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 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
;
458 static int slub_debug
;
461 static char *slub_debug_slabs
;
462 static int disable_higher_order_debug
;
465 * slub is about to manipulate internal object metadata. This memory lies
466 * outside the range of the allocated object, so accessing it would normally
467 * be reported by kasan as a bounds error. metadata_access_enable() is used
468 * to tell kasan that these accesses are OK.
470 static inline void metadata_access_enable(void)
472 kasan_disable_current();
475 static inline void metadata_access_disable(void)
477 kasan_enable_current();
484 /* Verify that a pointer has an address that is valid within a slab page */
485 static inline int check_valid_pointer(struct kmem_cache
*s
,
486 struct page
*page
, void *object
)
493 base
= page_address(page
);
494 object
= restore_red_left(s
, object
);
495 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
496 (object
- base
) % s
->size
) {
503 static void print_section(char *text
, u8
*addr
, unsigned int length
)
505 metadata_access_enable();
506 print_hex_dump(KERN_ERR
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
508 metadata_access_disable();
511 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
512 enum track_item alloc
)
517 p
= object
+ s
->offset
+ sizeof(void *);
519 p
= object
+ s
->inuse
;
524 static void set_track(struct kmem_cache
*s
, void *object
,
525 enum track_item alloc
, unsigned long addr
)
527 struct track
*p
= get_track(s
, object
, alloc
);
530 #ifdef CONFIG_STACKTRACE
531 struct stack_trace trace
;
534 trace
.nr_entries
= 0;
535 trace
.max_entries
= TRACK_ADDRS_COUNT
;
536 trace
.entries
= p
->addrs
;
538 metadata_access_enable();
539 save_stack_trace(&trace
);
540 metadata_access_disable();
542 /* See rant in lockdep.c */
543 if (trace
.nr_entries
!= 0 &&
544 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
547 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
551 p
->cpu
= smp_processor_id();
552 p
->pid
= current
->pid
;
555 memset(p
, 0, sizeof(struct track
));
558 static void init_tracking(struct kmem_cache
*s
, void *object
)
560 if (!(s
->flags
& SLAB_STORE_USER
))
563 set_track(s
, object
, TRACK_FREE
, 0UL);
564 set_track(s
, object
, TRACK_ALLOC
, 0UL);
567 static void print_track(const char *s
, struct track
*t
)
572 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
573 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
574 #ifdef CONFIG_STACKTRACE
577 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
579 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
586 static void print_tracking(struct kmem_cache
*s
, void *object
)
588 if (!(s
->flags
& SLAB_STORE_USER
))
591 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
592 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
595 static void print_page_info(struct page
*page
)
597 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
598 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
602 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
604 struct va_format vaf
;
610 pr_err("=============================================================================\n");
611 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
612 pr_err("-----------------------------------------------------------------------------\n\n");
614 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
618 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
620 struct va_format vaf
;
626 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
630 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
632 unsigned int off
; /* Offset of last byte */
633 u8
*addr
= page_address(page
);
635 print_tracking(s
, p
);
637 print_page_info(page
);
639 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
640 p
, p
- addr
, get_freepointer(s
, p
));
642 if (s
->flags
& SLAB_RED_ZONE
)
643 print_section("Redzone ", p
- s
->red_left_pad
, s
->red_left_pad
);
644 else if (p
> addr
+ 16)
645 print_section("Bytes b4 ", p
- 16, 16);
647 print_section("Object ", p
, min_t(unsigned long, s
->object_size
,
649 if (s
->flags
& SLAB_RED_ZONE
)
650 print_section("Redzone ", p
+ s
->object_size
,
651 s
->inuse
- s
->object_size
);
654 off
= s
->offset
+ sizeof(void *);
658 if (s
->flags
& SLAB_STORE_USER
)
659 off
+= 2 * sizeof(struct track
);
661 off
+= kasan_metadata_size(s
);
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 off
+= kasan_metadata_size(s
);
792 if (size_from_object(s
) == off
)
795 return check_bytes_and_report(s
, page
, p
, "Object padding",
796 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
799 /* Check the pad bytes at the end of a slab page */
800 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
808 if (!(s
->flags
& SLAB_POISON
))
811 start
= page_address(page
);
812 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
813 end
= start
+ length
;
814 remainder
= length
% s
->size
;
818 metadata_access_enable();
819 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
820 metadata_access_disable();
823 while (end
> fault
&& end
[-1] == POISON_INUSE
)
826 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
827 print_section("Padding ", end
- remainder
, remainder
);
829 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
833 static int check_object(struct kmem_cache
*s
, struct page
*page
,
834 void *object
, u8 val
)
837 u8
*endobject
= object
+ s
->object_size
;
839 if (s
->flags
& SLAB_RED_ZONE
) {
840 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
841 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
844 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
845 endobject
, val
, s
->inuse
- s
->object_size
))
848 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
849 check_bytes_and_report(s
, page
, p
, "Alignment padding",
850 endobject
, POISON_INUSE
,
851 s
->inuse
- s
->object_size
);
855 if (s
->flags
& SLAB_POISON
) {
856 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
857 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
858 POISON_FREE
, s
->object_size
- 1) ||
859 !check_bytes_and_report(s
, page
, p
, "Poison",
860 p
+ s
->object_size
- 1, POISON_END
, 1)))
863 * check_pad_bytes cleans up on its own.
865 check_pad_bytes(s
, page
, p
);
868 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
870 * Object and freepointer overlap. Cannot check
871 * freepointer while object is allocated.
875 /* Check free pointer validity */
876 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
877 object_err(s
, page
, p
, "Freepointer corrupt");
879 * No choice but to zap it and thus lose the remainder
880 * of the free objects in this slab. May cause
881 * another error because the object count is now wrong.
883 set_freepointer(s
, p
, NULL
);
889 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
893 VM_BUG_ON(!irqs_disabled());
895 if (!PageSlab(page
)) {
896 slab_err(s
, page
, "Not a valid slab page");
900 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
901 if (page
->objects
> maxobj
) {
902 slab_err(s
, page
, "objects %u > max %u",
903 page
->objects
, maxobj
);
906 if (page
->inuse
> page
->objects
) {
907 slab_err(s
, page
, "inuse %u > max %u",
908 page
->inuse
, page
->objects
);
911 /* Slab_pad_check fixes things up after itself */
912 slab_pad_check(s
, page
);
917 * Determine if a certain object on a page is on the freelist. Must hold the
918 * slab lock to guarantee that the chains are in a consistent state.
920 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
928 while (fp
&& nr
<= page
->objects
) {
931 if (!check_valid_pointer(s
, page
, fp
)) {
933 object_err(s
, page
, object
,
934 "Freechain corrupt");
935 set_freepointer(s
, object
, NULL
);
937 slab_err(s
, page
, "Freepointer corrupt");
938 page
->freelist
= NULL
;
939 page
->inuse
= page
->objects
;
940 slab_fix(s
, "Freelist cleared");
946 fp
= get_freepointer(s
, object
);
950 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
951 if (max_objects
> MAX_OBJS_PER_PAGE
)
952 max_objects
= MAX_OBJS_PER_PAGE
;
954 if (page
->objects
!= max_objects
) {
955 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
956 page
->objects
, max_objects
);
957 page
->objects
= max_objects
;
958 slab_fix(s
, "Number of objects adjusted.");
960 if (page
->inuse
!= page
->objects
- nr
) {
961 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
962 page
->inuse
, page
->objects
- nr
);
963 page
->inuse
= page
->objects
- nr
;
964 slab_fix(s
, "Object count adjusted.");
966 return search
== NULL
;
969 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
972 if (s
->flags
& SLAB_TRACE
) {
973 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
975 alloc
? "alloc" : "free",
980 print_section("Object ", (void *)object
,
988 * Tracking of fully allocated slabs for debugging purposes.
990 static void add_full(struct kmem_cache
*s
,
991 struct kmem_cache_node
*n
, struct page
*page
)
993 if (!(s
->flags
& SLAB_STORE_USER
))
996 lockdep_assert_held(&n
->list_lock
);
997 list_add(&page
->lru
, &n
->full
);
1000 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1002 if (!(s
->flags
& SLAB_STORE_USER
))
1005 lockdep_assert_held(&n
->list_lock
);
1006 list_del(&page
->lru
);
1009 /* Tracking of the number of slabs for debugging purposes */
1010 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1012 struct kmem_cache_node
*n
= get_node(s
, node
);
1014 return atomic_long_read(&n
->nr_slabs
);
1017 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1019 return atomic_long_read(&n
->nr_slabs
);
1022 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1024 struct kmem_cache_node
*n
= get_node(s
, node
);
1027 * May be called early in order to allocate a slab for the
1028 * kmem_cache_node structure. Solve the chicken-egg
1029 * dilemma by deferring the increment of the count during
1030 * bootstrap (see early_kmem_cache_node_alloc).
1033 atomic_long_inc(&n
->nr_slabs
);
1034 atomic_long_add(objects
, &n
->total_objects
);
1037 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1039 struct kmem_cache_node
*n
= get_node(s
, node
);
1041 atomic_long_dec(&n
->nr_slabs
);
1042 atomic_long_sub(objects
, &n
->total_objects
);
1045 /* Object debug checks for alloc/free paths */
1046 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1049 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1052 init_object(s
, object
, SLUB_RED_INACTIVE
);
1053 init_tracking(s
, object
);
1056 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1058 void *object
, unsigned long addr
)
1060 if (!check_slab(s
, page
))
1063 if (!check_valid_pointer(s
, page
, object
)) {
1064 object_err(s
, page
, object
, "Freelist Pointer check fails");
1068 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1074 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1076 void *object
, unsigned long addr
)
1078 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1079 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1083 /* Success perform special debug activities for allocs */
1084 if (s
->flags
& SLAB_STORE_USER
)
1085 set_track(s
, object
, TRACK_ALLOC
, addr
);
1086 trace(s
, page
, object
, 1);
1087 init_object(s
, object
, SLUB_RED_ACTIVE
);
1091 if (PageSlab(page
)) {
1093 * If this is a slab page then lets do the best we can
1094 * to avoid issues in the future. Marking all objects
1095 * as used avoids touching the remaining objects.
1097 slab_fix(s
, "Marking all objects used");
1098 page
->inuse
= page
->objects
;
1099 page
->freelist
= NULL
;
1104 static inline int free_consistency_checks(struct kmem_cache
*s
,
1105 struct page
*page
, void *object
, unsigned long addr
)
1107 if (!check_valid_pointer(s
, page
, object
)) {
1108 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1112 if (on_freelist(s
, page
, object
)) {
1113 object_err(s
, page
, object
, "Object already free");
1117 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1120 if (unlikely(s
!= page
->slab_cache
)) {
1121 if (!PageSlab(page
)) {
1122 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1124 } else if (!page
->slab_cache
) {
1125 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1129 object_err(s
, page
, object
,
1130 "page slab pointer corrupt.");
1136 /* Supports checking bulk free of a constructed freelist */
1137 static noinline
int free_debug_processing(
1138 struct kmem_cache
*s
, struct page
*page
,
1139 void *head
, void *tail
, int bulk_cnt
,
1142 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1143 void *object
= head
;
1145 unsigned long uninitialized_var(flags
);
1148 spin_lock_irqsave(&n
->list_lock
, flags
);
1151 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1152 if (!check_slab(s
, page
))
1159 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1160 if (!free_consistency_checks(s
, page
, object
, addr
))
1164 if (s
->flags
& SLAB_STORE_USER
)
1165 set_track(s
, object
, TRACK_FREE
, addr
);
1166 trace(s
, page
, object
, 0);
1167 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1168 init_object(s
, object
, SLUB_RED_INACTIVE
);
1170 /* Reached end of constructed freelist yet? */
1171 if (object
!= tail
) {
1172 object
= get_freepointer(s
, object
);
1178 if (cnt
!= bulk_cnt
)
1179 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1183 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1185 slab_fix(s
, "Object at 0x%p not freed", object
);
1189 static int __init
setup_slub_debug(char *str
)
1191 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1192 if (*str
++ != '=' || !*str
)
1194 * No options specified. Switch on full debugging.
1200 * No options but restriction on slabs. This means full
1201 * debugging for slabs matching a pattern.
1208 * Switch off all debugging measures.
1213 * Determine which debug features should be switched on
1215 for (; *str
&& *str
!= ','; str
++) {
1216 switch (tolower(*str
)) {
1218 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1221 slub_debug
|= SLAB_RED_ZONE
;
1224 slub_debug
|= SLAB_POISON
;
1227 slub_debug
|= SLAB_STORE_USER
;
1230 slub_debug
|= SLAB_TRACE
;
1233 slub_debug
|= SLAB_FAILSLAB
;
1237 * Avoid enabling debugging on caches if its minimum
1238 * order would increase as a result.
1240 disable_higher_order_debug
= 1;
1243 pr_err("slub_debug option '%c' unknown. skipped\n",
1250 slub_debug_slabs
= str
+ 1;
1255 __setup("slub_debug", setup_slub_debug
);
1257 unsigned long kmem_cache_flags(unsigned long object_size
,
1258 unsigned long flags
, const char *name
,
1259 void (*ctor
)(void *))
1262 * Enable debugging if selected on the kernel commandline.
