1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/notifier.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.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 */
196 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197 /* Use cmpxchg_double */
198 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201 * Tracking user of a slab.
203 #define TRACK_ADDRS_COUNT 16
205 unsigned long addr
; /* Called from address */
206 #ifdef CONFIG_STACKTRACE
207 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
209 int cpu
; /* Was running on cpu */
210 int pid
; /* Pid context */
211 unsigned long when
; /* When did the operation occur */
214 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
217 static int sysfs_slab_add(struct kmem_cache
*);
218 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
219 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
220 static void sysfs_slab_remove(struct kmem_cache
*s
);
222 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
226 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
229 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
231 #ifdef CONFIG_SLUB_STATS
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
236 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
240 /********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
249 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
250 unsigned long ptr_addr
)
252 #ifdef CONFIG_SLAB_FREELIST_HARDENED
253 return (void *)((unsigned long)ptr
^ s
->random
^ ptr_addr
);
259 /* Returns the freelist pointer recorded at location ptr_addr. */
260 static inline void *freelist_dereference(const struct kmem_cache
*s
,
263 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
264 (unsigned long)ptr_addr
);
267 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
269 return freelist_dereference(s
, object
+ s
->offset
);
272 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
275 prefetch(freelist_dereference(s
, object
+ s
->offset
));
278 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
280 unsigned long freepointer_addr
;
283 if (!debug_pagealloc_enabled())
284 return get_freepointer(s
, object
);
286 freepointer_addr
= (unsigned long)object
+ s
->offset
;
287 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
288 return freelist_ptr(s
, p
, freepointer_addr
);
291 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
293 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
295 #ifdef CONFIG_SLAB_FREELIST_HARDENED
296 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
299 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
302 /* Loop over all objects in a slab */
303 #define for_each_object(__p, __s, __addr, __objects) \
304 for (__p = fixup_red_left(__s, __addr); \
305 __p < (__addr) + (__objects) * (__s)->size; \
308 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
309 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
310 __idx <= __objects; \
311 __p += (__s)->size, __idx++)
313 /* Determine object index from a given position */
314 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
316 return (p
- addr
) / s
->size
;
319 static inline int order_objects(int order
, unsigned long size
, int reserved
)
321 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
324 static inline struct kmem_cache_order_objects
oo_make(int order
,
325 unsigned long size
, int reserved
)
327 struct kmem_cache_order_objects x
= {
328 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
334 static inline int oo_order(struct kmem_cache_order_objects x
)
336 return x
.x
>> OO_SHIFT
;
339 static inline int oo_objects(struct kmem_cache_order_objects x
)
341 return x
.x
& OO_MASK
;
345 * Per slab locking using the pagelock
347 static __always_inline
void slab_lock(struct page
*page
)
349 VM_BUG_ON_PAGE(PageTail(page
), page
);
350 bit_spin_lock(PG_locked
, &page
->flags
);
353 static __always_inline
void slab_unlock(struct page
*page
)
355 VM_BUG_ON_PAGE(PageTail(page
), page
);
356 __bit_spin_unlock(PG_locked
, &page
->flags
);
359 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
362 tmp
.counters
= counters_new
;
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_refcount. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_refcount, so
367 * be careful and only assign to the fields we need.
369 page
->frozen
= tmp
.frozen
;
370 page
->inuse
= tmp
.inuse
;
371 page
->objects
= tmp
.objects
;
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
376 void *freelist_old
, unsigned long counters_old
,
377 void *freelist_new
, unsigned long counters_new
,
380 VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s
->flags
& __CMPXCHG_DOUBLE
) {
384 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
385 freelist_old
, counters_old
,
386 freelist_new
, counters_new
))
392 if (page
->freelist
== freelist_old
&&
393 page
->counters
== counters_old
) {
394 page
->freelist
= freelist_new
;
395 set_page_slub_counters(page
, counters_new
);
403 stat(s
, CMPXCHG_DOUBLE_FAIL
);
405 #ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
412 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
413 void *freelist_old
, unsigned long counters_old
,
414 void *freelist_new
, unsigned long counters_new
,
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s
->flags
& __CMPXCHG_DOUBLE
) {
420 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
421 freelist_old
, counters_old
,
422 freelist_new
, counters_new
))
429 local_irq_save(flags
);
431 if (page
->freelist
== freelist_old
&&
432 page
->counters
== counters_old
) {
433 page
->freelist
= freelist_new
;
434 set_page_slub_counters(page
, counters_new
);
436 local_irq_restore(flags
);
440 local_irq_restore(flags
);
444 stat(s
, CMPXCHG_DOUBLE_FAIL
);
446 #ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
453 #ifdef CONFIG_SLUB_DEBUG
455 * Determine a map of object in use on a page.
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
460 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
463 void *addr
= page_address(page
);
465 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
466 set_bit(slab_index(p
, s
, addr
), map
);
469 static inline int size_from_object(struct kmem_cache
*s
)
471 if (s
->flags
& SLAB_RED_ZONE
)
472 return s
->size
- s
->red_left_pad
;
477 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
479 if (s
->flags
& SLAB_RED_ZONE
)
480 p
-= s
->red_left_pad
;
488 #if defined(CONFIG_SLUB_DEBUG_ON)
489 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
491 static slab_flags_t slub_debug
;
494 static char *slub_debug_slabs
;
495 static int disable_higher_order_debug
;
498 * slub is about to manipulate internal object metadata. This memory lies
499 * outside the range of the allocated object, so accessing it would normally
500 * be reported by kasan as a bounds error. metadata_access_enable() is used
501 * to tell kasan that these accesses are OK.
503 static inline void metadata_access_enable(void)
505 kasan_disable_current();
508 static inline void metadata_access_disable(void)
510 kasan_enable_current();
517 /* Verify that a pointer has an address that is valid within a slab page */
518 static inline int check_valid_pointer(struct kmem_cache
*s
,
519 struct page
*page
, void *object
)
526 base
= page_address(page
);
527 object
= restore_red_left(s
, object
);
528 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
529 (object
- base
) % s
->size
) {
536 static void print_section(char *level
, char *text
, u8
*addr
,
539 metadata_access_enable();
540 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
542 metadata_access_disable();
545 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
546 enum track_item alloc
)
551 p
= object
+ s
->offset
+ sizeof(void *);
553 p
= object
+ s
->inuse
;
558 static void set_track(struct kmem_cache
*s
, void *object
,
559 enum track_item alloc
, unsigned long addr
)
561 struct track
*p
= get_track(s
, object
, alloc
);
564 #ifdef CONFIG_STACKTRACE
565 struct stack_trace trace
;
568 trace
.nr_entries
= 0;
569 trace
.max_entries
= TRACK_ADDRS_COUNT
;
570 trace
.entries
= p
->addrs
;
572 metadata_access_enable();
573 save_stack_trace(&trace
);
574 metadata_access_disable();
576 /* See rant in lockdep.c */
577 if (trace
.nr_entries
!= 0 &&
578 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
581 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
585 p
->cpu
= smp_processor_id();
586 p
->pid
= current
->pid
;
589 memset(p
, 0, sizeof(struct track
));
592 static void init_tracking(struct kmem_cache
*s
, void *object
)
594 if (!(s
->flags
& SLAB_STORE_USER
))
597 set_track(s
, object
, TRACK_FREE
, 0UL);
598 set_track(s
, object
, TRACK_ALLOC
, 0UL);
601 static void print_track(const char *s
, struct track
*t
)
606 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
607 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
608 #ifdef CONFIG_STACKTRACE
611 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
613 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
620 static void print_tracking(struct kmem_cache
*s
, void *object
)
622 if (!(s
->flags
& SLAB_STORE_USER
))
625 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
626 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
629 static void print_page_info(struct page
*page
)
631 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
632 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
636 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
638 struct va_format vaf
;
644 pr_err("=============================================================================\n");
645 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
646 pr_err("-----------------------------------------------------------------------------\n\n");
648 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
652 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
654 struct va_format vaf
;
660 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
664 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
666 unsigned int off
; /* Offset of last byte */
667 u8
*addr
= page_address(page
);
669 print_tracking(s
, p
);
671 print_page_info(page
);
673 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
674 p
, p
- addr
, get_freepointer(s
, p
));
676 if (s
->flags
& SLAB_RED_ZONE
)
677 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
679 else if (p
> addr
+ 16)
680 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
682 print_section(KERN_ERR
, "Object ", p
,
683 min_t(unsigned long, s
->object_size
, PAGE_SIZE
));
684 if (s
->flags
& SLAB_RED_ZONE
)
685 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
686 s
->inuse
- s
->object_size
);
689 off
= s
->offset
+ sizeof(void *);
693 if (s
->flags
& SLAB_STORE_USER
)
694 off
+= 2 * sizeof(struct track
);
696 off
+= kasan_metadata_size(s
);
698 if (off
!= size_from_object(s
))
699 /* Beginning of the filler is the free pointer */
700 print_section(KERN_ERR
, "Padding ", p
+ off
,
701 size_from_object(s
) - off
);
706 void object_err(struct kmem_cache
*s
, struct page
*page
,
707 u8
*object
, char *reason
)
709 slab_bug(s
, "%s", reason
);
710 print_trailer(s
, page
, object
);
713 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
714 const char *fmt
, ...)
720 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
722 slab_bug(s
, "%s", buf
);
723 print_page_info(page
);
727 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
731 if (s
->flags
& SLAB_RED_ZONE
)
732 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
734 if (s
->flags
& __OBJECT_POISON
) {
735 memset(p
, POISON_FREE
, s
->object_size
- 1);
736 p
[s
->object_size
- 1] = POISON_END
;
739 if (s
->flags
& SLAB_RED_ZONE
)
740 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
743 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
744 void *from
, void *to
)
746 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
747 memset(from
, data
, to
- from
);
750 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
751 u8
*object
, char *what
,
752 u8
*start
, unsigned int value
, unsigned int bytes
)
757 metadata_access_enable();
758 fault
= memchr_inv(start
, value
, bytes
);
759 metadata_access_disable();
764 while (end
> fault
&& end
[-1] == value
)
767 slab_bug(s
, "%s overwritten", what
);
768 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
769 fault
, end
- 1, fault
[0], value
);
770 print_trailer(s
, page
, object
);
772 restore_bytes(s
, what
, value
, fault
, end
);
780 * Bytes of the object to be managed.
781 * If the freepointer may overlay the object then the free
782 * pointer is the first word of the object.
784 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
787 * object + s->object_size
788 * Padding to reach word boundary. This is also used for Redzoning.
789 * Padding is extended by another word if Redzoning is enabled and
790 * object_size == inuse.
792 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
793 * 0xcc (RED_ACTIVE) for objects in use.
796 * Meta data starts here.
798 * A. Free pointer (if we cannot overwrite object on free)
799 * B. Tracking data for SLAB_STORE_USER
800 * C. Padding to reach required alignment boundary or at mininum
801 * one word if debugging is on to be able to detect writes
802 * before the word boundary.
804 * Padding is done using 0x5a (POISON_INUSE)
807 * Nothing is used beyond s->size.
809 * If slabcaches are merged then the object_size and inuse boundaries are mostly
810 * ignored. And therefore no slab options that rely on these boundaries
811 * may be used with merged slabcaches.
814 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
816 unsigned long off
= s
->inuse
; /* The end of info */
819 /* Freepointer is placed after the object. */
820 off
+= sizeof(void *);
822 if (s
->flags
& SLAB_STORE_USER
)
823 /* We also have user information there */
824 off
+= 2 * sizeof(struct track
);
826 off
+= kasan_metadata_size(s
);
828 if (size_from_object(s
) == off
)
831 return check_bytes_and_report(s
, page
, p
, "Object padding",
832 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
835 /* Check the pad bytes at the end of a slab page */
836 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
844 if (!(s
->flags
& SLAB_POISON
))
847 start
= page_address(page
);
848 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
849 end
= start
+ length
;
850 remainder
= length
% s
->size
;
854 metadata_access_enable();
855 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
856 metadata_access_disable();
859 while (end
> fault
&& end
[-1] == POISON_INUSE
)
862 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
863 print_section(KERN_ERR
, "Padding ", end
- remainder
, remainder
);
865 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
869 static int check_object(struct kmem_cache
*s
, struct page
*page
,
870 void *object
, u8 val
)
873 u8
*endobject
= object
+ s
->object_size
;
875 if (s
->flags
& SLAB_RED_ZONE
) {
876 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
877 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
880 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
881 endobject
, val
, s
->inuse
- s
->object_size
))
884 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
885 check_bytes_and_report(s
, page
, p
, "Alignment padding",
886 endobject
, POISON_INUSE
,
887 s
->inuse
- s
->object_size
);
891 if (s
->flags
& SLAB_POISON
) {
892 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
893 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
894 POISON_FREE
, s
->object_size
- 1) ||
895 !check_bytes_and_report(s
, page
, p
, "Poison",
896 p
+ s
->object_size
- 1, POISON_END
, 1)))
899 * check_pad_bytes cleans up on its own.
901 check_pad_bytes(s
, page
, p
);
904 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
906 * Object and freepointer overlap. Cannot check
907 * freepointer while object is allocated.
911 /* Check free pointer validity */
912 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
913 object_err(s
, page
, p
, "Freepointer corrupt");
915 * No choice but to zap it and thus lose the remainder
916 * of the free objects in this slab. May cause
917 * another error because the object count is now wrong.
919 set_freepointer(s
, p
, NULL
);
925 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
929 VM_BUG_ON(!irqs_disabled());
931 if (!PageSlab(page
)) {
932 slab_err(s
, page
, "Not a valid slab page");
936 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
937 if (page
->objects
> maxobj
) {
938 slab_err(s
, page
, "objects %u > max %u",
939 page
->objects
, maxobj
);
942 if (page
->inuse
> page
->objects
) {
943 slab_err(s
, page
, "inuse %u > max %u",
944 page
->inuse
, page
->objects
);
947 /* Slab_pad_check fixes things up after itself */
948 slab_pad_check(s
, page
);
953 * Determine if a certain object on a page is on the freelist. Must hold the
954 * slab lock to guarantee that the chains are in a consistent state.
