1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache
*s
)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
126 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
128 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
129 p
+= s
->red_left_pad
;
134 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
136 #ifdef CONFIG_SLUB_CPU_PARTIAL
137 return !kmem_cache_debug(s
);
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
154 /* Enable to log cmpxchg failures */
155 #undef SLUB_DEBUG_CMPXCHG
158 * Mininum number of partial slabs. These will be left on the partial
159 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 #define MIN_PARTIAL 5
164 * Maximum number of desirable partial slabs.
165 * The existence of more partial slabs makes kmem_cache_shrink
166 * sort the partial list by the number of objects in use.
168 #define MAX_PARTIAL 10
170 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_STORE_USER)
174 * These debug flags cannot use CMPXCHG because there might be consistency
175 * issues when checking or reading debug information
177 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
182 * Debugging flags that require metadata to be stored in the slab. These get
183 * disabled when slub_debug=O is used and a cache's min order increases with
186 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189 #define OO_MASK ((1 << OO_SHIFT) - 1)
190 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192 /* Internal SLUB flags */
194 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
195 /* Use cmpxchg_double */
196 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199 * Tracking user of a slab.
201 #define TRACK_ADDRS_COUNT 16
203 unsigned long addr
; /* Called from address */
204 #ifdef CONFIG_STACKTRACE
205 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
207 int cpu
; /* Was running on cpu */
208 int pid
; /* Pid context */
209 unsigned long when
; /* When did the operation occur */
212 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
215 static int sysfs_slab_add(struct kmem_cache
*);
216 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
217 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
218 static void sysfs_slab_remove(struct kmem_cache
*s
);
220 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
221 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
223 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
224 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
227 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
248 unsigned long ptr_addr
)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr
^ s
->random
^
262 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 return freelist_dereference(s
, object
+ s
->offset
);
281 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
283 prefetch(object
+ s
->offset
);
286 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
288 unsigned long freepointer_addr
;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s
, object
);
294 freepointer_addr
= (unsigned long)object
+ s
->offset
;
295 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
296 return freelist_ptr(s
, p
, freepointer_addr
);
299 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
301 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 /* Determine object index from a given position */
317 static inline unsigned int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
319 return (kasan_reset_tag(p
) - addr
) / s
->size
;
322 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
324 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
327 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
330 struct kmem_cache_order_objects x
= {
331 (order
<< OO_SHIFT
) + order_objects(order
, size
)
337 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
339 return x
.x
>> OO_SHIFT
;
342 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
344 return x
.x
& OO_MASK
;
348 * Per slab locking using the pagelock
350 static __always_inline
void slab_lock(struct page
*page
)
352 VM_BUG_ON_PAGE(PageTail(page
), page
);
353 bit_spin_lock(PG_locked
, &page
->flags
);
356 static __always_inline
void slab_unlock(struct page
*page
)
358 VM_BUG_ON_PAGE(PageTail(page
), page
);
359 __bit_spin_unlock(PG_locked
, &page
->flags
);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
364 void *freelist_old
, unsigned long counters_old
,
365 void *freelist_new
, unsigned long counters_new
,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s
->flags
& __CMPXCHG_DOUBLE
) {
372 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
373 freelist_old
, counters_old
,
374 freelist_new
, counters_new
))
380 if (page
->freelist
== freelist_old
&&
381 page
->counters
== counters_old
) {
382 page
->freelist
= freelist_new
;
383 page
->counters
= counters_new
;
391 stat(s
, CMPXCHG_DOUBLE_FAIL
);
393 #ifdef SLUB_DEBUG_CMPXCHG
394 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
400 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
401 void *freelist_old
, unsigned long counters_old
,
402 void *freelist_new
, unsigned long counters_new
,
405 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
406 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
407 if (s
->flags
& __CMPXCHG_DOUBLE
) {
408 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
409 freelist_old
, counters_old
,
410 freelist_new
, counters_new
))
417 local_irq_save(flags
);
419 if (page
->freelist
== freelist_old
&&
420 page
->counters
== counters_old
) {
421 page
->freelist
= freelist_new
;
422 page
->counters
= counters_new
;
424 local_irq_restore(flags
);
428 local_irq_restore(flags
);
432 stat(s
, CMPXCHG_DOUBLE_FAIL
);
434 #ifdef SLUB_DEBUG_CMPXCHG
435 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
441 #ifdef CONFIG_SLUB_DEBUG
442 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
443 static DEFINE_SPINLOCK(object_map_lock
);
446 * Determine a map of object in use on a page.
448 * Node listlock must be held to guarantee that the page does
449 * not vanish from under us.
451 static unsigned long *get_map(struct kmem_cache
*s
, struct page
*page
)
452 __acquires(&object_map_lock
)
455 void *addr
= page_address(page
);
457 VM_BUG_ON(!irqs_disabled());
459 spin_lock(&object_map_lock
);
461 bitmap_zero(object_map
, page
->objects
);
463 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
464 set_bit(slab_index(p
, s
, addr
), object_map
);
469 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
471 VM_BUG_ON(map
!= object_map
);
472 lockdep_assert_held(&object_map_lock
);
474 spin_unlock(&object_map_lock
);
477 static inline unsigned int size_from_object(struct kmem_cache
*s
)
479 if (s
->flags
& SLAB_RED_ZONE
)
480 return s
->size
- s
->red_left_pad
;
485 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
487 if (s
->flags
& SLAB_RED_ZONE
)
488 p
-= s
->red_left_pad
;
496 #if defined(CONFIG_SLUB_DEBUG_ON)
497 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
499 static slab_flags_t slub_debug
;
502 static char *slub_debug_slabs
;
503 static int disable_higher_order_debug
;
506 * slub is about to manipulate internal object metadata. This memory lies
507 * outside the range of the allocated object, so accessing it would normally
508 * be reported by kasan as a bounds error. metadata_access_enable() is used
509 * to tell kasan that these accesses are OK.
511 static inline void metadata_access_enable(void)
513 kasan_disable_current();
516 static inline void metadata_access_disable(void)
518 kasan_enable_current();
525 /* Verify that a pointer has an address that is valid within a slab page */
526 static inline int check_valid_pointer(struct kmem_cache
*s
,
527 struct page
*page
, void *object
)
534 base
= page_address(page
);
535 object
= kasan_reset_tag(object
);
536 object
= restore_red_left(s
, object
);
537 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
538 (object
- base
) % s
->size
) {
545 static void print_section(char *level
, char *text
, u8
*addr
,
548 metadata_access_enable();
549 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
551 metadata_access_disable();
555 * See comment in calculate_sizes().
557 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
559 return s
->offset
>= s
->inuse
;
563 * Return offset of the end of info block which is inuse + free pointer if
564 * not overlapping with object.
566 static inline unsigned int get_info_end(struct kmem_cache
*s
)
568 if (freeptr_outside_object(s
))
569 return s
->inuse
+ sizeof(void *);
574 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
575 enum track_item alloc
)
579 p
= object
+ get_info_end(s
);
584 static void set_track(struct kmem_cache
*s
, void *object
,
585 enum track_item alloc
, unsigned long addr
)
587 struct track
*p
= get_track(s
, object
, alloc
);
590 #ifdef CONFIG_STACKTRACE
591 unsigned int nr_entries
;
593 metadata_access_enable();
594 nr_entries
= stack_trace_save(p
->addrs
, TRACK_ADDRS_COUNT
, 3);
595 metadata_access_disable();
597 if (nr_entries
< TRACK_ADDRS_COUNT
)
598 p
->addrs
[nr_entries
] = 0;
601 p
->cpu
= smp_processor_id();
602 p
->pid
= current
->pid
;
605 memset(p
, 0, sizeof(struct track
));
609 static void init_tracking(struct kmem_cache
*s
, void *object
)
611 if (!(s
->flags
& SLAB_STORE_USER
))
614 set_track(s
, object
, TRACK_FREE
, 0UL);
615 set_track(s
, object
, TRACK_ALLOC
, 0UL);
618 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
623 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
624 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
625 #ifdef CONFIG_STACKTRACE
628 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
630 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
637 static void print_tracking(struct kmem_cache
*s
, void *object
)
639 unsigned long pr_time
= jiffies
;
640 if (!(s
->flags
& SLAB_STORE_USER
))
643 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
644 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
647 static void print_page_info(struct page
*page
)
649 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
650 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
654 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
656 struct va_format vaf
;
662 pr_err("=============================================================================\n");
663 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
664 pr_err("-----------------------------------------------------------------------------\n\n");
666 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
670 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
672 struct va_format vaf
;
678 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
682 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
683 void *freelist
, void *nextfree
)
685 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
686 !check_valid_pointer(s
, page
, nextfree
)) {
687 object_err(s
, page
, freelist
, "Freechain corrupt");
689 slab_fix(s
, "Isolate corrupted freechain");
696 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
698 unsigned int off
; /* Offset of last byte */
699 u8
*addr
= page_address(page
);
701 print_tracking(s
, p
);
703 print_page_info(page
);
705 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
706 p
, p
- addr
, get_freepointer(s
, p
));
708 if (s
->flags
& SLAB_RED_ZONE
)
709 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
711 else if (p
> addr
+ 16)
712 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
714 print_section(KERN_ERR
, "Object ", p
,
715 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
716 if (s
->flags
& SLAB_RED_ZONE
)
717 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
718 s
->inuse
- s
->object_size
);
720 off
= get_info_end(s
);
722 if (s
->flags
& SLAB_STORE_USER
)
723 off
+= 2 * sizeof(struct track
);
725 off
+= kasan_metadata_size(s
);
727 if (off
!= size_from_object(s
))
728 /* Beginning of the filler is the free pointer */
729 print_section(KERN_ERR
, "Padding ", p
+ off
,
730 size_from_object(s
) - off
);
735 void object_err(struct kmem_cache
*s
, struct page
*page
,
736 u8
*object
, char *reason
)
738 slab_bug(s
, "%s", reason
);
739 print_trailer(s
, page
, object
);
742 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
743 const char *fmt
, ...)
749 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
751 slab_bug(s
, "%s", buf
);
752 print_page_info(page
);
756 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
760 if (s
->flags
& SLAB_RED_ZONE
)
761 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
763 if (s
->flags
& __OBJECT_POISON
) {
764 memset(p
, POISON_FREE
, s
->object_size
- 1);
765 p
[s
->object_size
- 1] = POISON_END
;
768 if (s
->flags
& SLAB_RED_ZONE
)
769 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
772 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
773 void *from
, void *to
)
775 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
776 memset(from
, data
, to
- from
);
779 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
780 u8
*object
, char *what
,
781 u8
*start
, unsigned int value
, unsigned int bytes
)
785 u8
*addr
= page_address(page
);
787 metadata_access_enable();
788 fault
= memchr_inv(start
, value
, bytes
);
789 metadata_access_disable();
794 while (end
> fault
&& end
[-1] == value
)
797 slab_bug(s
, "%s overwritten", what
);
798 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
799 fault
, end
- 1, fault
- addr
,
801 print_trailer(s
, page
, object
);
803 restore_bytes(s
, what
, value
, fault
, end
);
811 * Bytes of the object to be managed.
812 * If the freepointer may overlay the object then the free
813 * pointer is at the middle of the object.
815 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
818 * object + s->object_size
819 * Padding to reach word boundary. This is also used for Redzoning.
820 * Padding is extended by another word if Redzoning is enabled and
821 * object_size == inuse.
823 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
824 * 0xcc (RED_ACTIVE) for objects in use.
827 * Meta data starts here.
829 * A. Free pointer (if we cannot overwrite object on free)
830 * B. Tracking data for SLAB_STORE_USER
831 * C. Padding to reach required alignment boundary or at mininum
832 * one word if debugging is on to be able to detect writes
833 * before the word boundary.
835 * Padding is done using 0x5a (POISON_INUSE)
838 * Nothing is used beyond s->size.
840 * If slabcaches are merged then the object_size and inuse boundaries are mostly
841 * ignored. And therefore no slab options that rely on these boundaries
842 * may be used with merged slabcaches.
845 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
847 unsigned long off
= get_info_end(s
); /* The end of info */
849 if (s
->flags
& SLAB_STORE_USER
)
850 /* We also have user information there */
851 off
+= 2 * sizeof(struct track
);
853 off
+= kasan_metadata_size(s
);
855 if (size_from_object(s
) == off
)
858 return check_bytes_and_report(s
, page
, p
, "Object padding",
859 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
862 /* Check the pad bytes at the end of a slab page */
863 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
872 if (!(s
->flags
& SLAB_POISON
))
875 start
= page_address(page
);
876 length
= page_size(page
);
877 end
= start
+ length
;
878 remainder
= length
% s
->size
;
882 pad
= end
- remainder
;
883 metadata_access_enable();
884 fault
= memchr_inv(pad
, POISON_INUSE
, remainder
);
885 metadata_access_disable();
888 while (end
> fault
&& end
[-1] == POISON_INUSE
)
891 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
892 fault
, end
- 1, fault
- start
);
893 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
895 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
899 static int check_object(struct kmem_cache
*s
, struct page
*page
,
900 void *object
, u8 val
)
903 u8
*endobject
= object
+ s
->object_size
;
905 if (s
->flags
& SLAB_RED_ZONE
) {
906 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
907 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
910 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
911 endobject
, val
, s
->inuse
- s
->object_size
))
914 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
915 check_bytes_and_report(s
, page
, p
, "Alignment padding",
916 endobject
, POISON_INUSE
,
917 s
->inuse
- s
->object_size
);
921 if (s
->flags
& SLAB_POISON
) {
922 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
923 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
924 POISON_FREE
, s
->object_size
- 1) ||
925 !check_bytes_and_report(s
, page
, p
, "Poison",
926 p
+ s
->object_size
- 1, POISON_END
, 1)))
929 * check_pad_bytes cleans up on its own.
931 check_pad_bytes(s
, page
, p
);
934 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
936 * Object and freepointer overlap. Cannot check
937 * freepointer while object is allocated.
941 /* Check free pointer validity */
942 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
943 object_err(s
, page
, p
, "Freepointer corrupt");
945 * No choice but to zap it and thus lose the remainder
946 * of the free objects in this slab. May cause
947 * another error because the object count is now wrong.
949 set_freepointer(s
, p
, NULL
);
955 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
959 VM_BUG_ON(!irqs_disabled());
961 if (!PageSlab(page
)) {
962 slab_err(s
, page
, "Not a valid slab page");
966 maxobj
= order_objects(compound_order(page
), s
->size
);
967 if (page
->objects
> maxobj
) {
968 slab_err(s
, page
, "objects %u > max %u",
969 page
->objects
, maxobj
);
972 if (page
->inuse
> page
->objects
) {
973 slab_err(s
, page
, "inuse %u > max %u",
974 page
->inuse
, page
->objects
);
977 /* Slab_pad_check fixes things up after itself */
978 slab_pad_check(s
, page
);
983 * Determine if a certain object on a page is on the freelist. Must hold the
984 * slab lock to guarantee that the chains are in a consistent state.
