1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * 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
254 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
255 * Normally, this doesn't cause any issues, as both set_freepointer()
256 * and get_freepointer() are called with a pointer with the same tag.
257 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
258 * example, when __free_slub() iterates over objects in a cache, it
259 * passes untagged pointers to check_object(). check_object() in turns
260 * calls get_freepointer() with an untagged pointer, which causes the
261 * freepointer to be restored incorrectly.
263 return (void *)((unsigned long)ptr
^ s
->random
^
264 (unsigned long)kasan_reset_tag((void *)ptr_addr
));
270 /* Returns the freelist pointer recorded at location ptr_addr. */
271 static inline void *freelist_dereference(const struct kmem_cache
*s
,
274 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
275 (unsigned long)ptr_addr
);
278 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
280 return freelist_dereference(s
, object
+ s
->offset
);
283 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
285 prefetch(object
+ s
->offset
);
288 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
290 unsigned long freepointer_addr
;
293 if (!debug_pagealloc_enabled())
294 return get_freepointer(s
, object
);
296 freepointer_addr
= (unsigned long)object
+ s
->offset
;
297 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
298 return freelist_ptr(s
, p
, freepointer_addr
);
301 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
303 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
305 #ifdef CONFIG_SLAB_FREELIST_HARDENED
306 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
309 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
312 /* Loop over all objects in a slab */
313 #define for_each_object(__p, __s, __addr, __objects) \
314 for (__p = fixup_red_left(__s, __addr); \
315 __p < (__addr) + (__objects) * (__s)->size; \
318 /* Determine object index from a given position */
319 static inline unsigned int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
321 return (kasan_reset_tag(p
) - addr
) / s
->size
;
324 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
326 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
329 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
332 struct kmem_cache_order_objects x
= {
333 (order
<< OO_SHIFT
) + order_objects(order
, size
)
339 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
341 return x
.x
>> OO_SHIFT
;
344 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
346 return x
.x
& OO_MASK
;
350 * Per slab locking using the pagelock
352 static __always_inline
void slab_lock(struct page
*page
)
354 VM_BUG_ON_PAGE(PageTail(page
), page
);
355 bit_spin_lock(PG_locked
, &page
->flags
);
358 static __always_inline
void slab_unlock(struct page
*page
)
360 VM_BUG_ON_PAGE(PageTail(page
), page
);
361 __bit_spin_unlock(PG_locked
, &page
->flags
);
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
366 void *freelist_old
, unsigned long counters_old
,
367 void *freelist_new
, unsigned long counters_new
,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s
->flags
& __CMPXCHG_DOUBLE
) {
374 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
375 freelist_old
, counters_old
,
376 freelist_new
, counters_new
))
382 if (page
->freelist
== freelist_old
&&
383 page
->counters
== counters_old
) {
384 page
->freelist
= freelist_new
;
385 page
->counters
= counters_new
;
393 stat(s
, CMPXCHG_DOUBLE_FAIL
);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
402 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
403 void *freelist_old
, unsigned long counters_old
,
404 void *freelist_new
, unsigned long counters_new
,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s
->flags
& __CMPXCHG_DOUBLE
) {
410 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
411 freelist_old
, counters_old
,
412 freelist_new
, counters_new
))
419 local_irq_save(flags
);
421 if (page
->freelist
== freelist_old
&&
422 page
->counters
== counters_old
) {
423 page
->freelist
= freelist_new
;
424 page
->counters
= counters_new
;
426 local_irq_restore(flags
);
430 local_irq_restore(flags
);
434 stat(s
, CMPXCHG_DOUBLE_FAIL
);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
453 void *addr
= page_address(page
);
455 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
456 set_bit(slab_index(p
, s
, addr
), map
);
459 static inline unsigned int size_from_object(struct kmem_cache
*s
)
461 if (s
->flags
& SLAB_RED_ZONE
)
462 return s
->size
- s
->red_left_pad
;
467 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
469 if (s
->flags
& SLAB_RED_ZONE
)
470 p
-= s
->red_left_pad
;
478 #if defined(CONFIG_SLUB_DEBUG_ON)
479 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
481 static slab_flags_t slub_debug
;
484 static char *slub_debug_slabs
;
485 static int disable_higher_order_debug
;
488 * slub is about to manipulate internal object metadata. This memory lies
489 * outside the range of the allocated object, so accessing it would normally
490 * be reported by kasan as a bounds error. metadata_access_enable() is used
491 * to tell kasan that these accesses are OK.
493 static inline void metadata_access_enable(void)
495 kasan_disable_current();
498 static inline void metadata_access_disable(void)
500 kasan_enable_current();
507 /* Verify that a pointer has an address that is valid within a slab page */
508 static inline int check_valid_pointer(struct kmem_cache
*s
,
509 struct page
*page
, void *object
)
516 base
= page_address(page
);
517 object
= kasan_reset_tag(object
);
518 object
= restore_red_left(s
, object
);
519 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
520 (object
- base
) % s
->size
) {
527 static void print_section(char *level
, char *text
, u8
*addr
,
530 metadata_access_enable();
531 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
533 metadata_access_disable();
536 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
537 enum track_item alloc
)
542 p
= object
+ s
->offset
+ sizeof(void *);
544 p
= object
+ s
->inuse
;
549 static void set_track(struct kmem_cache
*s
, void *object
,
550 enum track_item alloc
, unsigned long addr
)
552 struct track
*p
= get_track(s
, object
, alloc
);
555 #ifdef CONFIG_STACKTRACE
556 unsigned int nr_entries
;
558 metadata_access_enable();
559 nr_entries
= stack_trace_save(p
->addrs
, TRACK_ADDRS_COUNT
, 3);
560 metadata_access_disable();
562 if (nr_entries
< TRACK_ADDRS_COUNT
)
563 p
->addrs
[nr_entries
] = 0;
566 p
->cpu
= smp_processor_id();
567 p
->pid
= current
->pid
;
570 memset(p
, 0, sizeof(struct track
));
574 static void init_tracking(struct kmem_cache
*s
, void *object
)
576 if (!(s
->flags
& SLAB_STORE_USER
))
579 set_track(s
, object
, TRACK_FREE
, 0UL);
580 set_track(s
, object
, TRACK_ALLOC
, 0UL);
583 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
588 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
589 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
590 #ifdef CONFIG_STACKTRACE
593 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
595 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
602 static void print_tracking(struct kmem_cache
*s
, void *object
)
604 unsigned long pr_time
= jiffies
;
605 if (!(s
->flags
& SLAB_STORE_USER
))
608 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
609 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
612 static void print_page_info(struct page
*page
)
614 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
615 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
619 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
621 struct va_format vaf
;
627 pr_err("=============================================================================\n");
628 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
629 pr_err("-----------------------------------------------------------------------------\n\n");
631 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
635 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
637 struct va_format vaf
;
643 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
647 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
649 unsigned int off
; /* Offset of last byte */
650 u8
*addr
= page_address(page
);
652 print_tracking(s
, p
);
654 print_page_info(page
);
656 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
657 p
, p
- addr
, get_freepointer(s
, p
));
659 if (s
->flags
& SLAB_RED_ZONE
)
660 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
662 else if (p
> addr
+ 16)
663 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
665 print_section(KERN_ERR
, "Object ", p
,
666 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
667 if (s
->flags
& SLAB_RED_ZONE
)
668 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
669 s
->inuse
- s
->object_size
);
672 off
= s
->offset
+ sizeof(void *);
676 if (s
->flags
& SLAB_STORE_USER
)
677 off
+= 2 * sizeof(struct track
);
679 off
+= kasan_metadata_size(s
);
681 if (off
!= size_from_object(s
))
682 /* Beginning of the filler is the free pointer */
683 print_section(KERN_ERR
, "Padding ", p
+ off
,
684 size_from_object(s
) - off
);
689 void object_err(struct kmem_cache
*s
, struct page
*page
,
690 u8
*object
, char *reason
)
692 slab_bug(s
, "%s", reason
);
693 print_trailer(s
, page
, object
);
696 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
697 const char *fmt
, ...)
703 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
705 slab_bug(s
, "%s", buf
);
706 print_page_info(page
);
710 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
714 if (s
->flags
& SLAB_RED_ZONE
)
715 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
717 if (s
->flags
& __OBJECT_POISON
) {
718 memset(p
, POISON_FREE
, s
->object_size
- 1);
719 p
[s
->object_size
- 1] = POISON_END
;
722 if (s
->flags
& SLAB_RED_ZONE
)
723 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
726 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
727 void *from
, void *to
)
729 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
730 memset(from
, data
, to
- from
);
733 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
734 u8
*object
, char *what
,
735 u8
*start
, unsigned int value
, unsigned int bytes
)
740 metadata_access_enable();
741 fault
= memchr_inv(start
, value
, bytes
);
742 metadata_access_disable();
747 while (end
> fault
&& end
[-1] == value
)
750 slab_bug(s
, "%s overwritten", what
);
751 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
752 fault
, end
- 1, fault
[0], value
);
753 print_trailer(s
, page
, object
);
755 restore_bytes(s
, what
, value
, fault
, end
);
763 * Bytes of the object to be managed.
764 * If the freepointer may overlay the object then the free
765 * pointer is the first word of the object.
767 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
770 * object + s->object_size
771 * Padding to reach word boundary. This is also used for Redzoning.
772 * Padding is extended by another word if Redzoning is enabled and
773 * object_size == inuse.
775 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
776 * 0xcc (RED_ACTIVE) for objects in use.
779 * Meta data starts here.
781 * A. Free pointer (if we cannot overwrite object on free)
782 * B. Tracking data for SLAB_STORE_USER
783 * C. Padding to reach required alignment boundary or at mininum
784 * one word if debugging is on to be able to detect writes
785 * before the word boundary.
787 * Padding is done using 0x5a (POISON_INUSE)
790 * Nothing is used beyond s->size.
792 * If slabcaches are merged then the object_size and inuse boundaries are mostly
793 * ignored. And therefore no slab options that rely on these boundaries
794 * may be used with merged slabcaches.