956 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
964 while (fp
&& nr
<= page
->objects
) {
967 if (!check_valid_pointer(s
, page
, fp
)) {
969 object_err(s
, page
, object
,
970 "Freechain corrupt");
971 set_freepointer(s
, object
, NULL
);
973 slab_err(s
, page
, "Freepointer corrupt");
974 page
->freelist
= NULL
;
975 page
->inuse
= page
->objects
;
976 slab_fix(s
, "Freelist cleared");
982 fp
= get_freepointer(s
, object
);
986 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
987 if (max_objects
> MAX_OBJS_PER_PAGE
)
988 max_objects
= MAX_OBJS_PER_PAGE
;
990 if (page
->objects
!= max_objects
) {
991 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
992 page
->objects
, max_objects
);
993 page
->objects
= max_objects
;
994 slab_fix(s
, "Number of objects adjusted.");
996 if (page
->inuse
!= page
->objects
- nr
) {
997 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
998 page
->inuse
, page
->objects
- nr
);
999 page
->inuse
= page
->objects
- nr
;
1000 slab_fix(s
, "Object count adjusted.");
1002 return search
== NULL
;
1005 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1008 if (s
->flags
& SLAB_TRACE
) {
1009 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1011 alloc
? "alloc" : "free",
1012 object
, page
->inuse
,
1016 print_section(KERN_INFO
, "Object ", (void *)object
,
1024 * Tracking of fully allocated slabs for debugging purposes.
1026 static void add_full(struct kmem_cache
*s
,
1027 struct kmem_cache_node
*n
, struct page
*page
)
1029 if (!(s
->flags
& SLAB_STORE_USER
))
1032 lockdep_assert_held(&n
->list_lock
);
1033 list_add(&page
->lru
, &n
->full
);
1036 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1038 if (!(s
->flags
& SLAB_STORE_USER
))
1041 lockdep_assert_held(&n
->list_lock
);
1042 list_del(&page
->lru
);
1045 /* Tracking of the number of slabs for debugging purposes */
1046 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1048 struct kmem_cache_node
*n
= get_node(s
, node
);
1050 return atomic_long_read(&n
->nr_slabs
);
1053 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1055 return atomic_long_read(&n
->nr_slabs
);
1058 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1060 struct kmem_cache_node
*n
= get_node(s
, node
);
1063 * May be called early in order to allocate a slab for the
1064 * kmem_cache_node structure. Solve the chicken-egg
1065 * dilemma by deferring the increment of the count during
1066 * bootstrap (see early_kmem_cache_node_alloc).
1069 atomic_long_inc(&n
->nr_slabs
);
1070 atomic_long_add(objects
, &n
->total_objects
);
1073 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1075 struct kmem_cache_node
*n
= get_node(s
, node
);
1077 atomic_long_dec(&n
->nr_slabs
);
1078 atomic_long_sub(objects
, &n
->total_objects
);
1081 /* Object debug checks for alloc/free paths */
1082 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1085 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1088 init_object(s
, object
, SLUB_RED_INACTIVE
);
1089 init_tracking(s
, object
);
1092 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1094 void *object
, unsigned long addr
)
1096 if (!check_slab(s
, page
))
1099 if (!check_valid_pointer(s
, page
, object
)) {
1100 object_err(s
, page
, object
, "Freelist Pointer check fails");
1104 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1110 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1112 void *object
, unsigned long addr
)
1114 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1115 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1119 /* Success perform special debug activities for allocs */
1120 if (s
->flags
& SLAB_STORE_USER
)
1121 set_track(s
, object
, TRACK_ALLOC
, addr
);
1122 trace(s
, page
, object
, 1);
1123 init_object(s
, object
, SLUB_RED_ACTIVE
);
1127 if (PageSlab(page
)) {
1129 * If this is a slab page then lets do the best we can
1130 * to avoid issues in the future. Marking all objects
1131 * as used avoids touching the remaining objects.
1133 slab_fix(s
, "Marking all objects used");
1134 page
->inuse
= page
->objects
;
1135 page
->freelist
= NULL
;
1140 static inline int free_consistency_checks(struct kmem_cache
*s
,
1141 struct page
*page
, void *object
, unsigned long addr
)
1143 if (!check_valid_pointer(s
, page
, object
)) {
1144 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1148 if (on_freelist(s
, page
, object
)) {
1149 object_err(s
, page
, object
, "Object already free");
1153 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1156 if (unlikely(s
!= page
->slab_cache
)) {
1157 if (!PageSlab(page
)) {
1158 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1160 } else if (!page
->slab_cache
) {
1161 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1165 object_err(s
, page
, object
,
1166 "page slab pointer corrupt.");
1172 /* Supports checking bulk free of a constructed freelist */
1173 static noinline
int free_debug_processing(
1174 struct kmem_cache
*s
, struct page
*page
,
1175 void *head
, void *tail
, int bulk_cnt
,
1178 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1179 void *object
= head
;
1181 unsigned long uninitialized_var(flags
);
1184 spin_lock_irqsave(&n
->list_lock
, flags
);
1187 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1188 if (!check_slab(s
, page
))
1195 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1196 if (!free_consistency_checks(s
, page
, object
, addr
))
1200 if (s
->flags
& SLAB_STORE_USER
)
1201 set_track(s
, object
, TRACK_FREE
, addr
);
1202 trace(s
, page
, object
, 0);
1203 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1204 init_object(s
, object
, SLUB_RED_INACTIVE
);
1206 /* Reached end of constructed freelist yet? */
1207 if (object
!= tail
) {
1208 object
= get_freepointer(s
, object
);
1214 if (cnt
!= bulk_cnt
)
1215 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1219 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1221 slab_fix(s
, "Object at 0x%p not freed", object
);
1225 static int __init
setup_slub_debug(char *str
)
1227 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1228 if (*str
++ != '=' || !*str
)
1230 * No options specified. Switch on full debugging.
1236 * No options but restriction on slabs. This means full
1237 * debugging for slabs matching a pattern.
1244 * Switch off all debugging measures.
1249 * Determine which debug features should be switched on
1251 for (; *str
&& *str
!= ','; str
++) {
1252 switch (tolower(*str
)) {
1254 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1257 slub_debug
|= SLAB_RED_ZONE
;
1260 slub_debug
|= SLAB_POISON
;
1263 slub_debug
|= SLAB_STORE_USER
;
1266 slub_debug
|= SLAB_TRACE
;
1269 slub_debug
|= SLAB_FAILSLAB
;
1273 * Avoid enabling debugging on caches if its minimum
1274 * order would increase as a result.
1276 disable_higher_order_debug
= 1;
1279 pr_err("slub_debug option '%c' unknown. skipped\n",
1286 slub_debug_slabs
= str
+ 1;
1291 __setup("slub_debug", setup_slub_debug
);
1293 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1294 slab_flags_t flags
, const char *name
,
1295 void (*ctor
)(void *))
1298 * Enable debugging if selected on the kernel commandline.
1300 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1301 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1302 flags
|= slub_debug
;
1306 #else /* !CONFIG_SLUB_DEBUG */
1307 static inline void setup_object_debug(struct kmem_cache
*s
,
1308 struct page
*page
, void *object
) {}
1310 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1311 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1313 static inline int free_debug_processing(
1314 struct kmem_cache
*s
, struct page
*page
,
1315 void *head
, void *tail
, int bulk_cnt
,
1316 unsigned long addr
) { return 0; }
1318 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1320 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1321 void *object
, u8 val
) { return 1; }
1322 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1323 struct page
*page
) {}
1324 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1325 struct page
*page
) {}
1326 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1327 slab_flags_t flags
, const char *name
,
1328 void (*ctor
)(void *))
1332 #define slub_debug 0
1334 #define disable_higher_order_debug 0
1336 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1338 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1340 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1342 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1345 #endif /* CONFIG_SLUB_DEBUG */
1348 * Hooks for other subsystems that check memory allocations. In a typical
1349 * production configuration these hooks all should produce no code at all.
1351 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1353 kmemleak_alloc(ptr
, size
, 1, flags
);
1354 kasan_kmalloc_large(ptr
, size
, flags
);
1357 static inline void kfree_hook(const void *x
)
1360 kasan_kfree_large(x
);
1363 static inline bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1365 kmemleak_free_recursive(x
, s
->flags
);
1368 * Trouble is that we may no longer disable interrupts in the fast path
1369 * So in order to make the debug calls that expect irqs to be
1370 * disabled we need to disable interrupts temporarily.
1372 #ifdef CONFIG_LOCKDEP
1374 unsigned long flags
;
1376 local_irq_save(flags
);
1377 debug_check_no_locks_freed(x
, s
->object_size
);
1378 local_irq_restore(flags
);
1381 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1382 debug_check_no_obj_freed(x
, s
->object_size
);
1384 /* KASAN might put x into memory quarantine, delaying its reuse */
1385 return kasan_slab_free(s
, x
);
1388 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1389 void **head
, void **tail
)
1392 * Compiler cannot detect this function can be removed if slab_free_hook()
1393 * evaluates to nothing. Thus, catch all relevant config debug options here.
1395 #if defined(CONFIG_LOCKDEP) || \
1396 defined(CONFIG_DEBUG_KMEMLEAK) || \
1397 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1398 defined(CONFIG_KASAN)
1402 void *old_tail
= *tail
? *tail
: *head
;
1404 /* Head and tail of the reconstructed freelist */
1410 next
= get_freepointer(s
, object
);
1411 /* If object's reuse doesn't have to be delayed */
1412 if (!slab_free_hook(s
, object
)) {
1413 /* Move object to the new freelist */
1414 set_freepointer(s
, object
, *head
);
1419 } while (object
!= old_tail
);
1424 return *head
!= NULL
;
1430 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1433 setup_object_debug(s
, page
, object
);
1434 kasan_init_slab_obj(s
, object
);
1435 if (unlikely(s
->ctor
)) {
1436 kasan_unpoison_object_data(s
, object
);
1438 kasan_poison_object_data(s
, object
);
1443 * Slab allocation and freeing
1445 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1446 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1449 int order
= oo_order(oo
);
1451 if (node
== NUMA_NO_NODE
)
1452 page
= alloc_pages(flags
, order
);
1454 page
= __alloc_pages_node(node
, flags
, order
);
1456 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1457 __free_pages(page
, order
);
1464 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1465 /* Pre-initialize the random sequence cache */
1466 static int init_cache_random_seq(struct kmem_cache
*s
)
1469 unsigned long i
, count
= oo_objects(s
->oo
);
1471 /* Bailout if already initialised */
1475 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1477 pr_err("SLUB: Unable to initialize free list for %s\n",
1482 /* Transform to an offset on the set of pages */
1483 if (s
->random_seq
) {
1484 for (i
= 0; i
< count
; i
++)
1485 s
->random_seq
[i
] *= s
->size
;
1490 /* Initialize each random sequence freelist per cache */
1491 static void __init
init_freelist_randomization(void)
1493 struct kmem_cache
*s
;
1495 mutex_lock(&slab_mutex
);
1497 list_for_each_entry(s
, &slab_caches
, list
)
1498 init_cache_random_seq(s
);
1500 mutex_unlock(&slab_mutex
);
1503 /* Get the next entry on the pre-computed freelist randomized */
1504 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1505 unsigned long *pos
, void *start
,
1506 unsigned long page_limit
,
1507 unsigned long freelist_count
)
1512 * If the target page allocation failed, the number of objects on the
1513 * page might be smaller than the usual size defined by the cache.
1516 idx
= s
->random_seq
[*pos
];
1518 if (*pos
>= freelist_count
)
1520 } while (unlikely(idx
>= page_limit
));
1522 return (char *)start
+ idx
;
1525 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1526 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1531 unsigned long idx
, pos
, page_limit
, freelist_count
;
1533 if (page
->objects
< 2 || !s
->random_seq
)
1536 freelist_count
= oo_objects(s
->oo
);
1537 pos
= get_random_int() % freelist_count
;
1539 page_limit
= page
->objects
* s
->size
;
1540 start
= fixup_red_left(s
, page_address(page
));
1542 /* First entry is used as the base of the freelist */
1543 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1545 page
->freelist
= cur
;
1547 for (idx
= 1; idx
< page
->objects
; idx
++) {
1548 setup_object(s
, page
, cur
);
1549 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1551 set_freepointer(s
, cur
, next
);
1554 setup_object(s
, page
, cur
);
1555 set_freepointer(s
, cur
, NULL
);
1560 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1564 static inline void init_freelist_randomization(void) { }
1565 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1569 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1571 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1574 struct kmem_cache_order_objects oo
= s
->oo
;
1580 flags
&= gfp_allowed_mask
;
1582 if (gfpflags_allow_blocking(flags
))
1585 flags
|= s
->allocflags
;
1588 * Let the initial higher-order allocation fail under memory pressure
1589 * so we fall-back to the minimum order allocation.
1591 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1592 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1593 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1595 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1596 if (unlikely(!page
)) {
1600 * Allocation may have failed due to fragmentation.
1601 * Try a lower order alloc if possible
1603 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1604 if (unlikely(!page
))
1606 stat(s
, ORDER_FALLBACK
);
1609 page
->objects
= oo_objects(oo
);
1611 order
= compound_order(page
);
1612 page
->slab_cache
= s
;
1613 __SetPageSlab(page
);
1614 if (page_is_pfmemalloc(page
))
1615 SetPageSlabPfmemalloc(page
);
1617 start
= page_address(page
);
1619 if (unlikely(s
->flags
& SLAB_POISON
))
1620 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1622 kasan_poison_slab(page
);
1624 shuffle
= shuffle_freelist(s
, page
);
1627 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1628 setup_object(s
, page
, p
);
1629 if (likely(idx
< page
->objects
))
1630 set_freepointer(s
, p
, p
+ s
->size
);
1632 set_freepointer(s
, p
, NULL
);
1634 page
->freelist
= fixup_red_left(s
, start
);
1637 page
->inuse
= page
->objects
;
1641 if (gfpflags_allow_blocking(flags
))
1642 local_irq_disable();
1646 mod_lruvec_page_state(page
,
1647 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1648 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1651 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1656 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1658 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1659 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1660 flags
&= ~GFP_SLAB_BUG_MASK
;
1661 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1662 invalid_mask
, &invalid_mask
, flags
, &flags
);
1666 return allocate_slab(s
,
1667 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1670 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1672 int order
= compound_order(page
);
1673 int pages
= 1 << order
;
1675 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1678 slab_pad_check(s
, page
);
1679 for_each_object(p
, s
, page_address(page
),
1681 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1684 mod_lruvec_page_state(page
,
1685 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1686 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1689 __ClearPageSlabPfmemalloc(page
);
1690 __ClearPageSlab(page
);
1692 page_mapcount_reset(page
);
1693 if (current
->reclaim_state
)
1694 current
->reclaim_state
->reclaimed_slab
+= pages
;
1695 memcg_uncharge_slab(page
, order
, s
);
1696 __free_pages(page
, order
);
1699 #define need_reserve_slab_rcu \
1700 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1702 static void rcu_free_slab(struct rcu_head
*h
)
1706 if (need_reserve_slab_rcu
)
1707 page
= virt_to_head_page(h
);
1709 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1711 __free_slab(page
->slab_cache
, page
);
1714 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1716 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1717 struct rcu_head
*head
;
1719 if (need_reserve_slab_rcu
) {
1720 int order
= compound_order(page
);
1721 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1723 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1724 head
= page_address(page
) + offset
;
1726 head
= &page
->rcu_head
;
1729 call_rcu(head
, rcu_free_slab
);
1731 __free_slab(s
, page
);
1734 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1736 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1741 * Management of partially allocated slabs.
