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
, unsigned long pr_time
)
606 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
607 s
, (void *)t
->addr
, pr_time
- 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 unsigned long pr_time
= jiffies
;
623 if (!(s
->flags
& SLAB_STORE_USER
))
626 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
627 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
630 static void print_page_info(struct page
*page
)
632 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
633 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
637 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
639 struct va_format vaf
;
645 pr_err("=============================================================================\n");
646 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
647 pr_err("-----------------------------------------------------------------------------\n\n");
649 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
653 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
655 struct va_format vaf
;
661 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
665 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
667 unsigned int off
; /* Offset of last byte */
668 u8
*addr
= page_address(page
);
670 print_tracking(s
, p
);
672 print_page_info(page
);
674 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
675 p
, p
- addr
, get_freepointer(s
, p
));
677 if (s
->flags
& SLAB_RED_ZONE
)
678 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
680 else if (p
> addr
+ 16)
681 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
683 print_section(KERN_ERR
, "Object ", p
,
684 min_t(unsigned long, s
->object_size
, PAGE_SIZE
));
685 if (s
->flags
& SLAB_RED_ZONE
)
686 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
687 s
->inuse
- s
->object_size
);
690 off
= s
->offset
+ sizeof(void *);
694 if (s
->flags
& SLAB_STORE_USER
)
695 off
+= 2 * sizeof(struct track
);
697 off
+= kasan_metadata_size(s
);
699 if (off
!= size_from_object(s
))
700 /* Beginning of the filler is the free pointer */
701 print_section(KERN_ERR
, "Padding ", p
+ off
,
702 size_from_object(s
) - off
);
707 void object_err(struct kmem_cache
*s
, struct page
*page
,
708 u8
*object
, char *reason
)
710 slab_bug(s
, "%s", reason
);
711 print_trailer(s
, page
, object
);
714 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
715 const char *fmt
, ...)
721 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
723 slab_bug(s
, "%s", buf
);
724 print_page_info(page
);
728 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
732 if (s
->flags
& SLAB_RED_ZONE
)
733 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
735 if (s
->flags
& __OBJECT_POISON
) {
736 memset(p
, POISON_FREE
, s
->object_size
- 1);
737 p
[s
->object_size
- 1] = POISON_END
;
740 if (s
->flags
& SLAB_RED_ZONE
)
741 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
744 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
745 void *from
, void *to
)
747 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
748 memset(from
, data
, to
- from
);
751 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
752 u8
*object
, char *what
,
753 u8
*start
, unsigned int value
, unsigned int bytes
)
758 metadata_access_enable();
759 fault
= memchr_inv(start
, value
, bytes
);
760 metadata_access_disable();
765 while (end
> fault
&& end
[-1] == value
)
768 slab_bug(s
, "%s overwritten", what
);
769 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
770 fault
, end
- 1, fault
[0], value
);
771 print_trailer(s
, page
, object
);
773 restore_bytes(s
, what
, value
, fault
, end
);
781 * Bytes of the object to be managed.
782 * If the freepointer may overlay the object then the free
783 * pointer is the first word of the object.
785 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
788 * object + s->object_size
789 * Padding to reach word boundary. This is also used for Redzoning.
790 * Padding is extended by another word if Redzoning is enabled and
791 * object_size == inuse.
793 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
794 * 0xcc (RED_ACTIVE) for objects in use.
797 * Meta data starts here.
799 * A. Free pointer (if we cannot overwrite object on free)
800 * B. Tracking data for SLAB_STORE_USER
801 * C. Padding to reach required alignment boundary or at mininum
802 * one word if debugging is on to be able to detect writes
803 * before the word boundary.
805 * Padding is done using 0x5a (POISON_INUSE)
808 * Nothing is used beyond s->size.
810 * If slabcaches are merged then the object_size and inuse boundaries are mostly
811 * ignored. And therefore no slab options that rely on these boundaries
812 * may be used with merged slabcaches.
815 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
817 unsigned long off
= s
->inuse
; /* The end of info */
820 /* Freepointer is placed after the object. */
821 off
+= sizeof(void *);
823 if (s
->flags
& SLAB_STORE_USER
)
824 /* We also have user information there */
825 off
+= 2 * sizeof(struct track
);
827 off
+= kasan_metadata_size(s
);
829 if (size_from_object(s
) == off
)
832 return check_bytes_and_report(s
, page
, p
, "Object padding",
833 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
836 /* Check the pad bytes at the end of a slab page */
837 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
846 if (!(s
->flags
& SLAB_POISON
))
849 start
= page_address(page
);
850 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
851 end
= start
+ length
;
852 remainder
= length
% s
->size
;
856 pad
= end
- remainder
;
857 metadata_access_enable();
858 fault
= memchr_inv(pad
, POISON_INUSE
, remainder
);
859 metadata_access_disable();
862 while (end
> fault
&& end
[-1] == POISON_INUSE
)
865 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
866 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
868 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
872 static int check_object(struct kmem_cache
*s
, struct page
*page
,
873 void *object
, u8 val
)
876 u8
*endobject
= object
+ s
->object_size
;
878 if (s
->flags
& SLAB_RED_ZONE
) {
879 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
880 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
883 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
884 endobject
, val
, s
->inuse
- s
->object_size
))
887 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
888 check_bytes_and_report(s
, page
, p
, "Alignment padding",
889 endobject
, POISON_INUSE
,
890 s
->inuse
- s
->object_size
);
894 if (s
->flags
& SLAB_POISON
) {
895 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
896 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
897 POISON_FREE
, s
->object_size
- 1) ||
898 !check_bytes_and_report(s
, page
, p
, "Poison",
899 p
+ s
->object_size
- 1, POISON_END
, 1)))
902 * check_pad_bytes cleans up on its own.
904 check_pad_bytes(s
, page
, p
);
907 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
909 * Object and freepointer overlap. Cannot check
910 * freepointer while object is allocated.
914 /* Check free pointer validity */
915 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
916 object_err(s
, page
, p
, "Freepointer corrupt");
918 * No choice but to zap it and thus lose the remainder
919 * of the free objects in this slab. May cause
920 * another error because the object count is now wrong.
922 set_freepointer(s
, p
, NULL
);
928 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
932 VM_BUG_ON(!irqs_disabled());
934 if (!PageSlab(page
)) {
935 slab_err(s
, page
, "Not a valid slab page");
939 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
940 if (page
->objects
> maxobj
) {
941 slab_err(s
, page
, "objects %u > max %u",
942 page
->objects
, maxobj
);
945 if (page
->inuse
> page
->objects
) {
946 slab_err(s
, page
, "inuse %u > max %u",
947 page
->inuse
, page
->objects
);
950 /* Slab_pad_check fixes things up after itself */
951 slab_pad_check(s
, page
);
956 * Determine if a certain object on a page is on the freelist. Must hold the
957 * slab lock to guarantee that the chains are in a consistent state.
959 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
967 while (fp
&& nr
<= page
->objects
) {
970 if (!check_valid_pointer(s
, page
, fp
)) {
972 object_err(s
, page
, object
,
973 "Freechain corrupt");
974 set_freepointer(s
, object
, NULL
);
976 slab_err(s
, page
, "Freepointer corrupt");
977 page
->freelist
= NULL
;
978 page
->inuse
= page
->objects
;
979 slab_fix(s
, "Freelist cleared");
985 fp
= get_freepointer(s
, object
);
989 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
990 if (max_objects
> MAX_OBJS_PER_PAGE
)
991 max_objects
= MAX_OBJS_PER_PAGE
;
993 if (page
->objects
!= max_objects
) {
994 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
995 page
->objects
, max_objects
);
996 page
->objects
= max_objects
;
997 slab_fix(s
, "Number of objects adjusted.");
999 if (page
->inuse
!= page
->objects
- nr
) {
1000 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1001 page
->inuse
, page
->objects
- nr
);
1002 page
->inuse
= page
->objects
- nr
;
1003 slab_fix(s
, "Object count adjusted.");
1005 return search
== NULL
;
1008 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1011 if (s
->flags
& SLAB_TRACE
) {
1012 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1014 alloc
? "alloc" : "free",
1015 object
, page
->inuse
,
1019 print_section(KERN_INFO
, "Object ", (void *)object
,
1027 * Tracking of fully allocated slabs for debugging purposes.
1029 static void add_full(struct kmem_cache
*s
,
1030 struct kmem_cache_node
*n
, struct page
*page
)
1032 if (!(s
->flags
& SLAB_STORE_USER
))
1035 lockdep_assert_held(&n
->list_lock
);
1036 list_add(&page
->lru
, &n
->full
);
1039 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1041 if (!(s
->flags
& SLAB_STORE_USER
))
1044 lockdep_assert_held(&n
->list_lock
);
1045 list_del(&page
->lru
);
1048 /* Tracking of the number of slabs for debugging purposes */
1049 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1051 struct kmem_cache_node
*n
= get_node(s
, node
);
1053 return atomic_long_read(&n
->nr_slabs
);
1056 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1058 return atomic_long_read(&n
->nr_slabs
);
1061 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1063 struct kmem_cache_node
*n
= get_node(s
, node
);
1066 * May be called early in order to allocate a slab for the
1067 * kmem_cache_node structure. Solve the chicken-egg
1068 * dilemma by deferring the increment of the count during
1069 * bootstrap (see early_kmem_cache_node_alloc).
1072 atomic_long_inc(&n
->nr_slabs
);
1073 atomic_long_add(objects
, &n
->total_objects
);
1076 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1078 struct kmem_cache_node
*n
= get_node(s
, node
);
1080 atomic_long_dec(&n
->nr_slabs
);
1081 atomic_long_sub(objects
, &n
->total_objects
);
1084 /* Object debug checks for alloc/free paths */
1085 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1088 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1091 init_object(s
, object
, SLUB_RED_INACTIVE
);
1092 init_tracking(s
, object
);
1095 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1097 void *object
, unsigned long addr
)
1099 if (!check_slab(s
, page
))
1102 if (!check_valid_pointer(s
, page
, object
)) {
1103 object_err(s
, page
, object
, "Freelist Pointer check fails");
1107 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1113 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1115 void *object
, unsigned long addr
)
1117 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1118 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1122 /* Success perform special debug activities for allocs */
1123 if (s
->flags
& SLAB_STORE_USER
)
1124 set_track(s
, object
, TRACK_ALLOC
, addr
);
1125 trace(s
, page
, object
, 1);
1126 init_object(s
, object
, SLUB_RED_ACTIVE
);
1130 if (PageSlab(page
)) {
1132 * If this is a slab page then lets do the best we can
1133 * to avoid issues in the future. Marking all objects
1134 * as used avoids touching the remaining objects.
1136 slab_fix(s
, "Marking all objects used");
1137 page
->inuse
= page
->objects
;
1138 page
->freelist
= NULL
;
1143 static inline int free_consistency_checks(struct kmem_cache
*s
,
1144 struct page
*page
, void *object
, unsigned long addr
)
1146 if (!check_valid_pointer(s
, page
, object
)) {
1147 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1151 if (on_freelist(s
, page
, object
)) {
1152 object_err(s
, page
, object
, "Object already free");
1156 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1159 if (unlikely(s
!= page
->slab_cache
)) {
1160 if (!PageSlab(page
)) {
1161 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1163 } else if (!page
->slab_cache
) {
1164 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1168 object_err(s
, page
, object
,
1169 "page slab pointer corrupt.");
1175 /* Supports checking bulk free of a constructed freelist */
1176 static noinline
int free_debug_processing(
1177 struct kmem_cache
*s
, struct page
*page
,
1178 void *head
, void *tail
, int bulk_cnt
,
1181 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1182 void *object
= head
;
1184 unsigned long uninitialized_var(flags
);
1187 spin_lock_irqsave(&n
->list_lock
, flags
);
1190 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1191 if (!check_slab(s
, page
))
1198 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1199 if (!free_consistency_checks(s
, page
, object
, addr
))
1203 if (s
->flags
& SLAB_STORE_USER
)
1204 set_track(s
, object
, TRACK_FREE
, addr
);
1205 trace(s
, page
, object
, 0);
1206 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1207 init_object(s
, object
, SLUB_RED_INACTIVE
);
1209 /* Reached end of constructed freelist yet? */
1210 if (object
!= tail
) {
1211 object
= get_freepointer(s
, object
);
1217 if (cnt
!= bulk_cnt
)
1218 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1222 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1224 slab_fix(s
, "Object at 0x%p not freed", object
);
1228 static int __init
setup_slub_debug(char *str
)
1230 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1231 if (*str
++ != '=' || !*str
)
1233 * No options specified. Switch on full debugging.
1239 * No options but restriction on slabs. This means full
1240 * debugging for slabs matching a pattern.
1247 * Switch off all debugging measures.
1252 * Determine which debug features should be switched on
1254 for (; *str
&& *str
!= ','; str
++) {
1255 switch (tolower(*str
)) {
1257 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1260 slub_debug
|= SLAB_RED_ZONE
;
1263 slub_debug
|= SLAB_POISON
;
1266 slub_debug
|= SLAB_STORE_USER
;
1269 slub_debug
|= SLAB_TRACE
;
1272 slub_debug
|= SLAB_FAILSLAB
;
1276 * Avoid enabling debugging on caches if its minimum
1277 * order would increase as a result.
1279 disable_higher_order_debug
= 1;
1282 pr_err("slub_debug option '%c' unknown. skipped\n",
1289 slub_debug_slabs
= str
+ 1;
1294 __setup("slub_debug", setup_slub_debug
);
1296 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1297 slab_flags_t flags
, const char *name
,
1298 void (*ctor
)(void *))
1301 * Enable debugging if selected on the kernel commandline.