986 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
994 while (fp
&& nr
<= page
->objects
) {
997 if (!check_valid_pointer(s
, page
, fp
)) {
999 object_err(s
, page
, object
,
1000 "Freechain corrupt");
1001 set_freepointer(s
, object
, NULL
);
1003 slab_err(s
, page
, "Freepointer corrupt");
1004 page
->freelist
= NULL
;
1005 page
->inuse
= page
->objects
;
1006 slab_fix(s
, "Freelist cleared");
1012 fp
= get_freepointer(s
, object
);
1016 max_objects
= order_objects(compound_order(page
), s
->size
);
1017 if (max_objects
> MAX_OBJS_PER_PAGE
)
1018 max_objects
= MAX_OBJS_PER_PAGE
;
1020 if (page
->objects
!= max_objects
) {
1021 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1022 page
->objects
, max_objects
);
1023 page
->objects
= max_objects
;
1024 slab_fix(s
, "Number of objects adjusted.");
1026 if (page
->inuse
!= page
->objects
- nr
) {
1027 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1028 page
->inuse
, page
->objects
- nr
);
1029 page
->inuse
= page
->objects
- nr
;
1030 slab_fix(s
, "Object count adjusted.");
1032 return search
== NULL
;
1035 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1038 if (s
->flags
& SLAB_TRACE
) {
1039 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1041 alloc
? "alloc" : "free",
1042 object
, page
->inuse
,
1046 print_section(KERN_INFO
, "Object ", (void *)object
,
1054 * Tracking of fully allocated slabs for debugging purposes.
1056 static void add_full(struct kmem_cache
*s
,
1057 struct kmem_cache_node
*n
, struct page
*page
)
1059 if (!(s
->flags
& SLAB_STORE_USER
))
1062 lockdep_assert_held(&n
->list_lock
);
1063 list_add(&page
->slab_list
, &n
->full
);
1066 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1068 if (!(s
->flags
& SLAB_STORE_USER
))
1071 lockdep_assert_held(&n
->list_lock
);
1072 list_del(&page
->slab_list
);
1075 /* Tracking of the number of slabs for debugging purposes */
1076 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1078 struct kmem_cache_node
*n
= get_node(s
, node
);
1080 return atomic_long_read(&n
->nr_slabs
);
1083 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1085 return atomic_long_read(&n
->nr_slabs
);
1088 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1090 struct kmem_cache_node
*n
= get_node(s
, node
);
1093 * May be called early in order to allocate a slab for the
1094 * kmem_cache_node structure. Solve the chicken-egg
1095 * dilemma by deferring the increment of the count during
1096 * bootstrap (see early_kmem_cache_node_alloc).
1099 atomic_long_inc(&n
->nr_slabs
);
1100 atomic_long_add(objects
, &n
->total_objects
);
1103 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1105 struct kmem_cache_node
*n
= get_node(s
, node
);
1107 atomic_long_dec(&n
->nr_slabs
);
1108 atomic_long_sub(objects
, &n
->total_objects
);
1111 /* Object debug checks for alloc/free paths */
1112 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1115 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1118 init_object(s
, object
, SLUB_RED_INACTIVE
);
1119 init_tracking(s
, object
);
1123 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1125 if (!(s
->flags
& SLAB_POISON
))
1128 metadata_access_enable();
1129 memset(addr
, POISON_INUSE
, page_size(page
));
1130 metadata_access_disable();
1133 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1134 struct page
*page
, void *object
)
1136 if (!check_slab(s
, page
))
1139 if (!check_valid_pointer(s
, page
, object
)) {
1140 object_err(s
, page
, object
, "Freelist Pointer check fails");
1144 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1150 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1152 void *object
, unsigned long addr
)
1154 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1155 if (!alloc_consistency_checks(s
, page
, object
))
1159 /* Success perform special debug activities for allocs */
1160 if (s
->flags
& SLAB_STORE_USER
)
1161 set_track(s
, object
, TRACK_ALLOC
, addr
);
1162 trace(s
, page
, object
, 1);
1163 init_object(s
, object
, SLUB_RED_ACTIVE
);
1167 if (PageSlab(page
)) {
1169 * If this is a slab page then lets do the best we can
1170 * to avoid issues in the future. Marking all objects
1171 * as used avoids touching the remaining objects.
1173 slab_fix(s
, "Marking all objects used");
1174 page
->inuse
= page
->objects
;
1175 page
->freelist
= NULL
;
1180 static inline int free_consistency_checks(struct kmem_cache
*s
,
1181 struct page
*page
, void *object
, unsigned long addr
)
1183 if (!check_valid_pointer(s
, page
, object
)) {
1184 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1188 if (on_freelist(s
, page
, object
)) {
1189 object_err(s
, page
, object
, "Object already free");
1193 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1196 if (unlikely(s
!= page
->slab_cache
)) {
1197 if (!PageSlab(page
)) {
1198 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1200 } else if (!page
->slab_cache
) {
1201 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1205 object_err(s
, page
, object
,
1206 "page slab pointer corrupt.");
1212 /* Supports checking bulk free of a constructed freelist */
1213 static noinline
int free_debug_processing(
1214 struct kmem_cache
*s
, struct page
*page
,
1215 void *head
, void *tail
, int bulk_cnt
,
1218 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1219 void *object
= head
;
1221 unsigned long flags
;
1224 spin_lock_irqsave(&n
->list_lock
, flags
);
1227 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1228 if (!check_slab(s
, page
))
1235 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1236 if (!free_consistency_checks(s
, page
, object
, addr
))
1240 if (s
->flags
& SLAB_STORE_USER
)
1241 set_track(s
, object
, TRACK_FREE
, addr
);
1242 trace(s
, page
, object
, 0);
1243 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1244 init_object(s
, object
, SLUB_RED_INACTIVE
);
1246 /* Reached end of constructed freelist yet? */
1247 if (object
!= tail
) {
1248 object
= get_freepointer(s
, object
);
1254 if (cnt
!= bulk_cnt
)
1255 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1259 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1261 slab_fix(s
, "Object at 0x%p not freed", object
);
1265 static int __init
setup_slub_debug(char *str
)
1267 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1268 if (*str
++ != '=' || !*str
)
1270 * No options specified. Switch on full debugging.
1276 * No options but restriction on slabs. This means full
1277 * debugging for slabs matching a pattern.
1284 * Switch off all debugging measures.
1289 * Determine which debug features should be switched on
1291 for (; *str
&& *str
!= ','; str
++) {
1292 switch (tolower(*str
)) {
1294 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1297 slub_debug
|= SLAB_RED_ZONE
;
1300 slub_debug
|= SLAB_POISON
;
1303 slub_debug
|= SLAB_STORE_USER
;
1306 slub_debug
|= SLAB_TRACE
;
1309 slub_debug
|= SLAB_FAILSLAB
;
1313 * Avoid enabling debugging on caches if its minimum
1314 * order would increase as a result.
1316 disable_higher_order_debug
= 1;
1319 pr_err("slub_debug option '%c' unknown. skipped\n",
1326 slub_debug_slabs
= str
+ 1;
1328 if ((static_branch_unlikely(&init_on_alloc
) ||
1329 static_branch_unlikely(&init_on_free
)) &&
1330 (slub_debug
& SLAB_POISON
))
1331 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1335 __setup("slub_debug", setup_slub_debug
);
1338 * kmem_cache_flags - apply debugging options to the cache
1339 * @object_size: the size of an object without meta data
1340 * @flags: flags to set
1341 * @name: name of the cache
1342 * @ctor: constructor function
1344 * Debug option(s) are applied to @flags. In addition to the debug
1345 * option(s), if a slab name (or multiple) is specified i.e.
1346 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1347 * then only the select slabs will receive the debug option(s).
1349 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1350 slab_flags_t flags
, const char *name
,
1351 void (*ctor
)(void *))
1356 /* If slub_debug = 0, it folds into the if conditional. */
1357 if (!slub_debug_slabs
)
1358 return flags
| slub_debug
;
1361 iter
= slub_debug_slabs
;
1366 end
= strchrnul(iter
, ',');
1368 glob
= strnchr(iter
, end
- iter
, '*');
1370 cmplen
= glob
- iter
;
1372 cmplen
= max_t(size_t, len
, (end
- iter
));
1374 if (!strncmp(name
, iter
, cmplen
)) {
1375 flags
|= slub_debug
;
1386 #else /* !CONFIG_SLUB_DEBUG */
1387 static inline void setup_object_debug(struct kmem_cache
*s
,
1388 struct page
*page
, void *object
) {}
1390 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1392 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1393 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1395 static inline int free_debug_processing(
1396 struct kmem_cache
*s
, struct page
*page
,
1397 void *head
, void *tail
, int bulk_cnt
,
1398 unsigned long addr
) { return 0; }
1400 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1402 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1403 void *object
, u8 val
) { return 1; }
1404 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1405 struct page
*page
) {}
1406 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1407 struct page
*page
) {}
1408 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1409 slab_flags_t flags
, const char *name
,
1410 void (*ctor
)(void *))
1414 #define slub_debug 0
1416 #define disable_higher_order_debug 0
1418 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1420 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1422 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1424 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1427 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1428 void *freelist
, void *nextfree
)
1432 #endif /* CONFIG_SLUB_DEBUG */
1435 * Hooks for other subsystems that check memory allocations. In a typical
1436 * production configuration these hooks all should produce no code at all.
1438 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1440 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1441 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1442 kmemleak_alloc(ptr
, size
, 1, flags
);
1446 static __always_inline
void kfree_hook(void *x
)
1449 kasan_kfree_large(x
, _RET_IP_
);
1452 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1454 kmemleak_free_recursive(x
, s
->flags
);
1457 * Trouble is that we may no longer disable interrupts in the fast path
1458 * So in order to make the debug calls that expect irqs to be
1459 * disabled we need to disable interrupts temporarily.
1461 #ifdef CONFIG_LOCKDEP
1463 unsigned long flags
;
1465 local_irq_save(flags
);
1466 debug_check_no_locks_freed(x
, s
->object_size
);
1467 local_irq_restore(flags
);
1470 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1471 debug_check_no_obj_freed(x
, s
->object_size
);
1473 /* KASAN might put x into memory quarantine, delaying its reuse */
1474 return kasan_slab_free(s
, x
, _RET_IP_
);
1477 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1478 void **head
, void **tail
)
1483 void *old_tail
= *tail
? *tail
: *head
;
1486 /* Head and tail of the reconstructed freelist */
1492 next
= get_freepointer(s
, object
);
1494 if (slab_want_init_on_free(s
)) {
1496 * Clear the object and the metadata, but don't touch
1499 memset(object
, 0, s
->object_size
);
1500 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
1502 memset((char *)object
+ s
->inuse
, 0,
1503 s
->size
- s
->inuse
- rsize
);
1506 /* If object's reuse doesn't have to be delayed */
1507 if (!slab_free_hook(s
, object
)) {
1508 /* Move object to the new freelist */
1509 set_freepointer(s
, object
, *head
);
1514 } while (object
!= old_tail
);
1519 return *head
!= NULL
;
1522 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1525 setup_object_debug(s
, page
, object
);
1526 object
= kasan_init_slab_obj(s
, object
);
1527 if (unlikely(s
->ctor
)) {
1528 kasan_unpoison_object_data(s
, object
);
1530 kasan_poison_object_data(s
, object
);
1536 * Slab allocation and freeing
1538 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1539 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1542 unsigned int order
= oo_order(oo
);
1544 if (node
== NUMA_NO_NODE
)
1545 page
= alloc_pages(flags
, order
);
1547 page
= __alloc_pages_node(node
, flags
, order
);
1549 if (page
&& charge_slab_page(page
, flags
, order
, s
)) {
1550 __free_pages(page
, order
);
1557 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1558 /* Pre-initialize the random sequence cache */
1559 static int init_cache_random_seq(struct kmem_cache
*s
)
1561 unsigned int count
= oo_objects(s
->oo
);
1564 /* Bailout if already initialised */
1568 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1570 pr_err("SLUB: Unable to initialize free list for %s\n",
1575 /* Transform to an offset on the set of pages */
1576 if (s
->random_seq
) {
1579 for (i
= 0; i
< count
; i
++)
1580 s
->random_seq
[i
] *= s
->size
;
1585 /* Initialize each random sequence freelist per cache */
1586 static void __init
init_freelist_randomization(void)
1588 struct kmem_cache
*s
;
1590 mutex_lock(&slab_mutex
);
1592 list_for_each_entry(s
, &slab_caches
, list
)
1593 init_cache_random_seq(s
);
1595 mutex_unlock(&slab_mutex
);
1598 /* Get the next entry on the pre-computed freelist randomized */
1599 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1600 unsigned long *pos
, void *start
,
1601 unsigned long page_limit
,
1602 unsigned long freelist_count
)
1607 * If the target page allocation failed, the number of objects on the
1608 * page might be smaller than the usual size defined by the cache.
1611 idx
= s
->random_seq
[*pos
];
1613 if (*pos
>= freelist_count
)
1615 } while (unlikely(idx
>= page_limit
));
1617 return (char *)start
+ idx
;
1620 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1621 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1626 unsigned long idx
, pos
, page_limit
, freelist_count
;
1628 if (page
->objects
< 2 || !s
->random_seq
)
1631 freelist_count
= oo_objects(s
->oo
);
1632 pos
= get_random_int() % freelist_count
;
1634 page_limit
= page
->objects
* s
->size
;
1635 start
= fixup_red_left(s
, page_address(page
));
1637 /* First entry is used as the base of the freelist */
1638 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1640 cur
= setup_object(s
, page
, cur
);
1641 page
->freelist
= cur
;
1643 for (idx
= 1; idx
< page
->objects
; idx
++) {
1644 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1646 next
= setup_object(s
, page
, next
);
1647 set_freepointer(s
, cur
, next
);
1650 set_freepointer(s
, cur
, NULL
);
1655 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1659 static inline void init_freelist_randomization(void) { }
1660 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1664 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1666 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1669 struct kmem_cache_order_objects oo
= s
->oo
;
1671 void *start
, *p
, *next
;
1675 flags
&= gfp_allowed_mask
;
1677 if (gfpflags_allow_blocking(flags
))
1680 flags
|= s
->allocflags
;
1683 * Let the initial higher-order allocation fail under memory pressure
1684 * so we fall-back to the minimum order allocation.
1686 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1687 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1688 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1690 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1691 if (unlikely(!page
)) {
1695 * Allocation may have failed due to fragmentation.