1264 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1265 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1266 flags
|= slub_debug
;
1270 #else /* !CONFIG_SLUB_DEBUG */
1271 static inline void setup_object_debug(struct kmem_cache
*s
,
1272 struct page
*page
, void *object
) {}
1274 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1275 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1277 static inline int free_debug_processing(
1278 struct kmem_cache
*s
, struct page
*page
,
1279 void *head
, void *tail
, int bulk_cnt
,
1280 unsigned long addr
) { return 0; }
1282 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1284 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1285 void *object
, u8 val
) { return 1; }
1286 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1287 struct page
*page
) {}
1288 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1289 struct page
*page
) {}
1290 unsigned long kmem_cache_flags(unsigned long object_size
,
1291 unsigned long flags
, const char *name
,
1292 void (*ctor
)(void *))
1296 #define slub_debug 0
1298 #define disable_higher_order_debug 0
1300 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1302 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1304 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1306 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1309 #endif /* CONFIG_SLUB_DEBUG */
1312 * Hooks for other subsystems that check memory allocations. In a typical
1313 * production configuration these hooks all should produce no code at all.
1315 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1317 kmemleak_alloc(ptr
, size
, 1, flags
);
1318 kasan_kmalloc_large(ptr
, size
, flags
);
1321 static inline void kfree_hook(const void *x
)
1324 kasan_kfree_large(x
);
1327 static inline void *slab_free_hook(struct kmem_cache
*s
, void *x
)
1331 kmemleak_free_recursive(x
, s
->flags
);
1334 * Trouble is that we may no longer disable interrupts in the fast path
1335 * So in order to make the debug calls that expect irqs to be
1336 * disabled we need to disable interrupts temporarily.
1338 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1340 unsigned long flags
;
1342 local_irq_save(flags
);
1343 kmemcheck_slab_free(s
, x
, s
->object_size
);
1344 debug_check_no_locks_freed(x
, s
->object_size
);
1345 local_irq_restore(flags
);
1348 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1349 debug_check_no_obj_freed(x
, s
->object_size
);
1351 freeptr
= get_freepointer(s
, x
);
1353 * kasan_slab_free() may put x into memory quarantine, delaying its
1354 * reuse. In this case the object's freelist pointer is changed.
1356 kasan_slab_free(s
, x
);
1360 static inline void slab_free_freelist_hook(struct kmem_cache
*s
,
1361 void *head
, void *tail
)
1364 * Compiler cannot detect this function can be removed if slab_free_hook()
1365 * evaluates to nothing. Thus, catch all relevant config debug options here.
1367 #if defined(CONFIG_KMEMCHECK) || \
1368 defined(CONFIG_LOCKDEP) || \
1369 defined(CONFIG_DEBUG_KMEMLEAK) || \
1370 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1371 defined(CONFIG_KASAN)
1373 void *object
= head
;
1374 void *tail_obj
= tail
? : head
;
1378 freeptr
= slab_free_hook(s
, object
);
1379 } while ((object
!= tail_obj
) && (object
= freeptr
));
1383 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1386 setup_object_debug(s
, page
, object
);
1387 if (unlikely(s
->ctor
)) {
1388 kasan_unpoison_object_data(s
, object
);
1390 kasan_poison_object_data(s
, object
);
1395 * Slab allocation and freeing
1397 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1398 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1401 int order
= oo_order(oo
);
1403 flags
|= __GFP_NOTRACK
;
1405 if (node
== NUMA_NO_NODE
)
1406 page
= alloc_pages(flags
, order
);
1408 page
= __alloc_pages_node(node
, flags
, order
);
1410 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1411 __free_pages(page
, order
);
1418 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1419 /* Pre-initialize the random sequence cache */
1420 static int init_cache_random_seq(struct kmem_cache
*s
)
1423 unsigned long i
, count
= oo_objects(s
->oo
);
1425 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1427 pr_err("SLUB: Unable to initialize free list for %s\n",
1432 /* Transform to an offset on the set of pages */
1433 if (s
->random_seq
) {
1434 for (i
= 0; i
< count
; i
++)
1435 s
->random_seq
[i
] *= s
->size
;
1440 /* Initialize each random sequence freelist per cache */
1441 static void __init
init_freelist_randomization(void)
1443 struct kmem_cache
*s
;
1445 mutex_lock(&slab_mutex
);
1447 list_for_each_entry(s
, &slab_caches
, list
)
1448 init_cache_random_seq(s
);
1450 mutex_unlock(&slab_mutex
);
1453 /* Get the next entry on the pre-computed freelist randomized */
1454 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1455 unsigned long *pos
, void *start
,
1456 unsigned long page_limit
,
1457 unsigned long freelist_count
)
1462 * If the target page allocation failed, the number of objects on the
1463 * page might be smaller than the usual size defined by the cache.
1466 idx
= s
->random_seq
[*pos
];
1468 if (*pos
>= freelist_count
)
1470 } while (unlikely(idx
>= page_limit
));
1472 return (char *)start
+ idx
;
1475 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1476 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1481 unsigned long idx
, pos
, page_limit
, freelist_count
;
1483 if (page
->objects
< 2 || !s
->random_seq
)
1486 freelist_count
= oo_objects(s
->oo
);
1487 pos
= get_random_int() % freelist_count
;
1489 page_limit
= page
->objects
* s
->size
;
1490 start
= fixup_red_left(s
, page_address(page
));
1492 /* First entry is used as the base of the freelist */
1493 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1495 page
->freelist
= cur
;
1497 for (idx
= 1; idx
< page
->objects
; idx
++) {
1498 setup_object(s
, page
, cur
);
1499 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1501 set_freepointer(s
, cur
, next
);
1504 setup_object(s
, page
, cur
);
1505 set_freepointer(s
, cur
, NULL
);
1510 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1514 static inline void init_freelist_randomization(void) { }
1515 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1519 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1521 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1524 struct kmem_cache_order_objects oo
= s
->oo
;
1530 flags
&= gfp_allowed_mask
;
1532 if (gfpflags_allow_blocking(flags
))
1535 flags
|= s
->allocflags
;
1538 * Let the initial higher-order allocation fail under memory pressure
1539 * so we fall-back to the minimum order allocation.
1541 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1542 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1543 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1545 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1546 if (unlikely(!page
)) {
1550 * Allocation may have failed due to fragmentation.
1551 * Try a lower order alloc if possible
1553 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1554 if (unlikely(!page
))
1556 stat(s
, ORDER_FALLBACK
);
1559 if (kmemcheck_enabled
&&
1560 !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1561 int pages
= 1 << oo_order(oo
);
1563 kmemcheck_alloc_shadow(page
, oo_order(oo
), alloc_gfp
, node
);
1566 * Objects from caches that have a constructor don't get
1567 * cleared when they're allocated, so we need to do it here.
1570 kmemcheck_mark_uninitialized_pages(page
, pages
);
1572 kmemcheck_mark_unallocated_pages(page
, pages
);
1575 page
->objects
= oo_objects(oo
);
1577 order
= compound_order(page
);
1578 page
->slab_cache
= s
;
1579 __SetPageSlab(page
);
1580 if (page_is_pfmemalloc(page
))
1581 SetPageSlabPfmemalloc(page
);
1583 start
= page_address(page
);
1585 if (unlikely(s
->flags
& SLAB_POISON
))
1586 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1588 kasan_poison_slab(page
);
1590 shuffle
= shuffle_freelist(s
, page
);
1593 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1594 setup_object(s
, page
, p
);
1595 if (likely(idx
< page
->objects
))
1596 set_freepointer(s
, p
, p
+ s
->size
);
1598 set_freepointer(s
, p
, NULL
);
1600 page
->freelist
= fixup_red_left(s
, start
);
1603 page
->inuse
= page
->objects
;
1607 if (gfpflags_allow_blocking(flags
))
1608 local_irq_disable();
1612 mod_zone_page_state(page_zone(page
),
1613 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1614 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1617 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1622 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1624 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1625 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1626 flags
&= ~GFP_SLAB_BUG_MASK
;
1627 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1628 invalid_mask
, &invalid_mask
, flags
, &flags
);
1631 return allocate_slab(s
,
1632 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1635 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1637 int order
= compound_order(page
);
1638 int pages
= 1 << order
;
1640 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1643 slab_pad_check(s
, page
);
1644 for_each_object(p
, s
, page_address(page
),
1646 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1649 kmemcheck_free_shadow(page
, compound_order(page
));
1651 mod_zone_page_state(page_zone(page
),
1652 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1653 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1656 __ClearPageSlabPfmemalloc(page
);
1657 __ClearPageSlab(page
);
1659 page_mapcount_reset(page
);
1660 if (current
->reclaim_state
)
1661 current
->reclaim_state
->reclaimed_slab
+= pages
;
1662 memcg_uncharge_slab(page
, order
, s
);
1663 __free_pages(page
, order
);
1666 #define need_reserve_slab_rcu \
1667 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1669 static void rcu_free_slab(struct rcu_head
*h
)
1673 if (need_reserve_slab_rcu
)
1674 page
= virt_to_head_page(h
);
1676 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1678 __free_slab(page
->slab_cache
, page
);
1681 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1683 if (unlikely(s
->flags
& SLAB_DESTROY_BY_RCU
)) {
1684 struct rcu_head
*head
;
1686 if (need_reserve_slab_rcu
) {
1687 int order
= compound_order(page
);
1688 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1690 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1691 head
= page_address(page
) + offset
;
1693 head
= &page
->rcu_head
;
1696 call_rcu(head
, rcu_free_slab
);
1698 __free_slab(s
, page
);
1701 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1703 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1708 * Management of partially allocated slabs.
1711 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1714 if (tail
== DEACTIVATE_TO_TAIL
)
1715 list_add_tail(&page
->lru
, &n
->partial
);
1717 list_add(&page
->lru
, &n
->partial
);
1720 static inline void add_partial(struct kmem_cache_node
*n
,
1721 struct page
*page
, int tail
)
1723 lockdep_assert_held(&n
->list_lock
);
1724 __add_partial(n
, page
, tail
);
1727 static inline void remove_partial(struct kmem_cache_node
*n
,
1730 lockdep_assert_held(&n
->list_lock
);
1731 list_del(&page
->lru
);
1736 * Remove slab from the partial list, freeze it and
1737 * return the pointer to the freelist.
1739 * Returns a list of objects or NULL if it fails.
1741 static inline void *acquire_slab(struct kmem_cache
*s
,
1742 struct kmem_cache_node
*n
, struct page
*page
,
1743 int mode
, int *objects
)
1746 unsigned long counters
;
1749 lockdep_assert_held(&n
->list_lock
);
1752 * Zap the freelist and set the frozen bit.
1753 * The old freelist is the list of objects for the
1754 * per cpu allocation list.
1756 freelist
= page
->freelist
;
1757 counters
= page
->counters
;
1758 new.counters
= counters
;
1759 *objects
= new.objects
- new.inuse
;
1761 new.inuse
= page
->objects
;
1762 new.freelist
= NULL
;
1764 new.freelist
= freelist
;
1767 VM_BUG_ON(new.frozen
);
1770 if (!__cmpxchg_double_slab(s
, page
,
1772 new.freelist
, new.counters
,
1776 remove_partial(n
, page
);
1781 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1782 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1785 * Try to allocate a partial slab from a specific node.
1787 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1788 struct kmem_cache_cpu
*c
, gfp_t flags
)
1790 struct page
*page
, *page2
;
1791 void *object
= NULL
;
1796 * Racy check. If we mistakenly see no partial slabs then we
1797 * just allocate an empty slab. If we mistakenly try to get a
1798 * partial slab and there is none available then get_partials()
1801 if (!n
|| !n
->nr_partial
)
1804 spin_lock(&n
->list_lock
);
1805 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1808 if (!pfmemalloc_match(page
, flags
))
1811 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1815 available
+= objects
;
1818 stat(s
, ALLOC_FROM_PARTIAL
);
1821 put_cpu_partial(s
, page
, 0);
1822 stat(s
, CPU_PARTIAL_NODE
);
1824 if (!kmem_cache_has_cpu_partial(s
)
1825 || available
> s
->cpu_partial
/ 2)
1829 spin_unlock(&n
->list_lock
);
1834 * Get a page from somewhere. Search in increasing NUMA distances.
1836 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1837 struct kmem_cache_cpu
*c
)
1840 struct zonelist
*zonelist
;
1843 enum zone_type high_zoneidx
= gfp_zone(flags
);
1845 unsigned int cpuset_mems_cookie
;
1848 * The defrag ratio allows a configuration of the tradeoffs between
1849 * inter node defragmentation and node local allocations. A lower
1850 * defrag_ratio increases the tendency to do local allocations
1851 * instead of attempting to obtain partial slabs from other nodes.
1853 * If the defrag_ratio is set to 0 then kmalloc() always
1854 * returns node local objects. If the ratio is higher then kmalloc()
1855 * may return off node objects because partial slabs are obtained
1856 * from other nodes and filled up.
1858 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1859 * (which makes defrag_ratio = 1000) then every (well almost)
1860 * allocation will first attempt to defrag slab caches on other nodes.
1861 * This means scanning over all nodes to look for partial slabs which
1862 * may be expensive if we do it every time we are trying to find a slab
1863 * with available objects.