1303 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1304 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1305 flags
|= slub_debug
;
1309 #else /* !CONFIG_SLUB_DEBUG */
1310 static inline void setup_object_debug(struct kmem_cache
*s
,
1311 struct page
*page
, void *object
) {}
1313 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1314 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1316 static inline int free_debug_processing(
1317 struct kmem_cache
*s
, struct page
*page
,
1318 void *head
, void *tail
, int bulk_cnt
,
1319 unsigned long addr
) { return 0; }
1321 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1323 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1324 void *object
, u8 val
) { return 1; }
1325 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1326 struct page
*page
) {}
1327 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1328 struct page
*page
) {}
1329 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1330 slab_flags_t flags
, const char *name
,
1331 void (*ctor
)(void *))
1335 #define slub_debug 0
1337 #define disable_higher_order_debug 0
1339 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1341 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1343 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1345 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1348 #endif /* CONFIG_SLUB_DEBUG */
1351 * Hooks for other subsystems that check memory allocations. In a typical
1352 * production configuration these hooks all should produce no code at all.
1354 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1356 kmemleak_alloc(ptr
, size
, 1, flags
);
1357 kasan_kmalloc_large(ptr
, size
, flags
);
1360 static __always_inline
void kfree_hook(void *x
)
1363 kasan_kfree_large(x
, _RET_IP_
);
1366 static __always_inline
void *slab_free_hook(struct kmem_cache
*s
, void *x
)
1370 kmemleak_free_recursive(x
, s
->flags
);
1373 * Trouble is that we may no longer disable interrupts in the fast path
1374 * So in order to make the debug calls that expect irqs to be
1375 * disabled we need to disable interrupts temporarily.
1377 #ifdef CONFIG_LOCKDEP
1379 unsigned long flags
;
1381 local_irq_save(flags
);
1382 debug_check_no_locks_freed(x
, s
->object_size
);
1383 local_irq_restore(flags
);
1386 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1387 debug_check_no_obj_freed(x
, s
->object_size
);
1389 freeptr
= get_freepointer(s
, x
);
1391 * kasan_slab_free() may put x into memory quarantine, delaying its
1392 * reuse. In this case the object's freelist pointer is changed.
1394 kasan_slab_free(s
, x
, _RET_IP_
);
1398 static inline void slab_free_freelist_hook(struct kmem_cache
*s
,
1399 void *head
, void *tail
)
1402 * Compiler cannot detect this function can be removed if slab_free_hook()
1403 * evaluates to nothing. Thus, catch all relevant config debug options here.
1405 #if defined(CONFIG_LOCKDEP) || \
1406 defined(CONFIG_DEBUG_KMEMLEAK) || \
1407 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1408 defined(CONFIG_KASAN)
1410 void *object
= head
;
1411 void *tail_obj
= tail
? : head
;
1415 freeptr
= slab_free_hook(s
, object
);
1416 } while ((object
!= tail_obj
) && (object
= freeptr
));
1420 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1423 setup_object_debug(s
, page
, object
);
1424 kasan_init_slab_obj(s
, object
);
1425 if (unlikely(s
->ctor
)) {
1426 kasan_unpoison_object_data(s
, object
);
1428 kasan_poison_object_data(s
, object
);
1433 * Slab allocation and freeing
1435 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1436 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1439 int order
= oo_order(oo
);
1441 if (node
== NUMA_NO_NODE
)
1442 page
= alloc_pages(flags
, order
);
1444 page
= __alloc_pages_node(node
, flags
, order
);
1446 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1447 __free_pages(page
, order
);
1454 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1455 /* Pre-initialize the random sequence cache */
1456 static int init_cache_random_seq(struct kmem_cache
*s
)
1459 unsigned long i
, count
= oo_objects(s
->oo
);
1461 /* Bailout if already initialised */
1465 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1467 pr_err("SLUB: Unable to initialize free list for %s\n",
1472 /* Transform to an offset on the set of pages */
1473 if (s
->random_seq
) {
1474 for (i
= 0; i
< count
; i
++)
1475 s
->random_seq
[i
] *= s
->size
;
1480 /* Initialize each random sequence freelist per cache */
1481 static void __init
init_freelist_randomization(void)
1483 struct kmem_cache
*s
;
1485 mutex_lock(&slab_mutex
);
1487 list_for_each_entry(s
, &slab_caches
, list
)
1488 init_cache_random_seq(s
);
1490 mutex_unlock(&slab_mutex
);
1493 /* Get the next entry on the pre-computed freelist randomized */
1494 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1495 unsigned long *pos
, void *start
,
1496 unsigned long page_limit
,
1497 unsigned long freelist_count
)
1502 * If the target page allocation failed, the number of objects on the
1503 * page might be smaller than the usual size defined by the cache.
1506 idx
= s
->random_seq
[*pos
];
1508 if (*pos
>= freelist_count
)
1510 } while (unlikely(idx
>= page_limit
));
1512 return (char *)start
+ idx
;
1515 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1516 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1521 unsigned long idx
, pos
, page_limit
, freelist_count
;
1523 if (page
->objects
< 2 || !s
->random_seq
)
1526 freelist_count
= oo_objects(s
->oo
);
1527 pos
= get_random_int() % freelist_count
;
1529 page_limit
= page
->objects
* s
->size
;
1530 start
= fixup_red_left(s
, page_address(page
));
1532 /* First entry is used as the base of the freelist */
1533 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1535 page
->freelist
= cur
;
1537 for (idx
= 1; idx
< page
->objects
; idx
++) {
1538 setup_object(s
, page
, cur
);
1539 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1541 set_freepointer(s
, cur
, next
);
1544 setup_object(s
, page
, cur
);
1545 set_freepointer(s
, cur
, NULL
);
1550 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1554 static inline void init_freelist_randomization(void) { }
1555 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1559 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1561 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1564 struct kmem_cache_order_objects oo
= s
->oo
;
1570 flags
&= gfp_allowed_mask
;
1572 if (gfpflags_allow_blocking(flags
))
1575 flags
|= s
->allocflags
;
1578 * Let the initial higher-order allocation fail under memory pressure
1579 * so we fall-back to the minimum order allocation.
1581 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1582 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1583 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1585 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1586 if (unlikely(!page
)) {
1590 * Allocation may have failed due to fragmentation.
1591 * Try a lower order alloc if possible
1593 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1594 if (unlikely(!page
))
1596 stat(s
, ORDER_FALLBACK
);
1599 page
->objects
= oo_objects(oo
);
1601 order
= compound_order(page
);
1602 page
->slab_cache
= s
;
1603 __SetPageSlab(page
);
1604 if (page_is_pfmemalloc(page
))
1605 SetPageSlabPfmemalloc(page
);
1607 start
= page_address(page
);
1609 if (unlikely(s
->flags
& SLAB_POISON
))
1610 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1612 kasan_poison_slab(page
);
1614 shuffle
= shuffle_freelist(s
, page
);
1617 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1618 setup_object(s
, page
, p
);
1619 if (likely(idx
< page
->objects
))
1620 set_freepointer(s
, p
, p
+ s
->size
);
1622 set_freepointer(s
, p
, NULL
);
1624 page
->freelist
= fixup_red_left(s
, start
);
1627 page
->inuse
= page
->objects
;
1631 if (gfpflags_allow_blocking(flags
))
1632 local_irq_disable();
1636 mod_lruvec_page_state(page
,
1637 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1638 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1641 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1646 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1648 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1649 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1650 flags
&= ~GFP_SLAB_BUG_MASK
;
1651 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1652 invalid_mask
, &invalid_mask
, flags
, &flags
);
1656 return allocate_slab(s
,
1657 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1660 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1662 int order
= compound_order(page
);
1663 int pages
= 1 << order
;
1665 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1668 slab_pad_check(s
, page
);
1669 for_each_object(p
, s
, page_address(page
),
1671 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1674 mod_lruvec_page_state(page
,
1675 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1676 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1679 __ClearPageSlabPfmemalloc(page
);
1680 __ClearPageSlab(page
);
1682 page_mapcount_reset(page
);
1683 if (current
->reclaim_state
)
1684 current
->reclaim_state
->reclaimed_slab
+= pages
;
1685 memcg_uncharge_slab(page
, order
, s
);
1686 __free_pages(page
, order
);
1689 #define need_reserve_slab_rcu \
1690 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1692 static void rcu_free_slab(struct rcu_head
*h
)
1696 if (need_reserve_slab_rcu
)
1697 page
= virt_to_head_page(h
);
1699 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1701 __free_slab(page
->slab_cache
, page
);
1704 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1706 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1707 struct rcu_head
*head
;
1709 if (need_reserve_slab_rcu
) {
1710 int order
= compound_order(page
);
1711 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1713 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1714 head
= page_address(page
) + offset
;
1716 head
= &page
->rcu_head
;
1719 call_rcu(head
, rcu_free_slab
);
1721 __free_slab(s
, page
);
1724 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1726 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1731 * Management of partially allocated slabs.
1734 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1737 if (tail
== DEACTIVATE_TO_TAIL
)
1738 list_add_tail(&page
->lru
, &n
->partial
);
1740 list_add(&page
->lru
, &n
->partial
);
1743 static inline void add_partial(struct kmem_cache_node
*n
,
1744 struct page
*page
, int tail
)
1746 lockdep_assert_held(&n
->list_lock
);
1747 __add_partial(n
, page
, tail
);
1750 static inline void remove_partial(struct kmem_cache_node
*n
,
1753 lockdep_assert_held(&n
->list_lock
);
1754 list_del(&page
->lru
);
1759 * Remove slab from the partial list, freeze it and
1760 * return the pointer to the freelist.
1762 * Returns a list of objects or NULL if it fails.
1764 static inline void *acquire_slab(struct kmem_cache
*s
,
1765 struct kmem_cache_node
*n
, struct page
*page
,
1766 int mode
, int *objects
)
1769 unsigned long counters
;
1772 lockdep_assert_held(&n
->list_lock
);
1775 * Zap the freelist and set the frozen bit.
1776 * The old freelist is the list of objects for the
1777 * per cpu allocation list.
1779 freelist
= page
->freelist
;
1780 counters
= page
->counters
;
1781 new.counters
= counters
;
1782 *objects
= new.objects
- new.inuse
;
1784 new.inuse
= page
->objects
;
1785 new.freelist
= NULL
;
1787 new.freelist
= freelist
;
1790 VM_BUG_ON(new.frozen
);
1793 if (!__cmpxchg_double_slab(s
, page
,
1795 new.freelist
, new.counters
,
1799 remove_partial(n
, page
);
1804 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1805 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1808 * Try to allocate a partial slab from a specific node.
1810 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1811 struct kmem_cache_cpu
*c
, gfp_t flags
)
1813 struct page
*page
, *page2
;
1814 void *object
= NULL
;
1819 * Racy check. If we mistakenly see no partial slabs then we
1820 * just allocate an empty slab. If we mistakenly try to get a
1821 * partial slab and there is none available then get_partials()
1824 if (!n
|| !n
->nr_partial
)
1827 spin_lock(&n
->list_lock
);
1828 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1831 if (!pfmemalloc_match(page
, flags
))
1834 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1838 available
+= objects
;
1841 stat(s
, ALLOC_FROM_PARTIAL
);
1844 put_cpu_partial(s
, page
, 0);
1845 stat(s
, CPU_PARTIAL_NODE
);
1847 if (!kmem_cache_has_cpu_partial(s
)
1848 || available
> slub_cpu_partial(s
) / 2)
1852 spin_unlock(&n
->list_lock
);
1857 * Get a page from somewhere. Search in increasing NUMA distances.
1859 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1860 struct kmem_cache_cpu
*c
)
1863 struct zonelist
*zonelist
;
1866 enum zone_type high_zoneidx
= gfp_zone(flags
);
1868 unsigned int cpuset_mems_cookie
;
1871 * The defrag ratio allows a configuration of the tradeoffs between
1872 * inter node defragmentation and node local allocations. A lower
1873 * defrag_ratio increases the tendency to do local allocations
1874 * instead of attempting to obtain partial slabs from other nodes.
1876 * If the defrag_ratio is set to 0 then kmalloc() always
1877 * returns node local objects. If the ratio is higher then kmalloc()
1878 * may return off node objects because partial slabs are obtained
1879 * from other nodes and filled up.
1881 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1882 * (which makes defrag_ratio = 1000) then every (well almost)
1883 * allocation will first attempt to defrag slab caches on other nodes.
1884 * This means scanning over all nodes to look for partial slabs which
1885 * may be expensive if we do it every time we are trying to find a slab
1886 * with available objects.
1888 if (!s
->remote_node_defrag_ratio
||
1889 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1893 cpuset_mems_cookie
= read_mems_allowed_begin();
1894 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1895 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1896 struct kmem_cache_node
*n
;
1898 n
= get_node(s
, zone_to_nid(zone
));
1900 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1901 n
->nr_partial
> s
->min_partial
) {
1902 object
= get_partial_node(s
, n
, c
, flags
);
1905 * Don't check read_mems_allowed_retry()
1906 * here - if mems_allowed was updated in
1907 * parallel, that was a harmless race
1908 * between allocation and the cpuset
1915 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1921 * Get a partial page, lock it and return it.
