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 #ifdef CONFIG_SLUB_DEBUG
118 #ifdef CONFIG_SLUB_DEBUG_ON
119 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled
);
121 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled
);
125 static inline bool kmem_cache_debug(struct kmem_cache
*s
)
127 return kmem_cache_debug_flags(s
, SLAB_DEBUG_FLAGS
);
130 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
132 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
))
133 p
+= s
->red_left_pad
;
138 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
140 #ifdef CONFIG_SLUB_CPU_PARTIAL
141 return !kmem_cache_debug(s
);
148 * Issues still to be resolved:
150 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
152 * - Variable sizing of the per node arrays
155 /* Enable to test recovery from slab corruption on boot */
156 #undef SLUB_RESILIENCY_TEST
158 /* Enable to log cmpxchg failures */
159 #undef SLUB_DEBUG_CMPXCHG
162 * Mininum number of partial slabs. These will be left on the partial
163 * lists even if they are empty. kmem_cache_shrink may reclaim them.
165 #define MIN_PARTIAL 5
168 * Maximum number of desirable partial slabs.
169 * The existence of more partial slabs makes kmem_cache_shrink
170 * sort the partial list by the number of objects in use.
172 #define MAX_PARTIAL 10
174 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_STORE_USER)
178 * These debug flags cannot use CMPXCHG because there might be consistency
179 * issues when checking or reading debug information
181 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
186 * Debugging flags that require metadata to be stored in the slab. These get
187 * disabled when slub_debug=O is used and a cache's min order increases with
190 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
193 #define OO_MASK ((1 << OO_SHIFT) - 1)
194 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
196 /* Internal SLUB flags */
198 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
199 /* Use cmpxchg_double */
200 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
203 * Tracking user of a slab.
205 #define TRACK_ADDRS_COUNT 16
207 unsigned long addr
; /* Called from address */
208 #ifdef CONFIG_STACKTRACE
209 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
211 int cpu
; /* Was running on cpu */
212 int pid
; /* Pid context */
213 unsigned long when
; /* When did the operation occur */
216 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
219 static int sysfs_slab_add(struct kmem_cache
*);
220 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
222 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
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/HW_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 swab((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 object
= kasan_reset_tag(object
);
279 return freelist_dereference(s
, object
+ s
->offset
);
282 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
284 prefetch(object
+ s
->offset
);
287 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
289 unsigned long freepointer_addr
;
292 if (!debug_pagealloc_enabled_static())
293 return get_freepointer(s
, object
);
295 freepointer_addr
= (unsigned long)object
+ s
->offset
;
296 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
297 return freelist_ptr(s
, p
, freepointer_addr
);
300 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
302 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
304 #ifdef CONFIG_SLAB_FREELIST_HARDENED
305 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
308 freeptr_addr
= (unsigned long)kasan_reset_tag((void *)freeptr_addr
);
309 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
312 /* Loop over all objects in a slab */
313 #define for_each_object(__p, __s, __addr, __objects) \
314 for (__p = fixup_red_left(__s, __addr); \
315 __p < (__addr) + (__objects) * (__s)->size; \
318 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
320 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
323 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
326 struct kmem_cache_order_objects x
= {
327 (order
<< OO_SHIFT
) + order_objects(order
, size
)
333 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
335 return x
.x
>> OO_SHIFT
;
338 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
340 return x
.x
& OO_MASK
;
344 * Per slab locking using the pagelock
346 static __always_inline
void slab_lock(struct page
*page
)
348 VM_BUG_ON_PAGE(PageTail(page
), page
);
349 bit_spin_lock(PG_locked
, &page
->flags
);
352 static __always_inline
void slab_unlock(struct page
*page
)
354 VM_BUG_ON_PAGE(PageTail(page
), page
);
355 __bit_spin_unlock(PG_locked
, &page
->flags
);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
360 void *freelist_old
, unsigned long counters_old
,
361 void *freelist_new
, unsigned long counters_new
,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s
->flags
& __CMPXCHG_DOUBLE
) {
368 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
369 freelist_old
, counters_old
,
370 freelist_new
, counters_new
))
376 if (page
->freelist
== freelist_old
&&
377 page
->counters
== counters_old
) {
378 page
->freelist
= freelist_new
;
379 page
->counters
= counters_new
;
387 stat(s
, CMPXCHG_DOUBLE_FAIL
);
389 #ifdef SLUB_DEBUG_CMPXCHG
390 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
396 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
397 void *freelist_old
, unsigned long counters_old
,
398 void *freelist_new
, unsigned long counters_new
,
401 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403 if (s
->flags
& __CMPXCHG_DOUBLE
) {
404 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
405 freelist_old
, counters_old
,
406 freelist_new
, counters_new
))
413 local_irq_save(flags
);
415 if (page
->freelist
== freelist_old
&&
416 page
->counters
== counters_old
) {
417 page
->freelist
= freelist_new
;
418 page
->counters
= counters_new
;
420 local_irq_restore(flags
);
424 local_irq_restore(flags
);
428 stat(s
, CMPXCHG_DOUBLE_FAIL
);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
437 #ifdef CONFIG_SLUB_DEBUG
438 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
439 static DEFINE_SPINLOCK(object_map_lock
);
442 * Determine a map of object in use on a page.
444 * Node listlock must be held to guarantee that the page does
445 * not vanish from under us.
447 static unsigned long *get_map(struct kmem_cache
*s
, struct page
*page
)
448 __acquires(&object_map_lock
)
451 void *addr
= page_address(page
);
453 VM_BUG_ON(!irqs_disabled());
455 spin_lock(&object_map_lock
);
457 bitmap_zero(object_map
, page
->objects
);
459 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
460 set_bit(__obj_to_index(s
, addr
, p
), object_map
);
465 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
467 VM_BUG_ON(map
!= object_map
);
468 spin_unlock(&object_map_lock
);
471 static inline unsigned int size_from_object(struct kmem_cache
*s
)
473 if (s
->flags
& SLAB_RED_ZONE
)
474 return s
->size
- s
->red_left_pad
;
479 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
481 if (s
->flags
& SLAB_RED_ZONE
)
482 p
-= s
->red_left_pad
;
490 #if defined(CONFIG_SLUB_DEBUG_ON)
491 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
493 static slab_flags_t slub_debug
;
496 static char *slub_debug_string
;
497 static int disable_higher_order_debug
;
500 * slub is about to manipulate internal object metadata. This memory lies
501 * outside the range of the allocated object, so accessing it would normally
502 * be reported by kasan as a bounds error. metadata_access_enable() is used
503 * to tell kasan that these accesses are OK.
505 static inline void metadata_access_enable(void)
507 kasan_disable_current();
510 static inline void metadata_access_disable(void)
512 kasan_enable_current();
519 /* Verify that a pointer has an address that is valid within a slab page */
520 static inline int check_valid_pointer(struct kmem_cache
*s
,
521 struct page
*page
, void *object
)
528 base
= page_address(page
);
529 object
= kasan_reset_tag(object
);
530 object
= restore_red_left(s
, object
);
531 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
532 (object
- base
) % s
->size
) {
539 static void print_section(char *level
, char *text
, u8
*addr
,
542 metadata_access_enable();
543 print_hex_dump(level
, kasan_reset_tag(text
), DUMP_PREFIX_ADDRESS
,
544 16, 1, addr
, length
, 1);
545 metadata_access_disable();
549 * See comment in calculate_sizes().
551 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
553 return s
->offset
>= s
->inuse
;
557 * Return offset of the end of info block which is inuse + free pointer if
558 * not overlapping with object.
560 static inline unsigned int get_info_end(struct kmem_cache
*s
)
562 if (freeptr_outside_object(s
))
563 return s
->inuse
+ sizeof(void *);
568 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
569 enum track_item alloc
)
573 p
= object
+ get_info_end(s
);
575 return kasan_reset_tag(p
+ alloc
);
578 static void set_track(struct kmem_cache
*s
, void *object
,
579 enum track_item alloc
, unsigned long addr
)
581 struct track
*p
= get_track(s
, object
, alloc
);
584 #ifdef CONFIG_STACKTRACE
585 unsigned int nr_entries
;
587 metadata_access_enable();
588 nr_entries
= stack_trace_save(kasan_reset_tag(p
->addrs
),
589 TRACK_ADDRS_COUNT
, 3);
590 metadata_access_disable();
592 if (nr_entries
< TRACK_ADDRS_COUNT
)
593 p
->addrs
[nr_entries
] = 0;
596 p
->cpu
= smp_processor_id();
597 p
->pid
= current
->pid
;
600 memset(p
, 0, sizeof(struct track
));
604 static void init_tracking(struct kmem_cache
*s
, void *object
)
606 if (!(s
->flags
& SLAB_STORE_USER
))
609 set_track(s
, object
, TRACK_FREE
, 0UL);
610 set_track(s
, object
, TRACK_ALLOC
, 0UL);
613 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
618 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
619 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
620 #ifdef CONFIG_STACKTRACE
623 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
625 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
632 void print_tracking(struct kmem_cache
*s
, void *object
)
634 unsigned long pr_time
= jiffies
;
635 if (!(s
->flags
& SLAB_STORE_USER
))
638 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
639 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
642 static void print_page_info(struct page
*page
)
644 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
645 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
649 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
651 struct va_format vaf
;
657 pr_err("=============================================================================\n");
658 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
659 pr_err("-----------------------------------------------------------------------------\n\n");
661 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
665 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
667 struct va_format vaf
;
673 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
677 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
678 void **freelist
, void *nextfree
)
680 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
681 !check_valid_pointer(s
, page
, nextfree
) && freelist
) {
682 object_err(s
, page
, *freelist
, "Freechain corrupt");
684 slab_fix(s
, "Isolate corrupted freechain");
691 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
693 unsigned int off
; /* Offset of last byte */
694 u8
*addr
= page_address(page
);
696 print_tracking(s
, p
);
698 print_page_info(page
);
700 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
701 p
, p
- addr
, get_freepointer(s
, p
));
703 if (s
->flags
& SLAB_RED_ZONE
)
704 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
706 else if (p
> addr
+ 16)
707 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
709 print_section(KERN_ERR
, "Object ", p
,
710 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
711 if (s
->flags
& SLAB_RED_ZONE
)
712 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
713 s
->inuse
- s
->object_size
);
715 off
= get_info_end(s
);
717 if (s
->flags
& SLAB_STORE_USER
)
718 off
+= 2 * sizeof(struct track
);
720 off
+= kasan_metadata_size(s
);
722 if (off
!= size_from_object(s
))
723 /* Beginning of the filler is the free pointer */
724 print_section(KERN_ERR
, "Padding ", p
+ off
,
725 size_from_object(s
) - off
);
730 void object_err(struct kmem_cache
*s
, struct page
*page
,
731 u8
*object
, char *reason
)
733 slab_bug(s
, "%s", reason
);
734 print_trailer(s
, page
, object
);
737 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
738 const char *fmt
, ...)
744 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
746 slab_bug(s
, "%s", buf
);
747 print_page_info(page
);
751 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
753 u8
*p
= kasan_reset_tag(object
);
755 if (s
->flags
& SLAB_RED_ZONE
)
756 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
758 if (s
->flags
& __OBJECT_POISON
) {
759 memset(p
, POISON_FREE
, s
->object_size
- 1);
760 p
[s
->object_size
- 1] = POISON_END
;
763 if (s
->flags
& SLAB_RED_ZONE
)
764 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
767 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
768 void *from
, void *to
)
770 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
771 memset(from
, data
, to
- from
);
774 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
775 u8
*object
, char *what
,
776 u8
*start
, unsigned int value
, unsigned int bytes
)
780 u8
*addr
= page_address(page
);
782 metadata_access_enable();
783 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
784 metadata_access_disable();
789 while (end
> fault
&& end
[-1] == value
)
792 slab_bug(s
, "%s overwritten", what
);
793 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
794 fault
, end
- 1, fault
- addr
,
796 print_trailer(s
, page
, object
);
798 restore_bytes(s
, what
, value
, fault
, end
);
806 * Bytes of the object to be managed.
807 * If the freepointer may overlay the object then the free
808 * pointer is at the middle of the object.
810 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
813 * object + s->object_size
814 * Padding to reach word boundary. This is also used for Redzoning.
815 * Padding is extended by another word if Redzoning is enabled and
816 * object_size == inuse.
818 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
819 * 0xcc (RED_ACTIVE) for objects in use.
822 * Meta data starts here.
824 * A. Free pointer (if we cannot overwrite object on free)
825 * B. Tracking data for SLAB_STORE_USER
826 * C. Padding to reach required alignment boundary or at mininum
827 * one word if debugging is on to be able to detect writes
828 * before the word boundary.
830 * Padding is done using 0x5a (POISON_INUSE)
833 * Nothing is used beyond s->size.
835 * If slabcaches are merged then the object_size and inuse boundaries are mostly
836 * ignored. And therefore no slab options that rely on these boundaries
837 * may be used with merged slabcaches.
840 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
842 unsigned long off
= get_info_end(s
); /* The end of info */
844 if (s
->flags
& SLAB_STORE_USER
)
845 /* We also have user information there */
846 off
+= 2 * sizeof(struct track
);
848 off
+= kasan_metadata_size(s
);
850 if (size_from_object(s
) == off
)
853 return check_bytes_and_report(s
, page
, p
, "Object padding",
854 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
857 /* Check the pad bytes at the end of a slab page */
858 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
867 if (!(s
->flags
& SLAB_POISON
))
870 start
= page_address(page
);
871 length
= page_size(page
);
872 end
= start
+ length
;
873 remainder
= length
% s
->size
;
877 pad
= end
- remainder
;
878 metadata_access_enable();
879 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
880 metadata_access_disable();
883 while (end
> fault
&& end
[-1] == POISON_INUSE
)
886 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
887 fault
, end
- 1, fault
- start
);
888 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
890 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
894 static int check_object(struct kmem_cache
*s
, struct page
*page
,
895 void *object
, u8 val
)
898 u8
*endobject
= object
+ s
->object_size
;
900 if (s
->flags
& SLAB_RED_ZONE
) {
901 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
902 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
905 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
906 endobject
, val
, s
->inuse
- s
->object_size
))
909 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
910 check_bytes_and_report(s
, page
, p
, "Alignment padding",
911 endobject
, POISON_INUSE
,
912 s
->inuse
- s
->object_size
);
916 if (s
->flags
& SLAB_POISON
) {
917 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
918 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
919 POISON_FREE
, s
->object_size
- 1) ||
920 !check_bytes_and_report(s
, page
, p
, "Poison",
921 p
+ s
->object_size
- 1, POISON_END
, 1)))
924 * check_pad_bytes cleans up on its own.
