2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects the second
56 * double word in the page struct. Meaning
57 * A. page->freelist -> List of object free in a page
58 * B. page->counters -> Counters of objects
59 * C. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list. The processor that froze the slab is the one who can
63 * perform list operations on the page. Other processors may put objects
64 * onto the freelist but the processor that froze the slab is the only
65 * one that can retrieve the objects from the page's freelist.
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 * Overloading of page flags that are otherwise used for LRU management.
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
119 static inline int kmem_cache_debug(struct kmem_cache
*s
)
121 #ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
128 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
130 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
131 p
+= s
->red_left_pad
;
136 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
138 #ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s
);
146 * Issues still to be resolved:
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 /* Enable to log cmpxchg failures */
157 #undef SLUB_DEBUG_CMPXCHG
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
163 #define MIN_PARTIAL 5
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
170 #define MAX_PARTIAL 10
172 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
179 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
188 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191 #define OO_MASK ((1 << OO_SHIFT) - 1)
192 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
194 /* Internal SLUB flags */
195 #define __OBJECT_POISON 0x80000000UL /* Poison object */
196 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
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
251 return (void *)((unsigned long)ptr
^ s
->random
^ ptr_addr
);
257 /* Returns the freelist pointer recorded at location ptr_addr. */
258 static inline void *freelist_dereference(const struct kmem_cache
*s
,
261 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
262 (unsigned long)ptr_addr
);
265 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
267 return freelist_dereference(s
, object
+ s
->offset
);
270 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
273 prefetch(freelist_dereference(s
, object
+ s
->offset
));
276 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
278 unsigned long freepointer_addr
;
281 if (!debug_pagealloc_enabled())
282 return get_freepointer(s
, object
);
284 freepointer_addr
= (unsigned long)object
+ s
->offset
;
285 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
286 return freelist_ptr(s
, p
, freepointer_addr
);
289 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
291 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
293 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
296 /* Loop over all objects in a slab */
297 #define for_each_object(__p, __s, __addr, __objects) \
298 for (__p = fixup_red_left(__s, __addr); \
299 __p < (__addr) + (__objects) * (__s)->size; \
302 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
303 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
304 __idx <= __objects; \
305 __p += (__s)->size, __idx++)
307 /* Determine object index from a given position */
308 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
310 return (p
- addr
) / s
->size
;
313 static inline int order_objects(int order
, unsigned long size
, int reserved
)
315 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
318 static inline struct kmem_cache_order_objects
oo_make(int order
,
319 unsigned long size
, int reserved
)
321 struct kmem_cache_order_objects x
= {
322 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
328 static inline int oo_order(struct kmem_cache_order_objects x
)
330 return x
.x
>> OO_SHIFT
;
333 static inline int oo_objects(struct kmem_cache_order_objects x
)
335 return x
.x
& OO_MASK
;
339 * Per slab locking using the pagelock
341 static __always_inline
void slab_lock(struct page
*page
)
343 VM_BUG_ON_PAGE(PageTail(page
), page
);
344 bit_spin_lock(PG_locked
, &page
->flags
);
347 static __always_inline
void slab_unlock(struct page
*page
)
349 VM_BUG_ON_PAGE(PageTail(page
), page
);
350 __bit_spin_unlock(PG_locked
, &page
->flags
);
353 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
356 tmp
.counters
= counters_new
;
358 * page->counters can cover frozen/inuse/objects as well
359 * as page->_refcount. If we assign to ->counters directly
360 * we run the risk of losing updates to page->_refcount, so
361 * be careful and only assign to the fields we need.
363 page
->frozen
= tmp
.frozen
;
364 page
->inuse
= tmp
.inuse
;
365 page
->objects
= tmp
.objects
;
368 /* Interrupts must be disabled (for the fallback code to work right) */
369 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
370 void *freelist_old
, unsigned long counters_old
,
371 void *freelist_new
, unsigned long counters_new
,
374 VM_BUG_ON(!irqs_disabled());
375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s
->flags
& __CMPXCHG_DOUBLE
) {
378 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
379 freelist_old
, counters_old
,
380 freelist_new
, counters_new
))
386 if (page
->freelist
== freelist_old
&&
387 page
->counters
== counters_old
) {
388 page
->freelist
= freelist_new
;
389 set_page_slub_counters(page
, counters_new
);
397 stat(s
, CMPXCHG_DOUBLE_FAIL
);
399 #ifdef SLUB_DEBUG_CMPXCHG
400 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
406 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
407 void *freelist_old
, unsigned long counters_old
,
408 void *freelist_new
, unsigned long counters_new
,
411 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
412 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
413 if (s
->flags
& __CMPXCHG_DOUBLE
) {
414 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
415 freelist_old
, counters_old
,
416 freelist_new
, counters_new
))
423 local_irq_save(flags
);
425 if (page
->freelist
== freelist_old
&&
426 page
->counters
== counters_old
) {
427 page
->freelist
= freelist_new
;
428 set_page_slub_counters(page
, counters_new
);
430 local_irq_restore(flags
);
434 local_irq_restore(flags
);
438 stat(s
, CMPXCHG_DOUBLE_FAIL
);
440 #ifdef SLUB_DEBUG_CMPXCHG
441 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
447 #ifdef CONFIG_SLUB_DEBUG
449 * Determine a map of object in use on a page.
451 * Node listlock must be held to guarantee that the page does
452 * not vanish from under us.
454 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
457 void *addr
= page_address(page
);
459 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
460 set_bit(slab_index(p
, s
, addr
), map
);
463 static inline int size_from_object(struct kmem_cache
*s
)
465 if (s
->flags
& SLAB_RED_ZONE
)
466 return s
->size
- s
->red_left_pad
;
471 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
473 if (s
->flags
& SLAB_RED_ZONE
)
474 p
-= s
->red_left_pad
;
482 #if defined(CONFIG_SLUB_DEBUG_ON)
483 static int slub_debug
= DEBUG_DEFAULT_FLAGS
;
485 static int slub_debug
;
488 static char *slub_debug_slabs
;
489 static int disable_higher_order_debug
;
492 * slub is about to manipulate internal object metadata. This memory lies
493 * outside the range of the allocated object, so accessing it would normally
494 * be reported by kasan as a bounds error. metadata_access_enable() is used
495 * to tell kasan that these accesses are OK.
497 static inline void metadata_access_enable(void)
499 kasan_disable_current();
502 static inline void metadata_access_disable(void)
504 kasan_enable_current();
511 /* Verify that a pointer has an address that is valid within a slab page */
512 static inline int check_valid_pointer(struct kmem_cache
*s
,
513 struct page
*page
, void *object
)
520 base
= page_address(page
);
521 object
= restore_red_left(s
, object
);
522 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
523 (object
- base
) % s
->size
) {
530 static void print_section(char *level
, char *text
, u8
*addr
,
533 metadata_access_enable();
534 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
536 metadata_access_disable();
539 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
540 enum track_item alloc
)
545 p
= object
+ s
->offset
+ sizeof(void *);
547 p
= object
+ s
->inuse
;
552 static void set_track(struct kmem_cache
*s
, void *object
,
553 enum track_item alloc
, unsigned long addr
)
555 struct track
*p
= get_track(s
, object
, alloc
);
558 #ifdef CONFIG_STACKTRACE
559 struct stack_trace trace
;
562 trace
.nr_entries
= 0;
563 trace
.max_entries
= TRACK_ADDRS_COUNT
;
564 trace
.entries
= p
->addrs
;
566 metadata_access_enable();
567 save_stack_trace(&trace
);
568 metadata_access_disable();
570 /* See rant in lockdep.c */
571 if (trace
.nr_entries
!= 0 &&
572 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
575 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
579 p
->cpu
= smp_processor_id();
580 p
->pid
= current
->pid
;
583 memset(p
, 0, sizeof(struct track
));
586 static void init_tracking(struct kmem_cache
*s
, void *object
)
588 if (!(s
->flags
& SLAB_STORE_USER
))
591 set_track(s
, object
, TRACK_FREE
, 0UL);
592 set_track(s
, object
, TRACK_ALLOC
, 0UL);
595 static void print_track(const char *s
, struct track
*t
)
600 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
601 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
602 #ifdef CONFIG_STACKTRACE
605 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
607 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
614 static void print_tracking(struct kmem_cache
*s
, void *object
)
616 if (!(s
->flags
& SLAB_STORE_USER
))
619 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
620 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
623 static void print_page_info(struct page
*page
)
625 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
626 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
630 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
632 struct va_format vaf
;
638 pr_err("=============================================================================\n");
639 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
640 pr_err("-----------------------------------------------------------------------------\n\n");
642 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
646 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
648 struct va_format vaf
;
654 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
658 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
660 unsigned int off
; /* Offset of last byte */
661 u8
*addr
= page_address(page
);
663 print_tracking(s
, p
);
665 print_page_info(page
);
667 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
668 p
, p
- addr
, get_freepointer(s
, p
));
670 if (s
->flags
& SLAB_RED_ZONE
)
671 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
673 else if (p
> addr
+ 16)
674 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
676 print_section(KERN_ERR
, "Object ", p
,
677 min_t(unsigned long, s
->object_size
, PAGE_SIZE
));
678 if (s
->flags
& SLAB_RED_ZONE
)
679 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
680 s
->inuse
- s
->object_size
);
683 off
= s
->offset
+ sizeof(void *);
687 if (s
->flags
& SLAB_STORE_USER
)
688 off
+= 2 * sizeof(struct track
);
690 off
+= kasan_metadata_size(s
);
692 if (off
!= size_from_object(s
))
693 /* Beginning of the filler is the free pointer */
694 print_section(KERN_ERR
, "Padding ", p
+ off
,
695 size_from_object(s
) - off
);
700 void object_err(struct kmem_cache
*s
, struct page
*page
,
701 u8
*object
, char *reason
)
703 slab_bug(s
, "%s", reason
);
704 print_trailer(s
, page
, object
);
707 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
708 const char *fmt
, ...)
714 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
716 slab_bug(s
, "%s", buf
);
717 print_page_info(page
);
721 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
725 if (s
->flags
& SLAB_RED_ZONE
)
726 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
728 if (s
->flags
& __OBJECT_POISON
) {
729 memset(p
, POISON_FREE
, s
->object_size
- 1);
730 p
[s
->object_size
- 1] = POISON_END
;
733 if (s
->flags
& SLAB_RED_ZONE
)
734 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
737 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
738 void *from
, void *to
)
740 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
741 memset(from
, data
, to
- from
);
744 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
745 u8
*object
, char *what
,
746 u8
*start
, unsigned int value
, unsigned int bytes
)
751 metadata_access_enable();
752 fault
= memchr_inv(start
, value
, bytes
);
753 metadata_access_disable();
758 while (end
> fault
&& end
[-1] == value
)
761 slab_bug(s
, "%s overwritten", what
);
762 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
763 fault
, end
- 1, fault
[0], value
);
764 print_trailer(s
, page
, object
);
766 restore_bytes(s
, what
, value
, fault
, end
);
774 * Bytes of the object to be managed.
775 * If the freepointer may overlay the object then the free
776 * pointer is the first word of the object.
778 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
781 * object + s->object_size
782 * Padding to reach word boundary. This is also used for Redzoning.
783 * Padding is extended by another word if Redzoning is enabled and
784 * object_size == inuse.
786 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
787 * 0xcc (RED_ACTIVE) for objects in use.
790 * Meta data starts here.
792 * A. Free pointer (if we cannot overwrite object on free)
793 * B. Tracking data for SLAB_STORE_USER
794 * C. Padding to reach required alignment boundary or at mininum
795 * one word if debugging is on to be able to detect writes
796 * before the word boundary.
798 * Padding is done using 0x5a (POISON_INUSE)
801 * Nothing is used beyond s->size.
803 * If slabcaches are merged then the object_size and inuse boundaries are mostly
804 * ignored. And therefore no slab options that rely on these boundaries
805 * may be used with merged slabcaches.
808 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
810 unsigned long off
= s
->inuse
; /* The end of info */
813 /* Freepointer is placed after the object. */
814 off
+= sizeof(void *);
816 if (s
->flags
& SLAB_STORE_USER
)
817 /* We also have user information there */
818 off
+= 2 * sizeof(struct track
);
820 off
+= kasan_metadata_size(s
);
822 if (size_from_object(s
) == off
)
825 return check_bytes_and_report(s
, page
, p
, "Object padding",
826 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
829 /* Check the pad bytes at the end of a slab page */
830 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
838 if (!(s
->flags
& SLAB_POISON
))
841 start
= page_address(page
);
842 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
843 end
= start
+ length
;
844 remainder
= length
% s
->size
;
848 metadata_access_enable();
849 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
850 metadata_access_disable();
853 while (end
> fault
&& end
[-1] == POISON_INUSE
)
856 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
857 print_section(KERN_ERR
, "Padding ", end
- remainder
, remainder
);
859 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
863 static int check_object(struct kmem_cache
*s
, struct page
*page
,
864 void *object
, u8 val
)
867 u8
*endobject
= object
+ s
->object_size
;
869 if (s
->flags
& SLAB_RED_ZONE
) {
870 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
871 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
874 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
875 endobject
, val
, s
->inuse
- s
->object_size
))
878 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
879 check_bytes_and_report(s
, page
, p
, "Alignment padding",
880 endobject
, POISON_INUSE
,
881 s
->inuse
- s
->object_size
);
885 if (s
->flags
& SLAB_POISON
) {
886 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
887 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
888 POISON_FREE
, s
->object_size
- 1) ||
889 !check_bytes_and_report(s
, page
, p
, "Poison",
890 p
+ s
->object_size
- 1, POISON_END
, 1)))
893 * check_pad_bytes cleans up on its own.
895 check_pad_bytes(s
, page
, p
);
898 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
900 * Object and freepointer overlap. Cannot check
901 * freepointer while object is allocated.
905 /* Check free pointer validity */
906 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
907 object_err(s
, page
, p
, "Freepointer corrupt");
909 * No choice but to zap it and thus lose the remainder
910 * of the free objects in this slab. May cause
911 * another error because the object count is now wrong.
913 set_freepointer(s
, p
, NULL
);
919 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
923 VM_BUG_ON(!irqs_disabled());
925 if (!PageSlab(page
)) {
926 slab_err(s
, page
, "Not a valid slab page");
930 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
931 if (page
->objects
> maxobj
) {
932 slab_err(s
, page
, "objects %u > max %u",
933 page
->objects
, maxobj
);
936 if (page
->inuse
> page
->objects
) {
937 slab_err(s
, page
, "inuse %u > max %u",
938 page
->inuse
, page
->objects
);
941 /* Slab_pad_check fixes things up after itself */
942 slab_pad_check(s
, page
);
947 * Determine if a certain object on a page is on the freelist. Must hold the
948 * slab lock to guarantee that the chains are in a consistent state.