1696 * Try a lower order alloc if possible
1698 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1699 if (unlikely(!page
))
1701 stat(s
, ORDER_FALLBACK
);
1704 page
->objects
= oo_objects(oo
);
1706 page
->slab_cache
= s
;
1707 __SetPageSlab(page
);
1708 if (page_is_pfmemalloc(page
))
1709 SetPageSlabPfmemalloc(page
);
1711 kasan_poison_slab(page
);
1713 start
= page_address(page
);
1715 setup_page_debug(s
, page
, start
);
1717 shuffle
= shuffle_freelist(s
, page
);
1720 start
= fixup_red_left(s
, start
);
1721 start
= setup_object(s
, page
, start
);
1722 page
->freelist
= start
;
1723 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1725 next
= setup_object(s
, page
, next
);
1726 set_freepointer(s
, p
, next
);
1729 set_freepointer(s
, p
, NULL
);
1732 page
->inuse
= page
->objects
;
1736 if (gfpflags_allow_blocking(flags
))
1737 local_irq_disable();
1741 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1746 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1748 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1749 flags
= kmalloc_fix_flags(flags
);
1751 return allocate_slab(s
,
1752 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1755 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1757 int order
= compound_order(page
);
1758 int pages
= 1 << order
;
1760 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1763 slab_pad_check(s
, page
);
1764 for_each_object(p
, s
, page_address(page
),
1766 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1769 __ClearPageSlabPfmemalloc(page
);
1770 __ClearPageSlab(page
);
1772 page
->mapping
= NULL
;
1773 if (current
->reclaim_state
)
1774 current
->reclaim_state
->reclaimed_slab
+= pages
;
1775 uncharge_slab_page(page
, order
, s
);
1776 __free_pages(page
, order
);
1779 static void rcu_free_slab(struct rcu_head
*h
)
1781 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1783 __free_slab(page
->slab_cache
, page
);
1786 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1788 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1789 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1791 __free_slab(s
, page
);
1794 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1796 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1801 * Management of partially allocated slabs.
1804 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1807 if (tail
== DEACTIVATE_TO_TAIL
)
1808 list_add_tail(&page
->slab_list
, &n
->partial
);
1810 list_add(&page
->slab_list
, &n
->partial
);
1813 static inline void add_partial(struct kmem_cache_node
*n
,
1814 struct page
*page
, int tail
)
1816 lockdep_assert_held(&n
->list_lock
);
1817 __add_partial(n
, page
, tail
);
1820 static inline void remove_partial(struct kmem_cache_node
*n
,
1823 lockdep_assert_held(&n
->list_lock
);
1824 list_del(&page
->slab_list
);
1829 * Remove slab from the partial list, freeze it and
1830 * return the pointer to the freelist.
1832 * Returns a list of objects or NULL if it fails.
1834 static inline void *acquire_slab(struct kmem_cache
*s
,
1835 struct kmem_cache_node
*n
, struct page
*page
,
1836 int mode
, int *objects
)
1839 unsigned long counters
;
1842 lockdep_assert_held(&n
->list_lock
);
1845 * Zap the freelist and set the frozen bit.
1846 * The old freelist is the list of objects for the
1847 * per cpu allocation list.
1849 freelist
= page
->freelist
;
1850 counters
= page
->counters
;
1851 new.counters
= counters
;
1852 *objects
= new.objects
- new.inuse
;
1854 new.inuse
= page
->objects
;
1855 new.freelist
= NULL
;
1857 new.freelist
= freelist
;
1860 VM_BUG_ON(new.frozen
);
1863 if (!__cmpxchg_double_slab(s
, page
,
1865 new.freelist
, new.counters
,
1869 remove_partial(n
, page
);
1874 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1875 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1878 * Try to allocate a partial slab from a specific node.
1880 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1881 struct kmem_cache_cpu
*c
, gfp_t flags
)
1883 struct page
*page
, *page2
;
1884 void *object
= NULL
;
1885 unsigned int available
= 0;
1889 * Racy check. If we mistakenly see no partial slabs then we
1890 * just allocate an empty slab. If we mistakenly try to get a
1891 * partial slab and there is none available then get_partials()
1894 if (!n
|| !n
->nr_partial
)
1897 spin_lock(&n
->list_lock
);
1898 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1901 if (!pfmemalloc_match(page
, flags
))
1904 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1908 available
+= objects
;
1911 stat(s
, ALLOC_FROM_PARTIAL
);
1914 put_cpu_partial(s
, page
, 0);
1915 stat(s
, CPU_PARTIAL_NODE
);
1917 if (!kmem_cache_has_cpu_partial(s
)
1918 || available
> slub_cpu_partial(s
) / 2)
1922 spin_unlock(&n
->list_lock
);
1927 * Get a page from somewhere. Search in increasing NUMA distances.
1929 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1930 struct kmem_cache_cpu
*c
)
1933 struct zonelist
*zonelist
;
1936 enum zone_type highest_zoneidx
= gfp_zone(flags
);
1938 unsigned int cpuset_mems_cookie
;
1941 * The defrag ratio allows a configuration of the tradeoffs between
1942 * inter node defragmentation and node local allocations. A lower
1943 * defrag_ratio increases the tendency to do local allocations
1944 * instead of attempting to obtain partial slabs from other nodes.
1946 * If the defrag_ratio is set to 0 then kmalloc() always
1947 * returns node local objects. If the ratio is higher then kmalloc()
1948 * may return off node objects because partial slabs are obtained
1949 * from other nodes and filled up.
1951 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1952 * (which makes defrag_ratio = 1000) then every (well almost)
1953 * allocation will first attempt to defrag slab caches on other nodes.
1954 * This means scanning over all nodes to look for partial slabs which
1955 * may be expensive if we do it every time we are trying to find a slab
1956 * with available objects.
1958 if (!s
->remote_node_defrag_ratio
||
1959 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1963 cpuset_mems_cookie
= read_mems_allowed_begin();
1964 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1965 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
1966 struct kmem_cache_node
*n
;
1968 n
= get_node(s
, zone_to_nid(zone
));
1970 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1971 n
->nr_partial
> s
->min_partial
) {
1972 object
= get_partial_node(s
, n
, c
, flags
);
1975 * Don't check read_mems_allowed_retry()
1976 * here - if mems_allowed was updated in
1977 * parallel, that was a harmless race
1978 * between allocation and the cpuset
1985 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1986 #endif /* CONFIG_NUMA */
1991 * Get a partial page, lock it and return it.
1993 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1994 struct kmem_cache_cpu
*c
)
1997 int searchnode
= node
;
1999 if (node
== NUMA_NO_NODE
)
2000 searchnode
= numa_mem_id();
2002 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
2003 if (object
|| node
!= NUMA_NO_NODE
)
2006 return get_any_partial(s
, flags
, c
);
2009 #ifdef CONFIG_PREEMPTION
2011 * Calculate the next globally unique transaction for disambiguation
2012 * during cmpxchg. The transactions start with the cpu number and are then
2013 * incremented by CONFIG_NR_CPUS.
2015 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2018 * No preemption supported therefore also no need to check for
2024 static inline unsigned long next_tid(unsigned long tid
)
2026 return tid
+ TID_STEP
;
2029 #ifdef SLUB_DEBUG_CMPXCHG
2030 static inline unsigned int tid_to_cpu(unsigned long tid
)
2032 return tid
% TID_STEP
;
2035 static inline unsigned long tid_to_event(unsigned long tid
)
2037 return tid
/ TID_STEP
;
2041 static inline unsigned int init_tid(int cpu
)
2046 static inline void note_cmpxchg_failure(const char *n
,
2047 const struct kmem_cache
*s
, unsigned long tid
)
2049 #ifdef SLUB_DEBUG_CMPXCHG
2050 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2052 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2054 #ifdef CONFIG_PREEMPTION
2055 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2056 pr_warn("due to cpu change %d -> %d\n",
2057 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2060 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2061 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2062 tid_to_event(tid
), tid_to_event(actual_tid
));
2064 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2065 actual_tid
, tid
, next_tid(tid
));
2067 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2070 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2074 for_each_possible_cpu(cpu
)
2075 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2079 * Remove the cpu slab
2081 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2082 void *freelist
, struct kmem_cache_cpu
*c
)
2084 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2085 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2087 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2089 int tail
= DEACTIVATE_TO_HEAD
;
2093 if (page
->freelist
) {
2094 stat(s
, DEACTIVATE_REMOTE_FREES
);
2095 tail
= DEACTIVATE_TO_TAIL
;
2099 * Stage one: Free all available per cpu objects back
2100 * to the page freelist while it is still frozen. Leave the
2103 * There is no need to take the list->lock because the page
2106 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2108 unsigned long counters
;
2111 * If 'nextfree' is invalid, it is possible that the object at
2112 * 'freelist' is already corrupted. So isolate all objects
2113 * starting at 'freelist'.
2115 if (freelist_corrupted(s
, page
, freelist
, nextfree
))
2119 prior
= page
->freelist
;
2120 counters
= page
->counters
;
2121 set_freepointer(s
, freelist
, prior
);
2122 new.counters
= counters
;
2124 VM_BUG_ON(!new.frozen
);
2126 } while (!__cmpxchg_double_slab(s
, page
,
2128 freelist
, new.counters
,
2129 "drain percpu freelist"));
2131 freelist
= nextfree
;
2135 * Stage two: Ensure that the page is unfrozen while the
2136 * list presence reflects the actual number of objects
2139 * We setup the list membership and then perform a cmpxchg
2140 * with the count. If there is a mismatch then the page
2141 * is not unfrozen but the page is on the wrong list.
2143 * Then we restart the process which may have to remove
2144 * the page from the list that we just put it on again
2145 * because the number of objects in the slab may have
2150 old
.freelist
= page
->freelist
;
2151 old
.counters
= page
->counters
;
2152 VM_BUG_ON(!old
.frozen
);
2154 /* Determine target state of the slab */
2155 new.counters
= old
.counters
;
2158 set_freepointer(s
, freelist
, old
.freelist
);
2159 new.freelist
= freelist
;
2161 new.freelist
= old
.freelist
;
2165 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2167 else if (new.freelist
) {
2172 * Taking the spinlock removes the possibility
2173 * that acquire_slab() will see a slab page that
2176 spin_lock(&n
->list_lock
);
2180 if (kmem_cache_debug(s
) && !lock
) {
2183 * This also ensures that the scanning of full
2184 * slabs from diagnostic functions will not see
2187 spin_lock(&n
->list_lock
);
2193 remove_partial(n
, page
);
2194 else if (l
== M_FULL
)
2195 remove_full(s
, n
, page
);
2198 add_partial(n
, page
, tail
);
2199 else if (m
== M_FULL
)
2200 add_full(s
, n
, page
);
2204 if (!__cmpxchg_double_slab(s
, page
,
2205 old
.freelist
, old
.counters
,
2206 new.freelist
, new.counters
,
2211 spin_unlock(&n
->list_lock
);
2215 else if (m
== M_FULL
)
2216 stat(s
, DEACTIVATE_FULL
);
2217 else if (m
== M_FREE
) {
2218 stat(s
, DEACTIVATE_EMPTY
);
2219 discard_slab(s
, page
);
2228 * Unfreeze all the cpu partial slabs.
2230 * This function must be called with interrupts disabled
2231 * for the cpu using c (or some other guarantee must be there
2232 * to guarantee no concurrent accesses).
2234 static void unfreeze_partials(struct kmem_cache
*s
,
2235 struct kmem_cache_cpu
*c
)
2237 #ifdef CONFIG_SLUB_CPU_PARTIAL
2238 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2239 struct page
*page
, *discard_page
= NULL
;
2241 while ((page
= slub_percpu_partial(c
))) {
2245 slub_set_percpu_partial(c
, page
);
2247 n2
= get_node(s
, page_to_nid(page
));
2250 spin_unlock(&n
->list_lock
);
2253 spin_lock(&n
->list_lock
);
2258 old
.freelist
= page
->freelist
;
2259 old
.counters
= page
->counters
;
2260 VM_BUG_ON(!old
.frozen
);
2262 new.counters
= old
.counters
;
2263 new.freelist
= old
.freelist
;
2267 } while (!__cmpxchg_double_slab(s
, page
,
2268 old
.freelist
, old
.counters
,
2269 new.freelist
, new.counters
,
2270 "unfreezing slab"));
2272 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2273 page
->next
= discard_page
;
2274 discard_page
= page
;
2276 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2277 stat(s
, FREE_ADD_PARTIAL
);
2282 spin_unlock(&n
->list_lock
);
2284 while (discard_page
) {
2285 page
= discard_page
;
2286 discard_page
= discard_page
->next
;
2288 stat(s
, DEACTIVATE_EMPTY
);
2289 discard_slab(s
, page
);
2292 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2296 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2297 * partial page slot if available.
2299 * If we did not find a slot then simply move all the partials to the
2300 * per node partial list.
2302 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2304 #ifdef CONFIG_SLUB_CPU_PARTIAL
2305 struct page
*oldpage
;
2313 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2316 pobjects
= oldpage
->pobjects
;
2317 pages
= oldpage
->pages
;
2318 if (drain
&& pobjects
> slub_cpu_partial(s
)) {
2319 unsigned long flags
;
2321 * partial array is full. Move the existing
2322 * set to the per node partial list.
2324 local_irq_save(flags
);
2325 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2326 local_irq_restore(flags
);
2330 stat(s
, CPU_PARTIAL_DRAIN
);
2335 pobjects
+= page
->objects
- page
->inuse
;
2337 page
->pages
= pages
;
2338 page
->pobjects
= pobjects
;
2339 page
->next
= oldpage
;
2341 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2343 if (unlikely(!slub_cpu_partial(s
))) {
2344 unsigned long flags
;
2346 local_irq_save(flags
);
2347 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2348 local_irq_restore(flags
);
2351 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2354 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2356 stat(s
, CPUSLAB_FLUSH
);
2357 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2359 c
->tid
= next_tid(c
->tid
);
2365 * Called from IPI handler with interrupts disabled.
2367 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2369 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2374 unfreeze_partials(s
, c
);
2377 static void flush_cpu_slab(void *d
)
2379 struct kmem_cache
*s
= d
;
2381 __flush_cpu_slab(s
, smp_processor_id());
2384 static bool has_cpu_slab(int cpu
, void *info
)
2386 struct kmem_cache
*s
= info
;
2387 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2389 return c
->page
|| slub_percpu_partial(c
);
2392 static void flush_all(struct kmem_cache
*s
)
2394 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1);
2398 * Use the cpu notifier to insure that the cpu slabs are flushed when
2401 static int slub_cpu_dead(unsigned int cpu
)
2403 struct kmem_cache
*s
;
2404 unsigned long flags
;
2406 mutex_lock(&slab_mutex
);
2407 list_for_each_entry(s
, &slab_caches
, list
) {
2408 local_irq_save(flags
);
2409 __flush_cpu_slab(s
, cpu
);
2410 local_irq_restore(flags
);
2412 mutex_unlock(&slab_mutex
);
2417 * Check if the objects in a per cpu structure fit numa
2418 * locality expectations.