926 check_pad_bytes(s
, page
, p
);
929 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
931 * Object and freepointer overlap. Cannot check
932 * freepointer while object is allocated.
936 /* Check free pointer validity */
937 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
938 object_err(s
, page
, p
, "Freepointer corrupt");
940 * No choice but to zap it and thus lose the remainder
941 * of the free objects in this slab. May cause
942 * another error because the object count is now wrong.
944 set_freepointer(s
, p
, NULL
);
950 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
954 VM_BUG_ON(!irqs_disabled());
956 if (!PageSlab(page
)) {
957 slab_err(s
, page
, "Not a valid slab page");
961 maxobj
= order_objects(compound_order(page
), s
->size
);
962 if (page
->objects
> maxobj
) {
963 slab_err(s
, page
, "objects %u > max %u",
964 page
->objects
, maxobj
);
967 if (page
->inuse
> page
->objects
) {
968 slab_err(s
, page
, "inuse %u > max %u",
969 page
->inuse
, page
->objects
);
972 /* Slab_pad_check fixes things up after itself */
973 slab_pad_check(s
, page
);
978 * Determine if a certain object on a page is on the freelist. Must hold the
979 * slab lock to guarantee that the chains are in a consistent state.
981 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
989 while (fp
&& nr
<= page
->objects
) {
992 if (!check_valid_pointer(s
, page
, fp
)) {
994 object_err(s
, page
, object
,
995 "Freechain corrupt");
996 set_freepointer(s
, object
, NULL
);
998 slab_err(s
, page
, "Freepointer corrupt");
999 page
->freelist
= NULL
;
1000 page
->inuse
= page
->objects
;
1001 slab_fix(s
, "Freelist cleared");
1007 fp
= get_freepointer(s
, object
);
1011 max_objects
= order_objects(compound_order(page
), s
->size
);
1012 if (max_objects
> MAX_OBJS_PER_PAGE
)
1013 max_objects
= MAX_OBJS_PER_PAGE
;
1015 if (page
->objects
!= max_objects
) {
1016 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1017 page
->objects
, max_objects
);
1018 page
->objects
= max_objects
;
1019 slab_fix(s
, "Number of objects adjusted.");
1021 if (page
->inuse
!= page
->objects
- nr
) {
1022 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1023 page
->inuse
, page
->objects
- nr
);
1024 page
->inuse
= page
->objects
- nr
;
1025 slab_fix(s
, "Object count adjusted.");
1027 return search
== NULL
;
1030 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1033 if (s
->flags
& SLAB_TRACE
) {
1034 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1036 alloc
? "alloc" : "free",
1037 object
, page
->inuse
,
1041 print_section(KERN_INFO
, "Object ", (void *)object
,
1049 * Tracking of fully allocated slabs for debugging purposes.
1051 static void add_full(struct kmem_cache
*s
,
1052 struct kmem_cache_node
*n
, struct page
*page
)
1054 if (!(s
->flags
& SLAB_STORE_USER
))
1057 lockdep_assert_held(&n
->list_lock
);
1058 list_add(&page
->slab_list
, &n
->full
);
1061 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1063 if (!(s
->flags
& SLAB_STORE_USER
))
1066 lockdep_assert_held(&n
->list_lock
);
1067 list_del(&page
->slab_list
);
1070 /* Tracking of the number of slabs for debugging purposes */
1071 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1073 struct kmem_cache_node
*n
= get_node(s
, node
);
1075 return atomic_long_read(&n
->nr_slabs
);
1078 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1080 return atomic_long_read(&n
->nr_slabs
);
1083 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1085 struct kmem_cache_node
*n
= get_node(s
, node
);
1088 * May be called early in order to allocate a slab for the
1089 * kmem_cache_node structure. Solve the chicken-egg
1090 * dilemma by deferring the increment of the count during
1091 * bootstrap (see early_kmem_cache_node_alloc).
1094 atomic_long_inc(&n
->nr_slabs
);
1095 atomic_long_add(objects
, &n
->total_objects
);
1098 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1100 struct kmem_cache_node
*n
= get_node(s
, node
);
1102 atomic_long_dec(&n
->nr_slabs
);
1103 atomic_long_sub(objects
, &n
->total_objects
);
1106 /* Object debug checks for alloc/free paths */
1107 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1110 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1113 init_object(s
, object
, SLUB_RED_INACTIVE
);
1114 init_tracking(s
, object
);
1118 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1120 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1123 metadata_access_enable();
1124 memset(kasan_reset_tag(addr
), POISON_INUSE
, page_size(page
));
1125 metadata_access_disable();
1128 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1129 struct page
*page
, void *object
)
1131 if (!check_slab(s
, page
))
1134 if (!check_valid_pointer(s
, page
, object
)) {
1135 object_err(s
, page
, object
, "Freelist Pointer check fails");
1139 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1145 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1147 void *object
, unsigned long addr
)
1149 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1150 if (!alloc_consistency_checks(s
, page
, object
))
1154 /* Success perform special debug activities for allocs */
1155 if (s
->flags
& SLAB_STORE_USER
)
1156 set_track(s
, object
, TRACK_ALLOC
, addr
);
1157 trace(s
, page
, object
, 1);
1158 init_object(s
, object
, SLUB_RED_ACTIVE
);
1162 if (PageSlab(page
)) {
1164 * If this is a slab page then lets do the best we can
1165 * to avoid issues in the future. Marking all objects
1166 * as used avoids touching the remaining objects.
1168 slab_fix(s
, "Marking all objects used");
1169 page
->inuse
= page
->objects
;
1170 page
->freelist
= NULL
;
1175 static inline int free_consistency_checks(struct kmem_cache
*s
,
1176 struct page
*page
, void *object
, unsigned long addr
)
1178 if (!check_valid_pointer(s
, page
, object
)) {
1179 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1183 if (on_freelist(s
, page
, object
)) {
1184 object_err(s
, page
, object
, "Object already free");
1188 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1191 if (unlikely(s
!= page
->slab_cache
)) {
1192 if (!PageSlab(page
)) {
1193 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1195 } else if (!page
->slab_cache
) {
1196 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1200 object_err(s
, page
, object
,
1201 "page slab pointer corrupt.");
1207 /* Supports checking bulk free of a constructed freelist */
1208 static noinline
int free_debug_processing(
1209 struct kmem_cache
*s
, struct page
*page
,
1210 void *head
, void *tail
, int bulk_cnt
,
1213 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1214 void *object
= head
;
1216 unsigned long flags
;
1219 spin_lock_irqsave(&n
->list_lock
, flags
);
1222 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1223 if (!check_slab(s
, page
))
1230 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1231 if (!free_consistency_checks(s
, page
, object
, addr
))
1235 if (s
->flags
& SLAB_STORE_USER
)
1236 set_track(s
, object
, TRACK_FREE
, addr
);
1237 trace(s
, page
, object
, 0);
1238 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1239 init_object(s
, object
, SLUB_RED_INACTIVE
);
1241 /* Reached end of constructed freelist yet? */
1242 if (object
!= tail
) {
1243 object
= get_freepointer(s
, object
);
1249 if (cnt
!= bulk_cnt
)
1250 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1254 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1256 slab_fix(s
, "Object at 0x%p not freed", object
);
1261 * Parse a block of slub_debug options. Blocks are delimited by ';'
1263 * @str: start of block
1264 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1265 * @slabs: return start of list of slabs, or NULL when there's no list
1266 * @init: assume this is initial parsing and not per-kmem-create parsing
1268 * returns the start of next block if there's any, or NULL
1271 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1273 bool higher_order_disable
= false;
1275 /* Skip any completely empty blocks */
1276 while (*str
&& *str
== ';')
1281 * No options but restriction on slabs. This means full
1282 * debugging for slabs matching a pattern.
1284 *flags
= DEBUG_DEFAULT_FLAGS
;
1289 /* Determine which debug features should be switched on */
1290 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1291 switch (tolower(*str
)) {
1296 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1299 *flags
|= SLAB_RED_ZONE
;
1302 *flags
|= SLAB_POISON
;
1305 *flags
|= SLAB_STORE_USER
;
1308 *flags
|= SLAB_TRACE
;
1311 *flags
|= SLAB_FAILSLAB
;
1315 * Avoid enabling debugging on caches if its minimum
1316 * order would increase as a result.
1318 higher_order_disable
= true;
1322 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1331 /* Skip over the slab list */
1332 while (*str
&& *str
!= ';')
1335 /* Skip any completely empty blocks */
1336 while (*str
&& *str
== ';')
1339 if (init
&& higher_order_disable
)
1340 disable_higher_order_debug
= 1;
1348 static int __init
setup_slub_debug(char *str
)
1353 bool global_slub_debug_changed
= false;
1354 bool slab_list_specified
= false;
1356 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1357 if (*str
++ != '=' || !*str
)
1359 * No options specified. Switch on full debugging.
1365 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1369 global_slub_debug_changed
= true;
1371 slab_list_specified
= true;
1376 * For backwards compatibility, a single list of flags with list of
1377 * slabs means debugging is only enabled for those slabs, so the global
1378 * slub_debug should be 0. We can extended that to multiple lists as
1379 * long as there is no option specifying flags without a slab list.
1381 if (slab_list_specified
) {
1382 if (!global_slub_debug_changed
)
1384 slub_debug_string
= saved_str
;
1387 if (slub_debug
!= 0 || slub_debug_string
)
1388 static_branch_enable(&slub_debug_enabled
);
1389 if ((static_branch_unlikely(&init_on_alloc
) ||
1390 static_branch_unlikely(&init_on_free
)) &&
1391 (slub_debug
& SLAB_POISON
))
1392 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1396 __setup("slub_debug", setup_slub_debug
);
1399 * kmem_cache_flags - apply debugging options to the cache
1400 * @object_size: the size of an object without meta data
1401 * @flags: flags to set
1402 * @name: name of the cache
1403 * @ctor: constructor function
1405 * Debug option(s) are applied to @flags. In addition to the debug
1406 * option(s), if a slab name (or multiple) is specified i.e.
1407 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1408 * then only the select slabs will receive the debug option(s).
1410 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1411 slab_flags_t flags
, const char *name
,
1412 void (*ctor
)(void *))
1417 slab_flags_t block_flags
;
1420 next_block
= slub_debug_string
;
1421 /* Go through all blocks of debug options, see if any matches our slab's name */
1422 while (next_block
) {
1423 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1426 /* Found a block that has a slab list, search it */
1431 end
= strchrnul(iter
, ',');
1432 if (next_block
&& next_block
< end
)
1433 end
= next_block
- 1;
1435 glob
= strnchr(iter
, end
- iter
, '*');
1437 cmplen
= glob
- iter
;
1439 cmplen
= max_t(size_t, len
, (end
- iter
));
1441 if (!strncmp(name
, iter
, cmplen
)) {
1442 flags
|= block_flags
;
1446 if (!*end
|| *end
== ';')
1452 return flags
| slub_debug
;
1454 #else /* !CONFIG_SLUB_DEBUG */
1455 static inline void setup_object_debug(struct kmem_cache
*s
,
1456 struct page
*page
, void *object
) {}
1458 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1460 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1461 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1463 static inline int free_debug_processing(
1464 struct kmem_cache
*s
, struct page
*page
,
1465 void *head
, void *tail
, int bulk_cnt
,
1466 unsigned long addr
) { return 0; }
1468 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1470 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1471 void *object
, u8 val
) { return 1; }
1472 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1473 struct page
*page
) {}
1474 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1475 struct page
*page
) {}
1476 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1477 slab_flags_t flags
, const char *name
,
1478 void (*ctor
)(void *))
1482 #define slub_debug 0
1484 #define disable_higher_order_debug 0
1486 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1488 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1490 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1492 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1495 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1496 void **freelist
, void *nextfree
)
1500 #endif /* CONFIG_SLUB_DEBUG */
1503 * Hooks for other subsystems that check memory allocations. In a typical
1504 * production configuration these hooks all should produce no code at all.
1506 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1508 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1509 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1510 kmemleak_alloc(ptr
, size
, 1, flags
);
1514 static __always_inline
void kfree_hook(void *x
)
1517 kasan_kfree_large(x
, _RET_IP_
);
1520 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1522 kmemleak_free_recursive(x
, s
->flags
);
1525 * Trouble is that we may no longer disable interrupts in the fast path
1526 * So in order to make the debug calls that expect irqs to be
1527 * disabled we need to disable interrupts temporarily.
1529 #ifdef CONFIG_LOCKDEP
1531 unsigned long flags
;
1533 local_irq_save(flags
);
1534 debug_check_no_locks_freed(x
, s
->object_size
);
1535 local_irq_restore(flags
);
1538 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1539 debug_check_no_obj_freed(x
, s
->object_size
);
1541 /* Use KCSAN to help debug racy use-after-free. */
1542 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1543 __kcsan_check_access(x
, s
->object_size
,
1544 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1546 /* KASAN might put x into memory quarantine, delaying its reuse */
1547 return kasan_slab_free(s
, x
, _RET_IP_
);
1550 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1551 void **head
, void **tail
)
1556 void *old_tail
= *tail
? *tail
: *head
;
1559 /* Head and tail of the reconstructed freelist */
1565 next
= get_freepointer(s
, object
);
1567 if (slab_want_init_on_free(s
)) {
1569 * Clear the object and the metadata, but don't touch
1572 memset(kasan_reset_tag(object
), 0, s
->object_size
);
1573 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
1575 memset((char *)kasan_reset_tag(object
) + s
->inuse
, 0,
1576 s
->size
- s
->inuse
- rsize
);
1579 /* If object's reuse doesn't have to be delayed */
1580 if (!slab_free_hook(s
, object
)) {
1581 /* Move object to the new freelist */
1582 set_freepointer(s
, object
, *head
);
1587 } while (object
!= old_tail
);
1592 return *head
!= NULL
;
1595 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1598 setup_object_debug(s
, page
, object
);
1599 object
= kasan_init_slab_obj(s
, object
);
1600 if (unlikely(s
->ctor
)) {
1601 kasan_unpoison_object_data(s
, object
);
1603 kasan_poison_object_data(s
, object
);
1609 * Slab allocation and freeing
1611 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1612 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1615 unsigned int order
= oo_order(oo
);
1617 if (node
== NUMA_NO_NODE
)
1618 page
= alloc_pages(flags
, order
);
1620 page
= __alloc_pages_node(node
, flags
, order
);
1625 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1626 /* Pre-initialize the random sequence cache */
1627 static int init_cache_random_seq(struct kmem_cache
*s
)
1629 unsigned int count
= oo_objects(s
->oo
);
1632 /* Bailout if already initialised */
1636 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1638 pr_err("SLUB: Unable to initialize free list for %s\n",
1643 /* Transform to an offset on the set of pages */
1644 if (s
->random_seq
) {
1647 for (i
= 0; i
< count
; i
++)
1648 s
->random_seq
[i
] *= s
->size
;
1653 /* Initialize each random sequence freelist per cache */
1654 static void __init
init_freelist_randomization(void)
1656 struct kmem_cache
*s
;
1658 mutex_lock(&slab_mutex
);
1660 list_for_each_entry(s
, &slab_caches
, list
)
1661 init_cache_random_seq(s
);
1663 mutex_unlock(&slab_mutex
);
1666 /* Get the next entry on the pre-computed freelist randomized */
1667 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1668 unsigned long *pos
, void *start
,
1669 unsigned long page_limit
,
1670 unsigned long freelist_count
)
1675 * If the target page allocation failed, the number of objects on the
1676 * page might be smaller than the usual size defined by the cache.