1744 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1747 if (tail
== DEACTIVATE_TO_TAIL
)
1748 list_add_tail(&page
->lru
, &n
->partial
);
1750 list_add(&page
->lru
, &n
->partial
);
1753 static inline void add_partial(struct kmem_cache_node
*n
,
1754 struct page
*page
, int tail
)
1756 lockdep_assert_held(&n
->list_lock
);
1757 __add_partial(n
, page
, tail
);
1760 static inline void remove_partial(struct kmem_cache_node
*n
,
1763 lockdep_assert_held(&n
->list_lock
);
1764 list_del(&page
->lru
);
1769 * Remove slab from the partial list, freeze it and
1770 * return the pointer to the freelist.
1772 * Returns a list of objects or NULL if it fails.
1774 static inline void *acquire_slab(struct kmem_cache
*s
,
1775 struct kmem_cache_node
*n
, struct page
*page
,
1776 int mode
, int *objects
)
1779 unsigned long counters
;
1782 lockdep_assert_held(&n
->list_lock
);
1785 * Zap the freelist and set the frozen bit.
1786 * The old freelist is the list of objects for the
1787 * per cpu allocation list.
1789 freelist
= page
->freelist
;
1790 counters
= page
->counters
;
1791 new.counters
= counters
;
1792 *objects
= new.objects
- new.inuse
;
1794 new.inuse
= page
->objects
;
1795 new.freelist
= NULL
;
1797 new.freelist
= freelist
;
1800 VM_BUG_ON(new.frozen
);
1803 if (!__cmpxchg_double_slab(s
, page
,
1805 new.freelist
, new.counters
,
1809 remove_partial(n
, page
);
1814 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1815 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1818 * Try to allocate a partial slab from a specific node.
1820 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1821 struct kmem_cache_cpu
*c
, gfp_t flags
)
1823 struct page
*page
, *page2
;
1824 void *object
= NULL
;
1825 unsigned int available
= 0;
1829 * Racy check. If we mistakenly see no partial slabs then we
1830 * just allocate an empty slab. If we mistakenly try to get a
1831 * partial slab and there is none available then get_partials()
1834 if (!n
|| !n
->nr_partial
)
1837 spin_lock(&n
->list_lock
);
1838 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1841 if (!pfmemalloc_match(page
, flags
))
1844 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1848 available
+= objects
;
1851 stat(s
, ALLOC_FROM_PARTIAL
);
1854 put_cpu_partial(s
, page
, 0);
1855 stat(s
, CPU_PARTIAL_NODE
);
1857 if (!kmem_cache_has_cpu_partial(s
)
1858 || available
> slub_cpu_partial(s
) / 2)
1862 spin_unlock(&n
->list_lock
);
1867 * Get a page from somewhere. Search in increasing NUMA distances.
1869 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1870 struct kmem_cache_cpu
*c
)
1873 struct zonelist
*zonelist
;
1876 enum zone_type high_zoneidx
= gfp_zone(flags
);
1878 unsigned int cpuset_mems_cookie
;
1881 * The defrag ratio allows a configuration of the tradeoffs between
1882 * inter node defragmentation and node local allocations. A lower
1883 * defrag_ratio increases the tendency to do local allocations
1884 * instead of attempting to obtain partial slabs from other nodes.
1886 * If the defrag_ratio is set to 0 then kmalloc() always
1887 * returns node local objects. If the ratio is higher then kmalloc()
1888 * may return off node objects because partial slabs are obtained
1889 * from other nodes and filled up.
1891 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1892 * (which makes defrag_ratio = 1000) then every (well almost)
1893 * allocation will first attempt to defrag slab caches on other nodes.
1894 * This means scanning over all nodes to look for partial slabs which
1895 * may be expensive if we do it every time we are trying to find a slab
1896 * with available objects.
1898 if (!s
->remote_node_defrag_ratio
||
1899 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1903 cpuset_mems_cookie
= read_mems_allowed_begin();
1904 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1905 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1906 struct kmem_cache_node
*n
;
1908 n
= get_node(s
, zone_to_nid(zone
));
1910 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1911 n
->nr_partial
> s
->min_partial
) {
1912 object
= get_partial_node(s
, n
, c
, flags
);
1915 * Don't check read_mems_allowed_retry()
1916 * here - if mems_allowed was updated in
1917 * parallel, that was a harmless race
1918 * between allocation and the cpuset
1925 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1931 * Get a partial page, lock it and return it.
1933 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1934 struct kmem_cache_cpu
*c
)
1937 int searchnode
= node
;
1939 if (node
== NUMA_NO_NODE
)
1940 searchnode
= numa_mem_id();
1941 else if (!node_present_pages(node
))
1942 searchnode
= node_to_mem_node(node
);
1944 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1945 if (object
|| node
!= NUMA_NO_NODE
)
1948 return get_any_partial(s
, flags
, c
);
1951 #ifdef CONFIG_PREEMPT
1953 * Calculate the next globally unique transaction for disambiguiation
1954 * during cmpxchg. The transactions start with the cpu number and are then
1955 * incremented by CONFIG_NR_CPUS.
1957 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1960 * No preemption supported therefore also no need to check for
1966 static inline unsigned long next_tid(unsigned long tid
)
1968 return tid
+ TID_STEP
;
1971 static inline unsigned int tid_to_cpu(unsigned long tid
)
1973 return tid
% TID_STEP
;
1976 static inline unsigned long tid_to_event(unsigned long tid
)
1978 return tid
/ TID_STEP
;
1981 static inline unsigned int init_tid(int cpu
)
1986 static inline void note_cmpxchg_failure(const char *n
,
1987 const struct kmem_cache
*s
, unsigned long tid
)
1989 #ifdef SLUB_DEBUG_CMPXCHG
1990 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1992 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1994 #ifdef CONFIG_PREEMPT
1995 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1996 pr_warn("due to cpu change %d -> %d\n",
1997 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2000 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2001 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2002 tid_to_event(tid
), tid_to_event(actual_tid
));
2004 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2005 actual_tid
, tid
, next_tid(tid
));
2007 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2010 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2014 for_each_possible_cpu(cpu
)
2015 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2019 * Remove the cpu slab
2021 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2022 void *freelist
, struct kmem_cache_cpu
*c
)
2024 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2025 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2027 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2029 int tail
= DEACTIVATE_TO_HEAD
;
2033 if (page
->freelist
) {
2034 stat(s
, DEACTIVATE_REMOTE_FREES
);
2035 tail
= DEACTIVATE_TO_TAIL
;
2039 * Stage one: Free all available per cpu objects back
2040 * to the page freelist while it is still frozen. Leave the
2043 * There is no need to take the list->lock because the page
2046 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2048 unsigned long counters
;
2051 prior
= page
->freelist
;
2052 counters
= page
->counters
;
2053 set_freepointer(s
, freelist
, prior
);
2054 new.counters
= counters
;
2056 VM_BUG_ON(!new.frozen
);
2058 } while (!__cmpxchg_double_slab(s
, page
,
2060 freelist
, new.counters
,
2061 "drain percpu freelist"));
2063 freelist
= nextfree
;
2067 * Stage two: Ensure that the page is unfrozen while the
2068 * list presence reflects the actual number of objects
2071 * We setup the list membership and then perform a cmpxchg
2072 * with the count. If there is a mismatch then the page
2073 * is not unfrozen but the page is on the wrong list.
2075 * Then we restart the process which may have to remove
2076 * the page from the list that we just put it on again
2077 * because the number of objects in the slab may have
2082 old
.freelist
= page
->freelist
;
2083 old
.counters
= page
->counters
;
2084 VM_BUG_ON(!old
.frozen
);
2086 /* Determine target state of the slab */
2087 new.counters
= old
.counters
;
2090 set_freepointer(s
, freelist
, old
.freelist
);
2091 new.freelist
= freelist
;
2093 new.freelist
= old
.freelist
;
2097 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2099 else if (new.freelist
) {
2104 * Taking the spinlock removes the possiblity
2105 * that acquire_slab() will see a slab page that
2108 spin_lock(&n
->list_lock
);
2112 if (kmem_cache_debug(s
) && !lock
) {
2115 * This also ensures that the scanning of full
2116 * slabs from diagnostic functions will not see
2119 spin_lock(&n
->list_lock
);
2127 remove_partial(n
, page
);
2129 else if (l
== M_FULL
)
2131 remove_full(s
, n
, page
);
2133 if (m
== M_PARTIAL
) {
2135 add_partial(n
, page
, tail
);
2138 } else if (m
== M_FULL
) {
2140 stat(s
, DEACTIVATE_FULL
);
2141 add_full(s
, n
, page
);
2147 if (!__cmpxchg_double_slab(s
, page
,
2148 old
.freelist
, old
.counters
,
2149 new.freelist
, new.counters
,
2154 spin_unlock(&n
->list_lock
);
2157 stat(s
, DEACTIVATE_EMPTY
);
2158 discard_slab(s
, page
);
2167 * Unfreeze all the cpu partial slabs.
2169 * This function must be called with interrupts disabled
2170 * for the cpu using c (or some other guarantee must be there
2171 * to guarantee no concurrent accesses).
2173 static void unfreeze_partials(struct kmem_cache
*s
,
2174 struct kmem_cache_cpu
*c
)
2176 #ifdef CONFIG_SLUB_CPU_PARTIAL
2177 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2178 struct page
*page
, *discard_page
= NULL
;
2180 while ((page
= c
->partial
)) {
2184 c
->partial
= page
->next
;
2186 n2
= get_node(s
, page_to_nid(page
));
2189 spin_unlock(&n
->list_lock
);
2192 spin_lock(&n
->list_lock
);
2197 old
.freelist
= page
->freelist
;
2198 old
.counters
= page
->counters
;
2199 VM_BUG_ON(!old
.frozen
);
2201 new.counters
= old
.counters
;
2202 new.freelist
= old
.freelist
;
2206 } while (!__cmpxchg_double_slab(s
, page
,
2207 old
.freelist
, old
.counters
,
2208 new.freelist
, new.counters
,
2209 "unfreezing slab"));
2211 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2212 page
->next
= discard_page
;
2213 discard_page
= page
;
2215 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2216 stat(s
, FREE_ADD_PARTIAL
);
2221 spin_unlock(&n
->list_lock
);
2223 while (discard_page
) {
2224 page
= discard_page
;
2225 discard_page
= discard_page
->next
;
2227 stat(s
, DEACTIVATE_EMPTY
);
2228 discard_slab(s
, page
);
2235 * Put a page that was just frozen (in __slab_free) into a partial page
2236 * slot if available. This is done without interrupts disabled and without
2237 * preemption disabled. The cmpxchg is racy and may put the partial page
2238 * onto a random cpus partial slot.
2240 * If we did not find a slot then simply move all the partials to the
2241 * per node partial list.
2243 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2245 #ifdef CONFIG_SLUB_CPU_PARTIAL
2246 struct page
*oldpage
;
2254 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2257 pobjects
= oldpage
->pobjects
;
2258 pages
= oldpage
->pages
;
2259 if (drain
&& pobjects
> s
->cpu_partial
) {
2260 unsigned long flags
;
2262 * partial array is full. Move the existing
2263 * set to the per node partial list.
2265 local_irq_save(flags
);
2266 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2267 local_irq_restore(flags
);
2271 stat(s
, CPU_PARTIAL_DRAIN
);
2276 pobjects
+= page
->objects
- page
->inuse
;
2278 page
->pages
= pages
;
2279 page
->pobjects
= pobjects
;
2280 page
->next
= oldpage
;
2282 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2284 if (unlikely(!s
->cpu_partial
)) {
2285 unsigned long flags
;
2287 local_irq_save(flags
);
2288 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2289 local_irq_restore(flags
);
2295 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2297 stat(s
, CPUSLAB_FLUSH
);
2298 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2300 c
->tid
= next_tid(c
->tid
);
2306 * Called from IPI handler with interrupts disabled.
2308 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2310 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2316 unfreeze_partials(s
, c
);
2320 static void flush_cpu_slab(void *d
)
2322 struct kmem_cache
*s
= d
;
2324 __flush_cpu_slab(s
, smp_processor_id());
2327 static bool has_cpu_slab(int cpu
, void *info
)
2329 struct kmem_cache
*s
= info
;
2330 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2332 return c
->page
|| slub_percpu_partial(c
);
2335 static void flush_all(struct kmem_cache
*s
)
2337 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2341 * Use the cpu notifier to insure that the cpu slabs are flushed when
2344 static int slub_cpu_dead(unsigned int cpu
)
2346 struct kmem_cache
*s
;
2347 unsigned long flags
;
2349 mutex_lock(&slab_mutex
);
2350 list_for_each_entry(s
, &slab_caches
, list
) {
2351 local_irq_save(flags
);
2352 __flush_cpu_slab(s
, cpu
);
2353 local_irq_restore(flags
);
2355 mutex_unlock(&slab_mutex
);
2360 * Check if the objects in a per cpu structure fit numa
2361 * locality expectations.