950 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
958 while (fp
&& nr
<= page
->objects
) {
961 if (!check_valid_pointer(s
, page
, fp
)) {
963 object_err(s
, page
, object
,
964 "Freechain corrupt");
965 set_freepointer(s
, object
, NULL
);
967 slab_err(s
, page
, "Freepointer corrupt");
968 page
->freelist
= NULL
;
969 page
->inuse
= page
->objects
;
970 slab_fix(s
, "Freelist cleared");
976 fp
= get_freepointer(s
, object
);
980 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
981 if (max_objects
> MAX_OBJS_PER_PAGE
)
982 max_objects
= MAX_OBJS_PER_PAGE
;
984 if (page
->objects
!= max_objects
) {
985 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
986 page
->objects
, max_objects
);
987 page
->objects
= max_objects
;
988 slab_fix(s
, "Number of objects adjusted.");
990 if (page
->inuse
!= page
->objects
- nr
) {
991 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
992 page
->inuse
, page
->objects
- nr
);
993 page
->inuse
= page
->objects
- nr
;
994 slab_fix(s
, "Object count adjusted.");
996 return search
== NULL
;
999 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1002 if (s
->flags
& SLAB_TRACE
) {
1003 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1005 alloc
? "alloc" : "free",
1006 object
, page
->inuse
,
1010 print_section(KERN_INFO
, "Object ", (void *)object
,
1018 * Tracking of fully allocated slabs for debugging purposes.
1020 static void add_full(struct kmem_cache
*s
,
1021 struct kmem_cache_node
*n
, struct page
*page
)
1023 if (!(s
->flags
& SLAB_STORE_USER
))
1026 lockdep_assert_held(&n
->list_lock
);
1027 list_add(&page
->lru
, &n
->full
);
1030 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1032 if (!(s
->flags
& SLAB_STORE_USER
))
1035 lockdep_assert_held(&n
->list_lock
);
1036 list_del(&page
->lru
);
1039 /* Tracking of the number of slabs for debugging purposes */
1040 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1042 struct kmem_cache_node
*n
= get_node(s
, node
);
1044 return atomic_long_read(&n
->nr_slabs
);
1047 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1049 return atomic_long_read(&n
->nr_slabs
);
1052 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1054 struct kmem_cache_node
*n
= get_node(s
, node
);
1057 * May be called early in order to allocate a slab for the
1058 * kmem_cache_node structure. Solve the chicken-egg
1059 * dilemma by deferring the increment of the count during
1060 * bootstrap (see early_kmem_cache_node_alloc).
1063 atomic_long_inc(&n
->nr_slabs
);
1064 atomic_long_add(objects
, &n
->total_objects
);
1067 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1069 struct kmem_cache_node
*n
= get_node(s
, node
);
1071 atomic_long_dec(&n
->nr_slabs
);
1072 atomic_long_sub(objects
, &n
->total_objects
);
1075 /* Object debug checks for alloc/free paths */
1076 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1079 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1082 init_object(s
, object
, SLUB_RED_INACTIVE
);
1083 init_tracking(s
, object
);
1086 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1088 void *object
, unsigned long addr
)
1090 if (!check_slab(s
, page
))
1093 if (!check_valid_pointer(s
, page
, object
)) {
1094 object_err(s
, page
, object
, "Freelist Pointer check fails");
1098 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1104 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1106 void *object
, unsigned long addr
)
1108 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1109 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1113 /* Success perform special debug activities for allocs */
1114 if (s
->flags
& SLAB_STORE_USER
)
1115 set_track(s
, object
, TRACK_ALLOC
, addr
);
1116 trace(s
, page
, object
, 1);
1117 init_object(s
, object
, SLUB_RED_ACTIVE
);
1121 if (PageSlab(page
)) {
1123 * If this is a slab page then lets do the best we can
1124 * to avoid issues in the future. Marking all objects
1125 * as used avoids touching the remaining objects.
1127 slab_fix(s
, "Marking all objects used");
1128 page
->inuse
= page
->objects
;
1129 page
->freelist
= NULL
;
1134 static inline int free_consistency_checks(struct kmem_cache
*s
,
1135 struct page
*page
, void *object
, unsigned long addr
)
1137 if (!check_valid_pointer(s
, page
, object
)) {
1138 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1142 if (on_freelist(s
, page
, object
)) {
1143 object_err(s
, page
, object
, "Object already free");
1147 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1150 if (unlikely(s
!= page
->slab_cache
)) {
1151 if (!PageSlab(page
)) {
1152 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1154 } else if (!page
->slab_cache
) {
1155 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1159 object_err(s
, page
, object
,
1160 "page slab pointer corrupt.");
1166 /* Supports checking bulk free of a constructed freelist */
1167 static noinline
int free_debug_processing(
1168 struct kmem_cache
*s
, struct page
*page
,
1169 void *head
, void *tail
, int bulk_cnt
,
1172 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1173 void *object
= head
;
1175 unsigned long uninitialized_var(flags
);
1178 spin_lock_irqsave(&n
->list_lock
, flags
);
1181 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1182 if (!check_slab(s
, page
))
1189 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1190 if (!free_consistency_checks(s
, page
, object
, addr
))
1194 if (s
->flags
& SLAB_STORE_USER
)
1195 set_track(s
, object
, TRACK_FREE
, addr
);
1196 trace(s
, page
, object
, 0);
1197 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1198 init_object(s
, object
, SLUB_RED_INACTIVE
);
1200 /* Reached end of constructed freelist yet? */
1201 if (object
!= tail
) {
1202 object
= get_freepointer(s
, object
);
1208 if (cnt
!= bulk_cnt
)
1209 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1213 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1215 slab_fix(s
, "Object at 0x%p not freed", object
);
1219 static int __init
setup_slub_debug(char *str
)
1221 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1222 if (*str
++ != '=' || !*str
)
1224 * No options specified. Switch on full debugging.
1230 * No options but restriction on slabs. This means full
1231 * debugging for slabs matching a pattern.
1238 * Switch off all debugging measures.
1243 * Determine which debug features should be switched on
1245 for (; *str
&& *str
!= ','; str
++) {
1246 switch (tolower(*str
)) {
1248 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1251 slub_debug
|= SLAB_RED_ZONE
;
1254 slub_debug
|= SLAB_POISON
;
1257 slub_debug
|= SLAB_STORE_USER
;
1260 slub_debug
|= SLAB_TRACE
;
1263 slub_debug
|= SLAB_FAILSLAB
;
1267 * Avoid enabling debugging on caches if its minimum
1268 * order would increase as a result.
1270 disable_higher_order_debug
= 1;
1273 pr_err("slub_debug option '%c' unknown. skipped\n",
1280 slub_debug_slabs
= str
+ 1;
1285 __setup("slub_debug", setup_slub_debug
);
1287 unsigned long kmem_cache_flags(unsigned long object_size
,
1288 unsigned long flags
, const char *name
,
1289 void (*ctor
)(void *))
1292 * Enable debugging if selected on the kernel commandline.
1294 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1295 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1296 flags
|= slub_debug
;
1300 #else /* !CONFIG_SLUB_DEBUG */
1301 static inline void setup_object_debug(struct kmem_cache
*s
,
1302 struct page
*page
, void *object
) {}
1304 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1305 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1307 static inline int free_debug_processing(
1308 struct kmem_cache
*s
, struct page
*page
,
1309 void *head
, void *tail
, int bulk_cnt
,
1310 unsigned long addr
) { return 0; }
1312 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1314 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1315 void *object
, u8 val
) { return 1; }
1316 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1317 struct page
*page
) {}
1318 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1319 struct page
*page
) {}
1320 unsigned long kmem_cache_flags(unsigned long object_size
,
1321 unsigned long flags
, const char *name
,
1322 void (*ctor
)(void *))
1326 #define slub_debug 0
1328 #define disable_higher_order_debug 0
1330 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1332 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1334 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1336 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1339 #endif /* CONFIG_SLUB_DEBUG */
1342 * Hooks for other subsystems that check memory allocations. In a typical
1343 * production configuration these hooks all should produce no code at all.
1345 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1347 kmemleak_alloc(ptr
, size
, 1, flags
);
1348 kasan_kmalloc_large(ptr
, size
, flags
);
1351 static inline void kfree_hook(const void *x
)
1354 kasan_kfree_large(x
);
1357 static inline void *slab_free_hook(struct kmem_cache
*s
, void *x
)
1361 kmemleak_free_recursive(x
, s
->flags
);
1364 * Trouble is that we may no longer disable interrupts in the fast path
1365 * So in order to make the debug calls that expect irqs to be
1366 * disabled we need to disable interrupts temporarily.
1368 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1370 unsigned long flags
;
1372 local_irq_save(flags
);
1373 kmemcheck_slab_free(s
, x
, s
->object_size
);
1374 debug_check_no_locks_freed(x
, s
->object_size
);
1375 local_irq_restore(flags
);
1378 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1379 debug_check_no_obj_freed(x
, s
->object_size
);
1381 freeptr
= get_freepointer(s
, x
);
1383 * kasan_slab_free() may put x into memory quarantine, delaying its
1384 * reuse. In this case the object's freelist pointer is changed.
1386 kasan_slab_free(s
, x
);
1390 static inline void slab_free_freelist_hook(struct kmem_cache
*s
,
1391 void *head
, void *tail
)
1394 * Compiler cannot detect this function can be removed if slab_free_hook()
1395 * evaluates to nothing. Thus, catch all relevant config debug options here.
1397 #if defined(CONFIG_KMEMCHECK) || \
1398 defined(CONFIG_LOCKDEP) || \
1399 defined(CONFIG_DEBUG_KMEMLEAK) || \
1400 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1401 defined(CONFIG_KASAN)
1403 void *object
= head
;
1404 void *tail_obj
= tail
? : head
;
1408 freeptr
= slab_free_hook(s
, object
);
1409 } while ((object
!= tail_obj
) && (object
= freeptr
));
1413 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1416 setup_object_debug(s
, page
, object
);
1417 kasan_init_slab_obj(s
, object
);
1418 if (unlikely(s
->ctor
)) {
1419 kasan_unpoison_object_data(s
, object
);
1421 kasan_poison_object_data(s
, object
);
1426 * Slab allocation and freeing
1428 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1429 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1432 int order
= oo_order(oo
);
1434 flags
|= __GFP_NOTRACK
;
1436 if (node
== NUMA_NO_NODE
)
1437 page
= alloc_pages(flags
, order
);
1439 page
= __alloc_pages_node(node
, flags
, order
);
1441 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1442 __free_pages(page
, order
);
1449 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1450 /* Pre-initialize the random sequence cache */
1451 static int init_cache_random_seq(struct kmem_cache
*s
)
1454 unsigned long i
, count
= oo_objects(s
->oo
);
1456 /* Bailout if already initialised */
1460 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1462 pr_err("SLUB: Unable to initialize free list for %s\n",
1467 /* Transform to an offset on the set of pages */
1468 if (s
->random_seq
) {
1469 for (i
= 0; i
< count
; i
++)
1470 s
->random_seq
[i
] *= s
->size
;
1475 /* Initialize each random sequence freelist per cache */
1476 static void __init
init_freelist_randomization(void)
1478 struct kmem_cache
*s
;
1480 mutex_lock(&slab_mutex
);
1482 list_for_each_entry(s
, &slab_caches
, list
)
1483 init_cache_random_seq(s
);
1485 mutex_unlock(&slab_mutex
);
1488 /* Get the next entry on the pre-computed freelist randomized */
1489 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1490 unsigned long *pos
, void *start
,
1491 unsigned long page_limit
,
1492 unsigned long freelist_count
)
1497 * If the target page allocation failed, the number of objects on the
1498 * page might be smaller than the usual size defined by the cache.
1501 idx
= s
->random_seq
[*pos
];
1503 if (*pos
>= freelist_count
)
1505 } while (unlikely(idx
>= page_limit
));
1507 return (char *)start
+ idx
;
1510 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1511 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1516 unsigned long idx
, pos
, page_limit
, freelist_count
;
1518 if (page
->objects
< 2 || !s
->random_seq
)
1521 freelist_count
= oo_objects(s
->oo
);
1522 pos
= get_random_int() % freelist_count
;
1524 page_limit
= page
->objects
* s
->size
;
1525 start
= fixup_red_left(s
, page_address(page
));
1527 /* First entry is used as the base of the freelist */
1528 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1530 page
->freelist
= cur
;
1532 for (idx
= 1; idx
< page
->objects
; idx
++) {
1533 setup_object(s
, page
, cur
);
1534 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1536 set_freepointer(s
, cur
, next
);
1539 setup_object(s
, page
, cur
);
1540 set_freepointer(s
, cur
, NULL
);
1545 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1549 static inline void init_freelist_randomization(void) { }
1550 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1554 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1556 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1559 struct kmem_cache_order_objects oo
= s
->oo
;
1565 flags
&= gfp_allowed_mask
;
1567 if (gfpflags_allow_blocking(flags
))
1570 flags
|= s
->allocflags
;
1573 * Let the initial higher-order allocation fail under memory pressure
1574 * so we fall-back to the minimum order allocation.
1576 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1577 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1578 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1580 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1581 if (unlikely(!page
)) {
1585 * Allocation may have failed due to fragmentation.
1586 * Try a lower order alloc if possible
1588 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1589 if (unlikely(!page
))
1591 stat(s
, ORDER_FALLBACK
);
1594 if (kmemcheck_enabled
&&
1595 !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1596 int pages
= 1 << oo_order(oo
);
1598 kmemcheck_alloc_shadow(page
, oo_order(oo
), alloc_gfp
, node
);
1601 * Objects from caches that have a constructor don't get
1602 * cleared when they're allocated, so we need to do it here.