1679 idx
= s
->random_seq
[*pos
];
1681 if (*pos
>= freelist_count
)
1683 } while (unlikely(idx
>= page_limit
));
1685 return (char *)start
+ idx
;
1688 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1689 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1694 unsigned long idx
, pos
, page_limit
, freelist_count
;
1696 if (page
->objects
< 2 || !s
->random_seq
)
1699 freelist_count
= oo_objects(s
->oo
);
1700 pos
= get_random_int() % freelist_count
;
1702 page_limit
= page
->objects
* s
->size
;
1703 start
= fixup_red_left(s
, page_address(page
));
1705 /* First entry is used as the base of the freelist */
1706 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1708 cur
= setup_object(s
, page
, cur
);
1709 page
->freelist
= cur
;
1711 for (idx
= 1; idx
< page
->objects
; idx
++) {
1712 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1714 next
= setup_object(s
, page
, next
);
1715 set_freepointer(s
, cur
, next
);
1718 set_freepointer(s
, cur
, NULL
);
1723 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1727 static inline void init_freelist_randomization(void) { }
1728 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1732 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1734 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1737 struct kmem_cache_order_objects oo
= s
->oo
;
1739 void *start
, *p
, *next
;
1743 flags
&= gfp_allowed_mask
;
1745 if (gfpflags_allow_blocking(flags
))
1748 flags
|= s
->allocflags
;
1751 * Let the initial higher-order allocation fail under memory pressure
1752 * so we fall-back to the minimum order allocation.
1754 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1755 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1756 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1758 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1759 if (unlikely(!page
)) {
1763 * Allocation may have failed due to fragmentation.
1764 * Try a lower order alloc if possible
1766 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1767 if (unlikely(!page
))
1769 stat(s
, ORDER_FALLBACK
);
1772 page
->objects
= oo_objects(oo
);
1774 account_slab_page(page
, oo_order(oo
), s
);
1776 page
->slab_cache
= s
;
1777 __SetPageSlab(page
);
1778 if (page_is_pfmemalloc(page
))
1779 SetPageSlabPfmemalloc(page
);
1781 kasan_poison_slab(page
);
1783 start
= page_address(page
);
1785 setup_page_debug(s
, page
, start
);
1787 shuffle
= shuffle_freelist(s
, page
);
1790 start
= fixup_red_left(s
, start
);
1791 start
= setup_object(s
, page
, start
);
1792 page
->freelist
= start
;
1793 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1795 next
= setup_object(s
, page
, next
);
1796 set_freepointer(s
, p
, next
);
1799 set_freepointer(s
, p
, NULL
);
1802 page
->inuse
= page
->objects
;
1806 if (gfpflags_allow_blocking(flags
))
1807 local_irq_disable();
1811 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1816 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1818 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1819 flags
= kmalloc_fix_flags(flags
);
1821 return allocate_slab(s
,
1822 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1825 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1827 int order
= compound_order(page
);
1828 int pages
= 1 << order
;
1830 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
1833 slab_pad_check(s
, page
);
1834 for_each_object(p
, s
, page_address(page
),
1836 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1839 __ClearPageSlabPfmemalloc(page
);
1840 __ClearPageSlab(page
);
1841 /* In union with page->mapping where page allocator expects NULL */
1842 page
->slab_cache
= NULL
;
1843 if (current
->reclaim_state
)
1844 current
->reclaim_state
->reclaimed_slab
+= pages
;
1845 unaccount_slab_page(page
, order
, s
);
1846 __free_pages(page
, order
);
1849 static void rcu_free_slab(struct rcu_head
*h
)
1851 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1853 __free_slab(page
->slab_cache
, page
);
1856 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1858 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1859 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1861 __free_slab(s
, page
);
1864 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1866 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1871 * Management of partially allocated slabs.
1874 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1877 if (tail
== DEACTIVATE_TO_TAIL
)
1878 list_add_tail(&page
->slab_list
, &n
->partial
);
1880 list_add(&page
->slab_list
, &n
->partial
);
1883 static inline void add_partial(struct kmem_cache_node
*n
,
1884 struct page
*page
, int tail
)
1886 lockdep_assert_held(&n
->list_lock
);
1887 __add_partial(n
, page
, tail
);
1890 static inline void remove_partial(struct kmem_cache_node
*n
,
1893 lockdep_assert_held(&n
->list_lock
);
1894 list_del(&page
->slab_list
);
1899 * Remove slab from the partial list, freeze it and
1900 * return the pointer to the freelist.
1902 * Returns a list of objects or NULL if it fails.
1904 static inline void *acquire_slab(struct kmem_cache
*s
,
1905 struct kmem_cache_node
*n
, struct page
*page
,
1906 int mode
, int *objects
)
1909 unsigned long counters
;
1912 lockdep_assert_held(&n
->list_lock
);
1915 * Zap the freelist and set the frozen bit.
1916 * The old freelist is the list of objects for the
1917 * per cpu allocation list.
1919 freelist
= page
->freelist
;
1920 counters
= page
->counters
;
1921 new.counters
= counters
;
1922 *objects
= new.objects
- new.inuse
;
1924 new.inuse
= page
->objects
;
1925 new.freelist
= NULL
;
1927 new.freelist
= freelist
;
1930 VM_BUG_ON(new.frozen
);
1933 if (!__cmpxchg_double_slab(s
, page
,
1935 new.freelist
, new.counters
,
1939 remove_partial(n
, page
);
1944 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1945 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1948 * Try to allocate a partial slab from a specific node.
1950 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1951 struct kmem_cache_cpu
*c
, gfp_t flags
)
1953 struct page
*page
, *page2
;
1954 void *object
= NULL
;
1955 unsigned int available
= 0;
1959 * Racy check. If we mistakenly see no partial slabs then we
1960 * just allocate an empty slab. If we mistakenly try to get a
1961 * partial slab and there is none available then get_partial()
1964 if (!n
|| !n
->nr_partial
)
1967 spin_lock(&n
->list_lock
);
1968 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1971 if (!pfmemalloc_match(page
, flags
))
1974 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1978 available
+= objects
;
1981 stat(s
, ALLOC_FROM_PARTIAL
);
1984 put_cpu_partial(s
, page
, 0);
1985 stat(s
, CPU_PARTIAL_NODE
);
1987 if (!kmem_cache_has_cpu_partial(s
)
1988 || available
> slub_cpu_partial(s
) / 2)
1992 spin_unlock(&n
->list_lock
);
1997 * Get a page from somewhere. Search in increasing NUMA distances.
1999 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2000 struct kmem_cache_cpu
*c
)
2003 struct zonelist
*zonelist
;
2006 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2008 unsigned int cpuset_mems_cookie
;
2011 * The defrag ratio allows a configuration of the tradeoffs between
2012 * inter node defragmentation and node local allocations. A lower
2013 * defrag_ratio increases the tendency to do local allocations
2014 * instead of attempting to obtain partial slabs from other nodes.
2016 * If the defrag_ratio is set to 0 then kmalloc() always
2017 * returns node local objects. If the ratio is higher then kmalloc()
2018 * may return off node objects because partial slabs are obtained
2019 * from other nodes and filled up.
2021 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2022 * (which makes defrag_ratio = 1000) then every (well almost)
2023 * allocation will first attempt to defrag slab caches on other nodes.
2024 * This means scanning over all nodes to look for partial slabs which
2025 * may be expensive if we do it every time we are trying to find a slab
2026 * with available objects.
2028 if (!s
->remote_node_defrag_ratio
||
2029 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2033 cpuset_mems_cookie
= read_mems_allowed_begin();
2034 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2035 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2036 struct kmem_cache_node
*n
;
2038 n
= get_node(s
, zone_to_nid(zone
));
2040 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2041 n
->nr_partial
> s
->min_partial
) {
2042 object
= get_partial_node(s
, n
, c
, flags
);
2045 * Don't check read_mems_allowed_retry()
2046 * here - if mems_allowed was updated in
2047 * parallel, that was a harmless race
2048 * between allocation and the cpuset
2055 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2056 #endif /* CONFIG_NUMA */
2061 * Get a partial page, lock it and return it.
2063 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2064 struct kmem_cache_cpu
*c
)
2067 int searchnode
= node
;
2069 if (node
== NUMA_NO_NODE
)
2070 searchnode
= numa_mem_id();
2072 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
2073 if (object
|| node
!= NUMA_NO_NODE
)
2076 return get_any_partial(s
, flags
, c
);
2079 #ifdef CONFIG_PREEMPTION
2081 * Calculate the next globally unique transaction for disambiguation
2082 * during cmpxchg. The transactions start with the cpu number and are then
2083 * incremented by CONFIG_NR_CPUS.
2085 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2088 * No preemption supported therefore also no need to check for
2094 static inline unsigned long next_tid(unsigned long tid
)
2096 return tid
+ TID_STEP
;
2099 #ifdef SLUB_DEBUG_CMPXCHG
2100 static inline unsigned int tid_to_cpu(unsigned long tid
)
2102 return tid
% TID_STEP
;
2105 static inline unsigned long tid_to_event(unsigned long tid
)
2107 return tid
/ TID_STEP
;
2111 static inline unsigned int init_tid(int cpu
)
2116 static inline void note_cmpxchg_failure(const char *n
,
2117 const struct kmem_cache
*s
, unsigned long tid
)
2119 #ifdef SLUB_DEBUG_CMPXCHG
2120 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2122 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2124 #ifdef CONFIG_PREEMPTION
2125 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2126 pr_warn("due to cpu change %d -> %d\n",
2127 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2130 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2131 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2132 tid_to_event(tid
), tid_to_event(actual_tid
));
2134 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2135 actual_tid
, tid
, next_tid(tid
));
2137 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2140 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2144 for_each_possible_cpu(cpu
)
2145 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2149 * Remove the cpu slab
2151 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2152 void *freelist
, struct kmem_cache_cpu
*c
)
2154 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2155 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2157 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2159 int tail
= DEACTIVATE_TO_HEAD
;
2163 if (page
->freelist
) {
2164 stat(s
, DEACTIVATE_REMOTE_FREES
);
2165 tail
= DEACTIVATE_TO_TAIL
;
2169 * Stage one: Free all available per cpu objects back
2170 * to the page freelist while it is still frozen. Leave the
2173 * There is no need to take the list->lock because the page
2176 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2178 unsigned long counters
;
2181 * If 'nextfree' is invalid, it is possible that the object at
2182 * 'freelist' is already corrupted. So isolate all objects
2183 * starting at 'freelist'.
2185 if (freelist_corrupted(s
, page
, &freelist
, nextfree
))
2189 prior
= page
->freelist
;
2190 counters
= page
->counters
;
2191 set_freepointer(s
, freelist
, prior
);
2192 new.counters
= counters
;
2194 VM_BUG_ON(!new.frozen
);
2196 } while (!__cmpxchg_double_slab(s
, page
,
2198 freelist
, new.counters
,
2199 "drain percpu freelist"));
2201 freelist
= nextfree
;
2205 * Stage two: Ensure that the page is unfrozen while the
2206 * list presence reflects the actual number of objects
2209 * We setup the list membership and then perform a cmpxchg
2210 * with the count. If there is a mismatch then the page
2211 * is not unfrozen but the page is on the wrong list.
2213 * Then we restart the process which may have to remove
2214 * the page from the list that we just put it on again
2215 * because the number of objects in the slab may have
2220 old
.freelist
= page
->freelist
;
2221 old
.counters
= page
->counters
;
2222 VM_BUG_ON(!old
.frozen
);
2224 /* Determine target state of the slab */
2225 new.counters
= old
.counters
;
2228 set_freepointer(s
, freelist
, old
.freelist
);
2229 new.freelist
= freelist
;
2231 new.freelist
= old
.freelist
;
2235 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2237 else if (new.freelist
) {
2242 * Taking the spinlock removes the possibility
2243 * that acquire_slab() will see a slab page that
2246 spin_lock(&n
->list_lock
);
2250 if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
) && !lock
) {
2253 * This also ensures that the scanning of full
2254 * slabs from diagnostic functions will not see
2257 spin_lock(&n
->list_lock
);
2263 remove_partial(n
, page
);
2264 else if (l
== M_FULL
)
2265 remove_full(s
, n
, page
);
2268 add_partial(n
, page
, tail
);
2269 else if (m
== M_FULL
)
2270 add_full(s
, n
, page
);
2274 if (!__cmpxchg_double_slab(s
, page
,
2275 old
.freelist
, old
.counters
,
2276 new.freelist
, new.counters
,
2281 spin_unlock(&n
->list_lock
);
2285 else if (m
== M_FULL
)
2286 stat(s
, DEACTIVATE_FULL
);
2287 else if (m
== M_FREE
) {
2288 stat(s
, DEACTIVATE_EMPTY
);
2289 discard_slab(s
, page
);
2298 * Unfreeze all the cpu partial slabs.
2300 * This function must be called with interrupts disabled
2301 * for the cpu using c (or some other guarantee must be there
2302 * to guarantee no concurrent accesses).