2363 static inline int node_match(struct page
*page
, int node
)
2366 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2372 #ifdef CONFIG_SLUB_DEBUG
2373 static int count_free(struct page
*page
)
2375 return page
->objects
- page
->inuse
;
2378 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2380 return atomic_long_read(&n
->total_objects
);
2382 #endif /* CONFIG_SLUB_DEBUG */
2384 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2385 static unsigned long count_partial(struct kmem_cache_node
*n
,
2386 int (*get_count
)(struct page
*))
2388 unsigned long flags
;
2389 unsigned long x
= 0;
2392 spin_lock_irqsave(&n
->list_lock
, flags
);
2393 list_for_each_entry(page
, &n
->partial
, lru
)
2394 x
+= get_count(page
);
2395 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2398 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2400 static noinline
void
2401 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2403 #ifdef CONFIG_SLUB_DEBUG
2404 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2405 DEFAULT_RATELIMIT_BURST
);
2407 struct kmem_cache_node
*n
;
2409 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2412 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2413 nid
, gfpflags
, &gfpflags
);
2414 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2415 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2418 if (oo_order(s
->min
) > get_order(s
->object_size
))
2419 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2422 for_each_kmem_cache_node(s
, node
, n
) {
2423 unsigned long nr_slabs
;
2424 unsigned long nr_objs
;
2425 unsigned long nr_free
;
2427 nr_free
= count_partial(n
, count_free
);
2428 nr_slabs
= node_nr_slabs(n
);
2429 nr_objs
= node_nr_objs(n
);
2431 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2432 node
, nr_slabs
, nr_objs
, nr_free
);
2437 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2438 int node
, struct kmem_cache_cpu
**pc
)
2441 struct kmem_cache_cpu
*c
= *pc
;
2444 freelist
= get_partial(s
, flags
, node
, c
);
2449 page
= new_slab(s
, flags
, node
);
2451 c
= raw_cpu_ptr(s
->cpu_slab
);
2456 * No other reference to the page yet so we can
2457 * muck around with it freely without cmpxchg
2459 freelist
= page
->freelist
;
2460 page
->freelist
= NULL
;
2462 stat(s
, ALLOC_SLAB
);
2471 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2473 if (unlikely(PageSlabPfmemalloc(page
)))
2474 return gfp_pfmemalloc_allowed(gfpflags
);
2480 * Check the page->freelist of a page and either transfer the freelist to the
2481 * per cpu freelist or deactivate the page.
2483 * The page is still frozen if the return value is not NULL.
2485 * If this function returns NULL then the page has been unfrozen.
2487 * This function must be called with interrupt disabled.
2489 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2492 unsigned long counters
;
2496 freelist
= page
->freelist
;
2497 counters
= page
->counters
;
2499 new.counters
= counters
;
2500 VM_BUG_ON(!new.frozen
);
2502 new.inuse
= page
->objects
;
2503 new.frozen
= freelist
!= NULL
;
2505 } while (!__cmpxchg_double_slab(s
, page
,
2514 * Slow path. The lockless freelist is empty or we need to perform
2517 * Processing is still very fast if new objects have been freed to the
2518 * regular freelist. In that case we simply take over the regular freelist
2519 * as the lockless freelist and zap the regular freelist.
2521 * If that is not working then we fall back to the partial lists. We take the
2522 * first element of the freelist as the object to allocate now and move the
2523 * rest of the freelist to the lockless freelist.
2525 * And if we were unable to get a new slab from the partial slab lists then
2526 * we need to allocate a new slab. This is the slowest path since it involves
2527 * a call to the page allocator and the setup of a new slab.
2529 * Version of __slab_alloc to use when we know that interrupts are
2530 * already disabled (which is the case for bulk allocation).
2532 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2533 unsigned long addr
, struct kmem_cache_cpu
*c
)
2543 if (unlikely(!node_match(page
, node
))) {
2544 int searchnode
= node
;
2546 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2547 searchnode
= node_to_mem_node(node
);
2549 if (unlikely(!node_match(page
, searchnode
))) {
2550 stat(s
, ALLOC_NODE_MISMATCH
);
2551 deactivate_slab(s
, page
, c
->freelist
, c
);
2557 * By rights, we should be searching for a slab page that was
2558 * PFMEMALLOC but right now, we are losing the pfmemalloc
2559 * information when the page leaves the per-cpu allocator
2561 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2562 deactivate_slab(s
, page
, c
->freelist
, c
);
2566 /* must check again c->freelist in case of cpu migration or IRQ */
2567 freelist
= c
->freelist
;
2571 freelist
= get_freelist(s
, page
);
2575 stat(s
, DEACTIVATE_BYPASS
);
2579 stat(s
, ALLOC_REFILL
);
2583 * freelist is pointing to the list of objects to be used.
2584 * page is pointing to the page from which the objects are obtained.
2585 * That page must be frozen for per cpu allocations to work.
2587 VM_BUG_ON(!c
->page
->frozen
);
2588 c
->freelist
= get_freepointer(s
, freelist
);
2589 c
->tid
= next_tid(c
->tid
);
2594 if (slub_percpu_partial(c
)) {
2595 page
= c
->page
= slub_percpu_partial(c
);
2596 slub_set_percpu_partial(c
, page
);
2597 stat(s
, CPU_PARTIAL_ALLOC
);
2601 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2603 if (unlikely(!freelist
)) {
2604 slab_out_of_memory(s
, gfpflags
, node
);
2609 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2612 /* Only entered in the debug case */
2613 if (kmem_cache_debug(s
) &&
2614 !alloc_debug_processing(s
, page
, freelist
, addr
))
2615 goto new_slab
; /* Slab failed checks. Next slab needed */
2617 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2622 * Another one that disabled interrupt and compensates for possible
2623 * cpu changes by refetching the per cpu area pointer.
2625 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2626 unsigned long addr
, struct kmem_cache_cpu
*c
)
2629 unsigned long flags
;
2631 local_irq_save(flags
);
2632 #ifdef CONFIG_PREEMPT
2634 * We may have been preempted and rescheduled on a different
2635 * cpu before disabling interrupts. Need to reload cpu area
2638 c
= this_cpu_ptr(s
->cpu_slab
);
2641 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2642 local_irq_restore(flags
);
2647 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2648 * have the fastpath folded into their functions. So no function call
2649 * overhead for requests that can be satisfied on the fastpath.
2651 * The fastpath works by first checking if the lockless freelist can be used.
2652 * If not then __slab_alloc is called for slow processing.
2654 * Otherwise we can simply pick the next object from the lockless free list.
2656 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2657 gfp_t gfpflags
, int node
, unsigned long addr
)
2660 struct kmem_cache_cpu
*c
;
2664 s
= slab_pre_alloc_hook(s
, gfpflags
);
2669 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2670 * enabled. We may switch back and forth between cpus while
2671 * reading from one cpu area. That does not matter as long
2672 * as we end up on the original cpu again when doing the cmpxchg.
2674 * We should guarantee that tid and kmem_cache are retrieved on
2675 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2676 * to check if it is matched or not.
2679 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2680 c
= raw_cpu_ptr(s
->cpu_slab
);
2681 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2682 unlikely(tid
!= READ_ONCE(c
->tid
)));
2685 * Irqless object alloc/free algorithm used here depends on sequence
2686 * of fetching cpu_slab's data. tid should be fetched before anything
2687 * on c to guarantee that object and page associated with previous tid
2688 * won't be used with current tid. If we fetch tid first, object and
2689 * page could be one associated with next tid and our alloc/free
2690 * request will be failed. In this case, we will retry. So, no problem.
2695 * The transaction ids are globally unique per cpu and per operation on
2696 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2697 * occurs on the right processor and that there was no operation on the
2698 * linked list in between.
2701 object
= c
->freelist
;
2703 if (unlikely(!object
|| !node_match(page
, node
))) {
2704 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2705 stat(s
, ALLOC_SLOWPATH
);
2707 void *next_object
= get_freepointer_safe(s
, object
);
2710 * The cmpxchg will only match if there was no additional
2711 * operation and if we are on the right processor.
2713 * The cmpxchg does the following atomically (without lock
2715 * 1. Relocate first pointer to the current per cpu area.
2716 * 2. Verify that tid and freelist have not been changed
2717 * 3. If they were not changed replace tid and freelist
2719 * Since this is without lock semantics the protection is only
2720 * against code executing on this cpu *not* from access by
2723 if (unlikely(!this_cpu_cmpxchg_double(
2724 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2726 next_object
, next_tid(tid
)))) {
2728 note_cmpxchg_failure("slab_alloc", s
, tid
);
2731 prefetch_freepointer(s
, next_object
);
2732 stat(s
, ALLOC_FASTPATH
);
2735 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2736 memset(object
, 0, s
->object_size
);
2738 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2743 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2744 gfp_t gfpflags
, unsigned long addr
)
2746 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2749 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2751 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2753 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2758 EXPORT_SYMBOL(kmem_cache_alloc
);
2760 #ifdef CONFIG_TRACING
2761 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2763 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2764 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2765 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2768 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2772 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2774 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2776 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2777 s
->object_size
, s
->size
, gfpflags
, node
);
2781 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2783 #ifdef CONFIG_TRACING
2784 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2786 int node
, size_t size
)
2788 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2790 trace_kmalloc_node(_RET_IP_
, ret
,
2791 size
, s
->size
, gfpflags
, node
);
2793 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2796 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2801 * Slow path handling. This may still be called frequently since objects
2802 * have a longer lifetime than the cpu slabs in most processing loads.
2804 * So we still attempt to reduce cache line usage. Just take the slab
2805 * lock and free the item. If there is no additional partial page
2806 * handling required then we can return immediately.
2808 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2809 void *head
, void *tail
, int cnt
,
2816 unsigned long counters
;
2817 struct kmem_cache_node
*n
= NULL
;
2818 unsigned long uninitialized_var(flags
);
2820 stat(s
, FREE_SLOWPATH
);
2822 if (kmem_cache_debug(s
) &&
2823 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2828 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2831 prior
= page
->freelist
;
2832 counters
= page
->counters
;
2833 set_freepointer(s
, tail
, prior
);
2834 new.counters
= counters
;
2835 was_frozen
= new.frozen
;
2837 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2839 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2842 * Slab was on no list before and will be
2844 * We can defer the list move and instead
2849 } else { /* Needs to be taken off a list */
2851 n
= get_node(s
, page_to_nid(page
));
2853 * Speculatively acquire the list_lock.
2854 * If the cmpxchg does not succeed then we may
2855 * drop the list_lock without any processing.
2857 * Otherwise the list_lock will synchronize with
2858 * other processors updating the list of slabs.
2860 spin_lock_irqsave(&n
->list_lock
, flags
);
2865 } while (!cmpxchg_double_slab(s
, page
,
2873 * If we just froze the page then put it onto the
2874 * per cpu partial list.
2876 if (new.frozen
&& !was_frozen
) {
2877 put_cpu_partial(s
, page
, 1);
2878 stat(s
, CPU_PARTIAL_FREE
);
2881 * The list lock was not taken therefore no list
2882 * activity can be necessary.
2885 stat(s
, FREE_FROZEN
);
2889 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2893 * Objects left in the slab. If it was not on the partial list before
2896 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2897 if (kmem_cache_debug(s
))
2898 remove_full(s
, n
, page
);
2899 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2900 stat(s
, FREE_ADD_PARTIAL
);
2902 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2908 * Slab on the partial list.
2910 remove_partial(n
, page
);
2911 stat(s
, FREE_REMOVE_PARTIAL
);
2913 /* Slab must be on the full list */
2914 remove_full(s
, n
, page
);
2917 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2919 discard_slab(s
, page
);
2923 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2924 * can perform fastpath freeing without additional function calls.
2926 * The fastpath is only possible if we are freeing to the current cpu slab
2927 * of this processor. This typically the case if we have just allocated
2930 * If fastpath is not possible then fall back to __slab_free where we deal
2931 * with all sorts of special processing.
2933 * Bulk free of a freelist with several objects (all pointing to the
2934 * same page) possible by specifying head and tail ptr, plus objects
2935 * count (cnt). Bulk free indicated by tail pointer being set.
2937 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2938 struct page
*page
, void *head
, void *tail
,
2939 int cnt
, unsigned long addr
)
2941 void *tail_obj
= tail
? : head
;
2942 struct kmem_cache_cpu
*c
;
2946 * Determine the currently cpus per cpu slab.
2947 * The cpu may change afterward. However that does not matter since
2948 * data is retrieved via this pointer. If we are on the same cpu
2949 * during the cmpxchg then the free will succeed.
2952 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2953 c
= raw_cpu_ptr(s
->cpu_slab
);
2954 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2955 unlikely(tid
!= READ_ONCE(c
->tid
)));
2957 /* Same with comment on barrier() in slab_alloc_node() */
2960 if (likely(page
== c
->page
)) {
2961 set_freepointer(s
, tail_obj
, c
->freelist
);
2963 if (unlikely(!this_cpu_cmpxchg_double(
2964 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2966 head
, next_tid(tid
)))) {
2968 note_cmpxchg_failure("slab_free", s
, tid
);
2971 stat(s
, FREE_FASTPATH
);
2973 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2977 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2978 void *head
, void *tail
, int cnt
,
2982 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2983 * to remove objects, whose reuse must be delayed.
2985 if (slab_free_freelist_hook(s
, &head
, &tail
))
2986 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2990 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2992 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
2996 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2998 s
= cache_from_obj(s
, x
);
3001 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3002 trace_kmem_cache_free(_RET_IP_
, x
);
3004 EXPORT_SYMBOL(kmem_cache_free
);
3006 struct detached_freelist
{
3011 struct kmem_cache
*s
;
3015 * This function progressively scans the array with free objects (with
3016 * a limited look ahead) and extract objects belonging to the same
3017 * page. It builds a detached freelist directly within the given
3018 * page/objects. This can happen without any need for
3019 * synchronization, because the objects are owned by running process.
3020 * The freelist is build up as a single linked list in the objects.
3021 * The idea is, that this detached freelist can then be bulk
3022 * transferred to the real freelist(s), but only requiring a single
3023 * synchronization primitive. Look ahead in the array is limited due
3024 * to performance reasons.