1865 if (!s
->remote_node_defrag_ratio
||
1866 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1870 cpuset_mems_cookie
= read_mems_allowed_begin();
1871 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1872 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1873 struct kmem_cache_node
*n
;
1875 n
= get_node(s
, zone_to_nid(zone
));
1877 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1878 n
->nr_partial
> s
->min_partial
) {
1879 object
= get_partial_node(s
, n
, c
, flags
);
1882 * Don't check read_mems_allowed_retry()
1883 * here - if mems_allowed was updated in
1884 * parallel, that was a harmless race
1885 * between allocation and the cpuset
1892 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1898 * Get a partial page, lock it and return it.
1900 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1901 struct kmem_cache_cpu
*c
)
1904 int searchnode
= node
;
1906 if (node
== NUMA_NO_NODE
)
1907 searchnode
= numa_mem_id();
1908 else if (!node_present_pages(node
))
1909 searchnode
= node_to_mem_node(node
);
1911 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1912 if (object
|| node
!= NUMA_NO_NODE
)
1915 return get_any_partial(s
, flags
, c
);
1918 #ifdef CONFIG_PREEMPT
1920 * Calculate the next globally unique transaction for disambiguiation
1921 * during cmpxchg. The transactions start with the cpu number and are then
1922 * incremented by CONFIG_NR_CPUS.
1924 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1927 * No preemption supported therefore also no need to check for
1933 static inline unsigned long next_tid(unsigned long tid
)
1935 return tid
+ TID_STEP
;
1938 static inline unsigned int tid_to_cpu(unsigned long tid
)
1940 return tid
% TID_STEP
;
1943 static inline unsigned long tid_to_event(unsigned long tid
)
1945 return tid
/ TID_STEP
;
1948 static inline unsigned int init_tid(int cpu
)
1953 static inline void note_cmpxchg_failure(const char *n
,
1954 const struct kmem_cache
*s
, unsigned long tid
)
1956 #ifdef SLUB_DEBUG_CMPXCHG
1957 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1959 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1961 #ifdef CONFIG_PREEMPT
1962 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1963 pr_warn("due to cpu change %d -> %d\n",
1964 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1967 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1968 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1969 tid_to_event(tid
), tid_to_event(actual_tid
));
1971 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1972 actual_tid
, tid
, next_tid(tid
));
1974 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1977 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1981 for_each_possible_cpu(cpu
)
1982 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1986 * Remove the cpu slab
1988 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1991 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1992 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1994 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1996 int tail
= DEACTIVATE_TO_HEAD
;
2000 if (page
->freelist
) {
2001 stat(s
, DEACTIVATE_REMOTE_FREES
);
2002 tail
= DEACTIVATE_TO_TAIL
;
2006 * Stage one: Free all available per cpu objects back
2007 * to the page freelist while it is still frozen. Leave the
2010 * There is no need to take the list->lock because the page
2013 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2015 unsigned long counters
;
2018 prior
= page
->freelist
;
2019 counters
= page
->counters
;
2020 set_freepointer(s
, freelist
, prior
);
2021 new.counters
= counters
;
2023 VM_BUG_ON(!new.frozen
);
2025 } while (!__cmpxchg_double_slab(s
, page
,
2027 freelist
, new.counters
,
2028 "drain percpu freelist"));
2030 freelist
= nextfree
;
2034 * Stage two: Ensure that the page is unfrozen while the
2035 * list presence reflects the actual number of objects
2038 * We setup the list membership and then perform a cmpxchg
2039 * with the count. If there is a mismatch then the page
2040 * is not unfrozen but the page is on the wrong list.
2042 * Then we restart the process which may have to remove
2043 * the page from the list that we just put it on again
2044 * because the number of objects in the slab may have
2049 old
.freelist
= page
->freelist
;
2050 old
.counters
= page
->counters
;
2051 VM_BUG_ON(!old
.frozen
);
2053 /* Determine target state of the slab */
2054 new.counters
= old
.counters
;
2057 set_freepointer(s
, freelist
, old
.freelist
);
2058 new.freelist
= freelist
;
2060 new.freelist
= old
.freelist
;
2064 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2066 else if (new.freelist
) {
2071 * Taking the spinlock removes the possiblity
2072 * that acquire_slab() will see a slab page that
2075 spin_lock(&n
->list_lock
);
2079 if (kmem_cache_debug(s
) && !lock
) {
2082 * This also ensures that the scanning of full
2083 * slabs from diagnostic functions will not see
2086 spin_lock(&n
->list_lock
);
2094 remove_partial(n
, page
);
2096 else if (l
== M_FULL
)
2098 remove_full(s
, n
, page
);
2100 if (m
== M_PARTIAL
) {
2102 add_partial(n
, page
, tail
);
2105 } else if (m
== M_FULL
) {
2107 stat(s
, DEACTIVATE_FULL
);
2108 add_full(s
, n
, page
);
2114 if (!__cmpxchg_double_slab(s
, page
,
2115 old
.freelist
, old
.counters
,
2116 new.freelist
, new.counters
,
2121 spin_unlock(&n
->list_lock
);
2124 stat(s
, DEACTIVATE_EMPTY
);
2125 discard_slab(s
, page
);
2131 * Unfreeze all the cpu partial slabs.
2133 * This function must be called with interrupts disabled
2134 * for the cpu using c (or some other guarantee must be there
2135 * to guarantee no concurrent accesses).
2137 static void unfreeze_partials(struct kmem_cache
*s
,
2138 struct kmem_cache_cpu
*c
)
2140 #ifdef CONFIG_SLUB_CPU_PARTIAL
2141 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2142 struct page
*page
, *discard_page
= NULL
;
2144 while ((page
= c
->partial
)) {
2148 c
->partial
= page
->next
;
2150 n2
= get_node(s
, page_to_nid(page
));
2153 spin_unlock(&n
->list_lock
);
2156 spin_lock(&n
->list_lock
);
2161 old
.freelist
= page
->freelist
;
2162 old
.counters
= page
->counters
;
2163 VM_BUG_ON(!old
.frozen
);
2165 new.counters
= old
.counters
;
2166 new.freelist
= old
.freelist
;
2170 } while (!__cmpxchg_double_slab(s
, page
,
2171 old
.freelist
, old
.counters
,
2172 new.freelist
, new.counters
,
2173 "unfreezing slab"));
2175 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2176 page
->next
= discard_page
;
2177 discard_page
= page
;
2179 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2180 stat(s
, FREE_ADD_PARTIAL
);
2185 spin_unlock(&n
->list_lock
);
2187 while (discard_page
) {
2188 page
= discard_page
;
2189 discard_page
= discard_page
->next
;
2191 stat(s
, DEACTIVATE_EMPTY
);
2192 discard_slab(s
, page
);
2199 * Put a page that was just frozen (in __slab_free) into a partial page
2200 * slot if available. This is done without interrupts disabled and without
2201 * preemption disabled. The cmpxchg is racy and may put the partial page
2202 * onto a random cpus partial slot.
2204 * If we did not find a slot then simply move all the partials to the
2205 * per node partial list.
2207 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2209 #ifdef CONFIG_SLUB_CPU_PARTIAL
2210 struct page
*oldpage
;
2218 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2221 pobjects
= oldpage
->pobjects
;
2222 pages
= oldpage
->pages
;
2223 if (drain
&& pobjects
> s
->cpu_partial
) {
2224 unsigned long flags
;
2226 * partial array is full. Move the existing
2227 * set to the per node partial list.
2229 local_irq_save(flags
);
2230 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2231 local_irq_restore(flags
);
2235 stat(s
, CPU_PARTIAL_DRAIN
);
2240 pobjects
+= page
->objects
- page
->inuse
;
2242 page
->pages
= pages
;
2243 page
->pobjects
= pobjects
;
2244 page
->next
= oldpage
;
2246 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2248 if (unlikely(!s
->cpu_partial
)) {
2249 unsigned long flags
;
2251 local_irq_save(flags
);
2252 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2253 local_irq_restore(flags
);
2259 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2261 stat(s
, CPUSLAB_FLUSH
);
2262 deactivate_slab(s
, c
->page
, c
->freelist
);
2264 c
->tid
= next_tid(c
->tid
);
2272 * Called from IPI handler with interrupts disabled.
2274 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2276 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2282 unfreeze_partials(s
, c
);
2286 static void flush_cpu_slab(void *d
)
2288 struct kmem_cache
*s
= d
;
2290 __flush_cpu_slab(s
, smp_processor_id());
2293 static bool has_cpu_slab(int cpu
, void *info
)
2295 struct kmem_cache
*s
= info
;
2296 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2298 return c
->page
|| c
->partial
;
2301 static void flush_all(struct kmem_cache
*s
)
2303 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2307 * Check if the objects in a per cpu structure fit numa
2308 * locality expectations.
2310 static inline int node_match(struct page
*page
, int node
)
2313 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2319 #ifdef CONFIG_SLUB_DEBUG
2320 static int count_free(struct page
*page
)
2322 return page
->objects
- page
->inuse
;
2325 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2327 return atomic_long_read(&n
->total_objects
);
2329 #endif /* CONFIG_SLUB_DEBUG */
2331 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2332 static unsigned long count_partial(struct kmem_cache_node
*n
,
2333 int (*get_count
)(struct page
*))
2335 unsigned long flags
;
2336 unsigned long x
= 0;
2339 spin_lock_irqsave(&n
->list_lock
, flags
);
2340 list_for_each_entry(page
, &n
->partial
, lru
)
2341 x
+= get_count(page
);
2342 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2345 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2347 static noinline
void
2348 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2350 #ifdef CONFIG_SLUB_DEBUG
2351 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2352 DEFAULT_RATELIMIT_BURST
);
2354 struct kmem_cache_node
*n
;
2356 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2359 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2360 nid
, gfpflags
, &gfpflags
);
2361 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2362 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2365 if (oo_order(s
->min
) > get_order(s
->object_size
))
2366 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2369 for_each_kmem_cache_node(s
, node
, n
) {
2370 unsigned long nr_slabs
;
2371 unsigned long nr_objs
;
2372 unsigned long nr_free
;
2374 nr_free
= count_partial(n
, count_free
);
2375 nr_slabs
= node_nr_slabs(n
);
2376 nr_objs
= node_nr_objs(n
);
2378 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2379 node
, nr_slabs
, nr_objs
, nr_free
);
2384 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2385 int node
, struct kmem_cache_cpu
**pc
)
2388 struct kmem_cache_cpu
*c
= *pc
;
2391 freelist
= get_partial(s
, flags
, node
, c
);
2396 page
= new_slab(s
, flags
, node
);
2398 c
= raw_cpu_ptr(s
->cpu_slab
);
2403 * No other reference to the page yet so we can
2404 * muck around with it freely without cmpxchg
2406 freelist
= page
->freelist
;
2407 page
->freelist
= NULL
;
2409 stat(s
, ALLOC_SLAB
);
2418 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2420 if (unlikely(PageSlabPfmemalloc(page
)))
2421 return gfp_pfmemalloc_allowed(gfpflags
);
2427 * Check the page->freelist of a page and either transfer the freelist to the
2428 * per cpu freelist or deactivate the page.
2430 * The page is still frozen if the return value is not NULL.
2432 * If this function returns NULL then the page has been unfrozen.
2434 * This function must be called with interrupt disabled.
2436 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2439 unsigned long counters
;
2443 freelist
= page
->freelist
;
2444 counters
= page
->counters
;
2446 new.counters
= counters
;
2447 VM_BUG_ON(!new.frozen
);
2449 new.inuse
= page
->objects
;
2450 new.frozen
= freelist
!= NULL
;
2452 } while (!__cmpxchg_double_slab(s
, page
,
2461 * Slow path. The lockless freelist is empty or we need to perform
2464 * Processing is still very fast if new objects have been freed to the
2465 * regular freelist. In that case we simply take over the regular freelist
2466 * as the lockless freelist and zap the regular freelist.
2468 * If that is not working then we fall back to the partial lists. We take the
2469 * first element of the freelist as the object to allocate now and move the
2470 * rest of the freelist to the lockless freelist.
2472 * And if we were unable to get a new slab from the partial slab lists then
2473 * we need to allocate a new slab. This is the slowest path since it involves
2474 * a call to the page allocator and the setup of a new slab.
2476 * Version of __slab_alloc to use when we know that interrupts are
2477 * already disabled (which is the case for bulk allocation).
2479 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2480 unsigned long addr
, struct kmem_cache_cpu
*c
)
2490 if (unlikely(!node_match(page
, node
))) {
2491 int searchnode
= node
;
2493 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2494 searchnode
= node_to_mem_node(node
);
2496 if (unlikely(!node_match(page
, searchnode
))) {
2497 stat(s
, ALLOC_NODE_MISMATCH
);
2498 deactivate_slab(s
, page
, c
->freelist
);
2506 * By rights, we should be searching for a slab page that was
2507 * PFMEMALLOC but right now, we are losing the pfmemalloc
2508 * information when the page leaves the per-cpu allocator
2510 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2511 deactivate_slab(s
, page
, c
->freelist
);
2517 /* must check again c->freelist in case of cpu migration or IRQ */
2518 freelist
= c
->freelist
;
2522 freelist
= get_freelist(s
, page
);
2526 stat(s
, DEACTIVATE_BYPASS
);
2530 stat(s
, ALLOC_REFILL
);
2534 * freelist is pointing to the list of objects to be used.