2304 static void unfreeze_partials(struct kmem_cache
*s
,
2305 struct kmem_cache_cpu
*c
)
2307 #ifdef CONFIG_SLUB_CPU_PARTIAL
2308 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2309 struct page
*page
, *discard_page
= NULL
;
2311 while ((page
= slub_percpu_partial(c
))) {
2315 slub_set_percpu_partial(c
, page
);
2317 n2
= get_node(s
, page_to_nid(page
));
2320 spin_unlock(&n
->list_lock
);
2323 spin_lock(&n
->list_lock
);
2328 old
.freelist
= page
->freelist
;
2329 old
.counters
= page
->counters
;
2330 VM_BUG_ON(!old
.frozen
);
2332 new.counters
= old
.counters
;
2333 new.freelist
= old
.freelist
;
2337 } while (!__cmpxchg_double_slab(s
, page
,
2338 old
.freelist
, old
.counters
,
2339 new.freelist
, new.counters
,
2340 "unfreezing slab"));
2342 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2343 page
->next
= discard_page
;
2344 discard_page
= page
;
2346 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2347 stat(s
, FREE_ADD_PARTIAL
);
2352 spin_unlock(&n
->list_lock
);
2354 while (discard_page
) {
2355 page
= discard_page
;
2356 discard_page
= discard_page
->next
;
2358 stat(s
, DEACTIVATE_EMPTY
);
2359 discard_slab(s
, page
);
2362 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2366 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2367 * partial page slot if available.
2369 * If we did not find a slot then simply move all the partials to the
2370 * per node partial list.
2372 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2374 #ifdef CONFIG_SLUB_CPU_PARTIAL
2375 struct page
*oldpage
;
2383 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2386 pobjects
= oldpage
->pobjects
;
2387 pages
= oldpage
->pages
;
2388 if (drain
&& pobjects
> slub_cpu_partial(s
)) {
2389 unsigned long flags
;
2391 * partial array is full. Move the existing
2392 * set to the per node partial list.
2394 local_irq_save(flags
);
2395 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2396 local_irq_restore(flags
);
2400 stat(s
, CPU_PARTIAL_DRAIN
);
2405 pobjects
+= page
->objects
- page
->inuse
;
2407 page
->pages
= pages
;
2408 page
->pobjects
= pobjects
;
2409 page
->next
= oldpage
;
2411 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2413 if (unlikely(!slub_cpu_partial(s
))) {
2414 unsigned long flags
;
2416 local_irq_save(flags
);
2417 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2418 local_irq_restore(flags
);
2421 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2424 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2426 stat(s
, CPUSLAB_FLUSH
);
2427 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2429 c
->tid
= next_tid(c
->tid
);
2435 * Called from IPI handler with interrupts disabled.
2437 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2439 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2444 unfreeze_partials(s
, c
);
2447 static void flush_cpu_slab(void *d
)
2449 struct kmem_cache
*s
= d
;
2451 __flush_cpu_slab(s
, smp_processor_id());
2454 static bool has_cpu_slab(int cpu
, void *info
)
2456 struct kmem_cache
*s
= info
;
2457 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2459 return c
->page
|| slub_percpu_partial(c
);
2462 static void flush_all(struct kmem_cache
*s
)
2464 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1);
2468 * Use the cpu notifier to insure that the cpu slabs are flushed when
2471 static int slub_cpu_dead(unsigned int cpu
)
2473 struct kmem_cache
*s
;
2474 unsigned long flags
;
2476 mutex_lock(&slab_mutex
);
2477 list_for_each_entry(s
, &slab_caches
, list
) {
2478 local_irq_save(flags
);
2479 __flush_cpu_slab(s
, cpu
);
2480 local_irq_restore(flags
);
2482 mutex_unlock(&slab_mutex
);
2487 * Check if the objects in a per cpu structure fit numa
2488 * locality expectations.
2490 static inline int node_match(struct page
*page
, int node
)
2493 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2499 #ifdef CONFIG_SLUB_DEBUG
2500 static int count_free(struct page
*page
)
2502 return page
->objects
- page
->inuse
;
2505 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2507 return atomic_long_read(&n
->total_objects
);
2509 #endif /* CONFIG_SLUB_DEBUG */
2511 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2512 static unsigned long count_partial(struct kmem_cache_node
*n
,
2513 int (*get_count
)(struct page
*))
2515 unsigned long flags
;
2516 unsigned long x
= 0;
2519 spin_lock_irqsave(&n
->list_lock
, flags
);
2520 list_for_each_entry(page
, &n
->partial
, slab_list
)
2521 x
+= get_count(page
);
2522 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2525 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2527 static noinline
void
2528 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2530 #ifdef CONFIG_SLUB_DEBUG
2531 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2532 DEFAULT_RATELIMIT_BURST
);
2534 struct kmem_cache_node
*n
;
2536 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2539 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2540 nid
, gfpflags
, &gfpflags
);
2541 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2542 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2545 if (oo_order(s
->min
) > get_order(s
->object_size
))
2546 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2549 for_each_kmem_cache_node(s
, node
, n
) {
2550 unsigned long nr_slabs
;
2551 unsigned long nr_objs
;
2552 unsigned long nr_free
;
2554 nr_free
= count_partial(n
, count_free
);
2555 nr_slabs
= node_nr_slabs(n
);
2556 nr_objs
= node_nr_objs(n
);
2558 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2559 node
, nr_slabs
, nr_objs
, nr_free
);
2564 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2565 int node
, struct kmem_cache_cpu
**pc
)
2568 struct kmem_cache_cpu
*c
= *pc
;
2571 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2573 freelist
= get_partial(s
, flags
, node
, c
);
2578 page
= new_slab(s
, flags
, node
);
2580 c
= raw_cpu_ptr(s
->cpu_slab
);
2585 * No other reference to the page yet so we can
2586 * muck around with it freely without cmpxchg
2588 freelist
= page
->freelist
;
2589 page
->freelist
= NULL
;
2591 stat(s
, ALLOC_SLAB
);
2599 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2601 if (unlikely(PageSlabPfmemalloc(page
)))
2602 return gfp_pfmemalloc_allowed(gfpflags
);
2608 * Check the page->freelist of a page and either transfer the freelist to the
2609 * per cpu freelist or deactivate the page.
2611 * The page is still frozen if the return value is not NULL.
2613 * If this function returns NULL then the page has been unfrozen.
2615 * This function must be called with interrupt disabled.
2617 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2620 unsigned long counters
;
2624 freelist
= page
->freelist
;
2625 counters
= page
->counters
;
2627 new.counters
= counters
;
2628 VM_BUG_ON(!new.frozen
);
2630 new.inuse
= page
->objects
;
2631 new.frozen
= freelist
!= NULL
;
2633 } while (!__cmpxchg_double_slab(s
, page
,
2642 * Slow path. The lockless freelist is empty or we need to perform
2645 * Processing is still very fast if new objects have been freed to the
2646 * regular freelist. In that case we simply take over the regular freelist
2647 * as the lockless freelist and zap the regular freelist.
2649 * If that is not working then we fall back to the partial lists. We take the
2650 * first element of the freelist as the object to allocate now and move the
2651 * rest of the freelist to the lockless freelist.
2653 * And if we were unable to get a new slab from the partial slab lists then
2654 * we need to allocate a new slab. This is the slowest path since it involves
2655 * a call to the page allocator and the setup of a new slab.
2657 * Version of __slab_alloc to use when we know that interrupts are
2658 * already disabled (which is the case for bulk allocation).
2660 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2661 unsigned long addr
, struct kmem_cache_cpu
*c
)
2666 stat(s
, ALLOC_SLOWPATH
);
2671 * if the node is not online or has no normal memory, just
2672 * ignore the node constraint
2674 if (unlikely(node
!= NUMA_NO_NODE
&&
2675 !node_state(node
, N_NORMAL_MEMORY
)))
2676 node
= NUMA_NO_NODE
;
2681 if (unlikely(!node_match(page
, node
))) {
2683 * same as above but node_match() being false already
2684 * implies node != NUMA_NO_NODE
2686 if (!node_state(node
, N_NORMAL_MEMORY
)) {
2687 node
= NUMA_NO_NODE
;
2690 stat(s
, ALLOC_NODE_MISMATCH
);
2691 deactivate_slab(s
, page
, c
->freelist
, c
);
2697 * By rights, we should be searching for a slab page that was
2698 * PFMEMALLOC but right now, we are losing the pfmemalloc
2699 * information when the page leaves the per-cpu allocator
2701 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2702 deactivate_slab(s
, page
, c
->freelist
, c
);
2706 /* must check again c->freelist in case of cpu migration or IRQ */
2707 freelist
= c
->freelist
;
2711 freelist
= get_freelist(s
, page
);
2715 stat(s
, DEACTIVATE_BYPASS
);
2719 stat(s
, ALLOC_REFILL
);
2723 * freelist is pointing to the list of objects to be used.
2724 * page is pointing to the page from which the objects are obtained.
2725 * That page must be frozen for per cpu allocations to work.
2727 VM_BUG_ON(!c
->page
->frozen
);
2728 c
->freelist
= get_freepointer(s
, freelist
);
2729 c
->tid
= next_tid(c
->tid
);
2734 if (slub_percpu_partial(c
)) {
2735 page
= c
->page
= slub_percpu_partial(c
);
2736 slub_set_percpu_partial(c
, page
);
2737 stat(s
, CPU_PARTIAL_ALLOC
);
2741 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2743 if (unlikely(!freelist
)) {
2744 slab_out_of_memory(s
, gfpflags
, node
);
2749 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2752 /* Only entered in the debug case */
2753 if (kmem_cache_debug(s
) &&
2754 !alloc_debug_processing(s
, page
, freelist
, addr
))
2755 goto new_slab
; /* Slab failed checks. Next slab needed */
2757 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2762 * Another one that disabled interrupt and compensates for possible
2763 * cpu changes by refetching the per cpu area pointer.
2765 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2766 unsigned long addr
, struct kmem_cache_cpu
*c
)
2769 unsigned long flags
;
2771 local_irq_save(flags
);
2772 #ifdef CONFIG_PREEMPTION
2774 * We may have been preempted and rescheduled on a different
2775 * cpu before disabling interrupts. Need to reload cpu area
2778 c
= this_cpu_ptr(s
->cpu_slab
);
2781 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2782 local_irq_restore(flags
);
2787 * If the object has been wiped upon free, make sure it's fully initialized by
2788 * zeroing out freelist pointer.
2790 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2793 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2794 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
2799 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2800 * have the fastpath folded into their functions. So no function call
2801 * overhead for requests that can be satisfied on the fastpath.
2803 * The fastpath works by first checking if the lockless freelist can be used.
2804 * If not then __slab_alloc is called for slow processing.
2806 * Otherwise we can simply pick the next object from the lockless free list.
2808 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2809 gfp_t gfpflags
, int node
, unsigned long addr
)
2812 struct kmem_cache_cpu
*c
;
2815 struct obj_cgroup
*objcg
= NULL
;
2817 s
= slab_pre_alloc_hook(s
, &objcg
, 1, gfpflags
);
2822 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2823 * enabled. We may switch back and forth between cpus while
2824 * reading from one cpu area. That does not matter as long
2825 * as we end up on the original cpu again when doing the cmpxchg.
2827 * We should guarantee that tid and kmem_cache are retrieved on
2828 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2829 * to check if it is matched or not.
2832 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2833 c
= raw_cpu_ptr(s
->cpu_slab
);
2834 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
2835 unlikely(tid
!= READ_ONCE(c
->tid
)));
2838 * Irqless object alloc/free algorithm used here depends on sequence
2839 * of fetching cpu_slab's data. tid should be fetched before anything
2840 * on c to guarantee that object and page associated with previous tid
2841 * won't be used with current tid. If we fetch tid first, object and
2842 * page could be one associated with next tid and our alloc/free
2843 * request will be failed. In this case, we will retry. So, no problem.
2848 * The transaction ids are globally unique per cpu and per operation on
2849 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2850 * occurs on the right processor and that there was no operation on the
2851 * linked list in between.
2854 object
= c
->freelist
;
2856 if (unlikely(!object
|| !page
|| !node_match(page
, node
))) {
2857 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2859 void *next_object
= get_freepointer_safe(s
, object
);
2862 * The cmpxchg will only match if there was no additional
2863 * operation and if we are on the right processor.
2865 * The cmpxchg does the following atomically (without lock
2867 * 1. Relocate first pointer to the current per cpu area.
2868 * 2. Verify that tid and freelist have not been changed
2869 * 3. If they were not changed replace tid and freelist
2871 * Since this is without lock semantics the protection is only
2872 * against code executing on this cpu *not* from access by
2875 if (unlikely(!this_cpu_cmpxchg_double(
2876 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2878 next_object
, next_tid(tid
)))) {
2880 note_cmpxchg_failure("slab_alloc", s
, tid
);
2883 prefetch_freepointer(s
, next_object
);
2884 stat(s
, ALLOC_FASTPATH
);
2887 maybe_wipe_obj_freeptr(s
, object
);
2889 if (unlikely(slab_want_init_on_alloc(gfpflags
, s
)) && object
)
2890 memset(kasan_reset_tag(object
), 0, s
->object_size
);
2892 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
);
2897 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2898 gfp_t gfpflags
, unsigned long addr
)
2900 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2903 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2905 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2907 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2912 EXPORT_SYMBOL(kmem_cache_alloc
);
2914 #ifdef CONFIG_TRACING
2915 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2917 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2918 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2919 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2922 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2926 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2928 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2930 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2931 s
->object_size
, s
->size
, gfpflags
, node
);
2935 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2937 #ifdef CONFIG_TRACING
2938 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2940 int node
, size_t size
)
2942 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2944 trace_kmalloc_node(_RET_IP_
, ret
,
2945 size
, s
->size
, gfpflags
, node
);
2947 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2950 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2952 #endif /* CONFIG_NUMA */
2955 * Slow path handling. This may still be called frequently since objects
2956 * have a longer lifetime than the cpu slabs in most processing loads.
2958 * So we still attempt to reduce cache line usage. Just take the slab
2959 * lock and free the item. If there is no additional partial page
2960 * handling required then we can return immediately.
2962 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2963 void *head
, void *tail
, int cnt
,
2970 unsigned long counters
;
2971 struct kmem_cache_node
*n
= NULL
;
2972 unsigned long flags
;
2974 stat(s
, FREE_SLOWPATH
);
2976 if (kmem_cache_debug(s
) &&
2977 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2982 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2985 prior
= page
->freelist
;
2986 counters
= page
->counters
;
2987 set_freepointer(s
, tail
, prior
);
2988 new.counters
= counters
;
2989 was_frozen
= new.frozen
;
2991 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2993 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2996 * Slab was on no list before and will be
2998 * We can defer the list move and instead
3003 } else { /* Needs to be taken off a list */
3005 n
= get_node(s
, page_to_nid(page
));
3007 * Speculatively acquire the list_lock.
3008 * If the cmpxchg does not succeed then we may
3009 * drop the list_lock without any processing.
3011 * Otherwise the list_lock will synchronize with
3012 * other processors updating the list of slabs.
