
2323 unsigned long flags
;
2324 unsigned long x
= 0;
2327 spin_lock_irqsave(&n
->list_lock
, flags
);
2328 list_for_each_entry(page
, &n
->partial
, lru
)
2329 x
+= get_count(page
);
2330 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2333 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2335 static noinline
void
2336 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2338 #ifdef CONFIG_SLUB_DEBUG
2339 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2340 DEFAULT_RATELIMIT_BURST
);
2342 struct kmem_cache_node
*n
;
2344 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2347 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2348 nid
, gfpflags
, &gfpflags
);
2349 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2350 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2353 if (oo_order(s
->min
) > get_order(s
->object_size
))
2354 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2357 for_each_kmem_cache_node(s
, node
, n
) {
2358 unsigned long nr_slabs
;
2359 unsigned long nr_objs
;
2360 unsigned long nr_free
;
2362 nr_free
= count_partial(n
, count_free
);
2363 nr_slabs
= node_nr_slabs(n
);
2364 nr_objs
= node_nr_objs(n
);
2366 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2367 node
, nr_slabs
, nr_objs
, nr_free
);
2372 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2373 int node
, struct kmem_cache_cpu
**pc
)
2376 struct kmem_cache_cpu
*c
= *pc
;
2379 freelist
= get_partial(s
, flags
, node
, c
);
2384 page
= new_slab(s
, flags
, node
);
2386 c
= raw_cpu_ptr(s
->cpu_slab
);
2391 * No other reference to the page yet so we can
2392 * muck around with it freely without cmpxchg
2394 freelist
= page
->freelist
;
2395 page
->freelist
= NULL
;
2397 stat(s
, ALLOC_SLAB
);
2406 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2408 if (unlikely(PageSlabPfmemalloc(page
)))
2409 return gfp_pfmemalloc_allowed(gfpflags
);
2415 * Check the page->freelist of a page and either transfer the freelist to the
2416 * per cpu freelist or deactivate the page.
2418 * The page is still frozen if the return value is not NULL.
2420 * If this function returns NULL then the page has been unfrozen.
2422 * This function must be called with interrupt disabled.
2424 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2427 unsigned long counters
;
2431 freelist
= page
->freelist
;
2432 counters
= page
->counters
;
2434 new.counters
= counters
;
2435 VM_BUG_ON(!new.frozen
);
2437 new.inuse
= page
->objects
;
2438 new.frozen
= freelist
!= NULL
;
2440 } while (!__cmpxchg_double_slab(s
, page
,
2449 * Slow path. The lockless freelist is empty or we need to perform
2452 * Processing is still very fast if new objects have been freed to the
2453 * regular freelist. In that case we simply take over the regular freelist
2454 * as the lockless freelist and zap the regular freelist.
2456 * If that is not working then we fall back to the partial lists. We take the
2457 * first element of the freelist as the object to allocate now and move the
2458 * rest of the freelist to the lockless freelist.
2460 * And if we were unable to get a new slab from the partial slab lists then
2461 * we need to allocate a new slab. This is the slowest path since it involves
2462 * a call to the page allocator and the setup of a new slab.
2464 * Version of __slab_alloc to use when we know that interrupts are
2465 * already disabled (which is the case for bulk allocation).
2467 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2468 unsigned long addr
, struct kmem_cache_cpu
*c
)
2478 if (unlikely(!node_match(page
, node
))) {
2479 int searchnode
= node
;
2481 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2482 searchnode
= node_to_mem_node(node
);
2484 if (unlikely(!node_match(page
, searchnode
))) {
2485 stat(s
, ALLOC_NODE_MISMATCH
);
2486 deactivate_slab(s
, page
, c
->freelist
);
2494 * By rights, we should be searching for a slab page that was
2495 * PFMEMALLOC but right now, we are losing the pfmemalloc
2496 * information when the page leaves the per-cpu allocator
2498 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2499 deactivate_slab(s
, page
, c
->freelist
);
2505 /* must check again c->freelist in case of cpu migration or IRQ */
2506 freelist
= c
->freelist
;
2510 freelist
= get_freelist(s
, page
);
2514 stat(s
, DEACTIVATE_BYPASS
);
2518 stat(s
, ALLOC_REFILL
);
2522 * freelist is pointing to the list of objects to be used.
2523 * page is pointing to the page from which the objects are obtained.
2524 * That page must be frozen for per cpu allocations to work.
2526 VM_BUG_ON(!c
->page
->frozen
);
2527 c
->freelist
= get_freepointer(s
, freelist
);
2528 c
->tid
= next_tid(c
->tid
);
2534 page
= c
->page
= c
->partial
;
2535 c
->partial
= page
->next
;
2536 stat(s
, CPU_PARTIAL_ALLOC
);
2541 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2543 if (unlikely(!freelist
)) {
2544 slab_out_of_memory(s
, gfpflags
, node
);
2549 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2552 /* Only entered in the debug case */
2553 if (kmem_cache_debug(s
) &&
2554 !alloc_debug_processing(s
, page
, freelist
, addr
))
2555 goto new_slab
; /* Slab failed checks. Next slab needed */
2557 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
2564 * Another one that disabled interrupt and compensates for possible
2565 * cpu changes by refetching the per cpu area pointer.
2567 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2568 unsigned long addr
, struct kmem_cache_cpu
*c
)
2571 unsigned long flags
;
2573 local_irq_save(flags
);
2574 #ifdef CONFIG_PREEMPT
2576 * We may have been preempted and rescheduled on a different
2577 * cpu before disabling interrupts. Need to reload cpu area
2580 c
= this_cpu_ptr(s
->cpu_slab
);
2583 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2584 local_irq_restore(flags
);
2589 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2590 * have the fastpath folded into their functions. So no function call
2591 * overhead for requests that can be satisfied on the fastpath.
2593 * The fastpath works by first checking if the lockless freelist can be used.
2594 * If not then __slab_alloc is called for slow processing.
2596 * Otherwise we can simply pick the next object from the lockless free list.
2598 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2599 gfp_t gfpflags
, int node
, unsigned long addr
)
2602 struct kmem_cache_cpu
*c
;
2606 s
= slab_pre_alloc_hook(s
, gfpflags
);
2611 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2612 * enabled. We may switch back and forth between cpus while
2613 * reading from one cpu area. That does not matter as long
2614 * as we end up on the original cpu again when doing the cmpxchg.
2616 * We should guarantee that tid and kmem_cache are retrieved on
2617 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2618 * to check if it is matched or not.
2621 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2622 c
= raw_cpu_ptr(s
->cpu_slab
);
2623 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2624 unlikely(tid
!= READ_ONCE(c
->tid
)));
2627 * Irqless object alloc/free algorithm used here depends on sequence
2628 * of fetching cpu_slab's data. tid should be fetched before anything
2629 * on c to guarantee that object and page associated with previous tid
2630 * won't be used with current tid. If we fetch tid first, object and
2631 * page could be one associated with next tid and our alloc/free
2632 * request will be failed. In this case, we will retry. So, no problem.
2637 * The transaction ids are globally unique per cpu and per operation on
2638 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2639 * occurs on the right processor and that there was no operation on the
2640 * linked list in between.
2643 object
= c
->freelist
;
2645 if (unlikely(!object
|| !node_match(page
, node
))) {
2646 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2647 stat(s
, ALLOC_SLOWPATH
);
2649 void *next_object
= get_freepointer_safe(s
, object
);
2652 * The cmpxchg will only match if there was no additional
2653 * operation and if we are on the right processor.
2655 * The cmpxchg does the following atomically (without lock
2657 * 1. Relocate first pointer to the current per cpu area.
2658 * 2. Verify that tid and freelist have not been changed
2659 * 3. If they were not changed replace tid and freelist
2661 * Since this is without lock semantics the protection is only
2662 * against code executing on this cpu *not* from access by
2665 if (unlikely(!this_cpu_cmpxchg_double(
2666 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2668 next_object
, next_tid(tid
)))) {
2670 note_cmpxchg_failure("slab_alloc", s
, tid
);
2673 prefetch_freepointer(s
, next_object
);
2674 stat(s
, ALLOC_FASTPATH
);
2677 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2678 memset(object
, 0, s
->object_size
);
2680 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2685 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2686 gfp_t gfpflags
, unsigned long addr
)
2688 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2691 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2693 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2695 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2700 EXPORT_SYMBOL(kmem_cache_alloc
);
2702 #ifdef CONFIG_TRACING
2703 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2705 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2706 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2707 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2710 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2714 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2716 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2718 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2719 s
->object_size
, s
->size
, gfpflags
, node
);
2723 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2725 #ifdef CONFIG_TRACING
2726 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2728 int node
, size_t size
)
2730 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2732 trace_kmalloc_node(_RET_IP_
, ret
,
2733 size
, s
->size
, gfpflags
, node
);
2735 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2738 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2743 * Slow path handling. This may still be called frequently since objects
2744 * have a longer lifetime than the cpu slabs in most processing loads.