2535 * page is pointing to the page from which the objects are obtained.
2536 * That page must be frozen for per cpu allocations to work.
2538 VM_BUG_ON(!c
->page
->frozen
);
2539 c
->freelist
= get_freepointer(s
, freelist
);
2540 c
->tid
= next_tid(c
->tid
);
2546 page
= c
->page
= c
->partial
;
2547 c
->partial
= page
->next
;
2548 stat(s
, CPU_PARTIAL_ALLOC
);
2553 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2555 if (unlikely(!freelist
)) {
2556 slab_out_of_memory(s
, gfpflags
, node
);
2561 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2564 /* Only entered in the debug case */
2565 if (kmem_cache_debug(s
) &&
2566 !alloc_debug_processing(s
, page
, freelist
, addr
))
2567 goto new_slab
; /* Slab failed checks. Next slab needed */
2569 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
2576 * Another one that disabled interrupt and compensates for possible
2577 * cpu changes by refetching the per cpu area pointer.
2579 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2580 unsigned long addr
, struct kmem_cache_cpu
*c
)
2583 unsigned long flags
;
2585 local_irq_save(flags
);
2586 #ifdef CONFIG_PREEMPT
2588 * We may have been preempted and rescheduled on a different
2589 * cpu before disabling interrupts. Need to reload cpu area
2592 c
= this_cpu_ptr(s
->cpu_slab
);
2595 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2596 local_irq_restore(flags
);
2601 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2602 * have the fastpath folded into their functions. So no function call
2603 * overhead for requests that can be satisfied on the fastpath.
2605 * The fastpath works by first checking if the lockless freelist can be used.
2606 * If not then __slab_alloc is called for slow processing.
2608 * Otherwise we can simply pick the next object from the lockless free list.
2610 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2611 gfp_t gfpflags
, int node
, unsigned long addr
)
2614 struct kmem_cache_cpu
*c
;
2618 s
= slab_pre_alloc_hook(s
, gfpflags
);
2623 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2624 * enabled. We may switch back and forth between cpus while
2625 * reading from one cpu area. That does not matter as long
2626 * as we end up on the original cpu again when doing the cmpxchg.
2628 * We should guarantee that tid and kmem_cache are retrieved on
2629 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2630 * to check if it is matched or not.
2633 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2634 c
= raw_cpu_ptr(s
->cpu_slab
);
2635 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2636 unlikely(tid
!= READ_ONCE(c
->tid
)));
2639 * Irqless object alloc/free algorithm used here depends on sequence
2640 * of fetching cpu_slab's data. tid should be fetched before anything
2641 * on c to guarantee that object and page associated with previous tid
2642 * won't be used with current tid. If we fetch tid first, object and
2643 * page could be one associated with next tid and our alloc/free
2644 * request will be failed. In this case, we will retry. So, no problem.
2649 * The transaction ids are globally unique per cpu and per operation on
2650 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2651 * occurs on the right processor and that there was no operation on the
2652 * linked list in between.
2655 object
= c
->freelist
;
2657 if (unlikely(!object
|| !node_match(page
, node
))) {
2658 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2659 stat(s
, ALLOC_SLOWPATH
);
2661 void *next_object
= get_freepointer_safe(s
, object
);
2664 * The cmpxchg will only match if there was no additional
2665 * operation and if we are on the right processor.
2667 * The cmpxchg does the following atomically (without lock
2669 * 1. Relocate first pointer to the current per cpu area.
2670 * 2. Verify that tid and freelist have not been changed
2671 * 3. If they were not changed replace tid and freelist
2673 * Since this is without lock semantics the protection is only
2674 * against code executing on this cpu *not* from access by
2677 if (unlikely(!this_cpu_cmpxchg_double(
2678 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2680 next_object
, next_tid(tid
)))) {
2682 note_cmpxchg_failure("slab_alloc", s
, tid
);
2685 prefetch_freepointer(s
, next_object
);
2686 stat(s
, ALLOC_FASTPATH
);
2689 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2690 memset(object
, 0, s
->object_size
);
2692 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2697 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2698 gfp_t gfpflags
, unsigned long addr
)
2700 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2703 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2705 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2707 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2712 EXPORT_SYMBOL(kmem_cache_alloc
);
2714 #ifdef CONFIG_TRACING
2715 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2717 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2718 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2719 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2722 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2726 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2728 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2730 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2731 s
->object_size
, s
->size
, gfpflags
, node
);
2735 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2737 #ifdef CONFIG_TRACING
2738 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2740 int node
, size_t size
)
2742 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2744 trace_kmalloc_node(_RET_IP_
, ret
,
2745 size
, s
->size
, gfpflags
, node
);
2747 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2750 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2755 * Slow path handling. This may still be called frequently since objects
2756 * have a longer lifetime than the cpu slabs in most processing loads.
2758 * So we still attempt to reduce cache line usage. Just take the slab
2759 * lock and free the item. If there is no additional partial page
2760 * handling required then we can return immediately.
2762 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2763 void *head
, void *tail
, int cnt
,
2770 unsigned long counters
;
2771 struct kmem_cache_node
*n
= NULL
;
2772 unsigned long uninitialized_var(flags
);
2774 stat(s
, FREE_SLOWPATH
);
2776 if (kmem_cache_debug(s
) &&
2777 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2782 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2785 prior
= page
->freelist
;
2786 counters
= page
->counters
;
2787 set_freepointer(s
, tail
, prior
);
2788 new.counters
= counters
;
2789 was_frozen
= new.frozen
;
2791 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2793 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2796 * Slab was on no list before and will be
2798 * We can defer the list move and instead
2803 } else { /* Needs to be taken off a list */
2805 n
= get_node(s
, page_to_nid(page
));
2807 * Speculatively acquire the list_lock.
2808 * If the cmpxchg does not succeed then we may
2809 * drop the list_lock without any processing.
2811 * Otherwise the list_lock will synchronize with
2812 * other processors updating the list of slabs.
2814 spin_lock_irqsave(&n
->list_lock
, flags
);
2819 } while (!cmpxchg_double_slab(s
, page
,
2827 * If we just froze the page then put it onto the
2828 * per cpu partial list.
2830 if (new.frozen
&& !was_frozen
) {
2831 put_cpu_partial(s
, page
, 1);
2832 stat(s
, CPU_PARTIAL_FREE
);
2835 * The list lock was not taken therefore no list
2836 * activity can be necessary.
2839 stat(s
, FREE_FROZEN
);
2843 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2847 * Objects left in the slab. If it was not on the partial list before
2850 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2851 if (kmem_cache_debug(s
))
2852 remove_full(s
, n
, page
);
2853 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2854 stat(s
, FREE_ADD_PARTIAL
);
2856 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2862 * Slab on the partial list.
2864 remove_partial(n
, page
);
2865 stat(s
, FREE_REMOVE_PARTIAL
);
2867 /* Slab must be on the full list */
2868 remove_full(s
, n
, page
);
2871 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2873 discard_slab(s
, page
);
2877 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2878 * can perform fastpath freeing without additional function calls.
2880 * The fastpath is only possible if we are freeing to the current cpu slab
2881 * of this processor. This typically the case if we have just allocated
2884 * If fastpath is not possible then fall back to __slab_free where we deal
2885 * with all sorts of special processing.
2887 * Bulk free of a freelist with several objects (all pointing to the
2888 * same page) possible by specifying head and tail ptr, plus objects
2889 * count (cnt). Bulk free indicated by tail pointer being set.
2891 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2892 struct page
*page
, void *head
, void *tail
,
2893 int cnt
, unsigned long addr
)
2895 void *tail_obj
= tail
? : head
;
2896 struct kmem_cache_cpu
*c
;
2900 * Determine the currently cpus per cpu slab.
2901 * The cpu may change afterward. However that does not matter since
2902 * data is retrieved via this pointer. If we are on the same cpu
2903 * during the cmpxchg then the free will succeed.
2906 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2907 c
= raw_cpu_ptr(s
->cpu_slab
);
2908 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2909 unlikely(tid
!= READ_ONCE(c
->tid
)));
2911 /* Same with comment on barrier() in slab_alloc_node() */
2914 if (likely(page
== c
->page
)) {
2915 set_freepointer(s
, tail_obj
, c
->freelist
);
2917 if (unlikely(!this_cpu_cmpxchg_double(
2918 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2920 head
, next_tid(tid
)))) {
2922 note_cmpxchg_failure("slab_free", s
, tid
);
2925 stat(s
, FREE_FASTPATH
);
2927 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2931 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2932 void *head
, void *tail
, int cnt
,
2935 slab_free_freelist_hook(s
, head
, tail
);
2937 * slab_free_freelist_hook() could have put the items into quarantine.
2938 * If so, no need to free them.
2940 if (s
->flags
& SLAB_KASAN
&& !(s
->flags
& SLAB_DESTROY_BY_RCU
))
2942 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2946 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2948 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
2952 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2954 s
= cache_from_obj(s
, x
);
2957 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
2958 trace_kmem_cache_free(_RET_IP_
, x
);
2960 EXPORT_SYMBOL(kmem_cache_free
);
2962 struct detached_freelist
{
2967 struct kmem_cache
*s
;
2971 * This function progressively scans the array with free objects (with
2972 * a limited look ahead) and extract objects belonging to the same
2973 * page. It builds a detached freelist directly within the given
2974 * page/objects. This can happen without any need for
2975 * synchronization, because the objects are owned by running process.
2976 * The freelist is build up as a single linked list in the objects.
2977 * The idea is, that this detached freelist can then be bulk
2978 * transferred to the real freelist(s), but only requiring a single
2979 * synchronization primitive. Look ahead in the array is limited due
2980 * to performance reasons.
2983 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
2984 void **p
, struct detached_freelist
*df
)
2986 size_t first_skipped_index
= 0;
2991 /* Always re-init detached_freelist */
2996 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2997 } while (!object
&& size
);
3002 page
= virt_to_head_page(object
);
3004 /* Handle kalloc'ed objects */
3005 if (unlikely(!PageSlab(page
))) {
3006 BUG_ON(!PageCompound(page
));
3008 __free_pages(page
, compound_order(page
));
3009 p
[size
] = NULL
; /* mark object processed */
3012 /* Derive kmem_cache from object */
3013 df
->s
= page
->slab_cache
;
3015 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3018 /* Start new detached freelist */
3020 set_freepointer(df
->s
, object
, NULL
);
3022 df
->freelist
= object
;
3023 p
[size
] = NULL
; /* mark object processed */
3029 continue; /* Skip processed objects */
3031 /* df->page is always set at this point */
3032 if (df
->page
== virt_to_head_page(object
)) {
3033 /* Opportunity build freelist */
3034 set_freepointer(df
->s
, object
, df
->freelist
);
3035 df
->freelist
= object
;
3037 p
[size
] = NULL
; /* mark object processed */
3042 /* Limit look ahead search */
3046 if (!first_skipped_index
)
3047 first_skipped_index
= size
+ 1;
3050 return first_skipped_index
;
3053 /* Note that interrupts must be enabled when calling this function. */
3054 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3060 struct detached_freelist df
;
3062 size
= build_detached_freelist(s
, size
, p
, &df
);
3063 if (unlikely(!df
.page
))
3066 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3067 } while (likely(size
));
3069 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3071 /* Note that interrupts must be enabled when calling this function. */
3072 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3075 struct kmem_cache_cpu
*c
;
3078 /* memcg and kmem_cache debug support */
3079 s
= slab_pre_alloc_hook(s
, flags
);
3083 * Drain objects in the per cpu slab, while disabling local
3084 * IRQs, which protects against PREEMPT and interrupts
3085 * handlers invoking normal fastpath.
3087 local_irq_disable();
3088 c
= this_cpu_ptr(s
->cpu_slab
);
3090 for (i
= 0; i
< size
; i
++) {
3091 void *object
= c
->freelist
;
3093 if (unlikely(!object
)) {
3095 * Invoking slow path likely have side-effect
3096 * of re-populating per CPU c->freelist
3098 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3100 if (unlikely(!p
[i
]))
3103 c
= this_cpu_ptr(s
->cpu_slab
);
3104 continue; /* goto for-loop */
3106 c
->freelist
= get_freepointer(s
, object
);
3109 c
->tid
= next_tid(c
->tid
);
3112 /* Clear memory outside IRQ disabled fastpath loop */
3113 if (unlikely(flags
& __GFP_ZERO
)) {
3116 for (j
= 0; j
< i
; j
++)
3117 memset(p
[j
], 0, s
->object_size
);
3120 /* memcg and kmem_cache debug support */
3121 slab_post_alloc_hook(s
, flags
, size
, p
);
3125 slab_post_alloc_hook(s
, flags
, i
, p
);
3126 __kmem_cache_free_bulk(s
, i
, p
);
3129 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3133 * Object placement in a slab is made very easy because we always start at
3134 * offset 0. If we tune the size of the object to the alignment then we can
3135 * get the required alignment by putting one properly sized object after
3138 * Notice that the allocation order determines the sizes of the per cpu
3139 * caches. Each processor has always one slab available for allocations.