797 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
799 unsigned long off
= s
->inuse
; /* The end of info */
802 /* Freepointer is placed after the object. */
803 off
+= sizeof(void *);
805 if (s
->flags
& SLAB_STORE_USER
)
806 /* We also have user information there */
807 off
+= 2 * sizeof(struct track
);
809 off
+= kasan_metadata_size(s
);
811 if (size_from_object(s
) == off
)
814 return check_bytes_and_report(s
, page
, p
, "Object padding",
815 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
818 /* Check the pad bytes at the end of a slab page */
819 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
828 if (!(s
->flags
& SLAB_POISON
))
831 start
= page_address(page
);
832 length
= PAGE_SIZE
<< compound_order(page
);
833 end
= start
+ length
;
834 remainder
= length
% s
->size
;
838 pad
= end
- remainder
;
839 metadata_access_enable();
840 fault
= memchr_inv(pad
, POISON_INUSE
, remainder
);
841 metadata_access_disable();
844 while (end
> fault
&& end
[-1] == POISON_INUSE
)
847 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
848 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
850 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
854 static int check_object(struct kmem_cache
*s
, struct page
*page
,
855 void *object
, u8 val
)
858 u8
*endobject
= object
+ s
->object_size
;
860 if (s
->flags
& SLAB_RED_ZONE
) {
861 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
862 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
865 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
866 endobject
, val
, s
->inuse
- s
->object_size
))
869 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
870 check_bytes_and_report(s
, page
, p
, "Alignment padding",
871 endobject
, POISON_INUSE
,
872 s
->inuse
- s
->object_size
);
876 if (s
->flags
& SLAB_POISON
) {
877 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
878 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
879 POISON_FREE
, s
->object_size
- 1) ||
880 !check_bytes_and_report(s
, page
, p
, "Poison",
881 p
+ s
->object_size
- 1, POISON_END
, 1)))
884 * check_pad_bytes cleans up on its own.
886 check_pad_bytes(s
, page
, p
);
889 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
891 * Object and freepointer overlap. Cannot check
892 * freepointer while object is allocated.
896 /* Check free pointer validity */
897 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
898 object_err(s
, page
, p
, "Freepointer corrupt");
900 * No choice but to zap it and thus lose the remainder
901 * of the free objects in this slab. May cause
902 * another error because the object count is now wrong.
904 set_freepointer(s
, p
, NULL
);
910 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
914 VM_BUG_ON(!irqs_disabled());
916 if (!PageSlab(page
)) {
917 slab_err(s
, page
, "Not a valid slab page");
921 maxobj
= order_objects(compound_order(page
), s
->size
);
922 if (page
->objects
> maxobj
) {
923 slab_err(s
, page
, "objects %u > max %u",
924 page
->objects
, maxobj
);
927 if (page
->inuse
> page
->objects
) {
928 slab_err(s
, page
, "inuse %u > max %u",
929 page
->inuse
, page
->objects
);
932 /* Slab_pad_check fixes things up after itself */
933 slab_pad_check(s
, page
);
938 * Determine if a certain object on a page is on the freelist. Must hold the
939 * slab lock to guarantee that the chains are in a consistent state.
941 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
949 while (fp
&& nr
<= page
->objects
) {
952 if (!check_valid_pointer(s
, page
, fp
)) {
954 object_err(s
, page
, object
,
955 "Freechain corrupt");
956 set_freepointer(s
, object
, NULL
);
958 slab_err(s
, page
, "Freepointer corrupt");
959 page
->freelist
= NULL
;
960 page
->inuse
= page
->objects
;
961 slab_fix(s
, "Freelist cleared");
967 fp
= get_freepointer(s
, object
);
971 max_objects
= order_objects(compound_order(page
), s
->size
);
972 if (max_objects
> MAX_OBJS_PER_PAGE
)
973 max_objects
= MAX_OBJS_PER_PAGE
;
975 if (page
->objects
!= max_objects
) {
976 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
977 page
->objects
, max_objects
);
978 page
->objects
= max_objects
;
979 slab_fix(s
, "Number of objects adjusted.");
981 if (page
->inuse
!= page
->objects
- nr
) {
982 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
983 page
->inuse
, page
->objects
- nr
);
984 page
->inuse
= page
->objects
- nr
;
985 slab_fix(s
, "Object count adjusted.");
987 return search
== NULL
;
990 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
993 if (s
->flags
& SLAB_TRACE
) {
994 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
996 alloc
? "alloc" : "free",
1001 print_section(KERN_INFO
, "Object ", (void *)object
,
1009 * Tracking of fully allocated slabs for debugging purposes.
1011 static void add_full(struct kmem_cache
*s
,
1012 struct kmem_cache_node
*n
, struct page
*page
)
1014 if (!(s
->flags
& SLAB_STORE_USER
))
1017 lockdep_assert_held(&n
->list_lock
);
1018 list_add(&page
->slab_list
, &n
->full
);
1021 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1023 if (!(s
->flags
& SLAB_STORE_USER
))
1026 lockdep_assert_held(&n
->list_lock
);
1027 list_del(&page
->slab_list
);
1030 /* Tracking of the number of slabs for debugging purposes */
1031 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1033 struct kmem_cache_node
*n
= get_node(s
, node
);
1035 return atomic_long_read(&n
->nr_slabs
);
1038 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1040 return atomic_long_read(&n
->nr_slabs
);
1043 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1045 struct kmem_cache_node
*n
= get_node(s
, node
);
1048 * May be called early in order to allocate a slab for the
1049 * kmem_cache_node structure. Solve the chicken-egg
1050 * dilemma by deferring the increment of the count during
1051 * bootstrap (see early_kmem_cache_node_alloc).
1054 atomic_long_inc(&n
->nr_slabs
);
1055 atomic_long_add(objects
, &n
->total_objects
);
1058 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1060 struct kmem_cache_node
*n
= get_node(s
, node
);
1062 atomic_long_dec(&n
->nr_slabs
);
1063 atomic_long_sub(objects
, &n
->total_objects
);
1066 /* Object debug checks for alloc/free paths */
1067 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1070 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1073 init_object(s
, object
, SLUB_RED_INACTIVE
);
1074 init_tracking(s
, object
);
1077 static void setup_page_debug(struct kmem_cache
*s
, void *addr
, int order
)
1079 if (!(s
->flags
& SLAB_POISON
))
1082 metadata_access_enable();
1083 memset(addr
, POISON_INUSE
, PAGE_SIZE
<< order
);
1084 metadata_access_disable();
1087 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1088 struct page
*page
, void *object
)
1090 if (!check_slab(s
, page
))
1093 if (!check_valid_pointer(s
, page
, object
)) {
1094 object_err(s
, page
, object
, "Freelist Pointer check fails");
1098 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1104 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1106 void *object
, unsigned long addr
)
1108 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1109 if (!alloc_consistency_checks(s
, page
, object
))
1113 /* Success perform special debug activities for allocs */
1114 if (s
->flags
& SLAB_STORE_USER
)
1115 set_track(s
, object
, TRACK_ALLOC
, addr
);
1116 trace(s
, page
, object
, 1);
1117 init_object(s
, object
, SLUB_RED_ACTIVE
);
1121 if (PageSlab(page
)) {
1123 * If this is a slab page then lets do the best we can
1124 * to avoid issues in the future. Marking all objects
1125 * as used avoids touching the remaining objects.
1127 slab_fix(s
, "Marking all objects used");
1128 page
->inuse
= page
->objects
;
1129 page
->freelist
= NULL
;
1134 static inline int free_consistency_checks(struct kmem_cache
*s
,
1135 struct page
*page
, void *object
, unsigned long addr
)
1137 if (!check_valid_pointer(s
, page
, object
)) {
1138 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1142 if (on_freelist(s
, page
, object
)) {
1143 object_err(s
, page
, object
, "Object already free");
1147 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1150 if (unlikely(s
!= page
->slab_cache
)) {
1151 if (!PageSlab(page
)) {
1152 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1154 } else if (!page
->slab_cache
) {
1155 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1159 object_err(s
, page
, object
,
1160 "page slab pointer corrupt.");
1166 /* Supports checking bulk free of a constructed freelist */
1167 static noinline
int free_debug_processing(
1168 struct kmem_cache
*s
, struct page
*page
,
1169 void *head
, void *tail
, int bulk_cnt
,
1172 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1173 void *object
= head
;
1175 unsigned long uninitialized_var(flags
);
1178 spin_lock_irqsave(&n
->list_lock
, flags
);
1181 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1182 if (!check_slab(s
, page
))
1189 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1190 if (!free_consistency_checks(s
, page
, object
, addr
))
1194 if (s
->flags
& SLAB_STORE_USER
)
1195 set_track(s
, object
, TRACK_FREE
, addr
);
1196 trace(s
, page
, object
, 0);
1197 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1198 init_object(s
, object
, SLUB_RED_INACTIVE
);
1200 /* Reached end of constructed freelist yet? */
1201 if (object
!= tail
) {
1202 object
= get_freepointer(s
, object
);
1208 if (cnt
!= bulk_cnt
)
1209 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1213 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1215 slab_fix(s
, "Object at 0x%p not freed", object
);
1219 static int __init
setup_slub_debug(char *str
)
1221 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1222 if (*str
++ != '=' || !*str
)
1224 * No options specified. Switch on full debugging.
1230 * No options but restriction on slabs. This means full
1231 * debugging for slabs matching a pattern.
1238 * Switch off all debugging measures.
1243 * Determine which debug features should be switched on
1245 for (; *str
&& *str
!= ','; str
++) {
1246 switch (tolower(*str
)) {
1248 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1251 slub_debug
|= SLAB_RED_ZONE
;
1254 slub_debug
|= SLAB_POISON
;
1257 slub_debug
|= SLAB_STORE_USER
;
1260 slub_debug
|= SLAB_TRACE
;
1263 slub_debug
|= SLAB_FAILSLAB
;
1267 * Avoid enabling debugging on caches if its minimum
1268 * order would increase as a result.
1270 disable_higher_order_debug
= 1;
1273 pr_err("slub_debug option '%c' unknown. skipped\n",
1280 slub_debug_slabs
= str
+ 1;
1285 __setup("slub_debug", setup_slub_debug
);
1288 * kmem_cache_flags - apply debugging options to the cache
1289 * @object_size: the size of an object without meta data
1290 * @flags: flags to set
1291 * @name: name of the cache
1292 * @ctor: constructor function
1294 * Debug option(s) are applied to @flags. In addition to the debug
1295 * option(s), if a slab name (or multiple) is specified i.e.
1296 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1297 * then only the select slabs will receive the debug option(s).