3014 spin_lock_irqsave(&n
->list_lock
, flags
);
3019 } while (!cmpxchg_double_slab(s
, page
,
3026 if (likely(was_frozen
)) {
3028 * The list lock was not taken therefore no list
3029 * activity can be necessary.
3031 stat(s
, FREE_FROZEN
);
3032 } else if (new.frozen
) {
3034 * If we just froze the page then put it onto the
3035 * per cpu partial list.
3037 put_cpu_partial(s
, page
, 1);
3038 stat(s
, CPU_PARTIAL_FREE
);
3044 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3048 * Objects left in the slab. If it was not on the partial list before
3051 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3052 remove_full(s
, n
, page
);
3053 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
3054 stat(s
, FREE_ADD_PARTIAL
);
3056 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3062 * Slab on the partial list.
3064 remove_partial(n
, page
);
3065 stat(s
, FREE_REMOVE_PARTIAL
);
3067 /* Slab must be on the full list */
3068 remove_full(s
, n
, page
);
3071 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3073 discard_slab(s
, page
);
3077 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3078 * can perform fastpath freeing without additional function calls.
3080 * The fastpath is only possible if we are freeing to the current cpu slab
3081 * of this processor. This typically the case if we have just allocated
3084 * If fastpath is not possible then fall back to __slab_free where we deal
3085 * with all sorts of special processing.
3087 * Bulk free of a freelist with several objects (all pointing to the
3088 * same page) possible by specifying head and tail ptr, plus objects
3089 * count (cnt). Bulk free indicated by tail pointer being set.
3091 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3092 struct page
*page
, void *head
, void *tail
,
3093 int cnt
, unsigned long addr
)
3095 void *tail_obj
= tail
? : head
;
3096 struct kmem_cache_cpu
*c
;
3099 memcg_slab_free_hook(s
, &head
, 1);
3102 * Determine the currently cpus per cpu slab.
3103 * The cpu may change afterward. However that does not matter since
3104 * data is retrieved via this pointer. If we are on the same cpu
3105 * during the cmpxchg then the free will succeed.
3108 tid
= this_cpu_read(s
->cpu_slab
->tid
);
3109 c
= raw_cpu_ptr(s
->cpu_slab
);
3110 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
3111 unlikely(tid
!= READ_ONCE(c
->tid
)));
3113 /* Same with comment on barrier() in slab_alloc_node() */
3116 if (likely(page
== c
->page
)) {
3117 void **freelist
= READ_ONCE(c
->freelist
);
3119 set_freepointer(s
, tail_obj
, freelist
);
3121 if (unlikely(!this_cpu_cmpxchg_double(
3122 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3124 head
, next_tid(tid
)))) {
3126 note_cmpxchg_failure("slab_free", s
, tid
);
3129 stat(s
, FREE_FASTPATH
);
3131 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3135 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3136 void *head
, void *tail
, int cnt
,
3140 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3141 * to remove objects, whose reuse must be delayed.
3143 if (slab_free_freelist_hook(s
, &head
, &tail
))
3144 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3147 #ifdef CONFIG_KASAN_GENERIC
3148 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3150 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3154 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3156 s
= cache_from_obj(s
, x
);
3159 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3160 trace_kmem_cache_free(_RET_IP_
, x
);
3162 EXPORT_SYMBOL(kmem_cache_free
);
3164 struct detached_freelist
{
3169 struct kmem_cache
*s
;
3173 * This function progressively scans the array with free objects (with
3174 * a limited look ahead) and extract objects belonging to the same
3175 * page. It builds a detached freelist directly within the given
3176 * page/objects. This can happen without any need for
3177 * synchronization, because the objects are owned by running process.
3178 * The freelist is build up as a single linked list in the objects.
3179 * The idea is, that this detached freelist can then be bulk
3180 * transferred to the real freelist(s), but only requiring a single
3181 * synchronization primitive. Look ahead in the array is limited due
3182 * to performance reasons.
3185 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3186 void **p
, struct detached_freelist
*df
)
3188 size_t first_skipped_index
= 0;
3193 /* Always re-init detached_freelist */
3198 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3199 } while (!object
&& size
);
3204 page
= virt_to_head_page(object
);
3206 /* Handle kalloc'ed objects */
3207 if (unlikely(!PageSlab(page
))) {
3208 BUG_ON(!PageCompound(page
));
3210 __free_pages(page
, compound_order(page
));
3211 p
[size
] = NULL
; /* mark object processed */
3214 /* Derive kmem_cache from object */
3215 df
->s
= page
->slab_cache
;
3217 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3220 /* Start new detached freelist */
3222 set_freepointer(df
->s
, object
, NULL
);
3224 df
->freelist
= object
;
3225 p
[size
] = NULL
; /* mark object processed */
3231 continue; /* Skip processed objects */
3233 /* df->page is always set at this point */
3234 if (df
->page
== virt_to_head_page(object
)) {
3235 /* Opportunity build freelist */
3236 set_freepointer(df
->s
, object
, df
->freelist
);
3237 df
->freelist
= object
;
3239 p
[size
] = NULL
; /* mark object processed */
3244 /* Limit look ahead search */
3248 if (!first_skipped_index
)
3249 first_skipped_index
= size
+ 1;
3252 return first_skipped_index
;
3255 /* Note that interrupts must be enabled when calling this function. */
3256 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3261 memcg_slab_free_hook(s
, p
, size
);
3263 struct detached_freelist df
;
3265 size
= build_detached_freelist(s
, size
, p
, &df
);
3269 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3270 } while (likely(size
));
3272 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3274 /* Note that interrupts must be enabled when calling this function. */
3275 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3278 struct kmem_cache_cpu
*c
;
3280 struct obj_cgroup
*objcg
= NULL
;
3282 /* memcg and kmem_cache debug support */
3283 s
= slab_pre_alloc_hook(s
, &objcg
, size
, flags
);
3287 * Drain objects in the per cpu slab, while disabling local
3288 * IRQs, which protects against PREEMPT and interrupts
3289 * handlers invoking normal fastpath.
3291 local_irq_disable();
3292 c
= this_cpu_ptr(s
->cpu_slab
);
3294 for (i
= 0; i
< size
; i
++) {
3295 void *object
= c
->freelist
;
3297 if (unlikely(!object
)) {
3299 * We may have removed an object from c->freelist using
3300 * the fastpath in the previous iteration; in that case,
3301 * c->tid has not been bumped yet.
3302 * Since ___slab_alloc() may reenable interrupts while
3303 * allocating memory, we should bump c->tid now.
3305 c
->tid
= next_tid(c
->tid
);
3308 * Invoking slow path likely have side-effect
3309 * of re-populating per CPU c->freelist
3311 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3313 if (unlikely(!p
[i
]))
3316 c
= this_cpu_ptr(s
->cpu_slab
);
3317 maybe_wipe_obj_freeptr(s
, p
[i
]);
3319 continue; /* goto for-loop */
3321 c
->freelist
= get_freepointer(s
, object
);
3323 maybe_wipe_obj_freeptr(s
, p
[i
]);
3325 c
->tid
= next_tid(c
->tid
);
3328 /* Clear memory outside IRQ disabled fastpath loop */
3329 if (unlikely(slab_want_init_on_alloc(flags
, s
))) {
3332 for (j
= 0; j
< i
; j
++)
3333 memset(kasan_reset_tag(p
[j
]), 0, s
->object_size
);
3336 /* memcg and kmem_cache debug support */
3337 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
);
3341 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
);
3342 __kmem_cache_free_bulk(s
, i
, p
);
3345 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3349 * Object placement in a slab is made very easy because we always start at
3350 * offset 0. If we tune the size of the object to the alignment then we can
3351 * get the required alignment by putting one properly sized object after
3354 * Notice that the allocation order determines the sizes of the per cpu
3355 * caches. Each processor has always one slab available for allocations.
3356 * Increasing the allocation order reduces the number of times that slabs
3357 * must be moved on and off the partial lists and is therefore a factor in
3362 * Mininum / Maximum order of slab pages. This influences locking overhead
3363 * and slab fragmentation. A higher order reduces the number of partial slabs
3364 * and increases the number of allocations possible without having to
3365 * take the list_lock.
3367 static unsigned int slub_min_order
;
3368 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3369 static unsigned int slub_min_objects
;
3372 * Calculate the order of allocation given an slab object size.
3374 * The order of allocation has significant impact on performance and other
3375 * system components. Generally order 0 allocations should be preferred since
3376 * order 0 does not cause fragmentation in the page allocator. Larger objects
3377 * be problematic to put into order 0 slabs because there may be too much
3378 * unused space left. We go to a higher order if more than 1/16th of the slab
3381 * In order to reach satisfactory performance we must ensure that a minimum
3382 * number of objects is in one slab. Otherwise we may generate too much
3383 * activity on the partial lists which requires taking the list_lock. This is
3384 * less a concern for large slabs though which are rarely used.
3386 * slub_max_order specifies the order where we begin to stop considering the
3387 * number of objects in a slab as critical. If we reach slub_max_order then
3388 * we try to keep the page order as low as possible. So we accept more waste
3389 * of space in favor of a small page order.
3391 * Higher order allocations also allow the placement of more objects in a
3392 * slab and thereby reduce object handling overhead. If the user has
3393 * requested a higher mininum order then we start with that one instead of
3394 * the smallest order which will fit the object.
3396 static inline unsigned int slab_order(unsigned int size
,
3397 unsigned int min_objects
, unsigned int max_order
,
3398 unsigned int fract_leftover
)
3400 unsigned int min_order
= slub_min_order
;
3403 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3404 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3406 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3407 order
<= max_order
; order
++) {
3409 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3412 rem
= slab_size
% size
;
3414 if (rem
<= slab_size
/ fract_leftover
)
3421 static inline int calculate_order(unsigned int size
)
3424 unsigned int min_objects
;
3425 unsigned int max_objects
;
3426 unsigned int nr_cpus
;
3429 * Attempt to find best configuration for a slab. This
3430 * works by first attempting to generate a layout with
3431 * the best configuration and backing off gradually.
3433 * First we increase the acceptable waste in a slab. Then
3434 * we reduce the minimum objects required in a slab.
3436 min_objects
= slub_min_objects
;
3439 * Some architectures will only update present cpus when
3440 * onlining them, so don't trust the number if it's just 1. But
3441 * we also don't want to use nr_cpu_ids always, as on some other
3442 * architectures, there can be many possible cpus, but never
3443 * onlined. Here we compromise between trying to avoid too high
3444 * order on systems that appear larger than they are, and too
3445 * low order on systems that appear smaller than they are.
3447 nr_cpus
= num_present_cpus();
3449 nr_cpus
= nr_cpu_ids
;
3450 min_objects
= 4 * (fls(nr_cpus
) + 1);
3452 max_objects
= order_objects(slub_max_order
, size
);
3453 min_objects
= min(min_objects
, max_objects
);
3455 while (min_objects
> 1) {
3456 unsigned int fraction
;
3459 while (fraction
>= 4) {
3460 order
= slab_order(size
, min_objects
,
3461 slub_max_order
, fraction
);
3462 if (order
<= slub_max_order
)
3470 * We were unable to place multiple objects in a slab. Now
3471 * lets see if we can place a single object there.
3473 order
= slab_order(size
, 1, slub_max_order
, 1);
3474 if (order
<= slub_max_order
)
3478 * Doh this slab cannot be placed using slub_max_order.
3480 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3481 if (order
< MAX_ORDER
)
3487 init_kmem_cache_node(struct kmem_cache_node
*n
)
3490 spin_lock_init(&n
->list_lock
);
3491 INIT_LIST_HEAD(&n
->partial
);
3492 #ifdef CONFIG_SLUB_DEBUG
3493 atomic_long_set(&n
->nr_slabs
, 0);
3494 atomic_long_set(&n
->total_objects
, 0);
3495 INIT_LIST_HEAD(&n
->full
);
3499 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3501 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3502 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3505 * Must align to double word boundary for the double cmpxchg
3506 * instructions to work; see __pcpu_double_call_return_bool().
3508 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3509 2 * sizeof(void *));
3514 init_kmem_cache_cpus(s
);
3519 static struct kmem_cache
*kmem_cache_node
;
3522 * No kmalloc_node yet so do it by hand. We know that this is the first
3523 * slab on the node for this slabcache. There are no concurrent accesses
3526 * Note that this function only works on the kmem_cache_node
3527 * when allocating for the kmem_cache_node. This is used for bootstrapping
3528 * memory on a fresh node that has no slab structures yet.
3530 static void early_kmem_cache_node_alloc(int node
)
3533 struct kmem_cache_node
*n
;
3535 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3537 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3540 if (page_to_nid(page
) != node
) {
3541 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3542 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3547 #ifdef CONFIG_SLUB_DEBUG
3548 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3549 init_tracking(kmem_cache_node
, n
);
3551 n
= kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3553 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3556 kmem_cache_node
->node
[node
] = n
;
3557 init_kmem_cache_node(n
);
3558 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3561 * No locks need to be taken here as it has just been
3562 * initialized and there is no concurrent access.
3564 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3567 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3570 struct kmem_cache_node
*n
;
3572 for_each_kmem_cache_node(s
, node
, n
) {
3573 s
->node
[node
] = NULL
;
3574 kmem_cache_free(kmem_cache_node
, n
);
3578 void __kmem_cache_release(struct kmem_cache
*s
)
3580 cache_random_seq_destroy(s
);
3581 free_percpu(s
->cpu_slab
);
3582 free_kmem_cache_nodes(s
);
3585 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3589 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3590 struct kmem_cache_node
*n
;
3592 if (slab_state
== DOWN
) {
3593 early_kmem_cache_node_alloc(node
);
3596 n
= kmem_cache_alloc_node(kmem_cache_node
,
3600 free_kmem_cache_nodes(s
);
3604 init_kmem_cache_node(n
);
3610 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3612 if (min
< MIN_PARTIAL
)
3614 else if (min
> MAX_PARTIAL
)
3616 s
->min_partial
= min
;
3619 static void set_cpu_partial(struct kmem_cache
*s
)
3621 #ifdef CONFIG_SLUB_CPU_PARTIAL
3623 * cpu_partial determined the maximum number of objects kept in the
3624 * per cpu partial lists of a processor.
3626 * Per cpu partial lists mainly contain slabs that just have one
3627 * object freed. If they are used for allocation then they can be
3628 * filled up again with minimal effort. The slab will never hit the
3629 * per node partial lists and therefore no locking will be required.
3631 * This setting also determines
3633 * A) The number of objects from per cpu partial slabs dumped to the
3634 * per node list when we reach the limit.