3140 * Increasing the allocation order reduces the number of times that slabs
3141 * must be moved on and off the partial lists and is therefore a factor in
3146 * Mininum / Maximum order of slab pages. This influences locking overhead
3147 * and slab fragmentation. A higher order reduces the number of partial slabs
3148 * and increases the number of allocations possible without having to
3149 * take the list_lock.
3151 static int slub_min_order
;
3152 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3153 static int slub_min_objects
;
3156 * Calculate the order of allocation given an slab object size.
3158 * The order of allocation has significant impact on performance and other
3159 * system components. Generally order 0 allocations should be preferred since
3160 * order 0 does not cause fragmentation in the page allocator. Larger objects
3161 * be problematic to put into order 0 slabs because there may be too much
3162 * unused space left. We go to a higher order if more than 1/16th of the slab
3165 * In order to reach satisfactory performance we must ensure that a minimum
3166 * number of objects is in one slab. Otherwise we may generate too much
3167 * activity on the partial lists which requires taking the list_lock. This is
3168 * less a concern for large slabs though which are rarely used.
3170 * slub_max_order specifies the order where we begin to stop considering the
3171 * number of objects in a slab as critical. If we reach slub_max_order then
3172 * we try to keep the page order as low as possible. So we accept more waste
3173 * of space in favor of a small page order.
3175 * Higher order allocations also allow the placement of more objects in a
3176 * slab and thereby reduce object handling overhead. If the user has
3177 * requested a higher mininum order then we start with that one instead of
3178 * the smallest order which will fit the object.
3180 static inline int slab_order(int size
, int min_objects
,
3181 int max_order
, int fract_leftover
, int reserved
)
3185 int min_order
= slub_min_order
;
3187 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3188 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3190 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3191 order
<= max_order
; order
++) {
3193 unsigned long slab_size
= PAGE_SIZE
<< order
;
3195 rem
= (slab_size
- reserved
) % size
;
3197 if (rem
<= slab_size
/ fract_leftover
)
3204 static inline int calculate_order(int size
, int reserved
)
3212 * Attempt to find best configuration for a slab. This
3213 * works by first attempting to generate a layout with
3214 * the best configuration and backing off gradually.
3216 * First we increase the acceptable waste in a slab. Then
3217 * we reduce the minimum objects required in a slab.
3219 min_objects
= slub_min_objects
;
3221 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3222 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3223 min_objects
= min(min_objects
, max_objects
);
3225 while (min_objects
> 1) {
3227 while (fraction
>= 4) {
3228 order
= slab_order(size
, min_objects
,
3229 slub_max_order
, fraction
, reserved
);
3230 if (order
<= slub_max_order
)
3238 * We were unable to place multiple objects in a slab. Now
3239 * lets see if we can place a single object there.
3241 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3242 if (order
<= slub_max_order
)
3246 * Doh this slab cannot be placed using slub_max_order.
3248 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3249 if (order
< MAX_ORDER
)
3255 init_kmem_cache_node(struct kmem_cache_node
*n
)
3258 spin_lock_init(&n
->list_lock
);
3259 INIT_LIST_HEAD(&n
->partial
);
3260 #ifdef CONFIG_SLUB_DEBUG
3261 atomic_long_set(&n
->nr_slabs
, 0);
3262 atomic_long_set(&n
->total_objects
, 0);
3263 INIT_LIST_HEAD(&n
->full
);
3267 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3269 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3270 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3273 * Must align to double word boundary for the double cmpxchg
3274 * instructions to work; see __pcpu_double_call_return_bool().
3276 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3277 2 * sizeof(void *));
3282 init_kmem_cache_cpus(s
);
3287 static struct kmem_cache
*kmem_cache_node
;
3290 * No kmalloc_node yet so do it by hand. We know that this is the first
3291 * slab on the node for this slabcache. There are no concurrent accesses
3294 * Note that this function only works on the kmem_cache_node
3295 * when allocating for the kmem_cache_node. This is used for bootstrapping
3296 * memory on a fresh node that has no slab structures yet.
3298 static void early_kmem_cache_node_alloc(int node
)
3301 struct kmem_cache_node
*n
;
3303 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3305 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3308 if (page_to_nid(page
) != node
) {
3309 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3310 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3315 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3318 kmem_cache_node
->node
[node
] = n
;
3319 #ifdef CONFIG_SLUB_DEBUG
3320 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3321 init_tracking(kmem_cache_node
, n
);
3323 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3325 init_kmem_cache_node(n
);
3326 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3329 * No locks need to be taken here as it has just been
3330 * initialized and there is no concurrent access.
3332 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3335 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3338 struct kmem_cache_node
*n
;
3340 for_each_kmem_cache_node(s
, node
, n
) {
3341 kmem_cache_free(kmem_cache_node
, n
);
3342 s
->node
[node
] = NULL
;
3346 void __kmem_cache_release(struct kmem_cache
*s
)
3348 cache_random_seq_destroy(s
);
3349 free_percpu(s
->cpu_slab
);
3350 free_kmem_cache_nodes(s
);
3353 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3357 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3358 struct kmem_cache_node
*n
;
3360 if (slab_state
== DOWN
) {
3361 early_kmem_cache_node_alloc(node
);
3364 n
= kmem_cache_alloc_node(kmem_cache_node
,
3368 free_kmem_cache_nodes(s
);
3373 init_kmem_cache_node(n
);
3378 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3380 if (min
< MIN_PARTIAL
)
3382 else if (min
> MAX_PARTIAL
)
3384 s
->min_partial
= min
;
3388 * calculate_sizes() determines the order and the distribution of data within
3391 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3393 unsigned long flags
= s
->flags
;
3394 size_t size
= s
->object_size
;
3398 * Round up object size to the next word boundary. We can only
3399 * place the free pointer at word boundaries and this determines
3400 * the possible location of the free pointer.
3402 size
= ALIGN(size
, sizeof(void *));
3404 #ifdef CONFIG_SLUB_DEBUG
3406 * Determine if we can poison the object itself. If the user of
3407 * the slab may touch the object after free or before allocation
3408 * then we should never poison the object itself.
3410 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_DESTROY_BY_RCU
) &&
3412 s
->flags
|= __OBJECT_POISON
;
3414 s
->flags
&= ~__OBJECT_POISON
;
3418 * If we are Redzoning then check if there is some space between the
3419 * end of the object and the free pointer. If not then add an
3420 * additional word to have some bytes to store Redzone information.
3422 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3423 size
+= sizeof(void *);
3427 * With that we have determined the number of bytes in actual use
3428 * by the object. This is the potential offset to the free pointer.
3432 if (((flags
& (SLAB_DESTROY_BY_RCU
| SLAB_POISON
)) ||
3435 * Relocate free pointer after the object if it is not
3436 * permitted to overwrite the first word of the object on
3439 * This is the case if we do RCU, have a constructor or
3440 * destructor or are poisoning the objects.
3443 size
+= sizeof(void *);
3446 #ifdef CONFIG_SLUB_DEBUG
3447 if (flags
& SLAB_STORE_USER
)
3449 * Need to store information about allocs and frees after
3452 size
+= 2 * sizeof(struct track
);
3455 kasan_cache_create(s
, &size
, &s
->flags
);
3456 #ifdef CONFIG_SLUB_DEBUG
3457 if (flags
& SLAB_RED_ZONE
) {
3459 * Add some empty padding so that we can catch
3460 * overwrites from earlier objects rather than let
3461 * tracking information or the free pointer be
3462 * corrupted if a user writes before the start
3465 size
+= sizeof(void *);
3467 s
->red_left_pad
= sizeof(void *);
3468 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3469 size
+= s
->red_left_pad
;
3474 * SLUB stores one object immediately after another beginning from
3475 * offset 0. In order to align the objects we have to simply size
3476 * each object to conform to the alignment.
3478 size
= ALIGN(size
, s
->align
);
3480 if (forced_order
>= 0)
3481 order
= forced_order
;
3483 order
= calculate_order(size
, s
->reserved
);
3490 s
->allocflags
|= __GFP_COMP
;
3492 if (s
->flags
& SLAB_CACHE_DMA
)
3493 s
->allocflags
|= GFP_DMA
;
3495 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3496 s
->allocflags
|= __GFP_RECLAIMABLE
;
3499 * Determine the number of objects per slab
3501 s
->oo
= oo_make(order
, size
, s
->reserved
);
3502 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3503 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3506 return !!oo_objects(s
->oo
);
3509 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3511 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3514 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_DESTROY_BY_RCU
))
3515 s
->reserved
= sizeof(struct rcu_head
);
3517 if (!calculate_sizes(s
, -1))
3519 if (disable_higher_order_debug
) {
3521 * Disable debugging flags that store metadata if the min slab
3524 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3525 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3527 if (!calculate_sizes(s
, -1))
3532 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3533 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3534 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3535 /* Enable fast mode */
3536 s
->flags
|= __CMPXCHG_DOUBLE
;
3540 * The larger the object size is, the more pages we want on the partial
3541 * list to avoid pounding the page allocator excessively.
3543 set_min_partial(s
, ilog2(s
->size
) / 2);
3546 * cpu_partial determined the maximum number of objects kept in the
3547 * per cpu partial lists of a processor.
3549 * Per cpu partial lists mainly contain slabs that just have one
3550 * object freed. If they are used for allocation then they can be
3551 * filled up again with minimal effort. The slab will never hit the
3552 * per node partial lists and therefore no locking will be required.
3554 * This setting also determines
3556 * A) The number of objects from per cpu partial slabs dumped to the
3557 * per node list when we reach the limit.
3558 * B) The number of objects in cpu partial slabs to extract from the
3559 * per node list when we run out of per cpu objects. We only fetch
3560 * 50% to keep some capacity around for frees.
3562 if (!kmem_cache_has_cpu_partial(s
))
3564 else if (s
->size
>= PAGE_SIZE
)
3566 else if (s
->size
>= 1024)
3568 else if (s
->size
>= 256)
3569 s
->cpu_partial
= 13;
3571 s
->cpu_partial
= 30;
3574 s
->remote_node_defrag_ratio
= 1000;
3577 /* Initialize the pre-computed randomized freelist if slab is up */
3578 if (slab_state
>= UP
) {
3579 if (init_cache_random_seq(s
))
3583 if (!init_kmem_cache_nodes(s
))
3586 if (alloc_kmem_cache_cpus(s
))
3589 free_kmem_cache_nodes(s
);
3591 if (flags
& SLAB_PANIC
)
3592 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3593 s
->name
, (unsigned long)s
->size
, s
->size
,
3594 oo_order(s
->oo
), s
->offset
, flags
);
3598 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3601 #ifdef CONFIG_SLUB_DEBUG
3602 void *addr
= page_address(page
);
3604 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3605 sizeof(long), GFP_ATOMIC
);
3608 slab_err(s
, page
, text
, s
->name
);
3611 get_map(s
, page
, map
);
3612 for_each_object(p
, s
, addr
, page
->objects
) {
3614 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3615 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3616 print_tracking(s
, p
);
3625 * Attempt to free all partial slabs on a node.
3626 * This is called from __kmem_cache_shutdown(). We must take list_lock
3627 * because sysfs file might still access partial list after the shutdowning.
3629 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3631 struct page
*page
, *h
;
3633 BUG_ON(irqs_disabled());
3634 spin_lock_irq(&n
->list_lock
);
3635 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3637 remove_partial(n
, page
);
3638 discard_slab(s
, page
);
3640 list_slab_objects(s
, page
,
3641 "Objects remaining in %s on __kmem_cache_shutdown()");
3644 spin_unlock_irq(&n
->list_lock
);
3648 * Release all resources used by a slab cache.
3650 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3653 struct kmem_cache_node
*n
;
3656 /* Attempt to free all objects */
3657 for_each_kmem_cache_node(s
, node
, n
) {
3659 if (n
->nr_partial
|| slabs_node(s
, node
))
3665 /********************************************************************
3667 *******************************************************************/
3669 static int __init
setup_slub_min_order(char *str
)
3671 get_option(&str
, &slub_min_order
);
3676 __setup("slub_min_order=", setup_slub_min_order
);
3678 static int __init
setup_slub_max_order(char *str
)
3680 get_option(&str
, &slub_max_order
);
3681 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3686 __setup("slub_max_order=", setup_slub_max_order
);
3688 static int __init
setup_slub_min_objects(char *str
)
3690 get_option(&str
, &slub_min_objects
);
3695 __setup("slub_min_objects=", setup_slub_min_objects
);
3697 void *__kmalloc(size_t size
, gfp_t flags
)
3699 struct kmem_cache
*s
;
3702 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3703 return kmalloc_large(size
, flags
);
3705 s
= kmalloc_slab(size
, flags
);
3707 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3710 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3712 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3714 kasan_kmalloc(s
, ret
, size
, flags
);
3718 EXPORT_SYMBOL(__kmalloc
);
3721 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3726 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
3727 page
= alloc_pages_node(node
, flags
, get_order(size
));
3729 ptr
= page_address(page
);
3731 kmalloc_large_node_hook(ptr
, size
, flags
);
3735 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3737 struct kmem_cache
*s
;
3740 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3741 ret
= kmalloc_large_node(size
, flags
, node
);
3743 trace_kmalloc_node(_RET_IP_
, ret
,
3744 size
, PAGE_SIZE
<< get_order(size
),
3750 s
= kmalloc_slab(size
, flags
);
3752 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3755 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3757 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3759 kasan_kmalloc(s
, ret
, size
, flags
);
3763 EXPORT_SYMBOL(__kmalloc_node
);
3766 static size_t __ksize(const void *object
)
3770 if (unlikely(object
== ZERO_SIZE_PTR
))
3773 page
= virt_to_head_page(object
);
3775 if (unlikely(!PageSlab(page
))) {
3776 WARN_ON(!PageCompound(page
));
3777 return PAGE_SIZE
<< compound_order(page
);
3780 return slab_ksize(page
->slab_cache
);
3783 size_t ksize(const void *object
)
3785 size_t size
= __ksize(object
);
3786 /* We assume that ksize callers could use whole allocated area,
3787 * so we need to unpoison this area.