1299 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1300 slab_flags_t flags
, const char *name
,
1301 void (*ctor
)(void *))
1306 /* If slub_debug = 0, it folds into the if conditional. */
1307 if (!slub_debug_slabs
)
1308 return flags
| slub_debug
;
1311 iter
= slub_debug_slabs
;
1316 end
= strchrnul(iter
, ',');
1318 glob
= strnchr(iter
, end
- iter
, '*');
1320 cmplen
= glob
- iter
;
1322 cmplen
= max_t(size_t, len
, (end
- iter
));
1324 if (!strncmp(name
, iter
, cmplen
)) {
1325 flags
|= slub_debug
;
1336 #else /* !CONFIG_SLUB_DEBUG */
1337 static inline void setup_object_debug(struct kmem_cache
*s
,
1338 struct page
*page
, void *object
) {}
1339 static inline void setup_page_debug(struct kmem_cache
*s
,
1340 void *addr
, int order
) {}
1342 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1343 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1345 static inline int free_debug_processing(
1346 struct kmem_cache
*s
, struct page
*page
,
1347 void *head
, void *tail
, int bulk_cnt
,
1348 unsigned long addr
) { return 0; }
1350 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1352 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1353 void *object
, u8 val
) { return 1; }
1354 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1355 struct page
*page
) {}
1356 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1357 struct page
*page
) {}
1358 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1359 slab_flags_t flags
, const char *name
,
1360 void (*ctor
)(void *))
1364 #define slub_debug 0
1366 #define disable_higher_order_debug 0
1368 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1370 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1372 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1374 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1377 #endif /* CONFIG_SLUB_DEBUG */
1380 * Hooks for other subsystems that check memory allocations. In a typical
1381 * production configuration these hooks all should produce no code at all.
1383 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1385 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1386 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1387 kmemleak_alloc(ptr
, size
, 1, flags
);
1391 static __always_inline
void kfree_hook(void *x
)
1394 kasan_kfree_large(x
, _RET_IP_
);
1397 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1399 kmemleak_free_recursive(x
, s
->flags
);
1402 * Trouble is that we may no longer disable interrupts in the fast path
1403 * So in order to make the debug calls that expect irqs to be
1404 * disabled we need to disable interrupts temporarily.
1406 #ifdef CONFIG_LOCKDEP
1408 unsigned long flags
;
1410 local_irq_save(flags
);
1411 debug_check_no_locks_freed(x
, s
->object_size
);
1412 local_irq_restore(flags
);
1415 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1416 debug_check_no_obj_freed(x
, s
->object_size
);
1418 /* KASAN might put x into memory quarantine, delaying its reuse */
1419 return kasan_slab_free(s
, x
, _RET_IP_
);
1422 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1423 void **head
, void **tail
)
1426 * Compiler cannot detect this function can be removed if slab_free_hook()
1427 * evaluates to nothing. Thus, catch all relevant config debug options here.
1429 #if defined(CONFIG_LOCKDEP) || \
1430 defined(CONFIG_DEBUG_KMEMLEAK) || \
1431 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1432 defined(CONFIG_KASAN)
1436 void *old_tail
= *tail
? *tail
: *head
;
1438 /* Head and tail of the reconstructed freelist */
1444 next
= get_freepointer(s
, object
);
1445 /* If object's reuse doesn't have to be delayed */
1446 if (!slab_free_hook(s
, object
)) {
1447 /* Move object to the new freelist */
1448 set_freepointer(s
, object
, *head
);
1453 } while (object
!= old_tail
);
1458 return *head
!= NULL
;
1464 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1467 setup_object_debug(s
, page
, object
);
1468 object
= kasan_init_slab_obj(s
, object
);
1469 if (unlikely(s
->ctor
)) {
1470 kasan_unpoison_object_data(s
, object
);
1472 kasan_poison_object_data(s
, object
);
1478 * Slab allocation and freeing
1480 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1481 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1484 unsigned int order
= oo_order(oo
);
1486 if (node
== NUMA_NO_NODE
)
1487 page
= alloc_pages(flags
, order
);
1489 page
= __alloc_pages_node(node
, flags
, order
);
1491 if (page
&& charge_slab_page(page
, flags
, order
, s
)) {
1492 __free_pages(page
, order
);
1499 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1500 /* Pre-initialize the random sequence cache */
1501 static int init_cache_random_seq(struct kmem_cache
*s
)
1503 unsigned int count
= oo_objects(s
->oo
);
1506 /* Bailout if already initialised */
1510 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1512 pr_err("SLUB: Unable to initialize free list for %s\n",
1517 /* Transform to an offset on the set of pages */
1518 if (s
->random_seq
) {
1521 for (i
= 0; i
< count
; i
++)
1522 s
->random_seq
[i
] *= s
->size
;
1527 /* Initialize each random sequence freelist per cache */
1528 static void __init
init_freelist_randomization(void)
1530 struct kmem_cache
*s
;
1532 mutex_lock(&slab_mutex
);
1534 list_for_each_entry(s
, &slab_caches
, list
)
1535 init_cache_random_seq(s
);
1537 mutex_unlock(&slab_mutex
);
1540 /* Get the next entry on the pre-computed freelist randomized */
1541 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1542 unsigned long *pos
, void *start
,
1543 unsigned long page_limit
,
1544 unsigned long freelist_count
)
1549 * If the target page allocation failed, the number of objects on the
1550 * page might be smaller than the usual size defined by the cache.
1553 idx
= s
->random_seq
[*pos
];
1555 if (*pos
>= freelist_count
)
1557 } while (unlikely(idx
>= page_limit
));
1559 return (char *)start
+ idx
;
1562 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1563 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1568 unsigned long idx
, pos
, page_limit
, freelist_count
;
1570 if (page
->objects
< 2 || !s
->random_seq
)
1573 freelist_count
= oo_objects(s
->oo
);
1574 pos
= get_random_int() % freelist_count
;
1576 page_limit
= page
->objects
* s
->size
;
1577 start
= fixup_red_left(s
, page_address(page
));
1579 /* First entry is used as the base of the freelist */
1580 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1582 cur
= setup_object(s
, page
, cur
);
1583 page
->freelist
= cur
;
1585 for (idx
= 1; idx
< page
->objects
; idx
++) {
1586 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1588 next
= setup_object(s
, page
, next
);
1589 set_freepointer(s
, cur
, next
);
1592 set_freepointer(s
, cur
, NULL
);
1597 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1601 static inline void init_freelist_randomization(void) { }
1602 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1606 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1608 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1611 struct kmem_cache_order_objects oo
= s
->oo
;
1613 void *start
, *p
, *next
;
1617 flags
&= gfp_allowed_mask
;
1619 if (gfpflags_allow_blocking(flags
))
1622 flags
|= s
->allocflags
;
1625 * Let the initial higher-order allocation fail under memory pressure
1626 * so we fall-back to the minimum order allocation.
1628 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1629 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1630 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1632 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1633 if (unlikely(!page
)) {
1637 * Allocation may have failed due to fragmentation.
1638 * Try a lower order alloc if possible
1640 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1641 if (unlikely(!page
))
1643 stat(s
, ORDER_FALLBACK
);
1646 page
->objects
= oo_objects(oo
);
1648 order
= compound_order(page
);
1649 page
->slab_cache
= s
;
1650 __SetPageSlab(page
);
1651 if (page_is_pfmemalloc(page
))
1652 SetPageSlabPfmemalloc(page
);
1654 kasan_poison_slab(page
);
1656 start
= page_address(page
);
1658 setup_page_debug(s
, start
, order
);
1660 shuffle
= shuffle_freelist(s
, page
);
1663 start
= fixup_red_left(s
, start
);
1664 start
= setup_object(s
, page
, start
);
1665 page
->freelist
= start
;
1666 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1668 next
= setup_object(s
, page
, next
);
1669 set_freepointer(s
, p
, next
);
1672 set_freepointer(s
, p
, NULL
);
1675 page
->inuse
= page
->objects
;
1679 if (gfpflags_allow_blocking(flags
))
1680 local_irq_disable();
1684 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1689 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1691 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1692 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1693 flags
&= ~GFP_SLAB_BUG_MASK
;
1694 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1695 invalid_mask
, &invalid_mask
, flags
, &flags
);
1699 return allocate_slab(s
,
1700 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1703 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1705 int order
= compound_order(page
);
1706 int pages
= 1 << order
;
1708 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1711 slab_pad_check(s
, page
);
1712 for_each_object(p
, s
, page_address(page
),
1714 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1717 __ClearPageSlabPfmemalloc(page
);
1718 __ClearPageSlab(page
);
1720 page
->mapping
= NULL
;
1721 if (current
->reclaim_state
)
1722 current
->reclaim_state
->reclaimed_slab
+= pages
;
1723 uncharge_slab_page(page
, order
, s
);
1724 __free_pages(page
, order
);
1727 static void rcu_free_slab(struct rcu_head
*h
)
1729 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1731 __free_slab(page
->slab_cache
, page
);
1734 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1736 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1737 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1739 __free_slab(s
, page
);
1742 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1744 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1749 * Management of partially allocated slabs.
1752 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1755 if (tail
== DEACTIVATE_TO_TAIL
)
1756 list_add_tail(&page
->slab_list
, &n
->partial
);
1758 list_add(&page
->slab_list
, &n
->partial
);
1761 static inline void add_partial(struct kmem_cache_node
*n
,
1762 struct page
*page
, int tail
)
1764 lockdep_assert_held(&n
->list_lock
);
1765 __add_partial(n
, page
, tail
);
1768 static inline void remove_partial(struct kmem_cache_node
*n
,
1771 lockdep_assert_held(&n
->list_lock
);
1772 list_del(&page
->slab_list
);
1777 * Remove slab from the partial list, freeze it and
1778 * return the pointer to the freelist.
1780 * Returns a list of objects or NULL if it fails.
1782 static inline void *acquire_slab(struct kmem_cache
*s
,
1783 struct kmem_cache_node
*n
, struct page
*page
,
1784 int mode
, int *objects
)
1787 unsigned long counters
;
1790 lockdep_assert_held(&n
->list_lock
);
1793 * Zap the freelist and set the frozen bit.
1794 * The old freelist is the list of objects for the
1795 * per cpu allocation list.
1797 freelist
= page
->freelist
;
1798 counters
= page
->counters
;
1799 new.counters
= counters
;
1800 *objects
= new.objects
- new.inuse
;
1802 new.inuse
= page
->objects
;
1803 new.freelist
= NULL
;
1805 new.freelist
= freelist
;
1808 VM_BUG_ON(new.frozen
);
1811 if (!__cmpxchg_double_slab(s
, page
,
1813 new.freelist
, new.counters
,
1817 remove_partial(n
, page
);
1822 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1823 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1826 * Try to allocate a partial slab from a specific node.