1923 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1924 struct kmem_cache_cpu
*c
)
1927 int searchnode
= node
;
1929 if (node
== NUMA_NO_NODE
)
1930 searchnode
= numa_mem_id();
1931 else if (!node_present_pages(node
))
1932 searchnode
= node_to_mem_node(node
);
1934 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1935 if (object
|| node
!= NUMA_NO_NODE
)
1938 return get_any_partial(s
, flags
, c
);
1941 #ifdef CONFIG_PREEMPT
1943 * Calculate the next globally unique transaction for disambiguiation
1944 * during cmpxchg. The transactions start with the cpu number and are then
1945 * incremented by CONFIG_NR_CPUS.
1947 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1950 * No preemption supported therefore also no need to check for
1956 static inline unsigned long next_tid(unsigned long tid
)
1958 return tid
+ TID_STEP
;
1961 static inline unsigned int tid_to_cpu(unsigned long tid
)
1963 return tid
% TID_STEP
;
1966 static inline unsigned long tid_to_event(unsigned long tid
)
1968 return tid
/ TID_STEP
;
1971 static inline unsigned int init_tid(int cpu
)
1976 static inline void note_cmpxchg_failure(const char *n
,
1977 const struct kmem_cache
*s
, unsigned long tid
)
1979 #ifdef SLUB_DEBUG_CMPXCHG
1980 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1982 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1984 #ifdef CONFIG_PREEMPT
1985 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1986 pr_warn("due to cpu change %d -> %d\n",
1987 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1990 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1991 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1992 tid_to_event(tid
), tid_to_event(actual_tid
));
1994 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1995 actual_tid
, tid
, next_tid(tid
));
1997 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2000 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2004 for_each_possible_cpu(cpu
)
2005 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2009 * Remove the cpu slab
2011 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2012 void *freelist
, struct kmem_cache_cpu
*c
)
2014 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2015 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2017 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2019 int tail
= DEACTIVATE_TO_HEAD
;
2023 if (page
->freelist
) {
2024 stat(s
, DEACTIVATE_REMOTE_FREES
);
2025 tail
= DEACTIVATE_TO_TAIL
;
2029 * Stage one: Free all available per cpu objects back
2030 * to the page freelist while it is still frozen. Leave the
2033 * There is no need to take the list->lock because the page
2036 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2038 unsigned long counters
;
2041 prior
= page
->freelist
;
2042 counters
= page
->counters
;
2043 set_freepointer(s
, freelist
, prior
);
2044 new.counters
= counters
;
2046 VM_BUG_ON(!new.frozen
);
2048 } while (!__cmpxchg_double_slab(s
, page
,
2050 freelist
, new.counters
,
2051 "drain percpu freelist"));
2053 freelist
= nextfree
;
2057 * Stage two: Ensure that the page is unfrozen while the
2058 * list presence reflects the actual number of objects
2061 * We setup the list membership and then perform a cmpxchg
2062 * with the count. If there is a mismatch then the page
2063 * is not unfrozen but the page is on the wrong list.
2065 * Then we restart the process which may have to remove
2066 * the page from the list that we just put it on again
2067 * because the number of objects in the slab may have
2072 old
.freelist
= page
->freelist
;
2073 old
.counters
= page
->counters
;
2074 VM_BUG_ON(!old
.frozen
);
2076 /* Determine target state of the slab */
2077 new.counters
= old
.counters
;
2080 set_freepointer(s
, freelist
, old
.freelist
);
2081 new.freelist
= freelist
;
2083 new.freelist
= old
.freelist
;
2087 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2089 else if (new.freelist
) {
2094 * Taking the spinlock removes the possiblity
2095 * that acquire_slab() will see a slab page that
2098 spin_lock(&n
->list_lock
);
2102 if (kmem_cache_debug(s
) && !lock
) {
2105 * This also ensures that the scanning of full
2106 * slabs from diagnostic functions will not see
2109 spin_lock(&n
->list_lock
);
2117 remove_partial(n
, page
);
2119 else if (l
== M_FULL
)
2121 remove_full(s
, n
, page
);
2123 if (m
== M_PARTIAL
) {
2125 add_partial(n
, page
, tail
);
2128 } else if (m
== M_FULL
) {
2130 stat(s
, DEACTIVATE_FULL
);
2131 add_full(s
, n
, page
);
2137 if (!__cmpxchg_double_slab(s
, page
,
2138 old
.freelist
, old
.counters
,
2139 new.freelist
, new.counters
,
2144 spin_unlock(&n
->list_lock
);
2147 stat(s
, DEACTIVATE_EMPTY
);
2148 discard_slab(s
, page
);
2157 * Unfreeze all the cpu partial slabs.
2159 * This function must be called with interrupts disabled
2160 * for the cpu using c (or some other guarantee must be there
2161 * to guarantee no concurrent accesses).
2163 static void unfreeze_partials(struct kmem_cache
*s
,
2164 struct kmem_cache_cpu
*c
)
2166 #ifdef CONFIG_SLUB_CPU_PARTIAL
2167 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2168 struct page
*page
, *discard_page
= NULL
;
2170 while ((page
= c
->partial
)) {
2174 c
->partial
= page
->next
;
2176 n2
= get_node(s
, page_to_nid(page
));
2179 spin_unlock(&n
->list_lock
);
2182 spin_lock(&n
->list_lock
);
2187 old
.freelist
= page
->freelist
;
2188 old
.counters
= page
->counters
;
2189 VM_BUG_ON(!old
.frozen
);
2191 new.counters
= old
.counters
;
2192 new.freelist
= old
.freelist
;
2196 } while (!__cmpxchg_double_slab(s
, page
,
2197 old
.freelist
, old
.counters
,
2198 new.freelist
, new.counters
,
2199 "unfreezing slab"));
2201 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2202 page
->next
= discard_page
;
2203 discard_page
= page
;
2205 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2206 stat(s
, FREE_ADD_PARTIAL
);
2211 spin_unlock(&n
->list_lock
);
2213 while (discard_page
) {
2214 page
= discard_page
;
2215 discard_page
= discard_page
->next
;
2217 stat(s
, DEACTIVATE_EMPTY
);
2218 discard_slab(s
, page
);
2225 * Put a page that was just frozen (in __slab_free) into a partial page
2226 * slot if available.
2228 * If we did not find a slot then simply move all the partials to the
2229 * per node partial list.
2231 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2233 #ifdef CONFIG_SLUB_CPU_PARTIAL
2234 struct page
*oldpage
;
2242 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2245 pobjects
= oldpage
->pobjects
;
2246 pages
= oldpage
->pages
;
2247 if (drain
&& pobjects
> s
->cpu_partial
) {
2248 unsigned long flags
;
2250 * partial array is full. Move the existing
2251 * set to the per node partial list.
2253 local_irq_save(flags
);
2254 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2255 local_irq_restore(flags
);
2259 stat(s
, CPU_PARTIAL_DRAIN
);
2264 pobjects
+= page
->objects
- page
->inuse
;
2266 page
->pages
= pages
;
2267 page
->pobjects
= pobjects
;
2268 page
->next
= oldpage
;
2270 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2272 if (unlikely(!s
->cpu_partial
)) {
2273 unsigned long flags
;
2275 local_irq_save(flags
);
2276 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2277 local_irq_restore(flags
);
2283 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2285 stat(s
, CPUSLAB_FLUSH
);
2286 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2288 c
->tid
= next_tid(c
->tid
);
2294 * Called from IPI handler with interrupts disabled.
2296 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2298 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2304 unfreeze_partials(s
, c
);
2308 static void flush_cpu_slab(void *d
)
2310 struct kmem_cache
*s
= d
;
2312 __flush_cpu_slab(s
, smp_processor_id());
2315 static bool has_cpu_slab(int cpu
, void *info
)
2317 struct kmem_cache
*s
= info
;
2318 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2320 return c
->page
|| slub_percpu_partial(c
);
2323 static void flush_all(struct kmem_cache
*s
)
2325 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2329 * Use the cpu notifier to insure that the cpu slabs are flushed when
2332 static int slub_cpu_dead(unsigned int cpu
)
2334 struct kmem_cache
*s
;
2335 unsigned long flags
;
2337 mutex_lock(&slab_mutex
);
2338 list_for_each_entry(s
, &slab_caches
, list
) {
2339 local_irq_save(flags
);
2340 __flush_cpu_slab(s
, cpu
);
2341 local_irq_restore(flags
);
2343 mutex_unlock(&slab_mutex
);
2348 * Check if the objects in a per cpu structure fit numa
2349 * locality expectations.
2351 static inline int node_match(struct page
*page
, int node
)
2354 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2360 #ifdef CONFIG_SLUB_DEBUG
2361 static int count_free(struct page
*page
)
2363 return page
->objects
- page
->inuse
;
2366 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2368 return atomic_long_read(&n
->total_objects
);
2370 #endif /* CONFIG_SLUB_DEBUG */
2372 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2373 static unsigned long count_partial(struct kmem_cache_node
*n
,
2374 int (*get_count
)(struct page
*))
2376 unsigned long flags
;
2377 unsigned long x
= 0;
2380 spin_lock_irqsave(&n
->list_lock
, flags
);
2381 list_for_each_entry(page
, &n
->partial
, lru
)
2382 x
+= get_count(page
);
2383 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2386 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2388 static noinline
void
2389 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2391 #ifdef CONFIG_SLUB_DEBUG
2392 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2393 DEFAULT_RATELIMIT_BURST
);
2395 struct kmem_cache_node
*n
;
2397 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2400 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2401 nid
, gfpflags
, &gfpflags
);
2402 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2403 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2406 if (oo_order(s
->min
) > get_order(s
->object_size
))
2407 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2410 for_each_kmem_cache_node(s
, node
, n
) {
2411 unsigned long nr_slabs
;
2412 unsigned long nr_objs
;
2413 unsigned long nr_free
;
2415 nr_free
= count_partial(n
, count_free
);
2416 nr_slabs
= node_nr_slabs(n
);
2417 nr_objs
= node_nr_objs(n
);
2419 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2420 node
, nr_slabs
, nr_objs
, nr_free
);
2425 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2426 int node
, struct kmem_cache_cpu
**pc
)
2429 struct kmem_cache_cpu
*c
= *pc
;
2432 freelist
= get_partial(s
, flags
, node
, c
);
2437 page
= new_slab(s
, flags
, node
);
2439 c
= raw_cpu_ptr(s
->cpu_slab
);
2444 * No other reference to the page yet so we can
2445 * muck around with it freely without cmpxchg
2447 freelist
= page
->freelist
;
2448 page
->freelist
= NULL
;
2450 stat(s
, ALLOC_SLAB
);
2459 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2461 if (unlikely(PageSlabPfmemalloc(page
)))
2462 return gfp_pfmemalloc_allowed(gfpflags
);
2468 * Check the page->freelist of a page and either transfer the freelist to the
2469 * per cpu freelist or deactivate the page.
2471 * The page is still frozen if the return value is not NULL.
2473 * If this function returns NULL then the page has been unfrozen.
2475 * This function must be called with interrupt disabled.
2477 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2480 unsigned long counters
;
2484 freelist
= page
->freelist
;
2485 counters
= page
->counters
;
2487 new.counters
= counters
;
2488 VM_BUG_ON(!new.frozen
);
2490 new.inuse
= page
->objects
;
2491 new.frozen
= freelist
!= NULL
;
2493 } while (!__cmpxchg_double_slab(s
, page
,
2502 * Slow path. The lockless freelist is empty or we need to perform
2505 * Processing is still very fast if new objects have been freed to the
2506 * regular freelist. In that case we simply take over the regular freelist
2507 * as the lockless freelist and zap the regular freelist.
2509 * If that is not working then we fall back to the partial lists. We take the
2510 * first element of the freelist as the object to allocate now and move the
2511 * rest of the freelist to the lockless freelist.
2513 * And if we were unable to get a new slab from the partial slab lists then
2514 * we need to allocate a new slab. This is the slowest path since it involves
2515 * a call to the page allocator and the setup of a new slab.
2517 * Version of __slab_alloc to use when we know that interrupts are
2518 * already disabled (which is the case for bulk allocation).
2520 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2521 unsigned long addr
, struct kmem_cache_cpu
*c
)
2531 if (unlikely(!node_match(page
, node
))) {
2532 int searchnode
= node
;
2534 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2535 searchnode
= node_to_mem_node(node
);
2537 if (unlikely(!node_match(page
, searchnode
))) {
2538 stat(s
, ALLOC_NODE_MISMATCH
);
2539 deactivate_slab(s
, page
, c
->freelist
, c
);
2545 * By rights, we should be searching for a slab page that was
2546 * PFMEMALLOC but right now, we are losing the pfmemalloc
2547 * information when the page leaves the per-cpu allocator
2549 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2550 deactivate_slab(s
, page
, c
->freelist
, c
);
2554 /* must check again c->freelist in case of cpu migration or IRQ */
2555 freelist
= c
->freelist
;
2559 freelist
= get_freelist(s
, page
);
2563 stat(s
, DEACTIVATE_BYPASS
);
2567 stat(s
, ALLOC_REFILL
);
2571 * freelist is pointing to the list of objects to be used.
2572 * page is pointing to the page from which the objects are obtained.
2573 * That page must be frozen for per cpu allocations to work.
2575 VM_BUG_ON(!c
->page
->frozen
);
2576 c
->freelist
= get_freepointer(s
, freelist
);
2577 c
->tid
= next_tid(c
->tid
);
2582 if (slub_percpu_partial(c
)) {
2583 page
= c
->page
= slub_percpu_partial(c
);
2584 slub_set_percpu_partial(c
, page
);
2585 stat(s
, CPU_PARTIAL_ALLOC
);
2589 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2591 if (unlikely(!freelist
)) {
2592 slab_out_of_memory(s
, gfpflags
, node
);
2597 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2600 /* Only entered in the debug case */
2601 if (kmem_cache_debug(s
) &&
2602 !alloc_debug_processing(s
, page
, freelist
, addr
))
2603 goto new_slab
; /* Slab failed checks. Next slab needed */
2605 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2610 * Another one that disabled interrupt and compensates for possible
2611 * cpu changes by refetching the per cpu area pointer.