3027 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3028 void **p
, struct detached_freelist
*df
)
3030 size_t first_skipped_index
= 0;
3035 /* Always re-init detached_freelist */
3040 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3041 } while (!object
&& size
);
3046 page
= virt_to_head_page(object
);
3048 /* Handle kalloc'ed objects */
3049 if (unlikely(!PageSlab(page
))) {
3050 BUG_ON(!PageCompound(page
));
3052 __free_pages(page
, compound_order(page
));
3053 p
[size
] = NULL
; /* mark object processed */
3056 /* Derive kmem_cache from object */
3057 df
->s
= page
->slab_cache
;
3059 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3062 /* Start new detached freelist */
3064 set_freepointer(df
->s
, object
, NULL
);
3066 df
->freelist
= object
;
3067 p
[size
] = NULL
; /* mark object processed */
3073 continue; /* Skip processed objects */
3075 /* df->page is always set at this point */
3076 if (df
->page
== virt_to_head_page(object
)) {
3077 /* Opportunity build freelist */
3078 set_freepointer(df
->s
, object
, df
->freelist
);
3079 df
->freelist
= object
;
3081 p
[size
] = NULL
; /* mark object processed */
3086 /* Limit look ahead search */
3090 if (!first_skipped_index
)
3091 first_skipped_index
= size
+ 1;
3094 return first_skipped_index
;
3097 /* Note that interrupts must be enabled when calling this function. */
3098 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3104 struct detached_freelist df
;
3106 size
= build_detached_freelist(s
, size
, p
, &df
);
3110 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3111 } while (likely(size
));
3113 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3115 /* Note that interrupts must be enabled when calling this function. */
3116 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3119 struct kmem_cache_cpu
*c
;
3122 /* memcg and kmem_cache debug support */
3123 s
= slab_pre_alloc_hook(s
, flags
);
3127 * Drain objects in the per cpu slab, while disabling local
3128 * IRQs, which protects against PREEMPT and interrupts
3129 * handlers invoking normal fastpath.
3131 local_irq_disable();
3132 c
= this_cpu_ptr(s
->cpu_slab
);
3134 for (i
= 0; i
< size
; i
++) {
3135 void *object
= c
->freelist
;
3137 if (unlikely(!object
)) {
3139 * Invoking slow path likely have side-effect
3140 * of re-populating per CPU c->freelist
3142 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3144 if (unlikely(!p
[i
]))
3147 c
= this_cpu_ptr(s
->cpu_slab
);
3148 continue; /* goto for-loop */
3150 c
->freelist
= get_freepointer(s
, object
);
3153 c
->tid
= next_tid(c
->tid
);
3156 /* Clear memory outside IRQ disabled fastpath loop */
3157 if (unlikely(flags
& __GFP_ZERO
)) {
3160 for (j
= 0; j
< i
; j
++)
3161 memset(p
[j
], 0, s
->object_size
);
3164 /* memcg and kmem_cache debug support */
3165 slab_post_alloc_hook(s
, flags
, size
, p
);
3169 slab_post_alloc_hook(s
, flags
, i
, p
);
3170 __kmem_cache_free_bulk(s
, i
, p
);
3173 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3177 * Object placement in a slab is made very easy because we always start at
3178 * offset 0. If we tune the size of the object to the alignment then we can
3179 * get the required alignment by putting one properly sized object after
3182 * Notice that the allocation order determines the sizes of the per cpu
3183 * caches. Each processor has always one slab available for allocations.
3184 * Increasing the allocation order reduces the number of times that slabs
3185 * must be moved on and off the partial lists and is therefore a factor in
3190 * Mininum / Maximum order of slab pages. This influences locking overhead
3191 * and slab fragmentation. A higher order reduces the number of partial slabs
3192 * and increases the number of allocations possible without having to
3193 * take the list_lock.
3195 static int slub_min_order
;
3196 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3197 static int slub_min_objects
;
3200 * Calculate the order of allocation given an slab object size.
3202 * The order of allocation has significant impact on performance and other
3203 * system components. Generally order 0 allocations should be preferred since
3204 * order 0 does not cause fragmentation in the page allocator. Larger objects
3205 * be problematic to put into order 0 slabs because there may be too much
3206 * unused space left. We go to a higher order if more than 1/16th of the slab
3209 * In order to reach satisfactory performance we must ensure that a minimum
3210 * number of objects is in one slab. Otherwise we may generate too much
3211 * activity on the partial lists which requires taking the list_lock. This is
3212 * less a concern for large slabs though which are rarely used.
3214 * slub_max_order specifies the order where we begin to stop considering the
3215 * number of objects in a slab as critical. If we reach slub_max_order then
3216 * we try to keep the page order as low as possible. So we accept more waste
3217 * of space in favor of a small page order.
3219 * Higher order allocations also allow the placement of more objects in a
3220 * slab and thereby reduce object handling overhead. If the user has
3221 * requested a higher mininum order then we start with that one instead of
3222 * the smallest order which will fit the object.
3224 static inline int slab_order(int size
, int min_objects
,
3225 int max_order
, int fract_leftover
, int reserved
)
3229 int min_order
= slub_min_order
;
3231 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3232 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3234 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3235 order
<= max_order
; order
++) {
3237 unsigned long slab_size
= PAGE_SIZE
<< order
;
3239 rem
= (slab_size
- reserved
) % size
;
3241 if (rem
<= slab_size
/ fract_leftover
)
3248 static inline int calculate_order(int size
, int reserved
)
3256 * Attempt to find best configuration for a slab. This
3257 * works by first attempting to generate a layout with
3258 * the best configuration and backing off gradually.
3260 * First we increase the acceptable waste in a slab. Then
3261 * we reduce the minimum objects required in a slab.
3263 min_objects
= slub_min_objects
;
3265 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3266 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3267 min_objects
= min(min_objects
, max_objects
);
3269 while (min_objects
> 1) {
3271 while (fraction
>= 4) {
3272 order
= slab_order(size
, min_objects
,
3273 slub_max_order
, fraction
, reserved
);
3274 if (order
<= slub_max_order
)
3282 * We were unable to place multiple objects in a slab. Now
3283 * lets see if we can place a single object there.
3285 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3286 if (order
<= slub_max_order
)
3290 * Doh this slab cannot be placed using slub_max_order.
3292 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3293 if (order
< MAX_ORDER
)
3299 init_kmem_cache_node(struct kmem_cache_node
*n
)
3302 spin_lock_init(&n
->list_lock
);
3303 INIT_LIST_HEAD(&n
->partial
);
3304 #ifdef CONFIG_SLUB_DEBUG
3305 atomic_long_set(&n
->nr_slabs
, 0);
3306 atomic_long_set(&n
->total_objects
, 0);
3307 INIT_LIST_HEAD(&n
->full
);
3311 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3313 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3314 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3317 * Must align to double word boundary for the double cmpxchg
3318 * instructions to work; see __pcpu_double_call_return_bool().
3320 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3321 2 * sizeof(void *));
3326 init_kmem_cache_cpus(s
);
3331 static struct kmem_cache
*kmem_cache_node
;
3334 * No kmalloc_node yet so do it by hand. We know that this is the first
3335 * slab on the node for this slabcache. There are no concurrent accesses
3338 * Note that this function only works on the kmem_cache_node
3339 * when allocating for the kmem_cache_node. This is used for bootstrapping
3340 * memory on a fresh node that has no slab structures yet.
3342 static void early_kmem_cache_node_alloc(int node
)
3345 struct kmem_cache_node
*n
;
3347 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3349 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3352 if (page_to_nid(page
) != node
) {
3353 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3354 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3359 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3362 kmem_cache_node
->node
[node
] = n
;
3363 #ifdef CONFIG_SLUB_DEBUG
3364 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3365 init_tracking(kmem_cache_node
, n
);
3367 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3369 init_kmem_cache_node(n
);
3370 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3373 * No locks need to be taken here as it has just been
3374 * initialized and there is no concurrent access.
3376 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3379 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3382 struct kmem_cache_node
*n
;
3384 for_each_kmem_cache_node(s
, node
, n
) {
3385 s
->node
[node
] = NULL
;
3386 kmem_cache_free(kmem_cache_node
, n
);
3390 void __kmem_cache_release(struct kmem_cache
*s
)
3392 cache_random_seq_destroy(s
);
3393 free_percpu(s
->cpu_slab
);
3394 free_kmem_cache_nodes(s
);
3397 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3401 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3402 struct kmem_cache_node
*n
;
3404 if (slab_state
== DOWN
) {
3405 early_kmem_cache_node_alloc(node
);
3408 n
= kmem_cache_alloc_node(kmem_cache_node
,
3412 free_kmem_cache_nodes(s
);
3416 init_kmem_cache_node(n
);
3422 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3424 if (min
< MIN_PARTIAL
)
3426 else if (min
> MAX_PARTIAL
)
3428 s
->min_partial
= min
;
3431 static void set_cpu_partial(struct kmem_cache
*s
)
3433 #ifdef CONFIG_SLUB_CPU_PARTIAL
3435 * cpu_partial determined the maximum number of objects kept in the
3436 * per cpu partial lists of a processor.
3438 * Per cpu partial lists mainly contain slabs that just have one
3439 * object freed. If they are used for allocation then they can be
3440 * filled up again with minimal effort. The slab will never hit the
3441 * per node partial lists and therefore no locking will be required.
3443 * This setting also determines
3445 * A) The number of objects from per cpu partial slabs dumped to the
3446 * per node list when we reach the limit.
3447 * B) The number of objects in cpu partial slabs to extract from the
3448 * per node list when we run out of per cpu objects. We only fetch
3449 * 50% to keep some capacity around for frees.
3451 if (!kmem_cache_has_cpu_partial(s
))
3453 else if (s
->size
>= PAGE_SIZE
)
3455 else if (s
->size
>= 1024)
3457 else if (s
->size
>= 256)
3458 s
->cpu_partial
= 13;
3460 s
->cpu_partial
= 30;
3465 * calculate_sizes() determines the order and the distribution of data within
3468 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3470 slab_flags_t flags
= s
->flags
;
3471 size_t size
= s
->object_size
;
3475 * Round up object size to the next word boundary. We can only
3476 * place the free pointer at word boundaries and this determines
3477 * the possible location of the free pointer.
3479 size
= ALIGN(size
, sizeof(void *));
3481 #ifdef CONFIG_SLUB_DEBUG
3483 * Determine if we can poison the object itself. If the user of
3484 * the slab may touch the object after free or before allocation
3485 * then we should never poison the object itself.
3487 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3489 s
->flags
|= __OBJECT_POISON
;
3491 s
->flags
&= ~__OBJECT_POISON
;
3495 * If we are Redzoning then check if there is some space between the
3496 * end of the object and the free pointer. If not then add an
3497 * additional word to have some bytes to store Redzone information.
3499 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3500 size
+= sizeof(void *);
3504 * With that we have determined the number of bytes in actual use
3505 * by the object. This is the potential offset to the free pointer.
3509 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3512 * Relocate free pointer after the object if it is not
3513 * permitted to overwrite the first word of the object on
3516 * This is the case if we do RCU, have a constructor or
3517 * destructor or are poisoning the objects.
3520 size
+= sizeof(void *);
3523 #ifdef CONFIG_SLUB_DEBUG
3524 if (flags
& SLAB_STORE_USER
)
3526 * Need to store information about allocs and frees after
3529 size
+= 2 * sizeof(struct track
);
3532 kasan_cache_create(s
, &size
, &s
->flags
);
3533 #ifdef CONFIG_SLUB_DEBUG
3534 if (flags
& SLAB_RED_ZONE
) {
3536 * Add some empty padding so that we can catch
3537 * overwrites from earlier objects rather than let
3538 * tracking information or the free pointer be
3539 * corrupted if a user writes before the start
3542 size
+= sizeof(void *);
3544 s
->red_left_pad
= sizeof(void *);
3545 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3546 size
+= s
->red_left_pad
;
3551 * SLUB stores one object immediately after another beginning from
3552 * offset 0. In order to align the objects we have to simply size
3553 * each object to conform to the alignment.
3555 size
= ALIGN(size
, s
->align
);
3557 if (forced_order
>= 0)
3558 order
= forced_order
;
3560 order
= calculate_order(size
, s
->reserved
);
3567 s
->allocflags
|= __GFP_COMP
;
3569 if (s
->flags
& SLAB_CACHE_DMA
)
3570 s
->allocflags
|= GFP_DMA
;
3572 if (s
->flags
& SLAB_CACHE_DMA32
)
3573 s
->allocflags
|= GFP_DMA32
;
3575 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3576 s
->allocflags
|= __GFP_RECLAIMABLE
;
3579 * Determine the number of objects per slab
3581 s
->oo
= oo_make(order
, size
, s
->reserved
);
3582 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3583 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3586 return !!oo_objects(s
->oo
);
3589 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3591 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3593 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3594 s
->random
= get_random_long();
3597 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_TYPESAFE_BY_RCU
))
3598 s
->reserved
= sizeof(struct rcu_head
);
3600 if (!calculate_sizes(s
, -1))
3602 if (disable_higher_order_debug
) {
3604 * Disable debugging flags that store metadata if the min slab
3607 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3608 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3610 if (!calculate_sizes(s
, -1))
3615 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3616 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3617 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3618 /* Enable fast mode */
3619 s
->flags
|= __CMPXCHG_DOUBLE
;
3623 * The larger the object size is, the more pages we want on the partial
3624 * list to avoid pounding the page allocator excessively.
3626 set_min_partial(s
, ilog2(s
->size
) / 2);
3631 s
->remote_node_defrag_ratio
= 1000;
3634 /* Initialize the pre-computed randomized freelist if slab is up */
3635 if (slab_state
>= UP
) {
3636 if (init_cache_random_seq(s
))
3640 if (!init_kmem_cache_nodes(s
))
3643 if (alloc_kmem_cache_cpus(s
))
3646 free_kmem_cache_nodes(s
);
3648 if (flags
& SLAB_PANIC
)
3649 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3650 s
->name
, (unsigned long)s
->size
, s
->size
,
3651 oo_order(s
->oo
), s
->offset
, (unsigned long)flags
);
3655 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3658 #ifdef CONFIG_SLUB_DEBUG
3659 void *addr
= page_address(page
);
3661 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3662 sizeof(long), GFP_ATOMIC
);
3665 slab_err(s
, page
, text
, s
->name
);
3668 get_map(s
, page
, map
);
3669 for_each_object(p
, s
, addr
, page
->objects
) {
3671 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3672 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3673 print_tracking(s
, p
);
3682 * Attempt to free all partial slabs on a node.
3683 * This is called from __kmem_cache_shutdown(). We must take list_lock
3684 * because sysfs file might still access partial list after the shutdowning.
3686 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3689 struct page
*page
, *h
;
3691 BUG_ON(irqs_disabled());
3692 spin_lock_irq(&n
->list_lock
);
3693 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3695 remove_partial(n
, page
);
3696 list_add(&page
->lru
, &discard
);
3698 list_slab_objects(s
, page
,
3699 "Objects remaining in %s on __kmem_cache_shutdown()");
3702 spin_unlock_irq(&n
->list_lock
);
3704 list_for_each_entry_safe(page
, h
, &discard
, lru
)
3705 discard_slab(s
, page
);
3709 * Release all resources used by a slab cache.