1605 kmemcheck_mark_uninitialized_pages(page
, pages
);
1607 kmemcheck_mark_unallocated_pages(page
, pages
);
1610 page
->objects
= oo_objects(oo
);
1612 order
= compound_order(page
);
1613 page
->slab_cache
= s
;
1614 __SetPageSlab(page
);
1615 if (page_is_pfmemalloc(page
))
1616 SetPageSlabPfmemalloc(page
);
1618 start
= page_address(page
);
1620 if (unlikely(s
->flags
& SLAB_POISON
))
1621 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1623 kasan_poison_slab(page
);
1625 shuffle
= shuffle_freelist(s
, page
);
1628 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1629 setup_object(s
, page
, p
);
1630 if (likely(idx
< page
->objects
))
1631 set_freepointer(s
, p
, p
+ s
->size
);
1633 set_freepointer(s
, p
, NULL
);
1635 page
->freelist
= fixup_red_left(s
, start
);
1638 page
->inuse
= page
->objects
;
1642 if (gfpflags_allow_blocking(flags
))
1643 local_irq_disable();
1647 mod_lruvec_page_state(page
,
1648 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1649 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1652 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1657 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1659 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1660 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1661 flags
&= ~GFP_SLAB_BUG_MASK
;
1662 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1663 invalid_mask
, &invalid_mask
, flags
, &flags
);
1667 return allocate_slab(s
,
1668 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1671 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1673 int order
= compound_order(page
);
1674 int pages
= 1 << order
;
1676 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1679 slab_pad_check(s
, page
);
1680 for_each_object(p
, s
, page_address(page
),
1682 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1685 kmemcheck_free_shadow(page
, compound_order(page
));
1687 mod_lruvec_page_state(page
,
1688 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1689 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1692 __ClearPageSlabPfmemalloc(page
);
1693 __ClearPageSlab(page
);
1695 page_mapcount_reset(page
);
1696 if (current
->reclaim_state
)
1697 current
->reclaim_state
->reclaimed_slab
+= pages
;
1698 memcg_uncharge_slab(page
, order
, s
);
1699 __free_pages(page
, order
);
1702 #define need_reserve_slab_rcu \
1703 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1705 static void rcu_free_slab(struct rcu_head
*h
)
1709 if (need_reserve_slab_rcu
)
1710 page
= virt_to_head_page(h
);
1712 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1714 __free_slab(page
->slab_cache
, page
);
1717 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1719 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1720 struct rcu_head
*head
;
1722 if (need_reserve_slab_rcu
) {
1723 int order
= compound_order(page
);
1724 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1726 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1727 head
= page_address(page
) + offset
;
1729 head
= &page
->rcu_head
;
1732 call_rcu(head
, rcu_free_slab
);
1734 __free_slab(s
, page
);
1737 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1739 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1744 * Management of partially allocated slabs.
1747 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1750 if (tail
== DEACTIVATE_TO_TAIL
)
1751 list_add_tail(&page
->lru
, &n
->partial
);
1753 list_add(&page
->lru
, &n
->partial
);
1756 static inline void add_partial(struct kmem_cache_node
*n
,
1757 struct page
*page
, int tail
)
1759 lockdep_assert_held(&n
->list_lock
);
1760 __add_partial(n
, page
, tail
);
1763 static inline void remove_partial(struct kmem_cache_node
*n
,
1766 lockdep_assert_held(&n
->list_lock
);
1767 list_del(&page
->lru
);
1772 * Remove slab from the partial list, freeze it and
1773 * return the pointer to the freelist.
1775 * Returns a list of objects or NULL if it fails.
1777 static inline void *acquire_slab(struct kmem_cache
*s
,
1778 struct kmem_cache_node
*n
, struct page
*page
,
1779 int mode
, int *objects
)
1782 unsigned long counters
;
1785 lockdep_assert_held(&n
->list_lock
);
1788 * Zap the freelist and set the frozen bit.
1789 * The old freelist is the list of objects for the
1790 * per cpu allocation list.
1792 freelist
= page
->freelist
;
1793 counters
= page
->counters
;
1794 new.counters
= counters
;
1795 *objects
= new.objects
- new.inuse
;
1797 new.inuse
= page
->objects
;
1798 new.freelist
= NULL
;
1800 new.freelist
= freelist
;
1803 VM_BUG_ON(new.frozen
);
1806 if (!__cmpxchg_double_slab(s
, page
,
1808 new.freelist
, new.counters
,
1812 remove_partial(n
, page
);
1817 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1818 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1821 * Try to allocate a partial slab from a specific node.
1823 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1824 struct kmem_cache_cpu
*c
, gfp_t flags
)
1826 struct page
*page
, *page2
;
1827 void *object
= NULL
;
1832 * Racy check. If we mistakenly see no partial slabs then we
1833 * just allocate an empty slab. If we mistakenly try to get a
1834 * partial slab and there is none available then get_partials()
1837 if (!n
|| !n
->nr_partial
)
1840 spin_lock(&n
->list_lock
);
1841 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1844 if (!pfmemalloc_match(page
, flags
))
1847 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1851 available
+= objects
;
1854 stat(s
, ALLOC_FROM_PARTIAL
);
1857 put_cpu_partial(s
, page
, 0);
1858 stat(s
, CPU_PARTIAL_NODE
);
1860 if (!kmem_cache_has_cpu_partial(s
)
1861 || available
> slub_cpu_partial(s
) / 2)
1865 spin_unlock(&n
->list_lock
);
1870 * Get a page from somewhere. Search in increasing NUMA distances.
1872 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1873 struct kmem_cache_cpu
*c
)
1876 struct zonelist
*zonelist
;
1879 enum zone_type high_zoneidx
= gfp_zone(flags
);
1881 unsigned int cpuset_mems_cookie
;
1884 * The defrag ratio allows a configuration of the tradeoffs between
1885 * inter node defragmentation and node local allocations. A lower
1886 * defrag_ratio increases the tendency to do local allocations
1887 * instead of attempting to obtain partial slabs from other nodes.
1889 * If the defrag_ratio is set to 0 then kmalloc() always
1890 * returns node local objects. If the ratio is higher then kmalloc()
1891 * may return off node objects because partial slabs are obtained
1892 * from other nodes and filled up.
1894 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1895 * (which makes defrag_ratio = 1000) then every (well almost)
1896 * allocation will first attempt to defrag slab caches on other nodes.
1897 * This means scanning over all nodes to look for partial slabs which
1898 * may be expensive if we do it every time we are trying to find a slab
1899 * with available objects.
1901 if (!s
->remote_node_defrag_ratio
||
1902 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1906 cpuset_mems_cookie
= read_mems_allowed_begin();
1907 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1908 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1909 struct kmem_cache_node
*n
;
1911 n
= get_node(s
, zone_to_nid(zone
));
1913 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1914 n
->nr_partial
> s
->min_partial
) {
1915 object
= get_partial_node(s
, n
, c
, flags
);
1918 * Don't check read_mems_allowed_retry()
1919 * here - if mems_allowed was updated in
1920 * parallel, that was a harmless race
1921 * between allocation and the cpuset
1928 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1934 * Get a partial page, lock it and return it.
1936 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1937 struct kmem_cache_cpu
*c
)
1940 int searchnode
= node
;
1942 if (node
== NUMA_NO_NODE
)
1943 searchnode
= numa_mem_id();
1944 else if (!node_present_pages(node
))
1945 searchnode
= node_to_mem_node(node
);
1947 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1948 if (object
|| node
!= NUMA_NO_NODE
)
1951 return get_any_partial(s
, flags
, c
);
1954 #ifdef CONFIG_PREEMPT
1956 * Calculate the next globally unique transaction for disambiguiation
1957 * during cmpxchg. The transactions start with the cpu number and are then
1958 * incremented by CONFIG_NR_CPUS.
1960 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1963 * No preemption supported therefore also no need to check for
1969 static inline unsigned long next_tid(unsigned long tid
)
1971 return tid
+ TID_STEP
;
1974 static inline unsigned int tid_to_cpu(unsigned long tid
)
1976 return tid
% TID_STEP
;
1979 static inline unsigned long tid_to_event(unsigned long tid
)
1981 return tid
/ TID_STEP
;
1984 static inline unsigned int init_tid(int cpu
)
1989 static inline void note_cmpxchg_failure(const char *n
,
1990 const struct kmem_cache
*s
, unsigned long tid
)
1992 #ifdef SLUB_DEBUG_CMPXCHG
1993 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1995 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1997 #ifdef CONFIG_PREEMPT
1998 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1999 pr_warn("due to cpu change %d -> %d\n",
2000 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2003 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2004 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2005 tid_to_event(tid
), tid_to_event(actual_tid
));
2007 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2008 actual_tid
, tid
, next_tid(tid
));
2010 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2013 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2017 for_each_possible_cpu(cpu
)
2018 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2022 * Remove the cpu slab
2024 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2025 void *freelist
, struct kmem_cache_cpu
*c
)
2027 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2028 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2030 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2032 int tail
= DEACTIVATE_TO_HEAD
;
2036 if (page
->freelist
) {
2037 stat(s
, DEACTIVATE_REMOTE_FREES
);
2038 tail
= DEACTIVATE_TO_TAIL
;
2042 * Stage one: Free all available per cpu objects back
2043 * to the page freelist while it is still frozen. Leave the
2046 * There is no need to take the list->lock because the page
2049 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2051 unsigned long counters
;
2054 prior
= page
->freelist
;
2055 counters
= page
->counters
;
2056 set_freepointer(s
, freelist
, prior
);
2057 new.counters
= counters
;
2059 VM_BUG_ON(!new.frozen
);
2061 } while (!__cmpxchg_double_slab(s
, page
,
2063 freelist
, new.counters
,
2064 "drain percpu freelist"));
2066 freelist
= nextfree
;
2070 * Stage two: Ensure that the page is unfrozen while the
2071 * list presence reflects the actual number of objects
2074 * We setup the list membership and then perform a cmpxchg
2075 * with the count. If there is a mismatch then the page
2076 * is not unfrozen but the page is on the wrong list.
2078 * Then we restart the process which may have to remove
2079 * the page from the list that we just put it on again
2080 * because the number of objects in the slab may have
2085 old
.freelist
= page
->freelist
;
2086 old
.counters
= page
->counters
;
2087 VM_BUG_ON(!old
.frozen
);
2089 /* Determine target state of the slab */
2090 new.counters
= old
.counters
;
2093 set_freepointer(s
, freelist
, old
.freelist
);
2094 new.freelist
= freelist
;
2096 new.freelist
= old
.freelist
;
2100 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2102 else if (new.freelist
) {
2107 * Taking the spinlock removes the possiblity
2108 * that acquire_slab() will see a slab page that
2111 spin_lock(&n
->list_lock
);
2115 if (kmem_cache_debug(s
) && !lock
) {
2118 * This also ensures that the scanning of full
2119 * slabs from diagnostic functions will not see
2122 spin_lock(&n
->list_lock
);
2130 remove_partial(n
, page
);
2132 else if (l
== M_FULL
)
2134 remove_full(s
, n
, page
);
2136 if (m
== M_PARTIAL
) {
2138 add_partial(n
, page
, tail
);
2141 } else if (m
== M_FULL
) {
2143 stat(s
, DEACTIVATE_FULL
);
2144 add_full(s
, n
, page
);
2150 if (!__cmpxchg_double_slab(s
, page
,
2151 old
.freelist
, old
.counters
,
2152 new.freelist
, new.counters
,
2157 spin_unlock(&n
->list_lock
);
2160 stat(s
, DEACTIVATE_EMPTY
);
2161 discard_slab(s
, page
);
2170 * Unfreeze all the cpu partial slabs.
2172 * This function must be called with interrupts disabled
2173 * for the cpu using c (or some other guarantee must be there
2174 * to guarantee no concurrent accesses).
2176 static void unfreeze_partials(struct kmem_cache
*s
,
2177 struct kmem_cache_cpu
*c
)
2179 #ifdef CONFIG_SLUB_CPU_PARTIAL
2180 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2181 struct page
*page
, *discard_page
= NULL
;
2183 while ((page
= c
->partial
)) {
2187 c
->partial
= page
->next
;
2189 n2
= get_node(s
, page_to_nid(page
));
2192 spin_unlock(&n
->list_lock
);
2195 spin_lock(&n
->list_lock
);
2200 old
.freelist
= page
->freelist
;
2201 old
.counters
= page
->counters
;
2202 VM_BUG_ON(!old
.frozen
);
2204 new.counters
= old
.counters
;
2205 new.freelist
= old
.freelist
;
2209 } while (!__cmpxchg_double_slab(s
, page
,
2210 old
.freelist
, old
.counters
,
2211 new.freelist
, new.counters
,
2212 "unfreezing slab"));
2214 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2215 page
->next
= discard_page
;
2216 discard_page
= page
;
2218 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2219 stat(s
, FREE_ADD_PARTIAL
);
2224 spin_unlock(&n
->list_lock
);
2226 while (discard_page
) {
2227 page
= discard_page
;
2228 discard_page
= discard_page
->next
;
2230 stat(s
, DEACTIVATE_EMPTY
);
2231 discard_slab(s
, page
);
2238 * Put a page that was just frozen (in __slab_free) into a partial page
2239 * slot if available. This is done without interrupts disabled and without
2240 * preemption disabled. The cmpxchg is racy and may put the partial page
2241 * onto a random cpus partial slot.
2243 * If we did not find a slot then simply move all the partials to the
2244 * per node partial list.
2246 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2248 #ifdef CONFIG_SLUB_CPU_PARTIAL
2249 struct page
*oldpage
;
2257 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2260 pobjects
= oldpage
->pobjects
;
2261 pages
= oldpage
->pages
;
2262 if (drain
&& pobjects
> s
->cpu_partial
) {
2263 unsigned long flags
;
2265 * partial array is full. Move the existing
2266 * set to the per node partial list.
2268 local_irq_save(flags
);
2269 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2270 local_irq_restore(flags
);
2274 stat(s
, CPU_PARTIAL_DRAIN
);
2279 pobjects
+= page
->objects
- page
->inuse
;
2281 page
->pages
= pages
;
2282 page
->pobjects
= pobjects
;
2283 page
->next
= oldpage
;
2285 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2287 if (unlikely(!s
->cpu_partial
)) {
2288 unsigned long flags
;
2290 local_irq_save(flags
);
2291 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2292 local_irq_restore(flags
);
2298 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2300 stat(s
, CPUSLAB_FLUSH
);
2301 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2303 c
->tid
= next_tid(c
->tid
);
2309 * Called from IPI handler with interrupts disabled.
2311 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2313 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2319 unfreeze_partials(s
, c
);
2323 static void flush_cpu_slab(void *d
)
2325 struct kmem_cache
*s
= d
;
2327 __flush_cpu_slab(s
, smp_processor_id());
2330 static bool has_cpu_slab(int cpu
, void *info
)
2332 struct kmem_cache
*s
= info
;
2333 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2335 return c
->page
|| slub_percpu_partial(c
);
2338 static void flush_all(struct kmem_cache
*s
)
2340 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2344 * Use the cpu notifier to insure that the cpu slabs are flushed when
2347 static int slub_cpu_dead(unsigned int cpu
)
2349 struct kmem_cache
*s
;
2350 unsigned long flags
;
2352 mutex_lock(&slab_mutex
);
2353 list_for_each_entry(s
, &slab_caches
, list
) {
2354 local_irq_save(flags
);
2355 __flush_cpu_slab(s
, cpu
);
2356 local_irq_restore(flags
);
2358 mutex_unlock(&slab_mutex
);
2363 * Check if the objects in a per cpu structure fit numa
2364 * locality expectations.