3635 * B) The number of objects in cpu partial slabs to extract from the
3636 * per node list when we run out of per cpu objects. We only fetch
3637 * 50% to keep some capacity around for frees.
3639 if (!kmem_cache_has_cpu_partial(s
))
3640 slub_set_cpu_partial(s
, 0);
3641 else if (s
->size
>= PAGE_SIZE
)
3642 slub_set_cpu_partial(s
, 2);
3643 else if (s
->size
>= 1024)
3644 slub_set_cpu_partial(s
, 6);
3645 else if (s
->size
>= 256)
3646 slub_set_cpu_partial(s
, 13);
3648 slub_set_cpu_partial(s
, 30);
3653 * calculate_sizes() determines the order and the distribution of data within
3656 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3658 slab_flags_t flags
= s
->flags
;
3659 unsigned int size
= s
->object_size
;
3660 unsigned int freepointer_area
;
3664 * Round up object size to the next word boundary. We can only
3665 * place the free pointer at word boundaries and this determines
3666 * the possible location of the free pointer.
3668 size
= ALIGN(size
, sizeof(void *));
3670 * This is the area of the object where a freepointer can be
3671 * safely written. If redzoning adds more to the inuse size, we
3672 * can't use that portion for writing the freepointer, so
3673 * s->offset must be limited within this for the general case.
3675 freepointer_area
= size
;
3677 #ifdef CONFIG_SLUB_DEBUG
3679 * Determine if we can poison the object itself. If the user of
3680 * the slab may touch the object after free or before allocation
3681 * then we should never poison the object itself.
3683 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3685 s
->flags
|= __OBJECT_POISON
;
3687 s
->flags
&= ~__OBJECT_POISON
;
3691 * If we are Redzoning then check if there is some space between the
3692 * end of the object and the free pointer. If not then add an
3693 * additional word to have some bytes to store Redzone information.
3695 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3696 size
+= sizeof(void *);
3700 * With that we have determined the number of bytes in actual use
3701 * by the object. This is the potential offset to the free pointer.
3705 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3708 * Relocate free pointer after the object if it is not
3709 * permitted to overwrite the first word of the object on
3712 * This is the case if we do RCU, have a constructor or
3713 * destructor or are poisoning the objects.
3715 * The assumption that s->offset >= s->inuse means free
3716 * pointer is outside of the object is used in the
3717 * freeptr_outside_object() function. If that is no
3718 * longer true, the function needs to be modified.
3721 size
+= sizeof(void *);
3722 } else if (freepointer_area
> sizeof(void *)) {
3724 * Store freelist pointer near middle of object to keep
3725 * it away from the edges of the object to avoid small
3726 * sized over/underflows from neighboring allocations.
3728 s
->offset
= ALIGN(freepointer_area
/ 2, sizeof(void *));
3731 #ifdef CONFIG_SLUB_DEBUG
3732 if (flags
& SLAB_STORE_USER
)
3734 * Need to store information about allocs and frees after
3737 size
+= 2 * sizeof(struct track
);
3740 kasan_cache_create(s
, &size
, &s
->flags
);
3741 #ifdef CONFIG_SLUB_DEBUG
3742 if (flags
& SLAB_RED_ZONE
) {
3744 * Add some empty padding so that we can catch
3745 * overwrites from earlier objects rather than let
3746 * tracking information or the free pointer be
3747 * corrupted if a user writes before the start
3750 size
+= sizeof(void *);
3752 s
->red_left_pad
= sizeof(void *);
3753 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3754 size
+= s
->red_left_pad
;
3759 * SLUB stores one object immediately after another beginning from
3760 * offset 0. In order to align the objects we have to simply size
3761 * each object to conform to the alignment.
3763 size
= ALIGN(size
, s
->align
);
3765 s
->reciprocal_size
= reciprocal_value(size
);
3766 if (forced_order
>= 0)
3767 order
= forced_order
;
3769 order
= calculate_order(size
);
3776 s
->allocflags
|= __GFP_COMP
;
3778 if (s
->flags
& SLAB_CACHE_DMA
)
3779 s
->allocflags
|= GFP_DMA
;
3781 if (s
->flags
& SLAB_CACHE_DMA32
)
3782 s
->allocflags
|= GFP_DMA32
;
3784 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3785 s
->allocflags
|= __GFP_RECLAIMABLE
;
3788 * Determine the number of objects per slab
3790 s
->oo
= oo_make(order
, size
);
3791 s
->min
= oo_make(get_order(size
), size
);
3792 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3795 return !!oo_objects(s
->oo
);
3798 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3800 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3801 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3802 s
->random
= get_random_long();
3805 if (!calculate_sizes(s
, -1))
3807 if (disable_higher_order_debug
) {
3809 * Disable debugging flags that store metadata if the min slab
3812 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3813 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3815 if (!calculate_sizes(s
, -1))
3820 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3821 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3822 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3823 /* Enable fast mode */
3824 s
->flags
|= __CMPXCHG_DOUBLE
;
3828 * The larger the object size is, the more pages we want on the partial
3829 * list to avoid pounding the page allocator excessively.
3831 set_min_partial(s
, ilog2(s
->size
) / 2);
3836 s
->remote_node_defrag_ratio
= 1000;
3839 /* Initialize the pre-computed randomized freelist if slab is up */
3840 if (slab_state
>= UP
) {
3841 if (init_cache_random_seq(s
))
3845 if (!init_kmem_cache_nodes(s
))
3848 if (alloc_kmem_cache_cpus(s
))
3851 free_kmem_cache_nodes(s
);
3856 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3859 #ifdef CONFIG_SLUB_DEBUG
3860 void *addr
= page_address(page
);
3864 slab_err(s
, page
, text
, s
->name
);
3867 map
= get_map(s
, page
);
3868 for_each_object(p
, s
, addr
, page
->objects
) {
3870 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
3871 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3872 print_tracking(s
, p
);
3881 * Attempt to free all partial slabs on a node.
3882 * This is called from __kmem_cache_shutdown(). We must take list_lock
3883 * because sysfs file might still access partial list after the shutdowning.
3885 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3888 struct page
*page
, *h
;
3890 BUG_ON(irqs_disabled());
3891 spin_lock_irq(&n
->list_lock
);
3892 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3894 remove_partial(n
, page
);
3895 list_add(&page
->slab_list
, &discard
);
3897 list_slab_objects(s
, page
,
3898 "Objects remaining in %s on __kmem_cache_shutdown()");
3901 spin_unlock_irq(&n
->list_lock
);
3903 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3904 discard_slab(s
, page
);
3907 bool __kmem_cache_empty(struct kmem_cache
*s
)
3910 struct kmem_cache_node
*n
;
3912 for_each_kmem_cache_node(s
, node
, n
)
3913 if (n
->nr_partial
|| slabs_node(s
, node
))
3919 * Release all resources used by a slab cache.
3921 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3924 struct kmem_cache_node
*n
;
3927 /* Attempt to free all objects */
3928 for_each_kmem_cache_node(s
, node
, n
) {
3930 if (n
->nr_partial
|| slabs_node(s
, node
))
3936 /********************************************************************
3938 *******************************************************************/
3940 static int __init
setup_slub_min_order(char *str
)
3942 get_option(&str
, (int *)&slub_min_order
);
3947 __setup("slub_min_order=", setup_slub_min_order
);
3949 static int __init
setup_slub_max_order(char *str
)
3951 get_option(&str
, (int *)&slub_max_order
);
3952 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
3957 __setup("slub_max_order=", setup_slub_max_order
);
3959 static int __init
setup_slub_min_objects(char *str
)
3961 get_option(&str
, (int *)&slub_min_objects
);
3966 __setup("slub_min_objects=", setup_slub_min_objects
);
3968 void *__kmalloc(size_t size
, gfp_t flags
)
3970 struct kmem_cache
*s
;
3973 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3974 return kmalloc_large(size
, flags
);
3976 s
= kmalloc_slab(size
, flags
);
3978 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3981 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3983 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3985 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3989 EXPORT_SYMBOL(__kmalloc
);
3992 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3996 unsigned int order
= get_order(size
);
3998 flags
|= __GFP_COMP
;
3999 page
= alloc_pages_node(node
, flags
, order
);
4001 ptr
= page_address(page
);
4002 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4003 PAGE_SIZE
<< order
);
4006 return kmalloc_large_node_hook(ptr
, size
, flags
);
4009 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4011 struct kmem_cache
*s
;
4014 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4015 ret
= kmalloc_large_node(size
, flags
, node
);
4017 trace_kmalloc_node(_RET_IP_
, ret
,
4018 size
, PAGE_SIZE
<< get_order(size
),
4024 s
= kmalloc_slab(size
, flags
);
4026 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4029 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
4031 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
4033 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4037 EXPORT_SYMBOL(__kmalloc_node
);
4038 #endif /* CONFIG_NUMA */
4040 #ifdef CONFIG_HARDENED_USERCOPY
4042 * Rejects incorrectly sized objects and objects that are to be copied
4043 * to/from userspace but do not fall entirely within the containing slab
4044 * cache's usercopy region.
4046 * Returns NULL if check passes, otherwise const char * to name of cache
4047 * to indicate an error.
4049 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4052 struct kmem_cache
*s
;
4053 unsigned int offset
;
4056 ptr
= kasan_reset_tag(ptr
);
4058 /* Find object and usable object size. */
4059 s
= page
->slab_cache
;
4061 /* Reject impossible pointers. */
4062 if (ptr
< page_address(page
))
4063 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4066 /* Find offset within object. */
4067 offset
= (ptr
- page_address(page
)) % s
->size
;
4069 /* Adjust for redzone and reject if within the redzone. */
4070 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4071 if (offset
< s
->red_left_pad
)
4072 usercopy_abort("SLUB object in left red zone",
4073 s
->name
, to_user
, offset
, n
);
4074 offset
-= s
->red_left_pad
;
4077 /* Allow address range falling entirely within usercopy region. */
4078 if (offset
>= s
->useroffset
&&
4079 offset
- s
->useroffset
<= s
->usersize
&&
4080 n
<= s
->useroffset
- offset
+ s
->usersize
)
4084 * If the copy is still within the allocated object, produce
4085 * a warning instead of rejecting the copy. This is intended
4086 * to be a temporary method to find any missing usercopy
4089 object_size
= slab_ksize(s
);
4090 if (usercopy_fallback
&&
4091 offset
<= object_size
&& n
<= object_size
- offset
) {
4092 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4096 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4098 #endif /* CONFIG_HARDENED_USERCOPY */
4100 size_t __ksize(const void *object
)
4104 if (unlikely(object
== ZERO_SIZE_PTR
))
4107 page
= virt_to_head_page(object
);
4109 if (unlikely(!PageSlab(page
))) {
4110 WARN_ON(!PageCompound(page
));
4111 return page_size(page
);
4114 return slab_ksize(page
->slab_cache
);
4116 EXPORT_SYMBOL(__ksize
);
4118 void kfree(const void *x
)
4121 void *object
= (void *)x
;
4123 trace_kfree(_RET_IP_
, x
);
4125 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4128 page
= virt_to_head_page(x
);
4129 if (unlikely(!PageSlab(page
))) {
4130 unsigned int order
= compound_order(page
);
4132 BUG_ON(!PageCompound(page
));
4134 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4135 -(PAGE_SIZE
<< order
));
4136 __free_pages(page
, order
);
4139 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4141 EXPORT_SYMBOL(kfree
);
4143 #define SHRINK_PROMOTE_MAX 32
4146 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4147 * up most to the head of the partial lists. New allocations will then
4148 * fill those up and thus they can be removed from the partial lists.
4150 * The slabs with the least items are placed last. This results in them
4151 * being allocated from last increasing the chance that the last objects
4152 * are freed in them.
4154 int __kmem_cache_shrink(struct kmem_cache
*s
)
4158 struct kmem_cache_node
*n
;
4161 struct list_head discard
;
4162 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4163 unsigned long flags
;
4167 for_each_kmem_cache_node(s
, node
, n
) {
4168 INIT_LIST_HEAD(&discard
);
4169 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4170 INIT_LIST_HEAD(promote
+ i
);
4172 spin_lock_irqsave(&n
->list_lock
, flags
);
4175 * Build lists of slabs to discard or promote.
4177 * Note that concurrent frees may occur while we hold the
4178 * list_lock. page->inuse here is the upper limit.
4180 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4181 int free
= page
->objects
- page
->inuse
;
4183 /* Do not reread page->inuse */
4186 /* We do not keep full slabs on the list */
4189 if (free
== page
->objects
) {
4190 list_move(&page
->slab_list
, &discard
);
4192 } else if (free
<= SHRINK_PROMOTE_MAX
)
4193 list_move(&page
->slab_list
, promote
+ free
- 1);
4197 * Promote the slabs filled up most to the head of the
4200 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4201 list_splice(promote
+ i
, &n
->partial
);
4203 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4205 /* Release empty slabs */
4206 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4207 discard_slab(s
, page
);
4209 if (slabs_node(s
, node
))
4216 static int slab_mem_going_offline_callback(void *arg
)
4218 struct kmem_cache
*s
;
4220 mutex_lock(&slab_mutex
);
4221 list_for_each_entry(s
, &slab_caches
, list
)
4222 __kmem_cache_shrink(s
);
4223 mutex_unlock(&slab_mutex
);
4228 static void slab_mem_offline_callback(void *arg
)
4230 struct kmem_cache_node
*n
;
4231 struct kmem_cache
*s
;
4232 struct memory_notify
*marg
= arg
;
4235 offline_node
= marg
->status_change_nid_normal
;
4238 * If the node still has available memory. we need kmem_cache_node
4241 if (offline_node
< 0)
4244 mutex_lock(&slab_mutex
);
4245 list_for_each_entry(s
, &slab_caches
, list
) {
4246 n
= get_node(s
, offline_node
);
4249 * if n->nr_slabs > 0, slabs still exist on the node
4250 * that is going down. We were unable to free them,
4251 * and offline_pages() function shouldn't call this
4252 * callback. So, we must fail.
4254 BUG_ON(slabs_node(s
, offline_node
));
4256 s
->node
[offline_node
] = NULL
;
4257 kmem_cache_free(kmem_cache_node
, n
);
4260 mutex_unlock(&slab_mutex
);
4263 static int slab_mem_going_online_callback(void *arg
)
4265 struct kmem_cache_node
*n
;
4266 struct kmem_cache
*s
;
4267 struct memory_notify
*marg
= arg
;
4268 int nid
= marg
->status_change_nid_normal
;
4272 * If the node's memory is already available, then kmem_cache_node is
4273 * already created. Nothing to do.