1828 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1829 struct kmem_cache_cpu
*c
, gfp_t flags
)
1831 struct page
*page
, *page2
;
1832 void *object
= NULL
;
1833 unsigned int available
= 0;
1837 * Racy check. If we mistakenly see no partial slabs then we
1838 * just allocate an empty slab. If we mistakenly try to get a
1839 * partial slab and there is none available then get_partials()
1842 if (!n
|| !n
->nr_partial
)
1845 spin_lock(&n
->list_lock
);
1846 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1849 if (!pfmemalloc_match(page
, flags
))
1852 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1856 available
+= objects
;
1859 stat(s
, ALLOC_FROM_PARTIAL
);
1862 put_cpu_partial(s
, page
, 0);
1863 stat(s
, CPU_PARTIAL_NODE
);
1865 if (!kmem_cache_has_cpu_partial(s
)
1866 || available
> slub_cpu_partial(s
) / 2)
1870 spin_unlock(&n
->list_lock
);
1875 * Get a page from somewhere. Search in increasing NUMA distances.
1877 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1878 struct kmem_cache_cpu
*c
)
1881 struct zonelist
*zonelist
;
1884 enum zone_type high_zoneidx
= gfp_zone(flags
);
1886 unsigned int cpuset_mems_cookie
;
1889 * The defrag ratio allows a configuration of the tradeoffs between
1890 * inter node defragmentation and node local allocations. A lower
1891 * defrag_ratio increases the tendency to do local allocations
1892 * instead of attempting to obtain partial slabs from other nodes.
1894 * If the defrag_ratio is set to 0 then kmalloc() always
1895 * returns node local objects. If the ratio is higher then kmalloc()
1896 * may return off node objects because partial slabs are obtained
1897 * from other nodes and filled up.
1899 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1900 * (which makes defrag_ratio = 1000) then every (well almost)
1901 * allocation will first attempt to defrag slab caches on other nodes.
1902 * This means scanning over all nodes to look for partial slabs which
1903 * may be expensive if we do it every time we are trying to find a slab
1904 * with available objects.
1906 if (!s
->remote_node_defrag_ratio
||
1907 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1911 cpuset_mems_cookie
= read_mems_allowed_begin();
1912 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1913 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1914 struct kmem_cache_node
*n
;
1916 n
= get_node(s
, zone_to_nid(zone
));
1918 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1919 n
->nr_partial
> s
->min_partial
) {
1920 object
= get_partial_node(s
, n
, c
, flags
);
1923 * Don't check read_mems_allowed_retry()
1924 * here - if mems_allowed was updated in
1925 * parallel, that was a harmless race
1926 * between allocation and the cpuset
1933 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1934 #endif /* CONFIG_NUMA */
1939 * Get a partial page, lock it and return it.
1941 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1942 struct kmem_cache_cpu
*c
)
1945 int searchnode
= node
;
1947 if (node
== NUMA_NO_NODE
)
1948 searchnode
= numa_mem_id();
1949 else if (!node_present_pages(node
))
1950 searchnode
= node_to_mem_node(node
);
1952 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1953 if (object
|| node
!= NUMA_NO_NODE
)
1956 return get_any_partial(s
, flags
, c
);
1959 #ifdef CONFIG_PREEMPT
1961 * Calculate the next globally unique transaction for disambiguiation
1962 * during cmpxchg. The transactions start with the cpu number and are then
1963 * incremented by CONFIG_NR_CPUS.
1965 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1968 * No preemption supported therefore also no need to check for
1974 static inline unsigned long next_tid(unsigned long tid
)
1976 return tid
+ TID_STEP
;
1979 static inline unsigned int tid_to_cpu(unsigned long tid
)
1981 return tid
% TID_STEP
;
1984 static inline unsigned long tid_to_event(unsigned long tid
)
1986 return tid
/ TID_STEP
;
1989 static inline unsigned int init_tid(int cpu
)
1994 static inline void note_cmpxchg_failure(const char *n
,
1995 const struct kmem_cache
*s
, unsigned long tid
)
1997 #ifdef SLUB_DEBUG_CMPXCHG
1998 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2000 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2002 #ifdef CONFIG_PREEMPT
2003 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2004 pr_warn("due to cpu change %d -> %d\n",
2005 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2008 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2009 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2010 tid_to_event(tid
), tid_to_event(actual_tid
));
2012 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2013 actual_tid
, tid
, next_tid(tid
));
2015 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2018 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2022 for_each_possible_cpu(cpu
)
2023 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2027 * Remove the cpu slab
2029 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2030 void *freelist
, struct kmem_cache_cpu
*c
)
2032 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2033 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2035 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2037 int tail
= DEACTIVATE_TO_HEAD
;
2041 if (page
->freelist
) {
2042 stat(s
, DEACTIVATE_REMOTE_FREES
);
2043 tail
= DEACTIVATE_TO_TAIL
;
2047 * Stage one: Free all available per cpu objects back
2048 * to the page freelist while it is still frozen. Leave the
2051 * There is no need to take the list->lock because the page
2054 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2056 unsigned long counters
;
2059 prior
= page
->freelist
;
2060 counters
= page
->counters
;
2061 set_freepointer(s
, freelist
, prior
);
2062 new.counters
= counters
;
2064 VM_BUG_ON(!new.frozen
);
2066 } while (!__cmpxchg_double_slab(s
, page
,
2068 freelist
, new.counters
,
2069 "drain percpu freelist"));
2071 freelist
= nextfree
;
2075 * Stage two: Ensure that the page is unfrozen while the
2076 * list presence reflects the actual number of objects
2079 * We setup the list membership and then perform a cmpxchg
2080 * with the count. If there is a mismatch then the page
2081 * is not unfrozen but the page is on the wrong list.
2083 * Then we restart the process which may have to remove
2084 * the page from the list that we just put it on again
2085 * because the number of objects in the slab may have
2090 old
.freelist
= page
->freelist
;
2091 old
.counters
= page
->counters
;
2092 VM_BUG_ON(!old
.frozen
);
2094 /* Determine target state of the slab */
2095 new.counters
= old
.counters
;
2098 set_freepointer(s
, freelist
, old
.freelist
);
2099 new.freelist
= freelist
;
2101 new.freelist
= old
.freelist
;
2105 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2107 else if (new.freelist
) {
2112 * Taking the spinlock removes the possibility
2113 * that acquire_slab() will see a slab page that
2116 spin_lock(&n
->list_lock
);
2120 if (kmem_cache_debug(s
) && !lock
) {
2123 * This also ensures that the scanning of full
2124 * slabs from diagnostic functions will not see
2127 spin_lock(&n
->list_lock
);
2133 remove_partial(n
, page
);
2134 else if (l
== M_FULL
)
2135 remove_full(s
, n
, page
);
2138 add_partial(n
, page
, tail
);
2139 else if (m
== M_FULL
)
2140 add_full(s
, n
, page
);
2144 if (!__cmpxchg_double_slab(s
, page
,
2145 old
.freelist
, old
.counters
,
2146 new.freelist
, new.counters
,
2151 spin_unlock(&n
->list_lock
);
2155 else if (m
== M_FULL
)
2156 stat(s
, DEACTIVATE_FULL
);
2157 else if (m
== M_FREE
) {
2158 stat(s
, DEACTIVATE_EMPTY
);
2159 discard_slab(s
, page
);
2168 * Unfreeze all the cpu partial slabs.
2170 * This function must be called with interrupts disabled
2171 * for the cpu using c (or some other guarantee must be there
2172 * to guarantee no concurrent accesses).
2174 static void unfreeze_partials(struct kmem_cache
*s
,
2175 struct kmem_cache_cpu
*c
)
2177 #ifdef CONFIG_SLUB_CPU_PARTIAL
2178 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2179 struct page
*page
, *discard_page
= NULL
;
2181 while ((page
= c
->partial
)) {
2185 c
->partial
= page
->next
;
2187 n2
= get_node(s
, page_to_nid(page
));
2190 spin_unlock(&n
->list_lock
);
2193 spin_lock(&n
->list_lock
);
2198 old
.freelist
= page
->freelist
;
2199 old
.counters
= page
->counters
;
2200 VM_BUG_ON(!old
.frozen
);
2202 new.counters
= old
.counters
;
2203 new.freelist
= old
.freelist
;
2207 } while (!__cmpxchg_double_slab(s
, page
,
2208 old
.freelist
, old
.counters
,
2209 new.freelist
, new.counters
,
2210 "unfreezing slab"));
2212 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2213 page
->next
= discard_page
;
2214 discard_page
= page
;
2216 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2217 stat(s
, FREE_ADD_PARTIAL
);
2222 spin_unlock(&n
->list_lock
);
2224 while (discard_page
) {
2225 page
= discard_page
;
2226 discard_page
= discard_page
->next
;
2228 stat(s
, DEACTIVATE_EMPTY
);
2229 discard_slab(s
, page
);
2232 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2236 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2237 * partial page slot if available.
2239 * If we did not find a slot then simply move all the partials to the
2240 * per node partial list.
2242 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2244 #ifdef CONFIG_SLUB_CPU_PARTIAL
2245 struct page
*oldpage
;
2253 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2256 pobjects
= oldpage
->pobjects
;
2257 pages
= oldpage
->pages
;
2258 if (drain
&& pobjects
> s
->cpu_partial
) {
2259 unsigned long flags
;
2261 * partial array is full. Move the existing
2262 * set to the per node partial list.
2264 local_irq_save(flags
);
2265 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2266 local_irq_restore(flags
);
2270 stat(s
, CPU_PARTIAL_DRAIN
);
2275 pobjects
+= page
->objects
- page
->inuse
;
2277 page
->pages
= pages
;
2278 page
->pobjects
= pobjects
;
2279 page
->next
= oldpage
;
2281 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2283 if (unlikely(!s
->cpu_partial
)) {
2284 unsigned long flags
;
2286 local_irq_save(flags
);
2287 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2288 local_irq_restore(flags
);
2291 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2294 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2296 stat(s
, CPUSLAB_FLUSH
);
2297 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2299 c
->tid
= next_tid(c
->tid
);
2305 * Called from IPI handler with interrupts disabled.
2307 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2309 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2314 unfreeze_partials(s
, c
);
2317 static void flush_cpu_slab(void *d
)
2319 struct kmem_cache
*s
= d
;
2321 __flush_cpu_slab(s
, smp_processor_id());
2324 static bool has_cpu_slab(int cpu
, void *info
)
2326 struct kmem_cache
*s
= info
;
2327 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2329 return c
->page
|| slub_percpu_partial(c
);
2332 static void flush_all(struct kmem_cache
*s
)
2334 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2338 * Use the cpu notifier to insure that the cpu slabs are flushed when
2341 static int slub_cpu_dead(unsigned int cpu
)
2343 struct kmem_cache
*s
;
2344 unsigned long flags
;
2346 mutex_lock(&slab_mutex
);
2347 list_for_each_entry(s
, &slab_caches
, list
) {
2348 local_irq_save(flags
);
2349 __flush_cpu_slab(s
, cpu
);
2350 local_irq_restore(flags
);
2352 mutex_unlock(&slab_mutex
);
2357 * Check if the objects in a per cpu structure fit numa
2358 * locality expectations.