2366 static inline int node_match(struct page
*page
, int node
)
2369 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2375 #ifdef CONFIG_SLUB_DEBUG
2376 static int count_free(struct page
*page
)
2378 return page
->objects
- page
->inuse
;
2381 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2383 return atomic_long_read(&n
->total_objects
);
2385 #endif /* CONFIG_SLUB_DEBUG */
2387 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2388 static unsigned long count_partial(struct kmem_cache_node
*n
,
2389 int (*get_count
)(struct page
*))
2391 unsigned long flags
;
2392 unsigned long x
= 0;
2395 spin_lock_irqsave(&n
->list_lock
, flags
);
2396 list_for_each_entry(page
, &n
->partial
, lru
)
2397 x
+= get_count(page
);
2398 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2401 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2403 static noinline
void
2404 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2406 #ifdef CONFIG_SLUB_DEBUG
2407 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2408 DEFAULT_RATELIMIT_BURST
);
2410 struct kmem_cache_node
*n
;
2412 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2415 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2416 nid
, gfpflags
, &gfpflags
);
2417 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2418 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2421 if (oo_order(s
->min
) > get_order(s
->object_size
))
2422 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2425 for_each_kmem_cache_node(s
, node
, n
) {
2426 unsigned long nr_slabs
;
2427 unsigned long nr_objs
;
2428 unsigned long nr_free
;
2430 nr_free
= count_partial(n
, count_free
);
2431 nr_slabs
= node_nr_slabs(n
);
2432 nr_objs
= node_nr_objs(n
);
2434 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2435 node
, nr_slabs
, nr_objs
, nr_free
);
2440 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2441 int node
, struct kmem_cache_cpu
**pc
)
2444 struct kmem_cache_cpu
*c
= *pc
;
2447 freelist
= get_partial(s
, flags
, node
, c
);
2452 page
= new_slab(s
, flags
, node
);
2454 c
= raw_cpu_ptr(s
->cpu_slab
);
2459 * No other reference to the page yet so we can
2460 * muck around with it freely without cmpxchg
2462 freelist
= page
->freelist
;
2463 page
->freelist
= NULL
;
2465 stat(s
, ALLOC_SLAB
);
2474 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2476 if (unlikely(PageSlabPfmemalloc(page
)))
2477 return gfp_pfmemalloc_allowed(gfpflags
);
2483 * Check the page->freelist of a page and either transfer the freelist to the
2484 * per cpu freelist or deactivate the page.
2486 * The page is still frozen if the return value is not NULL.
2488 * If this function returns NULL then the page has been unfrozen.
2490 * This function must be called with interrupt disabled.
2492 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2495 unsigned long counters
;
2499 freelist
= page
->freelist
;
2500 counters
= page
->counters
;
2502 new.counters
= counters
;
2503 VM_BUG_ON(!new.frozen
);
2505 new.inuse
= page
->objects
;
2506 new.frozen
= freelist
!= NULL
;
2508 } while (!__cmpxchg_double_slab(s
, page
,
2517 * Slow path. The lockless freelist is empty or we need to perform
2520 * Processing is still very fast if new objects have been freed to the
2521 * regular freelist. In that case we simply take over the regular freelist
2522 * as the lockless freelist and zap the regular freelist.
2524 * If that is not working then we fall back to the partial lists. We take the
2525 * first element of the freelist as the object to allocate now and move the
2526 * rest of the freelist to the lockless freelist.
2528 * And if we were unable to get a new slab from the partial slab lists then
2529 * we need to allocate a new slab. This is the slowest path since it involves
2530 * a call to the page allocator and the setup of a new slab.
2532 * Version of __slab_alloc to use when we know that interrupts are
2533 * already disabled (which is the case for bulk allocation).
2535 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2536 unsigned long addr
, struct kmem_cache_cpu
*c
)
2546 if (unlikely(!node_match(page
, node
))) {
2547 int searchnode
= node
;
2549 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2550 searchnode
= node_to_mem_node(node
);
2552 if (unlikely(!node_match(page
, searchnode
))) {
2553 stat(s
, ALLOC_NODE_MISMATCH
);
2554 deactivate_slab(s
, page
, c
->freelist
, c
);
2560 * By rights, we should be searching for a slab page that was
2561 * PFMEMALLOC but right now, we are losing the pfmemalloc
2562 * information when the page leaves the per-cpu allocator
2564 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2565 deactivate_slab(s
, page
, c
->freelist
, c
);
2569 /* must check again c->freelist in case of cpu migration or IRQ */
2570 freelist
= c
->freelist
;
2574 freelist
= get_freelist(s
, page
);
2578 stat(s
, DEACTIVATE_BYPASS
);
2582 stat(s
, ALLOC_REFILL
);
2586 * freelist is pointing to the list of objects to be used.
2587 * page is pointing to the page from which the objects are obtained.
2588 * That page must be frozen for per cpu allocations to work.
2590 VM_BUG_ON(!c
->page
->frozen
);
2591 c
->freelist
= get_freepointer(s
, freelist
);
2592 c
->tid
= next_tid(c
->tid
);
2597 if (slub_percpu_partial(c
)) {
2598 page
= c
->page
= slub_percpu_partial(c
);
2599 slub_set_percpu_partial(c
, page
);
2600 stat(s
, CPU_PARTIAL_ALLOC
);
2604 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2606 if (unlikely(!freelist
)) {
2607 slab_out_of_memory(s
, gfpflags
, node
);
2612 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2615 /* Only entered in the debug case */
2616 if (kmem_cache_debug(s
) &&
2617 !alloc_debug_processing(s
, page
, freelist
, addr
))
2618 goto new_slab
; /* Slab failed checks. Next slab needed */
2620 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2625 * Another one that disabled interrupt and compensates for possible
2626 * cpu changes by refetching the per cpu area pointer.
2628 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2629 unsigned long addr
, struct kmem_cache_cpu
*c
)
2632 unsigned long flags
;
2634 local_irq_save(flags
);
2635 #ifdef CONFIG_PREEMPT
2637 * We may have been preempted and rescheduled on a different
2638 * cpu before disabling interrupts. Need to reload cpu area
2641 c
= this_cpu_ptr(s
->cpu_slab
);
2644 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2645 local_irq_restore(flags
);
2650 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2651 * have the fastpath folded into their functions. So no function call
2652 * overhead for requests that can be satisfied on the fastpath.
2654 * The fastpath works by first checking if the lockless freelist can be used.
2655 * If not then __slab_alloc is called for slow processing.
2657 * Otherwise we can simply pick the next object from the lockless free list.
2659 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2660 gfp_t gfpflags
, int node
, unsigned long addr
)
2663 struct kmem_cache_cpu
*c
;
2667 s
= slab_pre_alloc_hook(s
, gfpflags
);
2672 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2673 * enabled. We may switch back and forth between cpus while
2674 * reading from one cpu area. That does not matter as long
2675 * as we end up on the original cpu again when doing the cmpxchg.
2677 * We should guarantee that tid and kmem_cache are retrieved on
2678 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2679 * to check if it is matched or not.
2682 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2683 c
= raw_cpu_ptr(s
->cpu_slab
);
2684 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2685 unlikely(tid
!= READ_ONCE(c
->tid
)));
2688 * Irqless object alloc/free algorithm used here depends on sequence
2689 * of fetching cpu_slab's data. tid should be fetched before anything
2690 * on c to guarantee that object and page associated with previous tid
2691 * won't be used with current tid. If we fetch tid first, object and
2692 * page could be one associated with next tid and our alloc/free
2693 * request will be failed. In this case, we will retry. So, no problem.
2698 * The transaction ids are globally unique per cpu and per operation on
2699 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2700 * occurs on the right processor and that there was no operation on the
2701 * linked list in between.
2704 object
= c
->freelist
;
2706 if (unlikely(!object
|| !node_match(page
, node
))) {
2707 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2708 stat(s
, ALLOC_SLOWPATH
);
2710 void *next_object
= get_freepointer_safe(s
, object
);
2713 * The cmpxchg will only match if there was no additional
2714 * operation and if we are on the right processor.
2716 * The cmpxchg does the following atomically (without lock
2718 * 1. Relocate first pointer to the current per cpu area.
2719 * 2. Verify that tid and freelist have not been changed
2720 * 3. If they were not changed replace tid and freelist
2722 * Since this is without lock semantics the protection is only
2723 * against code executing on this cpu *not* from access by
2726 if (unlikely(!this_cpu_cmpxchg_double(
2727 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2729 next_object
, next_tid(tid
)))) {
2731 note_cmpxchg_failure("slab_alloc", s
, tid
);
2734 prefetch_freepointer(s
, next_object
);
2735 stat(s
, ALLOC_FASTPATH
);
2738 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2739 memset(object
, 0, s
->object_size
);
2741 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2746 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2747 gfp_t gfpflags
, unsigned long addr
)
2749 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2752 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2754 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2756 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2761 EXPORT_SYMBOL(kmem_cache_alloc
);
2763 #ifdef CONFIG_TRACING
2764 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2766 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2767 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2768 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2771 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2775 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2777 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2779 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2780 s
->object_size
, s
->size
, gfpflags
, node
);
2784 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2786 #ifdef CONFIG_TRACING
2787 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2789 int node
, size_t size
)
2791 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2793 trace_kmalloc_node(_RET_IP_
, ret
,
2794 size
, s
->size
, gfpflags
, node
);
2796 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2799 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2804 * Slow path handling. This may still be called frequently since objects
2805 * have a longer lifetime than the cpu slabs in most processing loads.
2807 * So we still attempt to reduce cache line usage. Just take the slab
2808 * lock and free the item. If there is no additional partial page
2809 * handling required then we can return immediately.
2811 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2812 void *head
, void *tail
, int cnt
,
2819 unsigned long counters
;
2820 struct kmem_cache_node
*n
= NULL
;
2821 unsigned long uninitialized_var(flags
);
2823 stat(s
, FREE_SLOWPATH
);
2825 if (kmem_cache_debug(s
) &&
2826 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2831 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2834 prior
= page
->freelist
;
2835 counters
= page
->counters
;
2836 set_freepointer(s
, tail
, prior
);
2837 new.counters
= counters
;
2838 was_frozen
= new.frozen
;
2840 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2842 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2845 * Slab was on no list before and will be
2847 * We can defer the list move and instead
2852 } else { /* Needs to be taken off a list */
2854 n
= get_node(s
, page_to_nid(page
));
2856 * Speculatively acquire the list_lock.
2857 * If the cmpxchg does not succeed then we may
2858 * drop the list_lock without any processing.
2860 * Otherwise the list_lock will synchronize with
2861 * other processors updating the list of slabs.
2863 spin_lock_irqsave(&n
->list_lock
, flags
);
2868 } while (!cmpxchg_double_slab(s
, page
,
2876 * If we just froze the page then put it onto the
2877 * per cpu partial list.
2879 if (new.frozen
&& !was_frozen
) {
2880 put_cpu_partial(s
, page
, 1);
2881 stat(s
, CPU_PARTIAL_FREE
);
2884 * The list lock was not taken therefore no list
2885 * activity can be necessary.
2888 stat(s
, FREE_FROZEN
);
2892 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2896 * Objects left in the slab. If it was not on the partial list before
2899 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2900 if (kmem_cache_debug(s
))
2901 remove_full(s
, n
, page
);
2902 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2903 stat(s
, FREE_ADD_PARTIAL
);
2905 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2911 * Slab on the partial list.
2913 remove_partial(n
, page
);
2914 stat(s
, FREE_REMOVE_PARTIAL
);
2916 /* Slab must be on the full list */
2917 remove_full(s
, n
, page
);
2920 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2922 discard_slab(s
, page
);
2926 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2927 * can perform fastpath freeing without additional function calls.
2929 * The fastpath is only possible if we are freeing to the current cpu slab
2930 * of this processor. This typically the case if we have just allocated
2933 * If fastpath is not possible then fall back to __slab_free where we deal
2934 * with all sorts of special processing.
2936 * Bulk free of a freelist with several objects (all pointing to the
2937 * same page) possible by specifying head and tail ptr, plus objects
2938 * count (cnt). Bulk free indicated by tail pointer being set.
2940 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2941 struct page
*page
, void *head
, void *tail
,
2942 int cnt
, unsigned long addr
)
2944 void *tail_obj
= tail
? : head
;
2945 struct kmem_cache_cpu
*c
;
2949 * Determine the currently cpus per cpu slab.
2950 * The cpu may change afterward. However that does not matter since
2951 * data is retrieved via this pointer. If we are on the same cpu
2952 * during the cmpxchg then the free will succeed.
2955 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2956 c
= raw_cpu_ptr(s
->cpu_slab
);
2957 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2958 unlikely(tid
!= READ_ONCE(c
->tid
)));
2960 /* Same with comment on barrier() in slab_alloc_node() */
2963 if (likely(page
== c
->page
)) {
2964 set_freepointer(s
, tail_obj
, c
->freelist
);
2966 if (unlikely(!this_cpu_cmpxchg_double(
2967 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2969 head
, next_tid(tid
)))) {
2971 note_cmpxchg_failure("slab_free", s
, tid
);
2974 stat(s
, FREE_FASTPATH
);
2976 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2980 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2981 void *head
, void *tail
, int cnt
,
2984 slab_free_freelist_hook(s
, head
, tail
);
2986 * slab_free_freelist_hook() could have put the items into quarantine.
2987 * If so, no need to free them.
2989 if (s
->flags
& SLAB_KASAN
&& !(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
2991 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2995 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2997 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3001 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3003 s
= cache_from_obj(s
, x
);
3006 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3007 trace_kmem_cache_free(_RET_IP_
, x
);
3009 EXPORT_SYMBOL(kmem_cache_free
);
3011 struct detached_freelist
{
3016 struct kmem_cache
*s
;
3020 * This function progressively scans the array with free objects (with
3021 * a limited look ahead) and extract objects belonging to the same
3022 * page. It builds a detached freelist directly within the given
3023 * page/objects. This can happen without any need for
3024 * synchronization, because the objects are owned by running process.
3025 * The freelist is build up as a single linked list in the objects.