2613 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2614 unsigned long addr
, struct kmem_cache_cpu
*c
)
2617 unsigned long flags
;
2619 local_irq_save(flags
);
2620 #ifdef CONFIG_PREEMPT
2622 * We may have been preempted and rescheduled on a different
2623 * cpu before disabling interrupts. Need to reload cpu area
2626 c
= this_cpu_ptr(s
->cpu_slab
);
2629 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2630 local_irq_restore(flags
);
2635 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2636 * have the fastpath folded into their functions. So no function call
2637 * overhead for requests that can be satisfied on the fastpath.
2639 * The fastpath works by first checking if the lockless freelist can be used.
2640 * If not then __slab_alloc is called for slow processing.
2642 * Otherwise we can simply pick the next object from the lockless free list.
2644 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2645 gfp_t gfpflags
, int node
, unsigned long addr
)
2648 struct kmem_cache_cpu
*c
;
2652 s
= slab_pre_alloc_hook(s
, gfpflags
);
2657 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2658 * enabled. We may switch back and forth between cpus while
2659 * reading from one cpu area. That does not matter as long
2660 * as we end up on the original cpu again when doing the cmpxchg.
2662 * We should guarantee that tid and kmem_cache are retrieved on
2663 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2664 * to check if it is matched or not.
2667 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2668 c
= raw_cpu_ptr(s
->cpu_slab
);
2669 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2670 unlikely(tid
!= READ_ONCE(c
->tid
)));
2673 * Irqless object alloc/free algorithm used here depends on sequence
2674 * of fetching cpu_slab's data. tid should be fetched before anything
2675 * on c to guarantee that object and page associated with previous tid
2676 * won't be used with current tid. If we fetch tid first, object and
2677 * page could be one associated with next tid and our alloc/free
2678 * request will be failed. In this case, we will retry. So, no problem.
2683 * The transaction ids are globally unique per cpu and per operation on
2684 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2685 * occurs on the right processor and that there was no operation on the
2686 * linked list in between.
2689 object
= c
->freelist
;
2691 if (unlikely(!object
|| !node_match(page
, node
))) {
2692 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2693 stat(s
, ALLOC_SLOWPATH
);
2695 void *next_object
= get_freepointer_safe(s
, object
);
2698 * The cmpxchg will only match if there was no additional
2699 * operation and if we are on the right processor.
2701 * The cmpxchg does the following atomically (without lock
2703 * 1. Relocate first pointer to the current per cpu area.
2704 * 2. Verify that tid and freelist have not been changed
2705 * 3. If they were not changed replace tid and freelist
2707 * Since this is without lock semantics the protection is only
2708 * against code executing on this cpu *not* from access by
2711 if (unlikely(!this_cpu_cmpxchg_double(
2712 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2714 next_object
, next_tid(tid
)))) {
2716 note_cmpxchg_failure("slab_alloc", s
, tid
);
2719 prefetch_freepointer(s
, next_object
);
2720 stat(s
, ALLOC_FASTPATH
);
2723 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2724 memset(object
, 0, s
->object_size
);
2726 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2731 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2732 gfp_t gfpflags
, unsigned long addr
)
2734 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2737 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2739 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2741 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2746 EXPORT_SYMBOL(kmem_cache_alloc
);
2748 #ifdef CONFIG_TRACING
2749 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2751 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2752 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2753 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2756 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2760 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2762 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2764 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2765 s
->object_size
, s
->size
, gfpflags
, node
);
2769 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2771 #ifdef CONFIG_TRACING
2772 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2774 int node
, size_t size
)
2776 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2778 trace_kmalloc_node(_RET_IP_
, ret
,
2779 size
, s
->size
, gfpflags
, node
);
2781 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2784 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2789 * Slow path handling. This may still be called frequently since objects
2790 * have a longer lifetime than the cpu slabs in most processing loads.
2792 * So we still attempt to reduce cache line usage. Just take the slab
2793 * lock and free the item. If there is no additional partial page
2794 * handling required then we can return immediately.
2796 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2797 void *head
, void *tail
, int cnt
,
2804 unsigned long counters
;
2805 struct kmem_cache_node
*n
= NULL
;
2806 unsigned long uninitialized_var(flags
);
2808 stat(s
, FREE_SLOWPATH
);
2810 if (kmem_cache_debug(s
) &&
2811 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2816 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2819 prior
= page
->freelist
;
2820 counters
= page
->counters
;
2821 set_freepointer(s
, tail
, prior
);
2822 new.counters
= counters
;
2823 was_frozen
= new.frozen
;
2825 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2827 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2830 * Slab was on no list before and will be
2832 * We can defer the list move and instead
2837 } else { /* Needs to be taken off a list */
2839 n
= get_node(s
, page_to_nid(page
));
2841 * Speculatively acquire the list_lock.
2842 * If the cmpxchg does not succeed then we may
2843 * drop the list_lock without any processing.
2845 * Otherwise the list_lock will synchronize with
2846 * other processors updating the list of slabs.
2848 spin_lock_irqsave(&n
->list_lock
, flags
);
2853 } while (!cmpxchg_double_slab(s
, page
,
2861 * If we just froze the page then put it onto the
2862 * per cpu partial list.
2864 if (new.frozen
&& !was_frozen
) {
2865 put_cpu_partial(s
, page
, 1);
2866 stat(s
, CPU_PARTIAL_FREE
);
2869 * The list lock was not taken therefore no list
2870 * activity can be necessary.
2873 stat(s
, FREE_FROZEN
);
2877 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2881 * Objects left in the slab. If it was not on the partial list before
2884 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2885 if (kmem_cache_debug(s
))
2886 remove_full(s
, n
, page
);
2887 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2888 stat(s
, FREE_ADD_PARTIAL
);
2890 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2896 * Slab on the partial list.
2898 remove_partial(n
, page
);
2899 stat(s
, FREE_REMOVE_PARTIAL
);
2901 /* Slab must be on the full list */
2902 remove_full(s
, n
, page
);
2905 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2907 discard_slab(s
, page
);
2911 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2912 * can perform fastpath freeing without additional function calls.
2914 * The fastpath is only possible if we are freeing to the current cpu slab
2915 * of this processor. This typically the case if we have just allocated
2918 * If fastpath is not possible then fall back to __slab_free where we deal
2919 * with all sorts of special processing.
2921 * Bulk free of a freelist with several objects (all pointing to the
2922 * same page) possible by specifying head and tail ptr, plus objects
2923 * count (cnt). Bulk free indicated by tail pointer being set.
2925 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2926 struct page
*page
, void *head
, void *tail
,
2927 int cnt
, unsigned long addr
)
2929 void *tail_obj
= tail
? : head
;
2930 struct kmem_cache_cpu
*c
;
2934 * Determine the currently cpus per cpu slab.
2935 * The cpu may change afterward. However that does not matter since
2936 * data is retrieved via this pointer. If we are on the same cpu
2937 * during the cmpxchg then the free will succeed.
2940 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2941 c
= raw_cpu_ptr(s
->cpu_slab
);
2942 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2943 unlikely(tid
!= READ_ONCE(c
->tid
)));
2945 /* Same with comment on barrier() in slab_alloc_node() */
2948 if (likely(page
== c
->page
)) {
2949 set_freepointer(s
, tail_obj
, c
->freelist
);
2951 if (unlikely(!this_cpu_cmpxchg_double(
2952 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2954 head
, next_tid(tid
)))) {
2956 note_cmpxchg_failure("slab_free", s
, tid
);
2959 stat(s
, FREE_FASTPATH
);
2961 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2965 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2966 void *head
, void *tail
, int cnt
,
2969 slab_free_freelist_hook(s
, head
, tail
);
2971 * slab_free_freelist_hook() could have put the items into quarantine.
2972 * If so, no need to free them.
2974 if (s
->flags
& SLAB_KASAN
&& !(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
2976 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2980 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2982 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
2986 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2988 s
= cache_from_obj(s
, x
);
2991 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
2992 trace_kmem_cache_free(_RET_IP_
, x
);
2994 EXPORT_SYMBOL(kmem_cache_free
);
2996 struct detached_freelist
{
3001 struct kmem_cache
*s
;
3005 * This function progressively scans the array with free objects (with
3006 * a limited look ahead) and extract objects belonging to the same
3007 * page. It builds a detached freelist directly within the given
3008 * page/objects. This can happen without any need for
3009 * synchronization, because the objects are owned by running process.
3010 * The freelist is build up as a single linked list in the objects.
3011 * The idea is, that this detached freelist can then be bulk
3012 * transferred to the real freelist(s), but only requiring a single
3013 * synchronization primitive. Look ahead in the array is limited due
3014 * to performance reasons.
3017 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3018 void **p
, struct detached_freelist
*df
)
3020 size_t first_skipped_index
= 0;
3025 /* Always re-init detached_freelist */
3030 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3031 } while (!object
&& size
);
3036 page
= virt_to_head_page(object
);
3038 /* Handle kalloc'ed objects */
3039 if (unlikely(!PageSlab(page
))) {
3040 BUG_ON(!PageCompound(page
));
3042 __free_pages(page
, compound_order(page
));
3043 p
[size
] = NULL
; /* mark object processed */
3046 /* Derive kmem_cache from object */
3047 df
->s
= page
->slab_cache
;
3049 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3052 /* Start new detached freelist */
3054 set_freepointer(df
->s
, object
, NULL
);
3056 df
->freelist
= object
;
3057 p
[size
] = NULL
; /* mark object processed */
3063 continue; /* Skip processed objects */
3065 /* df->page is always set at this point */
3066 if (df
->page
== virt_to_head_page(object
)) {
3067 /* Opportunity build freelist */
3068 set_freepointer(df
->s
, object
, df
->freelist
);
3069 df
->freelist
= object
;
3071 p
[size
] = NULL
; /* mark object processed */
3076 /* Limit look ahead search */
3080 if (!first_skipped_index
)
3081 first_skipped_index
= size
+ 1;
3084 return first_skipped_index
;
3087 /* Note that interrupts must be enabled when calling this function. */
3088 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3094 struct detached_freelist df
;
3096 size
= build_detached_freelist(s
, size
, p
, &df
);
3100 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3101 } while (likely(size
));
3103 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3105 /* Note that interrupts must be enabled when calling this function. */
3106 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3109 struct kmem_cache_cpu
*c
;
3112 /* memcg and kmem_cache debug support */
3113 s
= slab_pre_alloc_hook(s
, flags
);
3117 * Drain objects in the per cpu slab, while disabling local
3118 * IRQs, which protects against PREEMPT and interrupts
3119 * handlers invoking normal fastpath.
3121 local_irq_disable();
3122 c
= this_cpu_ptr(s
->cpu_slab
);
3124 for (i
= 0; i
< size
; i
++) {
3125 void *object
= c
->freelist
;
3127 if (unlikely(!object
)) {
3129 * Invoking slow path likely have side-effect
3130 * of re-populating per CPU c->freelist
3132 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3134 if (unlikely(!p
[i
]))
3137 c
= this_cpu_ptr(s
->cpu_slab
);
3138 continue; /* goto for-loop */
3140 c
->freelist
= get_freepointer(s
, object
);
3143 c
->tid
= next_tid(c
->tid
);
3146 /* Clear memory outside IRQ disabled fastpath loop */
3147 if (unlikely(flags
& __GFP_ZERO
)) {
3150 for (j
= 0; j
< i
; j
++)
3151 memset(p
[j
], 0, s
->object_size
);
3154 /* memcg and kmem_cache debug support */
3155 slab_post_alloc_hook(s
, flags
, size
, p
);
3159 slab_post_alloc_hook(s
, flags
, i
, p
);
3160 __kmem_cache_free_bulk(s
, i
, p
);
3163 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3167 * Object placement in a slab is made very easy because we always start at
3168 * offset 0. If we tune the size of the object to the alignment then we can
3169 * get the required alignment by putting one properly sized object after
3172 * Notice that the allocation order determines the sizes of the per cpu
3173 * caches. Each processor has always one slab available for allocations.
3174 * Increasing the allocation order reduces the number of times that slabs
3175 * must be moved on and off the partial lists and is therefore a factor in
3180 * Mininum / Maximum order of slab pages. This influences locking overhead
3181 * and slab fragmentation. A higher order reduces the number of partial slabs
3182 * and increases the number of allocations possible without having to
3183 * take the list_lock.
3185 static int slub_min_order
;
3186 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3187 static int slub_min_objects
;
3190 * Calculate the order of allocation given an slab object size.
3192 * The order of allocation has significant impact on performance and other
3193 * system components. Generally order 0 allocations should be preferred since
3194 * order 0 does not cause fragmentation in the page allocator. Larger objects
3195 * be problematic to put into order 0 slabs because there may be too much
3196 * unused space left. We go to a higher order if more than 1/16th of the slab
3199 * In order to reach satisfactory performance we must ensure that a minimum
3200 * number of objects is in one slab. Otherwise we may generate too much
3201 * activity on the partial lists which requires taking the list_lock. This is
3202 * less a concern for large slabs though which are rarely used.
3204 * slub_max_order specifies the order where we begin to stop considering the
3205 * number of objects in a slab as critical. If we reach slub_max_order then
3206 * we try to keep the page order as low as possible. So we accept more waste
3207 * of space in favor of a small page order.