2360 static inline int node_match(struct page
*page
, int node
)
2363 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2369 #ifdef CONFIG_SLUB_DEBUG
2370 static int count_free(struct page
*page
)
2372 return page
->objects
- page
->inuse
;
2375 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2377 return atomic_long_read(&n
->total_objects
);
2379 #endif /* CONFIG_SLUB_DEBUG */
2381 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2382 static unsigned long count_partial(struct kmem_cache_node
*n
,
2383 int (*get_count
)(struct page
*))
2385 unsigned long flags
;
2386 unsigned long x
= 0;
2389 spin_lock_irqsave(&n
->list_lock
, flags
);
2390 list_for_each_entry(page
, &n
->partial
, slab_list
)
2391 x
+= get_count(page
);
2392 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2395 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2397 static noinline
void
2398 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2400 #ifdef CONFIG_SLUB_DEBUG
2401 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2402 DEFAULT_RATELIMIT_BURST
);
2404 struct kmem_cache_node
*n
;
2406 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2409 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2410 nid
, gfpflags
, &gfpflags
);
2411 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2412 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2415 if (oo_order(s
->min
) > get_order(s
->object_size
))
2416 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2419 for_each_kmem_cache_node(s
, node
, n
) {
2420 unsigned long nr_slabs
;
2421 unsigned long nr_objs
;
2422 unsigned long nr_free
;
2424 nr_free
= count_partial(n
, count_free
);
2425 nr_slabs
= node_nr_slabs(n
);
2426 nr_objs
= node_nr_objs(n
);
2428 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2429 node
, nr_slabs
, nr_objs
, nr_free
);
2434 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2435 int node
, struct kmem_cache_cpu
**pc
)
2438 struct kmem_cache_cpu
*c
= *pc
;
2441 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2443 freelist
= get_partial(s
, flags
, node
, c
);
2448 page
= new_slab(s
, flags
, node
);
2450 c
= raw_cpu_ptr(s
->cpu_slab
);
2455 * No other reference to the page yet so we can
2456 * muck around with it freely without cmpxchg
2458 freelist
= page
->freelist
;
2459 page
->freelist
= NULL
;
2461 stat(s
, ALLOC_SLAB
);
2469 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2471 if (unlikely(PageSlabPfmemalloc(page
)))
2472 return gfp_pfmemalloc_allowed(gfpflags
);
2478 * Check the page->freelist of a page and either transfer the freelist to the
2479 * per cpu freelist or deactivate the page.
2481 * The page is still frozen if the return value is not NULL.
2483 * If this function returns NULL then the page has been unfrozen.
2485 * This function must be called with interrupt disabled.
2487 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2490 unsigned long counters
;
2494 freelist
= page
->freelist
;
2495 counters
= page
->counters
;
2497 new.counters
= counters
;
2498 VM_BUG_ON(!new.frozen
);
2500 new.inuse
= page
->objects
;
2501 new.frozen
= freelist
!= NULL
;
2503 } while (!__cmpxchg_double_slab(s
, page
,
2512 * Slow path. The lockless freelist is empty or we need to perform
2515 * Processing is still very fast if new objects have been freed to the
2516 * regular freelist. In that case we simply take over the regular freelist
2517 * as the lockless freelist and zap the regular freelist.
2519 * If that is not working then we fall back to the partial lists. We take the
2520 * first element of the freelist as the object to allocate now and move the
2521 * rest of the freelist to the lockless freelist.
2523 * And if we were unable to get a new slab from the partial slab lists then
2524 * we need to allocate a new slab. This is the slowest path since it involves
2525 * a call to the page allocator and the setup of a new slab.
2527 * Version of __slab_alloc to use when we know that interrupts are
2528 * already disabled (which is the case for bulk allocation).
2530 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2531 unsigned long addr
, struct kmem_cache_cpu
*c
)
2541 if (unlikely(!node_match(page
, node
))) {
2542 int searchnode
= node
;
2544 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2545 searchnode
= node_to_mem_node(node
);
2547 if (unlikely(!node_match(page
, searchnode
))) {
2548 stat(s
, ALLOC_NODE_MISMATCH
);
2549 deactivate_slab(s
, page
, c
->freelist
, c
);
2555 * By rights, we should be searching for a slab page that was
2556 * PFMEMALLOC but right now, we are losing the pfmemalloc
2557 * information when the page leaves the per-cpu allocator
2559 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2560 deactivate_slab(s
, page
, c
->freelist
, c
);
2564 /* must check again c->freelist in case of cpu migration or IRQ */
2565 freelist
= c
->freelist
;
2569 freelist
= get_freelist(s
, page
);
2573 stat(s
, DEACTIVATE_BYPASS
);
2577 stat(s
, ALLOC_REFILL
);
2581 * freelist is pointing to the list of objects to be used.
2582 * page is pointing to the page from which the objects are obtained.
2583 * That page must be frozen for per cpu allocations to work.
2585 VM_BUG_ON(!c
->page
->frozen
);
2586 c
->freelist
= get_freepointer(s
, freelist
);
2587 c
->tid
= next_tid(c
->tid
);
2592 if (slub_percpu_partial(c
)) {
2593 page
= c
->page
= slub_percpu_partial(c
);
2594 slub_set_percpu_partial(c
, page
);
2595 stat(s
, CPU_PARTIAL_ALLOC
);
2599 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2601 if (unlikely(!freelist
)) {
2602 slab_out_of_memory(s
, gfpflags
, node
);
2607 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2610 /* Only entered in the debug case */
2611 if (kmem_cache_debug(s
) &&
2612 !alloc_debug_processing(s
, page
, freelist
, addr
))
2613 goto new_slab
; /* Slab failed checks. Next slab needed */
2615 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2620 * Another one that disabled interrupt and compensates for possible
2621 * cpu changes by refetching the per cpu area pointer.
2623 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2624 unsigned long addr
, struct kmem_cache_cpu
*c
)
2627 unsigned long flags
;
2629 local_irq_save(flags
);
2630 #ifdef CONFIG_PREEMPT
2632 * We may have been preempted and rescheduled on a different
2633 * cpu before disabling interrupts. Need to reload cpu area
2636 c
= this_cpu_ptr(s
->cpu_slab
);
2639 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2640 local_irq_restore(flags
);
2645 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2646 * have the fastpath folded into their functions. So no function call
2647 * overhead for requests that can be satisfied on the fastpath.
2649 * The fastpath works by first checking if the lockless freelist can be used.
2650 * If not then __slab_alloc is called for slow processing.
2652 * Otherwise we can simply pick the next object from the lockless free list.
2654 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2655 gfp_t gfpflags
, int node
, unsigned long addr
)
2658 struct kmem_cache_cpu
*c
;
2662 s
= slab_pre_alloc_hook(s
, gfpflags
);
2667 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2668 * enabled. We may switch back and forth between cpus while
2669 * reading from one cpu area. That does not matter as long
2670 * as we end up on the original cpu again when doing the cmpxchg.
2672 * We should guarantee that tid and kmem_cache are retrieved on
2673 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2674 * to check if it is matched or not.
2677 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2678 c
= raw_cpu_ptr(s
->cpu_slab
);
2679 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2680 unlikely(tid
!= READ_ONCE(c
->tid
)));
2683 * Irqless object alloc/free algorithm used here depends on sequence
2684 * of fetching cpu_slab's data. tid should be fetched before anything
2685 * on c to guarantee that object and page associated with previous tid
2686 * won't be used with current tid. If we fetch tid first, object and
2687 * page could be one associated with next tid and our alloc/free
2688 * request will be failed. In this case, we will retry. So, no problem.
2693 * The transaction ids are globally unique per cpu and per operation on
2694 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2695 * occurs on the right processor and that there was no operation on the
2696 * linked list in between.
2699 object
= c
->freelist
;
2701 if (unlikely(!object
|| !node_match(page
, node
))) {
2702 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2703 stat(s
, ALLOC_SLOWPATH
);
2705 void *next_object
= get_freepointer_safe(s
, object
);
2708 * The cmpxchg will only match if there was no additional
2709 * operation and if we are on the right processor.
2711 * The cmpxchg does the following atomically (without lock
2713 * 1. Relocate first pointer to the current per cpu area.
2714 * 2. Verify that tid and freelist have not been changed
2715 * 3. If they were not changed replace tid and freelist
2717 * Since this is without lock semantics the protection is only
2718 * against code executing on this cpu *not* from access by
2721 if (unlikely(!this_cpu_cmpxchg_double(
2722 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2724 next_object
, next_tid(tid
)))) {
2726 note_cmpxchg_failure("slab_alloc", s
, tid
);
2729 prefetch_freepointer(s
, next_object
);
2730 stat(s
, ALLOC_FASTPATH
);
2733 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2734 memset(object
, 0, s
->object_size
);
2736 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2741 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2742 gfp_t gfpflags
, unsigned long addr
)
2744 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2747 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2749 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2751 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2756 EXPORT_SYMBOL(kmem_cache_alloc
);
2758 #ifdef CONFIG_TRACING
2759 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2761 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2762 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2763 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2766 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2770 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2772 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2774 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2775 s
->object_size
, s
->size
, gfpflags
, node
);
2779 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2781 #ifdef CONFIG_TRACING
2782 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2784 int node
, size_t size
)
2786 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2788 trace_kmalloc_node(_RET_IP_
, ret
,
2789 size
, s
->size
, gfpflags
, node
);
2791 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2794 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2796 #endif /* CONFIG_NUMA */
2799 * Slow path handling. This may still be called frequently since objects
2800 * have a longer lifetime than the cpu slabs in most processing loads.
2802 * So we still attempt to reduce cache line usage. Just take the slab
2803 * lock and free the item. If there is no additional partial page
2804 * handling required then we can return immediately.