2420 static inline int node_match(struct page
*page
, int node
)
2423 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2429 #ifdef CONFIG_SLUB_DEBUG
2430 static int count_free(struct page
*page
)
2432 return page
->objects
- page
->inuse
;
2435 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2437 return atomic_long_read(&n
->total_objects
);
2439 #endif /* CONFIG_SLUB_DEBUG */
2441 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2442 static unsigned long count_partial(struct kmem_cache_node
*n
,
2443 int (*get_count
)(struct page
*))
2445 unsigned long flags
;
2446 unsigned long x
= 0;
2449 spin_lock_irqsave(&n
->list_lock
, flags
);
2450 list_for_each_entry(page
, &n
->partial
, slab_list
)
2451 x
+= get_count(page
);
2452 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2455 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2457 static noinline
void
2458 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2460 #ifdef CONFIG_SLUB_DEBUG
2461 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2462 DEFAULT_RATELIMIT_BURST
);
2464 struct kmem_cache_node
*n
;
2466 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2469 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2470 nid
, gfpflags
, &gfpflags
);
2471 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2472 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2475 if (oo_order(s
->min
) > get_order(s
->object_size
))
2476 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2479 for_each_kmem_cache_node(s
, node
, n
) {
2480 unsigned long nr_slabs
;
2481 unsigned long nr_objs
;
2482 unsigned long nr_free
;
2484 nr_free
= count_partial(n
, count_free
);
2485 nr_slabs
= node_nr_slabs(n
);
2486 nr_objs
= node_nr_objs(n
);
2488 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2489 node
, nr_slabs
, nr_objs
, nr_free
);
2494 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2495 int node
, struct kmem_cache_cpu
**pc
)
2498 struct kmem_cache_cpu
*c
= *pc
;
2501 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2503 freelist
= get_partial(s
, flags
, node
, c
);
2508 page
= new_slab(s
, flags
, node
);
2510 c
= raw_cpu_ptr(s
->cpu_slab
);
2515 * No other reference to the page yet so we can
2516 * muck around with it freely without cmpxchg
2518 freelist
= page
->freelist
;
2519 page
->freelist
= NULL
;
2521 stat(s
, ALLOC_SLAB
);
2529 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2531 if (unlikely(PageSlabPfmemalloc(page
)))
2532 return gfp_pfmemalloc_allowed(gfpflags
);
2538 * Check the page->freelist of a page and either transfer the freelist to the
2539 * per cpu freelist or deactivate the page.
2541 * The page is still frozen if the return value is not NULL.
2543 * If this function returns NULL then the page has been unfrozen.
2545 * This function must be called with interrupt disabled.
2547 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2550 unsigned long counters
;
2554 freelist
= page
->freelist
;
2555 counters
= page
->counters
;
2557 new.counters
= counters
;
2558 VM_BUG_ON(!new.frozen
);
2560 new.inuse
= page
->objects
;
2561 new.frozen
= freelist
!= NULL
;
2563 } while (!__cmpxchg_double_slab(s
, page
,
2572 * Slow path. The lockless freelist is empty or we need to perform
2575 * Processing is still very fast if new objects have been freed to the
2576 * regular freelist. In that case we simply take over the regular freelist
2577 * as the lockless freelist and zap the regular freelist.
2579 * If that is not working then we fall back to the partial lists. We take the
2580 * first element of the freelist as the object to allocate now and move the
2581 * rest of the freelist to the lockless freelist.
2583 * And if we were unable to get a new slab from the partial slab lists then
2584 * we need to allocate a new slab. This is the slowest path since it involves
2585 * a call to the page allocator and the setup of a new slab.
2587 * Version of __slab_alloc to use when we know that interrupts are
2588 * already disabled (which is the case for bulk allocation).
2590 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2591 unsigned long addr
, struct kmem_cache_cpu
*c
)
2599 * if the node is not online or has no normal memory, just
2600 * ignore the node constraint
2602 if (unlikely(node
!= NUMA_NO_NODE
&&
2603 !node_state(node
, N_NORMAL_MEMORY
)))
2604 node
= NUMA_NO_NODE
;
2609 if (unlikely(!node_match(page
, node
))) {
2611 * same as above but node_match() being false already
2612 * implies node != NUMA_NO_NODE
2614 if (!node_state(node
, N_NORMAL_MEMORY
)) {
2615 node
= NUMA_NO_NODE
;
2618 stat(s
, ALLOC_NODE_MISMATCH
);
2619 deactivate_slab(s
, page
, c
->freelist
, c
);
2625 * By rights, we should be searching for a slab page that was
2626 * PFMEMALLOC but right now, we are losing the pfmemalloc
2627 * information when the page leaves the per-cpu allocator
2629 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2630 deactivate_slab(s
, page
, c
->freelist
, c
);
2634 /* must check again c->freelist in case of cpu migration or IRQ */
2635 freelist
= c
->freelist
;
2639 freelist
= get_freelist(s
, page
);
2643 stat(s
, DEACTIVATE_BYPASS
);
2647 stat(s
, ALLOC_REFILL
);
2651 * freelist is pointing to the list of objects to be used.
2652 * page is pointing to the page from which the objects are obtained.
2653 * That page must be frozen for per cpu allocations to work.
2655 VM_BUG_ON(!c
->page
->frozen
);
2656 c
->freelist
= get_freepointer(s
, freelist
);
2657 c
->tid
= next_tid(c
->tid
);
2662 if (slub_percpu_partial(c
)) {
2663 page
= c
->page
= slub_percpu_partial(c
);
2664 slub_set_percpu_partial(c
, page
);
2665 stat(s
, CPU_PARTIAL_ALLOC
);
2669 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2671 if (unlikely(!freelist
)) {
2672 slab_out_of_memory(s
, gfpflags
, node
);
2677 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2680 /* Only entered in the debug case */
2681 if (kmem_cache_debug(s
) &&
2682 !alloc_debug_processing(s
, page
, freelist
, addr
))
2683 goto new_slab
; /* Slab failed checks. Next slab needed */
2685 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2690 * Another one that disabled interrupt and compensates for possible
2691 * cpu changes by refetching the per cpu area pointer.
2693 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2694 unsigned long addr
, struct kmem_cache_cpu
*c
)
2697 unsigned long flags
;
2699 local_irq_save(flags
);
2700 #ifdef CONFIG_PREEMPTION
2702 * We may have been preempted and rescheduled on a different
2703 * cpu before disabling interrupts. Need to reload cpu area
2706 c
= this_cpu_ptr(s
->cpu_slab
);
2709 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2710 local_irq_restore(flags
);
2715 * If the object has been wiped upon free, make sure it's fully initialized by
2716 * zeroing out freelist pointer.
2718 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2721 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2722 memset((void *)((char *)obj
+ s
->offset
), 0, sizeof(void *));
2726 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2727 * have the fastpath folded into their functions. So no function call
2728 * overhead for requests that can be satisfied on the fastpath.
2730 * The fastpath works by first checking if the lockless freelist can be used.
2731 * If not then __slab_alloc is called for slow processing.
2733 * Otherwise we can simply pick the next object from the lockless free list.
2735 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2736 gfp_t gfpflags
, int node
, unsigned long addr
)
2739 struct kmem_cache_cpu
*c
;
2743 s
= slab_pre_alloc_hook(s
, gfpflags
);
2748 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2749 * enabled. We may switch back and forth between cpus while
2750 * reading from one cpu area. That does not matter as long
2751 * as we end up on the original cpu again when doing the cmpxchg.
2753 * We should guarantee that tid and kmem_cache are retrieved on
2754 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2755 * to check if it is matched or not.
2758 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2759 c
= raw_cpu_ptr(s
->cpu_slab
);
2760 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
2761 unlikely(tid
!= READ_ONCE(c
->tid
)));
2764 * Irqless object alloc/free algorithm used here depends on sequence
2765 * of fetching cpu_slab's data. tid should be fetched before anything
2766 * on c to guarantee that object and page associated with previous tid
2767 * won't be used with current tid. If we fetch tid first, object and
2768 * page could be one associated with next tid and our alloc/free
2769 * request will be failed. In this case, we will retry. So, no problem.
2774 * The transaction ids are globally unique per cpu and per operation on
2775 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2776 * occurs on the right processor and that there was no operation on the
2777 * linked list in between.
2780 object
= c
->freelist
;
2782 if (unlikely(!object
|| !node_match(page
, node
))) {
2783 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2784 stat(s
, ALLOC_SLOWPATH
);
2786 void *next_object
= get_freepointer_safe(s
, object
);
2789 * The cmpxchg will only match if there was no additional
2790 * operation and if we are on the right processor.
2792 * The cmpxchg does the following atomically (without lock
2794 * 1. Relocate first pointer to the current per cpu area.
2795 * 2. Verify that tid and freelist have not been changed
2796 * 3. If they were not changed replace tid and freelist
2798 * Since this is without lock semantics the protection is only
2799 * against code executing on this cpu *not* from access by
2802 if (unlikely(!this_cpu_cmpxchg_double(
2803 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2805 next_object
, next_tid(tid
)))) {
2807 note_cmpxchg_failure("slab_alloc", s
, tid
);
2810 prefetch_freepointer(s
, next_object
);
2811 stat(s
, ALLOC_FASTPATH
);
2814 maybe_wipe_obj_freeptr(s
, object
);
2816 if (unlikely(slab_want_init_on_alloc(gfpflags
, s
)) && object
)
2817 memset(object
, 0, s
->object_size
);
2819 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2824 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2825 gfp_t gfpflags
, unsigned long addr
)
2827 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2830 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2832 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2834 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2839 EXPORT_SYMBOL(kmem_cache_alloc
);
2841 #ifdef CONFIG_TRACING
2842 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2844 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2845 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2846 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2849 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2853 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2855 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2857 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2858 s
->object_size
, s
->size
, gfpflags
, node
);
2862 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2864 #ifdef CONFIG_TRACING
2865 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2867 int node
, size_t size
)
2869 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2871 trace_kmalloc_node(_RET_IP_
, ret
,
2872 size
, s
->size
, gfpflags
, node
);
2874 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2877 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2879 #endif /* CONFIG_NUMA */
2882 * Slow path handling. This may still be called frequently since objects
2883 * have a longer lifetime than the cpu slabs in most processing loads.
2885 * So we still attempt to reduce cache line usage. Just take the slab
2886 * lock and free the item. If there is no additional partial page
2887 * handling required then we can return immediately.
2889 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2890 void *head
, void *tail
, int cnt
,
2897 unsigned long counters
;
2898 struct kmem_cache_node
*n
= NULL
;
2899 unsigned long flags
;
2901 stat(s
, FREE_SLOWPATH
);
2903 if (kmem_cache_debug(s
) &&
2904 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2909 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2912 prior
= page
->freelist
;
2913 counters
= page
->counters
;
2914 set_freepointer(s
, tail
, prior
);
2915 new.counters
= counters
;
2916 was_frozen
= new.frozen
;
2918 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2920 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2923 * Slab was on no list before and will be
2925 * We can defer the list move and instead
2930 } else { /* Needs to be taken off a list */
2932 n
= get_node(s
, page_to_nid(page
));
2934 * Speculatively acquire the list_lock.
2935 * If the cmpxchg does not succeed then we may
2936 * drop the list_lock without any processing.
2938 * Otherwise the list_lock will synchronize with
2939 * other processors updating the list of slabs.
2941 spin_lock_irqsave(&n
->list_lock
, flags
);
2946 } while (!cmpxchg_double_slab(s
, page
,
2954 * If we just froze the page then put it onto the
2955 * per cpu partial list.
2957 if (new.frozen
&& !was_frozen
) {
2958 put_cpu_partial(s
, page
, 1);
2959 stat(s
, CPU_PARTIAL_FREE
);
2962 * The list lock was not taken therefore no list
2963 * activity can be necessary.
2966 stat(s
, FREE_FROZEN
);
2970 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2974 * Objects left in the slab. If it was not on the partial list before
2977 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2978 remove_full(s
, n
, page
);
2979 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2980 stat(s
, FREE_ADD_PARTIAL
);
2982 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2988 * Slab on the partial list.
2990 remove_partial(n
, page
);
2991 stat(s
, FREE_REMOVE_PARTIAL
);
2993 /* Slab must be on the full list */
2994 remove_full(s
, n
, page
);
2997 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2999 discard_slab(s
, page
);
3003 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3004 * can perform fastpath freeing without additional function calls.
3006 * The fastpath is only possible if we are freeing to the current cpu slab
3007 * of this processor. This typically the case if we have just allocated
3010 * If fastpath is not possible then fall back to __slab_free where we deal
3011 * with all sorts of special processing.
3013 * Bulk free of a freelist with several objects (all pointing to the
3014 * same page) possible by specifying head and tail ptr, plus objects
3015 * count (cnt). Bulk free indicated by tail pointer being set.
3017 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3018 struct page
*page
, void *head
, void *tail
,
3019 int cnt
, unsigned long addr
)
3021 void *tail_obj
= tail
? : head
;
3022 struct kmem_cache_cpu
*c
;
3026 * Determine the currently cpus per cpu slab.
3027 * The cpu may change afterward. However that does not matter since
3028 * data is retrieved via this pointer. If we are on the same cpu
3029 * during the cmpxchg then the free will succeed.
3032 tid
= this_cpu_read(s
->cpu_slab
->tid
);
3033 c
= raw_cpu_ptr(s
->cpu_slab
);
3034 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
3035 unlikely(tid
!= READ_ONCE(c
->tid
)));
3037 /* Same with comment on barrier() in slab_alloc_node() */
3040 if (likely(page
== c
->page
)) {
3041 void **freelist
= READ_ONCE(c
->freelist
);
3043 set_freepointer(s
, tail_obj
, freelist
);
3045 if (unlikely(!this_cpu_cmpxchg_double(
3046 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3048 head
, next_tid(tid
)))) {
3050 note_cmpxchg_failure("slab_free", s
, tid
);
3053 stat(s
, FREE_FASTPATH
);
3055 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3059 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3060 void *head
, void *tail
, int cnt
,
3064 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3065 * to remove objects, whose reuse must be delayed.
3067 if (slab_free_freelist_hook(s
, &head
, &tail
))
3068 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3071 #ifdef CONFIG_KASAN_GENERIC
3072 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3074 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3078 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3080 s
= cache_from_obj(s
, x
);
3083 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3084 trace_kmem_cache_free(_RET_IP_
, x
);
3086 EXPORT_SYMBOL(kmem_cache_free
);
3088 struct detached_freelist
{
3093 struct kmem_cache
*s
;
3097 * This function progressively scans the array with free objects (with
3098 * a limited look ahead) and extract objects belonging to the same
3099 * page. It builds a detached freelist directly within the given
3100 * page/objects. This can happen without any need for
3101 * synchronization, because the objects are owned by running process.
3102 * The freelist is build up as a single linked list in the objects.
3103 * The idea is, that this detached freelist can then be bulk
3104 * transferred to the real freelist(s), but only requiring a single
3105 * synchronization primitive. Look ahead in the array is limited due
3106 * to performance reasons.
