1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.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:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache
*s
)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
126 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
128 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
129 p
+= s
->red_left_pad
;
134 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
136 #ifdef CONFIG_SLUB_CPU_PARTIAL
137 return !kmem_cache_debug(s
);
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
154 /* Enable to log cmpxchg failures */
155 #undef SLUB_DEBUG_CMPXCHG
158 * Mininum number of partial slabs. These will be left on the partial
159 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 #define MIN_PARTIAL 5
164 * Maximum number of desirable partial slabs.
165 * The existence of more partial slabs makes kmem_cache_shrink
166 * sort the partial list by the number of objects in use.
168 #define MAX_PARTIAL 10
170 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_STORE_USER)
174 * These debug flags cannot use CMPXCHG because there might be consistency
175 * issues when checking or reading debug information
177 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 * Debugging flags that require metadata to be stored in the slab. These get
183 * disabled when slub_debug=O is used and a cache's min order increases with
186 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189 #define OO_MASK ((1 << OO_SHIFT) - 1)
190 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192 /* Internal SLUB flags */
194 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
195 /* Use cmpxchg_double */
196 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199 * Tracking user of a slab.
201 #define TRACK_ADDRS_COUNT 16
203 unsigned long addr
; /* Called from address */
204 #ifdef CONFIG_STACKTRACE
205 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
207 int cpu
; /* Was running on cpu */
208 int pid
; /* Pid context */
209 unsigned long when
; /* When did the operation occur */
212 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
215 static int sysfs_slab_add(struct kmem_cache
*);
216 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
217 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
218 static void sysfs_slab_remove(struct kmem_cache
*s
);
220 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
221 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
223 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
224 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
227 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
248 unsigned long ptr_addr
)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr
^ s
->random
^
262 (unsigned long)kasan_reset_tag((void *)ptr_addr
));
268 /* Returns the freelist pointer recorded at location ptr_addr. */
269 static inline void *freelist_dereference(const struct kmem_cache
*s
,
272 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
273 (unsigned long)ptr_addr
);
276 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
278 return freelist_dereference(s
, object
+ s
->offset
);
281 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
283 prefetch(object
+ s
->offset
);
286 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
288 unsigned long freepointer_addr
;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s
, object
);
294 freepointer_addr
= (unsigned long)object
+ s
->offset
;
295 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
296 return freelist_ptr(s
, p
, freepointer_addr
);
299 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
301 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 /* Determine object index from a given position */
317 static inline unsigned int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
319 return (kasan_reset_tag(p
) - addr
) / s
->size
;
322 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
324 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
327 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
330 struct kmem_cache_order_objects x
= {
331 (order
<< OO_SHIFT
) + order_objects(order
, size
)
337 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
339 return x
.x
>> OO_SHIFT
;
342 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
344 return x
.x
& OO_MASK
;
348 * Per slab locking using the pagelock
350 static __always_inline
void slab_lock(struct page
*page
)
352 VM_BUG_ON_PAGE(PageTail(page
), page
);
353 bit_spin_lock(PG_locked
, &page
->flags
);
356 static __always_inline
void slab_unlock(struct page
*page
)
358 VM_BUG_ON_PAGE(PageTail(page
), page
);
359 __bit_spin_unlock(PG_locked
, &page
->flags
);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
364 void *freelist_old
, unsigned long counters_old
,
365 void *freelist_new
, unsigned long counters_new
,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s
->flags
& __CMPXCHG_DOUBLE
) {
372 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
373 freelist_old
, counters_old
,
374 freelist_new
, counters_new
))
380 if (page
->freelist
== freelist_old
&&
381 page
->counters
== counters_old
) {
382 page
->freelist
= freelist_new
;
383 page
->counters
= counters_new
;
391 stat(s
, CMPXCHG_DOUBLE_FAIL
);
393 #ifdef SLUB_DEBUG_CMPXCHG
394 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
400 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
401 void *freelist_old
, unsigned long counters_old
,
402 void *freelist_new
, unsigned long counters_new
,
405 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
406 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
407 if (s
->flags
& __CMPXCHG_DOUBLE
) {
408 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
409 freelist_old
, counters_old
,
410 freelist_new
, counters_new
))
417 local_irq_save(flags
);
419 if (page
->freelist
== freelist_old
&&
420 page
->counters
== counters_old
) {
421 page
->freelist
= freelist_new
;
422 page
->counters
= counters_new
;
424 local_irq_restore(flags
);
428 local_irq_restore(flags
);
432 stat(s
, CMPXCHG_DOUBLE_FAIL
);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
441 #ifdef CONFIG_SLUB_DEBUG
443 * Determine a map of object in use on a page.
445 * Node listlock must be held to guarantee that the page does
446 * not vanish from under us.
448 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
451 void *addr
= page_address(page
);
453 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
454 set_bit(slab_index(p
, s
, addr
), map
);
457 static inline unsigned int size_from_object(struct kmem_cache
*s
)
459 if (s
->flags
& SLAB_RED_ZONE
)
460 return s
->size
- s
->red_left_pad
;
465 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
467 if (s
->flags
& SLAB_RED_ZONE
)
468 p
-= s
->red_left_pad
;
476 #if defined(CONFIG_SLUB_DEBUG_ON)
477 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
479 static slab_flags_t slub_debug
;
482 static char *slub_debug_slabs
;
483 static int disable_higher_order_debug
;
486 * slub is about to manipulate internal object metadata. This memory lies
487 * outside the range of the allocated object, so accessing it would normally
488 * be reported by kasan as a bounds error. metadata_access_enable() is used
489 * to tell kasan that these accesses are OK.
491 static inline void metadata_access_enable(void)
493 kasan_disable_current();
496 static inline void metadata_access_disable(void)
498 kasan_enable_current();
505 /* Verify that a pointer has an address that is valid within a slab page */
506 static inline int check_valid_pointer(struct kmem_cache
*s
,
507 struct page
*page
, void *object
)
514 base
= page_address(page
);
515 object
= kasan_reset_tag(object
);
516 object
= restore_red_left(s
, object
);
517 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
518 (object
- base
) % s
->size
) {
525 static void print_section(char *level
, char *text
, u8
*addr
,
528 metadata_access_enable();
529 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
531 metadata_access_disable();
534 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
535 enum track_item alloc
)
540 p
= object
+ s
->offset
+ sizeof(void *);
542 p
= object
+ s
->inuse
;
547 static void set_track(struct kmem_cache
*s
, void *object
,
548 enum track_item alloc
, unsigned long addr
)
550 struct track
*p
= get_track(s
, object
, alloc
);
553 #ifdef CONFIG_STACKTRACE
554 unsigned int nr_entries
;
556 metadata_access_enable();
557 nr_entries
= stack_trace_save(p
->addrs
, TRACK_ADDRS_COUNT
, 3);
558 metadata_access_disable();
560 if (nr_entries
< TRACK_ADDRS_COUNT
)
561 p
->addrs
[nr_entries
] = 0;
564 p
->cpu
= smp_processor_id();
565 p
->pid
= current
->pid
;
568 memset(p
, 0, sizeof(struct track
));
572 static void init_tracking(struct kmem_cache
*s
, void *object
)
574 if (!(s
->flags
& SLAB_STORE_USER
))
577 set_track(s
, object
, TRACK_FREE
, 0UL);
578 set_track(s
, object
, TRACK_ALLOC
, 0UL);
581 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
586 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
587 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
588 #ifdef CONFIG_STACKTRACE
591 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
593 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
600 static void print_tracking(struct kmem_cache
*s
, void *object
)
602 unsigned long pr_time
= jiffies
;
603 if (!(s
->flags
& SLAB_STORE_USER
))
606 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
607 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
610 static void print_page_info(struct page
*page
)
612 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
613 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
617 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
619 struct va_format vaf
;
625 pr_err("=============================================================================\n");
626 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
627 pr_err("-----------------------------------------------------------------------------\n\n");
629 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
633 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
635 struct va_format vaf
;
641 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
645 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
647 unsigned int off
; /* Offset of last byte */
648 u8
*addr
= page_address(page
);
650 print_tracking(s
, p
);
652 print_page_info(page
);
654 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
655 p
, p
- addr
, get_freepointer(s
, p
));
657 if (s
->flags
& SLAB_RED_ZONE
)
658 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
660 else if (p
> addr
+ 16)
661 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
663 print_section(KERN_ERR
, "Object ", p
,
664 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
665 if (s
->flags
& SLAB_RED_ZONE
)
666 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
667 s
->inuse
- s
->object_size
);
670 off
= s
->offset
+ sizeof(void *);
674 if (s
->flags
& SLAB_STORE_USER
)
675 off
+= 2 * sizeof(struct track
);
677 off
+= kasan_metadata_size(s
);
679 if (off
!= size_from_object(s
))
680 /* Beginning of the filler is the free pointer */
681 print_section(KERN_ERR
, "Padding ", p
+ off
,
682 size_from_object(s
) - off
);
687 void object_err(struct kmem_cache
*s
, struct page
*page
,
688 u8
*object
, char *reason
)
690 slab_bug(s
, "%s", reason
);
691 print_trailer(s
, page
, object
);
694 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
695 const char *fmt
, ...)
701 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
703 slab_bug(s
, "%s", buf
);
704 print_page_info(page
);
708 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
712 if (s
->flags
& SLAB_RED_ZONE
)
713 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
715 if (s
->flags
& __OBJECT_POISON
) {
716 memset(p
, POISON_FREE
, s
->object_size
- 1);
717 p
[s
->object_size
- 1] = POISON_END
;
720 if (s
->flags
& SLAB_RED_ZONE
)
721 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
724 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
725 void *from
, void *to
)
727 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
728 memset(from
, data
, to
- from
);
731 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
732 u8
*object
, char *what
,
733 u8
*start
, unsigned int value
, unsigned int bytes
)
737 u8
*addr
= page_address(page
);
739 metadata_access_enable();
740 fault
= memchr_inv(start
, value
, bytes
);
741 metadata_access_disable();
746 while (end
> fault
&& end
[-1] == value
)
749 slab_bug(s
, "%s overwritten", what
);
750 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
751 fault
, end
- 1, fault
- addr
,
753 print_trailer(s
, page
, object
);
755 restore_bytes(s
, what
, value
, fault
, end
);
763 * Bytes of the object to be managed.
764 * If the freepointer may overlay the object then the free
765 * pointer is the first word of the object.
767 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
770 * object + s->object_size
771 * Padding to reach word boundary. This is also used for Redzoning.
772 * Padding is extended by another word if Redzoning is enabled and
773 * object_size == inuse.
775 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
776 * 0xcc (RED_ACTIVE) for objects in use.
779 * Meta data starts here.
781 * A. Free pointer (if we cannot overwrite object on free)
782 * B. Tracking data for SLAB_STORE_USER
783 * C. Padding to reach required alignment boundary or at mininum
784 * one word if debugging is on to be able to detect writes
785 * before the word boundary.
787 * Padding is done using 0x5a (POISON_INUSE)
790 * Nothing is used beyond s->size.
792 * If slabcaches are merged then the object_size and inuse boundaries are mostly
793 * ignored. And therefore no slab options that rely on these boundaries
794 * may be used with merged slabcaches.
797 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
799 unsigned long off
= s
->inuse
; /* The end of info */
802 /* Freepointer is placed after the object. */
803 off
+= sizeof(void *);
805 if (s
->flags
& SLAB_STORE_USER
)
806 /* We also have user information there */
807 off
+= 2 * sizeof(struct track
);
809 off
+= kasan_metadata_size(s
);
811 if (size_from_object(s
) == off
)
814 return check_bytes_and_report(s
, page
, p
, "Object padding",
815 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
818 /* Check the pad bytes at the end of a slab page */
819 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
828 if (!(s
->flags
& SLAB_POISON
))
831 start
= page_address(page
);
832 length
= page_size(page
);
833 end
= start
+ length
;
834 remainder
= length
% s
->size
;
838 pad
= end
- remainder
;
839 metadata_access_enable();
840 fault
= memchr_inv(pad
, POISON_INUSE
, remainder
);
841 metadata_access_disable();
844 while (end
> fault
&& end
[-1] == POISON_INUSE
)
847 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
848 fault
, end
- 1, fault
- start
);
849 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
851 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
855 static int check_object(struct kmem_cache
*s
, struct page
*page
,
856 void *object
, u8 val
)
859 u8
*endobject
= object
+ s
->object_size
;
861 if (s
->flags
& SLAB_RED_ZONE
) {
862 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
863 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
866 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
867 endobject
, val
, s
->inuse
- s
->object_size
))
870 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
871 check_bytes_and_report(s
, page
, p
, "Alignment padding",
872 endobject
, POISON_INUSE
,
873 s
->inuse
- s
->object_size
);
877 if (s
->flags
& SLAB_POISON
) {
878 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
879 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
880 POISON_FREE
, s
->object_size
- 1) ||
881 !check_bytes_and_report(s
, page
, p
, "Poison",
882 p
+ s
->object_size
- 1, POISON_END
, 1)))
885 * check_pad_bytes cleans up on its own.
887 check_pad_bytes(s
, page
, p
);
890 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
892 * Object and freepointer overlap. Cannot check
893 * freepointer while object is allocated.
897 /* Check free pointer validity */
898 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
899 object_err(s
, page
, p
, "Freepointer corrupt");
901 * No choice but to zap it and thus lose the remainder
902 * of the free objects in this slab. May cause
903 * another error because the object count is now wrong.
905 set_freepointer(s
, p
, NULL
);
911 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
915 VM_BUG_ON(!irqs_disabled());
917 if (!PageSlab(page
)) {
918 slab_err(s
, page
, "Not a valid slab page");
922 maxobj
= order_objects(compound_order(page
), s
->size
);
923 if (page
->objects
> maxobj
) {
924 slab_err(s
, page
, "objects %u > max %u",
925 page
->objects
, maxobj
);
928 if (page
->inuse
> page
->objects
) {
929 slab_err(s
, page
, "inuse %u > max %u",
930 page
->inuse
, page
->objects
);
933 /* Slab_pad_check fixes things up after itself */
934 slab_pad_check(s
, page
);
939 * Determine if a certain object on a page is on the freelist. Must hold the
940 * slab lock to guarantee that the chains are in a consistent state.