2806 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2807 void *head
, void *tail
, int cnt
,
2814 unsigned long counters
;
2815 struct kmem_cache_node
*n
= NULL
;
2816 unsigned long uninitialized_var(flags
);
2818 stat(s
, FREE_SLOWPATH
);
2820 if (kmem_cache_debug(s
) &&
2821 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2826 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2829 prior
= page
->freelist
;
2830 counters
= page
->counters
;
2831 set_freepointer(s
, tail
, prior
);
2832 new.counters
= counters
;
2833 was_frozen
= new.frozen
;
2835 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2837 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2840 * Slab was on no list before and will be
2842 * We can defer the list move and instead
2847 } else { /* Needs to be taken off a list */
2849 n
= get_node(s
, page_to_nid(page
));
2851 * Speculatively acquire the list_lock.
2852 * If the cmpxchg does not succeed then we may
2853 * drop the list_lock without any processing.
2855 * Otherwise the list_lock will synchronize with
2856 * other processors updating the list of slabs.
2858 spin_lock_irqsave(&n
->list_lock
, flags
);
2863 } while (!cmpxchg_double_slab(s
, page
,
2871 * If we just froze the page then put it onto the
2872 * per cpu partial list.
2874 if (new.frozen
&& !was_frozen
) {
2875 put_cpu_partial(s
, page
, 1);
2876 stat(s
, CPU_PARTIAL_FREE
);
2879 * The list lock was not taken therefore no list
2880 * activity can be necessary.
2883 stat(s
, FREE_FROZEN
);
2887 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2891 * Objects left in the slab. If it was not on the partial list before
2894 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2895 remove_full(s
, n
, page
);
2896 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2897 stat(s
, FREE_ADD_PARTIAL
);
2899 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2905 * Slab on the partial list.
2907 remove_partial(n
, page
);
2908 stat(s
, FREE_REMOVE_PARTIAL
);
2910 /* Slab must be on the full list */
2911 remove_full(s
, n
, page
);
2914 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2916 discard_slab(s
, page
);
2920 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2921 * can perform fastpath freeing without additional function calls.
2923 * The fastpath is only possible if we are freeing to the current cpu slab
2924 * of this processor. This typically the case if we have just allocated
2927 * If fastpath is not possible then fall back to __slab_free where we deal
2928 * with all sorts of special processing.
2930 * Bulk free of a freelist with several objects (all pointing to the
2931 * same page) possible by specifying head and tail ptr, plus objects
2932 * count (cnt). Bulk free indicated by tail pointer being set.
2934 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2935 struct page
*page
, void *head
, void *tail
,
2936 int cnt
, unsigned long addr
)
2938 void *tail_obj
= tail
? : head
;
2939 struct kmem_cache_cpu
*c
;
2943 * Determine the currently cpus per cpu slab.
2944 * The cpu may change afterward. However that does not matter since
2945 * data is retrieved via this pointer. If we are on the same cpu
2946 * during the cmpxchg then the free will succeed.
2949 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2950 c
= raw_cpu_ptr(s
->cpu_slab
);
2951 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2952 unlikely(tid
!= READ_ONCE(c
->tid
)));
2954 /* Same with comment on barrier() in slab_alloc_node() */
2957 if (likely(page
== c
->page
)) {
2958 set_freepointer(s
, tail_obj
, c
->freelist
);
2960 if (unlikely(!this_cpu_cmpxchg_double(
2961 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2963 head
, next_tid(tid
)))) {
2965 note_cmpxchg_failure("slab_free", s
, tid
);
2968 stat(s
, FREE_FASTPATH
);
2970 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2974 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2975 void *head
, void *tail
, int cnt
,
2979 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2980 * to remove objects, whose reuse must be delayed.
2982 if (slab_free_freelist_hook(s
, &head
, &tail
))
2983 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2986 #ifdef CONFIG_KASAN_GENERIC
2987 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2989 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
2993 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2995 s
= cache_from_obj(s
, x
);
2998 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
2999 trace_kmem_cache_free(_RET_IP_
, x
);
3001 EXPORT_SYMBOL(kmem_cache_free
);
3003 struct detached_freelist
{
3008 struct kmem_cache
*s
;
3012 * This function progressively scans the array with free objects (with
3013 * a limited look ahead) and extract objects belonging to the same
3014 * page. It builds a detached freelist directly within the given
3015 * page/objects. This can happen without any need for
3016 * synchronization, because the objects are owned by running process.
3017 * The freelist is build up as a single linked list in the objects.
3018 * The idea is, that this detached freelist can then be bulk
3019 * transferred to the real freelist(s), but only requiring a single
3020 * synchronization primitive. Look ahead in the array is limited due
3021 * to performance reasons.
3024 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3025 void **p
, struct detached_freelist
*df
)
3027 size_t first_skipped_index
= 0;
3032 /* Always re-init detached_freelist */
3037 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3038 } while (!object
&& size
);
3043 page
= virt_to_head_page(object
);
3045 /* Handle kalloc'ed objects */
3046 if (unlikely(!PageSlab(page
))) {
3047 BUG_ON(!PageCompound(page
));
3049 __free_pages(page
, compound_order(page
));
3050 p
[size
] = NULL
; /* mark object processed */
3053 /* Derive kmem_cache from object */
3054 df
->s
= page
->slab_cache
;
3056 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3059 /* Start new detached freelist */
3061 set_freepointer(df
->s
, object
, NULL
);
3063 df
->freelist
= object
;
3064 p
[size
] = NULL
; /* mark object processed */
3070 continue; /* Skip processed objects */
3072 /* df->page is always set at this point */
3073 if (df
->page
== virt_to_head_page(object
)) {
3074 /* Opportunity build freelist */
3075 set_freepointer(df
->s
, object
, df
->freelist
);
3076 df
->freelist
= object
;
3078 p
[size
] = NULL
; /* mark object processed */
3083 /* Limit look ahead search */
3087 if (!first_skipped_index
)
3088 first_skipped_index
= size
+ 1;
3091 return first_skipped_index
;
3094 /* Note that interrupts must be enabled when calling this function. */
3095 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3101 struct detached_freelist df
;
3103 size
= build_detached_freelist(s
, size
, p
, &df
);
3107 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3108 } while (likely(size
));
3110 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3112 /* Note that interrupts must be enabled when calling this function. */
3113 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3116 struct kmem_cache_cpu
*c
;
3119 /* memcg and kmem_cache debug support */
3120 s
= slab_pre_alloc_hook(s
, flags
);
3124 * Drain objects in the per cpu slab, while disabling local
3125 * IRQs, which protects against PREEMPT and interrupts
3126 * handlers invoking normal fastpath.
3128 local_irq_disable();
3129 c
= this_cpu_ptr(s
->cpu_slab
);
3131 for (i
= 0; i
< size
; i
++) {
3132 void *object
= c
->freelist
;
3134 if (unlikely(!object
)) {
3136 * Invoking slow path likely have side-effect
3137 * of re-populating per CPU c->freelist
3139 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3141 if (unlikely(!p
[i
]))
3144 c
= this_cpu_ptr(s
->cpu_slab
);
3145 continue; /* goto for-loop */
3147 c
->freelist
= get_freepointer(s
, object
);
3150 c
->tid
= next_tid(c
->tid
);
3153 /* Clear memory outside IRQ disabled fastpath loop */
3154 if (unlikely(flags
& __GFP_ZERO
)) {
3157 for (j
= 0; j
< i
; j
++)
3158 memset(p
[j
], 0, s
->object_size
);
3161 /* memcg and kmem_cache debug support */
3162 slab_post_alloc_hook(s
, flags
, size
, p
);
3166 slab_post_alloc_hook(s
, flags
, i
, p
);
3167 __kmem_cache_free_bulk(s
, i
, p
);
3170 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3174 * Object placement in a slab is made very easy because we always start at
3175 * offset 0. If we tune the size of the object to the alignment then we can
3176 * get the required alignment by putting one properly sized object after
3179 * Notice that the allocation order determines the sizes of the per cpu
3180 * caches. Each processor has always one slab available for allocations.
3181 * Increasing the allocation order reduces the number of times that slabs
3182 * must be moved on and off the partial lists and is therefore a factor in
3187 * Mininum / Maximum order of slab pages. This influences locking overhead
3188 * and slab fragmentation. A higher order reduces the number of partial slabs
3189 * and increases the number of allocations possible without having to
3190 * take the list_lock.
3192 static unsigned int slub_min_order
;
3193 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3194 static unsigned int slub_min_objects
;
3197 * Calculate the order of allocation given an slab object size.
3199 * The order of allocation has significant impact on performance and other
3200 * system components. Generally order 0 allocations should be preferred since
3201 * order 0 does not cause fragmentation in the page allocator. Larger objects
3202 * be problematic to put into order 0 slabs because there may be too much
3203 * unused space left. We go to a higher order if more than 1/16th of the slab
3206 * In order to reach satisfactory performance we must ensure that a minimum
3207 * number of objects is in one slab. Otherwise we may generate too much
3208 * activity on the partial lists which requires taking the list_lock. This is
3209 * less a concern for large slabs though which are rarely used.
3211 * slub_max_order specifies the order where we begin to stop considering the
3212 * number of objects in a slab as critical. If we reach slub_max_order then
3213 * we try to keep the page order as low as possible. So we accept more waste
3214 * of space in favor of a small page order.
3216 * Higher order allocations also allow the placement of more objects in a
3217 * slab and thereby reduce object handling overhead. If the user has
3218 * requested a higher mininum order then we start with that one instead of
3219 * the smallest order which will fit the object.
3221 static inline unsigned int slab_order(unsigned int size
,
3222 unsigned int min_objects
, unsigned int max_order
,
3223 unsigned int fract_leftover
)
3225 unsigned int min_order
= slub_min_order
;
3228 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3229 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3231 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3232 order
<= max_order
; order
++) {
3234 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3237 rem
= slab_size
% size
;
3239 if (rem
<= slab_size
/ fract_leftover
)
3246 static inline int calculate_order(unsigned int size
)
3249 unsigned int min_objects
;
3250 unsigned int max_objects
;
3253 * Attempt to find best configuration for a slab. This
3254 * works by first attempting to generate a layout with
3255 * the best configuration and backing off gradually.
3257 * First we increase the acceptable waste in a slab. Then
3258 * we reduce the minimum objects required in a slab.
3260 min_objects
= slub_min_objects
;
3262 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3263 max_objects
= order_objects(slub_max_order
, size
);
3264 min_objects
= min(min_objects
, max_objects
);
3266 while (min_objects
> 1) {
3267 unsigned int fraction
;
3270 while (fraction
>= 4) {
3271 order
= slab_order(size
, min_objects
,
3272 slub_max_order
, fraction
);
3273 if (order
<= slub_max_order
)
3281 * We were unable to place multiple objects in a slab. Now
3282 * lets see if we can place a single object there.