3109 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3110 void **p
, struct detached_freelist
*df
)
3112 size_t first_skipped_index
= 0;
3117 /* Always re-init detached_freelist */
3122 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3123 } while (!object
&& size
);
3128 page
= virt_to_head_page(object
);
3130 /* Handle kalloc'ed objects */
3131 if (unlikely(!PageSlab(page
))) {
3132 BUG_ON(!PageCompound(page
));
3134 __free_pages(page
, compound_order(page
));
3135 p
[size
] = NULL
; /* mark object processed */
3138 /* Derive kmem_cache from object */
3139 df
->s
= page
->slab_cache
;
3141 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3144 /* Start new detached freelist */
3146 set_freepointer(df
->s
, object
, NULL
);
3148 df
->freelist
= object
;
3149 p
[size
] = NULL
; /* mark object processed */
3155 continue; /* Skip processed objects */
3157 /* df->page is always set at this point */
3158 if (df
->page
== virt_to_head_page(object
)) {
3159 /* Opportunity build freelist */
3160 set_freepointer(df
->s
, object
, df
->freelist
);
3161 df
->freelist
= object
;
3163 p
[size
] = NULL
; /* mark object processed */
3168 /* Limit look ahead search */
3172 if (!first_skipped_index
)
3173 first_skipped_index
= size
+ 1;
3176 return first_skipped_index
;
3179 /* Note that interrupts must be enabled when calling this function. */
3180 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3186 struct detached_freelist df
;
3188 size
= build_detached_freelist(s
, size
, p
, &df
);
3192 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3193 } while (likely(size
));
3195 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3197 /* Note that interrupts must be enabled when calling this function. */
3198 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3201 struct kmem_cache_cpu
*c
;
3204 /* memcg and kmem_cache debug support */
3205 s
= slab_pre_alloc_hook(s
, flags
);
3209 * Drain objects in the per cpu slab, while disabling local
3210 * IRQs, which protects against PREEMPT and interrupts
3211 * handlers invoking normal fastpath.
3213 local_irq_disable();
3214 c
= this_cpu_ptr(s
->cpu_slab
);
3216 for (i
= 0; i
< size
; i
++) {
3217 void *object
= c
->freelist
;
3219 if (unlikely(!object
)) {
3221 * We may have removed an object from c->freelist using
3222 * the fastpath in the previous iteration; in that case,
3223 * c->tid has not been bumped yet.
3224 * Since ___slab_alloc() may reenable interrupts while
3225 * allocating memory, we should bump c->tid now.
3227 c
->tid
= next_tid(c
->tid
);
3230 * Invoking slow path likely have side-effect
3231 * of re-populating per CPU c->freelist
3233 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3235 if (unlikely(!p
[i
]))
3238 c
= this_cpu_ptr(s
->cpu_slab
);
3239 maybe_wipe_obj_freeptr(s
, p
[i
]);
3241 continue; /* goto for-loop */
3243 c
->freelist
= get_freepointer(s
, object
);
3245 maybe_wipe_obj_freeptr(s
, p
[i
]);
3247 c
->tid
= next_tid(c
->tid
);
3250 /* Clear memory outside IRQ disabled fastpath loop */
3251 if (unlikely(slab_want_init_on_alloc(flags
, s
))) {
3254 for (j
= 0; j
< i
; j
++)
3255 memset(p
[j
], 0, s
->object_size
);
3258 /* memcg and kmem_cache debug support */
3259 slab_post_alloc_hook(s
, flags
, size
, p
);
3263 slab_post_alloc_hook(s
, flags
, i
, p
);
3264 __kmem_cache_free_bulk(s
, i
, p
);
3267 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3271 * Object placement in a slab is made very easy because we always start at
3272 * offset 0. If we tune the size of the object to the alignment then we can
3273 * get the required alignment by putting one properly sized object after
3276 * Notice that the allocation order determines the sizes of the per cpu
3277 * caches. Each processor has always one slab available for allocations.
3278 * Increasing the allocation order reduces the number of times that slabs
3279 * must be moved on and off the partial lists and is therefore a factor in
3284 * Mininum / Maximum order of slab pages. This influences locking overhead
3285 * and slab fragmentation. A higher order reduces the number of partial slabs
3286 * and increases the number of allocations possible without having to
3287 * take the list_lock.
3289 static unsigned int slub_min_order
;
3290 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3291 static unsigned int slub_min_objects
;
3294 * Calculate the order of allocation given an slab object size.
3296 * The order of allocation has significant impact on performance and other
3297 * system components. Generally order 0 allocations should be preferred since
3298 * order 0 does not cause fragmentation in the page allocator. Larger objects
3299 * be problematic to put into order 0 slabs because there may be too much
3300 * unused space left. We go to a higher order if more than 1/16th of the slab
3303 * In order to reach satisfactory performance we must ensure that a minimum
3304 * number of objects is in one slab. Otherwise we may generate too much
3305 * activity on the partial lists which requires taking the list_lock. This is
3306 * less a concern for large slabs though which are rarely used.
3308 * slub_max_order specifies the order where we begin to stop considering the
3309 * number of objects in a slab as critical. If we reach slub_max_order then
3310 * we try to keep the page order as low as possible. So we accept more waste
3311 * of space in favor of a small page order.
3313 * Higher order allocations also allow the placement of more objects in a
3314 * slab and thereby reduce object handling overhead. If the user has
3315 * requested a higher mininum order then we start with that one instead of
3316 * the smallest order which will fit the object.
3318 static inline unsigned int slab_order(unsigned int size
,
3319 unsigned int min_objects
, unsigned int max_order
,
3320 unsigned int fract_leftover
)
3322 unsigned int min_order
= slub_min_order
;
3325 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3326 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3328 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3329 order
<= max_order
; order
++) {
3331 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3334 rem
= slab_size
% size
;
3336 if (rem
<= slab_size
/ fract_leftover
)
3343 static inline int calculate_order(unsigned int size
)
3346 unsigned int min_objects
;
3347 unsigned int max_objects
;
3350 * Attempt to find best configuration for a slab. This
3351 * works by first attempting to generate a layout with
3352 * the best configuration and backing off gradually.
3354 * First we increase the acceptable waste in a slab. Then
3355 * we reduce the minimum objects required in a slab.
3357 min_objects
= slub_min_objects
;
3359 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3360 max_objects
= order_objects(slub_max_order
, size
);
3361 min_objects
= min(min_objects
, max_objects
);
3363 while (min_objects
> 1) {
3364 unsigned int fraction
;
3367 while (fraction
>= 4) {
3368 order
= slab_order(size
, min_objects
,
3369 slub_max_order
, fraction
);
3370 if (order
<= slub_max_order
)
3378 * We were unable to place multiple objects in a slab. Now
3379 * lets see if we can place a single object there.
3381 order
= slab_order(size
, 1, slub_max_order
, 1);
3382 if (order
<= slub_max_order
)
3386 * Doh this slab cannot be placed using slub_max_order.
3388 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3389 if (order
< MAX_ORDER
)
3395 init_kmem_cache_node(struct kmem_cache_node
*n
)
3398 spin_lock_init(&n
->list_lock
);
3399 INIT_LIST_HEAD(&n
->partial
);
3400 #ifdef CONFIG_SLUB_DEBUG
3401 atomic_long_set(&n
->nr_slabs
, 0);
3402 atomic_long_set(&n
->total_objects
, 0);
3403 INIT_LIST_HEAD(&n
->full
);
3407 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3409 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3410 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3413 * Must align to double word boundary for the double cmpxchg
3414 * instructions to work; see __pcpu_double_call_return_bool().
3416 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3417 2 * sizeof(void *));
3422 init_kmem_cache_cpus(s
);
3427 static struct kmem_cache
*kmem_cache_node
;
3430 * No kmalloc_node yet so do it by hand. We know that this is the first
3431 * slab on the node for this slabcache. There are no concurrent accesses
3434 * Note that this function only works on the kmem_cache_node
3435 * when allocating for the kmem_cache_node. This is used for bootstrapping
3436 * memory on a fresh node that has no slab structures yet.
3438 static void early_kmem_cache_node_alloc(int node
)
3441 struct kmem_cache_node
*n
;
3443 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3445 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3448 if (page_to_nid(page
) != node
) {
3449 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3450 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3455 #ifdef CONFIG_SLUB_DEBUG
3456 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3457 init_tracking(kmem_cache_node
, n
);
3459 n
= kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3461 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3464 kmem_cache_node
->node
[node
] = n
;
3465 init_kmem_cache_node(n
);
3466 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3469 * No locks need to be taken here as it has just been
3470 * initialized and there is no concurrent access.
3472 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3475 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3478 struct kmem_cache_node
*n
;
3480 for_each_kmem_cache_node(s
, node
, n
) {
3481 s
->node
[node
] = NULL
;
3482 kmem_cache_free(kmem_cache_node
, n
);
3486 void __kmem_cache_release(struct kmem_cache
*s
)
3488 cache_random_seq_destroy(s
);
3489 free_percpu(s
->cpu_slab
);
3490 free_kmem_cache_nodes(s
);
3493 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3497 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3498 struct kmem_cache_node
*n
;
3500 if (slab_state
== DOWN
) {
3501 early_kmem_cache_node_alloc(node
);
3504 n
= kmem_cache_alloc_node(kmem_cache_node
,
3508 free_kmem_cache_nodes(s
);
3512 init_kmem_cache_node(n
);
3518 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3520 if (min
< MIN_PARTIAL
)
3522 else if (min
> MAX_PARTIAL
)
3524 s
->min_partial
= min
;
3527 static void set_cpu_partial(struct kmem_cache
*s
)
3529 #ifdef CONFIG_SLUB_CPU_PARTIAL
3531 * cpu_partial determined the maximum number of objects kept in the
3532 * per cpu partial lists of a processor.
3534 * Per cpu partial lists mainly contain slabs that just have one
3535 * object freed. If they are used for allocation then they can be
3536 * filled up again with minimal effort. The slab will never hit the
3537 * per node partial lists and therefore no locking will be required.
3539 * This setting also determines
3541 * A) The number of objects from per cpu partial slabs dumped to the
3542 * per node list when we reach the limit.
3543 * B) The number of objects in cpu partial slabs to extract from the
3544 * per node list when we run out of per cpu objects. We only fetch
3545 * 50% to keep some capacity around for frees.
3547 if (!kmem_cache_has_cpu_partial(s
))
3548 slub_set_cpu_partial(s
, 0);
3549 else if (s
->size
>= PAGE_SIZE
)
3550 slub_set_cpu_partial(s
, 2);
3551 else if (s
->size
>= 1024)
3552 slub_set_cpu_partial(s
, 6);
3553 else if (s
->size
>= 256)
3554 slub_set_cpu_partial(s
, 13);
3556 slub_set_cpu_partial(s
, 30);
3561 * calculate_sizes() determines the order and the distribution of data within
3564 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3566 slab_flags_t flags
= s
->flags
;
3567 unsigned int size
= s
->object_size
;
3568 unsigned int freepointer_area
;
3572 * Round up object size to the next word boundary. We can only
3573 * place the free pointer at word boundaries and this determines
3574 * the possible location of the free pointer.
3576 size
= ALIGN(size
, sizeof(void *));
3578 * This is the area of the object where a freepointer can be
3579 * safely written. If redzoning adds more to the inuse size, we
3580 * can't use that portion for writing the freepointer, so
3581 * s->offset must be limited within this for the general case.
3583 freepointer_area
= size
;
3585 #ifdef CONFIG_SLUB_DEBUG
3587 * Determine if we can poison the object itself. If the user of
3588 * the slab may touch the object after free or before allocation
3589 * then we should never poison the object itself.
3591 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3593 s
->flags
|= __OBJECT_POISON
;
3595 s
->flags
&= ~__OBJECT_POISON
;
3599 * If we are Redzoning then check if there is some space between the
3600 * end of the object and the free pointer. If not then add an
3601 * additional word to have some bytes to store Redzone information.
3603 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3604 size
+= sizeof(void *);
3608 * With that we have determined the number of bytes in actual use
3609 * by the object. This is the potential offset to the free pointer.
3613 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3616 * Relocate free pointer after the object if it is not
3617 * permitted to overwrite the first word of the object on
3620 * This is the case if we do RCU, have a constructor or
3621 * destructor or are poisoning the objects.
3623 * The assumption that s->offset >= s->inuse means free
3624 * pointer is outside of the object is used in the
3625 * freeptr_outside_object() function. If that is no
3626 * longer true, the function needs to be modified.
3629 size
+= sizeof(void *);
3630 } else if (freepointer_area
> sizeof(void *)) {
3632 * Store freelist pointer near middle of object to keep
3633 * it away from the edges of the object to avoid small
3634 * sized over/underflows from neighboring allocations.
3636 s
->offset
= ALIGN(freepointer_area
/ 2, sizeof(void *));
3639 #ifdef CONFIG_SLUB_DEBUG
3640 if (flags
& SLAB_STORE_USER
)
3642 * Need to store information about allocs and frees after
3645 size
+= 2 * sizeof(struct track
);
3648 kasan_cache_create(s
, &size
, &s
->flags
);
3649 #ifdef CONFIG_SLUB_DEBUG
3650 if (flags
& SLAB_RED_ZONE
) {
3652 * Add some empty padding so that we can catch
3653 * overwrites from earlier objects rather than let
3654 * tracking information or the free pointer be
3655 * corrupted if a user writes before the start
3658 size
+= sizeof(void *);
3660 s
->red_left_pad
= sizeof(void *);
3661 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3662 size
+= s
->red_left_pad
;
3667 * SLUB stores one object immediately after another beginning from
3668 * offset 0. In order to align the objects we have to simply size
3669 * each object to conform to the alignment.
3671 size
= ALIGN(size
, s
->align
);
3673 if (forced_order
>= 0)
3674 order
= forced_order
;
3676 order
= calculate_order(size
);
3683 s
->allocflags
|= __GFP_COMP
;
3685 if (s
->flags
& SLAB_CACHE_DMA
)
3686 s
->allocflags
|= GFP_DMA
;
3688 if (s
->flags
& SLAB_CACHE_DMA32
)
3689 s
->allocflags
|= GFP_DMA32
;
3691 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3692 s
->allocflags
|= __GFP_RECLAIMABLE
;
3695 * Determine the number of objects per slab
3697 s
->oo
= oo_make(order
, size
);
3698 s
->min
= oo_make(get_order(size
), size
);
3699 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3702 return !!oo_objects(s
->oo
);
3705 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3707 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3708 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3709 s
->random
= get_random_long();
3712 if (!calculate_sizes(s
, -1))
3714 if (disable_higher_order_debug
) {
3716 * Disable debugging flags that store metadata if the min slab
3719 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3720 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3722 if (!calculate_sizes(s
, -1))
3727 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3728 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3729 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3730 /* Enable fast mode */
3731 s
->flags
|= __CMPXCHG_DOUBLE
;
3735 * The larger the object size is, the more pages we want on the partial
3736 * list to avoid pounding the page allocator excessively.