942 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
950 while (fp
&& nr
<= page
->objects
) {
953 if (!check_valid_pointer(s
, page
, fp
)) {
955 object_err(s
, page
, object
,
956 "Freechain corrupt");
957 set_freepointer(s
, object
, NULL
);
959 slab_err(s
, page
, "Freepointer corrupt");
960 page
->freelist
= NULL
;
961 page
->inuse
= page
->objects
;
962 slab_fix(s
, "Freelist cleared");
968 fp
= get_freepointer(s
, object
);
972 max_objects
= order_objects(compound_order(page
), s
->size
);
973 if (max_objects
> MAX_OBJS_PER_PAGE
)
974 max_objects
= MAX_OBJS_PER_PAGE
;
976 if (page
->objects
!= max_objects
) {
977 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
978 page
->objects
, max_objects
);
979 page
->objects
= max_objects
;
980 slab_fix(s
, "Number of objects adjusted.");
982 if (page
->inuse
!= page
->objects
- nr
) {
983 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
984 page
->inuse
, page
->objects
- nr
);
985 page
->inuse
= page
->objects
- nr
;
986 slab_fix(s
, "Object count adjusted.");
988 return search
== NULL
;
991 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
994 if (s
->flags
& SLAB_TRACE
) {
995 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
997 alloc
? "alloc" : "free",
1002 print_section(KERN_INFO
, "Object ", (void *)object
,
1010 * Tracking of fully allocated slabs for debugging purposes.
1012 static void add_full(struct kmem_cache
*s
,
1013 struct kmem_cache_node
*n
, struct page
*page
)
1015 if (!(s
->flags
& SLAB_STORE_USER
))
1018 lockdep_assert_held(&n
->list_lock
);
1019 list_add(&page
->slab_list
, &n
->full
);
1022 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1024 if (!(s
->flags
& SLAB_STORE_USER
))
1027 lockdep_assert_held(&n
->list_lock
);
1028 list_del(&page
->slab_list
);
1031 /* Tracking of the number of slabs for debugging purposes */
1032 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1034 struct kmem_cache_node
*n
= get_node(s
, node
);
1036 return atomic_long_read(&n
->nr_slabs
);
1039 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1041 return atomic_long_read(&n
->nr_slabs
);
1044 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1046 struct kmem_cache_node
*n
= get_node(s
, node
);
1049 * May be called early in order to allocate a slab for the
1050 * kmem_cache_node structure. Solve the chicken-egg
1051 * dilemma by deferring the increment of the count during
1052 * bootstrap (see early_kmem_cache_node_alloc).
1055 atomic_long_inc(&n
->nr_slabs
);
1056 atomic_long_add(objects
, &n
->total_objects
);
1059 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1061 struct kmem_cache_node
*n
= get_node(s
, node
);
1063 atomic_long_dec(&n
->nr_slabs
);
1064 atomic_long_sub(objects
, &n
->total_objects
);
1067 /* Object debug checks for alloc/free paths */
1068 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1071 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1074 init_object(s
, object
, SLUB_RED_INACTIVE
);
1075 init_tracking(s
, object
);
1079 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1081 if (!(s
->flags
& SLAB_POISON
))
1084 metadata_access_enable();
1085 memset(addr
, POISON_INUSE
, page_size(page
));
1086 metadata_access_disable();
1089 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1090 struct page
*page
, void *object
)
1092 if (!check_slab(s
, page
))
1095 if (!check_valid_pointer(s
, page
, object
)) {
1096 object_err(s
, page
, object
, "Freelist Pointer check fails");
1100 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1106 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1108 void *object
, unsigned long addr
)
1110 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1111 if (!alloc_consistency_checks(s
, page
, object
))
1115 /* Success perform special debug activities for allocs */
1116 if (s
->flags
& SLAB_STORE_USER
)
1117 set_track(s
, object
, TRACK_ALLOC
, addr
);
1118 trace(s
, page
, object
, 1);
1119 init_object(s
, object
, SLUB_RED_ACTIVE
);
1123 if (PageSlab(page
)) {
1125 * If this is a slab page then lets do the best we can
1126 * to avoid issues in the future. Marking all objects
1127 * as used avoids touching the remaining objects.
1129 slab_fix(s
, "Marking all objects used");
1130 page
->inuse
= page
->objects
;
1131 page
->freelist
= NULL
;
1136 static inline int free_consistency_checks(struct kmem_cache
*s
,
1137 struct page
*page
, void *object
, unsigned long addr
)
1139 if (!check_valid_pointer(s
, page
, object
)) {
1140 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1144 if (on_freelist(s
, page
, object
)) {
1145 object_err(s
, page
, object
, "Object already free");
1149 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1152 if (unlikely(s
!= page
->slab_cache
)) {
1153 if (!PageSlab(page
)) {
1154 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1156 } else if (!page
->slab_cache
) {
1157 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1161 object_err(s
, page
, object
,
1162 "page slab pointer corrupt.");
1168 /* Supports checking bulk free of a constructed freelist */
1169 static noinline
int free_debug_processing(
1170 struct kmem_cache
*s
, struct page
*page
,
1171 void *head
, void *tail
, int bulk_cnt
,
1174 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1175 void *object
= head
;
1177 unsigned long uninitialized_var(flags
);
1180 spin_lock_irqsave(&n
->list_lock
, flags
);
1183 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1184 if (!check_slab(s
, page
))
1191 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1192 if (!free_consistency_checks(s
, page
, object
, addr
))
1196 if (s
->flags
& SLAB_STORE_USER
)
1197 set_track(s
, object
, TRACK_FREE
, addr
);
1198 trace(s
, page
, object
, 0);
1199 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1200 init_object(s
, object
, SLUB_RED_INACTIVE
);
1202 /* Reached end of constructed freelist yet? */
1203 if (object
!= tail
) {
1204 object
= get_freepointer(s
, object
);
1210 if (cnt
!= bulk_cnt
)
1211 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1215 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1217 slab_fix(s
, "Object at 0x%p not freed", object
);
1221 static int __init
setup_slub_debug(char *str
)
1223 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1224 if (*str
++ != '=' || !*str
)
1226 * No options specified. Switch on full debugging.
1232 * No options but restriction on slabs. This means full
1233 * debugging for slabs matching a pattern.
1240 * Switch off all debugging measures.
1245 * Determine which debug features should be switched on
1247 for (; *str
&& *str
!= ','; str
++) {
1248 switch (tolower(*str
)) {
1250 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1253 slub_debug
|= SLAB_RED_ZONE
;
1256 slub_debug
|= SLAB_POISON
;
1259 slub_debug
|= SLAB_STORE_USER
;
1262 slub_debug
|= SLAB_TRACE
;
1265 slub_debug
|= SLAB_FAILSLAB
;
1269 * Avoid enabling debugging on caches if its minimum
1270 * order would increase as a result.
1272 disable_higher_order_debug
= 1;
1275 pr_err("slub_debug option '%c' unknown. skipped\n",
1282 slub_debug_slabs
= str
+ 1;
1284 if ((static_branch_unlikely(&init_on_alloc
) ||
1285 static_branch_unlikely(&init_on_free
)) &&
1286 (slub_debug
& SLAB_POISON
))
1287 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1291 __setup("slub_debug", setup_slub_debug
);
1294 * kmem_cache_flags - apply debugging options to the cache
1295 * @object_size: the size of an object without meta data
1296 * @flags: flags to set
1297 * @name: name of the cache
1298 * @ctor: constructor function
1300 * Debug option(s) are applied to @flags. In addition to the debug
1301 * option(s), if a slab name (or multiple) is specified i.e.
1302 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1303 * then only the select slabs will receive the debug option(s).
1305 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1306 slab_flags_t flags
, const char *name
,
1307 void (*ctor
)(void *))
1312 /* If slub_debug = 0, it folds into the if conditional. */
1313 if (!slub_debug_slabs
)
1314 return flags
| slub_debug
;
1317 iter
= slub_debug_slabs
;
1322 end
= strchrnul(iter
, ',');
1324 glob
= strnchr(iter
, end
- iter
, '*');
1326 cmplen
= glob
- iter
;
1328 cmplen
= max_t(size_t, len
, (end
- iter
));
1330 if (!strncmp(name
, iter
, cmplen
)) {
1331 flags
|= slub_debug
;
1342 #else /* !CONFIG_SLUB_DEBUG */
1343 static inline void setup_object_debug(struct kmem_cache
*s
,
1344 struct page
*page
, void *object
) {}
1346 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1348 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1349 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1351 static inline int free_debug_processing(
1352 struct kmem_cache
*s
, struct page
*page
,
1353 void *head
, void *tail
, int bulk_cnt
,
1354 unsigned long addr
) { return 0; }
1356 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1358 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1359 void *object
, u8 val
) { return 1; }
1360 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1361 struct page
*page
) {}
1362 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1363 struct page
*page
) {}
1364 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1365 slab_flags_t flags
, const char *name
,
1366 void (*ctor
)(void *))
1370 #define slub_debug 0
1372 #define disable_higher_order_debug 0
1374 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1376 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1378 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1380 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1383 #endif /* CONFIG_SLUB_DEBUG */
1386 * Hooks for other subsystems that check memory allocations. In a typical
1387 * production configuration these hooks all should produce no code at all.
1389 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1391 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1392 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1393 kmemleak_alloc(ptr
, size
, 1, flags
);
1397 static __always_inline
void kfree_hook(void *x
)
1400 kasan_kfree_large(x
, _RET_IP_
);
1403 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1405 kmemleak_free_recursive(x
, s
->flags
);
1408 * Trouble is that we may no longer disable interrupts in the fast path
1409 * So in order to make the debug calls that expect irqs to be
1410 * disabled we need to disable interrupts temporarily.
1412 #ifdef CONFIG_LOCKDEP
1414 unsigned long flags
;
1416 local_irq_save(flags
);
1417 debug_check_no_locks_freed(x
, s
->object_size
);
1418 local_irq_restore(flags
);
1421 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1422 debug_check_no_obj_freed(x
, s
->object_size
);
1424 /* KASAN might put x into memory quarantine, delaying its reuse */
1425 return kasan_slab_free(s
, x
, _RET_IP_
);
1428 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1429 void **head
, void **tail
)
1434 void *old_tail
= *tail
? *tail
: *head
;
1437 /* Head and tail of the reconstructed freelist */
1443 next
= get_freepointer(s
, object
);
1445 if (slab_want_init_on_free(s
)) {
1447 * Clear the object and the metadata, but don't touch
1450 memset(object
, 0, s
->object_size
);
1451 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
1453 memset((char *)object
+ s
->inuse
, 0,
1454 s
->size
- s
->inuse
- rsize
);
1457 /* If object's reuse doesn't have to be delayed */
1458 if (!slab_free_hook(s
, object
)) {
1459 /* Move object to the new freelist */
1460 set_freepointer(s
, object
, *head
);
1465 } while (object
!= old_tail
);
1470 return *head
!= NULL
;
1473 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1476 setup_object_debug(s
, page
, object
);
1477 object
= kasan_init_slab_obj(s
, object
);
1478 if (unlikely(s
->ctor
)) {
1479 kasan_unpoison_object_data(s
, object
);
1481 kasan_poison_object_data(s
, object
);
1487 * Slab allocation and freeing
1489 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1490 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1493 unsigned int order
= oo_order(oo
);
1495 if (node
== NUMA_NO_NODE
)
1496 page
= alloc_pages(flags
, order
);
1498 page
= __alloc_pages_node(node
, flags
, order
);
1500 if (page
&& charge_slab_page(page
, flags
, order
, s
)) {
1501 __free_pages(page
, order
);
1508 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1509 /* Pre-initialize the random sequence cache */
1510 static int init_cache_random_seq(struct kmem_cache
*s
)
1512 unsigned int count
= oo_objects(s
->oo
);
1515 /* Bailout if already initialised */
1519 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1521 pr_err("SLUB: Unable to initialize free list for %s\n",
1526 /* Transform to an offset on the set of pages */
1527 if (s
->random_seq
) {
1530 for (i
= 0; i
< count
; i
++)
1531 s
->random_seq
[i
] *= s
->size
;
1536 /* Initialize each random sequence freelist per cache */
1537 static void __init
init_freelist_randomization(void)
1539 struct kmem_cache
*s
;
1541 mutex_lock(&slab_mutex
);
1543 list_for_each_entry(s
, &slab_caches
, list
)
1544 init_cache_random_seq(s
);
1546 mutex_unlock(&slab_mutex
);
1549 /* Get the next entry on the pre-computed freelist randomized */
1550 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1551 unsigned long *pos
, void *start
,
1552 unsigned long page_limit
,
1553 unsigned long freelist_count
)
1558 * If the target page allocation failed, the number of objects on the
1559 * page might be smaller than the usual size defined by the cache.
1562 idx
= s
->random_seq
[*pos
];
1564 if (*pos
>= freelist_count
)
1566 } while (unlikely(idx
>= page_limit
));
1568 return (char *)start
+ idx
;
1571 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1572 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1577 unsigned long idx
, pos
, page_limit
, freelist_count
;
1579 if (page
->objects
< 2 || !s
->random_seq
)
1582 freelist_count
= oo_objects(s
->oo
);
1583 pos
= get_random_int() % freelist_count
;
1585 page_limit
= page
->objects
* s
->size
;
1586 start
= fixup_red_left(s
, page_address(page
));
1588 /* First entry is used as the base of the freelist */
1589 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1591 cur
= setup_object(s
, page
, cur
);
1592 page
->freelist
= cur
;
1594 for (idx
= 1; idx
< page
->objects
; idx
++) {
1595 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1597 next
= setup_object(s
, page
, next
);
1598 set_freepointer(s
, cur
, next
);
1601 set_freepointer(s
, cur
, NULL
);
1606 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1610 static inline void init_freelist_randomization(void) { }
1611 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1615 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1617 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1620 struct kmem_cache_order_objects oo
= s
->oo
;
1622 void *start
, *p
, *next
;
1626 flags
&= gfp_allowed_mask
;
1628 if (gfpflags_allow_blocking(flags
))
1631 flags
|= s
->allocflags
;
1634 * Let the initial higher-order allocation fail under memory pressure
1635 * so we fall-back to the minimum order allocation.
1637 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1638 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1639 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1641 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1642 if (unlikely(!page
)) {
1646 * Allocation may have failed due to fragmentation.