4279 * We are bringing a node online. No memory is available yet. We must
4280 * allocate a kmem_cache_node structure in order to bring the node
4283 mutex_lock(&slab_mutex
);
4284 list_for_each_entry(s
, &slab_caches
, list
) {
4286 * XXX: kmem_cache_alloc_node will fallback to other nodes
4287 * since memory is not yet available from the node that
4290 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4295 init_kmem_cache_node(n
);
4299 mutex_unlock(&slab_mutex
);
4303 static int slab_memory_callback(struct notifier_block
*self
,
4304 unsigned long action
, void *arg
)
4309 case MEM_GOING_ONLINE
:
4310 ret
= slab_mem_going_online_callback(arg
);
4312 case MEM_GOING_OFFLINE
:
4313 ret
= slab_mem_going_offline_callback(arg
);
4316 case MEM_CANCEL_ONLINE
:
4317 slab_mem_offline_callback(arg
);
4320 case MEM_CANCEL_OFFLINE
:
4324 ret
= notifier_from_errno(ret
);
4330 static struct notifier_block slab_memory_callback_nb
= {
4331 .notifier_call
= slab_memory_callback
,
4332 .priority
= SLAB_CALLBACK_PRI
,
4335 /********************************************************************
4336 * Basic setup of slabs
4337 *******************************************************************/
4340 * Used for early kmem_cache structures that were allocated using
4341 * the page allocator. Allocate them properly then fix up the pointers
4342 * that may be pointing to the wrong kmem_cache structure.
4345 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4348 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4349 struct kmem_cache_node
*n
;
4351 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4354 * This runs very early, and only the boot processor is supposed to be
4355 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4358 __flush_cpu_slab(s
, smp_processor_id());
4359 for_each_kmem_cache_node(s
, node
, n
) {
4362 list_for_each_entry(p
, &n
->partial
, slab_list
)
4365 #ifdef CONFIG_SLUB_DEBUG
4366 list_for_each_entry(p
, &n
->full
, slab_list
)
4370 list_add(&s
->list
, &slab_caches
);
4374 void __init
kmem_cache_init(void)
4376 static __initdata
struct kmem_cache boot_kmem_cache
,
4377 boot_kmem_cache_node
;
4379 if (debug_guardpage_minorder())
4382 kmem_cache_node
= &boot_kmem_cache_node
;
4383 kmem_cache
= &boot_kmem_cache
;
4385 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4386 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4388 register_hotmemory_notifier(&slab_memory_callback_nb
);
4390 /* Able to allocate the per node structures */
4391 slab_state
= PARTIAL
;
4393 create_boot_cache(kmem_cache
, "kmem_cache",
4394 offsetof(struct kmem_cache
, node
) +
4395 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4396 SLAB_HWCACHE_ALIGN
, 0, 0);
4398 kmem_cache
= bootstrap(&boot_kmem_cache
);
4399 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4401 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4402 setup_kmalloc_cache_index_table();
4403 create_kmalloc_caches(0);
4405 /* Setup random freelists for each cache */
4406 init_freelist_randomization();
4408 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4411 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4413 slub_min_order
, slub_max_order
, slub_min_objects
,
4414 nr_cpu_ids
, nr_node_ids
);
4417 void __init
kmem_cache_init_late(void)
4422 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4423 slab_flags_t flags
, void (*ctor
)(void *))
4425 struct kmem_cache
*s
;
4427 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4432 * Adjust the object sizes so that we clear
4433 * the complete object on kzalloc.
4435 s
->object_size
= max(s
->object_size
, size
);
4436 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4438 if (sysfs_slab_alias(s
, name
)) {
4447 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4451 err
= kmem_cache_open(s
, flags
);
4455 /* Mutex is not taken during early boot */
4456 if (slab_state
<= UP
)
4459 err
= sysfs_slab_add(s
);
4461 __kmem_cache_release(s
);
4466 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4468 struct kmem_cache
*s
;
4471 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4472 return kmalloc_large(size
, gfpflags
);
4474 s
= kmalloc_slab(size
, gfpflags
);
4476 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4479 ret
= slab_alloc(s
, gfpflags
, caller
);
4481 /* Honor the call site pointer we received. */
4482 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4486 EXPORT_SYMBOL(__kmalloc_track_caller
);
4489 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4490 int node
, unsigned long caller
)
4492 struct kmem_cache
*s
;
4495 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4496 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4498 trace_kmalloc_node(caller
, ret
,
4499 size
, PAGE_SIZE
<< get_order(size
),
4505 s
= kmalloc_slab(size
, gfpflags
);
4507 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4510 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4512 /* Honor the call site pointer we received. */
4513 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4517 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4521 static int count_inuse(struct page
*page
)
4526 static int count_total(struct page
*page
)
4528 return page
->objects
;
4532 #ifdef CONFIG_SLUB_DEBUG
4533 static void validate_slab(struct kmem_cache
*s
, struct page
*page
)
4536 void *addr
= page_address(page
);
4541 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4544 /* Now we know that a valid freelist exists */
4545 map
= get_map(s
, page
);
4546 for_each_object(p
, s
, addr
, page
->objects
) {
4547 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), map
) ?
4548 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4550 if (!check_object(s
, page
, p
, val
))
4558 static int validate_slab_node(struct kmem_cache
*s
,
4559 struct kmem_cache_node
*n
)
4561 unsigned long count
= 0;
4563 unsigned long flags
;
4565 spin_lock_irqsave(&n
->list_lock
, flags
);
4567 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4568 validate_slab(s
, page
);
4571 if (count
!= n
->nr_partial
)
4572 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4573 s
->name
, count
, n
->nr_partial
);
4575 if (!(s
->flags
& SLAB_STORE_USER
))
4578 list_for_each_entry(page
, &n
->full
, slab_list
) {
4579 validate_slab(s
, page
);
4582 if (count
!= atomic_long_read(&n
->nr_slabs
))
4583 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4584 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4587 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4591 static long validate_slab_cache(struct kmem_cache
*s
)
4594 unsigned long count
= 0;
4595 struct kmem_cache_node
*n
;
4598 for_each_kmem_cache_node(s
, node
, n
)
4599 count
+= validate_slab_node(s
, n
);
4604 * Generate lists of code addresses where slabcache objects are allocated
4609 unsigned long count
;
4616 DECLARE_BITMAP(cpus
, NR_CPUS
);
4622 unsigned long count
;
4623 struct location
*loc
;
4626 static void free_loc_track(struct loc_track
*t
)
4629 free_pages((unsigned long)t
->loc
,
4630 get_order(sizeof(struct location
) * t
->max
));
4633 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4638 order
= get_order(sizeof(struct location
) * max
);
4640 l
= (void *)__get_free_pages(flags
, order
);
4645 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4653 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4654 const struct track
*track
)
4656 long start
, end
, pos
;
4658 unsigned long caddr
;
4659 unsigned long age
= jiffies
- track
->when
;
4665 pos
= start
+ (end
- start
+ 1) / 2;
4668 * There is nothing at "end". If we end up there
4669 * we need to add something to before end.
4674 caddr
= t
->loc
[pos
].addr
;
4675 if (track
->addr
== caddr
) {
4681 if (age
< l
->min_time
)
4683 if (age
> l
->max_time
)
4686 if (track
->pid
< l
->min_pid
)
4687 l
->min_pid
= track
->pid
;
4688 if (track
->pid
> l
->max_pid
)
4689 l
->max_pid
= track
->pid
;
4691 cpumask_set_cpu(track
->cpu
,
4692 to_cpumask(l
->cpus
));
4694 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4698 if (track
->addr
< caddr
)
4705 * Not found. Insert new tracking element.
4707 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4713 (t
->count
- pos
) * sizeof(struct location
));
4716 l
->addr
= track
->addr
;
4720 l
->min_pid
= track
->pid
;
4721 l
->max_pid
= track
->pid
;
4722 cpumask_clear(to_cpumask(l
->cpus
));
4723 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4724 nodes_clear(l
->nodes
);
4725 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4729 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4730 struct page
*page
, enum track_item alloc
)
4732 void *addr
= page_address(page
);
4736 map
= get_map(s
, page
);
4737 for_each_object(p
, s
, addr
, page
->objects
)
4738 if (!test_bit(__obj_to_index(s
, addr
, p
), map
))
4739 add_location(t
, s
, get_track(s
, p
, alloc
));
4743 static int list_locations(struct kmem_cache
*s
, char *buf
,
4744 enum track_item alloc
)
4748 struct loc_track t
= { 0, 0, NULL
};
4750 struct kmem_cache_node
*n
;
4752 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4754 return sysfs_emit(buf
, "Out of memory\n");
4756 /* Push back cpu slabs */
4759 for_each_kmem_cache_node(s
, node
, n
) {
4760 unsigned long flags
;
4763 if (!atomic_long_read(&n
->nr_slabs
))
4766 spin_lock_irqsave(&n
->list_lock
, flags
);
4767 list_for_each_entry(page
, &n
->partial
, slab_list
)
4768 process_slab(&t
, s
, page
, alloc
);
4769 list_for_each_entry(page
, &n
->full
, slab_list
)
4770 process_slab(&t
, s
, page
, alloc
);
4771 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4774 for (i
= 0; i
< t
.count
; i
++) {
4775 struct location
*l
= &t
.loc
[i
];
4777 len
+= sysfs_emit_at(buf
, len
, "%7ld ", l
->count
);
4780 len
+= sysfs_emit_at(buf
, len
, "%pS", (void *)l
->addr
);
4782 len
+= sysfs_emit_at(buf
, len
, "<not-available>");
4784 if (l
->sum_time
!= l
->min_time
)
4785 len
+= sysfs_emit_at(buf
, len
, " age=%ld/%ld/%ld",
4787 (long)div_u64(l
->sum_time
,
4791 len
+= sysfs_emit_at(buf
, len
, " age=%ld", l
->min_time
);
4793 if (l
->min_pid
!= l
->max_pid
)
4794 len
+= sysfs_emit_at(buf
, len
, " pid=%ld-%ld",
4795 l
->min_pid
, l
->max_pid
);
4797 len
+= sysfs_emit_at(buf
, len
, " pid=%ld",
4800 if (num_online_cpus() > 1 &&
4801 !cpumask_empty(to_cpumask(l
->cpus
)))
4802 len
+= sysfs_emit_at(buf
, len
, " cpus=%*pbl",
4803 cpumask_pr_args(to_cpumask(l
->cpus
)));
4805 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
4806 len
+= sysfs_emit_at(buf
, len
, " nodes=%*pbl",
4807 nodemask_pr_args(&l
->nodes
));
4809 len
+= sysfs_emit_at(buf
, len
, "\n");
4814 len
+= sysfs_emit_at(buf
, len
, "No data\n");
4818 #endif /* CONFIG_SLUB_DEBUG */
4820 #ifdef SLUB_RESILIENCY_TEST
4821 static void __init
resiliency_test(void)
4824 int type
= KMALLOC_NORMAL
;
4826 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4828 pr_err("SLUB resiliency testing\n");
4829 pr_err("-----------------------\n");
4830 pr_err("A. Corruption after allocation\n");
4832 p
= kzalloc(16, GFP_KERNEL
);
4834 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4837 validate_slab_cache(kmalloc_caches
[type
][4]);
4839 /* Hmmm... The next two are dangerous */
4840 p
= kzalloc(32, GFP_KERNEL
);
4841 p
[32 + sizeof(void *)] = 0x34;
4842 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4844 pr_err("If allocated object is overwritten then not detectable\n\n");
4846 validate_slab_cache(kmalloc_caches
[type
][5]);
4847 p
= kzalloc(64, GFP_KERNEL
);
4848 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4850 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4852 pr_err("If allocated object is overwritten then not detectable\n\n");
4853 validate_slab_cache(kmalloc_caches
[type
][6]);
4855 pr_err("\nB. Corruption after free\n");
4856 p
= kzalloc(128, GFP_KERNEL
);
4859 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4860 validate_slab_cache(kmalloc_caches
[type
][7]);
4862 p
= kzalloc(256, GFP_KERNEL
);
4865 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4866 validate_slab_cache(kmalloc_caches
[type
][8]);
4868 p
= kzalloc(512, GFP_KERNEL
);
4871 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4872 validate_slab_cache(kmalloc_caches
[type
][9]);
4876 static void resiliency_test(void) {};
4878 #endif /* SLUB_RESILIENCY_TEST */
4881 enum slab_stat_type
{
4882 SL_ALL
, /* All slabs */
4883 SL_PARTIAL
, /* Only partially allocated slabs */
4884 SL_CPU
, /* Only slabs used for cpu caches */
4885 SL_OBJECTS
, /* Determine allocated objects not slabs */
4886 SL_TOTAL
/* Determine object capacity not slabs */
4889 #define SO_ALL (1 << SL_ALL)
4890 #define SO_PARTIAL (1 << SL_PARTIAL)
4891 #define SO_CPU (1 << SL_CPU)
4892 #define SO_OBJECTS (1 << SL_OBJECTS)
4893 #define SO_TOTAL (1 << SL_TOTAL)
4896 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4898 static int __init
setup_slub_memcg_sysfs(char *str
)
4902 if (get_option(&str
, &v
) > 0)
4903 memcg_sysfs_enabled
= v
;
4908 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4911 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4912 char *buf
, unsigned long flags
)
4914 unsigned long total
= 0;
4917 unsigned long *nodes
;
4920 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4924 if (flags
& SO_CPU
) {
4927 for_each_possible_cpu(cpu
) {
4928 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4933 page
= READ_ONCE(c
->page
);
4937 node
= page_to_nid(page
);
4938 if (flags
& SO_TOTAL
)
4940 else if (flags
& SO_OBJECTS
)
4948 page
= slub_percpu_partial_read_once(c
);
4950 node
= page_to_nid(page
);
4951 if (flags
& SO_TOTAL
)
4953 else if (flags
& SO_OBJECTS
)
4964 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4965 * already held which will conflict with an existing lock order:
4967 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4969 * We don't really need mem_hotplug_lock (to hold off
4970 * slab_mem_going_offline_callback) here because slab's memory hot
4971 * unplug code doesn't destroy the kmem_cache->node[] data.