3209 * Higher order allocations also allow the placement of more objects in a
3210 * slab and thereby reduce object handling overhead. If the user has
3211 * requested a higher mininum order then we start with that one instead of
3212 * the smallest order which will fit the object.
3214 static inline int slab_order(int size
, int min_objects
,
3215 int max_order
, int fract_leftover
, int reserved
)
3219 int min_order
= slub_min_order
;
3221 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3222 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3224 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3225 order
<= max_order
; order
++) {
3227 unsigned long slab_size
= PAGE_SIZE
<< order
;
3229 rem
= (slab_size
- reserved
) % size
;
3231 if (rem
<= slab_size
/ fract_leftover
)
3238 static inline int calculate_order(int size
, int reserved
)
3246 * Attempt to find best configuration for a slab. This
3247 * works by first attempting to generate a layout with
3248 * the best configuration and backing off gradually.
3250 * First we increase the acceptable waste in a slab. Then
3251 * we reduce the minimum objects required in a slab.
3253 min_objects
= slub_min_objects
;
3255 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3256 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3257 min_objects
= min(min_objects
, max_objects
);
3259 while (min_objects
> 1) {
3261 while (fraction
>= 4) {
3262 order
= slab_order(size
, min_objects
,
3263 slub_max_order
, fraction
, reserved
);
3264 if (order
<= slub_max_order
)
3272 * We were unable to place multiple objects in a slab. Now
3273 * lets see if we can place a single object there.
3275 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3276 if (order
<= slub_max_order
)
3280 * Doh this slab cannot be placed using slub_max_order.
3282 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3283 if (order
< MAX_ORDER
)
3289 init_kmem_cache_node(struct kmem_cache_node
*n
)
3292 spin_lock_init(&n
->list_lock
);
3293 INIT_LIST_HEAD(&n
->partial
);
3294 #ifdef CONFIG_SLUB_DEBUG
3295 atomic_long_set(&n
->nr_slabs
, 0);
3296 atomic_long_set(&n
->total_objects
, 0);
3297 INIT_LIST_HEAD(&n
->full
);
3301 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3303 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3304 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3307 * Must align to double word boundary for the double cmpxchg
3308 * instructions to work; see __pcpu_double_call_return_bool().
3310 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3311 2 * sizeof(void *));
3316 init_kmem_cache_cpus(s
);
3321 static struct kmem_cache
*kmem_cache_node
;
3324 * No kmalloc_node yet so do it by hand. We know that this is the first
3325 * slab on the node for this slabcache. There are no concurrent accesses
3328 * Note that this function only works on the kmem_cache_node
3329 * when allocating for the kmem_cache_node. This is used for bootstrapping
3330 * memory on a fresh node that has no slab structures yet.
3332 static void early_kmem_cache_node_alloc(int node
)
3335 struct kmem_cache_node
*n
;
3337 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3339 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3342 if (page_to_nid(page
) != node
) {
3343 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3344 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3349 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3352 kmem_cache_node
->node
[node
] = n
;
3353 #ifdef CONFIG_SLUB_DEBUG
3354 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3355 init_tracking(kmem_cache_node
, n
);
3357 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3359 init_kmem_cache_node(n
);
3360 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3363 * No locks need to be taken here as it has just been
3364 * initialized and there is no concurrent access.
3366 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3369 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3372 struct kmem_cache_node
*n
;
3374 for_each_kmem_cache_node(s
, node
, n
) {
3375 s
->node
[node
] = NULL
;
3376 kmem_cache_free(kmem_cache_node
, n
);
3380 void __kmem_cache_release(struct kmem_cache
*s
)
3382 cache_random_seq_destroy(s
);
3383 free_percpu(s
->cpu_slab
);
3384 free_kmem_cache_nodes(s
);
3387 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3391 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3392 struct kmem_cache_node
*n
;
3394 if (slab_state
== DOWN
) {
3395 early_kmem_cache_node_alloc(node
);
3398 n
= kmem_cache_alloc_node(kmem_cache_node
,
3402 free_kmem_cache_nodes(s
);
3406 init_kmem_cache_node(n
);
3412 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3414 if (min
< MIN_PARTIAL
)
3416 else if (min
> MAX_PARTIAL
)
3418 s
->min_partial
= min
;
3421 static void set_cpu_partial(struct kmem_cache
*s
)
3423 #ifdef CONFIG_SLUB_CPU_PARTIAL
3425 * cpu_partial determined the maximum number of objects kept in the
3426 * per cpu partial lists of a processor.
3428 * Per cpu partial lists mainly contain slabs that just have one
3429 * object freed. If they are used for allocation then they can be
3430 * filled up again with minimal effort. The slab will never hit the
3431 * per node partial lists and therefore no locking will be required.
3433 * This setting also determines
3435 * A) The number of objects from per cpu partial slabs dumped to the
3436 * per node list when we reach the limit.
3437 * B) The number of objects in cpu partial slabs to extract from the
3438 * per node list when we run out of per cpu objects. We only fetch
3439 * 50% to keep some capacity around for frees.
3441 if (!kmem_cache_has_cpu_partial(s
))
3443 else if (s
->size
>= PAGE_SIZE
)
3445 else if (s
->size
>= 1024)
3447 else if (s
->size
>= 256)
3448 s
->cpu_partial
= 13;
3450 s
->cpu_partial
= 30;
3455 * calculate_sizes() determines the order and the distribution of data within
3458 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3460 slab_flags_t flags
= s
->flags
;
3461 size_t size
= s
->object_size
;
3465 * Round up object size to the next word boundary. We can only
3466 * place the free pointer at word boundaries and this determines
3467 * the possible location of the free pointer.
3469 size
= ALIGN(size
, sizeof(void *));
3471 #ifdef CONFIG_SLUB_DEBUG
3473 * Determine if we can poison the object itself. If the user of
3474 * the slab may touch the object after free or before allocation
3475 * then we should never poison the object itself.
3477 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3479 s
->flags
|= __OBJECT_POISON
;
3481 s
->flags
&= ~__OBJECT_POISON
;
3485 * If we are Redzoning then check if there is some space between the
3486 * end of the object and the free pointer. If not then add an
3487 * additional word to have some bytes to store Redzone information.
3489 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3490 size
+= sizeof(void *);
3494 * With that we have determined the number of bytes in actual use
3495 * by the object. This is the potential offset to the free pointer.
3499 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3502 * Relocate free pointer after the object if it is not
3503 * permitted to overwrite the first word of the object on
3506 * This is the case if we do RCU, have a constructor or
3507 * destructor or are poisoning the objects.
3510 size
+= sizeof(void *);
3513 #ifdef CONFIG_SLUB_DEBUG
3514 if (flags
& SLAB_STORE_USER
)
3516 * Need to store information about allocs and frees after
3519 size
+= 2 * sizeof(struct track
);
3522 kasan_cache_create(s
, &size
, &s
->flags
);
3523 #ifdef CONFIG_SLUB_DEBUG
3524 if (flags
& SLAB_RED_ZONE
) {
3526 * Add some empty padding so that we can catch
3527 * overwrites from earlier objects rather than let
3528 * tracking information or the free pointer be
3529 * corrupted if a user writes before the start
3532 size
+= sizeof(void *);
3534 s
->red_left_pad
= sizeof(void *);
3535 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3536 size
+= s
->red_left_pad
;
3541 * SLUB stores one object immediately after another beginning from
3542 * offset 0. In order to align the objects we have to simply size
3543 * each object to conform to the alignment.
3545 size
= ALIGN(size
, s
->align
);
3547 if (forced_order
>= 0)
3548 order
= forced_order
;
3550 order
= calculate_order(size
, s
->reserved
);
3557 s
->allocflags
|= __GFP_COMP
;
3559 if (s
->flags
& SLAB_CACHE_DMA
)
3560 s
->allocflags
|= GFP_DMA
;
3562 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3563 s
->allocflags
|= __GFP_RECLAIMABLE
;
3566 * Determine the number of objects per slab
3568 s
->oo
= oo_make(order
, size
, s
->reserved
);
3569 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3570 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3573 return !!oo_objects(s
->oo
);
3576 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3578 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3580 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3581 s
->random
= get_random_long();
3584 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_TYPESAFE_BY_RCU
))
3585 s
->reserved
= sizeof(struct rcu_head
);
3587 if (!calculate_sizes(s
, -1))
3589 if (disable_higher_order_debug
) {
3591 * Disable debugging flags that store metadata if the min slab
3594 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3595 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3597 if (!calculate_sizes(s
, -1))
3602 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3603 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3604 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3605 /* Enable fast mode */
3606 s
->flags
|= __CMPXCHG_DOUBLE
;
3610 * The larger the object size is, the more pages we want on the partial
3611 * list to avoid pounding the page allocator excessively.
3613 set_min_partial(s
, ilog2(s
->size
) / 2);
3618 s
->remote_node_defrag_ratio
= 1000;
3621 /* Initialize the pre-computed randomized freelist if slab is up */
3622 if (slab_state
>= UP
) {
3623 if (init_cache_random_seq(s
))
3627 if (!init_kmem_cache_nodes(s
))
3630 if (alloc_kmem_cache_cpus(s
))
3633 free_kmem_cache_nodes(s
);
3635 if (flags
& SLAB_PANIC
)
3636 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3637 s
->name
, (unsigned long)s
->size
, s
->size
,
3638 oo_order(s
->oo
), s
->offset
, (unsigned long)flags
);
3642 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3645 #ifdef CONFIG_SLUB_DEBUG
3646 void *addr
= page_address(page
);
3648 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3649 sizeof(long), GFP_ATOMIC
);
3652 slab_err(s
, page
, text
, s
->name
);
3655 get_map(s
, page
, map
);
3656 for_each_object(p
, s
, addr
, page
->objects
) {
3658 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3659 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3660 print_tracking(s
, p
);
3669 * Attempt to free all partial slabs on a node.
3670 * This is called from __kmem_cache_shutdown(). We must take list_lock
3671 * because sysfs file might still access partial list after the shutdowning.
3673 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3676 struct page
*page
, *h
;
3678 BUG_ON(irqs_disabled());
3679 spin_lock_irq(&n
->list_lock
);
3680 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3682 remove_partial(n
, page
);
3683 list_add(&page
->lru
, &discard
);
3685 list_slab_objects(s
, page
,
3686 "Objects remaining in %s on __kmem_cache_shutdown()");
3689 spin_unlock_irq(&n
->list_lock
);
3691 list_for_each_entry_safe(page
, h
, &discard
, lru
)
3692 discard_slab(s
, page
);
3696 * Release all resources used by a slab cache.
3698 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3701 struct kmem_cache_node
*n
;
3704 /* Attempt to free all objects */
3705 for_each_kmem_cache_node(s
, node
, n
) {
3707 if (n
->nr_partial
|| slabs_node(s
, node
))
3710 sysfs_slab_remove(s
);
3714 /********************************************************************
3716 *******************************************************************/
3718 static int __init
setup_slub_min_order(char *str
)
3720 get_option(&str
, &slub_min_order
);
3725 __setup("slub_min_order=", setup_slub_min_order
);
3727 static int __init
setup_slub_max_order(char *str
)
3729 get_option(&str
, &slub_max_order
);
3730 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3735 __setup("slub_max_order=", setup_slub_max_order
);
3737 static int __init
setup_slub_min_objects(char *str
)
3739 get_option(&str
, &slub_min_objects
);
3744 __setup("slub_min_objects=", setup_slub_min_objects
);
3746 void *__kmalloc(size_t size
, gfp_t flags
)
3748 struct kmem_cache
*s
;
3751 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3752 return kmalloc_large(size
, flags
);
3754 s
= kmalloc_slab(size
, flags
);
3756 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3759 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3761 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3763 kasan_kmalloc(s
, ret
, size
, flags
);
3767 EXPORT_SYMBOL(__kmalloc
);
3770 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3775 flags
|= __GFP_COMP
;
3776 page
= alloc_pages_node(node
, flags
, get_order(size
));
3778 ptr
= page_address(page
);
3780 kmalloc_large_node_hook(ptr
, size
, flags
);
3784 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3786 struct kmem_cache
*s
;
3789 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3790 ret
= kmalloc_large_node(size
, flags
, node
);
3792 trace_kmalloc_node(_RET_IP_
, ret
,
3793 size
, PAGE_SIZE
<< get_order(size
),
3799 s
= kmalloc_slab(size
, flags
);
3801 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3804 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3806 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3808 kasan_kmalloc(s
, ret
, size
, flags
);
3812 EXPORT_SYMBOL(__kmalloc_node
);
3815 #ifdef CONFIG_HARDENED_USERCOPY
3817 * Rejects incorrectly sized objects and objects that are to be copied
3818 * to/from userspace but do not fall entirely within the containing slab
3819 * cache's usercopy region.
3821 * Returns NULL if check passes, otherwise const char * to name of cache
3822 * to indicate an error.