1647 * Try a lower order alloc if possible
1649 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1650 if (unlikely(!page
))
1652 stat(s
, ORDER_FALLBACK
);
1655 page
->objects
= oo_objects(oo
);
1657 page
->slab_cache
= s
;
1658 __SetPageSlab(page
);
1659 if (page_is_pfmemalloc(page
))
1660 SetPageSlabPfmemalloc(page
);
1662 kasan_poison_slab(page
);
1664 start
= page_address(page
);
1666 setup_page_debug(s
, page
, start
);
1668 shuffle
= shuffle_freelist(s
, page
);
1671 start
= fixup_red_left(s
, start
);
1672 start
= setup_object(s
, page
, start
);
1673 page
->freelist
= start
;
1674 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1676 next
= setup_object(s
, page
, next
);
1677 set_freepointer(s
, p
, next
);
1680 set_freepointer(s
, p
, NULL
);
1683 page
->inuse
= page
->objects
;
1687 if (gfpflags_allow_blocking(flags
))
1688 local_irq_disable();
1692 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1697 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1699 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1700 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1701 flags
&= ~GFP_SLAB_BUG_MASK
;
1702 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1703 invalid_mask
, &invalid_mask
, flags
, &flags
);
1707 return allocate_slab(s
,
1708 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1711 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1713 int order
= compound_order(page
);
1714 int pages
= 1 << order
;
1716 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1719 slab_pad_check(s
, page
);
1720 for_each_object(p
, s
, page_address(page
),
1722 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1725 __ClearPageSlabPfmemalloc(page
);
1726 __ClearPageSlab(page
);
1728 page
->mapping
= NULL
;
1729 if (current
->reclaim_state
)
1730 current
->reclaim_state
->reclaimed_slab
+= pages
;
1731 uncharge_slab_page(page
, order
, s
);
1732 __free_pages(page
, order
);
1735 static void rcu_free_slab(struct rcu_head
*h
)
1737 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1739 __free_slab(page
->slab_cache
, page
);
1742 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1744 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1745 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1747 __free_slab(s
, page
);
1750 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1752 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1757 * Management of partially allocated slabs.
1760 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1763 if (tail
== DEACTIVATE_TO_TAIL
)
1764 list_add_tail(&page
->slab_list
, &n
->partial
);
1766 list_add(&page
->slab_list
, &n
->partial
);
1769 static inline void add_partial(struct kmem_cache_node
*n
,
1770 struct page
*page
, int tail
)
1772 lockdep_assert_held(&n
->list_lock
);
1773 __add_partial(n
, page
, tail
);
1776 static inline void remove_partial(struct kmem_cache_node
*n
,
1779 lockdep_assert_held(&n
->list_lock
);
1780 list_del(&page
->slab_list
);
1785 * Remove slab from the partial list, freeze it and
1786 * return the pointer to the freelist.
1788 * Returns a list of objects or NULL if it fails.
1790 static inline void *acquire_slab(struct kmem_cache
*s
,
1791 struct kmem_cache_node
*n
, struct page
*page
,
1792 int mode
, int *objects
)
1795 unsigned long counters
;
1798 lockdep_assert_held(&n
->list_lock
);
1801 * Zap the freelist and set the frozen bit.
1802 * The old freelist is the list of objects for the
1803 * per cpu allocation list.
1805 freelist
= page
->freelist
;
1806 counters
= page
->counters
;
1807 new.counters
= counters
;
1808 *objects
= new.objects
- new.inuse
;
1810 new.inuse
= page
->objects
;
1811 new.freelist
= NULL
;
1813 new.freelist
= freelist
;
1816 VM_BUG_ON(new.frozen
);
1819 if (!__cmpxchg_double_slab(s
, page
,
1821 new.freelist
, new.counters
,
1825 remove_partial(n
, page
);
1830 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1831 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1834 * Try to allocate a partial slab from a specific node.
1836 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1837 struct kmem_cache_cpu
*c
, gfp_t flags
)
1839 struct page
*page
, *page2
;
1840 void *object
= NULL
;
1841 unsigned int available
= 0;
1845 * Racy check. If we mistakenly see no partial slabs then we
1846 * just allocate an empty slab. If we mistakenly try to get a
1847 * partial slab and there is none available then get_partials()
1850 if (!n
|| !n
->nr_partial
)
1853 spin_lock(&n
->list_lock
);
1854 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1857 if (!pfmemalloc_match(page
, flags
))
1860 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1864 available
+= objects
;
1867 stat(s
, ALLOC_FROM_PARTIAL
);
1870 put_cpu_partial(s
, page
, 0);
1871 stat(s
, CPU_PARTIAL_NODE
);
1873 if (!kmem_cache_has_cpu_partial(s
)
1874 || available
> slub_cpu_partial(s
) / 2)
1878 spin_unlock(&n
->list_lock
);
1883 * Get a page from somewhere. Search in increasing NUMA distances.
1885 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1886 struct kmem_cache_cpu
*c
)
1889 struct zonelist
*zonelist
;
1892 enum zone_type high_zoneidx
= gfp_zone(flags
);
1894 unsigned int cpuset_mems_cookie
;
1897 * The defrag ratio allows a configuration of the tradeoffs between
1898 * inter node defragmentation and node local allocations. A lower
1899 * defrag_ratio increases the tendency to do local allocations
1900 * instead of attempting to obtain partial slabs from other nodes.
1902 * If the defrag_ratio is set to 0 then kmalloc() always
1903 * returns node local objects. If the ratio is higher then kmalloc()
1904 * may return off node objects because partial slabs are obtained
1905 * from other nodes and filled up.
1907 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1908 * (which makes defrag_ratio = 1000) then every (well almost)
1909 * allocation will first attempt to defrag slab caches on other nodes.
1910 * This means scanning over all nodes to look for partial slabs which
1911 * may be expensive if we do it every time we are trying to find a slab
1912 * with available objects.
1914 if (!s
->remote_node_defrag_ratio
||
1915 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1919 cpuset_mems_cookie
= read_mems_allowed_begin();
1920 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1921 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1922 struct kmem_cache_node
*n
;
1924 n
= get_node(s
, zone_to_nid(zone
));
1926 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1927 n
->nr_partial
> s
->min_partial
) {
1928 object
= get_partial_node(s
, n
, c
, flags
);
1931 * Don't check read_mems_allowed_retry()
1932 * here - if mems_allowed was updated in
1933 * parallel, that was a harmless race
1934 * between allocation and the cpuset
1941 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1942 #endif /* CONFIG_NUMA */
1947 * Get a partial page, lock it and return it.
1949 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1950 struct kmem_cache_cpu
*c
)
1953 int searchnode
= node
;
1955 if (node
== NUMA_NO_NODE
)
1956 searchnode
= numa_mem_id();
1957 else if (!node_present_pages(node
))
1958 searchnode
= node_to_mem_node(node
);
1960 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1961 if (object
|| node
!= NUMA_NO_NODE
)
1964 return get_any_partial(s
, flags
, c
);
1967 #ifdef CONFIG_PREEMPT
1969 * Calculate the next globally unique transaction for disambiguiation
1970 * during cmpxchg. The transactions start with the cpu number and are then
1971 * incremented by CONFIG_NR_CPUS.
1973 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1976 * No preemption supported therefore also no need to check for
1982 static inline unsigned long next_tid(unsigned long tid
)
1984 return tid
+ TID_STEP
;
1987 #ifdef SLUB_DEBUG_CMPXCHG
1988 static inline unsigned int tid_to_cpu(unsigned long tid
)
1990 return tid
% TID_STEP
;
1993 static inline unsigned long tid_to_event(unsigned long tid
)
1995 return tid
/ TID_STEP
;
1999 static inline unsigned int init_tid(int cpu
)
2004 static inline void note_cmpxchg_failure(const char *n
,
2005 const struct kmem_cache
*s
, unsigned long tid
)
2007 #ifdef SLUB_DEBUG_CMPXCHG
2008 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2010 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2012 #ifdef CONFIG_PREEMPT
2013 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2014 pr_warn("due to cpu change %d -> %d\n",
2015 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2018 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2019 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2020 tid_to_event(tid
), tid_to_event(actual_tid
));
2022 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2023 actual_tid
, tid
, next_tid(tid
));
2025 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2028 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2032 for_each_possible_cpu(cpu
)
2033 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2037 * Remove the cpu slab
2039 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2040 void *freelist
, struct kmem_cache_cpu
*c
)
2042 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2043 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2045 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2047 int tail
= DEACTIVATE_TO_HEAD
;
2051 if (page
->freelist
) {
2052 stat(s
, DEACTIVATE_REMOTE_FREES
);
2053 tail
= DEACTIVATE_TO_TAIL
;
2057 * Stage one: Free all available per cpu objects back
2058 * to the page freelist while it is still frozen. Leave the
2061 * There is no need to take the list->lock because the page
2064 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2066 unsigned long counters
;
2069 prior
= page
->freelist
;
2070 counters
= page
->counters
;
2071 set_freepointer(s
, freelist
, prior
);
2072 new.counters
= counters
;
2074 VM_BUG_ON(!new.frozen
);
2076 } while (!__cmpxchg_double_slab(s
, page
,
2078 freelist
, new.counters
,
2079 "drain percpu freelist"));
2081 freelist
= nextfree
;
2085 * Stage two: Ensure that the page is unfrozen while the
2086 * list presence reflects the actual number of objects
2089 * We setup the list membership and then perform a cmpxchg
2090 * with the count. If there is a mismatch then the page
2091 * is not unfrozen but the page is on the wrong list.
2093 * Then we restart the process which may have to remove
2094 * the page from the list that we just put it on again
2095 * because the number of objects in the slab may have
2100 old
.freelist
= page
->freelist
;
2101 old
.counters
= page
->counters
;
2102 VM_BUG_ON(!old
.frozen
);
2104 /* Determine target state of the slab */
2105 new.counters
= old
.counters
;
2108 set_freepointer(s
, freelist
, old
.freelist
);
2109 new.freelist
= freelist
;
2111 new.freelist
= old
.freelist
;
2115 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2117 else if (new.freelist
) {
2122 * Taking the spinlock removes the possibility
2123 * that acquire_slab() will see a slab page that
2126 spin_lock(&n
->list_lock
);
2130 if (kmem_cache_debug(s
) && !lock
) {
2133 * This also ensures that the scanning of full
2134 * slabs from diagnostic functions will not see
2137 spin_lock(&n
->list_lock
);
2143 remove_partial(n
, page
);
2144 else if (l
== M_FULL
)
2145 remove_full(s
, n
, page
);
2148 add_partial(n
, page
, tail
);
2149 else if (m
== M_FULL
)
2150 add_full(s
, n
, page
);
2154 if (!__cmpxchg_double_slab(s
, page
,
2155 old
.freelist
, old
.counters
,
2156 new.freelist
, new.counters
,
2161 spin_unlock(&n
->list_lock
);
2165 else if (m
== M_FULL
)
2166 stat(s
, DEACTIVATE_FULL
);
2167 else if (m
== M_FREE
) {
2168 stat(s
, DEACTIVATE_EMPTY
);
2169 discard_slab(s
, page
);
2178 * Unfreeze all the cpu partial slabs.
2180 * This function must be called with interrupts disabled
2181 * for the cpu using c (or some other guarantee must be there
2182 * to guarantee no concurrent accesses).
2184 static void unfreeze_partials(struct kmem_cache
*s
,
2185 struct kmem_cache_cpu
*c
)
2187 #ifdef CONFIG_SLUB_CPU_PARTIAL
2188 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2189 struct page
*page
, *discard_page
= NULL
;
2191 while ((page
= c
->partial
)) {
2195 c
->partial
= page
->next
;
2197 n2
= get_node(s
, page_to_nid(page
));
2200 spin_unlock(&n
->list_lock
);
2203 spin_lock(&n
->list_lock
);
2208 old
.freelist
= page
->freelist
;
2209 old
.counters
= page
->counters
;
2210 VM_BUG_ON(!old
.frozen
);
2212 new.counters
= old
.counters
;
2213 new.freelist
= old
.freelist
;
2217 } while (!__cmpxchg_double_slab(s
, page
,
2218 old
.freelist
, old
.counters
,
2219 new.freelist
, new.counters
,
2220 "unfreezing slab"));
2222 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2223 page
->next
= discard_page
;
2224 discard_page
= page
;
2226 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2227 stat(s
, FREE_ADD_PARTIAL
);
2232 spin_unlock(&n
->list_lock
);
2234 while (discard_page
) {
2235 page
= discard_page
;
2236 discard_page
= discard_page
->next
;
2238 stat(s
, DEACTIVATE_EMPTY
);
2239 discard_slab(s
, page
);
2242 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2246 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2247 * partial page slot if available.
2249 * If we did not find a slot then simply move all the partials to the
2250 * per node partial list.
2252 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2254 #ifdef CONFIG_SLUB_CPU_PARTIAL
2255 struct page
*oldpage
;
2263 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2266 pobjects
= oldpage
->pobjects
;
2267 pages
= oldpage
->pages
;
2268 if (drain
&& pobjects
> s
->cpu_partial
) {
2269 unsigned long flags
;
2271 * partial array is full. Move the existing
2272 * set to the per node partial list.
2274 local_irq_save(flags
);
2275 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2276 local_irq_restore(flags
);
2280 stat(s
, CPU_PARTIAL_DRAIN
);
2285 pobjects
+= page
->objects
- page
->inuse
;
2287 page
->pages
= pages
;
2288 page
->pobjects
= pobjects
;
2289 page
->next
= oldpage
;
2291 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2293 if (unlikely(!s
->cpu_partial
)) {
2294 unsigned long flags
;
2296 local_irq_save(flags
);
2297 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2298 local_irq_restore(flags
);
2301 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2304 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2306 stat(s
, CPUSLAB_FLUSH
);
2307 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2309 c
->tid
= next_tid(c
->tid
);
2315 * Called from IPI handler with interrupts disabled.
2317 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2319 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2324 unfreeze_partials(s
, c
);
2327 static void flush_cpu_slab(void *d
)
2329 struct kmem_cache
*s
= d
;
2331 __flush_cpu_slab(s
, smp_processor_id());
2334 static bool has_cpu_slab(int cpu
, void *info
)
2336 struct kmem_cache
*s
= info
;
2337 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2339 return c
->page
|| slub_percpu_partial(c
);
2342 static void flush_all(struct kmem_cache
*s
)
2344 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2348 * Use the cpu notifier to insure that the cpu slabs are flushed when
2351 static int slub_cpu_dead(unsigned int cpu
)
2353 struct kmem_cache
*s
;
2354 unsigned long flags
;
2356 mutex_lock(&slab_mutex
);
2357 list_for_each_entry(s
, &slab_caches
, list
) {
2358 local_irq_save(flags
);
2359 __flush_cpu_slab(s
, cpu
);
2360 local_irq_restore(flags
);
2362 mutex_unlock(&slab_mutex
);
2367 * Check if the objects in a per cpu structure fit numa
2368 * locality expectations.