2746 * So we still attempt to reduce cache line usage. Just take the slab
2747 * lock and free the item. If there is no additional partial page
2748 * handling required then we can return immediately.
2750 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2751 void *head
, void *tail
, int cnt
,
2758 unsigned long counters
;
2759 struct kmem_cache_node
*n
= NULL
;
2760 unsigned long uninitialized_var(flags
);
2762 stat(s
, FREE_SLOWPATH
);
2764 if (kmem_cache_debug(s
) &&
2765 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2770 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2773 prior
= page
->freelist
;
2774 counters
= page
->counters
;
2775 set_freepointer(s
, tail
, prior
);
2776 new.counters
= counters
;
2777 was_frozen
= new.frozen
;
2779 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2781 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2784 * Slab was on no list before and will be
2786 * We can defer the list move and instead
2791 } else { /* Needs to be taken off a list */
2793 n
= get_node(s
, page_to_nid(page
));
2795 * Speculatively acquire the list_lock.
2796 * If the cmpxchg does not succeed then we may
2797 * drop the list_lock without any processing.
2799 * Otherwise the list_lock will synchronize with
2800 * other processors updating the list of slabs.
2802 spin_lock_irqsave(&n
->list_lock
, flags
);
2807 } while (!cmpxchg_double_slab(s
, page
,
2815 * If we just froze the page then put it onto the
2816 * per cpu partial list.
2818 if (new.frozen
&& !was_frozen
) {
2819 put_cpu_partial(s
, page
, 1);
2820 stat(s
, CPU_PARTIAL_FREE
);
2823 * The list lock was not taken therefore no list
2824 * activity can be necessary.
2827 stat(s
, FREE_FROZEN
);
2831 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2835 * Objects left in the slab. If it was not on the partial list before
2838 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2839 if (kmem_cache_debug(s
))
2840 remove_full(s
, n
, page
);
2841 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2842 stat(s
, FREE_ADD_PARTIAL
);
2844 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2850 * Slab on the partial list.
2852 remove_partial(n
, page
);
2853 stat(s
, FREE_REMOVE_PARTIAL
);
2855 /* Slab must be on the full list */
2856 remove_full(s
, n
, page
);
2859 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2861 discard_slab(s
, page
);
2865 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2866 * can perform fastpath freeing without additional function calls.
2868 * The fastpath is only possible if we are freeing to the current cpu slab
2869 * of this processor. This typically the case if we have just allocated
2872 * If fastpath is not possible then fall back to __slab_free where we deal
2873 * with all sorts of special processing.
2875 * Bulk free of a freelist with several objects (all pointing to the
2876 * same page) possible by specifying head and tail ptr, plus objects
2877 * count (cnt). Bulk free indicated by tail pointer being set.
2879 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2880 void *head
, void *tail
, int cnt
,
2883 void *tail_obj
= tail
? : head
;
2884 struct kmem_cache_cpu
*c
;
2887 slab_free_freelist_hook(s
, head
, tail
);
2891 * Determine the currently cpus per cpu slab.
2892 * The cpu may change afterward. However that does not matter since
2893 * data is retrieved via this pointer. If we are on the same cpu
2894 * during the cmpxchg then the free will succeed.
2897 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2898 c
= raw_cpu_ptr(s
->cpu_slab
);
2899 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2900 unlikely(tid
!= READ_ONCE(c
->tid
)));
2902 /* Same with comment on barrier() in slab_alloc_node() */
2905 if (likely(page
== c
->page
)) {
2906 set_freepointer(s
, tail_obj
, c
->freelist
);
2908 if (unlikely(!this_cpu_cmpxchg_double(
2909 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2911 head
, next_tid(tid
)))) {
2913 note_cmpxchg_failure("slab_free", s
, tid
);
2916 stat(s
, FREE_FASTPATH
);
2918 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2922 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2924 s
= cache_from_obj(s
, x
);
2927 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
2928 trace_kmem_cache_free(_RET_IP_
, x
);
2930 EXPORT_SYMBOL(kmem_cache_free
);
2932 struct detached_freelist
{
2937 struct kmem_cache
*s
;
2941 * This function progressively scans the array with free objects (with
2942 * a limited look ahead) and extract objects belonging to the same
2943 * page. It builds a detached freelist directly within the given
2944 * page/objects. This can happen without any need for
2945 * synchronization, because the objects are owned by running process.
2946 * The freelist is build up as a single linked list in the objects.
2947 * The idea is, that this detached freelist can then be bulk
2948 * transferred to the real freelist(s), but only requiring a single
2949 * synchronization primitive. Look ahead in the array is limited due
2950 * to performance reasons.
2953 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
2954 void **p
, struct detached_freelist
*df
)
2956 size_t first_skipped_index
= 0;
2961 /* Always re-init detached_freelist */
2966 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2967 } while (!object
&& size
);
2972 page
= virt_to_head_page(object
);
2974 /* Handle kalloc'ed objects */
2975 if (unlikely(!PageSlab(page
))) {
2976 BUG_ON(!PageCompound(page
));
2978 __free_kmem_pages(page
, compound_order(page
));
2979 p
[size
] = NULL
; /* mark object processed */
2982 /* Derive kmem_cache from object */
2983 df
->s
= page
->slab_cache
;
2985 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
2988 /* Start new detached freelist */
2990 set_freepointer(df
->s
, object
, NULL
);
2992 df
->freelist
= object
;
2993 p
[size
] = NULL
; /* mark object processed */
2999 continue; /* Skip processed objects */
3001 /* df->page is always set at this point */
3002 if (df
->page
== virt_to_head_page(object
)) {
3003 /* Opportunity build freelist */
3004 set_freepointer(df
->s
, object
, df
->freelist
);
3005 df
->freelist
= object
;
3007 p
[size
] = NULL
; /* mark object processed */
3012 /* Limit look ahead search */
3016 if (!first_skipped_index
)
3017 first_skipped_index
= size
+ 1;
3020 return first_skipped_index
;
3023 /* Note that interrupts must be enabled when calling this function. */
3024 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3030 struct detached_freelist df
;
3032 size
= build_detached_freelist(s
, size
, p
, &df
);
3033 if (unlikely(!df
.page
))
3036 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3037 } while (likely(size
));
3039 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3041 /* Note that interrupts must be enabled when calling this function. */
3042 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3045 struct kmem_cache_cpu
*c
;
3048 /* memcg and kmem_cache debug support */
3049 s
= slab_pre_alloc_hook(s
, flags
);
3053 * Drain objects in the per cpu slab, while disabling local
3054 * IRQs, which protects against PREEMPT and interrupts
3055 * handlers invoking normal fastpath.
3057 local_irq_disable();
3058 c
= this_cpu_ptr(s
->cpu_slab
);
3060 for (i
= 0; i
< size
; i
++) {
3061 void *object
= c
->freelist
;
3063 if (unlikely(!object
)) {
3065 * Invoking slow path likely have side-effect
3066 * of re-populating per CPU c->freelist
3068 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3070 if (unlikely(!p
[i
]))
3073 c
= this_cpu_ptr(s
->cpu_slab
);
3074 continue; /* goto for-loop */
3076 c
->freelist
= get_freepointer(s
, object
);
3079 c
->tid
= next_tid(c
->tid
);
3082 /* Clear memory outside IRQ disabled fastpath loop */
3083 if (unlikely(flags
& __GFP_ZERO
)) {
3086 for (j
= 0; j
< i
; j
++)
3087 memset(p
[j
], 0, s
->object_size
);
3090 /* memcg and kmem_cache debug support */
3091 slab_post_alloc_hook(s
, flags
, size
, p
);
3095 slab_post_alloc_hook(s
, flags
, i
, p
);
3096 __kmem_cache_free_bulk(s
, i
, p
);
3099 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3103 * Object placement in a slab is made very easy because we always start at
3104 * offset 0. If we tune the size of the object to the alignment then we can
3105 * get the required alignment by putting one properly sized object after
3108 * Notice that the allocation order determines the sizes of the per cpu
3109 * caches. Each processor has always one slab available for allocations.
3110 * Increasing the allocation order reduces the number of times that slabs
3111 * must be moved on and off the partial lists and is therefore a factor in
3116 * Mininum / Maximum order of slab pages. This influences locking overhead
3117 * and slab fragmentation. A higher order reduces the number of partial slabs
3118 * and increases the number of allocations possible without having to
3119 * take the list_lock.
3121 static int slub_min_order
;
3122 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3123 static int slub_min_objects
;
3126 * Calculate the order of allocation given an slab object size.