3738 set_min_partial(s
, ilog2(s
->size
) / 2);
3743 s
->remote_node_defrag_ratio
= 1000;
3746 /* Initialize the pre-computed randomized freelist if slab is up */
3747 if (slab_state
>= UP
) {
3748 if (init_cache_random_seq(s
))
3752 if (!init_kmem_cache_nodes(s
))
3755 if (alloc_kmem_cache_cpus(s
))
3758 free_kmem_cache_nodes(s
);
3763 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3766 #ifdef CONFIG_SLUB_DEBUG
3767 void *addr
= page_address(page
);
3771 slab_err(s
, page
, text
, s
->name
);
3774 map
= get_map(s
, page
);
3775 for_each_object(p
, s
, addr
, page
->objects
) {
3777 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3778 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3779 print_tracking(s
, p
);
3788 * Attempt to free all partial slabs on a node.
3789 * This is called from __kmem_cache_shutdown(). We must take list_lock
3790 * because sysfs file might still access partial list after the shutdowning.
3792 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3795 struct page
*page
, *h
;
3797 BUG_ON(irqs_disabled());
3798 spin_lock_irq(&n
->list_lock
);
3799 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3801 remove_partial(n
, page
);
3802 list_add(&page
->slab_list
, &discard
);
3804 list_slab_objects(s
, page
,
3805 "Objects remaining in %s on __kmem_cache_shutdown()");
3808 spin_unlock_irq(&n
->list_lock
);
3810 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3811 discard_slab(s
, page
);
3814 bool __kmem_cache_empty(struct kmem_cache
*s
)
3817 struct kmem_cache_node
*n
;
3819 for_each_kmem_cache_node(s
, node
, n
)
3820 if (n
->nr_partial
|| slabs_node(s
, node
))
3826 * Release all resources used by a slab cache.
3828 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3831 struct kmem_cache_node
*n
;
3834 /* Attempt to free all objects */
3835 for_each_kmem_cache_node(s
, node
, n
) {
3837 if (n
->nr_partial
|| slabs_node(s
, node
))
3840 sysfs_slab_remove(s
);
3844 /********************************************************************
3846 *******************************************************************/
3848 static int __init
setup_slub_min_order(char *str
)
3850 get_option(&str
, (int *)&slub_min_order
);
3855 __setup("slub_min_order=", setup_slub_min_order
);
3857 static int __init
setup_slub_max_order(char *str
)
3859 get_option(&str
, (int *)&slub_max_order
);
3860 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
3865 __setup("slub_max_order=", setup_slub_max_order
);
3867 static int __init
setup_slub_min_objects(char *str
)
3869 get_option(&str
, (int *)&slub_min_objects
);
3874 __setup("slub_min_objects=", setup_slub_min_objects
);
3876 void *__kmalloc(size_t size
, gfp_t flags
)
3878 struct kmem_cache
*s
;
3881 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3882 return kmalloc_large(size
, flags
);
3884 s
= kmalloc_slab(size
, flags
);
3886 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3889 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3891 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3893 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3897 EXPORT_SYMBOL(__kmalloc
);
3900 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3904 unsigned int order
= get_order(size
);
3906 flags
|= __GFP_COMP
;
3907 page
= alloc_pages_node(node
, flags
, order
);
3909 ptr
= page_address(page
);
3910 mod_node_page_state(page_pgdat(page
), NR_SLAB_UNRECLAIMABLE
,
3914 return kmalloc_large_node_hook(ptr
, size
, flags
);
3917 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3919 struct kmem_cache
*s
;
3922 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3923 ret
= kmalloc_large_node(size
, flags
, node
);
3925 trace_kmalloc_node(_RET_IP_
, ret
,
3926 size
, PAGE_SIZE
<< get_order(size
),
3932 s
= kmalloc_slab(size
, flags
);
3934 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3937 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3939 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3941 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3945 EXPORT_SYMBOL(__kmalloc_node
);
3946 #endif /* CONFIG_NUMA */
3948 #ifdef CONFIG_HARDENED_USERCOPY
3950 * Rejects incorrectly sized objects and objects that are to be copied
3951 * to/from userspace but do not fall entirely within the containing slab
3952 * cache's usercopy region.
3954 * Returns NULL if check passes, otherwise const char * to name of cache
3955 * to indicate an error.
3957 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
3960 struct kmem_cache
*s
;
3961 unsigned int offset
;
3964 ptr
= kasan_reset_tag(ptr
);
3966 /* Find object and usable object size. */
3967 s
= page
->slab_cache
;
3969 /* Reject impossible pointers. */
3970 if (ptr
< page_address(page
))
3971 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
3974 /* Find offset within object. */
3975 offset
= (ptr
- page_address(page
)) % s
->size
;
3977 /* Adjust for redzone and reject if within the redzone. */
3978 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3979 if (offset
< s
->red_left_pad
)
3980 usercopy_abort("SLUB object in left red zone",
3981 s
->name
, to_user
, offset
, n
);
3982 offset
-= s
->red_left_pad
;
3985 /* Allow address range falling entirely within usercopy region. */
3986 if (offset
>= s
->useroffset
&&
3987 offset
- s
->useroffset
<= s
->usersize
&&
3988 n
<= s
->useroffset
- offset
+ s
->usersize
)
3992 * If the copy is still within the allocated object, produce
3993 * a warning instead of rejecting the copy. This is intended
3994 * to be a temporary method to find any missing usercopy
3997 object_size
= slab_ksize(s
);
3998 if (usercopy_fallback
&&
3999 offset
<= object_size
&& n
<= object_size
- offset
) {
4000 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4004 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4006 #endif /* CONFIG_HARDENED_USERCOPY */
4008 size_t __ksize(const void *object
)
4012 if (unlikely(object
== ZERO_SIZE_PTR
))
4015 page
= virt_to_head_page(object
);
4017 if (unlikely(!PageSlab(page
))) {
4018 WARN_ON(!PageCompound(page
));
4019 return page_size(page
);
4022 return slab_ksize(page
->slab_cache
);
4024 EXPORT_SYMBOL(__ksize
);
4026 void kfree(const void *x
)
4029 void *object
= (void *)x
;
4031 trace_kfree(_RET_IP_
, x
);
4033 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4036 page
= virt_to_head_page(x
);
4037 if (unlikely(!PageSlab(page
))) {
4038 unsigned int order
= compound_order(page
);
4040 BUG_ON(!PageCompound(page
));
4042 mod_node_page_state(page_pgdat(page
), NR_SLAB_UNRECLAIMABLE
,
4044 __free_pages(page
, order
);
4047 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4049 EXPORT_SYMBOL(kfree
);
4051 #define SHRINK_PROMOTE_MAX 32
4054 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4055 * up most to the head of the partial lists. New allocations will then
4056 * fill those up and thus they can be removed from the partial lists.
4058 * The slabs with the least items are placed last. This results in them
4059 * being allocated from last increasing the chance that the last objects
4060 * are freed in them.
4062 int __kmem_cache_shrink(struct kmem_cache
*s
)
4066 struct kmem_cache_node
*n
;
4069 struct list_head discard
;
4070 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4071 unsigned long flags
;
4075 for_each_kmem_cache_node(s
, node
, n
) {
4076 INIT_LIST_HEAD(&discard
);
4077 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4078 INIT_LIST_HEAD(promote
+ i
);
4080 spin_lock_irqsave(&n
->list_lock
, flags
);
4083 * Build lists of slabs to discard or promote.
4085 * Note that concurrent frees may occur while we hold the
4086 * list_lock. page->inuse here is the upper limit.
4088 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4089 int free
= page
->objects
- page
->inuse
;
4091 /* Do not reread page->inuse */
4094 /* We do not keep full slabs on the list */
4097 if (free
== page
->objects
) {
4098 list_move(&page
->slab_list
, &discard
);
4100 } else if (free
<= SHRINK_PROMOTE_MAX
)
4101 list_move(&page
->slab_list
, promote
+ free
- 1);
4105 * Promote the slabs filled up most to the head of the
4108 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4109 list_splice(promote
+ i
, &n
->partial
);
4111 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4113 /* Release empty slabs */
4114 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4115 discard_slab(s
, page
);
4117 if (slabs_node(s
, node
))
4125 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache
*s
)
4128 * Called with all the locks held after a sched RCU grace period.
4129 * Even if @s becomes empty after shrinking, we can't know that @s
4130 * doesn't have allocations already in-flight and thus can't
4131 * destroy @s until the associated memcg is released.
4133 * However, let's remove the sysfs files for empty caches here.
4134 * Each cache has a lot of interface files which aren't
4135 * particularly useful for empty draining caches; otherwise, we can
4136 * easily end up with millions of unnecessary sysfs files on
4137 * systems which have a lot of memory and transient cgroups.
4139 if (!__kmem_cache_shrink(s
))
4140 sysfs_slab_remove(s
);
4143 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4146 * Disable empty slabs caching. Used to avoid pinning offline
4147 * memory cgroups by kmem pages that can be freed.
4149 slub_set_cpu_partial(s
, 0);
4152 #endif /* CONFIG_MEMCG */
4154 static int slab_mem_going_offline_callback(void *arg
)
4156 struct kmem_cache
*s
;
4158 mutex_lock(&slab_mutex
);
4159 list_for_each_entry(s
, &slab_caches
, list
)
4160 __kmem_cache_shrink(s
);
4161 mutex_unlock(&slab_mutex
);
4166 static void slab_mem_offline_callback(void *arg
)
4168 struct kmem_cache_node
*n
;
4169 struct kmem_cache
*s
;
4170 struct memory_notify
*marg
= arg
;
4173 offline_node
= marg
->status_change_nid_normal
;
4176 * If the node still has available memory. we need kmem_cache_node
4179 if (offline_node
< 0)
4182 mutex_lock(&slab_mutex
);
4183 list_for_each_entry(s
, &slab_caches
, list
) {
4184 n
= get_node(s
, offline_node
);
4187 * if n->nr_slabs > 0, slabs still exist on the node
4188 * that is going down. We were unable to free them,
4189 * and offline_pages() function shouldn't call this
4190 * callback. So, we must fail.
4192 BUG_ON(slabs_node(s
, offline_node
));
4194 s
->node
[offline_node
] = NULL
;
4195 kmem_cache_free(kmem_cache_node
, n
);
4198 mutex_unlock(&slab_mutex
);
4201 static int slab_mem_going_online_callback(void *arg
)
4203 struct kmem_cache_node
*n
;
4204 struct kmem_cache
*s
;
4205 struct memory_notify
*marg
= arg
;
4206 int nid
= marg
->status_change_nid_normal
;
4210 * If the node's memory is already available, then kmem_cache_node is
4211 * already created. Nothing to do.
4217 * We are bringing a node online. No memory is available yet. We must
4218 * allocate a kmem_cache_node structure in order to bring the node
4221 mutex_lock(&slab_mutex
);
4222 list_for_each_entry(s
, &slab_caches
, list
) {
4224 * XXX: kmem_cache_alloc_node will fallback to other nodes
4225 * since memory is not yet available from the node that
4228 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4233 init_kmem_cache_node(n
);
4237 mutex_unlock(&slab_mutex
);
4241 static int slab_memory_callback(struct notifier_block
*self
,
4242 unsigned long action
, void *arg
)
4247 case MEM_GOING_ONLINE
:
4248 ret
= slab_mem_going_online_callback(arg
);
4250 case MEM_GOING_OFFLINE
:
4251 ret
= slab_mem_going_offline_callback(arg
);
4254 case MEM_CANCEL_ONLINE
:
4255 slab_mem_offline_callback(arg
);
4258 case MEM_CANCEL_OFFLINE
:
4262 ret
= notifier_from_errno(ret
);
4268 static struct notifier_block slab_memory_callback_nb
= {
4269 .notifier_call
= slab_memory_callback
,
4270 .priority
= SLAB_CALLBACK_PRI
,
4273 /********************************************************************
4274 * Basic setup of slabs
4275 *******************************************************************/
4278 * Used for early kmem_cache structures that were allocated using
4279 * the page allocator. Allocate them properly then fix up the pointers
4280 * that may be pointing to the wrong kmem_cache structure.
4283 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4286 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4287 struct kmem_cache_node
*n
;
4289 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4292 * This runs very early, and only the boot processor is supposed to be
4293 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4296 __flush_cpu_slab(s
, smp_processor_id());
4297 for_each_kmem_cache_node(s
, node
, n
) {
4300 list_for_each_entry(p
, &n
->partial
, slab_list
)
4303 #ifdef CONFIG_SLUB_DEBUG
4304 list_for_each_entry(p
, &n
->full
, slab_list
)
4308 slab_init_memcg_params(s
);
4309 list_add(&s
->list
, &slab_caches
);
4310 memcg_link_cache(s
, NULL
);
4314 void __init
kmem_cache_init(void)
4316 static __initdata
struct kmem_cache boot_kmem_cache
,
4317 boot_kmem_cache_node
;
4319 if (debug_guardpage_minorder())
4322 kmem_cache_node
= &boot_kmem_cache_node
;
4323 kmem_cache
= &boot_kmem_cache
;
4325 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4326 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4328 register_hotmemory_notifier(&slab_memory_callback_nb
);
4330 /* Able to allocate the per node structures */
4331 slab_state
= PARTIAL
;
4333 create_boot_cache(kmem_cache
, "kmem_cache",
4334 offsetof(struct kmem_cache
, node
) +
4335 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4336 SLAB_HWCACHE_ALIGN
, 0, 0);
4338 kmem_cache
= bootstrap(&boot_kmem_cache
);
4339 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4341 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4342 setup_kmalloc_cache_index_table();
4343 create_kmalloc_caches(0);
4345 /* Setup random freelists for each cache */
4346 init_freelist_randomization();
4348 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4351 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4353 slub_min_order
, slub_max_order
, slub_min_objects
,
4354 nr_cpu_ids
, nr_node_ids
);
4357 void __init
kmem_cache_init_late(void)
4362 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4363 slab_flags_t flags
, void (*ctor
)(void *))
4365 struct kmem_cache
*s
, *c
;
4367 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4372 * Adjust the object sizes so that we clear
4373 * the complete object on kzalloc.
4375 s
->object_size
= max(s
->object_size
, size
);
4376 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4378 for_each_memcg_cache(c
, s
) {
4379 c
->object_size
= s
->object_size
;
4380 c
->inuse
= max(c
->inuse
, ALIGN(size
, sizeof(void *)));
4383 if (sysfs_slab_alias(s
, name
)) {
4392 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4396 err
= kmem_cache_open(s
, flags
);
4400 /* Mutex is not taken during early boot */
4401 if (slab_state
<= UP
)
4404 memcg_propagate_slab_attrs(s
);
4405 err
= sysfs_slab_add(s
);
4407 __kmem_cache_release(s
);
4412 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4414 struct kmem_cache
*s
;
4417 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4418 return kmalloc_large(size
, gfpflags
);
4420 s
= kmalloc_slab(size
, gfpflags
);
4422 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4425 ret
= slab_alloc(s
, gfpflags
, caller
);
4427 /* Honor the call site pointer we received. */
4428 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4432 EXPORT_SYMBOL(__kmalloc_track_caller
);
4435 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4436 int node
, unsigned long caller
)
4438 struct kmem_cache
*s
;
4441 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4442 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4444 trace_kmalloc_node(caller
, ret
,
4445 size
, PAGE_SIZE
<< get_order(size
),
4451 s
= kmalloc_slab(size
, gfpflags
);
4453 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4456 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4458 /* Honor the call site pointer we received. */
4459 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4463 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4467 static int count_inuse(struct page
*page
)
4472 static int count_total(struct page
*page
)
4474 return page
->objects
;
4478 #ifdef CONFIG_SLUB_DEBUG
4479 static void validate_slab(struct kmem_cache
*s
, struct page
*page
)
4482 void *addr
= page_address(page
);
4487 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4490 /* Now we know that a valid freelist exists */
4491 map
= get_map(s
, page
);
4492 for_each_object(p
, s
, addr
, page
->objects
) {
4493 u8 val
= test_bit(slab_index(p
, s
, addr
), map
) ?