3711 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3714 struct kmem_cache_node
*n
;
3717 /* Attempt to free all objects */
3718 for_each_kmem_cache_node(s
, node
, n
) {
3720 if (n
->nr_partial
|| slabs_node(s
, node
))
3723 sysfs_slab_remove(s
);
3727 /********************************************************************
3729 *******************************************************************/
3731 static int __init
setup_slub_min_order(char *str
)
3733 get_option(&str
, &slub_min_order
);
3738 __setup("slub_min_order=", setup_slub_min_order
);
3740 static int __init
setup_slub_max_order(char *str
)
3742 get_option(&str
, &slub_max_order
);
3743 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3748 __setup("slub_max_order=", setup_slub_max_order
);
3750 static int __init
setup_slub_min_objects(char *str
)
3752 get_option(&str
, &slub_min_objects
);
3757 __setup("slub_min_objects=", setup_slub_min_objects
);
3759 void *__kmalloc(size_t size
, gfp_t flags
)
3761 struct kmem_cache
*s
;
3764 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3765 return kmalloc_large(size
, flags
);
3767 s
= kmalloc_slab(size
, flags
);
3769 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3772 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3774 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3776 kasan_kmalloc(s
, ret
, size
, flags
);
3780 EXPORT_SYMBOL(__kmalloc
);
3783 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3788 flags
|= __GFP_COMP
;
3789 page
= alloc_pages_node(node
, flags
, get_order(size
));
3791 ptr
= page_address(page
);
3793 kmalloc_large_node_hook(ptr
, size
, flags
);
3797 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3799 struct kmem_cache
*s
;
3802 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3803 ret
= kmalloc_large_node(size
, flags
, node
);
3805 trace_kmalloc_node(_RET_IP_
, ret
,
3806 size
, PAGE_SIZE
<< get_order(size
),
3812 s
= kmalloc_slab(size
, flags
);
3814 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3817 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3819 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3821 kasan_kmalloc(s
, ret
, size
, flags
);
3825 EXPORT_SYMBOL(__kmalloc_node
);
3828 #ifdef CONFIG_HARDENED_USERCOPY
3830 * Rejects objects that are incorrectly sized.
3832 * Returns NULL if check passes, otherwise const char * to name of cache
3833 * to indicate an error.
3835 const char *__check_heap_object(const void *ptr
, unsigned long n
,
3838 struct kmem_cache
*s
;
3839 unsigned long offset
;
3842 /* Find object and usable object size. */
3843 s
= page
->slab_cache
;
3844 object_size
= slab_ksize(s
);
3846 /* Reject impossible pointers. */
3847 if (ptr
< page_address(page
))
3850 /* Find offset within object. */
3851 offset
= (ptr
- page_address(page
)) % s
->size
;
3853 /* Adjust for redzone and reject if within the redzone. */
3854 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3855 if (offset
< s
->red_left_pad
)
3857 offset
-= s
->red_left_pad
;
3860 /* Allow address range falling entirely within object size. */
3861 if (offset
<= object_size
&& n
<= object_size
- offset
)
3866 #endif /* CONFIG_HARDENED_USERCOPY */
3868 static size_t __ksize(const void *object
)
3872 if (unlikely(object
== ZERO_SIZE_PTR
))
3875 page
= virt_to_head_page(object
);
3877 if (unlikely(!PageSlab(page
))) {
3878 WARN_ON(!PageCompound(page
));
3879 return PAGE_SIZE
<< compound_order(page
);
3882 return slab_ksize(page
->slab_cache
);
3885 size_t ksize(const void *object
)
3887 size_t size
= __ksize(object
);
3888 /* We assume that ksize callers could use whole allocated area,
3889 * so we need to unpoison this area.
3891 kasan_unpoison_shadow(object
, size
);
3894 EXPORT_SYMBOL(ksize
);
3896 void kfree(const void *x
)
3899 void *object
= (void *)x
;
3901 trace_kfree(_RET_IP_
, x
);
3903 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3906 page
= virt_to_head_page(x
);
3907 if (unlikely(!PageSlab(page
))) {
3908 BUG_ON(!PageCompound(page
));
3910 __free_pages(page
, compound_order(page
));
3913 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3915 EXPORT_SYMBOL(kfree
);
3917 #define SHRINK_PROMOTE_MAX 32
3920 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3921 * up most to the head of the partial lists. New allocations will then
3922 * fill those up and thus they can be removed from the partial lists.
3924 * The slabs with the least items are placed last. This results in them
3925 * being allocated from last increasing the chance that the last objects
3926 * are freed in them.
3928 int __kmem_cache_shrink(struct kmem_cache
*s
)
3932 struct kmem_cache_node
*n
;
3935 struct list_head discard
;
3936 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3937 unsigned long flags
;
3941 for_each_kmem_cache_node(s
, node
, n
) {
3942 INIT_LIST_HEAD(&discard
);
3943 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3944 INIT_LIST_HEAD(promote
+ i
);
3946 spin_lock_irqsave(&n
->list_lock
, flags
);
3949 * Build lists of slabs to discard or promote.
3951 * Note that concurrent frees may occur while we hold the
3952 * list_lock. page->inuse here is the upper limit.
3954 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3955 int free
= page
->objects
- page
->inuse
;
3957 /* Do not reread page->inuse */
3960 /* We do not keep full slabs on the list */
3963 if (free
== page
->objects
) {
3964 list_move(&page
->lru
, &discard
);
3966 } else if (free
<= SHRINK_PROMOTE_MAX
)
3967 list_move(&page
->lru
, promote
+ free
- 1);
3971 * Promote the slabs filled up most to the head of the
3974 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3975 list_splice(promote
+ i
, &n
->partial
);
3977 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3979 /* Release empty slabs */
3980 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3981 discard_slab(s
, page
);
3983 if (slabs_node(s
, node
))
3991 static void kmemcg_cache_deact_after_rcu(struct kmem_cache
*s
)
3994 * Called with all the locks held after a sched RCU grace period.
3995 * Even if @s becomes empty after shrinking, we can't know that @s
3996 * doesn't have allocations already in-flight and thus can't
3997 * destroy @s until the associated memcg is released.
3999 * However, let's remove the sysfs files for empty caches here.
4000 * Each cache has a lot of interface files which aren't
4001 * particularly useful for empty draining caches; otherwise, we can
4002 * easily end up with millions of unnecessary sysfs files on
4003 * systems which have a lot of memory and transient cgroups.
4005 if (!__kmem_cache_shrink(s
))
4006 sysfs_slab_remove(s
);
4009 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4012 * Disable empty slabs caching. Used to avoid pinning offline
4013 * memory cgroups by kmem pages that can be freed.
4015 slub_set_cpu_partial(s
, 0);
4019 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4020 * we have to make sure the change is visible before shrinking.
4022 slab_deactivate_memcg_cache_rcu_sched(s
, kmemcg_cache_deact_after_rcu
);
4026 static int slab_mem_going_offline_callback(void *arg
)
4028 struct kmem_cache
*s
;
4030 mutex_lock(&slab_mutex
);
4031 list_for_each_entry(s
, &slab_caches
, list
)
4032 __kmem_cache_shrink(s
);
4033 mutex_unlock(&slab_mutex
);
4038 static void slab_mem_offline_callback(void *arg
)
4040 struct kmem_cache_node
*n
;
4041 struct kmem_cache
*s
;
4042 struct memory_notify
*marg
= arg
;
4045 offline_node
= marg
->status_change_nid_normal
;
4048 * If the node still has available memory. we need kmem_cache_node
4051 if (offline_node
< 0)
4054 mutex_lock(&slab_mutex
);
4055 list_for_each_entry(s
, &slab_caches
, list
) {
4056 n
= get_node(s
, offline_node
);
4059 * if n->nr_slabs > 0, slabs still exist on the node
4060 * that is going down. We were unable to free them,
4061 * and offline_pages() function shouldn't call this
4062 * callback. So, we must fail.
4064 BUG_ON(slabs_node(s
, offline_node
));
4066 s
->node
[offline_node
] = NULL
;
4067 kmem_cache_free(kmem_cache_node
, n
);
4070 mutex_unlock(&slab_mutex
);
4073 static int slab_mem_going_online_callback(void *arg
)
4075 struct kmem_cache_node
*n
;
4076 struct kmem_cache
*s
;
4077 struct memory_notify
*marg
= arg
;
4078 int nid
= marg
->status_change_nid_normal
;
4082 * If the node's memory is already available, then kmem_cache_node is
4083 * already created. Nothing to do.
4089 * We are bringing a node online. No memory is available yet. We must
4090 * allocate a kmem_cache_node structure in order to bring the node
4093 mutex_lock(&slab_mutex
);
4094 list_for_each_entry(s
, &slab_caches
, list
) {
4096 * XXX: kmem_cache_alloc_node will fallback to other nodes
4097 * since memory is not yet available from the node that
4100 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4105 init_kmem_cache_node(n
);
4109 mutex_unlock(&slab_mutex
);
4113 static int slab_memory_callback(struct notifier_block
*self
,
4114 unsigned long action
, void *arg
)
4119 case MEM_GOING_ONLINE
:
4120 ret
= slab_mem_going_online_callback(arg
);
4122 case MEM_GOING_OFFLINE
:
4123 ret
= slab_mem_going_offline_callback(arg
);
4126 case MEM_CANCEL_ONLINE
:
4127 slab_mem_offline_callback(arg
);
4130 case MEM_CANCEL_OFFLINE
:
4134 ret
= notifier_from_errno(ret
);
4140 static struct notifier_block slab_memory_callback_nb
= {
4141 .notifier_call
= slab_memory_callback
,
4142 .priority
= SLAB_CALLBACK_PRI
,
4145 /********************************************************************
4146 * Basic setup of slabs
4147 *******************************************************************/
4150 * Used for early kmem_cache structures that were allocated using
4151 * the page allocator. Allocate them properly then fix up the pointers
4152 * that may be pointing to the wrong kmem_cache structure.
4155 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4158 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4159 struct kmem_cache_node
*n
;
4161 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4164 * This runs very early, and only the boot processor is supposed to be
4165 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4168 __flush_cpu_slab(s
, smp_processor_id());
4169 for_each_kmem_cache_node(s
, node
, n
) {
4172 list_for_each_entry(p
, &n
->partial
, lru
)
4175 #ifdef CONFIG_SLUB_DEBUG
4176 list_for_each_entry(p
, &n
->full
, lru
)
4180 slab_init_memcg_params(s
);
4181 list_add(&s
->list
, &slab_caches
);
4182 memcg_link_cache(s
);
4186 void __init
kmem_cache_init(void)
4188 static __initdata
struct kmem_cache boot_kmem_cache
,
4189 boot_kmem_cache_node
;
4191 if (debug_guardpage_minorder())
4194 kmem_cache_node
= &boot_kmem_cache_node
;
4195 kmem_cache
= &boot_kmem_cache
;
4197 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4198 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
4200 register_hotmemory_notifier(&slab_memory_callback_nb
);
4202 /* Able to allocate the per node structures */
4203 slab_state
= PARTIAL
;
4205 create_boot_cache(kmem_cache
, "kmem_cache",
4206 offsetof(struct kmem_cache
, node
) +
4207 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4208 SLAB_HWCACHE_ALIGN
);
4210 kmem_cache
= bootstrap(&boot_kmem_cache
);
4213 * Allocate kmem_cache_node properly from the kmem_cache slab.
4214 * kmem_cache_node is separately allocated so no need to
4215 * update any list pointers.
4217 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4219 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4220 setup_kmalloc_cache_index_table();
4221 create_kmalloc_caches(0);
4223 /* Setup random freelists for each cache */
4224 init_freelist_randomization();
4226 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4229 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%u, Nodes=%d\n",
4231 slub_min_order
, slub_max_order
, slub_min_objects
,
4232 nr_cpu_ids
, nr_node_ids
);
4235 void __init
kmem_cache_init_late(void)
4240 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
4241 slab_flags_t flags
, void (*ctor
)(void *))
4243 struct kmem_cache
*s
, *c
;
4245 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4250 * Adjust the object sizes so that we clear
4251 * the complete object on kzalloc.
4253 s
->object_size
= max(s
->object_size
, (int)size
);
4254 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4256 for_each_memcg_cache(c
, s
) {
4257 c
->object_size
= s
->object_size
;
4258 c
->inuse
= max_t(int, c
->inuse
,
4259 ALIGN(size
, sizeof(void *)));
4262 if (sysfs_slab_alias(s
, name
)) {
4271 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4275 err
= kmem_cache_open(s
, flags
);
4279 /* Mutex is not taken during early boot */
4280 if (slab_state
<= UP
)
4283 memcg_propagate_slab_attrs(s
);
4284 err
= sysfs_slab_add(s
);
4286 __kmem_cache_release(s
);
4291 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4293 struct kmem_cache
*s
;
4296 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4297 return kmalloc_large(size
, gfpflags
);
4299 s
= kmalloc_slab(size
, gfpflags
);
4301 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4304 ret
= slab_alloc(s
, gfpflags
, caller
);
4306 /* Honor the call site pointer we received. */
4307 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4313 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4314 int node
, unsigned long caller
)
4316 struct kmem_cache
*s
;
4319 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4320 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4322 trace_kmalloc_node(caller
, ret
,
4323 size
, PAGE_SIZE
<< get_order(size
),
4329 s
= kmalloc_slab(size
, gfpflags
);
4331 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4334 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4336 /* Honor the call site pointer we received. */
4337 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4344 static int count_inuse(struct page
*page
)
4349 static int count_total(struct page
*page
)
4351 return page
->objects
;
4355 #ifdef CONFIG_SLUB_DEBUG
4356 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4360 void *addr
= page_address(page
);
4362 if (!check_slab(s
, page
) ||
4363 !on_freelist(s
, page
, NULL
))
4366 /* Now we know that a valid freelist exists */
4367 bitmap_zero(map
, page
->objects
);
4369 get_map(s
, page
, map
);
4370 for_each_object(p
, s
, addr
, page
->objects
) {
4371 if (test_bit(slab_index(p
, s
, addr
), map
))
4372 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4376 for_each_object(p
, s
, addr
, page
->objects
)
4377 if (!test_bit(slab_index(p
, s
, addr
), map
))
4378 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4383 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4387 validate_slab(s
, page
, map
);
4391 static int validate_slab_node(struct kmem_cache
*s
,
4392 struct kmem_cache_node
*n
, unsigned long *map
)
4394 unsigned long count
= 0;
4396 unsigned long flags
;
4398 spin_lock_irqsave(&n
->list_lock
, flags
);
4400 list_for_each_entry(page
, &n
->partial
, lru
) {
4401 validate_slab_slab(s
, page
, map
);
4404 if (count
!= n
->nr_partial
)
4405 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4406 s
->name
, count
, n
->nr_partial
);
4408 if (!(s
->flags
& SLAB_STORE_USER
))
4411 list_for_each_entry(page
, &n
->full
, lru
) {
4412 validate_slab_slab(s
, page
, map
);
4415 if (count
!= atomic_long_read(&n
->nr_slabs
))
4416 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4417 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4420 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4424 static long validate_slab_cache(struct kmem_cache
*s
)
4427 unsigned long count
= 0;
4428 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4429 sizeof(unsigned long), GFP_KERNEL
);
4430 struct kmem_cache_node
*n
;
4436 for_each_kmem_cache_node(s
, node
, n
)
4437 count
+= validate_slab_node(s
, n
, map
);
4442 * Generate lists of code addresses where slabcache objects are allocated
4447 unsigned long count
;
4454 DECLARE_BITMAP(cpus
, NR_CPUS
);
4460 unsigned long count
;
4461 struct location
*loc
;
4464 static void free_loc_track(struct loc_track
*t
)
4467 free_pages((unsigned long)t
->loc
,
4468 get_order(sizeof(struct location
) * t
->max
));
4471 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4476 order
= get_order(sizeof(struct location
) * max
);
4478 l
= (void *)__get_free_pages(flags
, order
);
4483 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4491 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4492 const struct track
*track
)
4494 long start
, end
, pos
;
4496 unsigned long caddr
;
4497 unsigned long age
= jiffies
- track
->when
;
4503 pos
= start
+ (end
- start
+ 1) / 2;
4506 * There is nothing at "end". If we end up there
4507 * we need to add something to before end.