3128 * The order of allocation has significant impact on performance and other
3129 * system components. Generally order 0 allocations should be preferred since
3130 * order 0 does not cause fragmentation in the page allocator. Larger objects
3131 * be problematic to put into order 0 slabs because there may be too much
3132 * unused space left. We go to a higher order if more than 1/16th of the slab
3135 * In order to reach satisfactory performance we must ensure that a minimum
3136 * number of objects is in one slab. Otherwise we may generate too much
3137 * activity on the partial lists which requires taking the list_lock. This is
3138 * less a concern for large slabs though which are rarely used.
3140 * slub_max_order specifies the order where we begin to stop considering the
3141 * number of objects in a slab as critical. If we reach slub_max_order then
3142 * we try to keep the page order as low as possible. So we accept more waste
3143 * of space in favor of a small page order.
3145 * Higher order allocations also allow the placement of more objects in a
3146 * slab and thereby reduce object handling overhead. If the user has
3147 * requested a higher mininum order then we start with that one instead of
3148 * the smallest order which will fit the object.
3150 static inline int slab_order(int size
, int min_objects
,
3151 int max_order
, int fract_leftover
, int reserved
)
3155 int min_order
= slub_min_order
;
3157 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3158 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3160 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3161 order
<= max_order
; order
++) {
3163 unsigned long slab_size
= PAGE_SIZE
<< order
;
3165 rem
= (slab_size
- reserved
) % size
;
3167 if (rem
<= slab_size
/ fract_leftover
)
3174 static inline int calculate_order(int size
, int reserved
)
3182 * Attempt to find best configuration for a slab. This
3183 * works by first attempting to generate a layout with
3184 * the best configuration and backing off gradually.
3186 * First we increase the acceptable waste in a slab. Then
3187 * we reduce the minimum objects required in a slab.
3189 min_objects
= slub_min_objects
;
3191 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3192 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3193 min_objects
= min(min_objects
, max_objects
);
3195 while (min_objects
> 1) {
3197 while (fraction
>= 4) {
3198 order
= slab_order(size
, min_objects
,
3199 slub_max_order
, fraction
, reserved
);
3200 if (order
<= slub_max_order
)
3208 * We were unable to place multiple objects in a slab. Now
3209 * lets see if we can place a single object there.
3211 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3212 if (order
<= slub_max_order
)
3216 * Doh this slab cannot be placed using slub_max_order.
3218 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3219 if (order
< MAX_ORDER
)
3225 init_kmem_cache_node(struct kmem_cache_node
*n
)
3228 spin_lock_init(&n
->list_lock
);
3229 INIT_LIST_HEAD(&n
->partial
);
3230 #ifdef CONFIG_SLUB_DEBUG
3231 atomic_long_set(&n
->nr_slabs
, 0);
3232 atomic_long_set(&n
->total_objects
, 0);
3233 INIT_LIST_HEAD(&n
->full
);
3237 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3239 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3240 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3243 * Must align to double word boundary for the double cmpxchg
3244 * instructions to work; see __pcpu_double_call_return_bool().
3246 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3247 2 * sizeof(void *));
3252 init_kmem_cache_cpus(s
);
3257 static struct kmem_cache
*kmem_cache_node
;
3260 * No kmalloc_node yet so do it by hand. We know that this is the first
3261 * slab on the node for this slabcache. There are no concurrent accesses
3264 * Note that this function only works on the kmem_cache_node
3265 * when allocating for the kmem_cache_node. This is used for bootstrapping
3266 * memory on a fresh node that has no slab structures yet.
3268 static void early_kmem_cache_node_alloc(int node
)
3271 struct kmem_cache_node
*n
;
3273 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3275 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3278 if (page_to_nid(page
) != node
) {
3279 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3280 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3285 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3288 kmem_cache_node
->node
[node
] = n
;
3289 #ifdef CONFIG_SLUB_DEBUG
3290 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3291 init_tracking(kmem_cache_node
, n
);
3293 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3295 init_kmem_cache_node(n
);
3296 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3299 * No locks need to be taken here as it has just been
3300 * initialized and there is no concurrent access.
3302 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3305 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3308 struct kmem_cache_node
*n
;
3310 for_each_kmem_cache_node(s
, node
, n
) {
3311 kmem_cache_free(kmem_cache_node
, n
);
3312 s
->node
[node
] = NULL
;
3316 void __kmem_cache_release(struct kmem_cache
*s
)
3318 cache_random_seq_destroy(s
);
3319 free_percpu(s
->cpu_slab
);
3320 free_kmem_cache_nodes(s
);
3323 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3327 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3328 struct kmem_cache_node
*n
;
3330 if (slab_state
== DOWN
) {
3331 early_kmem_cache_node_alloc(node
);
3334 n
= kmem_cache_alloc_node(kmem_cache_node
,
3338 free_kmem_cache_nodes(s
);
3343 init_kmem_cache_node(n
);
3348 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3350 if (min
< MIN_PARTIAL
)
3352 else if (min
> MAX_PARTIAL
)
3354 s
->min_partial
= min
;
3358 * calculate_sizes() determines the order and the distribution of data within
3361 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3363 unsigned long flags
= s
->flags
;
3364 unsigned long size
= s
->object_size
;
3368 * Round up object size to the next word boundary. We can only
3369 * place the free pointer at word boundaries and this determines
3370 * the possible location of the free pointer.
3372 size
= ALIGN(size
, sizeof(void *));
3374 #ifdef CONFIG_SLUB_DEBUG
3376 * Determine if we can poison the object itself. If the user of
3377 * the slab may touch the object after free or before allocation
3378 * then we should never poison the object itself.
3380 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
3382 s
->flags
|= __OBJECT_POISON
;
3384 s
->flags
&= ~__OBJECT_POISON
;
3388 * If we are Redzoning then check if there is some space between the
3389 * end of the object and the free pointer. If not then add an
3390 * additional word to have some bytes to store Redzone information.
3392 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3393 size
+= sizeof(void *);
3397 * With that we have determined the number of bytes in actual use
3398 * by the object. This is the potential offset to the free pointer.
3402 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
3405 * Relocate free pointer after the object if it is not
3406 * permitted to overwrite the first word of the object on
3409 * This is the case if we do RCU, have a constructor or
3410 * destructor or are poisoning the objects.
3413 size
+= sizeof(void *);
3416 #ifdef CONFIG_SLUB_DEBUG
3417 if (flags
& SLAB_STORE_USER
)
3419 * Need to store information about allocs and frees after
3422 size
+= 2 * sizeof(struct track
);
3424 if (flags
& SLAB_RED_ZONE
) {
3426 * Add some empty padding so that we can catch
3427 * overwrites from earlier objects rather than let
3428 * tracking information or the free pointer be
3429 * corrupted if a user writes before the start
3432 size
+= sizeof(void *);
3434 s
->red_left_pad
= sizeof(void *);
3435 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3436 size
+= s
->red_left_pad
;
3441 * SLUB stores one object immediately after another beginning from
3442 * offset 0. In order to align the objects we have to simply size
3443 * each object to conform to the alignment.
3445 size
= ALIGN(size
, s
->align
);
3447 if (forced_order
>= 0)
3448 order
= forced_order
;
3450 order
= calculate_order(size
, s
->reserved
);
3457 s
->allocflags
|= __GFP_COMP
;
3459 if (s
->flags
& SLAB_CACHE_DMA
)
3460 s
->allocflags
|= GFP_DMA
;
3462 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3463 s
->allocflags
|= __GFP_RECLAIMABLE
;
3466 * Determine the number of objects per slab
3468 s
->oo
= oo_make(order
, size
, s
->reserved
);
3469 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3470 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3473 return !!oo_objects(s
->oo
);
3476 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3478 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3481 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
3482 s
->reserved
= sizeof(struct rcu_head
);
3484 if (!calculate_sizes(s
, -1))
3486 if (disable_higher_order_debug
) {
3488 * Disable debugging flags that store metadata if the min slab
3491 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3492 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3494 if (!calculate_sizes(s
, -1))
3499 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3500 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3501 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3502 /* Enable fast mode */
3503 s
->flags
|= __CMPXCHG_DOUBLE
;
3507 * The larger the object size is, the more pages we want on the partial
3508 * list to avoid pounding the page allocator excessively.
3510 set_min_partial(s
, ilog2(s
->size
) / 2);
3513 * cpu_partial determined the maximum number of objects kept in the
3514 * per cpu partial lists of a processor.
3516 * Per cpu partial lists mainly contain slabs that just have one
3517 * object freed. If they are used for allocation then they can be
3518 * filled up again with minimal effort. The slab will never hit the
3519 * per node partial lists and therefore no locking will be required.
3521 * This setting also determines
3523 * A) The number of objects from per cpu partial slabs dumped to the
3524 * per node list when we reach the limit.