3789 kasan_unpoison_shadow(object
, size
);
3792 EXPORT_SYMBOL(ksize
);
3794 void kfree(const void *x
)
3797 void *object
= (void *)x
;
3799 trace_kfree(_RET_IP_
, x
);
3801 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3804 page
= virt_to_head_page(x
);
3805 if (unlikely(!PageSlab(page
))) {
3806 BUG_ON(!PageCompound(page
));
3808 __free_pages(page
, compound_order(page
));
3811 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3813 EXPORT_SYMBOL(kfree
);
3815 #define SHRINK_PROMOTE_MAX 32
3818 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3819 * up most to the head of the partial lists. New allocations will then
3820 * fill those up and thus they can be removed from the partial lists.
3822 * The slabs with the least items are placed last. This results in them
3823 * being allocated from last increasing the chance that the last objects
3824 * are freed in them.
3826 int __kmem_cache_shrink(struct kmem_cache
*s
, bool deactivate
)
3830 struct kmem_cache_node
*n
;
3833 struct list_head discard
;
3834 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3835 unsigned long flags
;
3840 * Disable empty slabs caching. Used to avoid pinning offline
3841 * memory cgroups by kmem pages that can be freed.
3847 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3848 * so we have to make sure the change is visible.
3850 synchronize_sched();
3854 for_each_kmem_cache_node(s
, node
, n
) {
3855 INIT_LIST_HEAD(&discard
);
3856 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3857 INIT_LIST_HEAD(promote
+ i
);
3859 spin_lock_irqsave(&n
->list_lock
, flags
);
3862 * Build lists of slabs to discard or promote.
3864 * Note that concurrent frees may occur while we hold the
3865 * list_lock. page->inuse here is the upper limit.
3867 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3868 int free
= page
->objects
- page
->inuse
;
3870 /* Do not reread page->inuse */
3873 /* We do not keep full slabs on the list */
3876 if (free
== page
->objects
) {
3877 list_move(&page
->lru
, &discard
);
3879 } else if (free
<= SHRINK_PROMOTE_MAX
)
3880 list_move(&page
->lru
, promote
+ free
- 1);
3884 * Promote the slabs filled up most to the head of the
3887 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3888 list_splice(promote
+ i
, &n
->partial
);
3890 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3892 /* Release empty slabs */
3893 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3894 discard_slab(s
, page
);
3896 if (slabs_node(s
, node
))
3903 static int slab_mem_going_offline_callback(void *arg
)
3905 struct kmem_cache
*s
;
3907 mutex_lock(&slab_mutex
);
3908 list_for_each_entry(s
, &slab_caches
, list
)
3909 __kmem_cache_shrink(s
, false);
3910 mutex_unlock(&slab_mutex
);
3915 static void slab_mem_offline_callback(void *arg
)
3917 struct kmem_cache_node
*n
;
3918 struct kmem_cache
*s
;
3919 struct memory_notify
*marg
= arg
;
3922 offline_node
= marg
->status_change_nid_normal
;
3925 * If the node still has available memory. we need kmem_cache_node
3928 if (offline_node
< 0)
3931 mutex_lock(&slab_mutex
);
3932 list_for_each_entry(s
, &slab_caches
, list
) {
3933 n
= get_node(s
, offline_node
);
3936 * if n->nr_slabs > 0, slabs still exist on the node
3937 * that is going down. We were unable to free them,
3938 * and offline_pages() function shouldn't call this
3939 * callback. So, we must fail.
3941 BUG_ON(slabs_node(s
, offline_node
));
3943 s
->node
[offline_node
] = NULL
;
3944 kmem_cache_free(kmem_cache_node
, n
);
3947 mutex_unlock(&slab_mutex
);
3950 static int slab_mem_going_online_callback(void *arg
)
3952 struct kmem_cache_node
*n
;
3953 struct kmem_cache
*s
;
3954 struct memory_notify
*marg
= arg
;
3955 int nid
= marg
->status_change_nid_normal
;
3959 * If the node's memory is already available, then kmem_cache_node is
3960 * already created. Nothing to do.
3966 * We are bringing a node online. No memory is available yet. We must
3967 * allocate a kmem_cache_node structure in order to bring the node
3970 mutex_lock(&slab_mutex
);
3971 list_for_each_entry(s
, &slab_caches
, list
) {
3973 * XXX: kmem_cache_alloc_node will fallback to other nodes
3974 * since memory is not yet available from the node that
3977 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
3982 init_kmem_cache_node(n
);
3986 mutex_unlock(&slab_mutex
);
3990 static int slab_memory_callback(struct notifier_block
*self
,
3991 unsigned long action
, void *arg
)
3996 case MEM_GOING_ONLINE
:
3997 ret
= slab_mem_going_online_callback(arg
);
3999 case MEM_GOING_OFFLINE
:
4000 ret
= slab_mem_going_offline_callback(arg
);
4003 case MEM_CANCEL_ONLINE
:
4004 slab_mem_offline_callback(arg
);
4007 case MEM_CANCEL_OFFLINE
:
4011 ret
= notifier_from_errno(ret
);
4017 static struct notifier_block slab_memory_callback_nb
= {
4018 .notifier_call
= slab_memory_callback
,
4019 .priority
= SLAB_CALLBACK_PRI
,
4022 /********************************************************************
4023 * Basic setup of slabs
4024 *******************************************************************/
4027 * Used for early kmem_cache structures that were allocated using
4028 * the page allocator. Allocate them properly then fix up the pointers
4029 * that may be pointing to the wrong kmem_cache structure.
4032 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4035 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4036 struct kmem_cache_node
*n
;
4038 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4041 * This runs very early, and only the boot processor is supposed to be
4042 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4045 __flush_cpu_slab(s
, smp_processor_id());
4046 for_each_kmem_cache_node(s
, node
, n
) {
4049 list_for_each_entry(p
, &n
->partial
, lru
)
4052 #ifdef CONFIG_SLUB_DEBUG
4053 list_for_each_entry(p
, &n
->full
, lru
)
4057 slab_init_memcg_params(s
);
4058 list_add(&s
->list
, &slab_caches
);
4062 void __init
kmem_cache_init(void)
4064 static __initdata
struct kmem_cache boot_kmem_cache
,
4065 boot_kmem_cache_node
;
4067 if (debug_guardpage_minorder())
4070 kmem_cache_node
= &boot_kmem_cache_node
;
4071 kmem_cache
= &boot_kmem_cache
;
4073 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4074 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
4076 register_hotmemory_notifier(&slab_memory_callback_nb
);
4078 /* Able to allocate the per node structures */
4079 slab_state
= PARTIAL
;
4081 create_boot_cache(kmem_cache
, "kmem_cache",
4082 offsetof(struct kmem_cache
, node
) +
4083 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4084 SLAB_HWCACHE_ALIGN
);
4086 kmem_cache
= bootstrap(&boot_kmem_cache
);
4089 * Allocate kmem_cache_node properly from the kmem_cache slab.
4090 * kmem_cache_node is separately allocated so no need to
4091 * update any list pointers.
4093 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4095 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4096 setup_kmalloc_cache_index_table();
4097 create_kmalloc_caches(0);
4099 /* Setup random freelists for each cache */
4100 init_freelist_randomization();
4103 register_cpu_notifier(&slab_notifier
);
4106 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4108 slub_min_order
, slub_max_order
, slub_min_objects
,
4109 nr_cpu_ids
, nr_node_ids
);
4112 void __init
kmem_cache_init_late(void)
4117 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
4118 unsigned long flags
, void (*ctor
)(void *))
4120 struct kmem_cache
*s
, *c
;
4122 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4127 * Adjust the object sizes so that we clear
4128 * the complete object on kzalloc.
4130 s
->object_size
= max(s
->object_size
, (int)size
);
4131 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4133 for_each_memcg_cache(c
, s
) {
4134 c
->object_size
= s
->object_size
;
4135 c
->inuse
= max_t(int, c
->inuse
,
4136 ALIGN(size
, sizeof(void *)));
4139 if (sysfs_slab_alias(s
, name
)) {
4148 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
4152 err
= kmem_cache_open(s
, flags
);
4156 /* Mutex is not taken during early boot */
4157 if (slab_state
<= UP
)
4160 memcg_propagate_slab_attrs(s
);
4161 err
= sysfs_slab_add(s
);
4163 __kmem_cache_release(s
);
4170 * Use the cpu notifier to insure that the cpu slabs are flushed when
4173 static int slab_cpuup_callback(struct notifier_block
*nfb
,
4174 unsigned long action
, void *hcpu
)
4176 long cpu
= (long)hcpu
;
4177 struct kmem_cache
*s
;
4178 unsigned long flags
;
4181 case CPU_UP_CANCELED
:
4182 case CPU_UP_CANCELED_FROZEN
:
4184 case CPU_DEAD_FROZEN
:
4185 mutex_lock(&slab_mutex
);
4186 list_for_each_entry(s
, &slab_caches
, list
) {
4187 local_irq_save(flags
);
4188 __flush_cpu_slab(s
, cpu
);
4189 local_irq_restore(flags
);
4191 mutex_unlock(&slab_mutex
);
4199 static struct notifier_block slab_notifier
= {
4200 .notifier_call
= slab_cpuup_callback
4205 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4207 struct kmem_cache
*s
;
4210 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4211 return kmalloc_large(size
, gfpflags
);
4213 s
= kmalloc_slab(size
, gfpflags
);
4215 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4218 ret
= slab_alloc(s
, gfpflags
, caller
);
4220 /* Honor the call site pointer we received. */
4221 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4227 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4228 int node
, unsigned long caller
)
4230 struct kmem_cache
*s
;
4233 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4234 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4236 trace_kmalloc_node(caller
, ret
,
4237 size
, PAGE_SIZE
<< get_order(size
),
4243 s
= kmalloc_slab(size
, gfpflags
);
4245 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4248 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4250 /* Honor the call site pointer we received. */
4251 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4258 static int count_inuse(struct page
*page
)
4263 static int count_total(struct page
*page
)
4265 return page
->objects
;
4269 #ifdef CONFIG_SLUB_DEBUG
4270 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4274 void *addr
= page_address(page
);
4276 if (!check_slab(s
, page
) ||
4277 !on_freelist(s
, page
, NULL
))
4280 /* Now we know that a valid freelist exists */
4281 bitmap_zero(map
, page
->objects
);
4283 get_map(s
, page
, map
);
4284 for_each_object(p
, s
, addr
, page
->objects
) {
4285 if (test_bit(slab_index(p
, s
, addr
), map
))
4286 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4290 for_each_object(p
, s
, addr
, page
->objects
)
4291 if (!test_bit(slab_index(p
, s
, addr
), map
))
4292 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4297 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4301 validate_slab(s
, page
, map
);
4305 static int validate_slab_node(struct kmem_cache
*s
,
4306 struct kmem_cache_node
*n
, unsigned long *map
)
4308 unsigned long count
= 0;
4310 unsigned long flags
;
4312 spin_lock_irqsave(&n
->list_lock
, flags
);
4314 list_for_each_entry(page
, &n
->partial
, lru
) {
4315 validate_slab_slab(s
, page
, map
);
4318 if (count
!= n
->nr_partial
)
4319 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4320 s
->name
, count
, n
->nr_partial
);
4322 if (!(s
->flags
& SLAB_STORE_USER
))
4325 list_for_each_entry(page
, &n
->full
, lru
) {
4326 validate_slab_slab(s
, page
, map
);
4329 if (count
!= atomic_long_read(&n
->nr_slabs
))
4330 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4331 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4334 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4338 static long validate_slab_cache(struct kmem_cache
*s
)
4341 unsigned long count
= 0;
4342 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4343 sizeof(unsigned long), GFP_KERNEL
);
4344 struct kmem_cache_node
*n
;
4350 for_each_kmem_cache_node(s
, node
, n
)
4351 count
+= validate_slab_node(s
, n
, map
);
4356 * Generate lists of code addresses where slabcache objects are allocated
4361 unsigned long count
;
4368 DECLARE_BITMAP(cpus
, NR_CPUS
);
4374 unsigned long count
;
4375 struct location
*loc
;
4378 static void free_loc_track(struct loc_track
*t
)
4381 free_pages((unsigned long)t
->loc
,
4382 get_order(sizeof(struct location
) * t
->max
));
4385 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4390 order
= get_order(sizeof(struct location
) * max
);
4392 l
= (void *)__get_free_pages(flags
, order
);
4397 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4405 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4406 const struct track
*track
)
4408 long start
, end
, pos
;
4410 unsigned long caddr
;
4411 unsigned long age
= jiffies
- track
->when
;
4417 pos
= start
+ (end
- start
+ 1) / 2;
4420 * There is nothing at "end". If we end up there
4421 * we need to add something to before end.