4512 caddr
= t
->loc
[pos
].addr
;
4513 if (track
->addr
== caddr
) {
4519 if (age
< l
->min_time
)
4521 if (age
> l
->max_time
)
4524 if (track
->pid
< l
->min_pid
)
4525 l
->min_pid
= track
->pid
;
4526 if (track
->pid
> l
->max_pid
)
4527 l
->max_pid
= track
->pid
;
4529 cpumask_set_cpu(track
->cpu
,
4530 to_cpumask(l
->cpus
));
4532 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4536 if (track
->addr
< caddr
)
4543 * Not found. Insert new tracking element.
4545 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4551 (t
->count
- pos
) * sizeof(struct location
));
4554 l
->addr
= track
->addr
;
4558 l
->min_pid
= track
->pid
;
4559 l
->max_pid
= track
->pid
;
4560 cpumask_clear(to_cpumask(l
->cpus
));
4561 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4562 nodes_clear(l
->nodes
);
4563 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4567 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4568 struct page
*page
, enum track_item alloc
,
4571 void *addr
= page_address(page
);
4574 bitmap_zero(map
, page
->objects
);
4575 get_map(s
, page
, map
);
4577 for_each_object(p
, s
, addr
, page
->objects
)
4578 if (!test_bit(slab_index(p
, s
, addr
), map
))
4579 add_location(t
, s
, get_track(s
, p
, alloc
));
4582 static int list_locations(struct kmem_cache
*s
, char *buf
,
4583 enum track_item alloc
)
4587 struct loc_track t
= { 0, 0, NULL
};
4589 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4590 sizeof(unsigned long), GFP_KERNEL
);
4591 struct kmem_cache_node
*n
;
4593 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4596 return sprintf(buf
, "Out of memory\n");
4598 /* Push back cpu slabs */
4601 for_each_kmem_cache_node(s
, node
, n
) {
4602 unsigned long flags
;
4605 if (!atomic_long_read(&n
->nr_slabs
))
4608 spin_lock_irqsave(&n
->list_lock
, flags
);
4609 list_for_each_entry(page
, &n
->partial
, lru
)
4610 process_slab(&t
, s
, page
, alloc
, map
);
4611 list_for_each_entry(page
, &n
->full
, lru
)
4612 process_slab(&t
, s
, page
, alloc
, map
);
4613 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4616 for (i
= 0; i
< t
.count
; i
++) {
4617 struct location
*l
= &t
.loc
[i
];
4619 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4621 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4624 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4626 len
+= sprintf(buf
+ len
, "<not-available>");
4628 if (l
->sum_time
!= l
->min_time
) {
4629 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4631 (long)div_u64(l
->sum_time
, l
->count
),
4634 len
+= sprintf(buf
+ len
, " age=%ld",
4637 if (l
->min_pid
!= l
->max_pid
)
4638 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4639 l
->min_pid
, l
->max_pid
);
4641 len
+= sprintf(buf
+ len
, " pid=%ld",
4644 if (num_online_cpus() > 1 &&
4645 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4646 len
< PAGE_SIZE
- 60)
4647 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4649 cpumask_pr_args(to_cpumask(l
->cpus
)));
4651 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4652 len
< PAGE_SIZE
- 60)
4653 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4655 nodemask_pr_args(&l
->nodes
));
4657 len
+= sprintf(buf
+ len
, "\n");
4663 len
+= sprintf(buf
, "No data\n");
4668 #ifdef SLUB_RESILIENCY_TEST
4669 static void __init
resiliency_test(void)
4673 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4675 pr_err("SLUB resiliency testing\n");
4676 pr_err("-----------------------\n");
4677 pr_err("A. Corruption after allocation\n");
4679 p
= kzalloc(16, GFP_KERNEL
);
4681 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4684 validate_slab_cache(kmalloc_caches
[4]);
4686 /* Hmmm... The next two are dangerous */
4687 p
= kzalloc(32, GFP_KERNEL
);
4688 p
[32 + sizeof(void *)] = 0x34;
4689 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4691 pr_err("If allocated object is overwritten then not detectable\n\n");
4693 validate_slab_cache(kmalloc_caches
[5]);
4694 p
= kzalloc(64, GFP_KERNEL
);
4695 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4697 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4699 pr_err("If allocated object is overwritten then not detectable\n\n");
4700 validate_slab_cache(kmalloc_caches
[6]);
4702 pr_err("\nB. Corruption after free\n");
4703 p
= kzalloc(128, GFP_KERNEL
);
4706 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4707 validate_slab_cache(kmalloc_caches
[7]);
4709 p
= kzalloc(256, GFP_KERNEL
);
4712 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4713 validate_slab_cache(kmalloc_caches
[8]);
4715 p
= kzalloc(512, GFP_KERNEL
);
4718 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4719 validate_slab_cache(kmalloc_caches
[9]);
4723 static void resiliency_test(void) {};
4728 enum slab_stat_type
{
4729 SL_ALL
, /* All slabs */
4730 SL_PARTIAL
, /* Only partially allocated slabs */
4731 SL_CPU
, /* Only slabs used for cpu caches */
4732 SL_OBJECTS
, /* Determine allocated objects not slabs */
4733 SL_TOTAL
/* Determine object capacity not slabs */
4736 #define SO_ALL (1 << SL_ALL)
4737 #define SO_PARTIAL (1 << SL_PARTIAL)
4738 #define SO_CPU (1 << SL_CPU)
4739 #define SO_OBJECTS (1 << SL_OBJECTS)
4740 #define SO_TOTAL (1 << SL_TOTAL)
4743 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4745 static int __init
setup_slub_memcg_sysfs(char *str
)
4749 if (get_option(&str
, &v
) > 0)
4750 memcg_sysfs_enabled
= v
;
4755 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4758 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4759 char *buf
, unsigned long flags
)
4761 unsigned long total
= 0;
4764 unsigned long *nodes
;
4766 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4770 if (flags
& SO_CPU
) {
4773 for_each_possible_cpu(cpu
) {
4774 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4779 page
= READ_ONCE(c
->page
);
4783 node
= page_to_nid(page
);
4784 if (flags
& SO_TOTAL
)
4786 else if (flags
& SO_OBJECTS
)
4794 page
= slub_percpu_partial_read_once(c
);
4796 node
= page_to_nid(page
);
4797 if (flags
& SO_TOTAL
)
4799 else if (flags
& SO_OBJECTS
)
4810 #ifdef CONFIG_SLUB_DEBUG
4811 if (flags
& SO_ALL
) {
4812 struct kmem_cache_node
*n
;
4814 for_each_kmem_cache_node(s
, node
, n
) {
4816 if (flags
& SO_TOTAL
)
4817 x
= atomic_long_read(&n
->total_objects
);
4818 else if (flags
& SO_OBJECTS
)
4819 x
= atomic_long_read(&n
->total_objects
) -
4820 count_partial(n
, count_free
);
4822 x
= atomic_long_read(&n
->nr_slabs
);
4829 if (flags
& SO_PARTIAL
) {
4830 struct kmem_cache_node
*n
;
4832 for_each_kmem_cache_node(s
, node
, n
) {
4833 if (flags
& SO_TOTAL
)
4834 x
= count_partial(n
, count_total
);
4835 else if (flags
& SO_OBJECTS
)
4836 x
= count_partial(n
, count_inuse
);
4843 x
= sprintf(buf
, "%lu", total
);
4845 for (node
= 0; node
< nr_node_ids
; node
++)
4847 x
+= sprintf(buf
+ x
, " N%d=%lu",
4852 return x
+ sprintf(buf
+ x
, "\n");
4855 #ifdef CONFIG_SLUB_DEBUG
4856 static int any_slab_objects(struct kmem_cache
*s
)
4859 struct kmem_cache_node
*n
;
4861 for_each_kmem_cache_node(s
, node
, n
)
4862 if (atomic_long_read(&n
->total_objects
))
4869 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4870 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4872 struct slab_attribute
{
4873 struct attribute attr
;
4874 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4875 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4878 #define SLAB_ATTR_RO(_name) \
4879 static struct slab_attribute _name##_attr = \
4880 __ATTR(_name, 0400, _name##_show, NULL)
4882 #define SLAB_ATTR(_name) \
4883 static struct slab_attribute _name##_attr = \
4884 __ATTR(_name, 0600, _name##_show, _name##_store)
4886 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4888 return sprintf(buf
, "%d\n", s
->size
);
4890 SLAB_ATTR_RO(slab_size
);
4892 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4894 return sprintf(buf
, "%d\n", s
->align
);
4896 SLAB_ATTR_RO(align
);
4898 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4900 return sprintf(buf
, "%d\n", s
->object_size
);
4902 SLAB_ATTR_RO(object_size
);
4904 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4906 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4908 SLAB_ATTR_RO(objs_per_slab
);
4910 static ssize_t
order_store(struct kmem_cache
*s
,
4911 const char *buf
, size_t length
)
4913 unsigned long order
;
4916 err
= kstrtoul(buf
, 10, &order
);
4920 if (order
> slub_max_order
|| order
< slub_min_order
)
4923 calculate_sizes(s
, order
);
4927 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4929 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4933 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4935 return sprintf(buf
, "%lu\n", s
->min_partial
);
4938 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4944 err
= kstrtoul(buf
, 10, &min
);
4948 set_min_partial(s
, min
);
4951 SLAB_ATTR(min_partial
);
4953 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4955 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4958 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4961 unsigned int objects
;
4964 err
= kstrtouint(buf
, 10, &objects
);
4967 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4970 slub_set_cpu_partial(s
, objects
);
4974 SLAB_ATTR(cpu_partial
);
4976 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4980 return sprintf(buf
, "%pS\n", s
->ctor
);
4984 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4986 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4988 SLAB_ATTR_RO(aliases
);
4990 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4992 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4994 SLAB_ATTR_RO(partial
);
4996 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4998 return show_slab_objects(s
, buf
, SO_CPU
);
5000 SLAB_ATTR_RO(cpu_slabs
);
5002 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5004 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5006 SLAB_ATTR_RO(objects
);
5008 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5010 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5012 SLAB_ATTR_RO(objects_partial
);
5014 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5021 for_each_online_cpu(cpu
) {
5024 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5027 pages
+= page
->pages
;
5028 objects
+= page
->pobjects
;
5032 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5035 for_each_online_cpu(cpu
) {
5038 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5040 if (page
&& len
< PAGE_SIZE
- 20)
5041 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5042 page
->pobjects
, page
->pages
);
5045 return len
+ sprintf(buf
+ len
, "\n");
5047 SLAB_ATTR_RO(slabs_cpu_partial
);
5049 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5051 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5054 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5055 const char *buf
, size_t length
)
5057 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5059 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5062 SLAB_ATTR(reclaim_account
);
5064 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5066 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5068 SLAB_ATTR_RO(hwcache_align
);
5070 #ifdef CONFIG_ZONE_DMA
5071 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5073 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5075 SLAB_ATTR_RO(cache_dma
);
5078 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5080 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5082 SLAB_ATTR_RO(destroy_by_rcu
);
5084 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
5086 return sprintf(buf
, "%d\n", s
->reserved
);
5088 SLAB_ATTR_RO(reserved
);
5090 #ifdef CONFIG_SLUB_DEBUG
5091 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5093 return show_slab_objects(s
, buf
, SO_ALL
);
5095 SLAB_ATTR_RO(slabs
);
5097 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5099 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5101 SLAB_ATTR_RO(total_objects
);
5103 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5105 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5108 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5109 const char *buf
, size_t length
)
5111 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5112 if (buf
[0] == '1') {
5113 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5114 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5118 SLAB_ATTR(sanity_checks
);
5120 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5122 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5125 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5129 * Tracing a merged cache is going to give confusing results
5130 * as well as cause other issues like converting a mergeable
5131 * cache into an umergeable one.