3525 * B) The number of objects in cpu partial slabs to extract from the
3526 * per node list when we run out of per cpu objects. We only fetch
3527 * 50% to keep some capacity around for frees.
3529 if (!kmem_cache_has_cpu_partial(s
))
3531 else if (s
->size
>= PAGE_SIZE
)
3533 else if (s
->size
>= 1024)
3535 else if (s
->size
>= 256)
3536 s
->cpu_partial
= 13;
3538 s
->cpu_partial
= 30;
3541 s
->remote_node_defrag_ratio
= 1000;
3544 /* Initialize the pre-computed randomized freelist if slab is up */
3545 if (slab_state
>= UP
) {
3546 if (init_cache_random_seq(s
))
3550 if (!init_kmem_cache_nodes(s
))
3553 if (alloc_kmem_cache_cpus(s
))
3556 free_kmem_cache_nodes(s
);
3558 if (flags
& SLAB_PANIC
)
3559 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3560 s
->name
, (unsigned long)s
->size
, s
->size
,
3561 oo_order(s
->oo
), s
->offset
, flags
);
3565 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3568 #ifdef CONFIG_SLUB_DEBUG
3569 void *addr
= page_address(page
);
3571 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3572 sizeof(long), GFP_ATOMIC
);
3575 slab_err(s
, page
, text
, s
->name
);
3578 get_map(s
, page
, map
);
3579 for_each_object(p
, s
, addr
, page
->objects
) {
3581 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3582 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3583 print_tracking(s
, p
);
3592 * Attempt to free all partial slabs on a node.
3593 * This is called from __kmem_cache_shutdown(). We must take list_lock
3594 * because sysfs file might still access partial list after the shutdowning.
3596 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3598 struct page
*page
, *h
;
3600 BUG_ON(irqs_disabled());
3601 spin_lock_irq(&n
->list_lock
);
3602 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3604 remove_partial(n
, page
);
3605 discard_slab(s
, page
);
3607 list_slab_objects(s
, page
,
3608 "Objects remaining in %s on __kmem_cache_shutdown()");
3611 spin_unlock_irq(&n
->list_lock
);
3615 * Release all resources used by a slab cache.
3617 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3620 struct kmem_cache_node
*n
;
3623 /* Attempt to free all objects */
3624 for_each_kmem_cache_node(s
, node
, n
) {
3626 if (n
->nr_partial
|| slabs_node(s
, node
))
3632 /********************************************************************
3634 *******************************************************************/
3636 static int __init
setup_slub_min_order(char *str
)
3638 get_option(&str
, &slub_min_order
);
3643 __setup("slub_min_order=", setup_slub_min_order
);
3645 static int __init
setup_slub_max_order(char *str
)
3647 get_option(&str
, &slub_max_order
);
3648 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3653 __setup("slub_max_order=", setup_slub_max_order
);
3655 static int __init
setup_slub_min_objects(char *str
)
3657 get_option(&str
, &slub_min_objects
);
3662 __setup("slub_min_objects=", setup_slub_min_objects
);
3664 void *__kmalloc(size_t size
, gfp_t flags
)
3666 struct kmem_cache
*s
;
3669 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3670 return kmalloc_large(size
, flags
);
3672 s
= kmalloc_slab(size
, flags
);
3674 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3677 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3679 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3681 kasan_kmalloc(s
, ret
, size
, flags
);
3685 EXPORT_SYMBOL(__kmalloc
);
3688 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3693 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
3694 page
= alloc_kmem_pages_node(node
, flags
, get_order(size
));
3696 ptr
= page_address(page
);
3698 kmalloc_large_node_hook(ptr
, size
, flags
);
3702 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3704 struct kmem_cache
*s
;
3707 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3708 ret
= kmalloc_large_node(size
, flags
, node
);
3710 trace_kmalloc_node(_RET_IP_
, ret
,
3711 size
, PAGE_SIZE
<< get_order(size
),
3717 s
= kmalloc_slab(size
, flags
);
3719 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3722 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3724 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3726 kasan_kmalloc(s
, ret
, size
, flags
);
3730 EXPORT_SYMBOL(__kmalloc_node
);
3733 static size_t __ksize(const void *object
)
3737 if (unlikely(object
== ZERO_SIZE_PTR
))
3740 page
= virt_to_head_page(object
);
3742 if (unlikely(!PageSlab(page
))) {
3743 WARN_ON(!PageCompound(page
));
3744 return PAGE_SIZE
<< compound_order(page
);
3747 return slab_ksize(page
->slab_cache
);
3750 size_t ksize(const void *object
)
3752 size_t size
= __ksize(object
);
3753 /* We assume that ksize callers could use whole allocated area,
3754 * so we need to unpoison this area.
3756 kasan_unpoison_shadow(object
, size
);
3759 EXPORT_SYMBOL(ksize
);
3761 void kfree(const void *x
)
3764 void *object
= (void *)x
;
3766 trace_kfree(_RET_IP_
, x
);
3768 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3771 page
= virt_to_head_page(x
);
3772 if (unlikely(!PageSlab(page
))) {
3773 BUG_ON(!PageCompound(page
));
3775 __free_kmem_pages(page
, compound_order(page
));
3778 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3780 EXPORT_SYMBOL(kfree
);
3782 #define SHRINK_PROMOTE_MAX 32
3785 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3786 * up most to the head of the partial lists. New allocations will then
3787 * fill those up and thus they can be removed from the partial lists.
3789 * The slabs with the least items are placed last. This results in them
3790 * being allocated from last increasing the chance that the last objects
3791 * are freed in them.
3793 int __kmem_cache_shrink(struct kmem_cache
*s
, bool deactivate
)
3797 struct kmem_cache_node
*n
;
3800 struct list_head discard
;
3801 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3802 unsigned long flags
;
3807 * Disable empty slabs caching. Used to avoid pinning offline
3808 * memory cgroups by kmem pages that can be freed.
3814 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3815 * so we have to make sure the change is visible.
3817 synchronize_sched();
3821 for_each_kmem_cache_node(s
, node
, n
) {
3822 INIT_LIST_HEAD(&discard
);
3823 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3824 INIT_LIST_HEAD(promote
+ i
);
3826 spin_lock_irqsave(&n
->list_lock
, flags
);
3829 * Build lists of slabs to discard or promote.
3831 * Note that concurrent frees may occur while we hold the
3832 * list_lock. page->inuse here is the upper limit.
3834 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3835 int free
= page
->objects
- page
->inuse
;
3837 /* Do not reread page->inuse */
3840 /* We do not keep full slabs on the list */
3843 if (free
== page
->objects
) {
3844 list_move(&page
->lru
, &discard
);
3846 } else if (free
<= SHRINK_PROMOTE_MAX
)
3847 list_move(&page
->lru
, promote
+ free
- 1);
3851 * Promote the slabs filled up most to the head of the
3854 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3855 list_splice(promote
+ i
, &n
->partial
);
3857 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3859 /* Release empty slabs */
3860 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3861 discard_slab(s
, page
);
3863 if (slabs_node(s
, node
))
3870 static int slab_mem_going_offline_callback(void *arg
)
3872 struct kmem_cache
*s
;
3874 mutex_lock(&slab_mutex
);
3875 list_for_each_entry(s
, &slab_caches
, list
)
3876 __kmem_cache_shrink(s
, false);
3877 mutex_unlock(&slab_mutex
);
3882 static void slab_mem_offline_callback(void *arg
)
3884 struct kmem_cache_node
*n
;
3885 struct kmem_cache
*s
;
3886 struct memory_notify
*marg
= arg
;
3889 offline_node
= marg
->status_change_nid_normal
;
3892 * If the node still has available memory. we need kmem_cache_node
3895 if (offline_node
< 0)
3898 mutex_lock(&slab_mutex
);
3899 list_for_each_entry(s
, &slab_caches
, list
) {
3900 n
= get_node(s
, offline_node
);
3903 * if n->nr_slabs > 0, slabs still exist on the node
3904 * that is going down. We were unable to free them,
3905 * and offline_pages() function shouldn't call this
3906 * callback. So, we must fail.
3908 BUG_ON(slabs_node(s
, offline_node
));
3910 s
->node
[offline_node
] = NULL
;
3911 kmem_cache_free(kmem_cache_node
, n
);
3914 mutex_unlock(&slab_mutex
);
3917 static int slab_mem_going_online_callback(void *arg
)
3919 struct kmem_cache_node
*n
;
3920 struct kmem_cache
*s
;
3921 struct memory_notify
*marg
= arg
;
3922 int nid
= marg
->status_change_nid_normal
;
3926 * If the node's memory is already available, then kmem_cache_node is
3927 * already created. Nothing to do.