4494 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4496 if (!check_object(s
, page
, p
, val
))
4504 static int validate_slab_node(struct kmem_cache
*s
,
4505 struct kmem_cache_node
*n
)
4507 unsigned long count
= 0;
4509 unsigned long flags
;
4511 spin_lock_irqsave(&n
->list_lock
, flags
);
4513 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4514 validate_slab(s
, page
);
4517 if (count
!= n
->nr_partial
)
4518 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4519 s
->name
, count
, n
->nr_partial
);
4521 if (!(s
->flags
& SLAB_STORE_USER
))
4524 list_for_each_entry(page
, &n
->full
, slab_list
) {
4525 validate_slab(s
, page
);
4528 if (count
!= atomic_long_read(&n
->nr_slabs
))
4529 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4530 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4533 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4537 static long validate_slab_cache(struct kmem_cache
*s
)
4540 unsigned long count
= 0;
4541 struct kmem_cache_node
*n
;
4544 for_each_kmem_cache_node(s
, node
, n
)
4545 count
+= validate_slab_node(s
, n
);
4550 * Generate lists of code addresses where slabcache objects are allocated
4555 unsigned long count
;
4562 DECLARE_BITMAP(cpus
, NR_CPUS
);
4568 unsigned long count
;
4569 struct location
*loc
;
4572 static void free_loc_track(struct loc_track
*t
)
4575 free_pages((unsigned long)t
->loc
,
4576 get_order(sizeof(struct location
) * t
->max
));
4579 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4584 order
= get_order(sizeof(struct location
) * max
);
4586 l
= (void *)__get_free_pages(flags
, order
);
4591 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4599 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4600 const struct track
*track
)
4602 long start
, end
, pos
;
4604 unsigned long caddr
;
4605 unsigned long age
= jiffies
- track
->when
;
4611 pos
= start
+ (end
- start
+ 1) / 2;
4614 * There is nothing at "end". If we end up there
4615 * we need to add something to before end.
4620 caddr
= t
->loc
[pos
].addr
;
4621 if (track
->addr
== caddr
) {
4627 if (age
< l
->min_time
)
4629 if (age
> l
->max_time
)
4632 if (track
->pid
< l
->min_pid
)
4633 l
->min_pid
= track
->pid
;
4634 if (track
->pid
> l
->max_pid
)
4635 l
->max_pid
= track
->pid
;
4637 cpumask_set_cpu(track
->cpu
,
4638 to_cpumask(l
->cpus
));
4640 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4644 if (track
->addr
< caddr
)
4651 * Not found. Insert new tracking element.
4653 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4659 (t
->count
- pos
) * sizeof(struct location
));
4662 l
->addr
= track
->addr
;
4666 l
->min_pid
= track
->pid
;
4667 l
->max_pid
= track
->pid
;
4668 cpumask_clear(to_cpumask(l
->cpus
));
4669 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4670 nodes_clear(l
->nodes
);
4671 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4675 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4676 struct page
*page
, enum track_item alloc
)
4678 void *addr
= page_address(page
);
4682 map
= get_map(s
, page
);
4683 for_each_object(p
, s
, addr
, page
->objects
)
4684 if (!test_bit(slab_index(p
, s
, addr
), map
))
4685 add_location(t
, s
, get_track(s
, p
, alloc
));
4689 static int list_locations(struct kmem_cache
*s
, char *buf
,
4690 enum track_item alloc
)
4694 struct loc_track t
= { 0, 0, NULL
};
4696 struct kmem_cache_node
*n
;
4698 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4700 return sprintf(buf
, "Out of memory\n");
4702 /* Push back cpu slabs */
4705 for_each_kmem_cache_node(s
, node
, n
) {
4706 unsigned long flags
;
4709 if (!atomic_long_read(&n
->nr_slabs
))
4712 spin_lock_irqsave(&n
->list_lock
, flags
);
4713 list_for_each_entry(page
, &n
->partial
, slab_list
)
4714 process_slab(&t
, s
, page
, alloc
);
4715 list_for_each_entry(page
, &n
->full
, slab_list
)
4716 process_slab(&t
, s
, page
, alloc
);
4717 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4720 for (i
= 0; i
< t
.count
; i
++) {
4721 struct location
*l
= &t
.loc
[i
];
4723 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4725 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4728 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4730 len
+= sprintf(buf
+ len
, "<not-available>");
4732 if (l
->sum_time
!= l
->min_time
) {
4733 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4735 (long)div_u64(l
->sum_time
, l
->count
),
4738 len
+= sprintf(buf
+ len
, " age=%ld",
4741 if (l
->min_pid
!= l
->max_pid
)
4742 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4743 l
->min_pid
, l
->max_pid
);
4745 len
+= sprintf(buf
+ len
, " pid=%ld",
4748 if (num_online_cpus() > 1 &&
4749 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4750 len
< PAGE_SIZE
- 60)
4751 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4753 cpumask_pr_args(to_cpumask(l
->cpus
)));
4755 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4756 len
< PAGE_SIZE
- 60)
4757 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4759 nodemask_pr_args(&l
->nodes
));
4761 len
+= sprintf(buf
+ len
, "\n");
4766 len
+= sprintf(buf
, "No data\n");
4769 #endif /* CONFIG_SLUB_DEBUG */
4771 #ifdef SLUB_RESILIENCY_TEST
4772 static void __init
resiliency_test(void)
4775 int type
= KMALLOC_NORMAL
;
4777 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4779 pr_err("SLUB resiliency testing\n");
4780 pr_err("-----------------------\n");
4781 pr_err("A. Corruption after allocation\n");
4783 p
= kzalloc(16, GFP_KERNEL
);
4785 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4788 validate_slab_cache(kmalloc_caches
[type
][4]);
4790 /* Hmmm... The next two are dangerous */
4791 p
= kzalloc(32, GFP_KERNEL
);
4792 p
[32 + sizeof(void *)] = 0x34;
4793 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4795 pr_err("If allocated object is overwritten then not detectable\n\n");
4797 validate_slab_cache(kmalloc_caches
[type
][5]);
4798 p
= kzalloc(64, GFP_KERNEL
);
4799 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4801 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4803 pr_err("If allocated object is overwritten then not detectable\n\n");
4804 validate_slab_cache(kmalloc_caches
[type
][6]);
4806 pr_err("\nB. Corruption after free\n");
4807 p
= kzalloc(128, GFP_KERNEL
);
4810 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4811 validate_slab_cache(kmalloc_caches
[type
][7]);
4813 p
= kzalloc(256, GFP_KERNEL
);
4816 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4817 validate_slab_cache(kmalloc_caches
[type
][8]);
4819 p
= kzalloc(512, GFP_KERNEL
);
4822 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4823 validate_slab_cache(kmalloc_caches
[type
][9]);
4827 static void resiliency_test(void) {};
4829 #endif /* SLUB_RESILIENCY_TEST */
4832 enum slab_stat_type
{
4833 SL_ALL
, /* All slabs */
4834 SL_PARTIAL
, /* Only partially allocated slabs */
4835 SL_CPU
, /* Only slabs used for cpu caches */
4836 SL_OBJECTS
, /* Determine allocated objects not slabs */
4837 SL_TOTAL
/* Determine object capacity not slabs */
4840 #define SO_ALL (1 << SL_ALL)
4841 #define SO_PARTIAL (1 << SL_PARTIAL)
4842 #define SO_CPU (1 << SL_CPU)
4843 #define SO_OBJECTS (1 << SL_OBJECTS)
4844 #define SO_TOTAL (1 << SL_TOTAL)
4847 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4849 static int __init
setup_slub_memcg_sysfs(char *str
)
4853 if (get_option(&str
, &v
) > 0)
4854 memcg_sysfs_enabled
= v
;
4859 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4862 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4863 char *buf
, unsigned long flags
)
4865 unsigned long total
= 0;
4868 unsigned long *nodes
;
4870 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4874 if (flags
& SO_CPU
) {
4877 for_each_possible_cpu(cpu
) {
4878 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4883 page
= READ_ONCE(c
->page
);
4887 node
= page_to_nid(page
);
4888 if (flags
& SO_TOTAL
)
4890 else if (flags
& SO_OBJECTS
)
4898 page
= slub_percpu_partial_read_once(c
);
4900 node
= page_to_nid(page
);
4901 if (flags
& SO_TOTAL
)
4903 else if (flags
& SO_OBJECTS
)
4914 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4915 * already held which will conflict with an existing lock order:
4917 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4919 * We don't really need mem_hotplug_lock (to hold off
4920 * slab_mem_going_offline_callback) here because slab's memory hot
4921 * unplug code doesn't destroy the kmem_cache->node[] data.
4924 #ifdef CONFIG_SLUB_DEBUG
4925 if (flags
& SO_ALL
) {
4926 struct kmem_cache_node
*n
;
4928 for_each_kmem_cache_node(s
, node
, n
) {
4930 if (flags
& SO_TOTAL
)
4931 x
= atomic_long_read(&n
->total_objects
);
4932 else if (flags
& SO_OBJECTS
)
4933 x
= atomic_long_read(&n
->total_objects
) -
4934 count_partial(n
, count_free
);
4936 x
= atomic_long_read(&n
->nr_slabs
);
4943 if (flags
& SO_PARTIAL
) {
4944 struct kmem_cache_node
*n
;
4946 for_each_kmem_cache_node(s
, node
, n
) {
4947 if (flags
& SO_TOTAL
)
4948 x
= count_partial(n
, count_total
);
4949 else if (flags
& SO_OBJECTS
)
4950 x
= count_partial(n
, count_inuse
);
4957 x
= sprintf(buf
, "%lu", total
);
4959 for (node
= 0; node
< nr_node_ids
; node
++)
4961 x
+= sprintf(buf
+ x
, " N%d=%lu",
4965 return x
+ sprintf(buf
+ x
, "\n");
4968 #ifdef CONFIG_SLUB_DEBUG
4969 static int any_slab_objects(struct kmem_cache
*s
)
4972 struct kmem_cache_node
*n
;
4974 for_each_kmem_cache_node(s
, node
, n
)
4975 if (atomic_long_read(&n
->total_objects
))
4982 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4983 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4985 struct slab_attribute
{
4986 struct attribute attr
;
4987 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4988 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4991 #define SLAB_ATTR_RO(_name) \
4992 static struct slab_attribute _name##_attr = \
4993 __ATTR(_name, 0400, _name##_show, NULL)
4995 #define SLAB_ATTR(_name) \
4996 static struct slab_attribute _name##_attr = \
4997 __ATTR(_name, 0600, _name##_show, _name##_store)
4999 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5001 return sprintf(buf
, "%u\n", s
->size
);
5003 SLAB_ATTR_RO(slab_size
);
5005 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5007 return sprintf(buf
, "%u\n", s
->align
);
5009 SLAB_ATTR_RO(align
);
5011 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5013 return sprintf(buf
, "%u\n", s
->object_size
);
5015 SLAB_ATTR_RO(object_size
);
5017 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5019 return sprintf(buf
, "%u\n", oo_objects(s
->oo
));
5021 SLAB_ATTR_RO(objs_per_slab
);
5023 static ssize_t
order_store(struct kmem_cache
*s
,
5024 const char *buf
, size_t length
)
5029 err
= kstrtouint(buf
, 10, &order
);
5033 if (order
> slub_max_order
|| order
< slub_min_order
)
5036 calculate_sizes(s
, order
);
5040 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5042 return sprintf(buf
, "%u\n", oo_order(s
->oo
));
5046 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5048 return sprintf(buf
, "%lu\n", s
->min_partial
);
5051 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5057 err
= kstrtoul(buf
, 10, &min
);
5061 set_min_partial(s
, min
);
5064 SLAB_ATTR(min_partial
);
5066 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5068 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
5071 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5074 unsigned int objects
;
5077 err
= kstrtouint(buf
, 10, &objects
);
5080 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5083 slub_set_cpu_partial(s
, objects
);
5087 SLAB_ATTR(cpu_partial
);
5089 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5093 return sprintf(buf
, "%pS\n", s
->ctor
);
5097 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5099 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5101 SLAB_ATTR_RO(aliases
);
5103 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5105 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5107 SLAB_ATTR_RO(partial
);
5109 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5111 return show_slab_objects(s
, buf
, SO_CPU
);
5113 SLAB_ATTR_RO(cpu_slabs
);
5115 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5117 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5119 SLAB_ATTR_RO(objects
);
5121 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5123 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5125 SLAB_ATTR_RO(objects_partial
);
5127 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5134 for_each_online_cpu(cpu
) {
5137 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5140 pages
+= page
->pages
;
5141 objects
+= page
->pobjects
;
5145 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5148 for_each_online_cpu(cpu
) {
5151 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5153 if (page
&& len
< PAGE_SIZE
- 20)
5154 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5155 page
->pobjects
, page
->pages
);
5158 return len
+ sprintf(buf
+ len
, "\n");
5160 SLAB_ATTR_RO(slabs_cpu_partial
);
5162 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5164 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5167 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5168 const char *buf
, size_t length
)
5170 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5172 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5175 SLAB_ATTR(reclaim_account
);
5177 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5179 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5181 SLAB_ATTR_RO(hwcache_align
);
5183 #ifdef CONFIG_ZONE_DMA
5184 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5186 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5188 SLAB_ATTR_RO(cache_dma
);
5191 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5193 return sprintf(buf
, "%u\n", s
->usersize
);
5195 SLAB_ATTR_RO(usersize
);
5197 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5199 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5201 SLAB_ATTR_RO(destroy_by_rcu
);
5203 #ifdef CONFIG_SLUB_DEBUG
5204 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5206 return show_slab_objects(s
, buf
, SO_ALL
);
5208 SLAB_ATTR_RO(slabs
);
5210 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5212 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5214 SLAB_ATTR_RO(total_objects
);
5216 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5218 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5221 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5222 const char *buf
, size_t length
)
5224 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5225 if (buf
[0] == '1') {
5226 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5227 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5231 SLAB_ATTR(sanity_checks
);
5233 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5235 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5238 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5242 * Tracing a merged cache is going to give confusing results
5243 * as well as cause other issues like converting a mergeable
5244 * cache into an umergeable one.