5133 if (s
->refcount
> 1)
5136 s
->flags
&= ~SLAB_TRACE
;
5137 if (buf
[0] == '1') {
5138 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5139 s
->flags
|= SLAB_TRACE
;
5145 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5147 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5150 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5151 const char *buf
, size_t length
)
5153 if (any_slab_objects(s
))
5156 s
->flags
&= ~SLAB_RED_ZONE
;
5157 if (buf
[0] == '1') {
5158 s
->flags
|= SLAB_RED_ZONE
;
5160 calculate_sizes(s
, -1);
5163 SLAB_ATTR(red_zone
);
5165 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5167 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5170 static ssize_t
poison_store(struct kmem_cache
*s
,
5171 const char *buf
, size_t length
)
5173 if (any_slab_objects(s
))
5176 s
->flags
&= ~SLAB_POISON
;
5177 if (buf
[0] == '1') {
5178 s
->flags
|= SLAB_POISON
;
5180 calculate_sizes(s
, -1);
5185 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5187 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5190 static ssize_t
store_user_store(struct kmem_cache
*s
,
5191 const char *buf
, size_t length
)
5193 if (any_slab_objects(s
))
5196 s
->flags
&= ~SLAB_STORE_USER
;
5197 if (buf
[0] == '1') {
5198 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5199 s
->flags
|= SLAB_STORE_USER
;
5201 calculate_sizes(s
, -1);
5204 SLAB_ATTR(store_user
);
5206 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5211 static ssize_t
validate_store(struct kmem_cache
*s
,
5212 const char *buf
, size_t length
)
5216 if (buf
[0] == '1') {
5217 ret
= validate_slab_cache(s
);
5223 SLAB_ATTR(validate
);
5225 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5227 if (!(s
->flags
& SLAB_STORE_USER
))
5229 return list_locations(s
, buf
, TRACK_ALLOC
);
5231 SLAB_ATTR_RO(alloc_calls
);
5233 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5235 if (!(s
->flags
& SLAB_STORE_USER
))
5237 return list_locations(s
, buf
, TRACK_FREE
);
5239 SLAB_ATTR_RO(free_calls
);
5240 #endif /* CONFIG_SLUB_DEBUG */
5242 #ifdef CONFIG_FAILSLAB
5243 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5245 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5248 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5251 if (s
->refcount
> 1)
5254 s
->flags
&= ~SLAB_FAILSLAB
;
5256 s
->flags
|= SLAB_FAILSLAB
;
5259 SLAB_ATTR(failslab
);
5262 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5267 static ssize_t
shrink_store(struct kmem_cache
*s
,
5268 const char *buf
, size_t length
)
5271 kmem_cache_shrink(s
);
5279 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5281 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
5284 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5285 const char *buf
, size_t length
)
5287 unsigned long ratio
;
5290 err
= kstrtoul(buf
, 10, &ratio
);
5295 s
->remote_node_defrag_ratio
= ratio
* 10;
5299 SLAB_ATTR(remote_node_defrag_ratio
);
5302 #ifdef CONFIG_SLUB_STATS
5303 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5305 unsigned long sum
= 0;
5308 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5313 for_each_online_cpu(cpu
) {
5314 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5320 len
= sprintf(buf
, "%lu", sum
);
5323 for_each_online_cpu(cpu
) {
5324 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5325 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5329 return len
+ sprintf(buf
+ len
, "\n");
5332 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5336 for_each_online_cpu(cpu
)
5337 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5340 #define STAT_ATTR(si, text) \
5341 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5343 return show_stat(s, buf, si); \
5345 static ssize_t text##_store(struct kmem_cache *s, \
5346 const char *buf, size_t length) \
5348 if (buf[0] != '0') \
5350 clear_stat(s, si); \
5355 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5356 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5357 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5358 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5359 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5360 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5361 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5362 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5363 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5364 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5365 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5366 STAT_ATTR(FREE_SLAB
, free_slab
);
5367 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5368 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5369 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5370 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5371 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5372 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5373 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5374 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5375 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5376 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5377 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5378 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5379 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5380 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5383 static struct attribute
*slab_attrs
[] = {
5384 &slab_size_attr
.attr
,
5385 &object_size_attr
.attr
,
5386 &objs_per_slab_attr
.attr
,
5388 &min_partial_attr
.attr
,
5389 &cpu_partial_attr
.attr
,
5391 &objects_partial_attr
.attr
,
5393 &cpu_slabs_attr
.attr
,
5397 &hwcache_align_attr
.attr
,
5398 &reclaim_account_attr
.attr
,
5399 &destroy_by_rcu_attr
.attr
,
5401 &reserved_attr
.attr
,
5402 &slabs_cpu_partial_attr
.attr
,
5403 #ifdef CONFIG_SLUB_DEBUG
5404 &total_objects_attr
.attr
,
5406 &sanity_checks_attr
.attr
,
5408 &red_zone_attr
.attr
,
5410 &store_user_attr
.attr
,
5411 &validate_attr
.attr
,
5412 &alloc_calls_attr
.attr
,
5413 &free_calls_attr
.attr
,
5415 #ifdef CONFIG_ZONE_DMA
5416 &cache_dma_attr
.attr
,
5419 &remote_node_defrag_ratio_attr
.attr
,
5421 #ifdef CONFIG_SLUB_STATS
5422 &alloc_fastpath_attr
.attr
,
5423 &alloc_slowpath_attr
.attr
,
5424 &free_fastpath_attr
.attr
,
5425 &free_slowpath_attr
.attr
,
5426 &free_frozen_attr
.attr
,
5427 &free_add_partial_attr
.attr
,
5428 &free_remove_partial_attr
.attr
,
5429 &alloc_from_partial_attr
.attr
,
5430 &alloc_slab_attr
.attr
,
5431 &alloc_refill_attr
.attr
,
5432 &alloc_node_mismatch_attr
.attr
,
5433 &free_slab_attr
.attr
,
5434 &cpuslab_flush_attr
.attr
,
5435 &deactivate_full_attr
.attr
,
5436 &deactivate_empty_attr
.attr
,
5437 &deactivate_to_head_attr
.attr
,
5438 &deactivate_to_tail_attr
.attr
,
5439 &deactivate_remote_frees_attr
.attr
,
5440 &deactivate_bypass_attr
.attr
,
5441 &order_fallback_attr
.attr
,
5442 &cmpxchg_double_fail_attr
.attr
,
5443 &cmpxchg_double_cpu_fail_attr
.attr
,
5444 &cpu_partial_alloc_attr
.attr
,
5445 &cpu_partial_free_attr
.attr
,
5446 &cpu_partial_node_attr
.attr
,
5447 &cpu_partial_drain_attr
.attr
,
5449 #ifdef CONFIG_FAILSLAB
5450 &failslab_attr
.attr
,
5456 static const struct attribute_group slab_attr_group
= {
5457 .attrs
= slab_attrs
,
5460 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5461 struct attribute
*attr
,
5464 struct slab_attribute
*attribute
;
5465 struct kmem_cache
*s
;
5468 attribute
= to_slab_attr(attr
);
5471 if (!attribute
->show
)
5474 err
= attribute
->show(s
, buf
);
5479 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5480 struct attribute
*attr
,
5481 const char *buf
, size_t len
)
5483 struct slab_attribute
*attribute
;
5484 struct kmem_cache
*s
;
5487 attribute
= to_slab_attr(attr
);
5490 if (!attribute
->store
)
5493 err
= attribute
->store(s
, buf
, len
);
5495 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5496 struct kmem_cache
*c
;
5498 mutex_lock(&slab_mutex
);
5499 if (s
->max_attr_size
< len
)
5500 s
->max_attr_size
= len
;
5503 * This is a best effort propagation, so this function's return
5504 * value will be determined by the parent cache only. This is
5505 * basically because not all attributes will have a well
5506 * defined semantics for rollbacks - most of the actions will
5507 * have permanent effects.
5509 * Returning the error value of any of the children that fail
5510 * is not 100 % defined, in the sense that users seeing the
5511 * error code won't be able to know anything about the state of
5514 * Only returning the error code for the parent cache at least
5515 * has well defined semantics. The cache being written to
5516 * directly either failed or succeeded, in which case we loop
5517 * through the descendants with best-effort propagation.
5519 for_each_memcg_cache(c
, s
)
5520 attribute
->store(c
, buf
, len
);
5521 mutex_unlock(&slab_mutex
);
5527 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5531 char *buffer
= NULL
;
5532 struct kmem_cache
*root_cache
;
5534 if (is_root_cache(s
))
5537 root_cache
= s
->memcg_params
.root_cache
;
5540 * This mean this cache had no attribute written. Therefore, no point
5541 * in copying default values around
5543 if (!root_cache
->max_attr_size
)
5546 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5549 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5552 if (!attr
|| !attr
->store
|| !attr
->show
)
5556 * It is really bad that we have to allocate here, so we will
5557 * do it only as a fallback. If we actually allocate, though,
5558 * we can just use the allocated buffer until the end.
5560 * Most of the slub attributes will tend to be very small in
5561 * size, but sysfs allows buffers up to a page, so they can
5562 * theoretically happen.
5566 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5569 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5570 if (WARN_ON(!buffer
))
5575 len
= attr
->show(root_cache
, buf
);
5577 attr
->store(s
, buf
, len
);
5581 free_page((unsigned long)buffer
);
5585 static void kmem_cache_release(struct kobject
*k
)
5587 slab_kmem_cache_release(to_slab(k
));
5590 static const struct sysfs_ops slab_sysfs_ops
= {
5591 .show
= slab_attr_show
,
5592 .store
= slab_attr_store
,
5595 static struct kobj_type slab_ktype
= {
5596 .sysfs_ops
= &slab_sysfs_ops
,
5597 .release
= kmem_cache_release
,
5600 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5602 struct kobj_type
*ktype
= get_ktype(kobj
);
5604 if (ktype
== &slab_ktype
)
5609 static const struct kset_uevent_ops slab_uevent_ops
= {
5610 .filter
= uevent_filter
,
5613 static struct kset
*slab_kset
;
5615 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5618 if (!is_root_cache(s
))
5619 return s
->memcg_params
.root_cache
->memcg_kset
;
5624 #define ID_STR_LENGTH 64
5626 /* Create a unique string id for a slab cache:
5628 * Format :[flags-]size
5630 static char *create_unique_id(struct kmem_cache
*s
)
5632 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5639 * First flags affecting slabcache operations. We will only
5640 * get here for aliasable slabs so we do not need to support
5641 * too many flags. The flags here must cover all flags that
5642 * are matched during merging to guarantee that the id is
5645 if (s
->flags
& SLAB_CACHE_DMA
)
5647 if (s
->flags
& SLAB_CACHE_DMA32
)
5649 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5651 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5653 if (s
->flags
& SLAB_ACCOUNT
)
5657 p
+= sprintf(p
, "%07d", s
->size
);
5659 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5663 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5665 struct kmem_cache
*s
=
5666 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5668 if (!s
->kobj
.state_in_sysfs
)
5670 * For a memcg cache, this may be called during
5671 * deactivation and again on shutdown. Remove only once.
5672 * A cache is never shut down before deactivation is
5673 * complete, so no need to worry about synchronization.
5678 kset_unregister(s
->memcg_kset
);
5680 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5682 kobject_put(&s
->kobj
);
5685 static int sysfs_slab_add(struct kmem_cache
*s
)
5689 struct kset
*kset
= cache_kset(s
);
5690 int unmergeable
= slab_unmergeable(s
);
5692 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5695 kobject_init(&s
->kobj
, &slab_ktype
);
5699 if (!unmergeable
&& disable_higher_order_debug
&&
5700 (slub_debug
& DEBUG_METADATA_FLAGS
))
5705 * Slabcache can never be merged so we can use the name proper.
5706 * This is typically the case for debug situations. In that
5707 * case we can catch duplicate names easily.
5709 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5713 * Create a unique name for the slab as a target
5716 name
= create_unique_id(s
);
5719 s
->kobj
.kset
= kset
;
5720 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5724 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5729 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5730 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5731 if (!s
->memcg_kset
) {
5738 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5740 /* Setup first alias */
5741 sysfs_slab_alias(s
, s
->name
);
5748 kobject_del(&s
->kobj
);
5752 static void sysfs_slab_remove(struct kmem_cache
*s
)
5754 if (slab_state
< FULL
)
5756 * Sysfs has not been setup yet so no need to remove the
5761 kobject_get(&s
->kobj
);
5762 schedule_work(&s
->kobj_remove_work
);
5765 void sysfs_slab_unlink(struct kmem_cache
*s
)
5767 if (slab_state
>= FULL
)
5768 kobject_del(&s
->kobj
);
5771 void sysfs_slab_release(struct kmem_cache
*s
)
5773 if (slab_state
>= FULL
)
5774 kobject_put(&s
->kobj
);
5778 * Need to buffer aliases during bootup until sysfs becomes
5779 * available lest we lose that information.
5781 struct saved_alias
{
5782 struct kmem_cache
*s
;
5784 struct saved_alias
*next
;
5787 static struct saved_alias
*alias_list
;
5789 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5791 struct saved_alias
*al
;
5793 if (slab_state
== FULL
) {
5795 * If we have a leftover link then remove it.
5797 sysfs_remove_link(&slab_kset
->kobj
, name
);
5798 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5801 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5807 al
->next
= alias_list
;
5812 static int __init
slab_sysfs_init(void)
5814 struct kmem_cache
*s
;
5817 mutex_lock(&slab_mutex
);
5819 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5821 mutex_unlock(&slab_mutex
);
5822 pr_err("Cannot register slab subsystem.\n");
5828 list_for_each_entry(s
, &slab_caches
, list
) {
5829 err
= sysfs_slab_add(s
);
5831 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5835 while (alias_list
) {
5836 struct saved_alias
*al
= alias_list
;
5838 alias_list
= alias_list
->next
;
5839 err
= sysfs_slab_alias(al
->s
, al
->name
);
5841 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5846 mutex_unlock(&slab_mutex
);
5851 __initcall(slab_sysfs_init
);
5852 #endif /* CONFIG_SYSFS */
5855 * The /proc/slabinfo ABI
5857 #ifdef CONFIG_SLUB_DEBUG
5858 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5860 unsigned long nr_slabs
= 0;
5861 unsigned long nr_objs
= 0;
5862 unsigned long nr_free
= 0;
5864 struct kmem_cache_node
*n
;
5866 for_each_kmem_cache_node(s
, node
, n
) {
5867 nr_slabs
+= node_nr_slabs(n
);
5868 nr_objs
+= node_nr_objs(n
);
5869 nr_free
+= count_partial(n
, count_free
);
5872 sinfo
->active_objs
= nr_objs
- nr_free
;
5873 sinfo
->num_objs
= nr_objs
;
5874 sinfo
->active_slabs
= nr_slabs
;
5875 sinfo
->num_slabs
= nr_slabs
;
5876 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5877 sinfo
->cache_order
= oo_order(s
->oo
);
5880 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5884 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5885 size_t count
, loff_t
*ppos
)
5889 #endif /* CONFIG_SLUB_DEBUG */