4974 #ifdef CONFIG_SLUB_DEBUG
4975 if (flags
& SO_ALL
) {
4976 struct kmem_cache_node
*n
;
4978 for_each_kmem_cache_node(s
, node
, n
) {
4980 if (flags
& SO_TOTAL
)
4981 x
= atomic_long_read(&n
->total_objects
);
4982 else if (flags
& SO_OBJECTS
)
4983 x
= atomic_long_read(&n
->total_objects
) -
4984 count_partial(n
, count_free
);
4986 x
= atomic_long_read(&n
->nr_slabs
);
4993 if (flags
& SO_PARTIAL
) {
4994 struct kmem_cache_node
*n
;
4996 for_each_kmem_cache_node(s
, node
, n
) {
4997 if (flags
& SO_TOTAL
)
4998 x
= count_partial(n
, count_total
);
4999 else if (flags
& SO_OBJECTS
)
5000 x
= count_partial(n
, count_inuse
);
5008 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5010 for (node
= 0; node
< nr_node_ids
; node
++) {
5012 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5016 len
+= sysfs_emit_at(buf
, len
, "\n");
5022 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5023 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5025 struct slab_attribute
{
5026 struct attribute attr
;
5027 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5028 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5031 #define SLAB_ATTR_RO(_name) \
5032 static struct slab_attribute _name##_attr = \
5033 __ATTR(_name, 0400, _name##_show, NULL)
5035 #define SLAB_ATTR(_name) \
5036 static struct slab_attribute _name##_attr = \
5037 __ATTR(_name, 0600, _name##_show, _name##_store)
5039 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5041 return sysfs_emit(buf
, "%u\n", s
->size
);
5043 SLAB_ATTR_RO(slab_size
);
5045 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5047 return sysfs_emit(buf
, "%u\n", s
->align
);
5049 SLAB_ATTR_RO(align
);
5051 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5053 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5055 SLAB_ATTR_RO(object_size
);
5057 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5059 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5061 SLAB_ATTR_RO(objs_per_slab
);
5063 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5065 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5067 SLAB_ATTR_RO(order
);
5069 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5071 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5074 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5080 err
= kstrtoul(buf
, 10, &min
);
5084 set_min_partial(s
, min
);
5087 SLAB_ATTR(min_partial
);
5089 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5091 return sysfs_emit(buf
, "%u\n", slub_cpu_partial(s
));
5094 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5097 unsigned int objects
;
5100 err
= kstrtouint(buf
, 10, &objects
);
5103 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5106 slub_set_cpu_partial(s
, objects
);
5110 SLAB_ATTR(cpu_partial
);
5112 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5116 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5120 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5122 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5124 SLAB_ATTR_RO(aliases
);
5126 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5128 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5130 SLAB_ATTR_RO(partial
);
5132 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5134 return show_slab_objects(s
, buf
, SO_CPU
);
5136 SLAB_ATTR_RO(cpu_slabs
);
5138 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5140 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5142 SLAB_ATTR_RO(objects
);
5144 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5146 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5148 SLAB_ATTR_RO(objects_partial
);
5150 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5157 for_each_online_cpu(cpu
) {
5160 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5163 pages
+= page
->pages
;
5164 objects
+= page
->pobjects
;
5168 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, pages
);
5171 for_each_online_cpu(cpu
) {
5174 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5176 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5177 cpu
, page
->pobjects
, page
->pages
);
5180 len
+= sysfs_emit_at(buf
, len
, "\n");
5184 SLAB_ATTR_RO(slabs_cpu_partial
);
5186 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5188 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5190 SLAB_ATTR_RO(reclaim_account
);
5192 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5194 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5196 SLAB_ATTR_RO(hwcache_align
);
5198 #ifdef CONFIG_ZONE_DMA
5199 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5201 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5203 SLAB_ATTR_RO(cache_dma
);
5206 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5208 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5210 SLAB_ATTR_RO(usersize
);
5212 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5214 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5216 SLAB_ATTR_RO(destroy_by_rcu
);
5218 #ifdef CONFIG_SLUB_DEBUG
5219 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5221 return show_slab_objects(s
, buf
, SO_ALL
);
5223 SLAB_ATTR_RO(slabs
);
5225 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5227 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5229 SLAB_ATTR_RO(total_objects
);
5231 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5233 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5235 SLAB_ATTR_RO(sanity_checks
);
5237 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5239 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5241 SLAB_ATTR_RO(trace
);
5243 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5245 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5248 SLAB_ATTR_RO(red_zone
);
5250 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5252 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5255 SLAB_ATTR_RO(poison
);
5257 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5259 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5262 SLAB_ATTR_RO(store_user
);
5264 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5269 static ssize_t
validate_store(struct kmem_cache
*s
,
5270 const char *buf
, size_t length
)
5274 if (buf
[0] == '1') {
5275 ret
= validate_slab_cache(s
);
5281 SLAB_ATTR(validate
);
5283 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5285 if (!(s
->flags
& SLAB_STORE_USER
))
5287 return list_locations(s
, buf
, TRACK_ALLOC
);
5289 SLAB_ATTR_RO(alloc_calls
);
5291 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5293 if (!(s
->flags
& SLAB_STORE_USER
))
5295 return list_locations(s
, buf
, TRACK_FREE
);
5297 SLAB_ATTR_RO(free_calls
);
5298 #endif /* CONFIG_SLUB_DEBUG */
5300 #ifdef CONFIG_FAILSLAB
5301 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5303 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5305 SLAB_ATTR_RO(failslab
);
5308 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5313 static ssize_t
shrink_store(struct kmem_cache
*s
,
5314 const char *buf
, size_t length
)
5317 kmem_cache_shrink(s
);
5325 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5327 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5330 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5331 const char *buf
, size_t length
)
5336 err
= kstrtouint(buf
, 10, &ratio
);
5342 s
->remote_node_defrag_ratio
= ratio
* 10;
5346 SLAB_ATTR(remote_node_defrag_ratio
);
5349 #ifdef CONFIG_SLUB_STATS
5350 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5352 unsigned long sum
= 0;
5355 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5360 for_each_online_cpu(cpu
) {
5361 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5367 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5370 for_each_online_cpu(cpu
) {
5372 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5377 len
+= sysfs_emit_at(buf
, len
, "\n");
5382 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5386 for_each_online_cpu(cpu
)
5387 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5390 #define STAT_ATTR(si, text) \
5391 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5393 return show_stat(s, buf, si); \
5395 static ssize_t text##_store(struct kmem_cache *s, \
5396 const char *buf, size_t length) \
5398 if (buf[0] != '0') \
5400 clear_stat(s, si); \
5405 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5406 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5407 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5408 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5409 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5410 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5411 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5412 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5413 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5414 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5415 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5416 STAT_ATTR(FREE_SLAB
, free_slab
);
5417 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5418 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5419 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5420 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5421 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5422 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5423 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5424 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5425 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5426 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5427 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5428 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5429 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5430 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5431 #endif /* CONFIG_SLUB_STATS */
5433 static struct attribute
*slab_attrs
[] = {
5434 &slab_size_attr
.attr
,
5435 &object_size_attr
.attr
,
5436 &objs_per_slab_attr
.attr
,
5438 &min_partial_attr
.attr
,
5439 &cpu_partial_attr
.attr
,
5441 &objects_partial_attr
.attr
,
5443 &cpu_slabs_attr
.attr
,
5447 &hwcache_align_attr
.attr
,
5448 &reclaim_account_attr
.attr
,
5449 &destroy_by_rcu_attr
.attr
,
5451 &slabs_cpu_partial_attr
.attr
,
5452 #ifdef CONFIG_SLUB_DEBUG
5453 &total_objects_attr
.attr
,
5455 &sanity_checks_attr
.attr
,
5457 &red_zone_attr
.attr
,
5459 &store_user_attr
.attr
,
5460 &validate_attr
.attr
,
5461 &alloc_calls_attr
.attr
,
5462 &free_calls_attr
.attr
,
5464 #ifdef CONFIG_ZONE_DMA
5465 &cache_dma_attr
.attr
,
5468 &remote_node_defrag_ratio_attr
.attr
,
5470 #ifdef CONFIG_SLUB_STATS
5471 &alloc_fastpath_attr
.attr
,
5472 &alloc_slowpath_attr
.attr
,
5473 &free_fastpath_attr
.attr
,
5474 &free_slowpath_attr
.attr
,
5475 &free_frozen_attr
.attr
,
5476 &free_add_partial_attr
.attr
,
5477 &free_remove_partial_attr
.attr
,
5478 &alloc_from_partial_attr
.attr
,
5479 &alloc_slab_attr
.attr
,
5480 &alloc_refill_attr
.attr
,
5481 &alloc_node_mismatch_attr
.attr
,
5482 &free_slab_attr
.attr
,
5483 &cpuslab_flush_attr
.attr
,
5484 &deactivate_full_attr
.attr
,
5485 &deactivate_empty_attr
.attr
,
5486 &deactivate_to_head_attr
.attr
,
5487 &deactivate_to_tail_attr
.attr
,
5488 &deactivate_remote_frees_attr
.attr
,
5489 &deactivate_bypass_attr
.attr
,
5490 &order_fallback_attr
.attr
,
5491 &cmpxchg_double_fail_attr
.attr
,
5492 &cmpxchg_double_cpu_fail_attr
.attr
,
5493 &cpu_partial_alloc_attr
.attr
,
5494 &cpu_partial_free_attr
.attr
,
5495 &cpu_partial_node_attr
.attr
,
5496 &cpu_partial_drain_attr
.attr
,
5498 #ifdef CONFIG_FAILSLAB
5499 &failslab_attr
.attr
,
5501 &usersize_attr
.attr
,
5506 static const struct attribute_group slab_attr_group
= {
5507 .attrs
= slab_attrs
,
5510 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5511 struct attribute
*attr
,
5514 struct slab_attribute
*attribute
;
5515 struct kmem_cache
*s
;
5518 attribute
= to_slab_attr(attr
);
5521 if (!attribute
->show
)
5524 err
= attribute
->show(s
, buf
);
5529 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5530 struct attribute
*attr
,
5531 const char *buf
, size_t len
)
5533 struct slab_attribute
*attribute
;
5534 struct kmem_cache
*s
;
5537 attribute
= to_slab_attr(attr
);
5540 if (!attribute
->store
)
5543 err
= attribute
->store(s
, buf
, len
);
5547 static void kmem_cache_release(struct kobject
*k
)
5549 slab_kmem_cache_release(to_slab(k
));
5552 static const struct sysfs_ops slab_sysfs_ops
= {
5553 .show
= slab_attr_show
,
5554 .store
= slab_attr_store
,
5557 static struct kobj_type slab_ktype
= {
5558 .sysfs_ops
= &slab_sysfs_ops
,
5559 .release
= kmem_cache_release
,
5562 static struct kset
*slab_kset
;
5564 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5569 #define ID_STR_LENGTH 64
5571 /* Create a unique string id for a slab cache:
5573 * Format :[flags-]size
5575 static char *create_unique_id(struct kmem_cache
*s
)
5577 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5584 * First flags affecting slabcache operations. We will only
5585 * get here for aliasable slabs so we do not need to support
5586 * too many flags. The flags here must cover all flags that
5587 * are matched during merging to guarantee that the id is
5590 if (s
->flags
& SLAB_CACHE_DMA
)
5592 if (s
->flags
& SLAB_CACHE_DMA32
)
5594 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5596 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5598 if (s
->flags
& SLAB_ACCOUNT
)
5602 p
+= sprintf(p
, "%07u", s
->size
);
5604 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5608 static int sysfs_slab_add(struct kmem_cache
*s
)
5612 struct kset
*kset
= cache_kset(s
);
5613 int unmergeable
= slab_unmergeable(s
);
5616 kobject_init(&s
->kobj
, &slab_ktype
);
5620 if (!unmergeable
&& disable_higher_order_debug
&&
5621 (slub_debug
& DEBUG_METADATA_FLAGS
))
5626 * Slabcache can never be merged so we can use the name proper.
5627 * This is typically the case for debug situations. In that
5628 * case we can catch duplicate names easily.
5630 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5634 * Create a unique name for the slab as a target
5637 name
= create_unique_id(s
);
5640 s
->kobj
.kset
= kset
;
5641 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5645 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5650 /* Setup first alias */
5651 sysfs_slab_alias(s
, s
->name
);
5658 kobject_del(&s
->kobj
);
5662 void sysfs_slab_unlink(struct kmem_cache
*s
)
5664 if (slab_state
>= FULL
)
5665 kobject_del(&s
->kobj
);
5668 void sysfs_slab_release(struct kmem_cache
*s
)
5670 if (slab_state
>= FULL
)
5671 kobject_put(&s
->kobj
);
5675 * Need to buffer aliases during bootup until sysfs becomes
5676 * available lest we lose that information.
5678 struct saved_alias
{
5679 struct kmem_cache
*s
;
5681 struct saved_alias
*next
;
5684 static struct saved_alias
*alias_list
;
5686 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5688 struct saved_alias
*al
;
5690 if (slab_state
== FULL
) {
5692 * If we have a leftover link then remove it.
5694 sysfs_remove_link(&slab_kset
->kobj
, name
);
5695 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5698 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5704 al
->next
= alias_list
;
5709 static int __init
slab_sysfs_init(void)
5711 struct kmem_cache
*s
;
5714 mutex_lock(&slab_mutex
);
5716 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
5718 mutex_unlock(&slab_mutex
);
5719 pr_err("Cannot register slab subsystem.\n");
5725 list_for_each_entry(s
, &slab_caches
, list
) {
5726 err
= sysfs_slab_add(s
);
5728 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5732 while (alias_list
) {
5733 struct saved_alias
*al
= alias_list
;
5735 alias_list
= alias_list
->next
;
5736 err
= sysfs_slab_alias(al
->s
, al
->name
);
5738 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5743 mutex_unlock(&slab_mutex
);
5748 __initcall(slab_sysfs_init
);
5749 #endif /* CONFIG_SYSFS */
5752 * The /proc/slabinfo ABI
5754 #ifdef CONFIG_SLUB_DEBUG
5755 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5757 unsigned long nr_slabs
= 0;
5758 unsigned long nr_objs
= 0;
5759 unsigned long nr_free
= 0;
5761 struct kmem_cache_node
*n
;
5763 for_each_kmem_cache_node(s
, node
, n
) {
5764 nr_slabs
+= node_nr_slabs(n
);
5765 nr_objs
+= node_nr_objs(n
);
5766 nr_free
+= count_partial(n
, count_free
);
5769 sinfo
->active_objs
= nr_objs
- nr_free
;
5770 sinfo
->num_objs
= nr_objs
;
5771 sinfo
->active_slabs
= nr_slabs
;
5772 sinfo
->num_slabs
= nr_slabs
;
5773 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5774 sinfo
->cache_order
= oo_order(s
->oo
);
5777 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5781 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5782 size_t count
, loff_t
*ppos
)
5786 #endif /* CONFIG_SLUB_DEBUG */