3824 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
3827 struct kmem_cache
*s
;
3828 unsigned long offset
;
3831 /* Find object and usable object size. */
3832 s
= page
->slab_cache
;
3834 /* Reject impossible pointers. */
3835 if (ptr
< page_address(page
))
3836 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
3839 /* Find offset within object. */
3840 offset
= (ptr
- page_address(page
)) % s
->size
;
3842 /* Adjust for redzone and reject if within the redzone. */
3843 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3844 if (offset
< s
->red_left_pad
)
3845 usercopy_abort("SLUB object in left red zone",
3846 s
->name
, to_user
, offset
, n
);
3847 offset
-= s
->red_left_pad
;
3850 /* Allow address range falling entirely within usercopy region. */
3851 if (offset
>= s
->useroffset
&&
3852 offset
- s
->useroffset
<= s
->usersize
&&
3853 n
<= s
->useroffset
- offset
+ s
->usersize
)
3857 * If the copy is still within the allocated object, produce
3858 * a warning instead of rejecting the copy. This is intended
3859 * to be a temporary method to find any missing usercopy
3862 object_size
= slab_ksize(s
);
3863 if (usercopy_fallback
&&
3864 offset
<= object_size
&& n
<= object_size
- offset
) {
3865 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
3869 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
3871 #endif /* CONFIG_HARDENED_USERCOPY */
3873 static size_t __ksize(const void *object
)
3877 if (unlikely(object
== ZERO_SIZE_PTR
))
3880 page
= virt_to_head_page(object
);
3882 if (unlikely(!PageSlab(page
))) {
3883 WARN_ON(!PageCompound(page
));
3884 return PAGE_SIZE
<< compound_order(page
);
3887 return slab_ksize(page
->slab_cache
);
3890 size_t ksize(const void *object
)
3892 size_t size
= __ksize(object
);
3893 /* We assume that ksize callers could use whole allocated area,
3894 * so we need to unpoison this area.
3896 kasan_unpoison_shadow(object
, size
);
3899 EXPORT_SYMBOL(ksize
);
3901 void kfree(const void *x
)
3904 void *object
= (void *)x
;
3906 trace_kfree(_RET_IP_
, x
);
3908 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3911 page
= virt_to_head_page(x
);
3912 if (unlikely(!PageSlab(page
))) {
3913 BUG_ON(!PageCompound(page
));
3915 __free_pages(page
, compound_order(page
));
3918 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3920 EXPORT_SYMBOL(kfree
);
3922 #define SHRINK_PROMOTE_MAX 32
3925 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3926 * up most to the head of the partial lists. New allocations will then
3927 * fill those up and thus they can be removed from the partial lists.
3929 * The slabs with the least items are placed last. This results in them
3930 * being allocated from last increasing the chance that the last objects
3931 * are freed in them.
3933 int __kmem_cache_shrink(struct kmem_cache
*s
)
3937 struct kmem_cache_node
*n
;
3940 struct list_head discard
;
3941 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3942 unsigned long flags
;
3946 for_each_kmem_cache_node(s
, node
, n
) {
3947 INIT_LIST_HEAD(&discard
);
3948 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3949 INIT_LIST_HEAD(promote
+ i
);
3951 spin_lock_irqsave(&n
->list_lock
, flags
);
3954 * Build lists of slabs to discard or promote.
3956 * Note that concurrent frees may occur while we hold the
3957 * list_lock. page->inuse here is the upper limit.
3959 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3960 int free
= page
->objects
- page
->inuse
;
3962 /* Do not reread page->inuse */
3965 /* We do not keep full slabs on the list */
3968 if (free
== page
->objects
) {
3969 list_move(&page
->lru
, &discard
);
3971 } else if (free
<= SHRINK_PROMOTE_MAX
)
3972 list_move(&page
->lru
, promote
+ free
- 1);
3976 * Promote the slabs filled up most to the head of the
3979 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3980 list_splice(promote
+ i
, &n
->partial
);
3982 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3984 /* Release empty slabs */
3985 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3986 discard_slab(s
, page
);
3988 if (slabs_node(s
, node
))
3996 static void kmemcg_cache_deact_after_rcu(struct kmem_cache
*s
)
3999 * Called with all the locks held after a sched RCU grace period.
4000 * Even if @s becomes empty after shrinking, we can't know that @s
4001 * doesn't have allocations already in-flight and thus can't
4002 * destroy @s until the associated memcg is released.
4004 * However, let's remove the sysfs files for empty caches here.
4005 * Each cache has a lot of interface files which aren't
4006 * particularly useful for empty draining caches; otherwise, we can
4007 * easily end up with millions of unnecessary sysfs files on
4008 * systems which have a lot of memory and transient cgroups.
4010 if (!__kmem_cache_shrink(s
))
4011 sysfs_slab_remove(s
);
4014 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4017 * Disable empty slabs caching. Used to avoid pinning offline
4018 * memory cgroups by kmem pages that can be freed.
4020 slub_set_cpu_partial(s
, 0);
4024 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4025 * we have to make sure the change is visible before shrinking.
4027 slab_deactivate_memcg_cache_rcu_sched(s
, kmemcg_cache_deact_after_rcu
);
4031 static int slab_mem_going_offline_callback(void *arg
)
4033 struct kmem_cache
*s
;
4035 mutex_lock(&slab_mutex
);
4036 list_for_each_entry(s
, &slab_caches
, list
)
4037 __kmem_cache_shrink(s
);
4038 mutex_unlock(&slab_mutex
);
4043 static void slab_mem_offline_callback(void *arg
)
4045 struct kmem_cache_node
*n
;
4046 struct kmem_cache
*s
;
4047 struct memory_notify
*marg
= arg
;
4050 offline_node
= marg
->status_change_nid_normal
;
4053 * If the node still has available memory. we need kmem_cache_node
4056 if (offline_node
< 0)
4059 mutex_lock(&slab_mutex
);
4060 list_for_each_entry(s
, &slab_caches
, list
) {
4061 n
= get_node(s
, offline_node
);
4064 * if n->nr_slabs > 0, slabs still exist on the node
4065 * that is going down. We were unable to free them,
4066 * and offline_pages() function shouldn't call this
4067 * callback. So, we must fail.
4069 BUG_ON(slabs_node(s
, offline_node
));
4071 s
->node
[offline_node
] = NULL
;
4072 kmem_cache_free(kmem_cache_node
, n
);
4075 mutex_unlock(&slab_mutex
);
4078 static int slab_mem_going_online_callback(void *arg
)
4080 struct kmem_cache_node
*n
;
4081 struct kmem_cache
*s
;
4082 struct memory_notify
*marg
= arg
;
4083 int nid
= marg
->status_change_nid_normal
;
4087 * If the node's memory is already available, then kmem_cache_node is
4088 * already created. Nothing to do.
4094 * We are bringing a node online. No memory is available yet. We must
4095 * allocate a kmem_cache_node structure in order to bring the node
4098 mutex_lock(&slab_mutex
);
4099 list_for_each_entry(s
, &slab_caches
, list
) {
4101 * XXX: kmem_cache_alloc_node will fallback to other nodes
4102 * since memory is not yet available from the node that
4105 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4110 init_kmem_cache_node(n
);
4114 mutex_unlock(&slab_mutex
);
4118 static int slab_memory_callback(struct notifier_block
*self
,
4119 unsigned long action
, void *arg
)
4124 case MEM_GOING_ONLINE
:
4125 ret
= slab_mem_going_online_callback(arg
);
4127 case MEM_GOING_OFFLINE
:
4128 ret
= slab_mem_going_offline_callback(arg
);
4131 case MEM_CANCEL_ONLINE
:
4132 slab_mem_offline_callback(arg
);
4135 case MEM_CANCEL_OFFLINE
:
4139 ret
= notifier_from_errno(ret
);
4145 static struct notifier_block slab_memory_callback_nb
= {
4146 .notifier_call
= slab_memory_callback
,
4147 .priority
= SLAB_CALLBACK_PRI
,
4150 /********************************************************************
4151 * Basic setup of slabs
4152 *******************************************************************/
4155 * Used for early kmem_cache structures that were allocated using
4156 * the page allocator. Allocate them properly then fix up the pointers
4157 * that may be pointing to the wrong kmem_cache structure.
4160 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4163 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4164 struct kmem_cache_node
*n
;
4166 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4169 * This runs very early, and only the boot processor is supposed to be
4170 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4173 __flush_cpu_slab(s
, smp_processor_id());
4174 for_each_kmem_cache_node(s
, node
, n
) {
4177 list_for_each_entry(p
, &n
->partial
, lru
)
4180 #ifdef CONFIG_SLUB_DEBUG
4181 list_for_each_entry(p
, &n
->full
, lru
)
4185 slab_init_memcg_params(s
);
4186 list_add(&s
->list
, &slab_caches
);
4187 memcg_link_cache(s
);
4191 void __init
kmem_cache_init(void)
4193 static __initdata
struct kmem_cache boot_kmem_cache
,
4194 boot_kmem_cache_node
;
4196 if (debug_guardpage_minorder())
4199 kmem_cache_node
= &boot_kmem_cache_node
;
4200 kmem_cache
= &boot_kmem_cache
;
4202 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4203 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4205 register_hotmemory_notifier(&slab_memory_callback_nb
);
4207 /* Able to allocate the per node structures */
4208 slab_state
= PARTIAL
;
4210 create_boot_cache(kmem_cache
, "kmem_cache",
4211 offsetof(struct kmem_cache
, node
) +
4212 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4213 SLAB_HWCACHE_ALIGN
, 0, 0);
4215 kmem_cache
= bootstrap(&boot_kmem_cache
);
4218 * Allocate kmem_cache_node properly from the kmem_cache slab.
4219 * kmem_cache_node is separately allocated so no need to
4220 * update any list pointers.
4222 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4224 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4225 setup_kmalloc_cache_index_table();
4226 create_kmalloc_caches(0);
4228 /* Setup random freelists for each cache */
4229 init_freelist_randomization();
4231 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4234 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%u, Nodes=%d\n",
4236 slub_min_order
, slub_max_order
, slub_min_objects
,
4237 nr_cpu_ids
, nr_node_ids
);
4240 void __init
kmem_cache_init_late(void)
4245 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4246 slab_flags_t flags
, void (*ctor
)(void *))
4248 struct kmem_cache
*s
, *c
;
4250 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4255 * Adjust the object sizes so that we clear
4256 * the complete object on kzalloc.
4258 s
->object_size
= max(s
->object_size
, (int)size
);
4259 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4261 for_each_memcg_cache(c
, s
) {
4262 c
->object_size
= s
->object_size
;
4263 c
->inuse
= max_t(int, c
->inuse
,
4264 ALIGN(size
, sizeof(void *)));
4267 if (sysfs_slab_alias(s
, name
)) {
4276 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4280 err
= kmem_cache_open(s
, flags
);
4284 /* Mutex is not taken during early boot */
4285 if (slab_state
<= UP
)
4288 memcg_propagate_slab_attrs(s
);
4289 err
= sysfs_slab_add(s
);
4291 __kmem_cache_release(s
);
4296 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4298 struct kmem_cache
*s
;
4301 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4302 return kmalloc_large(size
, gfpflags
);
4304 s
= kmalloc_slab(size
, gfpflags
);
4306 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4309 ret
= slab_alloc(s
, gfpflags
, caller
);
4311 /* Honor the call site pointer we received. */
4312 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4318 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4319 int node
, unsigned long caller
)
4321 struct kmem_cache
*s
;
4324 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4325 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4327 trace_kmalloc_node(caller
, ret
,
4328 size
, PAGE_SIZE
<< get_order(size
),
4334 s
= kmalloc_slab(size
, gfpflags
);
4336 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4339 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4341 /* Honor the call site pointer we received. */
4342 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4349 static int count_inuse(struct page
*page
)
4354 static int count_total(struct page
*page
)
4356 return page
->objects
;
4360 #ifdef CONFIG_SLUB_DEBUG
4361 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4365 void *addr
= page_address(page
);
4367 if (!check_slab(s
, page
) ||
4368 !on_freelist(s
, page
, NULL
))
4371 /* Now we know that a valid freelist exists */
4372 bitmap_zero(map
, page
->objects
);
4374 get_map(s
, page
, map
);
4375 for_each_object(p
, s
, addr
, page
->objects
) {
4376 if (test_bit(slab_index(p
, s
, addr
), map
))
4377 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4381 for_each_object(p
, s
, addr
, page
->objects
)
4382 if (!test_bit(slab_index(p
, s
, addr
), map
))
4383 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4388 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4392 validate_slab(s
, page
, map
);
4396 static int validate_slab_node(struct kmem_cache
*s
,
4397 struct kmem_cache_node
*n
, unsigned long *map
)
4399 unsigned long count
= 0;
4401 unsigned long flags
;
4403 spin_lock_irqsave(&n
->list_lock
, flags
);
4405 list_for_each_entry(page
, &n
->partial
, lru
) {
4406 validate_slab_slab(s
, page
, map
);
4409 if (count
!= n
->nr_partial
)
4410 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4411 s
->name
, count
, n
->nr_partial
);
4413 if (!(s
->flags
& SLAB_STORE_USER
))
4416 list_for_each_entry(page
, &n
->full
, lru
) {
4417 validate_slab_slab(s
, page
, map
);
4420 if (count
!= atomic_long_read(&n
->nr_slabs
))
4421 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4422 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4425 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4429 static long validate_slab_cache(struct kmem_cache
*s
)
4432 unsigned long count
= 0;
4433 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4434 sizeof(unsigned long), GFP_KERNEL
);
4435 struct kmem_cache_node
*n
;
4441 for_each_kmem_cache_node(s
, node
, n
)
4442 count
+= validate_slab_node(s
, n
, map
);
4447 * Generate lists of code addresses where slabcache objects are allocated
4452 unsigned long count
;
4459 DECLARE_BITMAP(cpus
, NR_CPUS
);
4465 unsigned long count
;
4466 struct location
*loc
;
4469 static void free_loc_track(struct loc_track
*t
)
4472 free_pages((unsigned long)t
->loc
,
4473 get_order(sizeof(struct location
) * t
->max
));
4476 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4481 order
= get_order(sizeof(struct location
) * max
);
4483 l
= (void *)__get_free_pages(flags
, order
);
4488 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4496 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4497 const struct track
*track
)
4499 long start
, end
, pos
;
4501 unsigned long caddr
;
4502 unsigned long age
= jiffies
- track
->when
;
4508 pos
= start
+ (end
- start
+ 1) / 2;
4511 * There is nothing at "end". If we end up there
4512 * we need to add something to before end.