3933 * We are bringing a node online. No memory is available yet. We must
3934 * allocate a kmem_cache_node structure in order to bring the node
3937 mutex_lock(&slab_mutex
);
3938 list_for_each_entry(s
, &slab_caches
, list
) {
3940 * XXX: kmem_cache_alloc_node will fallback to other nodes
3941 * since memory is not yet available from the node that
3944 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3949 init_kmem_cache_node(n
);
3953 mutex_unlock(&slab_mutex
);
3957 static int slab_memory_callback(struct notifier_block
*self
,
3958 unsigned long action
, void *arg
)
3963 case MEM_GOING_ONLINE
:
3964 ret
= slab_mem_going_online_callback(arg
);
3966 case MEM_GOING_OFFLINE
:
3967 ret
= slab_mem_going_offline_callback(arg
);
3970 case MEM_CANCEL_ONLINE
:
3971 slab_mem_offline_callback(arg
);
3974 case MEM_CANCEL_OFFLINE
:
3978 ret
= notifier_from_errno(ret
);
3984 static struct notifier_block slab_memory_callback_nb
= {
3985 .notifier_call
= slab_memory_callback
,
3986 .priority
= SLAB_CALLBACK_PRI
,
3989 /********************************************************************
3990 * Basic setup of slabs
3991 *******************************************************************/
3994 * Used for early kmem_cache structures that were allocated using
3995 * the page allocator. Allocate them properly then fix up the pointers
3996 * that may be pointing to the wrong kmem_cache structure.
3999 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4002 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4003 struct kmem_cache_node
*n
;
4005 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4008 * This runs very early, and only the boot processor is supposed to be
4009 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4012 __flush_cpu_slab(s
, smp_processor_id());
4013 for_each_kmem_cache_node(s
, node
, n
) {
4016 list_for_each_entry(p
, &n
->partial
, lru
)
4019 #ifdef CONFIG_SLUB_DEBUG
4020 list_for_each_entry(p
, &n
->full
, lru
)
4024 slab_init_memcg_params(s
);
4025 list_add(&s
->list
, &slab_caches
);
4029 void __init
kmem_cache_init(void)
4031 static __initdata
struct kmem_cache boot_kmem_cache
,
4032 boot_kmem_cache_node
;
4034 if (debug_guardpage_minorder())
4037 kmem_cache_node
= &boot_kmem_cache_node
;
4038 kmem_cache
= &boot_kmem_cache
;
4040 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4041 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
4043 register_hotmemory_notifier(&slab_memory_callback_nb
);
4045 /* Able to allocate the per node structures */
4046 slab_state
= PARTIAL
;
4048 create_boot_cache(kmem_cache
, "kmem_cache",
4049 offsetof(struct kmem_cache
, node
) +
4050 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4051 SLAB_HWCACHE_ALIGN
);
4053 kmem_cache
= bootstrap(&boot_kmem_cache
);
4056 * Allocate kmem_cache_node properly from the kmem_cache slab.
4057 * kmem_cache_node is separately allocated so no need to
4058 * update any list pointers.
4060 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4062 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4063 setup_kmalloc_cache_index_table();
4064 create_kmalloc_caches(0);
4066 /* Setup random freelists for each cache */
4067 init_freelist_randomization();
4070 register_cpu_notifier(&slab_notifier
);
4073 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4075 slub_min_order
, slub_max_order
, slub_min_objects
,
4076 nr_cpu_ids
, nr_node_ids
);
4079 void __init
kmem_cache_init_late(void)
4084 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
4085 unsigned long flags
, void (*ctor
)(void *))
4087 struct kmem_cache
*s
, *c
;
4089 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4094 * Adjust the object sizes so that we clear
4095 * the complete object on kzalloc.
4097 s
->object_size
= max(s
->object_size
, (int)size
);
4098 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4100 for_each_memcg_cache(c
, s
) {
4101 c
->object_size
= s
->object_size
;
4102 c
->inuse
= max_t(int, c
->inuse
,
4103 ALIGN(size
, sizeof(void *)));
4106 if (sysfs_slab_alias(s
, name
)) {
4115 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
4119 err
= kmem_cache_open(s
, flags
);
4123 /* Mutex is not taken during early boot */
4124 if (slab_state
<= UP
)
4127 memcg_propagate_slab_attrs(s
);
4128 err
= sysfs_slab_add(s
);
4130 __kmem_cache_release(s
);
4137 * Use the cpu notifier to insure that the cpu slabs are flushed when
4140 static int slab_cpuup_callback(struct notifier_block
*nfb
,
4141 unsigned long action
, void *hcpu
)
4143 long cpu
= (long)hcpu
;
4144 struct kmem_cache
*s
;
4145 unsigned long flags
;
4148 case CPU_UP_CANCELED
:
4149 case CPU_UP_CANCELED_FROZEN
:
4151 case CPU_DEAD_FROZEN
:
4152 mutex_lock(&slab_mutex
);
4153 list_for_each_entry(s
, &slab_caches
, list
) {
4154 local_irq_save(flags
);
4155 __flush_cpu_slab(s
, cpu
);
4156 local_irq_restore(flags
);
4158 mutex_unlock(&slab_mutex
);
4166 static struct notifier_block slab_notifier
= {
4167 .notifier_call
= slab_cpuup_callback
4172 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4174 struct kmem_cache
*s
;
4177 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4178 return kmalloc_large(size
, gfpflags
);
4180 s
= kmalloc_slab(size
, gfpflags
);
4182 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4185 ret
= slab_alloc(s
, gfpflags
, caller
);
4187 /* Honor the call site pointer we received. */
4188 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4194 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4195 int node
, unsigned long caller
)
4197 struct kmem_cache
*s
;
4200 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4201 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4203 trace_kmalloc_node(caller
, ret
,
4204 size
, PAGE_SIZE
<< get_order(size
),
4210 s
= kmalloc_slab(size
, gfpflags
);
4212 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4215 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4217 /* Honor the call site pointer we received. */
4218 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4225 static int count_inuse(struct page
*page
)
4230 static int count_total(struct page
*page
)
4232 return page
->objects
;
4236 #ifdef CONFIG_SLUB_DEBUG
4237 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4241 void *addr
= page_address(page
);
4243 if (!check_slab(s
, page
) ||
4244 !on_freelist(s
, page
, NULL
))
4247 /* Now we know that a valid freelist exists */
4248 bitmap_zero(map
, page
->objects
);
4250 get_map(s
, page
, map
);
4251 for_each_object(p
, s
, addr
, page
->objects
) {
4252 if (test_bit(slab_index(p
, s
, addr
), map
))
4253 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4257 for_each_object(p
, s
, addr
, page
->objects
)
4258 if (!test_bit(slab_index(p
, s
, addr
), map
))
4259 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4264 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4268 validate_slab(s
, page
, map
);
4272 static int validate_slab_node(struct kmem_cache
*s
,
4273 struct kmem_cache_node
*n
, unsigned long *map
)
4275 unsigned long count
= 0;
4277 unsigned long flags
;
4279 spin_lock_irqsave(&n
->list_lock
, flags
);
4281 list_for_each_entry(page
, &n
->partial
, lru
) {
4282 validate_slab_slab(s
, page
, map
);
4285 if (count
!= n
->nr_partial
)
4286 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4287 s
->name
, count
, n
->nr_partial
);
4289 if (!(s
->flags
& SLAB_STORE_USER
))
4292 list_for_each_entry(page
, &n
->full
, lru
) {
4293 validate_slab_slab(s
, page
, map
);
4296 if (count
!= atomic_long_read(&n
->nr_slabs
))
4297 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4298 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4301 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4305 static long validate_slab_cache(struct kmem_cache
*s
)
4308 unsigned long count
= 0;
4309 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4310 sizeof(unsigned long), GFP_KERNEL
);
4311 struct kmem_cache_node
*n
;
4317 for_each_kmem_cache_node(s
, node
, n
)
4318 count
+= validate_slab_node(s
, n
, map
);
4323 * Generate lists of code addresses where slabcache objects are allocated
4328 unsigned long count
;
4335 DECLARE_BITMAP(cpus
, NR_CPUS
);
4341 unsigned long count
;
4342 struct location
*loc
;
4345 static void free_loc_track(struct loc_track
*t
)
4348 free_pages((unsigned long)t
->loc
,
4349 get_order(sizeof(struct location
) * t
->max
));
4352 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4357 order
= get_order(sizeof(struct location
) * max
);
4359 l
= (void *)__get_free_pages(flags
, order
);
4364 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4372 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4373 const struct track
*track
)
4375 long start
, end
, pos
;
4377 unsigned long caddr
;
4378 unsigned long age
= jiffies
- track
->when
;
4384 pos
= start
+ (end
- start
+ 1) / 2;
4387 * There is nothing at "end". If we end up there
4388 * we need to add something to before end.