2370 static inline int node_match(struct page
*page
, int node
)
2373 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2379 #ifdef CONFIG_SLUB_DEBUG
2380 static int count_free(struct page
*page
)
2382 return page
->objects
- page
->inuse
;
2385 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2387 return atomic_long_read(&n
->total_objects
);
2389 #endif /* CONFIG_SLUB_DEBUG */
2391 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2392 static unsigned long count_partial(struct kmem_cache_node
*n
,
2393 int (*get_count
)(struct page
*))
2395 unsigned long flags
;
2396 unsigned long x
= 0;
2399 spin_lock_irqsave(&n
->list_lock
, flags
);
2400 list_for_each_entry(page
, &n
->partial
, slab_list
)
2401 x
+= get_count(page
);
2402 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2405 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2407 static noinline
void
2408 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2410 #ifdef CONFIG_SLUB_DEBUG
2411 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2412 DEFAULT_RATELIMIT_BURST
);
2414 struct kmem_cache_node
*n
;
2416 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2419 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2420 nid
, gfpflags
, &gfpflags
);
2421 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2422 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2425 if (oo_order(s
->min
) > get_order(s
->object_size
))
2426 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2429 for_each_kmem_cache_node(s
, node
, n
) {
2430 unsigned long nr_slabs
;
2431 unsigned long nr_objs
;
2432 unsigned long nr_free
;
2434 nr_free
= count_partial(n
, count_free
);
2435 nr_slabs
= node_nr_slabs(n
);
2436 nr_objs
= node_nr_objs(n
);
2438 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2439 node
, nr_slabs
, nr_objs
, nr_free
);
2444 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2445 int node
, struct kmem_cache_cpu
**pc
)
2448 struct kmem_cache_cpu
*c
= *pc
;
2451 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2453 freelist
= get_partial(s
, flags
, node
, c
);
2458 page
= new_slab(s
, flags
, node
);
2460 c
= raw_cpu_ptr(s
->cpu_slab
);
2465 * No other reference to the page yet so we can
2466 * muck around with it freely without cmpxchg
2468 freelist
= page
->freelist
;
2469 page
->freelist
= NULL
;
2471 stat(s
, ALLOC_SLAB
);
2479 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2481 if (unlikely(PageSlabPfmemalloc(page
)))
2482 return gfp_pfmemalloc_allowed(gfpflags
);
2488 * Check the page->freelist of a page and either transfer the freelist to the
2489 * per cpu freelist or deactivate the page.
2491 * The page is still frozen if the return value is not NULL.
2493 * If this function returns NULL then the page has been unfrozen.
2495 * This function must be called with interrupt disabled.
2497 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2500 unsigned long counters
;
2504 freelist
= page
->freelist
;
2505 counters
= page
->counters
;
2507 new.counters
= counters
;
2508 VM_BUG_ON(!new.frozen
);
2510 new.inuse
= page
->objects
;
2511 new.frozen
= freelist
!= NULL
;
2513 } while (!__cmpxchg_double_slab(s
, page
,
2522 * Slow path. The lockless freelist is empty or we need to perform
2525 * Processing is still very fast if new objects have been freed to the
2526 * regular freelist. In that case we simply take over the regular freelist
2527 * as the lockless freelist and zap the regular freelist.
2529 * If that is not working then we fall back to the partial lists. We take the
2530 * first element of the freelist as the object to allocate now and move the
2531 * rest of the freelist to the lockless freelist.
2533 * And if we were unable to get a new slab from the partial slab lists then
2534 * we need to allocate a new slab. This is the slowest path since it involves
2535 * a call to the page allocator and the setup of a new slab.
2537 * Version of __slab_alloc to use when we know that interrupts are
2538 * already disabled (which is the case for bulk allocation).
2540 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2541 unsigned long addr
, struct kmem_cache_cpu
*c
)
2551 if (unlikely(!node_match(page
, node
))) {
2552 int searchnode
= node
;
2554 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2555 searchnode
= node_to_mem_node(node
);
2557 if (unlikely(!node_match(page
, searchnode
))) {
2558 stat(s
, ALLOC_NODE_MISMATCH
);
2559 deactivate_slab(s
, page
, c
->freelist
, c
);
2565 * By rights, we should be searching for a slab page that was
2566 * PFMEMALLOC but right now, we are losing the pfmemalloc
2567 * information when the page leaves the per-cpu allocator
2569 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2570 deactivate_slab(s
, page
, c
->freelist
, c
);
2574 /* must check again c->freelist in case of cpu migration or IRQ */
2575 freelist
= c
->freelist
;
2579 freelist
= get_freelist(s
, page
);
2583 stat(s
, DEACTIVATE_BYPASS
);
2587 stat(s
, ALLOC_REFILL
);
2591 * freelist is pointing to the list of objects to be used.
2592 * page is pointing to the page from which the objects are obtained.
2593 * That page must be frozen for per cpu allocations to work.
2595 VM_BUG_ON(!c
->page
->frozen
);
2596 c
->freelist
= get_freepointer(s
, freelist
);
2597 c
->tid
= next_tid(c
->tid
);
2602 if (slub_percpu_partial(c
)) {
2603 page
= c
->page
= slub_percpu_partial(c
);
2604 slub_set_percpu_partial(c
, page
);
2605 stat(s
, CPU_PARTIAL_ALLOC
);
2609 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2611 if (unlikely(!freelist
)) {
2612 slab_out_of_memory(s
, gfpflags
, node
);
2617 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2620 /* Only entered in the debug case */
2621 if (kmem_cache_debug(s
) &&
2622 !alloc_debug_processing(s
, page
, freelist
, addr
))
2623 goto new_slab
; /* Slab failed checks. Next slab needed */
2625 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2630 * Another one that disabled interrupt and compensates for possible
2631 * cpu changes by refetching the per cpu area pointer.
2633 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2634 unsigned long addr
, struct kmem_cache_cpu
*c
)
2637 unsigned long flags
;
2639 local_irq_save(flags
);
2640 #ifdef CONFIG_PREEMPT
2642 * We may have been preempted and rescheduled on a different
2643 * cpu before disabling interrupts. Need to reload cpu area
2646 c
= this_cpu_ptr(s
->cpu_slab
);
2649 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2650 local_irq_restore(flags
);
2655 * If the object has been wiped upon free, make sure it's fully initialized by
2656 * zeroing out freelist pointer.
2658 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2661 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2662 memset((void *)((char *)obj
+ s
->offset
), 0, sizeof(void *));
2666 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2667 * have the fastpath folded into their functions. So no function call
2668 * overhead for requests that can be satisfied on the fastpath.
2670 * The fastpath works by first checking if the lockless freelist can be used.
2671 * If not then __slab_alloc is called for slow processing.
2673 * Otherwise we can simply pick the next object from the lockless free list.
2675 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2676 gfp_t gfpflags
, int node
, unsigned long addr
)
2679 struct kmem_cache_cpu
*c
;
2683 s
= slab_pre_alloc_hook(s
, gfpflags
);
2688 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2689 * enabled. We may switch back and forth between cpus while
2690 * reading from one cpu area. That does not matter as long
2691 * as we end up on the original cpu again when doing the cmpxchg.
2693 * We should guarantee that tid and kmem_cache are retrieved on
2694 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2695 * to check if it is matched or not.
2698 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2699 c
= raw_cpu_ptr(s
->cpu_slab
);
2700 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2701 unlikely(tid
!= READ_ONCE(c
->tid
)));
2704 * Irqless object alloc/free algorithm used here depends on sequence
2705 * of fetching cpu_slab's data. tid should be fetched before anything
2706 * on c to guarantee that object and page associated with previous tid
2707 * won't be used with current tid. If we fetch tid first, object and
2708 * page could be one associated with next tid and our alloc/free
2709 * request will be failed. In this case, we will retry. So, no problem.
2714 * The transaction ids are globally unique per cpu and per operation on
2715 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2716 * occurs on the right processor and that there was no operation on the
2717 * linked list in between.
2720 object
= c
->freelist
;
2722 if (unlikely(!object
|| !node_match(page
, node
))) {
2723 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2724 stat(s
, ALLOC_SLOWPATH
);
2726 void *next_object
= get_freepointer_safe(s
, object
);
2729 * The cmpxchg will only match if there was no additional
2730 * operation and if we are on the right processor.
2732 * The cmpxchg does the following atomically (without lock
2734 * 1. Relocate first pointer to the current per cpu area.
2735 * 2. Verify that tid and freelist have not been changed
2736 * 3. If they were not changed replace tid and freelist
2738 * Since this is without lock semantics the protection is only
2739 * against code executing on this cpu *not* from access by
2742 if (unlikely(!this_cpu_cmpxchg_double(
2743 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2745 next_object
, next_tid(tid
)))) {
2747 note_cmpxchg_failure("slab_alloc", s
, tid
);
2750 prefetch_freepointer(s
, next_object
);
2751 stat(s
, ALLOC_FASTPATH
);
2754 maybe_wipe_obj_freeptr(s
, object
);
2756 if (unlikely(slab_want_init_on_alloc(gfpflags
, s
)) && object
)
2757 memset(object
, 0, s
->object_size
);
2759 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2764 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2765 gfp_t gfpflags
, unsigned long addr
)
2767 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2770 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2772 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2774 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2779 EXPORT_SYMBOL(kmem_cache_alloc
);
2781 #ifdef CONFIG_TRACING
2782 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2784 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2785 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2786 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2789 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2793 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2795 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2797 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2798 s
->object_size
, s
->size
, gfpflags
, node
);
2802 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2804 #ifdef CONFIG_TRACING
2805 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2807 int node
, size_t size
)
2809 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2811 trace_kmalloc_node(_RET_IP_
, ret
,
2812 size
, s
->size
, gfpflags
, node
);
2814 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2817 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2819 #endif /* CONFIG_NUMA */
2822 * Slow path handling. This may still be called frequently since objects
2823 * have a longer lifetime than the cpu slabs in most processing loads.
2825 * So we still attempt to reduce cache line usage. Just take the slab
2826 * lock and free the item. If there is no additional partial page
2827 * handling required then we can return immediately.
2829 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2830 void *head
, void *tail
, int cnt
,
2837 unsigned long counters
;
2838 struct kmem_cache_node
*n
= NULL
;
2839 unsigned long uninitialized_var(flags
);
2841 stat(s
, FREE_SLOWPATH
);
2843 if (kmem_cache_debug(s
) &&
2844 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2849 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2852 prior
= page
->freelist
;
2853 counters
= page
->counters
;
2854 set_freepointer(s
, tail
, prior
);
2855 new.counters
= counters
;
2856 was_frozen
= new.frozen
;
2858 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2860 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2863 * Slab was on no list before and will be
2865 * We can defer the list move and instead
2870 } else { /* Needs to be taken off a list */
2872 n
= get_node(s
, page_to_nid(page
));
2874 * Speculatively acquire the list_lock.
2875 * If the cmpxchg does not succeed then we may
2876 * drop the list_lock without any processing.
2878 * Otherwise the list_lock will synchronize with
2879 * other processors updating the list of slabs.
2881 spin_lock_irqsave(&n
->list_lock
, flags
);
2886 } while (!cmpxchg_double_slab(s
, page
,
2894 * If we just froze the page then put it onto the
2895 * per cpu partial list.
2897 if (new.frozen
&& !was_frozen
) {
2898 put_cpu_partial(s
, page
, 1);
2899 stat(s
, CPU_PARTIAL_FREE
);
2902 * The list lock was not taken therefore no list
2903 * activity can be necessary.
2906 stat(s
, FREE_FROZEN
);
2910 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2914 * Objects left in the slab. If it was not on the partial list before
2917 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2918 remove_full(s
, n
, page
);
2919 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2920 stat(s
, FREE_ADD_PARTIAL
);
2922 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2928 * Slab on the partial list.
2930 remove_partial(n
, page
);
2931 stat(s
, FREE_REMOVE_PARTIAL
);
2933 /* Slab must be on the full list */
2934 remove_full(s
, n
, page
);
2937 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2939 discard_slab(s
, page
);
2943 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2944 * can perform fastpath freeing without additional function calls.
2946 * The fastpath is only possible if we are freeing to the current cpu slab
2947 * of this processor. This typically the case if we have just allocated
2950 * If fastpath is not possible then fall back to __slab_free where we deal
2951 * with all sorts of special processing.
2953 * Bulk free of a freelist with several objects (all pointing to the
2954 * same page) possible by specifying head and tail ptr, plus objects
2955 * count (cnt). Bulk free indicated by tail pointer being set.
2957 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2958 struct page
*page
, void *head
, void *tail
,
2959 int cnt
, unsigned long addr
)
2961 void *tail_obj
= tail
? : head
;
2962 struct kmem_cache_cpu
*c
;
2966 * Determine the currently cpus per cpu slab.
2967 * The cpu may change afterward. However that does not matter since
2968 * data is retrieved via this pointer. If we are on the same cpu
2969 * during the cmpxchg then the free will succeed.
2972 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2973 c
= raw_cpu_ptr(s
->cpu_slab
);
2974 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2975 unlikely(tid
!= READ_ONCE(c
->tid
)));
2977 /* Same with comment on barrier() in slab_alloc_node() */
2980 if (likely(page
== c
->page
)) {
2981 set_freepointer(s
, tail_obj
, c
->freelist
);
2983 if (unlikely(!this_cpu_cmpxchg_double(
2984 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2986 head
, next_tid(tid
)))) {
2988 note_cmpxchg_failure("slab_free", s
, tid
);
2991 stat(s
, FREE_FASTPATH
);
2993 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2997 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2998 void *head
, void *tail
, int cnt
,
3002 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3003 * to remove objects, whose reuse must be delayed.
3005 if (slab_free_freelist_hook(s
, &head
, &tail
))
3006 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3009 #ifdef CONFIG_KASAN_GENERIC
3010 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3012 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3016 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3018 s
= cache_from_obj(s
, x
);
3021 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3022 trace_kmem_cache_free(_RET_IP_
, x
);
3024 EXPORT_SYMBOL(kmem_cache_free
);
3026 struct detached_freelist
{
3031 struct kmem_cache
*s
;
3035 * This function progressively scans the array with free objects (with
3036 * a limited look ahead) and extract objects belonging to the same
3037 * page. It builds a detached freelist directly within the given
3038 * page/objects. This can happen without any need for
3039 * synchronization, because the objects are owned by running process.
3040 * The freelist is build up as a single linked list in the objects.
3041 * The idea is, that this detached freelist can then be bulk
3042 * transferred to the real freelist(s), but only requiring a single
3043 * synchronization primitive. Look ahead in the array is limited due
3044 * to performance reasons.