3026 * The idea is, that this detached freelist can then be bulk
3027 * transferred to the real freelist(s), but only requiring a single
3028 * synchronization primitive. Look ahead in the array is limited due
3029 * to performance reasons.
3032 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3033 void **p
, struct detached_freelist
*df
)
3035 size_t first_skipped_index
= 0;
3040 /* Always re-init detached_freelist */
3045 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3046 } while (!object
&& size
);
3051 page
= virt_to_head_page(object
);
3053 /* Handle kalloc'ed objects */
3054 if (unlikely(!PageSlab(page
))) {
3055 BUG_ON(!PageCompound(page
));
3057 __free_pages(page
, compound_order(page
));
3058 p
[size
] = NULL
; /* mark object processed */
3061 /* Derive kmem_cache from object */
3062 df
->s
= page
->slab_cache
;
3064 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3067 /* Start new detached freelist */
3069 set_freepointer(df
->s
, object
, NULL
);
3071 df
->freelist
= object
;
3072 p
[size
] = NULL
; /* mark object processed */
3078 continue; /* Skip processed objects */
3080 /* df->page is always set at this point */
3081 if (df
->page
== virt_to_head_page(object
)) {
3082 /* Opportunity build freelist */
3083 set_freepointer(df
->s
, object
, df
->freelist
);
3084 df
->freelist
= object
;
3086 p
[size
] = NULL
; /* mark object processed */
3091 /* Limit look ahead search */
3095 if (!first_skipped_index
)
3096 first_skipped_index
= size
+ 1;
3099 return first_skipped_index
;
3102 /* Note that interrupts must be enabled when calling this function. */
3103 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3109 struct detached_freelist df
;
3111 size
= build_detached_freelist(s
, size
, p
, &df
);
3115 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3116 } while (likely(size
));
3118 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3120 /* Note that interrupts must be enabled when calling this function. */
3121 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3124 struct kmem_cache_cpu
*c
;
3127 /* memcg and kmem_cache debug support */
3128 s
= slab_pre_alloc_hook(s
, flags
);
3132 * Drain objects in the per cpu slab, while disabling local
3133 * IRQs, which protects against PREEMPT and interrupts
3134 * handlers invoking normal fastpath.
3136 local_irq_disable();
3137 c
= this_cpu_ptr(s
->cpu_slab
);
3139 for (i
= 0; i
< size
; i
++) {
3140 void *object
= c
->freelist
;
3142 if (unlikely(!object
)) {
3144 * Invoking slow path likely have side-effect
3145 * of re-populating per CPU c->freelist
3147 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3149 if (unlikely(!p
[i
]))
3152 c
= this_cpu_ptr(s
->cpu_slab
);
3153 continue; /* goto for-loop */
3155 c
->freelist
= get_freepointer(s
, object
);
3158 c
->tid
= next_tid(c
->tid
);
3161 /* Clear memory outside IRQ disabled fastpath loop */
3162 if (unlikely(flags
& __GFP_ZERO
)) {
3165 for (j
= 0; j
< i
; j
++)
3166 memset(p
[j
], 0, s
->object_size
);
3169 /* memcg and kmem_cache debug support */
3170 slab_post_alloc_hook(s
, flags
, size
, p
);
3174 slab_post_alloc_hook(s
, flags
, i
, p
);
3175 __kmem_cache_free_bulk(s
, i
, p
);
3178 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3182 * Object placement in a slab is made very easy because we always start at
3183 * offset 0. If we tune the size of the object to the alignment then we can
3184 * get the required alignment by putting one properly sized object after
3187 * Notice that the allocation order determines the sizes of the per cpu
3188 * caches. Each processor has always one slab available for allocations.
3189 * Increasing the allocation order reduces the number of times that slabs
3190 * must be moved on and off the partial lists and is therefore a factor in
3195 * Mininum / Maximum order of slab pages. This influences locking overhead
3196 * and slab fragmentation. A higher order reduces the number of partial slabs
3197 * and increases the number of allocations possible without having to
3198 * take the list_lock.
3200 static int slub_min_order
;
3201 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3202 static int slub_min_objects
;
3205 * Calculate the order of allocation given an slab object size.
3207 * The order of allocation has significant impact on performance and other
3208 * system components. Generally order 0 allocations should be preferred since
3209 * order 0 does not cause fragmentation in the page allocator. Larger objects
3210 * be problematic to put into order 0 slabs because there may be too much
3211 * unused space left. We go to a higher order if more than 1/16th of the slab
3214 * In order to reach satisfactory performance we must ensure that a minimum
3215 * number of objects is in one slab. Otherwise we may generate too much
3216 * activity on the partial lists which requires taking the list_lock. This is
3217 * less a concern for large slabs though which are rarely used.
3219 * slub_max_order specifies the order where we begin to stop considering the
3220 * number of objects in a slab as critical. If we reach slub_max_order then
3221 * we try to keep the page order as low as possible. So we accept more waste
3222 * of space in favor of a small page order.
3224 * Higher order allocations also allow the placement of more objects in a
3225 * slab and thereby reduce object handling overhead. If the user has
3226 * requested a higher mininum order then we start with that one instead of
3227 * the smallest order which will fit the object.
3229 static inline int slab_order(int size
, int min_objects
,
3230 int max_order
, int fract_leftover
, int reserved
)
3234 int min_order
= slub_min_order
;
3236 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3237 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3239 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3240 order
<= max_order
; order
++) {
3242 unsigned long slab_size
= PAGE_SIZE
<< order
;
3244 rem
= (slab_size
- reserved
) % size
;
3246 if (rem
<= slab_size
/ fract_leftover
)
3253 static inline int calculate_order(int size
, int reserved
)
3261 * Attempt to find best configuration for a slab. This
3262 * works by first attempting to generate a layout with
3263 * the best configuration and backing off gradually.
3265 * First we increase the acceptable waste in a slab. Then
3266 * we reduce the minimum objects required in a slab.
3268 min_objects
= slub_min_objects
;
3270 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3271 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3272 min_objects
= min(min_objects
, max_objects
);
3274 while (min_objects
> 1) {
3276 while (fraction
>= 4) {
3277 order
= slab_order(size
, min_objects
,
3278 slub_max_order
, fraction
, reserved
);
3279 if (order
<= slub_max_order
)
3287 * We were unable to place multiple objects in a slab. Now
3288 * lets see if we can place a single object there.
3290 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3291 if (order
<= slub_max_order
)
3295 * Doh this slab cannot be placed using slub_max_order.
3297 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3298 if (order
< MAX_ORDER
)
3304 init_kmem_cache_node(struct kmem_cache_node
*n
)
3307 spin_lock_init(&n
->list_lock
);
3308 INIT_LIST_HEAD(&n
->partial
);
3309 #ifdef CONFIG_SLUB_DEBUG
3310 atomic_long_set(&n
->nr_slabs
, 0);
3311 atomic_long_set(&n
->total_objects
, 0);
3312 INIT_LIST_HEAD(&n
->full
);
3316 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3318 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3319 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3322 * Must align to double word boundary for the double cmpxchg
3323 * instructions to work; see __pcpu_double_call_return_bool().
3325 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3326 2 * sizeof(void *));
3331 init_kmem_cache_cpus(s
);
3336 static struct kmem_cache
*kmem_cache_node
;
3339 * No kmalloc_node yet so do it by hand. We know that this is the first
3340 * slab on the node for this slabcache. There are no concurrent accesses
3343 * Note that this function only works on the kmem_cache_node
3344 * when allocating for the kmem_cache_node. This is used for bootstrapping
3345 * memory on a fresh node that has no slab structures yet.
3347 static void early_kmem_cache_node_alloc(int node
)
3350 struct kmem_cache_node
*n
;
3352 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3354 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3357 if (page_to_nid(page
) != node
) {
3358 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3359 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3364 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3367 kmem_cache_node
->node
[node
] = n
;
3368 #ifdef CONFIG_SLUB_DEBUG
3369 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3370 init_tracking(kmem_cache_node
, n
);
3372 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3374 init_kmem_cache_node(n
);
3375 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3378 * No locks need to be taken here as it has just been
3379 * initialized and there is no concurrent access.
3381 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3384 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3387 struct kmem_cache_node
*n
;
3389 for_each_kmem_cache_node(s
, node
, n
) {
3390 s
->node
[node
] = NULL
;
3391 kmem_cache_free(kmem_cache_node
, n
);
3395 void __kmem_cache_release(struct kmem_cache
*s
)
3397 cache_random_seq_destroy(s
);
3398 free_percpu(s
->cpu_slab
);
3399 free_kmem_cache_nodes(s
);
3402 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3406 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3407 struct kmem_cache_node
*n
;
3409 if (slab_state
== DOWN
) {
3410 early_kmem_cache_node_alloc(node
);
3413 n
= kmem_cache_alloc_node(kmem_cache_node
,
3417 free_kmem_cache_nodes(s
);
3421 init_kmem_cache_node(n
);
3427 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3429 if (min
< MIN_PARTIAL
)
3431 else if (min
> MAX_PARTIAL
)
3433 s
->min_partial
= min
;
3436 static void set_cpu_partial(struct kmem_cache
*s
)
3438 #ifdef CONFIG_SLUB_CPU_PARTIAL
3440 * cpu_partial determined the maximum number of objects kept in the
3441 * per cpu partial lists of a processor.
3443 * Per cpu partial lists mainly contain slabs that just have one
3444 * object freed. If they are used for allocation then they can be
3445 * filled up again with minimal effort. The slab will never hit the
3446 * per node partial lists and therefore no locking will be required.
3448 * This setting also determines
3450 * A) The number of objects from per cpu partial slabs dumped to the
3451 * per node list when we reach the limit.
3452 * B) The number of objects in cpu partial slabs to extract from the
3453 * per node list when we run out of per cpu objects. We only fetch
3454 * 50% to keep some capacity around for frees.
3456 if (!kmem_cache_has_cpu_partial(s
))
3458 else if (s
->size
>= PAGE_SIZE
)
3460 else if (s
->size
>= 1024)
3462 else if (s
->size
>= 256)
3463 s
->cpu_partial
= 13;
3465 s
->cpu_partial
= 30;
3470 * calculate_sizes() determines the order and the distribution of data within
3473 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3475 unsigned long flags
= s
->flags
;
3476 size_t size
= s
->object_size
;
3480 * Round up object size to the next word boundary. We can only
3481 * place the free pointer at word boundaries and this determines
3482 * the possible location of the free pointer.
3484 size
= ALIGN(size
, sizeof(void *));
3486 #ifdef CONFIG_SLUB_DEBUG
3488 * Determine if we can poison the object itself. If the user of
3489 * the slab may touch the object after free or before allocation
3490 * then we should never poison the object itself.
3492 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3494 s
->flags
|= __OBJECT_POISON
;
3496 s
->flags
&= ~__OBJECT_POISON
;
3500 * If we are Redzoning then check if there is some space between the
3501 * end of the object and the free pointer. If not then add an
3502 * additional word to have some bytes to store Redzone information.
3504 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3505 size
+= sizeof(void *);
3509 * With that we have determined the number of bytes in actual use
3510 * by the object. This is the potential offset to the free pointer.
3514 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3517 * Relocate free pointer after the object if it is not
3518 * permitted to overwrite the first word of the object on
3521 * This is the case if we do RCU, have a constructor or
3522 * destructor or are poisoning the objects.
3525 size
+= sizeof(void *);
3528 #ifdef CONFIG_SLUB_DEBUG
3529 if (flags
& SLAB_STORE_USER
)
3531 * Need to store information about allocs and frees after
3534 size
+= 2 * sizeof(struct track
);
3537 kasan_cache_create(s
, &size
, &s
->flags
);
3538 #ifdef CONFIG_SLUB_DEBUG
3539 if (flags
& SLAB_RED_ZONE
) {
3541 * Add some empty padding so that we can catch
3542 * overwrites from earlier objects rather than let
3543 * tracking information or the free pointer be
3544 * corrupted if a user writes before the start
3547 size
+= sizeof(void *);
3549 s
->red_left_pad
= sizeof(void *);
3550 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3551 size
+= s
->red_left_pad
;
3556 * SLUB stores one object immediately after another beginning from
3557 * offset 0. In order to align the objects we have to simply size
3558 * each object to conform to the alignment.
3560 size
= ALIGN(size
, s
->align
);
3562 if (forced_order
>= 0)
3563 order
= forced_order
;
3565 order
= calculate_order(size
, s
->reserved
);
3572 s
->allocflags
|= __GFP_COMP
;
3574 if (s
->flags
& SLAB_CACHE_DMA
)
3575 s
->allocflags
|= GFP_DMA
;
3577 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3578 s
->allocflags
|= __GFP_RECLAIMABLE
;
3581 * Determine the number of objects per slab
3583 s
->oo
= oo_make(order
, size
, s
->reserved
);
3584 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3585 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3588 return !!oo_objects(s
->oo
);
3591 static int kmem_cache_open(struct kmem_cache
*s
, unsigned long flags
)
3593 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3595 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3596 s
->random
= get_random_long();
3599 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_TYPESAFE_BY_RCU
))
3600 s
->reserved
= sizeof(struct rcu_head
);
3602 if (!calculate_sizes(s
, -1))
3604 if (disable_higher_order_debug
) {
3606 * Disable debugging flags that store metadata if the min slab
3609 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3610 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3612 if (!calculate_sizes(s
, -1))
3617 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3618 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3619 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3620 /* Enable fast mode */
3621 s
->flags
|= __CMPXCHG_DOUBLE
;
3625 * The larger the object size is, the more pages we want on the partial
3626 * list to avoid pounding the page allocator excessively.
3628 set_min_partial(s
, ilog2(s
->size
) / 2);
3633 s
->remote_node_defrag_ratio
= 1000;
3636 /* Initialize the pre-computed randomized freelist if slab is up */
3637 if (slab_state
>= UP
) {
3638 if (init_cache_random_seq(s
))
3642 if (!init_kmem_cache_nodes(s
))
3645 if (alloc_kmem_cache_cpus(s
))
3648 free_kmem_cache_nodes(s
);
3650 if (flags
& SLAB_PANIC
)
3651 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3652 s
->name
, (unsigned long)s
->size
, s
->size
,
3653 oo_order(s
->oo
), s
->offset
, flags
);
3657 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3660 #ifdef CONFIG_SLUB_DEBUG
3661 void *addr
= page_address(page
);
3663 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3664 sizeof(long), GFP_ATOMIC
);
3667 slab_err(s
, page
, text
, s
->name
);
3670 get_map(s
, page
, map
);
3671 for_each_object(p
, s
, addr
, page
->objects
) {
3673 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3674 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3675 print_tracking(s
, p
);
3684 * Attempt to free all partial slabs on a node.