4393 caddr
= t
->loc
[pos
].addr
;
4394 if (track
->addr
== caddr
) {
4400 if (age
< l
->min_time
)
4402 if (age
> l
->max_time
)
4405 if (track
->pid
< l
->min_pid
)
4406 l
->min_pid
= track
->pid
;
4407 if (track
->pid
> l
->max_pid
)
4408 l
->max_pid
= track
->pid
;
4410 cpumask_set_cpu(track
->cpu
,
4411 to_cpumask(l
->cpus
));
4413 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4417 if (track
->addr
< caddr
)
4424 * Not found. Insert new tracking element.
4426 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4432 (t
->count
- pos
) * sizeof(struct location
));
4435 l
->addr
= track
->addr
;
4439 l
->min_pid
= track
->pid
;
4440 l
->max_pid
= track
->pid
;
4441 cpumask_clear(to_cpumask(l
->cpus
));
4442 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4443 nodes_clear(l
->nodes
);
4444 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4448 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4449 struct page
*page
, enum track_item alloc
,
4452 void *addr
= page_address(page
);
4455 bitmap_zero(map
, page
->objects
);
4456 get_map(s
, page
, map
);
4458 for_each_object(p
, s
, addr
, page
->objects
)
4459 if (!test_bit(slab_index(p
, s
, addr
), map
))
4460 add_location(t
, s
, get_track(s
, p
, alloc
));
4463 static int list_locations(struct kmem_cache
*s
, char *buf
,
4464 enum track_item alloc
)
4468 struct loc_track t
= { 0, 0, NULL
};
4470 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4471 sizeof(unsigned long), GFP_KERNEL
);
4472 struct kmem_cache_node
*n
;
4474 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4477 return sprintf(buf
, "Out of memory\n");
4479 /* Push back cpu slabs */
4482 for_each_kmem_cache_node(s
, node
, n
) {
4483 unsigned long flags
;
4486 if (!atomic_long_read(&n
->nr_slabs
))
4489 spin_lock_irqsave(&n
->list_lock
, flags
);
4490 list_for_each_entry(page
, &n
->partial
, lru
)
4491 process_slab(&t
, s
, page
, alloc
, map
);
4492 list_for_each_entry(page
, &n
->full
, lru
)
4493 process_slab(&t
, s
, page
, alloc
, map
);
4494 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4497 for (i
= 0; i
< t
.count
; i
++) {
4498 struct location
*l
= &t
.loc
[i
];
4500 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4502 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4505 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4507 len
+= sprintf(buf
+ len
, "<not-available>");
4509 if (l
->sum_time
!= l
->min_time
) {
4510 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4512 (long)div_u64(l
->sum_time
, l
->count
),
4515 len
+= sprintf(buf
+ len
, " age=%ld",
4518 if (l
->min_pid
!= l
->max_pid
)
4519 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4520 l
->min_pid
, l
->max_pid
);
4522 len
+= sprintf(buf
+ len
, " pid=%ld",
4525 if (num_online_cpus() > 1 &&
4526 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4527 len
< PAGE_SIZE
- 60)
4528 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4530 cpumask_pr_args(to_cpumask(l
->cpus
)));
4532 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4533 len
< PAGE_SIZE
- 60)
4534 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4536 nodemask_pr_args(&l
->nodes
));
4538 len
+= sprintf(buf
+ len
, "\n");
4544 len
+= sprintf(buf
, "No data\n");
4549 #ifdef SLUB_RESILIENCY_TEST
4550 static void __init
resiliency_test(void)
4554 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4556 pr_err("SLUB resiliency testing\n");
4557 pr_err("-----------------------\n");
4558 pr_err("A. Corruption after allocation\n");
4560 p
= kzalloc(16, GFP_KERNEL
);
4562 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4565 validate_slab_cache(kmalloc_caches
[4]);
4567 /* Hmmm... The next two are dangerous */
4568 p
= kzalloc(32, GFP_KERNEL
);
4569 p
[32 + sizeof(void *)] = 0x34;
4570 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4572 pr_err("If allocated object is overwritten then not detectable\n\n");
4574 validate_slab_cache(kmalloc_caches
[5]);
4575 p
= kzalloc(64, GFP_KERNEL
);
4576 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4578 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4580 pr_err("If allocated object is overwritten then not detectable\n\n");
4581 validate_slab_cache(kmalloc_caches
[6]);
4583 pr_err("\nB. Corruption after free\n");
4584 p
= kzalloc(128, GFP_KERNEL
);
4587 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4588 validate_slab_cache(kmalloc_caches
[7]);
4590 p
= kzalloc(256, GFP_KERNEL
);
4593 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4594 validate_slab_cache(kmalloc_caches
[8]);
4596 p
= kzalloc(512, GFP_KERNEL
);
4599 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4600 validate_slab_cache(kmalloc_caches
[9]);
4604 static void resiliency_test(void) {};
4609 enum slab_stat_type
{
4610 SL_ALL
, /* All slabs */
4611 SL_PARTIAL
, /* Only partially allocated slabs */
4612 SL_CPU
, /* Only slabs used for cpu caches */
4613 SL_OBJECTS
, /* Determine allocated objects not slabs */
4614 SL_TOTAL
/* Determine object capacity not slabs */
4617 #define SO_ALL (1 << SL_ALL)
4618 #define SO_PARTIAL (1 << SL_PARTIAL)
4619 #define SO_CPU (1 << SL_CPU)
4620 #define SO_OBJECTS (1 << SL_OBJECTS)
4621 #define SO_TOTAL (1 << SL_TOTAL)
4623 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4624 char *buf
, unsigned long flags
)
4626 unsigned long total
= 0;
4629 unsigned long *nodes
;
4631 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4635 if (flags
& SO_CPU
) {
4638 for_each_possible_cpu(cpu
) {
4639 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4644 page
= READ_ONCE(c
->page
);
4648 node
= page_to_nid(page
);
4649 if (flags
& SO_TOTAL
)
4651 else if (flags
& SO_OBJECTS
)
4659 page
= READ_ONCE(c
->partial
);
4661 node
= page_to_nid(page
);
4662 if (flags
& SO_TOTAL
)
4664 else if (flags
& SO_OBJECTS
)
4675 #ifdef CONFIG_SLUB_DEBUG
4676 if (flags
& SO_ALL
) {
4677 struct kmem_cache_node
*n
;
4679 for_each_kmem_cache_node(s
, node
, n
) {
4681 if (flags
& SO_TOTAL
)
4682 x
= atomic_long_read(&n
->total_objects
);
4683 else if (flags
& SO_OBJECTS
)
4684 x
= atomic_long_read(&n
->total_objects
) -
4685 count_partial(n
, count_free
);
4687 x
= atomic_long_read(&n
->nr_slabs
);
4694 if (flags
& SO_PARTIAL
) {
4695 struct kmem_cache_node
*n
;
4697 for_each_kmem_cache_node(s
, node
, n
) {
4698 if (flags
& SO_TOTAL
)
4699 x
= count_partial(n
, count_total
);
4700 else if (flags
& SO_OBJECTS
)
4701 x
= count_partial(n
, count_inuse
);
4708 x
= sprintf(buf
, "%lu", total
);
4710 for (node
= 0; node
< nr_node_ids
; node
++)
4712 x
+= sprintf(buf
+ x
, " N%d=%lu",
4717 return x
+ sprintf(buf
+ x
, "\n");
4720 #ifdef CONFIG_SLUB_DEBUG
4721 static int any_slab_objects(struct kmem_cache
*s
)
4724 struct kmem_cache_node
*n
;
4726 for_each_kmem_cache_node(s
, node
, n
)
4727 if (atomic_long_read(&n
->total_objects
))
4734 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4735 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4737 struct slab_attribute
{
4738 struct attribute attr
;
4739 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4740 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4743 #define SLAB_ATTR_RO(_name) \
4744 static struct slab_attribute _name##_attr = \
4745 __ATTR(_name, 0400, _name##_show, NULL)
4747 #define SLAB_ATTR(_name) \
4748 static struct slab_attribute _name##_attr = \
4749 __ATTR(_name, 0600, _name##_show, _name##_store)
4751 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4753 return sprintf(buf
, "%d\n", s
->size
);
4755 SLAB_ATTR_RO(slab_size
);
4757 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4759 return sprintf(buf
, "%d\n", s
->align
);
4761 SLAB_ATTR_RO(align
);
4763 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4765 return sprintf(buf
, "%d\n", s
->object_size
);
4767 SLAB_ATTR_RO(object_size
);
4769 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4771 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4773 SLAB_ATTR_RO(objs_per_slab
);
4775 static ssize_t
order_store(struct kmem_cache
*s
,
4776 const char *buf
, size_t length
)
4778 unsigned long order
;
4781 err
= kstrtoul(buf
, 10, &order
);
4785 if (order
> slub_max_order
|| order
< slub_min_order
)
4788 calculate_sizes(s
, order
);
4792 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4794 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4798 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4800 return sprintf(buf
, "%lu\n", s
->min_partial
);
4803 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4809 err
= kstrtoul(buf
, 10, &min
);
4813 set_min_partial(s
, min
);
4816 SLAB_ATTR(min_partial
);
4818 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4820 return sprintf(buf
, "%u\n", s
->cpu_partial
);
4823 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4826 unsigned long objects
;
4829 err
= kstrtoul(buf
, 10, &objects
);
4832 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4835 s
->cpu_partial
= objects
;
4839 SLAB_ATTR(cpu_partial
);
4841 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4845 return sprintf(buf
, "%pS\n", s
->ctor
);
4849 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4851 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4853 SLAB_ATTR_RO(aliases
);
4855 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4857 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4859 SLAB_ATTR_RO(partial
);
4861 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4863 return show_slab_objects(s
, buf
, SO_CPU
);
4865 SLAB_ATTR_RO(cpu_slabs
);
4867 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4869 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4871 SLAB_ATTR_RO(objects
);
4873 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4875 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4877 SLAB_ATTR_RO(objects_partial
);
4879 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4886 for_each_online_cpu(cpu
) {
4887 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
)->partial
;
4890 pages
+= page
->pages
;
4891 objects
+= page
->pobjects
;
4895 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
4898 for_each_online_cpu(cpu
) {
4899 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
) ->partial
;
4901 if (page
&& len
< PAGE_SIZE
- 20)
4902 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
4903 page
->pobjects
, page
->pages
);
4906 return len
+ sprintf(buf
+ len
, "\n");
4908 SLAB_ATTR_RO(slabs_cpu_partial