3047 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3048 void **p
, struct detached_freelist
*df
)
3050 size_t first_skipped_index
= 0;
3055 /* Always re-init detached_freelist */
3060 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3061 } while (!object
&& size
);
3066 page
= virt_to_head_page(object
);
3068 /* Handle kalloc'ed objects */
3069 if (unlikely(!PageSlab(page
))) {
3070 BUG_ON(!PageCompound(page
));
3072 __free_pages(page
, compound_order(page
));
3073 p
[size
] = NULL
; /* mark object processed */
3076 /* Derive kmem_cache from object */
3077 df
->s
= page
->slab_cache
;
3079 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3082 /* Start new detached freelist */
3084 set_freepointer(df
->s
, object
, NULL
);
3086 df
->freelist
= object
;
3087 p
[size
] = NULL
; /* mark object processed */
3093 continue; /* Skip processed objects */
3095 /* df->page is always set at this point */
3096 if (df
->page
== virt_to_head_page(object
)) {
3097 /* Opportunity build freelist */
3098 set_freepointer(df
->s
, object
, df
->freelist
);
3099 df
->freelist
= object
;
3101 p
[size
] = NULL
; /* mark object processed */
3106 /* Limit look ahead search */
3110 if (!first_skipped_index
)
3111 first_skipped_index
= size
+ 1;
3114 return first_skipped_index
;
3117 /* Note that interrupts must be enabled when calling this function. */
3118 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3124 struct detached_freelist df
;
3126 size
= build_detached_freelist(s
, size
, p
, &df
);
3130 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3131 } while (likely(size
));
3133 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3135 /* Note that interrupts must be enabled when calling this function. */
3136 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3139 struct kmem_cache_cpu
*c
;
3142 /* memcg and kmem_cache debug support */
3143 s
= slab_pre_alloc_hook(s
, flags
);
3147 * Drain objects in the per cpu slab, while disabling local
3148 * IRQs, which protects against PREEMPT and interrupts
3149 * handlers invoking normal fastpath.
3151 local_irq_disable();
3152 c
= this_cpu_ptr(s
->cpu_slab
);
3154 for (i
= 0; i
< size
; i
++) {
3155 void *object
= c
->freelist
;
3157 if (unlikely(!object
)) {
3159 * Invoking slow path likely have side-effect
3160 * of re-populating per CPU c->freelist
3162 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3164 if (unlikely(!p
[i
]))
3167 c
= this_cpu_ptr(s
->cpu_slab
);
3168 maybe_wipe_obj_freeptr(s
, p
[i
]);
3170 continue; /* goto for-loop */
3172 c
->freelist
= get_freepointer(s
, object
);
3174 maybe_wipe_obj_freeptr(s
, p
[i
]);
3176 c
->tid
= next_tid(c
->tid
);
3179 /* Clear memory outside IRQ disabled fastpath loop */
3180 if (unlikely(slab_want_init_on_alloc(flags
, s
))) {
3183 for (j
= 0; j
< i
; j
++)
3184 memset(p
[j
], 0, s
->object_size
);
3187 /* memcg and kmem_cache debug support */
3188 slab_post_alloc_hook(s
, flags
, size
, p
);
3192 slab_post_alloc_hook(s
, flags
, i
, p
);
3193 __kmem_cache_free_bulk(s
, i
, p
);
3196 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3200 * Object placement in a slab is made very easy because we always start at
3201 * offset 0. If we tune the size of the object to the alignment then we can
3202 * get the required alignment by putting one properly sized object after
3205 * Notice that the allocation order determines the sizes of the per cpu
3206 * caches. Each processor has always one slab available for allocations.
3207 * Increasing the allocation order reduces the number of times that slabs
3208 * must be moved on and off the partial lists and is therefore a factor in
3213 * Mininum / Maximum order of slab pages. This influences locking overhead
3214 * and slab fragmentation. A higher order reduces the number of partial slabs
3215 * and increases the number of allocations possible without having to
3216 * take the list_lock.
3218 static unsigned int slub_min_order
;
3219 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3220 static unsigned int slub_min_objects
;
3223 * Calculate the order of allocation given an slab object size.
3225 * The order of allocation has significant impact on performance and other
3226 * system components. Generally order 0 allocations should be preferred since
3227 * order 0 does not cause fragmentation in the page allocator. Larger objects
3228 * be problematic to put into order 0 slabs because there may be too much
3229 * unused space left. We go to a higher order if more than 1/16th of the slab
3232 * In order to reach satisfactory performance we must ensure that a minimum
3233 * number of objects is in one slab. Otherwise we may generate too much
3234 * activity on the partial lists which requires taking the list_lock. This is
3235 * less a concern for large slabs though which are rarely used.
3237 * slub_max_order specifies the order where we begin to stop considering the
3238 * number of objects in a slab as critical. If we reach slub_max_order then
3239 * we try to keep the page order as low as possible. So we accept more waste
3240 * of space in favor of a small page order.
3242 * Higher order allocations also allow the placement of more objects in a
3243 * slab and thereby reduce object handling overhead. If the user has
3244 * requested a higher mininum order then we start with that one instead of
3245 * the smallest order which will fit the object.
3247 static inline unsigned int slab_order(unsigned int size
,
3248 unsigned int min_objects
, unsigned int max_order
,
3249 unsigned int fract_leftover
)
3251 unsigned int min_order
= slub_min_order
;
3254 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3255 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3257 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3258 order
<= max_order
; order
++) {
3260 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3263 rem
= slab_size
% size
;
3265 if (rem
<= slab_size
/ fract_leftover
)
3272 static inline int calculate_order(unsigned int size
)
3275 unsigned int min_objects
;
3276 unsigned int max_objects
;
3279 * Attempt to find best configuration for a slab. This
3280 * works by first attempting to generate a layout with
3281 * the best configuration and backing off gradually.
3283 * First we increase the acceptable waste in a slab. Then
3284 * we reduce the minimum objects required in a slab.
3286 min_objects
= slub_min_objects
;
3288 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3289 max_objects
= order_objects(slub_max_order
, size
);
3290 min_objects
= min(min_objects
, max_objects
);
3292 while (min_objects
> 1) {
3293 unsigned int fraction
;
3296 while (fraction
>= 4) {
3297 order
= slab_order(size
, min_objects
,
3298 slub_max_order
, fraction
);
3299 if (order
<= slub_max_order
)
3307 * We were unable to place multiple objects in a slab. Now
3308 * lets see if we can place a single object there.
3310 order
= slab_order(size
, 1, slub_max_order
, 1);
3311 if (order
<= slub_max_order
)
3315 * Doh this slab cannot be placed using slub_max_order.
3317 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3318 if (order
< MAX_ORDER
)
3324 init_kmem_cache_node(struct kmem_cache_node
*n
)
3327 spin_lock_init(&n
->list_lock
);
3328 INIT_LIST_HEAD(&n
->partial
);
3329 #ifdef CONFIG_SLUB_DEBUG
3330 atomic_long_set(&n
->nr_slabs
, 0);
3331 atomic_long_set(&n
->total_objects
, 0);
3332 INIT_LIST_HEAD(&n
->full
);
3336 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3338 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3339 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3342 * Must align to double word boundary for the double cmpxchg
3343 * instructions to work; see __pcpu_double_call_return_bool().
3345 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3346 2 * sizeof(void *));
3351 init_kmem_cache_cpus(s
);
3356 static struct kmem_cache
*kmem_cache_node
;
3359 * No kmalloc_node yet so do it by hand. We know that this is the first
3360 * slab on the node for this slabcache. There are no concurrent accesses
3363 * Note that this function only works on the kmem_cache_node
3364 * when allocating for the kmem_cache_node. This is used for bootstrapping
3365 * memory on a fresh node that has no slab structures yet.
3367 static void early_kmem_cache_node_alloc(int node
)
3370 struct kmem_cache_node
*n
;
3372 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3374 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3377 if (page_to_nid(page
) != node
) {
3378 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3379 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3384 #ifdef CONFIG_SLUB_DEBUG
3385 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3386 init_tracking(kmem_cache_node
, n
);
3388 n
= kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3390 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3393 kmem_cache_node
->node
[node
] = n
;
3394 init_kmem_cache_node(n
);
3395 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3398 * No locks need to be taken here as it has just been
3399 * initialized and there is no concurrent access.
3401 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3404 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3407 struct kmem_cache_node
*n
;
3409 for_each_kmem_cache_node(s
, node
, n
) {
3410 s
->node
[node
] = NULL
;
3411 kmem_cache_free(kmem_cache_node
, n
);
3415 void __kmem_cache_release(struct kmem_cache
*s
)
3417 cache_random_seq_destroy(s
);
3418 free_percpu(s
->cpu_slab
);
3419 free_kmem_cache_nodes(s
);
3422 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3426 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3427 struct kmem_cache_node
*n
;
3429 if (slab_state
== DOWN
) {
3430 early_kmem_cache_node_alloc(node
);
3433 n
= kmem_cache_alloc_node(kmem_cache_node
,
3437 free_kmem_cache_nodes(s
);
3441 init_kmem_cache_node(n
);
3447 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3449 if (min
< MIN_PARTIAL
)
3451 else if (min
> MAX_PARTIAL
)
3453 s
->min_partial
= min
;
3456 static void set_cpu_partial(struct kmem_cache
*s
)
3458 #ifdef CONFIG_SLUB_CPU_PARTIAL
3460 * cpu_partial determined the maximum number of objects kept in the
3461 * per cpu partial lists of a processor.
3463 * Per cpu partial lists mainly contain slabs that just have one
3464 * object freed. If they are used for allocation then they can be
3465 * filled up again with minimal effort. The slab will never hit the
3466 * per node partial lists and therefore no locking will be required.
3468 * This setting also determines
3470 * A) The number of objects from per cpu partial slabs dumped to the
3471 * per node list when we reach the limit.
3472 * B) The number of objects in cpu partial slabs to extract from the
3473 * per node list when we run out of per cpu objects. We only fetch
3474 * 50% to keep some capacity around for frees.
3476 if (!kmem_cache_has_cpu_partial(s
))
3478 else if (s
->size
>= PAGE_SIZE
)
3480 else if (s
->size
>= 1024)
3482 else if (s
->size
>= 256)
3483 s
->cpu_partial
= 13;
3485 s
->cpu_partial
= 30;
3490 * calculate_sizes() determines the order and the distribution of data within
3493 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3495 slab_flags_t flags
= s
->flags
;
3496 unsigned int size
= s
->object_size
;
3500 * Round up object size to the next word boundary. We can only
3501 * place the free pointer at word boundaries and this determines
3502 * the possible location of the free pointer.
3504 size
= ALIGN(size
, sizeof(void *));
3506 #ifdef CONFIG_SLUB_DEBUG
3508 * Determine if we can poison the object itself. If the user of
3509 * the slab may touch the object after free or before allocation
3510 * then we should never poison the object itself.
3512 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3514 s
->flags
|= __OBJECT_POISON
;
3516 s
->flags
&= ~__OBJECT_POISON
;
3520 * If we are Redzoning then check if there is some space between the
3521 * end of the object and the free pointer. If not then add an
3522 * additional word to have some bytes to store Redzone information.
3524 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3525 size
+= sizeof(void *);
3529 * With that we have determined the number of bytes in actual use
3530 * by the object. This is the potential offset to the free pointer.
3534 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3537 * Relocate free pointer after the object if it is not
3538 * permitted to overwrite the first word of the object on
3541 * This is the case if we do RCU, have a constructor or
3542 * destructor or are poisoning the objects.
3545 size
+= sizeof(void *);
3548 #ifdef CONFIG_SLUB_DEBUG
3549 if (flags
& SLAB_STORE_USER
)
3551 * Need to store information about allocs and frees after
3554 size
+= 2 * sizeof(struct track
);
3557 kasan_cache_create(s
, &size
, &s
->flags
);
3558 #ifdef CONFIG_SLUB_DEBUG
3559 if (flags
& SLAB_RED_ZONE
) {
3561 * Add some empty padding so that we can catch
3562 * overwrites from earlier objects rather than let
3563 * tracking information or the free pointer be
3564 * corrupted if a user writes before the start
3567 size
+= sizeof(void *);
3569 s
->red_left_pad
= sizeof(void *);
3570 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3571 size
+= s
->red_left_pad
;
3576 * SLUB stores one object immediately after another beginning from
3577 * offset 0. In order to align the objects we have to simply size
3578 * each object to conform to the alignment.
3580 size
= ALIGN(size
, s
->align
);
3582 if (forced_order
>= 0)
3583 order
= forced_order
;
3585 order
= calculate_order(size
);
3592 s
->allocflags
|= __GFP_COMP
;
3594 if (s
->flags
& SLAB_CACHE_DMA
)
3595 s
->allocflags
|= GFP_DMA
;
3597 if (s
->flags
& SLAB_CACHE_DMA32
)
3598 s
->allocflags
|= GFP_DMA32
;
3600 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3601 s
->allocflags
|= __GFP_RECLAIMABLE
;
3604 * Determine the number of objects per slab
3606 s
->oo
= oo_make(order
, size
);
3607 s
->min
= oo_make(get_order(size
), size
);
3608 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3611 return !!oo_objects(s
->oo
);
3614 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3616 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3617 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3618 s
->random
= get_random_long();
3621 if (!calculate_sizes(s
, -1))
3623 if (disable_higher_order_debug
) {
3625 * Disable debugging flags that store metadata if the min slab
3628 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3629 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3631 if (!calculate_sizes(s
, -1))
3636 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3637 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3638 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3639 /* Enable fast mode */
3640 s
->flags
|= __CMPXCHG_DOUBLE
;
3644 * The larger the object size is, the more pages we want on the partial
3645 * list to avoid pounding the page allocator excessively.
3647 set_min_partial(s
, ilog2(s
->size
) / 2);
3652 s
->remote_node_defrag_ratio
= 1000;
3655 /* Initialize the pre-computed randomized freelist if slab is up */
3656 if (slab_state
>= UP
) {
3657 if (init_cache_random_seq(s
))
3661 if (!init_kmem_cache_nodes(s
))
3664 if (alloc_kmem_cache_cpus(s
))
3667 free_kmem_cache_nodes(s
);
3672 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3675 #ifdef CONFIG_SLUB_DEBUG
3676 void *addr
= page_address(page
);
3678 unsigned long *map
= bitmap_zalloc(page
->objects
, GFP_ATOMIC
);
3681 slab_err(s
, page
, text
, s
->name
);
3684 get_map(s
, page
, map
);
3685 for_each_object(p
, s
, addr
, page
->objects
) {
3687 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3688 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3689 print_tracking(s
, p
);
3698 * Attempt to free all partial slabs on a node.
3699 * This is called from __kmem_cache_shutdown(). We must take list_lock
3700 * because sysfs file might still access partial list after the shutdowning.
3702 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3705 struct page
*page
, *h
;
3707 BUG_ON(irqs_disabled());
3708 spin_lock_irq(&n
->list_lock
);
3709 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3711 remove_partial(n
, page
);
3712 list_add(&page
->slab_list
, &discard
);
3714 list_slab_objects(s
, page
,
3715 "Objects remaining in %s on __kmem_cache_shutdown()");
3718 spin_unlock_irq(&n
->list_lock
);
3720 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3721 discard_slab(s
, page
);
3724 bool __kmem_cache_empty(struct kmem_cache
*s
)
3727 struct kmem_cache_node
*n
;
3729 for_each_kmem_cache_node(s
, node
, n
)
3730 if (n
->nr_partial
|| slabs_node(s
, node
))
3736 * Release all resources used by a slab cache.