4426 caddr
= t
->loc
[pos
].addr
;
4427 if (track
->addr
== caddr
) {
4433 if (age
< l
->min_time
)
4435 if (age
> l
->max_time
)
4438 if (track
->pid
< l
->min_pid
)
4439 l
->min_pid
= track
->pid
;
4440 if (track
->pid
> l
->max_pid
)
4441 l
->max_pid
= track
->pid
;
4443 cpumask_set_cpu(track
->cpu
,
4444 to_cpumask(l
->cpus
));
4446 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4450 if (track
->addr
< caddr
)
4457 * Not found. Insert new tracking element.
4459 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4465 (t
->count
- pos
) * sizeof(struct location
));
4468 l
->addr
= track
->addr
;
4472 l
->min_pid
= track
->pid
;
4473 l
->max_pid
= track
->pid
;
4474 cpumask_clear(to_cpumask(l
->cpus
));
4475 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4476 nodes_clear(l
->nodes
);
4477 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4481 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4482 struct page
*page
, enum track_item alloc
,
4485 void *addr
= page_address(page
);
4488 bitmap_zero(map
, page
->objects
);
4489 get_map(s
, page
, map
);
4491 for_each_object(p
, s
, addr
, page
->objects
)
4492 if (!test_bit(slab_index(p
, s
, addr
), map
))
4493 add_location(t
, s
, get_track(s
, p
, alloc
));
4496 static int list_locations(struct kmem_cache
*s
, char *buf
,
4497 enum track_item alloc
)
4501 struct loc_track t
= { 0, 0, NULL
};
4503 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4504 sizeof(unsigned long), GFP_KERNEL
);
4505 struct kmem_cache_node
*n
;
4507 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4510 return sprintf(buf
, "Out of memory\n");
4512 /* Push back cpu slabs */
4515 for_each_kmem_cache_node(s
, node
, n
) {
4516 unsigned long flags
;
4519 if (!atomic_long_read(&n
->nr_slabs
))
4522 spin_lock_irqsave(&n
->list_lock
, flags
);
4523 list_for_each_entry(page
, &n
->partial
, lru
)
4524 process_slab(&t
, s
, page
, alloc
, map
);
4525 list_for_each_entry(page
, &n
->full
, lru
)
4526 process_slab(&t
, s
, page
, alloc
, map
);
4527 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4530 for (i
= 0; i
< t
.count
; i
++) {
4531 struct location
*l
= &t
.loc
[i
];
4533 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4535 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4538 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4540 len
+= sprintf(buf
+ len
, "<not-available>");
4542 if (l
->sum_time
!= l
->min_time
) {
4543 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4545 (long)div_u64(l
->sum_time
, l
->count
),
4548 len
+= sprintf(buf
+ len
, " age=%ld",
4551 if (l
->min_pid
!= l
->max_pid
)
4552 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4553 l
->min_pid
, l
->max_pid
);
4555 len
+= sprintf(buf
+ len
, " pid=%ld",
4558 if (num_online_cpus() > 1 &&
4559 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4560 len
< PAGE_SIZE
- 60)
4561 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4563 cpumask_pr_args(to_cpumask(l
->cpus
)));
4565 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4566 len
< PAGE_SIZE
- 60)
4567 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4569 nodemask_pr_args(&l
->nodes
));
4571 len
+= sprintf(buf
+ len
, "\n");
4577 len
+= sprintf(buf
, "No data\n");
4582 #ifdef SLUB_RESILIENCY_TEST
4583 static void __init
resiliency_test(void)
4587 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4589 pr_err("SLUB resiliency testing\n");
4590 pr_err("-----------------------\n");
4591 pr_err("A. Corruption after allocation\n");
4593 p
= kzalloc(16, GFP_KERNEL
);
4595 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4598 validate_slab_cache(kmalloc_caches
[4]);
4600 /* Hmmm... The next two are dangerous */
4601 p
= kzalloc(32, GFP_KERNEL
);
4602 p
[32 + sizeof(void *)] = 0x34;
4603 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4605 pr_err("If allocated object is overwritten then not detectable\n\n");
4607 validate_slab_cache(kmalloc_caches
[5]);
4608 p
= kzalloc(64, GFP_KERNEL
);
4609 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4611 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4613 pr_err("If allocated object is overwritten then not detectable\n\n");
4614 validate_slab_cache(kmalloc_caches
[6]);
4616 pr_err("\nB. Corruption after free\n");
4617 p
= kzalloc(128, GFP_KERNEL
);
4620 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4621 validate_slab_cache(kmalloc_caches
[7]);
4623 p
= kzalloc(256, GFP_KERNEL
);
4626 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4627 validate_slab_cache(kmalloc_caches
[8]);
4629 p
= kzalloc(512, GFP_KERNEL
);
4632 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4633 validate_slab_cache(kmalloc_caches
[9]);
4637 static void resiliency_test(void) {};
4642 enum slab_stat_type
{
4643 SL_ALL
, /* All slabs */
4644 SL_PARTIAL
, /* Only partially allocated slabs */
4645 SL_CPU
, /* Only slabs used for cpu caches */
4646 SL_OBJECTS
, /* Determine allocated objects not slabs */
4647 SL_TOTAL
/* Determine object capacity not slabs */
4650 #define SO_ALL (1 << SL_ALL)
4651 #define SO_PARTIAL (1 << SL_PARTIAL)
4652 #define SO_CPU (1 << SL_CPU)
4653 #define SO_OBJECTS (1 << SL_OBJECTS)
4654 #define SO_TOTAL (1 << SL_TOTAL)
4656 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4657 char *buf
, unsigned long flags
)
4659 unsigned long total
= 0;
4662 unsigned long *nodes
;
4664 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4668 if (flags
& SO_CPU
) {
4671 for_each_possible_cpu(cpu
) {
4672 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4677 page
= READ_ONCE(c
->page
);
4681 node
= page_to_nid(page
);
4682 if (flags
& SO_TOTAL
)
4684 else if (flags
& SO_OBJECTS
)
4692 page
= READ_ONCE(c
->partial
);
4694 node
= page_to_nid(page
);
4695 if (flags
& SO_TOTAL
)
4697 else if (flags
& SO_OBJECTS
)
4708 #ifdef CONFIG_SLUB_DEBUG
4709 if (flags
& SO_ALL
) {
4710 struct kmem_cache_node
*n
;
4712 for_each_kmem_cache_node(s
, node
, n
) {
4714 if (flags
& SO_TOTAL
)
4715 x
= atomic_long_read(&n
->total_objects
);
4716 else if (flags
& SO_OBJECTS
)
4717 x
= atomic_long_read(&n
->total_objects
) -
4718 count_partial(n
, count_free
);
4720 x
= atomic_long_read(&n
->nr_slabs
);
4727 if (flags
& SO_PARTIAL
) {
4728 struct kmem_cache_node
*n
;
4730 for_each_kmem_cache_node(s
, node
, n
) {
4731 if (flags
& SO_TOTAL
)
4732 x
= count_partial(n
, count_total
);
4733 else if (flags
& SO_OBJECTS
)
4734 x
= count_partial(n
, count_inuse
);
4741 x
= sprintf(buf
, "%lu", total
);
4743 for (node
= 0; node
< nr_node_ids
; node
++)
4745 x
+= sprintf(buf
+ x
, " N%d=%lu",
4750 return x
+ sprintf(buf
+ x
, "\n");
4753 #ifdef CONFIG_SLUB_DEBUG
4754 static int any_slab_objects(struct kmem_cache
*s
)
4757 struct kmem_cache_node
*n
;
4759 for_each_kmem_cache_node(s
, node
, n
)
4760 if (atomic_long_read(&n
->total_objects
))
4767 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4768 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4770 struct slab_attribute
{
4771 struct attribute attr
;
4772 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4773 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4776 #define SLAB_ATTR_RO(_name) \
4777 static struct slab_attribute _name##_attr = \
4778 __ATTR(_name, 0400, _name##_show, NULL)
4780 #define SLAB_ATTR(_name) \
4781 static struct slab_attribute _name##_attr = \
4782 __ATTR(_name, 0600, _name##_show, _name##_store)
4784 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4786 return sprintf(buf
, "%d\n", s
->size
);
4788 SLAB_ATTR_RO(slab_size
);
4790 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4792 return sprintf(buf
, "%d\n", s
->align
);
4794 SLAB_ATTR_RO(align
);
4796 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4798 return sprintf(buf
, "%d\n", s
->object_size
);
4800 SLAB_ATTR_RO(object_size
);
4802 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4804 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4806 SLAB_ATTR_RO(objs_per_slab
);
4808 static ssize_t
order_store(struct kmem_cache
*s
,
4809 const char *buf
, size_t length
)
4811 unsigned long order
;
4814 err
= kstrtoul(buf
, 10, &order
);
4818 if (order
> slub_max_order
|| order
< slub_min_order
)
4821 calculate_sizes(s
, order
);
4825 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4827 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4831 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4833 return sprintf(buf
, "%lu\n", s
->min_partial
);
4836 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4842 err
= kstrtoul(buf
, 10, &min
);
4846 set_min_partial(s
, min
);
4849 SLAB_ATTR(min_partial
);
4851 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4853 return sprintf(buf
, "%u\n", s
->cpu_partial
);
4856 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4859 unsigned long objects
;
4862 err
= kstrtoul(buf
, 10, &objects
);
4865 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4868 s
->cpu_partial
= objects
;
4872 SLAB_ATTR(cpu_partial
);
4874 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4878 return sprintf(buf
, "%pS\n", s
->ctor
);
4882 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4884 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4886 SLAB_ATTR_RO(aliases
);
4888 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4890 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4892 SLAB_ATTR_RO(partial
);
4894 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4896 return show_slab_objects(s
, buf
, SO_CPU
);
4898 SLAB_ATTR_RO(cpu_slabs
);
4900 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4902 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4904 SLAB_ATTR_RO(objects
);
4906 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4908 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
4910 SLAB_ATTR_RO(objects_partial
);
4912 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4919 for_each_online_cpu(cpu
) {
4920 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
)->partial
;
4923 pages
+= page
->pages
;
4924 objects
+= page
->pobjects
;
4928 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
4931 for_each_online_cpu(cpu
) {
4932 struct page
*page
= per_cpu_ptr(s
->cpu_slab
, cpu
) ->partial
;
4934 if (page
&& len
< PAGE_SIZE
- 20)
4935 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
4936 page
->pobjects
, page
->pages
);
4939 return len
+ sprintf(buf
+ len
, "\n");
4941 SLAB_ATTR_RO(slabs_cpu_partial
);
4943 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
4945 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
4948 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
4949 const char *buf
, size_t length
)
4951 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
4953 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
4956 SLAB_ATTR(reclaim_account
);
4958 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
4960 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
4962 SLAB_ATTR_RO(hwcache_align
);
4964 #ifdef CONFIG_ZONE_DMA
4965 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
4967 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
4969 SLAB_ATTR_RO(cache_dma
);
4972 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
4974 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_DESTROY_BY_RCU
));
4976 SLAB_ATTR_RO(destroy_by_rcu
);
4978 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
4980 return sprintf(buf
, "%d\n", s
->reserved
);
4982 SLAB_ATTR_RO(reserved
);
4984 #ifdef CONFIG_SLUB_DEBUG
4985 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
4987 return show_slab_objects(s
, buf
, SO_ALL
);
4989 SLAB_ATTR_RO(slabs
);
4991 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
4993 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
4995 SLAB_ATTR_RO(total_objects
);
4997 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
4999 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5002 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5003 const char *buf
, size_t length
)
5005 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5006 if (buf
[0] == '1') {
5007 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5008 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5012 SLAB_ATTR(sanity_checks
);
5014 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5016 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5019 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5023 * Tracing a merged cache is going to give confusing results
5024 * as well as cause other issues like converting a mergeable
5025 * cache into an umergeable one.