3284 order
= slab_order(size
, 1, slub_max_order
, 1);
3285 if (order
<= slub_max_order
)
3289 * Doh this slab cannot be placed using slub_max_order.
3291 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3292 if (order
< MAX_ORDER
)
3298 init_kmem_cache_node(struct kmem_cache_node
*n
)
3301 spin_lock_init(&n
->list_lock
);
3302 INIT_LIST_HEAD(&n
->partial
);
3303 #ifdef CONFIG_SLUB_DEBUG
3304 atomic_long_set(&n
->nr_slabs
, 0);
3305 atomic_long_set(&n
->total_objects
, 0);
3306 INIT_LIST_HEAD(&n
->full
);
3310 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3312 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3313 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3316 * Must align to double word boundary for the double cmpxchg
3317 * instructions to work; see __pcpu_double_call_return_bool().
3319 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3320 2 * sizeof(void *));
3325 init_kmem_cache_cpus(s
);
3330 static struct kmem_cache
*kmem_cache_node
;
3333 * No kmalloc_node yet so do it by hand. We know that this is the first
3334 * slab on the node for this slabcache. There are no concurrent accesses
3337 * Note that this function only works on the kmem_cache_node
3338 * when allocating for the kmem_cache_node. This is used for bootstrapping
3339 * memory on a fresh node that has no slab structures yet.
3341 static void early_kmem_cache_node_alloc(int node
)
3344 struct kmem_cache_node
*n
;
3346 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3348 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3351 if (page_to_nid(page
) != node
) {
3352 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3353 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3358 #ifdef CONFIG_SLUB_DEBUG
3359 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3360 init_tracking(kmem_cache_node
, n
);
3362 n
= kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3364 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3367 kmem_cache_node
->node
[node
] = n
;
3368 init_kmem_cache_node(n
);
3369 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3372 * No locks need to be taken here as it has just been
3373 * initialized and there is no concurrent access.
3375 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3378 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3381 struct kmem_cache_node
*n
;
3383 for_each_kmem_cache_node(s
, node
, n
) {
3384 s
->node
[node
] = NULL
;
3385 kmem_cache_free(kmem_cache_node
, n
);
3389 void __kmem_cache_release(struct kmem_cache
*s
)
3391 cache_random_seq_destroy(s
);
3392 free_percpu(s
->cpu_slab
);
3393 free_kmem_cache_nodes(s
);
3396 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3400 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3401 struct kmem_cache_node
*n
;
3403 if (slab_state
== DOWN
) {
3404 early_kmem_cache_node_alloc(node
);
3407 n
= kmem_cache_alloc_node(kmem_cache_node
,
3411 free_kmem_cache_nodes(s
);
3415 init_kmem_cache_node(n
);
3421 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3423 if (min
< MIN_PARTIAL
)
3425 else if (min
> MAX_PARTIAL
)
3427 s
->min_partial
= min
;
3430 static void set_cpu_partial(struct kmem_cache
*s
)
3432 #ifdef CONFIG_SLUB_CPU_PARTIAL
3434 * cpu_partial determined the maximum number of objects kept in the
3435 * per cpu partial lists of a processor.
3437 * Per cpu partial lists mainly contain slabs that just have one
3438 * object freed. If they are used for allocation then they can be
3439 * filled up again with minimal effort. The slab will never hit the
3440 * per node partial lists and therefore no locking will be required.
3442 * This setting also determines
3444 * A) The number of objects from per cpu partial slabs dumped to the
3445 * per node list when we reach the limit.
3446 * B) The number of objects in cpu partial slabs to extract from the
3447 * per node list when we run out of per cpu objects. We only fetch
3448 * 50% to keep some capacity around for frees.
3450 if (!kmem_cache_has_cpu_partial(s
))
3452 else if (s
->size
>= PAGE_SIZE
)
3454 else if (s
->size
>= 1024)
3456 else if (s
->size
>= 256)
3457 s
->cpu_partial
= 13;
3459 s
->cpu_partial
= 30;
3464 * calculate_sizes() determines the order and the distribution of data within
3467 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3469 slab_flags_t flags
= s
->flags
;
3470 unsigned int size
= s
->object_size
;
3474 * Round up object size to the next word boundary. We can only
3475 * place the free pointer at word boundaries and this determines
3476 * the possible location of the free pointer.
3478 size
= ALIGN(size
, sizeof(void *));
3480 #ifdef CONFIG_SLUB_DEBUG
3482 * Determine if we can poison the object itself. If the user of
3483 * the slab may touch the object after free or before allocation
3484 * then we should never poison the object itself.
3486 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3488 s
->flags
|= __OBJECT_POISON
;
3490 s
->flags
&= ~__OBJECT_POISON
;
3494 * If we are Redzoning then check if there is some space between the
3495 * end of the object and the free pointer. If not then add an
3496 * additional word to have some bytes to store Redzone information.
3498 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3499 size
+= sizeof(void *);
3503 * With that we have determined the number of bytes in actual use
3504 * by the object. This is the potential offset to the free pointer.
3508 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3511 * Relocate free pointer after the object if it is not
3512 * permitted to overwrite the first word of the object on
3515 * This is the case if we do RCU, have a constructor or
3516 * destructor or are poisoning the objects.
3519 size
+= sizeof(void *);
3522 #ifdef CONFIG_SLUB_DEBUG
3523 if (flags
& SLAB_STORE_USER
)
3525 * Need to store information about allocs and frees after
3528 size
+= 2 * sizeof(struct track
);
3531 kasan_cache_create(s
, &size
, &s
->flags
);
3532 #ifdef CONFIG_SLUB_DEBUG
3533 if (flags
& SLAB_RED_ZONE
) {
3535 * Add some empty padding so that we can catch
3536 * overwrites from earlier objects rather than let
3537 * tracking information or the free pointer be
3538 * corrupted if a user writes before the start
3541 size
+= sizeof(void *);
3543 s
->red_left_pad
= sizeof(void *);
3544 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3545 size
+= s
->red_left_pad
;
3550 * SLUB stores one object immediately after another beginning from
3551 * offset 0. In order to align the objects we have to simply size
3552 * each object to conform to the alignment.
3554 size
= ALIGN(size
, s
->align
);
3556 if (forced_order
>= 0)
3557 order
= forced_order
;
3559 order
= calculate_order(size
);
3566 s
->allocflags
|= __GFP_COMP
;
3568 if (s
->flags
& SLAB_CACHE_DMA
)
3569 s
->allocflags
|= GFP_DMA
;
3571 if (s
->flags
& SLAB_CACHE_DMA32
)
3572 s
->allocflags
|= GFP_DMA32
;
3574 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3575 s
->allocflags
|= __GFP_RECLAIMABLE
;
3578 * Determine the number of objects per slab
3580 s
->oo
= oo_make(order
, size
);
3581 s
->min
= oo_make(get_order(size
), size
);
3582 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3585 return !!oo_objects(s
->oo
);
3588 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3590 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3591 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3592 s
->random
= get_random_long();
3595 if (!calculate_sizes(s
, -1))
3597 if (disable_higher_order_debug
) {
3599 * Disable debugging flags that store metadata if the min slab
3602 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3603 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3605 if (!calculate_sizes(s
, -1))
3610 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3611 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3612 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3613 /* Enable fast mode */
3614 s
->flags
|= __CMPXCHG_DOUBLE
;
3618 * The larger the object size is, the more pages we want on the partial
3619 * list to avoid pounding the page allocator excessively.
3621 set_min_partial(s
, ilog2(s
->size
) / 2);
3626 s
->remote_node_defrag_ratio
= 1000;
3629 /* Initialize the pre-computed randomized freelist if slab is up */
3630 if (slab_state
>= UP
) {
3631 if (init_cache_random_seq(s
))
3635 if (!init_kmem_cache_nodes(s
))
3638 if (alloc_kmem_cache_cpus(s
))
3641 free_kmem_cache_nodes(s
);
3646 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3649 #ifdef CONFIG_SLUB_DEBUG
3650 void *addr
= page_address(page
);
3652 unsigned long *map
= bitmap_zalloc(page
->objects
, GFP_ATOMIC
);
3655 slab_err(s
, page
, text
, s
->name
);
3658 get_map(s
, page
, map
);
3659 for_each_object(p
, s
, addr
, page
->objects
) {
3661 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3662 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3663 print_tracking(s
, p
);
3672 * Attempt to free all partial slabs on a node.
3673 * This is called from __kmem_cache_shutdown(). We must take list_lock
3674 * because sysfs file might still access partial list after the shutdowning.
3676 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3679 struct page
*page
, *h
;
3681 BUG_ON(irqs_disabled());
3682 spin_lock_irq(&n
->list_lock
);
3683 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3685 remove_partial(n
, page
);
3686 list_add(&page
->slab_list
, &discard
);
3688 list_slab_objects(s
, page
,
3689 "Objects remaining in %s on __kmem_cache_shutdown()");
3692 spin_unlock_irq(&n
->list_lock
);
3694 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3695 discard_slab(s
, page
);
3698 bool __kmem_cache_empty(struct kmem_cache
*s
)
3701 struct kmem_cache_node
*n
;
3703 for_each_kmem_cache_node(s
, node
, n
)
3704 if (n
->nr_partial
|| slabs_node(s
, node
))
3710 * Release all resources used by a slab cache.
3712 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3715 struct kmem_cache_node
*n
;
3718 /* Attempt to free all objects */
3719 for_each_kmem_cache_node(s
, node
, n
) {
3721 if (n
->nr_partial
|| slabs_node(s
, node
))
3724 sysfs_slab_remove(s
);
3728 /********************************************************************
3730 *******************************************************************/
3732 static int __init
setup_slub_min_order(char *str
)
3734 get_option(&str
, (int *)&slub_min_order
);
3739 __setup("slub_min_order=", setup_slub_min_order
);
3741 static int __init
setup_slub_max_order(char *str
)
3743 get_option(&str
, (int *)&slub_max_order
);
3744 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
3749 __setup("slub_max_order=", setup_slub_max_order
);
3751 static int __init
setup_slub_min_objects(char *str
)
3753 get_option(&str
, (int *)&slub_min_objects
);
3758 __setup("slub_min_objects=", setup_slub_min_objects
);
3760 void *__kmalloc(size_t size
, gfp_t flags
)
3762 struct kmem_cache
*s
;
3765 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3766 return kmalloc_large(size
, flags
);
3768 s
= kmalloc_slab(size
, flags
);
3770 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3773 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3775 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3777 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3781 EXPORT_SYMBOL(__kmalloc
);
3784 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3789 flags
|= __GFP_COMP
;
3790 page
= alloc_pages_node(node
, flags
, get_order(size
));
3792 ptr
= page_address(page
);
3794 return kmalloc_large_node_hook(ptr
, size
, flags
);
3797 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3799 struct kmem_cache
*s
;
3802 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3803 ret
= kmalloc_large_node(size
, flags
, node
);
3805 trace_kmalloc_node(_RET_IP_
, ret
,
3806 size
, PAGE_SIZE
<< get_order(size
),
3812 s
= kmalloc_slab(size
, flags
);
3814 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3817 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3819 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3821 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
3825 EXPORT_SYMBOL(__kmalloc_node
);
3826 #endif /* CONFIG_NUMA */
3828 #ifdef CONFIG_HARDENED_USERCOPY
3830 * Rejects incorrectly sized objects and objects that are to be copied
3831 * to/from userspace but do not fall entirely within the containing slab
3832 * cache's usercopy region.