3738 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3741 struct kmem_cache_node
*n
;
3744 /* Attempt to free all objects */
3745 for_each_kmem_cache_node(s
, node
, n
) {
3747 if (n
->nr_partial
|| slabs_node(s
, node
))
3750 sysfs_slab_remove(s
);
3754 /********************************************************************
3756 *******************************************************************/
3758 static int __init
setup_slub_min_order(char *str
)
3760 get_option(&str
, (int *)&slub_min_order
);
3765 __setup("slub_min_order=", setup_slub_min_order
);
3767 static int __init
setup_slub_max_order(char *str
)
3769 get_option(&str
, (int *)&slub_max_order
);
3770 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
3775 __setup("slub_max_order=", setup_slub_max_order
);
3777 static int __init
setup_slub_min_objects(char *str
)
3779 get_option(&str
, (int *)&slub_min_objects
);
3784 __setup("slub_min_objects=", setup_slub_min_objects
);
3786 void *__kmalloc(size_t size
, gfp_t flags
)
3788 struct kmem_cache
*s
;
3791 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3792 return kmalloc_large(size
, flags
);
3794 s
= kmalloc_slab(size
, flags
);
3796 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3799 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3801 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3803 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3807 EXPORT_SYMBOL(__kmalloc
);
3810 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3814 unsigned int order
= get_order(size
);
3816 flags
|= __GFP_COMP
;
3817 page
= alloc_pages_node(node
, flags
, order
);
3819 ptr
= page_address(page
);
3820 mod_node_page_state(page_pgdat(page
), NR_SLAB_UNRECLAIMABLE
,
3824 return kmalloc_large_node_hook(ptr
, size
, flags
);
3827 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3829 struct kmem_cache
*s
;
3832 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3833 ret
= kmalloc_large_node(size
, flags
, node
);
3835 trace_kmalloc_node(_RET_IP_
, ret
,
3836 size
, PAGE_SIZE
<< get_order(size
),
3842 s
= kmalloc_slab(size
, flags
);
3844 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3847 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3849 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3851 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3855 EXPORT_SYMBOL(__kmalloc_node
);
3856 #endif /* CONFIG_NUMA */
3858 #ifdef CONFIG_HARDENED_USERCOPY
3860 * Rejects incorrectly sized objects and objects that are to be copied
3861 * to/from userspace but do not fall entirely within the containing slab
3862 * cache's usercopy region.
3864 * Returns NULL if check passes, otherwise const char * to name of cache
3865 * to indicate an error.
3867 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
3870 struct kmem_cache
*s
;
3871 unsigned int offset
;
3874 ptr
= kasan_reset_tag(ptr
);
3876 /* Find object and usable object size. */
3877 s
= page
->slab_cache
;
3879 /* Reject impossible pointers. */
3880 if (ptr
< page_address(page
))
3881 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
3884 /* Find offset within object. */
3885 offset
= (ptr
- page_address(page
)) % s
->size
;
3887 /* Adjust for redzone and reject if within the redzone. */
3888 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3889 if (offset
< s
->red_left_pad
)
3890 usercopy_abort("SLUB object in left red zone",
3891 s
->name
, to_user
, offset
, n
);
3892 offset
-= s
->red_left_pad
;
3895 /* Allow address range falling entirely within usercopy region. */
3896 if (offset
>= s
->useroffset
&&
3897 offset
- s
->useroffset
<= s
->usersize
&&
3898 n
<= s
->useroffset
- offset
+ s
->usersize
)
3902 * If the copy is still within the allocated object, produce
3903 * a warning instead of rejecting the copy. This is intended
3904 * to be a temporary method to find any missing usercopy
3907 object_size
= slab_ksize(s
);
3908 if (usercopy_fallback
&&
3909 offset
<= object_size
&& n
<= object_size
- offset
) {
3910 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
3914 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
3916 #endif /* CONFIG_HARDENED_USERCOPY */
3918 size_t __ksize(const void *object
)
3922 if (unlikely(object
== ZERO_SIZE_PTR
))
3925 page
= virt_to_head_page(object
);
3927 if (unlikely(!PageSlab(page
))) {
3928 WARN_ON(!PageCompound(page
));
3929 return page_size(page
);
3932 return slab_ksize(page
->slab_cache
);
3934 EXPORT_SYMBOL(__ksize
);
3936 void kfree(const void *x
)
3939 void *object
= (void *)x
;
3941 trace_kfree(_RET_IP_
, x
);
3943 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3946 page
= virt_to_head_page(x
);
3947 if (unlikely(!PageSlab(page
))) {
3948 unsigned int order
= compound_order(page
);
3950 BUG_ON(!PageCompound(page
));
3952 mod_node_page_state(page_pgdat(page
), NR_SLAB_UNRECLAIMABLE
,
3954 __free_pages(page
, order
);
3957 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3959 EXPORT_SYMBOL(kfree
);
3961 #define SHRINK_PROMOTE_MAX 32
3964 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3965 * up most to the head of the partial lists. New allocations will then
3966 * fill those up and thus they can be removed from the partial lists.
3968 * The slabs with the least items are placed last. This results in them
3969 * being allocated from last increasing the chance that the last objects
3970 * are freed in them.
3972 int __kmem_cache_shrink(struct kmem_cache
*s
)
3976 struct kmem_cache_node
*n
;
3979 struct list_head discard
;
3980 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3981 unsigned long flags
;
3985 for_each_kmem_cache_node(s
, node
, n
) {
3986 INIT_LIST_HEAD(&discard
);
3987 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3988 INIT_LIST_HEAD(promote
+ i
);
3990 spin_lock_irqsave(&n
->list_lock
, flags
);
3993 * Build lists of slabs to discard or promote.
3995 * Note that concurrent frees may occur while we hold the
3996 * list_lock. page->inuse here is the upper limit.
3998 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
3999 int free
= page
->objects
- page
->inuse
;
4001 /* Do not reread page->inuse */
4004 /* We do not keep full slabs on the list */
4007 if (free
== page
->objects
) {
4008 list_move(&page
->slab_list
, &discard
);
4010 } else if (free
<= SHRINK_PROMOTE_MAX
)
4011 list_move(&page
->slab_list
, promote
+ free
- 1);
4015 * Promote the slabs filled up most to the head of the
4018 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4019 list_splice(promote
+ i
, &n
->partial
);
4021 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4023 /* Release empty slabs */
4024 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4025 discard_slab(s
, page
);
4027 if (slabs_node(s
, node
))
4035 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache
*s
)
4038 * Called with all the locks held after a sched RCU grace period.
4039 * Even if @s becomes empty after shrinking, we can't know that @s
4040 * doesn't have allocations already in-flight and thus can't
4041 * destroy @s until the associated memcg is released.
4043 * However, let's remove the sysfs files for empty caches here.
4044 * Each cache has a lot of interface files which aren't
4045 * particularly useful for empty draining caches; otherwise, we can
4046 * easily end up with millions of unnecessary sysfs files on
4047 * systems which have a lot of memory and transient cgroups.
4049 if (!__kmem_cache_shrink(s
))
4050 sysfs_slab_remove(s
);
4053 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4056 * Disable empty slabs caching. Used to avoid pinning offline
4057 * memory cgroups by kmem pages that can be freed.
4059 slub_set_cpu_partial(s
, 0);
4062 #endif /* CONFIG_MEMCG */
4064 static int slab_mem_going_offline_callback(void *arg
)
4066 struct kmem_cache
*s
;
4068 mutex_lock(&slab_mutex
);
4069 list_for_each_entry(s
, &slab_caches
, list
)
4070 __kmem_cache_shrink(s
);
4071 mutex_unlock(&slab_mutex
);
4076 static void slab_mem_offline_callback(void *arg
)
4078 struct kmem_cache_node
*n
;
4079 struct kmem_cache
*s
;
4080 struct memory_notify
*marg
= arg
;
4083 offline_node
= marg
->status_change_nid_normal
;
4086 * If the node still has available memory. we need kmem_cache_node
4089 if (offline_node
< 0)
4092 mutex_lock(&slab_mutex
);
4093 list_for_each_entry(s
, &slab_caches
, list
) {
4094 n
= get_node(s
, offline_node
);
4097 * if n->nr_slabs > 0, slabs still exist on the node
4098 * that is going down. We were unable to free them,
4099 * and offline_pages() function shouldn't call this
4100 * callback. So, we must fail.
4102 BUG_ON(slabs_node(s
, offline_node
));
4104 s
->node
[offline_node
] = NULL
;
4105 kmem_cache_free(kmem_cache_node
, n
);
4108 mutex_unlock(&slab_mutex
);
4111 static int slab_mem_going_online_callback(void *arg
)
4113 struct kmem_cache_node
*n
;
4114 struct kmem_cache
*s
;
4115 struct memory_notify
*marg
= arg
;
4116 int nid
= marg
->status_change_nid_normal
;
4120 * If the node's memory is already available, then kmem_cache_node is
4121 * already created. Nothing to do.
4127 * We are bringing a node online. No memory is available yet. We must
4128 * allocate a kmem_cache_node structure in order to bring the node
4131 mutex_lock(&slab_mutex
);
4132 list_for_each_entry(s
, &slab_caches
, list
) {
4134 * XXX: kmem_cache_alloc_node will fallback to other nodes
4135 * since memory is not yet available from the node that
4138 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4143 init_kmem_cache_node(n
);
4147 mutex_unlock(&slab_mutex
);
4151 static int slab_memory_callback(struct notifier_block
*self
,
4152 unsigned long action
, void *arg
)
4157 case MEM_GOING_ONLINE
:
4158 ret
= slab_mem_going_online_callback(arg
);
4160 case MEM_GOING_OFFLINE
:
4161 ret
= slab_mem_going_offline_callback(arg
);
4164 case MEM_CANCEL_ONLINE
:
4165 slab_mem_offline_callback(arg
);
4168 case MEM_CANCEL_OFFLINE
:
4172 ret
= notifier_from_errno(ret
);
4178 static struct notifier_block slab_memory_callback_nb
= {
4179 .notifier_call
= slab_memory_callback
,
4180 .priority
= SLAB_CALLBACK_PRI
,
4183 /********************************************************************
4184 * Basic setup of slabs
4185 *******************************************************************/
4188 * Used for early kmem_cache structures that were allocated using
4189 * the page allocator. Allocate them properly then fix up the pointers
4190 * that may be pointing to the wrong kmem_cache structure.
4193 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4196 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4197 struct kmem_cache_node
*n
;
4199 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4202 * This runs very early, and only the boot processor is supposed to be
4203 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4206 __flush_cpu_slab(s
, smp_processor_id());
4207 for_each_kmem_cache_node(s
, node
, n
) {
4210 list_for_each_entry(p
, &n
->partial
, slab_list
)
4213 #ifdef CONFIG_SLUB_DEBUG
4214 list_for_each_entry(p
, &n
->full
, slab_list
)
4218 slab_init_memcg_params(s
);
4219 list_add(&s
->list
, &slab_caches
);
4220 memcg_link_cache(s
, NULL
);
4224 void __init
kmem_cache_init(void)
4226 static __initdata
struct kmem_cache boot_kmem_cache
,
4227 boot_kmem_cache_node
;
4229 if (debug_guardpage_minorder())
4232 kmem_cache_node
= &boot_kmem_cache_node
;
4233 kmem_cache
= &boot_kmem_cache
;
4235 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4236 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4238 register_hotmemory_notifier(&slab_memory_callback_nb
);
4240 /* Able to allocate the per node structures */
4241 slab_state
= PARTIAL
;
4243 create_boot_cache(kmem_cache
, "kmem_cache",
4244 offsetof(struct kmem_cache
, node
) +
4245 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4246 SLAB_HWCACHE_ALIGN
, 0, 0);
4248 kmem_cache
= bootstrap(&boot_kmem_cache
);
4249 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4251 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4252 setup_kmalloc_cache_index_table();
4253 create_kmalloc_caches(0);
4255 /* Setup random freelists for each cache */
4256 init_freelist_randomization();
4258 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4261 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4263 slub_min_order
, slub_max_order
, slub_min_objects
,
4264 nr_cpu_ids
, nr_node_ids
);
4267 void __init
kmem_cache_init_late(void)
4272 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4273 slab_flags_t flags
, void (*ctor
)(void *))
4275 struct kmem_cache
*s
, *c
;
4277 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4282 * Adjust the object sizes so that we clear
4283 * the complete object on kzalloc.
4285 s
->object_size
= max(s
->object_size
, size
);
4286 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4288 for_each_memcg_cache(c
, s
) {
4289 c
->object_size
= s
->object_size
;
4290 c
->inuse
= max(c
->inuse
, ALIGN(size
, sizeof(void *)));
4293 if (sysfs_slab_alias(s
, name
)) {
4302 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4306 err
= kmem_cache_open(s
, flags
);
4310 /* Mutex is not taken during early boot */
4311 if (slab_state
<= UP
)
4314 memcg_propagate_slab_attrs(s
);
4315 err
= sysfs_slab_add(s
);
4317 __kmem_cache_release(s
);
4322 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4324 struct kmem_cache
*s
;
4327 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4328 return kmalloc_large(size
, gfpflags
);
4330 s
= kmalloc_slab(size
, gfpflags
);
4332 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4335 ret
= slab_alloc(s
, gfpflags
, caller
);
4337 /* Honor the call site pointer we received. */
4338 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4344 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4345 int node
, unsigned long caller
)
4347 struct kmem_cache
*s
;
4350 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4351 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4353 trace_kmalloc_node(caller
, ret
,
4354 size
, PAGE_SIZE
<< get_order(size
),
4360 s
= kmalloc_slab(size
, gfpflags
);
4362 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4365 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4367 /* Honor the call site pointer we received. */
4368 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4375 static int count_inuse(struct page
*page
)
4380 static int count_total(struct page
*page
)
4382 return page
->objects
;
4386 #ifdef CONFIG_SLUB_DEBUG
4387 static void validate_slab(struct kmem_cache
*s
, struct page
*page
,
4391 void *addr
= page_address(page
);
4393 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4396 /* Now we know that a valid freelist exists */
4397 bitmap_zero(map
, page
->objects
);
4399 get_map(s
, page
, map
);
4400 for_each_object(p
, s
, addr
, page
->objects
) {
4401 u8 val
= test_bit(slab_index(p
, s
, addr
), map
) ?