5027 if (s
->refcount
> 1)
5030 s
->flags
&= ~SLAB_TRACE
;
5031 if (buf
[0] == '1') {
5032 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5033 s
->flags
|= SLAB_TRACE
;
5039 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5041 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5044 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5045 const char *buf
, size_t length
)
5047 if (any_slab_objects(s
))
5050 s
->flags
&= ~SLAB_RED_ZONE
;
5051 if (buf
[0] == '1') {
5052 s
->flags
|= SLAB_RED_ZONE
;
5054 calculate_sizes(s
, -1);
5057 SLAB_ATTR(red_zone
);
5059 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5061 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5064 static ssize_t
poison_store(struct kmem_cache
*s
,
5065 const char *buf
, size_t length
)
5067 if (any_slab_objects(s
))
5070 s
->flags
&= ~SLAB_POISON
;
5071 if (buf
[0] == '1') {
5072 s
->flags
|= SLAB_POISON
;
5074 calculate_sizes(s
, -1);
5079 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5081 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5084 static ssize_t
store_user_store(struct kmem_cache
*s
,
5085 const char *buf
, size_t length
)
5087 if (any_slab_objects(s
))
5090 s
->flags
&= ~SLAB_STORE_USER
;
5091 if (buf
[0] == '1') {
5092 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5093 s
->flags
|= SLAB_STORE_USER
;
5095 calculate_sizes(s
, -1);
5098 SLAB_ATTR(store_user
);
5100 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5105 static ssize_t
validate_store(struct kmem_cache
*s
,
5106 const char *buf
, size_t length
)
5110 if (buf
[0] == '1') {
5111 ret
= validate_slab_cache(s
);
5117 SLAB_ATTR(validate
);
5119 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5121 if (!(s
->flags
& SLAB_STORE_USER
))
5123 return list_locations(s
, buf
, TRACK_ALLOC
);
5125 SLAB_ATTR_RO(alloc_calls
);
5127 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5129 if (!(s
->flags
& SLAB_STORE_USER
))
5131 return list_locations(s
, buf
, TRACK_FREE
);
5133 SLAB_ATTR_RO(free_calls
);
5134 #endif /* CONFIG_SLUB_DEBUG */
5136 #ifdef CONFIG_FAILSLAB
5137 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5139 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5142 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5145 if (s
->refcount
> 1)
5148 s
->flags
&= ~SLAB_FAILSLAB
;
5150 s
->flags
|= SLAB_FAILSLAB
;
5153 SLAB_ATTR(failslab
);
5156 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5161 static ssize_t
shrink_store(struct kmem_cache
*s
,
5162 const char *buf
, size_t length
)
5165 kmem_cache_shrink(s
);
5173 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5175 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
5178 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5179 const char *buf
, size_t length
)
5181 unsigned long ratio
;
5184 err
= kstrtoul(buf
, 10, &ratio
);
5189 s
->remote_node_defrag_ratio
= ratio
* 10;
5193 SLAB_ATTR(remote_node_defrag_ratio
);
5196 #ifdef CONFIG_SLUB_STATS
5197 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5199 unsigned long sum
= 0;
5202 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5207 for_each_online_cpu(cpu
) {
5208 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5214 len
= sprintf(buf
, "%lu", sum
);
5217 for_each_online_cpu(cpu
) {
5218 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5219 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5223 return len
+ sprintf(buf
+ len
, "\n");
5226 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5230 for_each_online_cpu(cpu
)
5231 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5234 #define STAT_ATTR(si, text) \
5235 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5237 return show_stat(s, buf, si); \
5239 static ssize_t text##_store(struct kmem_cache *s, \
5240 const char *buf, size_t length) \
5242 if (buf[0] != '0') \
5244 clear_stat(s, si); \
5249 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5250 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5251 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5252 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5253 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5254 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5255 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5256 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5257 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5258 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5259 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5260 STAT_ATTR(FREE_SLAB
, free_slab
);
5261 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5262 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5263 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5264 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5265 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5266 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5267 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5268 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5269 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5270 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5271 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5272 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5273 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5274 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5277 static struct attribute
*slab_attrs
[] = {
5278 &slab_size_attr
.attr
,
5279 &object_size_attr
.attr
,
5280 &objs_per_slab_attr
.attr
,
5282 &min_partial_attr
.attr
,
5283 &cpu_partial_attr
.attr
,
5285 &objects_partial_attr
.attr
,
5287 &cpu_slabs_attr
.attr
,
5291 &hwcache_align_attr
.attr
,
5292 &reclaim_account_attr
.attr
,
5293 &destroy_by_rcu_attr
.attr
,
5295 &reserved_attr
.attr
,
5296 &slabs_cpu_partial_attr
.attr
,
5297 #ifdef CONFIG_SLUB_DEBUG
5298 &total_objects_attr
.attr
,
5300 &sanity_checks_attr
.attr
,
5302 &red_zone_attr
.attr
,
5304 &store_user_attr
.attr
,
5305 &validate_attr
.attr
,
5306 &alloc_calls_attr
.attr
,
5307 &free_calls_attr
.attr
,
5309 #ifdef CONFIG_ZONE_DMA
5310 &cache_dma_attr
.attr
,
5313 &remote_node_defrag_ratio_attr
.attr
,
5315 #ifdef CONFIG_SLUB_STATS
5316 &alloc_fastpath_attr
.attr
,
5317 &alloc_slowpath_attr
.attr
,
5318 &free_fastpath_attr
.attr
,
5319 &free_slowpath_attr
.attr
,
5320 &free_frozen_attr
.attr
,
5321 &free_add_partial_attr
.attr
,
5322 &free_remove_partial_attr
.attr
,
5323 &alloc_from_partial_attr
.attr
,
5324 &alloc_slab_attr
.attr
,
5325 &alloc_refill_attr
.attr
,
5326 &alloc_node_mismatch_attr
.attr
,
5327 &free_slab_attr
.attr
,
5328 &cpuslab_flush_attr
.attr
,
5329 &deactivate_full_attr
.attr
,
5330 &deactivate_empty_attr
.attr
,
5331 &deactivate_to_head_attr
.attr
,
5332 &deactivate_to_tail_attr
.attr
,
5333 &deactivate_remote_frees_attr
.attr
,
5334 &deactivate_bypass_attr
.attr
,
5335 &order_fallback_attr
.attr
,
5336 &cmpxchg_double_fail_attr
.attr
,
5337 &cmpxchg_double_cpu_fail_attr
.attr
,
5338 &cpu_partial_alloc_attr
.attr
,
5339 &cpu_partial_free_attr
.attr
,
5340 &cpu_partial_node_attr
.attr
,
5341 &cpu_partial_drain_attr
.attr
,
5343 #ifdef CONFIG_FAILSLAB
5344 &failslab_attr
.attr
,
5350 static struct attribute_group slab_attr_group
= {
5351 .attrs
= slab_attrs
,
5354 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5355 struct attribute
*attr
,
5358 struct slab_attribute
*attribute
;
5359 struct kmem_cache
*s
;
5362 attribute
= to_slab_attr(attr
);
5365 if (!attribute
->show
)
5368 err
= attribute
->show(s
, buf
);
5373 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5374 struct attribute
*attr
,
5375 const char *buf
, size_t len
)
5377 struct slab_attribute
*attribute
;
5378 struct kmem_cache
*s
;
5381 attribute
= to_slab_attr(attr
);
5384 if (!attribute
->store
)
5387 err
= attribute
->store(s
, buf
, len
);
5389 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5390 struct kmem_cache
*c
;
5392 mutex_lock(&slab_mutex
);
5393 if (s
->max_attr_size
< len
)
5394 s
->max_attr_size
= len
;
5397 * This is a best effort propagation, so this function's return
5398 * value will be determined by the parent cache only. This is
5399 * basically because not all attributes will have a well
5400 * defined semantics for rollbacks - most of the actions will
5401 * have permanent effects.
5403 * Returning the error value of any of the children that fail
5404 * is not 100 % defined, in the sense that users seeing the
5405 * error code won't be able to know anything about the state of
5408 * Only returning the error code for the parent cache at least
5409 * has well defined semantics. The cache being written to
5410 * directly either failed or succeeded, in which case we loop
5411 * through the descendants with best-effort propagation.
5413 for_each_memcg_cache(c
, s
)
5414 attribute
->store(c
, buf
, len
);
5415 mutex_unlock(&slab_mutex
);
5421 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5425 char *buffer
= NULL
;
5426 struct kmem_cache
*root_cache
;
5428 if (is_root_cache(s
))
5431 root_cache
= s
->memcg_params
.root_cache
;
5434 * This mean this cache had no attribute written. Therefore, no point
5435 * in copying default values around
5437 if (!root_cache
->max_attr_size
)
5440 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5443 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5445 if (!attr
|| !attr
->store
|| !attr
->show
)
5449 * It is really bad that we have to allocate here, so we will
5450 * do it only as a fallback. If we actually allocate, though,
5451 * we can just use the allocated buffer until the end.
5453 * Most of the slub attributes will tend to be very small in
5454 * size, but sysfs allows buffers up to a page, so they can
5455 * theoretically happen.
5459 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5462 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5463 if (WARN_ON(!buffer
))
5468 attr
->show(root_cache
, buf
);
5469 attr
->store(s
, buf
, strlen(buf
));
5473 free_page((unsigned long)buffer
);
5477 static void kmem_cache_release(struct kobject
*k
)
5479 slab_kmem_cache_release(to_slab(k
));
5482 static const struct sysfs_ops slab_sysfs_ops
= {
5483 .show
= slab_attr_show
,
5484 .store
= slab_attr_store
,
5487 static struct kobj_type slab_ktype
= {
5488 .sysfs_ops
= &slab_sysfs_ops
,
5489 .release
= kmem_cache_release
,
5492 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5494 struct kobj_type
*ktype
= get_ktype(kobj
);
5496 if (ktype
== &slab_ktype
)
5501 static const struct kset_uevent_ops slab_uevent_ops
= {
5502 .filter
= uevent_filter
,
5505 static struct kset
*slab_kset
;
5507 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5510 if (!is_root_cache(s
))
5511 return s
->memcg_params
.root_cache
->memcg_kset
;
5516 #define ID_STR_LENGTH 64
5518 /* Create a unique string id for a slab cache:
5520 * Format :[flags-]size
5522 static char *create_unique_id(struct kmem_cache
*s
)
5524 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5531 * First flags affecting slabcache operations. We will only
5532 * get here for aliasable slabs so we do not need to support
5533 * too many flags. The flags here must cover all flags that
5534 * are matched during merging to guarantee that the id is
5537 if (s
->flags
& SLAB_CACHE_DMA
)
5539 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5541 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5543 if (!(s
->flags
& SLAB_NOTRACK
))
5545 if (s
->flags
& SLAB_ACCOUNT
)
5549 p
+= sprintf(p
, "%07d", s
->size
);
5551 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5555 static int sysfs_slab_add(struct kmem_cache
*s
)
5559 int unmergeable
= slab_unmergeable(s
);
5563 * Slabcache can never be merged so we can use the name proper.
5564 * This is typically the case for debug situations. In that
5565 * case we can catch duplicate names easily.
5567 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5571 * Create a unique name for the slab as a target
5574 name
= create_unique_id(s
);
5577 s
->kobj
.kset
= cache_kset(s
);
5578 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5582 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5587 if (is_root_cache(s
)) {
5588 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5589 if (!s
->memcg_kset
) {
5596 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5598 /* Setup first alias */
5599 sysfs_slab_alias(s
, s
->name
);
5606 kobject_del(&s
->kobj
);
5610 void sysfs_slab_remove(struct kmem_cache
*s
)
5612 if (slab_state
< FULL
)
5614 * Sysfs has not been setup yet so no need to remove the
5620 kset_unregister(s
->memcg_kset
);
5622 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5623 kobject_del(&s
->kobj
);
5624 kobject_put(&s
->kobj
);
5628 * Need to buffer aliases during bootup until sysfs becomes
5629 * available lest we lose that information.
5631 struct saved_alias
{
5632 struct kmem_cache
*s
;
5634 struct saved_alias
*next
;
5637 static struct saved_alias
*alias_list
;
5639 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5641 struct saved_alias
*al
;
5643 if (slab_state
== FULL
) {
5645 * If we have a leftover link then remove it.
5647 sysfs_remove_link(&slab_kset
->kobj
, name
);
5648 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5651 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5657 al
->next
= alias_list
;
5662 static int __init
slab_sysfs_init(void)
5664 struct kmem_cache
*s
;
5667 mutex_lock(&slab_mutex
);
5669 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5671 mutex_unlock(&slab_mutex
);
5672 pr_err("Cannot register slab subsystem.\n");
5678 list_for_each_entry(s
, &slab_caches
, list
) {
5679 err
= sysfs_slab_add(s
);
5681 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5685 while (alias_list
) {
5686 struct saved_alias
*al
= alias_list
;
5688 alias_list
= alias_list
->next
;
5689 err
= sysfs_slab_alias(al
->s
, al
->name
);
5691 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5696 mutex_unlock(&slab_mutex
);
5701 __initcall(slab_sysfs_init
);
5702 #endif /* CONFIG_SYSFS */
5705 * The /proc/slabinfo ABI
5707 #ifdef CONFIG_SLABINFO
5708 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5710 unsigned long nr_slabs
= 0;
5711 unsigned long nr_objs
= 0;
5712 unsigned long nr_free
= 0;
5714 struct kmem_cache_node
*n
;
5716 for_each_kmem_cache_node(s
, node
, n
) {
5717 nr_slabs
+= node_nr_slabs(n
);
5718 nr_objs
+= node_nr_objs(n
);
5719 nr_free
+= count_partial(n
, count_free
);
5722 sinfo
->active_objs
= nr_objs
- nr_free
;
5723 sinfo
->num_objs
= nr_objs
;
5724 sinfo
->active_slabs
= nr_slabs
;
5725 sinfo
->num_slabs
= nr_slabs
;
5726 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5727 sinfo
->cache_order
= oo_order(s
->oo
);
5730 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5734 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5735 size_t count
, loff_t
*ppos
)
5739 #endif /* CONFIG_SLABINFO */