);
4910 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4912 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4915 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4916 const char *buf
, size_t length
)
4918 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4920 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4923 SLAB_ATTR(reclaim_account
);
4925 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4927 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4929 SLAB_ATTR_RO(hwcache_align
);
4931 #ifdef CONFIG_ZONE_DMA
4932 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4934 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4936 SLAB_ATTR_RO(cache_dma
);
4939 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4941 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4943 SLAB_ATTR_RO(destroy_by_rcu
);
4945 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4947 return sprintf(buf
, "%d\n", s
->reserved
);
4949 SLAB_ATTR_RO(reserved
);
4951 #ifdef CONFIG_SLUB_DEBUG
4952 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4954 return show_slab_objects(s
, buf
, SO_ALL
);
4956 SLAB_ATTR_RO(slabs
);
4958 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4960 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4962 SLAB_ATTR_RO(total_objects
);
4964 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4966 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
4969 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
4970 const char *buf
, size_t length
)
4972 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
4973 if (buf
[0] == '1') {
4974 s
->flags
&= ~__CMPXCHG_DOUBLE
;
4975 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
4979 SLAB_ATTR(sanity_checks
);
4981 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
4983 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
4986 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
4990 * Tracing a merged cache is going to give confusing results
4991 * as well as cause other issues like converting a mergeable
4992 * cache into an umergeable one.
4994 if (s
->refcount
> 1)
4997 s
->flags
&= ~SLAB_TRACE
;
4998 if (buf
[0] == '1') {
4999 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5000 s
->flags
|= SLAB_TRACE
;
5006 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5008 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5011 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5012 const char *buf
, size_t length
)
5014 if (any_slab_objects(s
))
5017 s
->flags
&= ~SLAB_RED_ZONE
;
5018 if (buf
[0] == '1') {
5019 s
->flags
|= SLAB_RED_ZONE
;
5021 calculate_sizes(s
, -1);
5024 SLAB_ATTR(red_zone
);
5026 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5028 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5031 static ssize_t
poison_store(struct kmem_cache
*s
,
5032 const char *buf
, size_t length
)
5034 if (any_slab_objects(s
))
5037 s
->flags
&= ~SLAB_POISON
;
5038 if (buf
[0] == '1') {
5039 s
->flags
|= SLAB_POISON
;
5041 calculate_sizes(s
, -1);
5046 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5048 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5051 static ssize_t
store_user_store(struct kmem_cache
*s
,
5052 const char *buf
, size_t length
)
5054 if (any_slab_objects(s
))
5057 s
->flags
&= ~SLAB_STORE_USER
;
5058 if (buf
[0] == '1') {
5059 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5060 s
->flags
|= SLAB_STORE_USER
;
5062 calculate_sizes(s
, -1);
5065 SLAB_ATTR(store_user
);
5067 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5072 static ssize_t
validate_store(struct kmem_cache
*s
,
5073 const char *buf
, size_t length
)
5077 if (buf
[0] == '1') {
5078 ret
= validate_slab_cache(s
);
5084 SLAB_ATTR(validate
);
5086 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5088 if (!(s
->flags
& SLAB_STORE_USER
))
5090 return list_locations(s
, buf
, TRACK_ALLOC
);
5092 SLAB_ATTR_RO(alloc_calls
);
5094 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5096 if (!(s
->flags
& SLAB_STORE_USER
))
5098 return list_locations(s
, buf
, TRACK_FREE
);
5100 SLAB_ATTR_RO(free_calls
);
5101 #endif /* CONFIG_SLUB_DEBUG */
5103 #ifdef CONFIG_FAILSLAB
5104 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5106 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5109 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5112 if (s
->refcount
> 1)
5115 s
->flags
&= ~SLAB_FAILSLAB
;
5117 s
->flags
|= SLAB_FAILSLAB
;
5120 SLAB_ATTR(failslab
);
5123 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5128 static ssize_t
shrink_store(struct kmem_cache
*s
,
5129 const char *buf
, size_t length
)
5132 kmem_cache_shrink(s
);
5140 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5142 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
5145 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5146 const char *buf
, size_t length
)
5148 unsigned long ratio
;
5151 err
= kstrtoul(buf
, 10, &ratio
);
5156 s
->remote_node_defrag_ratio
= ratio
* 10;
5160 SLAB_ATTR(remote_node_defrag_ratio
);
5163 #ifdef CONFIG_SLUB_STATS
5164 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5166 unsigned long sum
= 0;
5169 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5174 for_each_online_cpu(cpu
) {
5175 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5181 len
= sprintf(buf
, "%lu", sum
);
5184 for_each_online_cpu(cpu
) {
5185 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5186 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5190 return len
+ sprintf(buf
+ len
, "\n");
5193 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5197 for_each_online_cpu(cpu
)
5198 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5201 #define STAT_ATTR(si, text) \
5202 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5204 return show_stat(s, buf, si); \
5206 static ssize_t text##_store(struct kmem_cache *s, \
5207 const char *buf, size_t length) \
5209 if (buf[0] != '0') \
5211 clear_stat(s, si); \
5216 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5217 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5218 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5219 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5220 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5221 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5222 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5223 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5224 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5225 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5226 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5227 STAT_ATTR(FREE_SLAB
, free_slab
);
5228 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5229 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5230 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5231 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5232 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5233 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5234 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5235 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5236 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5237 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5238 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5239 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5240 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5241 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5244 static struct attribute
*slab_attrs
[] = {
5245 &slab_size_attr
.attr
,
5246 &object_size_attr
.attr
,
5247 &objs_per_slab_attr
.attr
,
5249 &min_partial_attr
.attr
,
5250 &cpu_partial_attr
.attr
,
5252 &objects_partial_attr
.attr
,
5254 &cpu_slabs_attr
.attr
,
5258 &hwcache_align_attr
.attr
,
5259 &reclaim_account_attr
.attr
,
5260 &destroy_by_rcu_attr
.attr
,
5262 &reserved_attr
.attr
,
5263 &slabs_cpu_partial_attr
.attr
,
5264 #ifdef CONFIG_SLUB_DEBUG
5265 &total_objects_attr
.attr
,
5267 &sanity_checks_attr
.attr
,
5269 &red_zone_attr
.attr
,
5271 &store_user_attr
.attr
,
5272 &validate_attr
.attr
,
5273 &alloc_calls_attr
.attr
,
5274 &free_calls_attr
.attr
,
5276 #ifdef CONFIG_ZONE_DMA
5277 &cache_dma_attr
.attr
,
5280 &remote_node_defrag_ratio_attr
.attr
,
5282 #ifdef CONFIG_SLUB_STATS
5283 &alloc_fastpath_attr
.attr
,
5284 &alloc_slowpath_attr
.attr
,
5285 &free_fastpath_attr
.attr
,
5286 &free_slowpath_attr
.attr
,
5287 &free_frozen_attr
.attr
,
5288 &free_add_partial_attr
.attr
,
5289 &free_remove_partial_attr
.attr
,
5290 &alloc_from_partial_attr
.attr
,
5291 &alloc_slab_attr
.attr
,
5292 &alloc_refill_attr
.attr
,
5293 &alloc_node_mismatch_attr
.attr
,
5294 &free_slab_attr
.attr
,
5295 &cpuslab_flush_attr
.attr
,
5296 &deactivate_full_attr
.attr
,
5297 &deactivate_empty_attr
.attr
,
5298 &deactivate_to_head_attr
.attr
,
5299 &deactivate_to_tail_attr
.attr
,
5300 &deactivate_remote_frees_attr
.attr
,
5301 &deactivate_bypass_attr
.attr
,
5302 &order_fallback_attr
.attr
,
5303 &cmpxchg_double_fail_attr
.attr
,
5304 &cmpxchg_double_cpu_fail_attr
.attr
,
5305 &cpu_partial_alloc_attr
.attr
,
5306 &cpu_partial_free_attr
.attr
,
5307 &cpu_partial_node_attr
.attr
,
5308 &cpu_partial_drain_attr
.attr
,
5310 #ifdef CONFIG_FAILSLAB
5311 &failslab_attr
.attr
,
5317 static struct attribute_group slab_attr_group
= {
5318 .attrs
= slab_attrs
,
5321 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5322 struct attribute
*attr
,
5325 struct slab_attribute
*attribute
;
5326 struct kmem_cache
*s
;
5329 attribute
= to_slab_attr(attr
);
5332 if (!attribute
->show
)
5335 err
= attribute
->show(s
, buf
);
5340 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5341 struct attribute
*attr
,
5342 const char *buf
, size_t len
)
5344 struct slab_attribute
*attribute
;
5345 struct kmem_cache
*s
;
5348 attribute
= to_slab_attr(attr
);
5351 if (!attribute
->store
)
5354 err
= attribute
->store(s
, buf
, len
);
5356 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5357 struct kmem_cache
*c
;
5359 mutex_lock(&slab_mutex
);
5360 if (s
->max_attr_size
< len
)
5361 s
->max_attr_size
= len
;
5364 * This is a best effort propagation, so this function's return
5365 * value will be determined by the parent cache only. This is
5366 * basically because not all attributes will have a well
5367 * defined semantics for rollbacks - most of the actions will
5368 * have permanent effects.