4402 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4404 if (!check_object(s
, page
, p
, val
))
4409 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4413 validate_slab(s
, page
, map
);
4417 static int validate_slab_node(struct kmem_cache
*s
,
4418 struct kmem_cache_node
*n
, unsigned long *map
)
4420 unsigned long count
= 0;
4422 unsigned long flags
;
4424 spin_lock_irqsave(&n
->list_lock
, flags
);
4426 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4427 validate_slab_slab(s
, page
, map
);
4430 if (count
!= n
->nr_partial
)
4431 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4432 s
->name
, count
, n
->nr_partial
);
4434 if (!(s
->flags
& SLAB_STORE_USER
))
4437 list_for_each_entry(page
, &n
->full
, slab_list
) {
4438 validate_slab_slab(s
, page
, map
);
4441 if (count
!= atomic_long_read(&n
->nr_slabs
))
4442 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4443 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4446 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4450 static long validate_slab_cache(struct kmem_cache
*s
)
4453 unsigned long count
= 0;
4454 struct kmem_cache_node
*n
;
4455 unsigned long *map
= bitmap_alloc(oo_objects(s
->max
), GFP_KERNEL
);
4461 for_each_kmem_cache_node(s
, node
, n
)
4462 count
+= validate_slab_node(s
, n
, map
);
4467 * Generate lists of code addresses where slabcache objects are allocated
4472 unsigned long count
;
4479 DECLARE_BITMAP(cpus
, NR_CPUS
);
4485 unsigned long count
;
4486 struct location
*loc
;
4489 static void free_loc_track(struct loc_track
*t
)
4492 free_pages((unsigned long)t
->loc
,
4493 get_order(sizeof(struct location
) * t
->max
));
4496 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4501 order
= get_order(sizeof(struct location
) * max
);
4503 l
= (void *)__get_free_pages(flags
, order
);
4508 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4516 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4517 const struct track
*track
)
4519 long start
, end
, pos
;
4521 unsigned long caddr
;
4522 unsigned long age
= jiffies
- track
->when
;
4528 pos
= start
+ (end
- start
+ 1) / 2;
4531 * There is nothing at "end". If we end up there
4532 * we need to add something to before end.
4537 caddr
= t
->loc
[pos
].addr
;
4538 if (track
->addr
== caddr
) {
4544 if (age
< l
->min_time
)
4546 if (age
> l
->max_time
)
4549 if (track
->pid
< l
->min_pid
)
4550 l
->min_pid
= track
->pid
;
4551 if (track
->pid
> l
->max_pid
)
4552 l
->max_pid
= track
->pid
;
4554 cpumask_set_cpu(track
->cpu
,
4555 to_cpumask(l
->cpus
));
4557 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4561 if (track
->addr
< caddr
)
4568 * Not found. Insert new tracking element.
4570 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4576 (t
->count
- pos
) * sizeof(struct location
));
4579 l
->addr
= track
->addr
;
4583 l
->min_pid
= track
->pid
;
4584 l
->max_pid
= track
->pid
;
4585 cpumask_clear(to_cpumask(l
->cpus
));
4586 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4587 nodes_clear(l
->nodes
);
4588 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4592 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4593 struct page
*page
, enum track_item alloc
,
4596 void *addr
= page_address(page
);
4599 bitmap_zero(map
, page
->objects
);
4600 get_map(s
, page
, map
);
4602 for_each_object(p
, s
, addr
, page
->objects
)
4603 if (!test_bit(slab_index(p
, s
, addr
), map
))
4604 add_location(t
, s
, get_track(s
, p
, alloc
));
4607 static int list_locations(struct kmem_cache
*s
, char *buf
,
4608 enum track_item alloc
)
4612 struct loc_track t
= { 0, 0, NULL
};
4614 struct kmem_cache_node
*n
;
4615 unsigned long *map
= bitmap_alloc(oo_objects(s
->max
), GFP_KERNEL
);
4617 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4620 return sprintf(buf
, "Out of memory\n");
4622 /* Push back cpu slabs */
4625 for_each_kmem_cache_node(s
, node
, n
) {
4626 unsigned long flags
;
4629 if (!atomic_long_read(&n
->nr_slabs
))
4632 spin_lock_irqsave(&n
->list_lock
, flags
);
4633 list_for_each_entry(page
, &n
->partial
, slab_list
)
4634 process_slab(&t
, s
, page
, alloc
, map
);
4635 list_for_each_entry(page
, &n
->full
, slab_list
)
4636 process_slab(&t
, s
, page
, alloc
, map
);
4637 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4640 for (i
= 0; i
< t
.count
; i
++) {
4641 struct location
*l
= &t
.loc
[i
];
4643 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4645 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4648 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4650 len
+= sprintf(buf
+ len
, "<not-available>");
4652 if (l
->sum_time
!= l
->min_time
) {
4653 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4655 (long)div_u64(l
->sum_time
, l
->count
),
4658 len
+= sprintf(buf
+ len
, " age=%ld",
4661 if (l
->min_pid
!= l
->max_pid
)
4662 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4663 l
->min_pid
, l
->max_pid
);
4665 len
+= sprintf(buf
+ len
, " pid=%ld",
4668 if (num_online_cpus() > 1 &&
4669 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4670 len
< PAGE_SIZE
- 60)
4671 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4673 cpumask_pr_args(to_cpumask(l
->cpus
)));
4675 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4676 len
< PAGE_SIZE
- 60)
4677 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4679 nodemask_pr_args(&l
->nodes
));
4681 len
+= sprintf(buf
+ len
, "\n");
4687 len
+= sprintf(buf
, "No data\n");
4690 #endif /* CONFIG_SLUB_DEBUG */
4692 #ifdef SLUB_RESILIENCY_TEST
4693 static void __init
resiliency_test(void)
4696 int type
= KMALLOC_NORMAL
;
4698 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4700 pr_err("SLUB resiliency testing\n");
4701 pr_err("-----------------------\n");
4702 pr_err("A. Corruption after allocation\n");
4704 p
= kzalloc(16, GFP_KERNEL
);
4706 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4709 validate_slab_cache(kmalloc_caches
[type
][4]);
4711 /* Hmmm... The next two are dangerous */
4712 p
= kzalloc(32, GFP_KERNEL
);
4713 p
[32 + sizeof(void *)] = 0x34;
4714 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4716 pr_err("If allocated object is overwritten then not detectable\n\n");
4718 validate_slab_cache(kmalloc_caches
[type
][5]);
4719 p
= kzalloc(64, GFP_KERNEL
);
4720 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4722 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4724 pr_err("If allocated object is overwritten then not detectable\n\n");
4725 validate_slab_cache(kmalloc_caches
[type
][6]);
4727 pr_err("\nB. Corruption after free\n");
4728 p
= kzalloc(128, GFP_KERNEL
);
4731 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4732 validate_slab_cache(kmalloc_caches
[type
][7]);
4734 p
= kzalloc(256, GFP_KERNEL
);
4737 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4738 validate_slab_cache(kmalloc_caches
[type
][8]);
4740 p
= kzalloc(512, GFP_KERNEL
);
4743 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4744 validate_slab_cache(kmalloc_caches
[type
][9]);
4748 static void resiliency_test(void) {};
4750 #endif /* SLUB_RESILIENCY_TEST */
4753 enum slab_stat_type
{
4754 SL_ALL
, /* All slabs */
4755 SL_PARTIAL
, /* Only partially allocated slabs */
4756 SL_CPU
, /* Only slabs used for cpu caches */
4757 SL_OBJECTS
, /* Determine allocated objects not slabs */
4758 SL_TOTAL
/* Determine object capacity not slabs */
4761 #define SO_ALL (1 << SL_ALL)
4762 #define SO_PARTIAL (1 << SL_PARTIAL)
4763 #define SO_CPU (1 << SL_CPU)
4764 #define SO_OBJECTS (1 << SL_OBJECTS)
4765 #define SO_TOTAL (1 << SL_TOTAL)
4768 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4770 static int __init
setup_slub_memcg_sysfs(char *str
)
4774 if (get_option(&str
, &v
) > 0)
4775 memcg_sysfs_enabled
= v
;
4780 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4783 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4784 char *buf
, unsigned long flags
)
4786 unsigned long total
= 0;
4789 unsigned long *nodes
;
4791 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4795 if (flags
& SO_CPU
) {
4798 for_each_possible_cpu(cpu
) {
4799 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4804 page
= READ_ONCE(c
->page
);
4808 node
= page_to_nid(page
);
4809 if (flags
& SO_TOTAL
)
4811 else if (flags
& SO_OBJECTS
)
4819 page
= slub_percpu_partial_read_once(c
);
4821 node
= page_to_nid(page
);
4822 if (flags
& SO_TOTAL
)
4824 else if (flags
& SO_OBJECTS
)
4835 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4836 * already held which will conflict with an existing lock order:
4838 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4840 * We don't really need mem_hotplug_lock (to hold off
4841 * slab_mem_going_offline_callback) here because slab's memory hot
4842 * unplug code doesn't destroy the kmem_cache->node[] data.
4845 #ifdef CONFIG_SLUB_DEBUG
4846 if (flags
& SO_ALL
) {
4847 struct kmem_cache_node
*n
;
4849 for_each_kmem_cache_node(s
, node
, n
) {
4851 if (flags
& SO_TOTAL
)
4852 x
= atomic_long_read(&n
->total_objects
);
4853 else if (flags
& SO_OBJECTS
)
4854 x
= atomic_long_read(&n
->total_objects
) -
4855 count_partial(n
, count_free
);
4857 x
= atomic_long_read(&n
->nr_slabs
);
4864 if (flags
& SO_PARTIAL
) {
4865 struct kmem_cache_node
*n
;
4867 for_each_kmem_cache_node(s
, node
, n
) {
4868 if (flags
& SO_TOTAL
)
4869 x
= count_partial(n
, count_total
);
4870 else if (flags
& SO_OBJECTS
)
4871 x
= count_partial(n
, count_inuse
);
4878 x
= sprintf(buf
, "%lu", total
);
4880 for (node
= 0; node
< nr_node_ids
; node
++)
4882 x
+= sprintf(buf
+ x
, " N%d=%lu",
4886 return x
+ sprintf(buf
+ x
, "\n");
4889 #ifdef CONFIG_SLUB_DEBUG
4890 static int any_slab_objects(struct kmem_cache
*s
)
4893 struct kmem_cache_node
*n
;
4895 for_each_kmem_cache_node(s
, node
, n
)
4896 if (atomic_long_read(&n
->total_objects
))
4903 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4904 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4906 struct slab_attribute
{
4907 struct attribute attr
;
4908 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4909 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4912 #define SLAB_ATTR_RO(_name) \
4913 static struct slab_attribute _name##_attr = \
4914 __ATTR(_name, 0400, _name##_show, NULL)
4916 #define SLAB_ATTR(_name) \
4917 static struct slab_attribute _name##_attr = \
4918 __ATTR(_name, 0600, _name##_show, _name##_store)
4920 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4922 return sprintf(buf
, "%u\n", s
->size
);
4924 SLAB_ATTR_RO(slab_size
);
4926 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4928 return sprintf(buf
, "%u\n", s
->align
);
4930 SLAB_ATTR_RO(align
);
4932 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4934 return sprintf(buf
, "%u\n", s
->object_size
);
4936 SLAB_ATTR_RO(object_size
);
4938 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4940 return sprintf(buf
, "%u\n", oo_objects(s
->oo
));
4942 SLAB_ATTR_RO(objs_per_slab
);
4944 static ssize_t
order_store(struct kmem_cache
*s
,
4945 const char *buf
, size_t length
)
4950 err
= kstrtouint(buf
, 10, &order
);
4954 if (order
> slub_max_order
|| order
< slub_min_order
)
4957 calculate_sizes(s
, order
);
4961 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4963 return sprintf(buf
, "%u\n", oo_order(s
->oo
));
4967 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4969 return sprintf(buf
, "%lu\n", s
->min_partial
);
4972 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4978 err
= kstrtoul(buf
, 10, &min
);
4982 set_min_partial(s
, min
);
4985 SLAB_ATTR(min_partial
);
4987 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4989 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4992 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4995 unsigned int objects
;
4998 err
= kstrtouint(buf
, 10, &objects
);
5001 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5004 slub_set_cpu_partial(s
, objects
);
5008 SLAB_ATTR(cpu_partial
);
5010 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5014 return sprintf(buf
, "%pS\n", s
->ctor
);
5018 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5020 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5022 SLAB_ATTR_RO(aliases
);
5024 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5026 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5028 SLAB_ATTR_RO(partial
);
5030 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5032 return show_slab_objects(s
, buf
, SO_CPU
);
5034 SLAB_ATTR_RO(cpu_slabs
);
5036 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5038 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5040 SLAB_ATTR_RO(objects
);
5042 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5044 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5046 SLAB_ATTR_RO(objects_partial
);
5048 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5055 for_each_online_cpu(cpu
) {
5058 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5061 pages
+= page
->pages
;
5062 objects
+= page
->pobjects
;
5066 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5069 for_each_online_cpu(cpu
) {
5072 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5074 if (page
&& len
< PAGE_SIZE
- 20)
5075 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5076 page
->pobjects
, page
->pages
);
5079 return len
+ sprintf(buf
+ len
, "\n");
5081 SLAB_ATTR_RO(slabs_cpu_partial
);
5083 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5085 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5088 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5089 const char *buf
, size_t length
)
5091 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5093 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5096 SLAB_ATTR(reclaim_account
);
5098 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5100 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5102 SLAB_ATTR_RO(hwcache_align
);
5104 #ifdef CONFIG_ZONE_DMA
5105 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5107 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5109 SLAB_ATTR_RO(cache_dma
);
5112 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5114 return sprintf(buf
, "%u\n", s
->usersize
);
5116 SLAB_ATTR_RO(usersize
);
5118 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5120 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5122 SLAB_ATTR_RO(destroy_by_rcu
);
5124 #ifdef CONFIG_SLUB_DEBUG
5125 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5127 return show_slab_objects(s
, buf
, SO_ALL
);
5129 SLAB_ATTR_RO(slabs
);
5131 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5133 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5135 SLAB_ATTR_RO(total_objects
);
5137 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5139 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5142 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5143 const char *buf
, size_t length
)
5145 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5146 if (buf
[0] == '1') {
5147 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5148 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5152 SLAB_ATTR(sanity_checks
);
5154 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5156 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5159 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5163 * Tracing a merged cache is going to give confusing results
5164 * as well as cause other issues like converting a mergeable
5165 * cache into an umergeable one.