4517 caddr
= t
->loc
[pos
].addr
;
4518 if (track
->addr
== caddr
) {
4524 if (age
< l
->min_time
)
4526 if (age
> l
->max_time
)
4529 if (track
->pid
< l
->min_pid
)
4530 l
->min_pid
= track
->pid
;
4531 if (track
->pid
> l
->max_pid
)
4532 l
->max_pid
= track
->pid
;
4534 cpumask_set_cpu(track
->cpu
,
4535 to_cpumask(l
->cpus
));
4537 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4541 if (track
->addr
< caddr
)
4548 * Not found. Insert new tracking element.
4550 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4556 (t
->count
- pos
) * sizeof(struct location
));
4559 l
->addr
= track
->addr
;
4563 l
->min_pid
= track
->pid
;
4564 l
->max_pid
= track
->pid
;
4565 cpumask_clear(to_cpumask(l
->cpus
));
4566 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4567 nodes_clear(l
->nodes
);
4568 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4572 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4573 struct page
*page
, enum track_item alloc
,
4576 void *addr
= page_address(page
);
4579 bitmap_zero(map
, page
->objects
);
4580 get_map(s
, page
, map
);
4582 for_each_object(p
, s
, addr
, page
->objects
)
4583 if (!test_bit(slab_index(p
, s
, addr
), map
))
4584 add_location(t
, s
, get_track(s
, p
, alloc
));
4587 static int list_locations(struct kmem_cache
*s
, char *buf
,
4588 enum track_item alloc
)
4592 struct loc_track t
= { 0, 0, NULL
};
4594 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4595 sizeof(unsigned long), GFP_KERNEL
);
4596 struct kmem_cache_node
*n
;
4598 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4601 return sprintf(buf
, "Out of memory\n");
4603 /* Push back cpu slabs */
4606 for_each_kmem_cache_node(s
, node
, n
) {
4607 unsigned long flags
;
4610 if (!atomic_long_read(&n
->nr_slabs
))
4613 spin_lock_irqsave(&n
->list_lock
, flags
);
4614 list_for_each_entry(page
, &n
->partial
, lru
)
4615 process_slab(&t
, s
, page
, alloc
, map
);
4616 list_for_each_entry(page
, &n
->full
, lru
)
4617 process_slab(&t
, s
, page
, alloc
, map
);
4618 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4621 for (i
= 0; i
< t
.count
; i
++) {
4622 struct location
*l
= &t
.loc
[i
];
4624 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4626 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4629 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4631 len
+= sprintf(buf
+ len
, "<not-available>");
4633 if (l
->sum_time
!= l
->min_time
) {
4634 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4636 (long)div_u64(l
->sum_time
, l
->count
),
4639 len
+= sprintf(buf
+ len
, " age=%ld",
4642 if (l
->min_pid
!= l
->max_pid
)
4643 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4644 l
->min_pid
, l
->max_pid
);
4646 len
+= sprintf(buf
+ len
, " pid=%ld",
4649 if (num_online_cpus() > 1 &&
4650 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4651 len
< PAGE_SIZE
- 60)
4652 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4654 cpumask_pr_args(to_cpumask(l
->cpus
)));
4656 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4657 len
< PAGE_SIZE
- 60)
4658 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4660 nodemask_pr_args(&l
->nodes
));
4662 len
+= sprintf(buf
+ len
, "\n");
4668 len
+= sprintf(buf
, "No data\n");
4673 #ifdef SLUB_RESILIENCY_TEST
4674 static void __init
resiliency_test(void)
4678 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4680 pr_err("SLUB resiliency testing\n");
4681 pr_err("-----------------------\n");
4682 pr_err("A. Corruption after allocation\n");
4684 p
= kzalloc(16, GFP_KERNEL
);
4686 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4689 validate_slab_cache(kmalloc_caches
[4]);
4691 /* Hmmm... The next two are dangerous */
4692 p
= kzalloc(32, GFP_KERNEL
);
4693 p
[32 + sizeof(void *)] = 0x34;
4694 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4696 pr_err("If allocated object is overwritten then not detectable\n\n");
4698 validate_slab_cache(kmalloc_caches
[5]);
4699 p
= kzalloc(64, GFP_KERNEL
);
4700 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4702 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4704 pr_err("If allocated object is overwritten then not detectable\n\n");
4705 validate_slab_cache(kmalloc_caches
[6]);
4707 pr_err("\nB. Corruption after free\n");
4708 p
= kzalloc(128, GFP_KERNEL
);
4711 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4712 validate_slab_cache(kmalloc_caches
[7]);
4714 p
= kzalloc(256, GFP_KERNEL
);
4717 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4718 validate_slab_cache(kmalloc_caches
[8]);
4720 p
= kzalloc(512, GFP_KERNEL
);
4723 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4724 validate_slab_cache(kmalloc_caches
[9]);
4728 static void resiliency_test(void) {};
4733 enum slab_stat_type
{
4734 SL_ALL
, /* All slabs */
4735 SL_PARTIAL
, /* Only partially allocated slabs */
4736 SL_CPU
, /* Only slabs used for cpu caches */
4737 SL_OBJECTS
, /* Determine allocated objects not slabs */
4738 SL_TOTAL
/* Determine object capacity not slabs */
4741 #define SO_ALL (1 << SL_ALL)
4742 #define SO_PARTIAL (1 << SL_PARTIAL)
4743 #define SO_CPU (1 << SL_CPU)
4744 #define SO_OBJECTS (1 << SL_OBJECTS)
4745 #define SO_TOTAL (1 << SL_TOTAL)
4748 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4750 static int __init
setup_slub_memcg_sysfs(char *str
)
4754 if (get_option(&str
, &v
) > 0)
4755 memcg_sysfs_enabled
= v
;
4760 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4763 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4764 char *buf
, unsigned long flags
)
4766 unsigned long total
= 0;
4769 unsigned long *nodes
;
4771 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4775 if (flags
& SO_CPU
) {
4778 for_each_possible_cpu(cpu
) {
4779 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4784 page
= READ_ONCE(c
->page
);
4788 node
= page_to_nid(page
);
4789 if (flags
& SO_TOTAL
)
4791 else if (flags
& SO_OBJECTS
)
4799 page
= slub_percpu_partial_read_once(c
);
4801 node
= page_to_nid(page
);
4802 if (flags
& SO_TOTAL
)
4804 else if (flags
& SO_OBJECTS
)
4815 #ifdef CONFIG_SLUB_DEBUG
4816 if (flags
& SO_ALL
) {
4817 struct kmem_cache_node
*n
;
4819 for_each_kmem_cache_node(s
, node
, n
) {
4821 if (flags
& SO_TOTAL
)
4822 x
= atomic_long_read(&n
->total_objects
);
4823 else if (flags
& SO_OBJECTS
)
4824 x
= atomic_long_read(&n
->total_objects
) -
4825 count_partial(n
, count_free
);
4827 x
= atomic_long_read(&n
->nr_slabs
);
4834 if (flags
& SO_PARTIAL
) {
4835 struct kmem_cache_node
*n
;
4837 for_each_kmem_cache_node(s
, node
, n
) {
4838 if (flags
& SO_TOTAL
)
4839 x
= count_partial(n
, count_total
);
4840 else if (flags
& SO_OBJECTS
)
4841 x
= count_partial(n
, count_inuse
);
4848 x
= sprintf(buf
, "%lu", total
);
4850 for (node
= 0; node
< nr_node_ids
; node
++)
4852 x
+= sprintf(buf
+ x
, " N%d=%lu",
4857 return x
+ sprintf(buf
+ x
, "\n");
4860 #ifdef CONFIG_SLUB_DEBUG
4861 static int any_slab_objects(struct kmem_cache
*s
)
4864 struct kmem_cache_node
*n
;
4866 for_each_kmem_cache_node(s
, node
, n
)
4867 if (atomic_long_read(&n
->total_objects
))
4874 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4875 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4877 struct slab_attribute
{
4878 struct attribute attr
;
4879 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4880 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4883 #define SLAB_ATTR_RO(_name) \
4884 static struct slab_attribute _name##_attr = \
4885 __ATTR(_name, 0400, _name##_show, NULL)
4887 #define SLAB_ATTR(_name) \
4888 static struct slab_attribute _name##_attr = \
4889 __ATTR(_name, 0600, _name##_show, _name##_store)
4891 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4893 return sprintf(buf
, "%d\n", s
->size
);
4895 SLAB_ATTR_RO(slab_size
);
4897 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4899 return sprintf(buf
, "%u\n", s
->align
);
4901 SLAB_ATTR_RO(align
);
4903 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4905 return sprintf(buf
, "%d\n", s
->object_size
);
4907 SLAB_ATTR_RO(object_size
);
4909 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4911 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4913 SLAB_ATTR_RO(objs_per_slab
);
4915 static ssize_t
order_store(struct kmem_cache
*s
,
4916 const char *buf
, size_t length
)
4918 unsigned long order
;
4921 err
= kstrtoul(buf
, 10, &order
);
4925 if (order
> slub_max_order
|| order
< slub_min_order
)
4928 calculate_sizes(s
, order
);
4932 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4934 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4938 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4940 return sprintf(buf
, "%lu\n", s
->min_partial
);
4943 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4949 err
= kstrtoul(buf
, 10, &min
);
4953 set_min_partial(s
, min
);
4956 SLAB_ATTR(min_partial
);
4958 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4960 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4963 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4966 unsigned long objects
;
4969 err
= kstrtoul(buf
, 10, &objects
);
4972 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4975 slub_set_cpu_partial(s
, objects
);
4979 SLAB_ATTR(cpu_partial
);
4981 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4985 return sprintf(buf
, "%pS\n", s
->ctor
);
4989 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4991 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4993 SLAB_ATTR_RO(aliases
);
4995 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4997 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4999 SLAB_ATTR_RO(partial
);
5001 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5003 return show_slab_objects(s
, buf
, SO_CPU
);
5005 SLAB_ATTR_RO(cpu_slabs
);
5007 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5009 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5011 SLAB_ATTR_RO(objects
);
5013 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5015 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5017 SLAB_ATTR_RO(objects_partial
);
5019 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5026 for_each_online_cpu(cpu
) {
5029 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5032 pages
+= page
->pages
;
5033 objects
+= page
->pobjects
;
5037 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5040 for_each_online_cpu(cpu
) {
5043 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5045 if (page
&& len
< PAGE_SIZE
- 20)
5046 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5047 page
->pobjects
, page
->pages
);
5050 return len
+ sprintf(buf
+ len
, "\n");
5052 SLAB_ATTR_RO(slabs_cpu_partial
);
5054 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5056 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5059 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5060 const char *buf
, size_t length
)
5062 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5064 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5067 SLAB_ATTR(reclaim_account
);
5069 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5071 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5073 SLAB_ATTR_RO(hwcache_align
);
5075 #ifdef CONFIG_ZONE_DMA
5076 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5078 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5080 SLAB_ATTR_RO(cache_dma
);
5083 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5085 return sprintf(buf
, "%zu\n", s
->usersize
);
5087 SLAB_ATTR_RO(usersize
);
5089 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5091 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5093 SLAB_ATTR_RO(destroy_by_rcu
);
5095 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
5097 return sprintf(buf
, "%u\n", s
->reserved
);
5099 SLAB_ATTR_RO(reserved
);
5101 #ifdef CONFIG_SLUB_DEBUG
5102 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5104 return show_slab_objects(s
, buf
, SO_ALL
);
5106 SLAB_ATTR_RO(slabs
);
5108 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5110 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5112 SLAB_ATTR_RO(total_objects
);
5114 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5116 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5119 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5120 const char *buf
, size_t length
)
5122 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5123 if (buf
[0] == '1') {
5124 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5125 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5129 SLAB_ATTR(sanity_checks
);
5131 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5133 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5136 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5140 * Tracing a merged cache is going to give confusing results
5141 * as well as cause other issues like converting a mergeable
5142 * cache into an umergeable one.