3685 * This is called from __kmem_cache_shutdown(). We must take list_lock
3686 * because sysfs file might still access partial list after the shutdowning.
3688 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3691 struct page
*page
, *h
;
3693 BUG_ON(irqs_disabled());
3694 spin_lock_irq(&n
->list_lock
);
3695 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3697 remove_partial(n
, page
);
3698 list_add(&page
->lru
, &discard
);
3700 list_slab_objects(s
, page
,
3701 "Objects remaining in %s on __kmem_cache_shutdown()");
3704 spin_unlock_irq(&n
->list_lock
);
3706 list_for_each_entry_safe(page
, h
, &discard
, lru
)
3707 discard_slab(s
, page
);
3711 * Release all resources used by a slab cache.
3713 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3716 struct kmem_cache_node
*n
;
3719 /* Attempt to free all objects */
3720 for_each_kmem_cache_node(s
, node
, n
) {
3722 if (n
->nr_partial
|| slabs_node(s
, node
))
3725 sysfs_slab_remove(s
);
3729 /********************************************************************
3731 *******************************************************************/
3733 static int __init
setup_slub_min_order(char *str
)
3735 get_option(&str
, &slub_min_order
);
3740 __setup("slub_min_order=", setup_slub_min_order
);
3742 static int __init
setup_slub_max_order(char *str
)
3744 get_option(&str
, &slub_max_order
);
3745 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3750 __setup("slub_max_order=", setup_slub_max_order
);
3752 static int __init
setup_slub_min_objects(char *str
)
3754 get_option(&str
, &slub_min_objects
);
3759 __setup("slub_min_objects=", setup_slub_min_objects
);
3761 void *__kmalloc(size_t size
, gfp_t flags
)
3763 struct kmem_cache
*s
;
3766 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3767 return kmalloc_large(size
, flags
);
3769 s
= kmalloc_slab(size
, flags
);
3771 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3774 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3776 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3778 kasan_kmalloc(s
, ret
, size
, flags
);
3782 EXPORT_SYMBOL(__kmalloc
);
3785 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3790 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
3791 page
= alloc_pages_node(node
, flags
, get_order(size
));
3793 ptr
= page_address(page
);
3795 kmalloc_large_node_hook(ptr
, size
, flags
);
3799 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3801 struct kmem_cache
*s
;
3804 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3805 ret
= kmalloc_large_node(size
, flags
, node
);
3807 trace_kmalloc_node(_RET_IP_
, ret
,
3808 size
, PAGE_SIZE
<< get_order(size
),
3814 s
= kmalloc_slab(size
, flags
);
3816 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3819 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3821 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3823 kasan_kmalloc(s
, ret
, size
, flags
);
3827 EXPORT_SYMBOL(__kmalloc_node
);
3830 #ifdef CONFIG_HARDENED_USERCOPY
3832 * Rejects objects that are incorrectly sized.
3834 * Returns NULL if check passes, otherwise const char * to name of cache
3835 * to indicate an error.
3837 const char *__check_heap_object(const void *ptr
, unsigned long n
,
3840 struct kmem_cache
*s
;
3841 unsigned long offset
;
3844 /* Find object and usable object size. */
3845 s
= page
->slab_cache
;
3846 object_size
= slab_ksize(s
);
3848 /* Reject impossible pointers. */
3849 if (ptr
< page_address(page
))
3852 /* Find offset within object. */
3853 offset
= (ptr
- page_address(page
)) % s
->size
;
3855 /* Adjust for redzone and reject if within the redzone. */
3856 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3857 if (offset
< s
->red_left_pad
)
3859 offset
-= s
->red_left_pad
;
3862 /* Allow address range falling entirely within object size. */
3863 if (offset
<= object_size
&& n
<= object_size
- offset
)
3868 #endif /* CONFIG_HARDENED_USERCOPY */
3870 static size_t __ksize(const void *object
)
3874 if (unlikely(object
== ZERO_SIZE_PTR
))
3877 page
= virt_to_head_page(object
);
3879 if (unlikely(!PageSlab(page
))) {
3880 WARN_ON(!PageCompound(page
));
3881 return PAGE_SIZE
<< compound_order(page
);
3884 return slab_ksize(page
->slab_cache
);
3887 size_t ksize(const void *object
)
3889 size_t size
= __ksize(object
);
3890 /* We assume that ksize callers could use whole allocated area,
3891 * so we need to unpoison this area.
3893 kasan_unpoison_shadow(object
, size
);
3896 EXPORT_SYMBOL(ksize
);
3898 void kfree(const void *x
)
3901 void *object
= (void *)x
;
3903 trace_kfree(_RET_IP_
, x
);
3905 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3908 page
= virt_to_head_page(x
);
3909 if (unlikely(!PageSlab(page
))) {
3910 BUG_ON(!PageCompound(page
));
3912 __free_pages(page
, compound_order(page
));
3915 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3917 EXPORT_SYMBOL(kfree
);
3919 #define SHRINK_PROMOTE_MAX 32
3922 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3923 * up most to the head of the partial lists. New allocations will then
3924 * fill those up and thus they can be removed from the partial lists.
3926 * The slabs with the least items are placed last. This results in them
3927 * being allocated from last increasing the chance that the last objects
3928 * are freed in them.
3930 int __kmem_cache_shrink(struct kmem_cache
*s
)
3934 struct kmem_cache_node
*n
;
3937 struct list_head discard
;
3938 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3939 unsigned long flags
;
3943 for_each_kmem_cache_node(s
, node
, n
) {
3944 INIT_LIST_HEAD(&discard
);
3945 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3946 INIT_LIST_HEAD(promote
+ i
);
3948 spin_lock_irqsave(&n
->list_lock
, flags
);
3951 * Build lists of slabs to discard or promote.
3953 * Note that concurrent frees may occur while we hold the
3954 * list_lock. page->inuse here is the upper limit.
3956 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3957 int free
= page
->objects
- page
->inuse
;
3959 /* Do not reread page->inuse */
3962 /* We do not keep full slabs on the list */
3965 if (free
== page
->objects
) {
3966 list_move(&page
->lru
, &discard
);
3968 } else if (free
<= SHRINK_PROMOTE_MAX
)
3969 list_move(&page
->lru
, promote
+ free
- 1);
3973 * Promote the slabs filled up most to the head of the
3976 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3977 list_splice(promote
+ i
, &n
->partial
);
3979 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3981 /* Release empty slabs */
3982 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3983 discard_slab(s
, page
);
3985 if (slabs_node(s
, node
))
3993 static void kmemcg_cache_deact_after_rcu(struct kmem_cache
*s
)
3996 * Called with all the locks held after a sched RCU grace period.
3997 * Even if @s becomes empty after shrinking, we can't know that @s
3998 * doesn't have allocations already in-flight and thus can't
3999 * destroy @s until the associated memcg is released.
4001 * However, let's remove the sysfs files for empty caches here.
4002 * Each cache has a lot of interface files which aren't
4003 * particularly useful for empty draining caches; otherwise, we can
4004 * easily end up with millions of unnecessary sysfs files on
4005 * systems which have a lot of memory and transient cgroups.
4007 if (!__kmem_cache_shrink(s
))
4008 sysfs_slab_remove(s
);
4011 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4014 * Disable empty slabs caching. Used to avoid pinning offline
4015 * memory cgroups by kmem pages that can be freed.
4017 slub_set_cpu_partial(s
, 0);
4021 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4022 * we have to make sure the change is visible before shrinking.
4024 slab_deactivate_memcg_cache_rcu_sched(s
, kmemcg_cache_deact_after_rcu
);
4028 static int slab_mem_going_offline_callback(void *arg
)
4030 struct kmem_cache
*s
;
4032 mutex_lock(&slab_mutex
);
4033 list_for_each_entry(s
, &slab_caches
, list
)
4034 __kmem_cache_shrink(s
);
4035 mutex_unlock(&slab_mutex
);
4040 static void slab_mem_offline_callback(void *arg
)
4042 struct kmem_cache_node
*n
;
4043 struct kmem_cache
*s
;
4044 struct memory_notify
*marg
= arg
;
4047 offline_node
= marg
->status_change_nid_normal
;
4050 * If the node still has available memory. we need kmem_cache_node
4053 if (offline_node
< 0)
4056 mutex_lock(&slab_mutex
);
4057 list_for_each_entry(s
, &slab_caches
, list
) {
4058 n
= get_node(s
, offline_node
);
4061 * if n->nr_slabs > 0, slabs still exist on the node
4062 * that is going down. We were unable to free them,
4063 * and offline_pages() function shouldn't call this
4064 * callback. So, we must fail.
4066 BUG_ON(slabs_node(s
, offline_node
));
4068 s
->node
[offline_node
] = NULL
;
4069 kmem_cache_free(kmem_cache_node
, n
);
4072 mutex_unlock(&slab_mutex
);
4075 static int slab_mem_going_online_callback(void *arg
)
4077 struct kmem_cache_node
*n
;
4078 struct kmem_cache
*s
;
4079 struct memory_notify
*marg
= arg
;
4080 int nid
= marg
->status_change_nid_normal
;
4084 * If the node's memory is already available, then kmem_cache_node is
4085 * already created. Nothing to do.
4091 * We are bringing a node online. No memory is available yet. We must
4092 * allocate a kmem_cache_node structure in order to bring the node
4095 mutex_lock(&slab_mutex
);
4096 list_for_each_entry(s
, &slab_caches
, list
) {
4098 * XXX: kmem_cache_alloc_node will fallback to other nodes
4099 * since memory is not yet available from the node that
4102 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4107 init_kmem_cache_node(n
);
4111 mutex_unlock(&slab_mutex
);
4115 static int slab_memory_callback(struct notifier_block
*self
,
4116 unsigned long action
, void *arg
)
4121 case MEM_GOING_ONLINE
:
4122 ret
= slab_mem_going_online_callback(arg
);
4124 case MEM_GOING_OFFLINE
:
4125 ret
= slab_mem_going_offline_callback(arg
);
4128 case MEM_CANCEL_ONLINE
:
4129 slab_mem_offline_callback(arg
);
4132 case MEM_CANCEL_OFFLINE
:
4136 ret
= notifier_from_errno(ret
);
4142 static struct notifier_block slab_memory_callback_nb
= {
4143 .notifier_call
= slab_memory_callback
,
4144 .priority
= SLAB_CALLBACK_PRI
,
4147 /********************************************************************
4148 * Basic setup of slabs
4149 *******************************************************************/
4152 * Used for early kmem_cache structures that were allocated using
4153 * the page allocator. Allocate them properly then fix up the pointers
4154 * that may be pointing to the wrong kmem_cache structure.
4157 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4160 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4161 struct kmem_cache_node
*n
;
4163 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4166 * This runs very early, and only the boot processor is supposed to be
4167 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4170 __flush_cpu_slab(s
, smp_processor_id());
4171 for_each_kmem_cache_node(s
, node
, n
) {
4174 list_for_each_entry(p
, &n
->partial
, lru
)
4177 #ifdef CONFIG_SLUB_DEBUG
4178 list_for_each_entry(p
, &n
->full
, lru
)
4182 slab_init_memcg_params(s
);
4183 list_add(&s
->list
, &slab_caches
);
4184 memcg_link_cache(s
);
4188 void __init
kmem_cache_init(void)
4190 static __initdata
struct kmem_cache boot_kmem_cache
,
4191 boot_kmem_cache_node
;
4193 if (debug_guardpage_minorder())
4196 kmem_cache_node
= &boot_kmem_cache_node
;
4197 kmem_cache
= &boot_kmem_cache
;
4199 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4200 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
4202 register_hotmemory_notifier(&slab_memory_callback_nb
);
4204 /* Able to allocate the per node structures */
4205 slab_state
= PARTIAL
;
4207 create_boot_cache(kmem_cache
, "kmem_cache",
4208 offsetof(struct kmem_cache
, node
) +
4209 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4210 SLAB_HWCACHE_ALIGN
);
4212 kmem_cache
= bootstrap(&boot_kmem_cache
);
4215 * Allocate kmem_cache_node properly from the kmem_cache slab.
4216 * kmem_cache_node is separately allocated so no need to
4217 * update any list pointers.
4219 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4221 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4222 setup_kmalloc_cache_index_table();
4223 create_kmalloc_caches(0);
4225 /* Setup random freelists for each cache */
4226 init_freelist_randomization();
4228 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4231 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4233 slub_min_order
, slub_max_order
, slub_min_objects
,
4234 nr_cpu_ids
, nr_node_ids
);
4237 void __init
kmem_cache_init_late(void)
4242 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
4243 unsigned long flags
, void (*ctor
)(void *))
4245 struct kmem_cache
*s
, *c
;
4247 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4252 * Adjust the object sizes so that we clear
4253 * the complete object on kzalloc.