5167 if (s
->refcount
> 1)
5170 s
->flags
&= ~SLAB_TRACE
;
5171 if (buf
[0] == '1') {
5172 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5173 s
->flags
|= SLAB_TRACE
;
5179 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5181 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5184 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5185 const char *buf
, size_t length
)
5187 if (any_slab_objects(s
))
5190 s
->flags
&= ~SLAB_RED_ZONE
;
5191 if (buf
[0] == '1') {
5192 s
->flags
|= SLAB_RED_ZONE
;
5194 calculate_sizes(s
, -1);
5197 SLAB_ATTR(red_zone
);
5199 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5201 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5204 static ssize_t
poison_store(struct kmem_cache
*s
,
5205 const char *buf
, size_t length
)
5207 if (any_slab_objects(s
))
5210 s
->flags
&= ~SLAB_POISON
;
5211 if (buf
[0] == '1') {
5212 s
->flags
|= SLAB_POISON
;
5214 calculate_sizes(s
, -1);
5219 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5221 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5224 static ssize_t
store_user_store(struct kmem_cache
*s
,
5225 const char *buf
, size_t length
)
5227 if (any_slab_objects(s
))
5230 s
->flags
&= ~SLAB_STORE_USER
;
5231 if (buf
[0] == '1') {
5232 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5233 s
->flags
|= SLAB_STORE_USER
;
5235 calculate_sizes(s
, -1);
5238 SLAB_ATTR(store_user
);
5240 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5245 static ssize_t
validate_store(struct kmem_cache
*s
,
5246 const char *buf
, size_t length
)
5250 if (buf
[0] == '1') {
5251 ret
= validate_slab_cache(s
);
5257 SLAB_ATTR(validate
);
5259 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5261 if (!(s
->flags
& SLAB_STORE_USER
))
5263 return list_locations(s
, buf
, TRACK_ALLOC
);
5265 SLAB_ATTR_RO(alloc_calls
);
5267 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5269 if (!(s
->flags
& SLAB_STORE_USER
))
5271 return list_locations(s
, buf
, TRACK_FREE
);
5273 SLAB_ATTR_RO(free_calls
);
5274 #endif /* CONFIG_SLUB_DEBUG */
5276 #ifdef CONFIG_FAILSLAB
5277 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5279 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5282 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5285 if (s
->refcount
> 1)
5288 s
->flags
&= ~SLAB_FAILSLAB
;
5290 s
->flags
|= SLAB_FAILSLAB
;
5293 SLAB_ATTR(failslab
);
5296 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5301 static ssize_t
shrink_store(struct kmem_cache
*s
,
5302 const char *buf
, size_t length
)
5305 kmem_cache_shrink_all(s
);
5313 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5315 return sprintf(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5318 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5319 const char *buf
, size_t length
)
5324 err
= kstrtouint(buf
, 10, &ratio
);
5330 s
->remote_node_defrag_ratio
= ratio
* 10;
5334 SLAB_ATTR(remote_node_defrag_ratio
);
5337 #ifdef CONFIG_SLUB_STATS
5338 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5340 unsigned long sum
= 0;
5343 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5348 for_each_online_cpu(cpu
) {
5349 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5355 len
= sprintf(buf
, "%lu", sum
);
5358 for_each_online_cpu(cpu
) {
5359 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5360 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5364 return len
+ sprintf(buf
+ len
, "\n");
5367 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5371 for_each_online_cpu(cpu
)
5372 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5375 #define STAT_ATTR(si, text) \
5376 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5378 return show_stat(s, buf, si); \
5380 static ssize_t text##_store(struct kmem_cache *s, \
5381 const char *buf, size_t length) \
5383 if (buf[0] != '0') \
5385 clear_stat(s, si); \
5390 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5391 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5392 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5393 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5394 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5395 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5396 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5397 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5398 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5399 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5400 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5401 STAT_ATTR(FREE_SLAB
, free_slab
);
5402 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5403 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5404 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5405 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5406 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5407 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5408 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5409 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5410 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5411 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5412 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5413 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5414 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5415 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5416 #endif /* CONFIG_SLUB_STATS */
5418 static struct attribute
*slab_attrs
[] = {
5419 &slab_size_attr
.attr
,
5420 &object_size_attr
.attr
,
5421 &objs_per_slab_attr
.attr
,
5423 &min_partial_attr
.attr
,
5424 &cpu_partial_attr
.attr
,
5426 &objects_partial_attr
.attr
,
5428 &cpu_slabs_attr
.attr
,
5432 &hwcache_align_attr
.attr
,
5433 &reclaim_account_attr
.attr
,
5434 &destroy_by_rcu_attr
.attr
,
5436 &slabs_cpu_partial_attr
.attr
,
5437 #ifdef CONFIG_SLUB_DEBUG
5438 &total_objects_attr
.attr
,
5440 &sanity_checks_attr
.attr
,
5442 &red_zone_attr
.attr
,
5444 &store_user_attr
.attr
,
5445 &validate_attr
.attr
,
5446 &alloc_calls_attr
.attr
,
5447 &free_calls_attr
.attr
,
5449 #ifdef CONFIG_ZONE_DMA
5450 &cache_dma_attr
.attr
,
5453 &remote_node_defrag_ratio_attr
.attr
,
5455 #ifdef CONFIG_SLUB_STATS
5456 &alloc_fastpath_attr
.attr
,
5457 &alloc_slowpath_attr
.attr
,
5458 &free_fastpath_attr
.attr
,
5459 &free_slowpath_attr
.attr
,
5460 &free_frozen_attr
.attr
,
5461 &free_add_partial_attr
.attr
,
5462 &free_remove_partial_attr
.attr
,
5463 &alloc_from_partial_attr
.attr
,
5464 &alloc_slab_attr
.attr
,
5465 &alloc_refill_attr
.attr
,
5466 &alloc_node_mismatch_attr
.attr
,
5467 &free_slab_attr
.attr
,
5468 &cpuslab_flush_attr
.attr
,
5469 &deactivate_full_attr
.attr
,
5470 &deactivate_empty_attr
.attr
,
5471 &deactivate_to_head_attr
.attr
,
5472 &deactivate_to_tail_attr
.attr
,
5473 &deactivate_remote_frees_attr
.attr
,
5474 &deactivate_bypass_attr
.attr
,
5475 &order_fallback_attr
.attr
,
5476 &cmpxchg_double_fail_attr
.attr
,
5477 &cmpxchg_double_cpu_fail_attr
.attr
,
5478 &cpu_partial_alloc_attr
.attr
,
5479 &cpu_partial_free_attr
.attr
,
5480 &cpu_partial_node_attr
.attr
,
5481 &cpu_partial_drain_attr
.attr
,
5483 #ifdef CONFIG_FAILSLAB
5484 &failslab_attr
.attr
,
5486 &usersize_attr
.attr
,
5491 static const struct attribute_group slab_attr_group
= {
5492 .attrs
= slab_attrs
,
5495 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5496 struct attribute
*attr
,
5499 struct slab_attribute
*attribute
;
5500 struct kmem_cache
*s
;
5503 attribute
= to_slab_attr(attr
);
5506 if (!attribute
->show
)
5509 err
= attribute
->show(s
, buf
);
5514 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5515 struct attribute
*attr
,
5516 const char *buf
, size_t len
)
5518 struct slab_attribute
*attribute
;
5519 struct kmem_cache
*s
;
5522 attribute
= to_slab_attr(attr
);
5525 if (!attribute
->store
)
5528 err
= attribute
->store(s
, buf
, len
);
5530 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5531 struct kmem_cache
*c
;
5533 mutex_lock(&slab_mutex
);
5534 if (s
->max_attr_size
< len
)
5535 s
->max_attr_size
= len
;
5538 * This is a best effort propagation, so this function's return
5539 * value will be determined by the parent cache only. This is
5540 * basically because not all attributes will have a well
5541 * defined semantics for rollbacks - most of the actions will
5542 * have permanent effects.
5544 * Returning the error value of any of the children that fail
5545 * is not 100 % defined, in the sense that users seeing the
5546 * error code won't be able to know anything about the state of
5549 * Only returning the error code for the parent cache at least
5550 * has well defined semantics. The cache being written to
5551 * directly either failed or succeeded, in which case we loop
5552 * through the descendants with best-effort propagation.
5554 for_each_memcg_cache(c
, s
)
5555 attribute
->store(c
, buf
, len
);
5556 mutex_unlock(&slab_mutex
);
5562 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5566 char *buffer
= NULL
;
5567 struct kmem_cache
*root_cache
;
5569 if (is_root_cache(s
))
5572 root_cache
= s
->memcg_params
.root_cache
;
5575 * This mean this cache had no attribute written. Therefore, no point
5576 * in copying default values around
5578 if (!root_cache
->max_attr_size
)
5581 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5584 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5587 if (!attr
|| !attr
->store
|| !attr
->show
)
5591 * It is really bad that we have to allocate here, so we will
5592 * do it only as a fallback. If we actually allocate, though,
5593 * we can just use the allocated buffer until the end.
5595 * Most of the slub attributes will tend to be very small in
5596 * size, but sysfs allows buffers up to a page, so they can
5597 * theoretically happen.
5601 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5604 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5605 if (WARN_ON(!buffer
))
5610 len
= attr
->show(root_cache
, buf
);
5612 attr
->store(s
, buf
, len
);
5616 free_page((unsigned long)buffer
);
5617 #endif /* CONFIG_MEMCG */
5620 static void kmem_cache_release(struct kobject
*k
)
5622 slab_kmem_cache_release(to_slab(k
));
5625 static const struct sysfs_ops slab_sysfs_ops
= {
5626 .show
= slab_attr_show
,
5627 .store
= slab_attr_store
,
5630 static struct kobj_type slab_ktype
= {
5631 .sysfs_ops
= &slab_sysfs_ops
,
5632 .release
= kmem_cache_release
,
5635 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5637 struct kobj_type
*ktype
= get_ktype(kobj
);
5639 if (ktype
== &slab_ktype
)
5644 static const struct kset_uevent_ops slab_uevent_ops
= {
5645 .filter
= uevent_filter
,
5648 static struct kset
*slab_kset
;
5650 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5653 if (!is_root_cache(s
))
5654 return s
->memcg_params
.root_cache
->memcg_kset
;
5659 #define ID_STR_LENGTH 64
5661 /* Create a unique string id for a slab cache:
5663 * Format :[flags-]size
5665 static char *create_unique_id(struct kmem_cache
*s
)
5667 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5674 * First flags affecting slabcache operations. We will only
5675 * get here for aliasable slabs so we do not need to support
5676 * too many flags. The flags here must cover all flags that
5677 * are matched during merging to guarantee that the id is
5680 if (s
->flags
& SLAB_CACHE_DMA
)
5682 if (s
->flags
& SLAB_CACHE_DMA32
)
5684 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5686 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5688 if (s
->flags
& SLAB_ACCOUNT
)
5692 p
+= sprintf(p
, "%07u", s
->size
);
5694 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5698 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5700 struct kmem_cache
*s
=
5701 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5703 if (!s
->kobj
.state_in_sysfs
)
5705 * For a memcg cache, this may be called during
5706 * deactivation and again on shutdown. Remove only once.
5707 * A cache is never shut down before deactivation is
5708 * complete, so no need to worry about synchronization.
5713 kset_unregister(s
->memcg_kset
);
5715 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5717 kobject_put(&s
->kobj
);
5720 static int sysfs_slab_add(struct kmem_cache
*s
)
5724 struct kset
*kset
= cache_kset(s
);
5725 int unmergeable
= slab_unmergeable(s
);
5727 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5730 kobject_init(&s
->kobj
, &slab_ktype
);
5734 if (!unmergeable
&& disable_higher_order_debug
&&
5735 (slub_debug
& DEBUG_METADATA_FLAGS
))
5740 * Slabcache can never be merged so we can use the name proper.
5741 * This is typically the case for debug situations. In that
5742 * case we can catch duplicate names easily.
5744 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5748 * Create a unique name for the slab as a target
5751 name
= create_unique_id(s
);
5754 s
->kobj
.kset
= kset
;
5755 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5759 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5764 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5765 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5766 if (!s
->memcg_kset
) {
5773 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5775 /* Setup first alias */
5776 sysfs_slab_alias(s
, s
->name
);
5783 kobject_del(&s
->kobj
);
5787 static void sysfs_slab_remove(struct kmem_cache
*s
)
5789 if (slab_state
< FULL
)
5791 * Sysfs has not been setup yet so no need to remove the
5796 kobject_get(&s
->kobj
);
5797 schedule_work(&s
->kobj_remove_work
);
5800 void sysfs_slab_unlink(struct kmem_cache
*s
)
5802 if (slab_state
>= FULL
)
5803 kobject_del(&s
->kobj
);
5806 void sysfs_slab_release(struct kmem_cache
*s
)
5808 if (slab_state
>= FULL
)
5809 kobject_put(&s
->kobj
);
5813 * Need to buffer aliases during bootup until sysfs becomes
5814 * available lest we lose that information.
5816 struct saved_alias
{
5817 struct kmem_cache
*s
;
5819 struct saved_alias
*next
;
5822 static struct saved_alias
*alias_list
;
5824 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5826 struct saved_alias
*al
;
5828 if (slab_state
== FULL
) {
5830 * If we have a leftover link then remove it.
5832 sysfs_remove_link(&slab_kset
->kobj
, name
);
5833 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5836 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5842 al
->next
= alias_list
;
5847 static int __init
slab_sysfs_init(void)
5849 struct kmem_cache
*s
;
5852 mutex_lock(&slab_mutex
);
5854 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5856 mutex_unlock(&slab_mutex
);
5857 pr_err("Cannot register slab subsystem.\n");
5863 list_for_each_entry(s
, &slab_caches
, list
) {
5864 err
= sysfs_slab_add(s
);
5866 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5870 while (alias_list
) {
5871 struct saved_alias
*al
= alias_list
;
5873 alias_list
= alias_list
->next
;
5874 err
= sysfs_slab_alias(al
->s
, al
->name
);
5876 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5881 mutex_unlock(&slab_mutex
);
5886 __initcall(slab_sysfs_init
);
5887 #endif /* CONFIG_SYSFS */
5890 * The /proc/slabinfo ABI
5892 #ifdef CONFIG_SLUB_DEBUG
5893 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5895 unsigned long nr_slabs
= 0;
5896 unsigned long nr_objs
= 0;
5897 unsigned long nr_free
= 0;
5899 struct kmem_cache_node
*n
;
5901 for_each_kmem_cache_node(s
, node
, n
) {
5902 nr_slabs
+= node_nr_slabs(n
);
5903 nr_objs
+= node_nr_objs(n
);
5904 nr_free
+= count_partial(n
, count_free
);
5907 sinfo
->active_objs
= nr_objs
- nr_free
;
5908 sinfo
->num_objs
= nr_objs
;
5909 sinfo
->active_slabs
= nr_slabs
;
5910 sinfo
->num_slabs
= nr_slabs
;
5911 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5912 sinfo
->cache_order
= oo_order(s
->oo
);
5915 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5919 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5920 size_t count
, loff_t
*ppos
)
5924 #endif /* CONFIG_SLUB_DEBUG */