5370 * Returning the error value of any of the children that fail
5371 * is not 100 % defined, in the sense that users seeing the
5372 * error code won't be able to know anything about the state of
5375 * Only returning the error code for the parent cache at least
5376 * has well defined semantics. The cache being written to
5377 * directly either failed or succeeded, in which case we loop
5378 * through the descendants with best-effort propagation.
5380 for_each_memcg_cache(c
, s
)
5381 attribute
->store(c
, buf
, len
);
5382 mutex_unlock(&slab_mutex
);
5388 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5392 char *buffer
= NULL
;
5393 struct kmem_cache
*root_cache
;
5395 if (is_root_cache(s
))
5398 root_cache
= s
->memcg_params
.root_cache
;
5401 * This mean this cache had no attribute written. Therefore, no point
5402 * in copying default values around
5404 if (!root_cache
->max_attr_size
)
5407 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5410 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5412 if (!attr
|| !attr
->store
|| !attr
->show
)
5416 * It is really bad that we have to allocate here, so we will
5417 * do it only as a fallback. If we actually allocate, though,
5418 * we can just use the allocated buffer until the end.
5420 * Most of the slub attributes will tend to be very small in
5421 * size, but sysfs allows buffers up to a page, so they can
5422 * theoretically happen.
5426 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5429 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5430 if (WARN_ON(!buffer
))
5435 attr
->show(root_cache
, buf
);
5436 attr
->store(s
, buf
, strlen(buf
));
5440 free_page((unsigned long)buffer
);
5444 static void kmem_cache_release(struct kobject
*k
)
5446 slab_kmem_cache_release(to_slab(k
));
5449 static const struct sysfs_ops slab_sysfs_ops
= {
5450 .show
= slab_attr_show
,
5451 .store
= slab_attr_store
,
5454 static struct kobj_type slab_ktype
= {
5455 .sysfs_ops
= &slab_sysfs_ops
,
5456 .release
= kmem_cache_release
,
5459 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5461 struct kobj_type
*ktype
= get_ktype(kobj
);
5463 if (ktype
== &slab_ktype
)
5468 static const struct kset_uevent_ops slab_uevent_ops
= {
5469 .filter
= uevent_filter
,
5472 static struct kset
*slab_kset
;
5474 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5477 if (!is_root_cache(s
))
5478 return s
->memcg_params
.root_cache
->memcg_kset
;
5483 #define ID_STR_LENGTH 64
5485 /* Create a unique string id for a slab cache:
5487 * Format :[flags-]size
5489 static char *create_unique_id(struct kmem_cache
*s
)
5491 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5498 * First flags affecting slabcache operations. We will only
5499 * get here for aliasable slabs so we do not need to support
5500 * too many flags. The flags here must cover all flags that
5501 * are matched during merging to guarantee that the id is
5504 if (s
->flags
& SLAB_CACHE_DMA
)
5506 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5508 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5510 if (!(s
->flags
& SLAB_NOTRACK
))
5512 if (s
->flags
& SLAB_ACCOUNT
)
5516 p
+= sprintf(p
, "%07d", s
->size
);
5518 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5522 static int sysfs_slab_add(struct kmem_cache
*s
)
5526 int unmergeable
= slab_unmergeable(s
);
5530 * Slabcache can never be merged so we can use the name proper.
5531 * This is typically the case for debug situations. In that
5532 * case we can catch duplicate names easily.
5534 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5538 * Create a unique name for the slab as a target
5541 name
= create_unique_id(s
);
5544 s
->kobj
.kset
= cache_kset(s
);
5545 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5549 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5554 if (is_root_cache(s
)) {
5555 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5556 if (!s
->memcg_kset
) {
5563 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5565 /* Setup first alias */
5566 sysfs_slab_alias(s
, s
->name
);
5573 kobject_del(&s
->kobj
);
5577 void sysfs_slab_remove(struct kmem_cache
*s
)
5579 if (slab_state
< FULL
)
5581 * Sysfs has not been setup yet so no need to remove the
5587 kset_unregister(s
->memcg_kset
);
5589 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5590 kobject_del(&s
->kobj
);
5591 kobject_put(&s
->kobj
);
5595 * Need to buffer aliases during bootup until sysfs becomes
5596 * available lest we lose that information.
5598 struct saved_alias
{
5599 struct kmem_cache
*s
;
5601 struct saved_alias
*next
;
5604 static struct saved_alias
*alias_list
;
5606 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5608 struct saved_alias
*al
;
5610 if (slab_state
== FULL
) {
5612 * If we have a leftover link then remove it.
5614 sysfs_remove_link(&slab_kset
->kobj
, name
);
5615 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5618 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5624 al
->next
= alias_list
;
5629 static int __init
slab_sysfs_init(void)
5631 struct kmem_cache
*s
;
5634 mutex_lock(&slab_mutex
);
5636 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5638 mutex_unlock(&slab_mutex
);
5639 pr_err("Cannot register slab subsystem.\n");
5645 list_for_each_entry(s
, &slab_caches
, list
) {
5646 err
= sysfs_slab_add(s
);
5648 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5652 while (alias_list
) {
5653 struct saved_alias
*al
= alias_list
;
5655 alias_list
= alias_list
->next
;
5656 err
= sysfs_slab_alias(al
->s
, al
->name
);
5658 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5663 mutex_unlock(&slab_mutex
);
5668 __initcall(slab_sysfs_init
);
5669 #endif /* CONFIG_SYSFS */
5672 * The /proc/slabinfo ABI
5674 #ifdef CONFIG_SLABINFO
5675 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5677 unsigned long nr_slabs
= 0;
5678 unsigned long nr_objs
= 0;
5679 unsigned long nr_free
= 0;
5681 struct kmem_cache_node
*n
;
5683 for_each_kmem_cache_node(s
, node
, n
) {
5684 nr_slabs
+= node_nr_slabs(n
);
5685 nr_objs
+= node_nr_objs(n
);
5686 nr_free
+= count_partial(n
, count_free
);
5689 sinfo
->active_objs
= nr_objs
- nr_free
;
5690 sinfo
->num_objs
= nr_objs
;
5691 sinfo
->active_slabs
= nr_slabs
;
5692 sinfo
->num_slabs
= nr_slabs
;
5693 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5694 sinfo
->cache_order
= oo_order(s
->oo
);
5697 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5701 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5702 size_t count
, loff_t
*ppos
)
5706 #endif /* CONFIG_SLABINFO */