4255 s
->object_size
= max(s
->object_size
, (int)size
);
4256 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4258 for_each_memcg_cache(c
, s
) {
4259 c
->object_size
= s
->object_size
;
4260 c
->inuse
= max_t(int, c
->inuse
,
4261 ALIGN(size
, sizeof(void *)));
4264 if (sysfs_slab_alias(s
, name
)) {
4273 int __kmem_cache_create(struct kmem_cache
*s
, unsigned long flags
)
4277 err
= kmem_cache_open(s
, flags
);
4281 /* Mutex is not taken during early boot */
4282 if (slab_state
<= UP
)
4285 memcg_propagate_slab_attrs(s
);
4286 err
= sysfs_slab_add(s
);
4288 __kmem_cache_release(s
);
4293 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4295 struct kmem_cache
*s
;
4298 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4299 return kmalloc_large(size
, gfpflags
);
4301 s
= kmalloc_slab(size
, gfpflags
);
4303 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4306 ret
= slab_alloc(s
, gfpflags
, caller
);
4308 /* Honor the call site pointer we received. */
4309 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4315 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4316 int node
, unsigned long caller
)
4318 struct kmem_cache
*s
;
4321 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4322 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4324 trace_kmalloc_node(caller
, ret
,
4325 size
, PAGE_SIZE
<< get_order(size
),
4331 s
= kmalloc_slab(size
, gfpflags
);
4333 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4336 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4338 /* Honor the call site pointer we received. */
4339 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4346 static int count_inuse(struct page
*page
)
4351 static int count_total(struct page
*page
)
4353 return page
->objects
;
4357 #ifdef CONFIG_SLUB_DEBUG
4358 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4362 void *addr
= page_address(page
);
4364 if (!check_slab(s
, page
) ||
4365 !on_freelist(s
, page
, NULL
))
4368 /* Now we know that a valid freelist exists */
4369 bitmap_zero(map
, page
->objects
);
4371 get_map(s
, page
, map
);
4372 for_each_object(p
, s
, addr
, page
->objects
) {
4373 if (test_bit(slab_index(p
, s
, addr
), map
))
4374 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4378 for_each_object(p
, s
, addr
, page
->objects
)
4379 if (!test_bit(slab_index(p
, s
, addr
), map
))
4380 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4385 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4389 validate_slab(s
, page
, map
);
4393 static int validate_slab_node(struct kmem_cache
*s
,
4394 struct kmem_cache_node
*n
, unsigned long *map
)
4396 unsigned long count
= 0;
4398 unsigned long flags
;
4400 spin_lock_irqsave(&n
->list_lock
, flags
);
4402 list_for_each_entry(page
, &n
->partial
, lru
) {
4403 validate_slab_slab(s
, page
, map
);
4406 if (count
!= n
->nr_partial
)
4407 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4408 s
->name
, count
, n
->nr_partial
);
4410 if (!(s
->flags
& SLAB_STORE_USER
))
4413 list_for_each_entry(page
, &n
->full
, lru
) {
4414 validate_slab_slab(s
, page
, map
);
4417 if (count
!= atomic_long_read(&n
->nr_slabs
))
4418 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4419 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4422 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4426 static long validate_slab_cache(struct kmem_cache
*s
)
4429 unsigned long count
= 0;
4430 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4431 sizeof(unsigned long), GFP_KERNEL
);
4432 struct kmem_cache_node
*n
;
4438 for_each_kmem_cache_node(s
, node
, n
)
4439 count
+= validate_slab_node(s
, n
, map
);
4444 * Generate lists of code addresses where slabcache objects are allocated
4449 unsigned long count
;
4456 DECLARE_BITMAP(cpus
, NR_CPUS
);
4462 unsigned long count
;
4463 struct location
*loc
;
4466 static void free_loc_track(struct loc_track
*t
)
4469 free_pages((unsigned long)t
->loc
,
4470 get_order(sizeof(struct location
) * t
->max
));
4473 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4478 order
= get_order(sizeof(struct location
) * max
);
4480 l
= (void *)__get_free_pages(flags
, order
);
4485 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4493 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4494 const struct track
*track
)
4496 long start
, end
, pos
;
4498 unsigned long caddr
;
4499 unsigned long age
= jiffies
- track
->when
;
4505 pos
= start
+ (end
- start
+ 1) / 2;
4508 * There is nothing at "end". If we end up there
4509 * we need to add something to before end.
4514 caddr
= t
->loc
[pos
].addr
;
4515 if (track
->addr
== caddr
) {
4521 if (age
< l
->min_time
)
4523 if (age
> l
->max_time
)
4526 if (track
->pid
< l
->min_pid
)
4527 l
->min_pid
= track
->pid
;
4528 if (track
->pid
> l
->max_pid
)
4529 l
->max_pid
= track
->pid
;
4531 cpumask_set_cpu(track
->cpu
,
4532 to_cpumask(l
->cpus
));
4534 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4538 if (track
->addr
< caddr
)
4545 * Not found. Insert new tracking element.
4547 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4553 (t
->count
- pos
) * sizeof(struct location
));
4556 l
->addr
= track
->addr
;
4560 l
->min_pid
= track
->pid
;
4561 l
->max_pid
= track
->pid
;
4562 cpumask_clear(to_cpumask(l
->cpus
));
4563 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4564 nodes_clear(l
->nodes
);
4565 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4569 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4570 struct page
*page
, enum track_item alloc
,
4573 void *addr
= page_address(page
);
4576 bitmap_zero(map
, page
->objects
);
4577 get_map(s
, page
, map
);
4579 for_each_object(p
, s
, addr
, page
->objects
)
4580 if (!test_bit(slab_index(p
, s
, addr
), map
))
4581 add_location(t
, s
, get_track(s
, p
, alloc
));
4584 static int list_locations(struct kmem_cache
*s
, char *buf
,
4585 enum track_item alloc
)
4589 struct loc_track t
= { 0, 0, NULL
};
4591 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4592 sizeof(unsigned long), GFP_KERNEL
);
4593 struct kmem_cache_node
*n
;
4595 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4598 return sprintf(buf
, "Out of memory\n");
4600 /* Push back cpu slabs */
4603 for_each_kmem_cache_node(s
, node
, n
) {
4604 unsigned long flags
;
4607 if (!atomic_long_read(&n
->nr_slabs
))
4610 spin_lock_irqsave(&n
->list_lock
, flags
);
4611 list_for_each_entry(page
, &n
->partial
, lru
)
4612 process_slab(&t
, s
, page
, alloc
, map
);
4613 list_for_each_entry(page
, &n
->full
, lru
)
4614 process_slab(&t
, s
, page
, alloc
, map
);
4615 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4618 for (i
= 0; i
< t
.count
; i
++) {
4619 struct location
*l
= &t
.loc
[i
];
4621 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4623 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4626 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4628 len
+= sprintf(buf
+ len
, "<not-available>");
4630 if (l
->sum_time
!= l
->min_time
) {
4631 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4633 (long)div_u64(l
->sum_time
, l
->count
),
4636 len
+= sprintf(buf
+ len
, " age=%ld",
4639 if (l
->min_pid
!= l
->max_pid
)
4640 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4641 l
->min_pid
, l
->max_pid
);
4643 len
+= sprintf(buf
+ len
, " pid=%ld",
4646 if (num_online_cpus() > 1 &&
4647 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4648 len
< PAGE_SIZE
- 60)
4649 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4651 cpumask_pr_args(to_cpumask(l
->cpus
)));
4653 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4654 len
< PAGE_SIZE
- 60)
4655 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4657 nodemask_pr_args(&l
->nodes
));
4659 len
+= sprintf(buf
+ len
, "\n");
4665 len
+= sprintf(buf
, "No data\n");
4670 #ifdef SLUB_RESILIENCY_TEST
4671 static void __init
resiliency_test(void)
4675 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4677 pr_err("SLUB resiliency testing\n");
4678 pr_err("-----------------------\n");
4679 pr_err("A. Corruption after allocation\n");
4681 p
= kzalloc(16, GFP_KERNEL
);
4683 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4686 validate_slab_cache(kmalloc_caches
[4]);
4688 /* Hmmm... The next two are dangerous */
4689 p
= kzalloc(32, GFP_KERNEL
);
4690 p
[32 + sizeof(void *)] = 0x34;
4691 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4693 pr_err("If allocated object is overwritten then not detectable\n\n");
4695 validate_slab_cache(kmalloc_caches
[5]);
4696 p
= kzalloc(64, GFP_KERNEL
);
4697 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4699 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4701 pr_err("If allocated object is overwritten then not detectable\n\n");
4702 validate_slab_cache(kmalloc_caches
[6]);
4704 pr_err("\nB. Corruption after free\n");
4705 p
= kzalloc(128, GFP_KERNEL
);
4708 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4709 validate_slab_cache(kmalloc_caches
[7]);
4711 p
= kzalloc(256, GFP_KERNEL
);
4714 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4715 validate_slab_cache(kmalloc_caches
[8]);
4717 p
= kzalloc(512, GFP_KERNEL
);
4720 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4721 validate_slab_cache(kmalloc_caches
[9]);
4725 static void resiliency_test(void) {};
4730 enum slab_stat_type
{
4731 SL_ALL
, /* All slabs */
4732 SL_PARTIAL
, /* Only partially allocated slabs */
4733 SL_CPU
, /* Only slabs used for cpu caches */
4734 SL_OBJECTS
, /* Determine allocated objects not slabs */
4735 SL_TOTAL
/* Determine object capacity not slabs */
4738 #define SO_ALL (1 << SL_ALL)
4739 #define SO_PARTIAL (1 << SL_PARTIAL)
4740 #define SO_CPU (1 << SL_CPU)
4741 #define SO_OBJECTS (1 << SL_OBJECTS)
4742 #define SO_TOTAL (1 << SL_TOTAL)
4745 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4747 static int __init
setup_slub_memcg_sysfs(char *str
)
4751 if (get_option(&str
, &v
) > 0)
4752 memcg_sysfs_enabled
= v
;
4757 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4760 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4761 char *buf
, unsigned long flags
)
4763 unsigned long total
= 0;
4766 unsigned long *nodes
;
4768 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4772 if (flags
& SO_CPU
) {
4775 for_each_possible_cpu(cpu
) {
4776 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4781 page
= READ_ONCE(c
->page
);
4785 node
= page_to_nid(page
);
4786 if (flags
& SO_TOTAL
)
4788 else if (flags
& SO_OBJECTS
)
4796 page
= slub_percpu_partial_read_once(c
);
4798 node
= page_to_nid(page
);
4799 if (flags
& SO_TOTAL
)
4801 else if (flags
& SO_OBJECTS
)
4812 #ifdef CONFIG_SLUB_DEBUG
4813 if (flags
& SO_ALL
) {
4814 struct kmem_cache_node
*n
;
4816 for_each_kmem_cache_node(s
, node
, n
) {
4818 if (flags
& SO_TOTAL
)
4819 x
= atomic_long_read(&n
->total_objects
);
4820 else if (flags
& SO_OBJECTS
)
4821 x
= atomic_long_read(&n
->total_objects
) -
4822 count_partial(n
, count_free
);
4824 x
= atomic_long_read(&n
->nr_slabs
);
4831 if (flags
& SO_PARTIAL
) {
4832 struct kmem_cache_node
*n
;
4834 for_each_kmem_cache_node(s
, node
, n
) {
4835 if (flags
& SO_TOTAL
)
4836 x
= count_partial(n
, count_total
);
4837 else if (flags
& SO_OBJECTS
)
4838 x
= count_partial(n
, count_inuse
);
4845 x
= sprintf(buf
, "%lu", total
);
4847 for (node
= 0; node
< nr_node_ids
; node
++)
4849 x
+= sprintf(buf
+ x
, " N%d=%lu",
4854 return x
+ sprintf(buf
+ x
, "\n");
4857 #ifdef CONFIG_SLUB_DEBUG
4858 static int any_slab_objects(struct kmem_cache
*s
)
4861 struct kmem_cache_node
*n
;
4863 for_each_kmem_cache_node(s
, node
, n
)
4864 if (atomic_long_read(&n
->total_objects
))
4871 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4872 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4874 struct slab_attribute
{
4875 struct attribute attr
;
4876 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4877 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4880 #define SLAB_ATTR_RO(_name) \
4881 static struct slab_attribute _name##_attr = \
4882 __ATTR(_name, 0400, _name##_show, NULL)
4884 #define SLAB_ATTR(_name) \
4885 static struct slab_attribute _name##_attr = \
4886 __ATTR(_name, 0600, _name##_show, _name##_store)
4888 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4890 return sprintf(buf
, "%d\n", s
->size
);
4892 SLAB_ATTR_RO(slab_size
);
4894 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4896 return sprintf(buf
, "%d\n", s
->align
);
4898 SLAB_ATTR_RO(align
);
4900 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4902 return sprintf(buf
, "%d\n", s
->object_size
);
4904 SLAB_ATTR_RO(object_size
);
4906 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4908 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4910 SLAB_ATTR_RO(objs_per_slab
);
4912 static ssize_t
order_store(struct kmem_cache
*s
,
4913 const char *buf
, size_t length
)
4915 unsigned long order
;
4918 err
= kstrtoul(buf
, 10, &order
);
4922 if (order
> slub_max_order
|| order
< slub_min_order
)
4925 calculate_sizes(s
, order
);
4929 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4931 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4935 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4937 return sprintf(buf
, "%lu\n", s
->min_partial
);
4940 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4946 err
= kstrtoul(buf
, 10, &min
);
4950 set_min_partial(s
, min
);
4953 SLAB_ATTR(min_partial
);
4955 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4957 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4960 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4963 unsigned long objects
;
4966 err
= kstrtoul(buf
, 10, &objects
);
4969 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4972 slub_set_cpu_partial(s
, objects
);
4976 SLAB_ATTR(cpu_partial
);
4978 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4982 return sprintf(buf
, "%pS\n", s
->ctor
);
4986 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4988 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4990 SLAB_ATTR_RO(aliases
);
4992 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4994 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4996 SLAB_ATTR_RO(partial
);
4998 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5000 return show_slab_objects(s
, buf
, SO_CPU
);
5002 SLAB_ATTR_RO(cpu_slabs
);
5004 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5006 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5008 SLAB_ATTR_RO(objects
);
5010 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5012 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5014 SLAB_ATTR_RO(objects_partial
);
5016 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5023 for_each_online_cpu(cpu
) {
5026 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5029 pages
+= page
->pages
;
5030 objects
+= page
->pobjects
;
5034 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5037 for_each_online_cpu(cpu
) {
5040 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5042 if (page
&& len
< PAGE_SIZE
- 20)
5043 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5044 page
->pobjects
, page
->pages
);
5047 return len
+ sprintf(buf
+ len
, "\n");
5049 SLAB_ATTR_RO(slabs_cpu_partial
);
5051 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5053 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5056 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5057 const char *buf
, size_t length
)
5059 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5061 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5064 SLAB_ATTR(reclaim_account
);
5066 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5068 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5070 SLAB_ATTR_RO(hwcache_align
);
5072 #ifdef CONFIG_ZONE_DMA
5073 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5075 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5077 SLAB_ATTR_RO(cache_dma
);
5080 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5082 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5084 SLAB_ATTR_RO(destroy_by_rcu
);
5086 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
5088 return sprintf(buf
, "%d\n", s
->reserved
);
5090 SLAB_ATTR_RO(reserved
);
5092 #ifdef CONFIG_SLUB_DEBUG
5093 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5095 return show_slab_objects(s
, buf
, SO_ALL
);
5097 SLAB_ATTR_RO(slabs
);
5099 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5101 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5103 SLAB_ATTR_RO(total_objects
);
5105 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5107 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5110 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5111 const char *buf
, size_t length
)
5113 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5114 if (buf
[0] == '1') {
5115 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5116 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5120 SLAB_ATTR(sanity_checks
);
5122 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5124 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5127 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5131 * Tracing a merged cache is going to give confusing results
5132 * as well as cause other issues like converting a mergeable
5133 * cache into an umergeable one.