5144 if (s
->refcount
> 1)
5147 s
->flags
&= ~SLAB_TRACE
;
5148 if (buf
[0] == '1') {
5149 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5150 s
->flags
|= SLAB_TRACE
;
5156 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5158 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5161 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5162 const char *buf
, size_t length
)
5164 if (any_slab_objects(s
))
5167 s
->flags
&= ~SLAB_RED_ZONE
;
5168 if (buf
[0] == '1') {
5169 s
->flags
|= SLAB_RED_ZONE
;
5171 calculate_sizes(s
, -1);
5174 SLAB_ATTR(red_zone
);
5176 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5178 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5181 static ssize_t
poison_store(struct kmem_cache
*s
,
5182 const char *buf
, size_t length
)
5184 if (any_slab_objects(s
))
5187 s
->flags
&= ~SLAB_POISON
;
5188 if (buf
[0] == '1') {
5189 s
->flags
|= SLAB_POISON
;
5191 calculate_sizes(s
, -1);
5196 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5198 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5201 static ssize_t
store_user_store(struct kmem_cache
*s
,
5202 const char *buf
, size_t length
)
5204 if (any_slab_objects(s
))
5207 s
->flags
&= ~SLAB_STORE_USER
;
5208 if (buf
[0] == '1') {
5209 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5210 s
->flags
|= SLAB_STORE_USER
;
5212 calculate_sizes(s
, -1);
5215 SLAB_ATTR(store_user
);
5217 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5222 static ssize_t
validate_store(struct kmem_cache
*s
,
5223 const char *buf
, size_t length
)
5227 if (buf
[0] == '1') {
5228 ret
= validate_slab_cache(s
);
5234 SLAB_ATTR(validate
);
5236 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5238 if (!(s
->flags
& SLAB_STORE_USER
))
5240 return list_locations(s
, buf
, TRACK_ALLOC
);
5242 SLAB_ATTR_RO(alloc_calls
);
5244 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5246 if (!(s
->flags
& SLAB_STORE_USER
))
5248 return list_locations(s
, buf
, TRACK_FREE
);
5250 SLAB_ATTR_RO(free_calls
);
5251 #endif /* CONFIG_SLUB_DEBUG */
5253 #ifdef CONFIG_FAILSLAB
5254 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5256 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5259 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5262 if (s
->refcount
> 1)
5265 s
->flags
&= ~SLAB_FAILSLAB
;
5267 s
->flags
|= SLAB_FAILSLAB
;
5270 SLAB_ATTR(failslab
);
5273 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5278 static ssize_t
shrink_store(struct kmem_cache
*s
,
5279 const char *buf
, size_t length
)
5282 kmem_cache_shrink(s
);
5290 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5292 return sprintf(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5295 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5296 const char *buf
, size_t length
)
5301 err
= kstrtouint(buf
, 10, &ratio
);
5307 s
->remote_node_defrag_ratio
= ratio
* 10;
5311 SLAB_ATTR(remote_node_defrag_ratio
);
5314 #ifdef CONFIG_SLUB_STATS
5315 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5317 unsigned long sum
= 0;
5320 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5325 for_each_online_cpu(cpu
) {
5326 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5332 len
= sprintf(buf
, "%lu", sum
);
5335 for_each_online_cpu(cpu
) {
5336 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5337 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5341 return len
+ sprintf(buf
+ len
, "\n");
5344 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5348 for_each_online_cpu(cpu
)
5349 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5352 #define STAT_ATTR(si, text) \
5353 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5355 return show_stat(s, buf, si); \
5357 static ssize_t text##_store(struct kmem_cache *s, \
5358 const char *buf, size_t length) \
5360 if (buf[0] != '0') \
5362 clear_stat(s, si); \
5367 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5368 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5369 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5370 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5371 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5372 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5373 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5374 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5375 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5376 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5377 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5378 STAT_ATTR(FREE_SLAB
, free_slab
);
5379 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5380 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5381 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5382 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5383 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5384 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5385 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5386 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5387 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5388 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5389 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5390 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5391 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5392 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5395 static struct attribute
*slab_attrs
[] = {
5396 &slab_size_attr
.attr
,
5397 &object_size_attr
.attr
,
5398 &objs_per_slab_attr
.attr
,
5400 &min_partial_attr
.attr
,
5401 &cpu_partial_attr
.attr
,
5403 &objects_partial_attr
.attr
,
5405 &cpu_slabs_attr
.attr
,
5409 &hwcache_align_attr
.attr
,
5410 &reclaim_account_attr
.attr
,
5411 &destroy_by_rcu_attr
.attr
,
5413 &reserved_attr
.attr
,
5414 &slabs_cpu_partial_attr
.attr
,
5415 #ifdef CONFIG_SLUB_DEBUG
5416 &total_objects_attr
.attr
,
5418 &sanity_checks_attr
.attr
,
5420 &red_zone_attr
.attr
,
5422 &store_user_attr
.attr
,
5423 &validate_attr
.attr
,
5424 &alloc_calls_attr
.attr
,
5425 &free_calls_attr
.attr
,
5427 #ifdef CONFIG_ZONE_DMA
5428 &cache_dma_attr
.attr
,
5431 &remote_node_defrag_ratio_attr
.attr
,
5433 #ifdef CONFIG_SLUB_STATS
5434 &alloc_fastpath_attr
.attr
,
5435 &alloc_slowpath_attr
.attr
,
5436 &free_fastpath_attr
.attr
,
5437 &free_slowpath_attr
.attr
,
5438 &free_frozen_attr
.attr
,
5439 &free_add_partial_attr
.attr
,
5440 &free_remove_partial_attr
.attr
,
5441 &alloc_from_partial_attr
.attr
,
5442 &alloc_slab_attr
.attr
,
5443 &alloc_refill_attr
.attr
,
5444 &alloc_node_mismatch_attr
.attr
,
5445 &free_slab_attr
.attr
,
5446 &cpuslab_flush_attr
.attr
,
5447 &deactivate_full_attr
.attr
,
5448 &deactivate_empty_attr
.attr
,
5449 &deactivate_to_head_attr
.attr
,
5450 &deactivate_to_tail_attr
.attr
,
5451 &deactivate_remote_frees_attr
.attr
,
5452 &deactivate_bypass_attr
.attr
,
5453 &order_fallback_attr
.attr
,
5454 &cmpxchg_double_fail_attr
.attr
,
5455 &cmpxchg_double_cpu_fail_attr
.attr
,
5456 &cpu_partial_alloc_attr
.attr
,
5457 &cpu_partial_free_attr
.attr
,
5458 &cpu_partial_node_attr
.attr
,
5459 &cpu_partial_drain_attr
.attr
,
5461 #ifdef CONFIG_FAILSLAB
5462 &failslab_attr
.attr
,
5464 &usersize_attr
.attr
,
5469 static const struct attribute_group slab_attr_group
= {
5470 .attrs
= slab_attrs
,
5473 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5474 struct attribute
*attr
,
5477 struct slab_attribute
*attribute
;
5478 struct kmem_cache
*s
;
5481 attribute
= to_slab_attr(attr
);
5484 if (!attribute
->show
)
5487 err
= attribute
->show(s
, buf
);
5492 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5493 struct attribute
*attr
,
5494 const char *buf
, size_t len
)
5496 struct slab_attribute
*attribute
;
5497 struct kmem_cache
*s
;
5500 attribute
= to_slab_attr(attr
);
5503 if (!attribute
->store
)
5506 err
= attribute
->store(s
, buf
, len
);
5508 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5509 struct kmem_cache
*c
;
5511 mutex_lock(&slab_mutex
);
5512 if (s
->max_attr_size
< len
)
5513 s
->max_attr_size
= len
;
5516 * This is a best effort propagation, so this function's return
5517 * value will be determined by the parent cache only. This is
5518 * basically because not all attributes will have a well
5519 * defined semantics for rollbacks - most of the actions will
5520 * have permanent effects.
5522 * Returning the error value of any of the children that fail
5523 * is not 100 % defined, in the sense that users seeing the
5524 * error code won't be able to know anything about the state of
5527 * Only returning the error code for the parent cache at least
5528 * has well defined semantics. The cache being written to
5529 * directly either failed or succeeded, in which case we loop
5530 * through the descendants with best-effort propagation.
5532 for_each_memcg_cache(c
, s
)
5533 attribute
->store(c
, buf
, len
);
5534 mutex_unlock(&slab_mutex
);
5540 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5544 char *buffer
= NULL
;
5545 struct kmem_cache
*root_cache
;
5547 if (is_root_cache(s
))
5550 root_cache
= s
->memcg_params
.root_cache
;
5553 * This mean this cache had no attribute written. Therefore, no point
5554 * in copying default values around
5556 if (!root_cache
->max_attr_size
)
5559 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5562 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5565 if (!attr
|| !attr
->store
|| !attr
->show
)
5569 * It is really bad that we have to allocate here, so we will
5570 * do it only as a fallback. If we actually allocate, though,
5571 * we can just use the allocated buffer until the end.
5573 * Most of the slub attributes will tend to be very small in
5574 * size, but sysfs allows buffers up to a page, so they can
5575 * theoretically happen.
5579 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5582 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5583 if (WARN_ON(!buffer
))
5588 len
= attr
->show(root_cache
, buf
);
5590 attr
->store(s
, buf
, len
);
5594 free_page((unsigned long)buffer
);
5598 static void kmem_cache_release(struct kobject
*k
)
5600 slab_kmem_cache_release(to_slab(k
));
5603 static const struct sysfs_ops slab_sysfs_ops
= {
5604 .show
= slab_attr_show
,
5605 .store
= slab_attr_store
,
5608 static struct kobj_type slab_ktype
= {
5609 .sysfs_ops
= &slab_sysfs_ops
,
5610 .release
= kmem_cache_release
,
5613 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5615 struct kobj_type
*ktype
= get_ktype(kobj
);
5617 if (ktype
== &slab_ktype
)
5622 static const struct kset_uevent_ops slab_uevent_ops
= {
5623 .filter
= uevent_filter
,
5626 static struct kset
*slab_kset
;
5628 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5631 if (!is_root_cache(s
))
5632 return s
->memcg_params
.root_cache
->memcg_kset
;
5637 #define ID_STR_LENGTH 64
5639 /* Create a unique string id for a slab cache:
5641 * Format :[flags-]size
5643 static char *create_unique_id(struct kmem_cache
*s
)
5645 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5652 * First flags affecting slabcache operations. We will only
5653 * get here for aliasable slabs so we do not need to support
5654 * too many flags. The flags here must cover all flags that
5655 * are matched during merging to guarantee that the id is
5658 if (s
->flags
& SLAB_CACHE_DMA
)
5660 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5662 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5664 if (s
->flags
& SLAB_ACCOUNT
)
5668 p
+= sprintf(p
, "%07d", s
->size
);
5670 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5674 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5676 struct kmem_cache
*s
=
5677 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5679 if (!s
->kobj
.state_in_sysfs
)
5681 * For a memcg cache, this may be called during
5682 * deactivation and again on shutdown. Remove only once.
5683 * A cache is never shut down before deactivation is
5684 * complete, so no need to worry about synchronization.
5689 kset_unregister(s
->memcg_kset
);
5691 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5692 kobject_del(&s
->kobj
);
5694 kobject_put(&s
->kobj
);
5697 static int sysfs_slab_add(struct kmem_cache
*s
)
5701 struct kset
*kset
= cache_kset(s
);
5702 int unmergeable
= slab_unmergeable(s
);
5704 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5707 kobject_init(&s
->kobj
, &slab_ktype
);
5711 if (!unmergeable
&& disable_higher_order_debug
&&
5712 (slub_debug
& DEBUG_METADATA_FLAGS
))
5717 * Slabcache can never be merged so we can use the name proper.
5718 * This is typically the case for debug situations. In that
5719 * case we can catch duplicate names easily.
5721 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5725 * Create a unique name for the slab as a target
5728 name
= create_unique_id(s
);
5731 s
->kobj
.kset
= kset
;
5732 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5736 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5741 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5742 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5743 if (!s
->memcg_kset
) {
5750 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5752 /* Setup first alias */
5753 sysfs_slab_alias(s
, s
->name
);
5760 kobject_del(&s
->kobj
);
5764 static void sysfs_slab_remove(struct kmem_cache
*s
)
5766 if (slab_state
< FULL
)
5768 * Sysfs has not been setup yet so no need to remove the
5773 kobject_get(&s
->kobj
);
5774 schedule_work(&s
->kobj_remove_work
);
5777 void sysfs_slab_release(struct kmem_cache
*s
)
5779 if (slab_state
>= FULL
)
5780 kobject_put(&s
->kobj
);
5784 * Need to buffer aliases during bootup until sysfs becomes
5785 * available lest we lose that information.
5787 struct saved_alias
{
5788 struct kmem_cache
*s
;
5790 struct saved_alias
*next
;
5793 static struct saved_alias
*alias_list
;
5795 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5797 struct saved_alias
*al
;
5799 if (slab_state
== FULL
) {
5801 * If we have a leftover link then remove it.
5803 sysfs_remove_link(&slab_kset
->kobj
, name
);
5804 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5807 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5813 al
->next
= alias_list
;
5818 static int __init
slab_sysfs_init(void)
5820 struct kmem_cache
*s
;
5823 mutex_lock(&slab_mutex
);
5825 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5827 mutex_unlock(&slab_mutex
);
5828 pr_err("Cannot register slab subsystem.\n");
5834 list_for_each_entry(s
, &slab_caches
, list
) {
5835 err
= sysfs_slab_add(s
);
5837 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5841 while (alias_list
) {
5842 struct saved_alias
*al
= alias_list
;
5844 alias_list
= alias_list
->next
;
5845 err
= sysfs_slab_alias(al
->s
, al
->name
);
5847 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5852 mutex_unlock(&slab_mutex
);
5857 __initcall(slab_sysfs_init
);
5858 #endif /* CONFIG_SYSFS */
5861 * The /proc/slabinfo ABI
5863 #ifdef CONFIG_SLUB_DEBUG
5864 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5866 unsigned long nr_slabs
= 0;
5867 unsigned long nr_objs
= 0;
5868 unsigned long nr_free
= 0;
5870 struct kmem_cache_node
*n
;
5872 for_each_kmem_cache_node(s
, node
, n
) {
5873 nr_slabs
+= node_nr_slabs(n
);
5874 nr_objs
+= node_nr_objs(n
);
5875 nr_free
+= count_partial(n
, count_free
);
5878 sinfo
->active_objs
= nr_objs
- nr_free
;
5879 sinfo
->num_objs
= nr_objs
;
5880 sinfo
->active_slabs
= nr_slabs
;
5881 sinfo
->num_slabs
= nr_slabs
;
5882 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5883 sinfo
->cache_order
= oo_order(s
->oo
);
5886 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5890 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5891 size_t count
, loff_t
*ppos
)
5895 #endif /* CONFIG_SLUB_DEBUG */