5246 if (s
->refcount
> 1)
5249 s
->flags
&= ~SLAB_TRACE
;
5250 if (buf
[0] == '1') {
5251 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5252 s
->flags
|= SLAB_TRACE
;
5258 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5260 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5263 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5264 const char *buf
, size_t length
)
5266 if (any_slab_objects(s
))
5269 s
->flags
&= ~SLAB_RED_ZONE
;
5270 if (buf
[0] == '1') {
5271 s
->flags
|= SLAB_RED_ZONE
;
5273 calculate_sizes(s
, -1);
5276 SLAB_ATTR(red_zone
);
5278 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5280 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5283 static ssize_t
poison_store(struct kmem_cache
*s
,
5284 const char *buf
, size_t length
)
5286 if (any_slab_objects(s
))
5289 s
->flags
&= ~SLAB_POISON
;
5290 if (buf
[0] == '1') {
5291 s
->flags
|= SLAB_POISON
;
5293 calculate_sizes(s
, -1);
5298 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5300 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5303 static ssize_t
store_user_store(struct kmem_cache
*s
,
5304 const char *buf
, size_t length
)
5306 if (any_slab_objects(s
))
5309 s
->flags
&= ~SLAB_STORE_USER
;
5310 if (buf
[0] == '1') {
5311 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5312 s
->flags
|= SLAB_STORE_USER
;
5314 calculate_sizes(s
, -1);
5317 SLAB_ATTR(store_user
);
5319 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5324 static ssize_t
validate_store(struct kmem_cache
*s
,
5325 const char *buf
, size_t length
)
5329 if (buf
[0] == '1') {
5330 ret
= validate_slab_cache(s
);
5336 SLAB_ATTR(validate
);
5338 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5340 if (!(s
->flags
& SLAB_STORE_USER
))
5342 return list_locations(s
, buf
, TRACK_ALLOC
);
5344 SLAB_ATTR_RO(alloc_calls
);
5346 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5348 if (!(s
->flags
& SLAB_STORE_USER
))
5350 return list_locations(s
, buf
, TRACK_FREE
);
5352 SLAB_ATTR_RO(free_calls
);
5353 #endif /* CONFIG_SLUB_DEBUG */
5355 #ifdef CONFIG_FAILSLAB
5356 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5358 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5361 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5364 if (s
->refcount
> 1)
5367 s
->flags
&= ~SLAB_FAILSLAB
;
5369 s
->flags
|= SLAB_FAILSLAB
;
5372 SLAB_ATTR(failslab
);
5375 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5380 static ssize_t
shrink_store(struct kmem_cache
*s
,
5381 const char *buf
, size_t length
)
5384 kmem_cache_shrink_all(s
);
5392 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5394 return sprintf(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5397 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5398 const char *buf
, size_t length
)
5403 err
= kstrtouint(buf
, 10, &ratio
);
5409 s
->remote_node_defrag_ratio
= ratio
* 10;
5413 SLAB_ATTR(remote_node_defrag_ratio
);
5416 #ifdef CONFIG_SLUB_STATS
5417 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5419 unsigned long sum
= 0;
5422 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5427 for_each_online_cpu(cpu
) {
5428 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5434 len
= sprintf(buf
, "%lu", sum
);
5437 for_each_online_cpu(cpu
) {
5438 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5439 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5443 return len
+ sprintf(buf
+ len
, "\n");
5446 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5450 for_each_online_cpu(cpu
)
5451 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5454 #define STAT_ATTR(si, text) \
5455 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5457 return show_stat(s, buf, si); \
5459 static ssize_t text##_store(struct kmem_cache *s, \
5460 const char *buf, size_t length) \
5462 if (buf[0] != '0') \
5464 clear_stat(s, si); \
5469 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5470 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5471 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5472 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5473 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5474 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5475 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5476 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5477 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5478 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5479 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5480 STAT_ATTR(FREE_SLAB
, free_slab
);
5481 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5482 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5483 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5484 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5485 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5486 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5487 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5488 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5489 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5490 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5491 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5492 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5493 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5494 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5495 #endif /* CONFIG_SLUB_STATS */
5497 static struct attribute
*slab_attrs
[] = {
5498 &slab_size_attr
.attr
,
5499 &object_size_attr
.attr
,
5500 &objs_per_slab_attr
.attr
,
5502 &min_partial_attr
.attr
,
5503 &cpu_partial_attr
.attr
,
5505 &objects_partial_attr
.attr
,
5507 &cpu_slabs_attr
.attr
,
5511 &hwcache_align_attr
.attr
,
5512 &reclaim_account_attr
.attr
,
5513 &destroy_by_rcu_attr
.attr
,
5515 &slabs_cpu_partial_attr
.attr
,
5516 #ifdef CONFIG_SLUB_DEBUG
5517 &total_objects_attr
.attr
,
5519 &sanity_checks_attr
.attr
,
5521 &red_zone_attr
.attr
,
5523 &store_user_attr
.attr
,
5524 &validate_attr
.attr
,
5525 &alloc_calls_attr
.attr
,
5526 &free_calls_attr
.attr
,
5528 #ifdef CONFIG_ZONE_DMA
5529 &cache_dma_attr
.attr
,
5532 &remote_node_defrag_ratio_attr
.attr
,
5534 #ifdef CONFIG_SLUB_STATS
5535 &alloc_fastpath_attr
.attr
,
5536 &alloc_slowpath_attr
.attr
,
5537 &free_fastpath_attr
.attr
,
5538 &free_slowpath_attr
.attr
,
5539 &free_frozen_attr
.attr
,
5540 &free_add_partial_attr
.attr
,
5541 &free_remove_partial_attr
.attr
,
5542 &alloc_from_partial_attr
.attr
,
5543 &alloc_slab_attr
.attr
,
5544 &alloc_refill_attr
.attr
,
5545 &alloc_node_mismatch_attr
.attr
,
5546 &free_slab_attr
.attr
,
5547 &cpuslab_flush_attr
.attr
,
5548 &deactivate_full_attr
.attr
,
5549 &deactivate_empty_attr
.attr
,
5550 &deactivate_to_head_attr
.attr
,
5551 &deactivate_to_tail_attr
.attr
,
5552 &deactivate_remote_frees_attr
.attr
,
5553 &deactivate_bypass_attr
.attr
,
5554 &order_fallback_attr
.attr
,
5555 &cmpxchg_double_fail_attr
.attr
,
5556 &cmpxchg_double_cpu_fail_attr
.attr
,
5557 &cpu_partial_alloc_attr
.attr
,
5558 &cpu_partial_free_attr
.attr
,
5559 &cpu_partial_node_attr
.attr
,
5560 &cpu_partial_drain_attr
.attr
,
5562 #ifdef CONFIG_FAILSLAB
5563 &failslab_attr
.attr
,
5565 &usersize_attr
.attr
,
5570 static const struct attribute_group slab_attr_group
= {
5571 .attrs
= slab_attrs
,
5574 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5575 struct attribute
*attr
,
5578 struct slab_attribute
*attribute
;
5579 struct kmem_cache
*s
;
5582 attribute
= to_slab_attr(attr
);
5585 if (!attribute
->show
)
5588 err
= attribute
->show(s
, buf
);
5593 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5594 struct attribute
*attr
,
5595 const char *buf
, size_t len
)
5597 struct slab_attribute
*attribute
;
5598 struct kmem_cache
*s
;
5601 attribute
= to_slab_attr(attr
);
5604 if (!attribute
->store
)
5607 err
= attribute
->store(s
, buf
, len
);
5609 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5610 struct kmem_cache
*c
;
5612 mutex_lock(&slab_mutex
);
5613 if (s
->max_attr_size
< len
)
5614 s
->max_attr_size
= len
;
5617 * This is a best effort propagation, so this function's return
5618 * value will be determined by the parent cache only. This is
5619 * basically because not all attributes will have a well
5620 * defined semantics for rollbacks - most of the actions will
5621 * have permanent effects.
5623 * Returning the error value of any of the children that fail
5624 * is not 100 % defined, in the sense that users seeing the
5625 * error code won't be able to know anything about the state of
5628 * Only returning the error code for the parent cache at least
5629 * has well defined semantics. The cache being written to
5630 * directly either failed or succeeded, in which case we loop
5631 * through the descendants with best-effort propagation.
5633 for_each_memcg_cache(c
, s
)
5634 attribute
->store(c
, buf
, len
);
5635 mutex_unlock(&slab_mutex
);
5641 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5645 char *buffer
= NULL
;
5646 struct kmem_cache
*root_cache
;
5648 if (is_root_cache(s
))
5651 root_cache
= s
->memcg_params
.root_cache
;
5654 * This mean this cache had no attribute written. Therefore, no point
5655 * in copying default values around
5657 if (!root_cache
->max_attr_size
)
5660 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5663 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5666 if (!attr
|| !attr
->store
|| !attr
->show
)
5670 * It is really bad that we have to allocate here, so we will
5671 * do it only as a fallback. If we actually allocate, though,
5672 * we can just use the allocated buffer until the end.
5674 * Most of the slub attributes will tend to be very small in
5675 * size, but sysfs allows buffers up to a page, so they can
5676 * theoretically happen.
5680 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
) &&
5681 !IS_ENABLED(CONFIG_SLUB_STATS
))
5684 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5685 if (WARN_ON(!buffer
))
5690 len
= attr
->show(root_cache
, buf
);
5692 attr
->store(s
, buf
, len
);
5696 free_page((unsigned long)buffer
);
5697 #endif /* CONFIG_MEMCG */
5700 static void kmem_cache_release(struct kobject
*k
)
5702 slab_kmem_cache_release(to_slab(k
));
5705 static const struct sysfs_ops slab_sysfs_ops
= {
5706 .show
= slab_attr_show
,
5707 .store
= slab_attr_store
,
5710 static struct kobj_type slab_ktype
= {
5711 .sysfs_ops
= &slab_sysfs_ops
,
5712 .release
= kmem_cache_release
,
5715 static struct kset
*slab_kset
;
5717 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5720 if (!is_root_cache(s
))
5721 return s
->memcg_params
.root_cache
->memcg_kset
;
5726 #define ID_STR_LENGTH 64
5728 /* Create a unique string id for a slab cache:
5730 * Format :[flags-]size
5732 static char *create_unique_id(struct kmem_cache
*s
)
5734 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5741 * First flags affecting slabcache operations. We will only
5742 * get here for aliasable slabs so we do not need to support
5743 * too many flags. The flags here must cover all flags that
5744 * are matched during merging to guarantee that the id is
5747 if (s
->flags
& SLAB_CACHE_DMA
)
5749 if (s
->flags
& SLAB_CACHE_DMA32
)
5751 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5753 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5755 if (s
->flags
& SLAB_ACCOUNT
)
5759 p
+= sprintf(p
, "%07u", s
->size
);
5761 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5765 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5767 struct kmem_cache
*s
=
5768 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5770 if (!s
->kobj
.state_in_sysfs
)
5772 * For a memcg cache, this may be called during
5773 * deactivation and again on shutdown. Remove only once.
5774 * A cache is never shut down before deactivation is
5775 * complete, so no need to worry about synchronization.
5780 kset_unregister(s
->memcg_kset
);
5783 kobject_put(&s
->kobj
);
5786 static int sysfs_slab_add(struct kmem_cache
*s
)
5790 struct kset
*kset
= cache_kset(s
);
5791 int unmergeable
= slab_unmergeable(s
);
5793 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5796 kobject_init(&s
->kobj
, &slab_ktype
);
5800 if (!unmergeable
&& disable_higher_order_debug
&&
5801 (slub_debug
& DEBUG_METADATA_FLAGS
))
5806 * Slabcache can never be merged so we can use the name proper.
5807 * This is typically the case for debug situations. In that
5808 * case we can catch duplicate names easily.
5810 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5814 * Create a unique name for the slab as a target
5817 name
= create_unique_id(s
);
5820 s
->kobj
.kset
= kset
;
5821 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5823 kobject_put(&s
->kobj
);
5827 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5832 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5833 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5834 if (!s
->memcg_kset
) {
5842 /* Setup first alias */
5843 sysfs_slab_alias(s
, s
->name
);
5850 kobject_del(&s
->kobj
);
5854 static void sysfs_slab_remove(struct kmem_cache
*s
)
5856 if (slab_state
< FULL
)
5858 * Sysfs has not been setup yet so no need to remove the
5863 kobject_get(&s
->kobj
);
5864 schedule_work(&s
->kobj_remove_work
);
5867 void sysfs_slab_unlink(struct kmem_cache
*s
)
5869 if (slab_state
>= FULL
)
5870 kobject_del(&s
->kobj
);
5873 void sysfs_slab_release(struct kmem_cache
*s
)
5875 if (slab_state
>= FULL
)
5876 kobject_put(&s
->kobj
);
5880 * Need to buffer aliases during bootup until sysfs becomes
5881 * available lest we lose that information.
5883 struct saved_alias
{
5884 struct kmem_cache
*s
;
5886 struct saved_alias
*next
;
5889 static struct saved_alias
*alias_list
;
5891 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5893 struct saved_alias
*al
;
5895 if (slab_state
== FULL
) {
5897 * If we have a leftover link then remove it.
5899 sysfs_remove_link(&slab_kset
->kobj
, name
);
5900 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5903 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5909 al
->next
= alias_list
;
5914 static int __init
slab_sysfs_init(void)
5916 struct kmem_cache
*s
;
5919 mutex_lock(&slab_mutex
);
5921 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
5923 mutex_unlock(&slab_mutex
);
5924 pr_err("Cannot register slab subsystem.\n");
5930 list_for_each_entry(s
, &slab_caches
, list
) {
5931 err
= sysfs_slab_add(s
);
5933 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5937 while (alias_list
) {
5938 struct saved_alias
*al
= alias_list
;
5940 alias_list
= alias_list
->next
;
5941 err
= sysfs_slab_alias(al
->s
, al
->name
);
5943 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5948 mutex_unlock(&slab_mutex
);
5953 __initcall(slab_sysfs_init
);
5954 #endif /* CONFIG_SYSFS */
5957 * The /proc/slabinfo ABI
5959 #ifdef CONFIG_SLUB_DEBUG
5960 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5962 unsigned long nr_slabs
= 0;
5963 unsigned long nr_objs
= 0;
5964 unsigned long nr_free
= 0;
5966 struct kmem_cache_node
*n
;
5968 for_each_kmem_cache_node(s
, node
, n
) {
5969 nr_slabs
+= node_nr_slabs(n
);
5970 nr_objs
+= node_nr_objs(n
);
5971 nr_free
+= count_partial(n
, count_free
);
5974 sinfo
->active_objs
= nr_objs
- nr_free
;
5975 sinfo
->num_objs
= nr_objs
;
5976 sinfo
->active_slabs
= nr_slabs
;
5977 sinfo
->num_slabs
= nr_slabs
;
5978 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5979 sinfo
->cache_order
= oo_order(s
->oo
);
5982 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5986 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5987 size_t count
, loff_t
*ppos
)
5991 #endif /* CONFIG_SLUB_DEBUG */