5135 if (s
->refcount
> 1)
5138 s
->flags
&= ~SLAB_TRACE
;
5139 if (buf
[0] == '1') {
5140 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5141 s
->flags
|= SLAB_TRACE
;
5147 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5149 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5152 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5153 const char *buf
, size_t length
)
5155 if (any_slab_objects(s
))
5158 s
->flags
&= ~SLAB_RED_ZONE
;
5159 if (buf
[0] == '1') {
5160 s
->flags
|= SLAB_RED_ZONE
;
5162 calculate_sizes(s
, -1);
5165 SLAB_ATTR(red_zone
);
5167 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5169 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5172 static ssize_t
poison_store(struct kmem_cache
*s
,
5173 const char *buf
, size_t length
)
5175 if (any_slab_objects(s
))
5178 s
->flags
&= ~SLAB_POISON
;
5179 if (buf
[0] == '1') {
5180 s
->flags
|= SLAB_POISON
;
5182 calculate_sizes(s
, -1);
5187 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5189 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5192 static ssize_t
store_user_store(struct kmem_cache
*s
,
5193 const char *buf
, size_t length
)
5195 if (any_slab_objects(s
))
5198 s
->flags
&= ~SLAB_STORE_USER
;
5199 if (buf
[0] == '1') {
5200 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5201 s
->flags
|= SLAB_STORE_USER
;
5203 calculate_sizes(s
, -1);
5206 SLAB_ATTR(store_user
);
5208 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5213 static ssize_t
validate_store(struct kmem_cache
*s
,
5214 const char *buf
, size_t length
)
5218 if (buf
[0] == '1') {
5219 ret
= validate_slab_cache(s
);
5225 SLAB_ATTR(validate
);
5227 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5229 if (!(s
->flags
& SLAB_STORE_USER
))
5231 return list_locations(s
, buf
, TRACK_ALLOC
);
5233 SLAB_ATTR_RO(alloc_calls
);
5235 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5237 if (!(s
->flags
& SLAB_STORE_USER
))
5239 return list_locations(s
, buf
, TRACK_FREE
);
5241 SLAB_ATTR_RO(free_calls
);
5242 #endif /* CONFIG_SLUB_DEBUG */
5244 #ifdef CONFIG_FAILSLAB
5245 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5247 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5250 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5253 if (s
->refcount
> 1)
5256 s
->flags
&= ~SLAB_FAILSLAB
;
5258 s
->flags
|= SLAB_FAILSLAB
;
5261 SLAB_ATTR(failslab
);
5264 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5269 static ssize_t
shrink_store(struct kmem_cache
*s
,
5270 const char *buf
, size_t length
)
5273 kmem_cache_shrink(s
);
5281 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5283 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
5286 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5287 const char *buf
, size_t length
)
5289 unsigned long ratio
;
5292 err
= kstrtoul(buf
, 10, &ratio
);
5297 s
->remote_node_defrag_ratio
= ratio
* 10;
5301 SLAB_ATTR(remote_node_defrag_ratio
);
5304 #ifdef CONFIG_SLUB_STATS
5305 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5307 unsigned long sum
= 0;
5310 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5315 for_each_online_cpu(cpu
) {
5316 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5322 len
= sprintf(buf
, "%lu", sum
);
5325 for_each_online_cpu(cpu
) {
5326 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5327 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5331 return len
+ sprintf(buf
+ len
, "\n");
5334 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5338 for_each_online_cpu(cpu
)
5339 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5342 #define STAT_ATTR(si, text) \
5343 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5345 return show_stat(s, buf, si); \
5347 static ssize_t text##_store(struct kmem_cache *s, \
5348 const char *buf, size_t length) \
5350 if (buf[0] != '0') \
5352 clear_stat(s, si); \
5357 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5358 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5359 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5360 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5361 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5362 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5363 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5364 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5365 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5366 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5367 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5368 STAT_ATTR(FREE_SLAB
, free_slab
);
5369 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5370 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5371 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5372 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5373 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5374 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5375 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5376 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5377 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5378 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5379 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5380 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5381 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5382 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5385 static struct attribute
*slab_attrs
[] = {
5386 &slab_size_attr
.attr
,
5387 &object_size_attr
.attr
,
5388 &objs_per_slab_attr
.attr
,
5390 &min_partial_attr
.attr
,
5391 &cpu_partial_attr
.attr
,
5393 &objects_partial_attr
.attr
,
5395 &cpu_slabs_attr
.attr
,
5399 &hwcache_align_attr
.attr
,
5400 &reclaim_account_attr
.attr
,
5401 &destroy_by_rcu_attr
.attr
,
5403 &reserved_attr
.attr
,
5404 &slabs_cpu_partial_attr
.attr
,
5405 #ifdef CONFIG_SLUB_DEBUG
5406 &total_objects_attr
.attr
,
5408 &sanity_checks_attr
.attr
,
5410 &red_zone_attr
.attr
,
5412 &store_user_attr
.attr
,
5413 &validate_attr
.attr
,
5414 &alloc_calls_attr
.attr
,
5415 &free_calls_attr
.attr
,
5417 #ifdef CONFIG_ZONE_DMA
5418 &cache_dma_attr
.attr
,
5421 &remote_node_defrag_ratio_attr
.attr
,
5423 #ifdef CONFIG_SLUB_STATS
5424 &alloc_fastpath_attr
.attr
,
5425 &alloc_slowpath_attr
.attr
,
5426 &free_fastpath_attr
.attr
,
5427 &free_slowpath_attr
.attr
,
5428 &free_frozen_attr
.attr
,
5429 &free_add_partial_attr
.attr
,
5430 &free_remove_partial_attr
.attr
,
5431 &alloc_from_partial_attr
.attr
,
5432 &alloc_slab_attr
.attr
,
5433 &alloc_refill_attr
.attr
,
5434 &alloc_node_mismatch_attr
.attr
,
5435 &free_slab_attr
.attr
,
5436 &cpuslab_flush_attr
.attr
,
5437 &deactivate_full_attr
.attr
,
5438 &deactivate_empty_attr
.attr
,
5439 &deactivate_to_head_attr
.attr
,
5440 &deactivate_to_tail_attr
.attr
,
5441 &deactivate_remote_frees_attr
.attr
,
5442 &deactivate_bypass_attr
.attr
,
5443 &order_fallback_attr
.attr
,
5444 &cmpxchg_double_fail_attr
.attr
,
5445 &cmpxchg_double_cpu_fail_attr
.attr
,
5446 &cpu_partial_alloc_attr
.attr
,
5447 &cpu_partial_free_attr
.attr
,
5448 &cpu_partial_node_attr
.attr
,
5449 &cpu_partial_drain_attr
.attr
,
5451 #ifdef CONFIG_FAILSLAB
5452 &failslab_attr
.attr
,
5458 static struct attribute_group slab_attr_group
= {
5459 .attrs
= slab_attrs
,
5462 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5463 struct attribute
*attr
,
5466 struct slab_attribute
*attribute
;
5467 struct kmem_cache
*s
;
5470 attribute
= to_slab_attr(attr
);
5473 if (!attribute
->show
)
5476 err
= attribute
->show(s
, buf
);
5481 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5482 struct attribute
*attr
,
5483 const char *buf
, size_t len
)
5485 struct slab_attribute
*attribute
;
5486 struct kmem_cache
*s
;
5489 attribute
= to_slab_attr(attr
);
5492 if (!attribute
->store
)
5495 err
= attribute
->store(s
, buf
, len
);
5497 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5498 struct kmem_cache
*c
;
5500 mutex_lock(&slab_mutex
);
5501 if (s
->max_attr_size
< len
)
5502 s
->max_attr_size
= len
;
5505 * This is a best effort propagation, so this function's return
5506 * value will be determined by the parent cache only. This is
5507 * basically because not all attributes will have a well
5508 * defined semantics for rollbacks - most of the actions will
5509 * have permanent effects.
5511 * Returning the error value of any of the children that fail
5512 * is not 100 % defined, in the sense that users seeing the
5513 * error code won't be able to know anything about the state of
5516 * Only returning the error code for the parent cache at least
5517 * has well defined semantics. The cache being written to
5518 * directly either failed or succeeded, in which case we loop
5519 * through the descendants with best-effort propagation.
5521 for_each_memcg_cache(c
, s
)
5522 attribute
->store(c
, buf
, len
);
5523 mutex_unlock(&slab_mutex
);
5529 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5533 char *buffer
= NULL
;
5534 struct kmem_cache
*root_cache
;
5536 if (is_root_cache(s
))
5539 root_cache
= s
->memcg_params
.root_cache
;
5542 * This mean this cache had no attribute written. Therefore, no point
5543 * in copying default values around
5545 if (!root_cache
->max_attr_size
)
5548 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5551 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5554 if (!attr
|| !attr
->store
|| !attr
->show
)
5558 * It is really bad that we have to allocate here, so we will
5559 * do it only as a fallback. If we actually allocate, though,
5560 * we can just use the allocated buffer until the end.
5562 * Most of the slub attributes will tend to be very small in
5563 * size, but sysfs allows buffers up to a page, so they can
5564 * theoretically happen.
5568 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5571 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5572 if (WARN_ON(!buffer
))
5577 len
= attr
->show(root_cache
, buf
);
5579 attr
->store(s
, buf
, len
);
5583 free_page((unsigned long)buffer
);
5587 static void kmem_cache_release(struct kobject
*k
)
5589 slab_kmem_cache_release(to_slab(k
));
5592 static const struct sysfs_ops slab_sysfs_ops
= {
5593 .show
= slab_attr_show
,
5594 .store
= slab_attr_store
,
5597 static struct kobj_type slab_ktype
= {
5598 .sysfs_ops
= &slab_sysfs_ops
,
5599 .release
= kmem_cache_release
,
5602 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5604 struct kobj_type
*ktype
= get_ktype(kobj
);
5606 if (ktype
== &slab_ktype
)
5611 static const struct kset_uevent_ops slab_uevent_ops
= {
5612 .filter
= uevent_filter
,
5615 static struct kset
*slab_kset
;
5617 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5620 if (!is_root_cache(s
))
5621 return s
->memcg_params
.root_cache
->memcg_kset
;
5626 #define ID_STR_LENGTH 64
5628 /* Create a unique string id for a slab cache:
5630 * Format :[flags-]size
5632 static char *create_unique_id(struct kmem_cache
*s
)
5634 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5641 * First flags affecting slabcache operations. We will only
5642 * get here for aliasable slabs so we do not need to support
5643 * too many flags. The flags here must cover all flags that
5644 * are matched during merging to guarantee that the id is
5647 if (s
->flags
& SLAB_CACHE_DMA
)
5649 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5651 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5653 if (!(s
->flags
& SLAB_NOTRACK
))
5655 if (s
->flags
& SLAB_ACCOUNT
)
5659 p
+= sprintf(p
, "%07d", s
->size
);
5661 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5665 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5667 struct kmem_cache
*s
=
5668 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5670 if (!s
->kobj
.state_in_sysfs
)
5672 * For a memcg cache, this may be called during
5673 * deactivation and again on shutdown. Remove only once.
5674 * A cache is never shut down before deactivation is
5675 * complete, so no need to worry about synchronization.
5680 kset_unregister(s
->memcg_kset
);
5682 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5683 kobject_del(&s
->kobj
);
5685 kobject_put(&s
->kobj
);
5688 static int sysfs_slab_add(struct kmem_cache
*s
)
5692 struct kset
*kset
= cache_kset(s
);
5693 int unmergeable
= slab_unmergeable(s
);
5695 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5698 kobject_init(&s
->kobj
, &slab_ktype
);
5704 * Slabcache can never be merged so we can use the name proper.
5705 * This is typically the case for debug situations. In that
5706 * case we can catch duplicate names easily.
5708 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5712 * Create a unique name for the slab as a target
5715 name
= create_unique_id(s
);
5718 s
->kobj
.kset
= kset
;
5719 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5723 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5728 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5729 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5730 if (!s
->memcg_kset
) {
5737 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5739 /* Setup first alias */
5740 sysfs_slab_alias(s
, s
->name
);
5747 kobject_del(&s
->kobj
);
5751 static void sysfs_slab_remove(struct kmem_cache
*s
)
5753 if (slab_state
< FULL
)
5755 * Sysfs has not been setup yet so no need to remove the
5760 kobject_get(&s
->kobj
);
5761 schedule_work(&s
->kobj_remove_work
);
5764 void sysfs_slab_release(struct kmem_cache
*s
)
5766 if (slab_state
>= FULL
)
5767 kobject_put(&s
->kobj
);
5771 * Need to buffer aliases during bootup until sysfs becomes
5772 * available lest we lose that information.
5774 struct saved_alias
{
5775 struct kmem_cache
*s
;
5777 struct saved_alias
*next
;
5780 static struct saved_alias
*alias_list
;
5782 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5784 struct saved_alias
*al
;
5786 if (slab_state
== FULL
) {
5788 * If we have a leftover link then remove it.
5790 sysfs_remove_link(&slab_kset
->kobj
, name
);
5791 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5794 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5800 al
->next
= alias_list
;
5805 static int __init
slab_sysfs_init(void)
5807 struct kmem_cache
*s
;
5810 mutex_lock(&slab_mutex
);
5812 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5814 mutex_unlock(&slab_mutex
);
5815 pr_err("Cannot register slab subsystem.\n");
5821 list_for_each_entry(s
, &slab_caches
, list
) {
5822 err
= sysfs_slab_add(s
);
5824 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5828 while (alias_list
) {
5829 struct saved_alias
*al
= alias_list
;
5831 alias_list
= alias_list
->next
;
5832 err
= sysfs_slab_alias(al
->s
, al
->name
);
5834 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5839 mutex_unlock(&slab_mutex
);
5844 __initcall(slab_sysfs_init
);
5845 #endif /* CONFIG_SYSFS */
5848 * The /proc/slabinfo ABI
5850 #ifdef CONFIG_SLABINFO
5851 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5853 unsigned long nr_slabs
= 0;
5854 unsigned long nr_objs
= 0;
5855 unsigned long nr_free
= 0;
5857 struct kmem_cache_node
*n
;
5859 for_each_kmem_cache_node(s
, node
, n
) {
5860 nr_slabs
+= node_nr_slabs(n
);
5861 nr_objs
+= node_nr_objs(n
);
5862 nr_free
+= count_partial(n
, count_free
);
5865 sinfo
->active_objs
= nr_objs
- nr_free
;
5866 sinfo
->num_objs
= nr_objs
;
5867 sinfo
->active_slabs
= nr_slabs
;
5868 sinfo
->num_slabs
= nr_slabs
;
5869 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5870 sinfo
->cache_order
= oo_order(s
->oo
);
5873 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5877 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5878 size_t count
, loff_t
*ppos
)
5882 #endif /* CONFIG_SLABINFO */