3834 * Returns NULL if check passes, otherwise const char * to name of cache
3835 * to indicate an error.
3837 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
3840 struct kmem_cache
*s
;
3841 unsigned int offset
;
3844 ptr
= kasan_reset_tag(ptr
);
3846 /* Find object and usable object size. */
3847 s
= page
->slab_cache
;
3849 /* Reject impossible pointers. */
3850 if (ptr
< page_address(page
))
3851 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
3854 /* Find offset within object. */
3855 offset
= (ptr
- page_address(page
)) % s
->size
;
3857 /* Adjust for redzone and reject if within the redzone. */
3858 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3859 if (offset
< s
->red_left_pad
)
3860 usercopy_abort("SLUB object in left red zone",
3861 s
->name
, to_user
, offset
, n
);
3862 offset
-= s
->red_left_pad
;
3865 /* Allow address range falling entirely within usercopy region. */
3866 if (offset
>= s
->useroffset
&&
3867 offset
- s
->useroffset
<= s
->usersize
&&
3868 n
<= s
->useroffset
- offset
+ s
->usersize
)
3872 * If the copy is still within the allocated object, produce
3873 * a warning instead of rejecting the copy. This is intended
3874 * to be a temporary method to find any missing usercopy
3877 object_size
= slab_ksize(s
);
3878 if (usercopy_fallback
&&
3879 offset
<= object_size
&& n
<= object_size
- offset
) {
3880 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
3884 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
3886 #endif /* CONFIG_HARDENED_USERCOPY */
3888 size_t __ksize(const void *object
)
3892 if (unlikely(object
== ZERO_SIZE_PTR
))
3895 page
= virt_to_head_page(object
);
3897 if (unlikely(!PageSlab(page
))) {
3898 WARN_ON(!PageCompound(page
));
3899 return PAGE_SIZE
<< compound_order(page
);
3902 return slab_ksize(page
->slab_cache
);
3904 EXPORT_SYMBOL(__ksize
);
3906 void kfree(const void *x
)
3909 void *object
= (void *)x
;
3911 trace_kfree(_RET_IP_
, x
);
3913 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3916 page
= virt_to_head_page(x
);
3917 if (unlikely(!PageSlab(page
))) {
3918 BUG_ON(!PageCompound(page
));
3920 __free_pages(page
, compound_order(page
));
3923 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3925 EXPORT_SYMBOL(kfree
);
3927 #define SHRINK_PROMOTE_MAX 32
3930 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3931 * up most to the head of the partial lists. New allocations will then
3932 * fill those up and thus they can be removed from the partial lists.
3934 * The slabs with the least items are placed last. This results in them
3935 * being allocated from last increasing the chance that the last objects
3936 * are freed in them.
3938 int __kmem_cache_shrink(struct kmem_cache
*s
)
3942 struct kmem_cache_node
*n
;
3945 struct list_head discard
;
3946 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3947 unsigned long flags
;
3951 for_each_kmem_cache_node(s
, node
, n
) {
3952 INIT_LIST_HEAD(&discard
);
3953 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3954 INIT_LIST_HEAD(promote
+ i
);
3956 spin_lock_irqsave(&n
->list_lock
, flags
);
3959 * Build lists of slabs to discard or promote.
3961 * Note that concurrent frees may occur while we hold the
3962 * list_lock. page->inuse here is the upper limit.
3964 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
3965 int free
= page
->objects
- page
->inuse
;
3967 /* Do not reread page->inuse */
3970 /* We do not keep full slabs on the list */
3973 if (free
== page
->objects
) {
3974 list_move(&page
->slab_list
, &discard
);
3976 } else if (free
<= SHRINK_PROMOTE_MAX
)
3977 list_move(&page
->slab_list
, promote
+ free
- 1);
3981 * Promote the slabs filled up most to the head of the
3984 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3985 list_splice(promote
+ i
, &n
->partial
);
3987 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3989 /* Release empty slabs */
3990 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
3991 discard_slab(s
, page
);
3993 if (slabs_node(s
, node
))
4001 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache
*s
)
4004 * Called with all the locks held after a sched RCU grace period.
4005 * Even if @s becomes empty after shrinking, we can't know that @s
4006 * doesn't have allocations already in-flight and thus can't
4007 * destroy @s until the associated memcg is released.
4009 * However, let's remove the sysfs files for empty caches here.
4010 * Each cache has a lot of interface files which aren't
4011 * particularly useful for empty draining caches; otherwise, we can
4012 * easily end up with millions of unnecessary sysfs files on
4013 * systems which have a lot of memory and transient cgroups.
4015 if (!__kmem_cache_shrink(s
))
4016 sysfs_slab_remove(s
);
4019 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4022 * Disable empty slabs caching. Used to avoid pinning offline
4023 * memory cgroups by kmem pages that can be freed.
4025 slub_set_cpu_partial(s
, 0);
4028 #endif /* CONFIG_MEMCG */
4030 static int slab_mem_going_offline_callback(void *arg
)
4032 struct kmem_cache
*s
;
4034 mutex_lock(&slab_mutex
);
4035 list_for_each_entry(s
, &slab_caches
, list
)
4036 __kmem_cache_shrink(s
);
4037 mutex_unlock(&slab_mutex
);
4042 static void slab_mem_offline_callback(void *arg
)
4044 struct kmem_cache_node
*n
;
4045 struct kmem_cache
*s
;
4046 struct memory_notify
*marg
= arg
;
4049 offline_node
= marg
->status_change_nid_normal
;
4052 * If the node still has available memory. we need kmem_cache_node
4055 if (offline_node
< 0)
4058 mutex_lock(&slab_mutex
);
4059 list_for_each_entry(s
, &slab_caches
, list
) {
4060 n
= get_node(s
, offline_node
);
4063 * if n->nr_slabs > 0, slabs still exist on the node
4064 * that is going down. We were unable to free them,
4065 * and offline_pages() function shouldn't call this
4066 * callback. So, we must fail.
4068 BUG_ON(slabs_node(s
, offline_node
));
4070 s
->node
[offline_node
] = NULL
;
4071 kmem_cache_free(kmem_cache_node
, n
);
4074 mutex_unlock(&slab_mutex
);
4077 static int slab_mem_going_online_callback(void *arg
)
4079 struct kmem_cache_node
*n
;
4080 struct kmem_cache
*s
;
4081 struct memory_notify
*marg
= arg
;
4082 int nid
= marg
->status_change_nid_normal
;
4086 * If the node's memory is already available, then kmem_cache_node is
4087 * already created. Nothing to do.
4093 * We are bringing a node online. No memory is available yet. We must
4094 * allocate a kmem_cache_node structure in order to bring the node
4097 mutex_lock(&slab_mutex
);
4098 list_for_each_entry(s
, &slab_caches
, list
) {
4100 * XXX: kmem_cache_alloc_node will fallback to other nodes
4101 * since memory is not yet available from the node that
4104 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4109 init_kmem_cache_node(n
);
4113 mutex_unlock(&slab_mutex
);
4117 static int slab_memory_callback(struct notifier_block
*self
,
4118 unsigned long action
, void *arg
)
4123 case MEM_GOING_ONLINE
:
4124 ret
= slab_mem_going_online_callback(arg
);
4126 case MEM_GOING_OFFLINE
:
4127 ret
= slab_mem_going_offline_callback(arg
);
4130 case MEM_CANCEL_ONLINE
:
4131 slab_mem_offline_callback(arg
);
4134 case MEM_CANCEL_OFFLINE
:
4138 ret
= notifier_from_errno(ret
);
4144 static struct notifier_block slab_memory_callback_nb
= {
4145 .notifier_call
= slab_memory_callback
,
4146 .priority
= SLAB_CALLBACK_PRI
,
4149 /********************************************************************
4150 * Basic setup of slabs
4151 *******************************************************************/
4154 * Used for early kmem_cache structures that were allocated using
4155 * the page allocator. Allocate them properly then fix up the pointers
4156 * that may be pointing to the wrong kmem_cache structure.
4159 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4162 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4163 struct kmem_cache_node
*n
;
4165 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4168 * This runs very early, and only the boot processor is supposed to be
4169 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4172 __flush_cpu_slab(s
, smp_processor_id());
4173 for_each_kmem_cache_node(s
, node
, n
) {
4176 list_for_each_entry(p
, &n
->partial
, slab_list
)
4179 #ifdef CONFIG_SLUB_DEBUG
4180 list_for_each_entry(p
, &n
->full
, slab_list
)
4184 slab_init_memcg_params(s
);
4185 list_add(&s
->list
, &slab_caches
);
4186 memcg_link_cache(s
, NULL
);
4190 void __init
kmem_cache_init(void)
4192 static __initdata
struct kmem_cache boot_kmem_cache
,
4193 boot_kmem_cache_node
;
4195 if (debug_guardpage_minorder())
4198 kmem_cache_node
= &boot_kmem_cache_node
;
4199 kmem_cache
= &boot_kmem_cache
;
4201 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4202 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4204 register_hotmemory_notifier(&slab_memory_callback_nb
);
4206 /* Able to allocate the per node structures */
4207 slab_state
= PARTIAL
;
4209 create_boot_cache(kmem_cache
, "kmem_cache",
4210 offsetof(struct kmem_cache
, node
) +
4211 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4212 SLAB_HWCACHE_ALIGN
, 0, 0);
4214 kmem_cache
= bootstrap(&boot_kmem_cache
);
4215 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4217 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4218 setup_kmalloc_cache_index_table();
4219 create_kmalloc_caches(0);
4221 /* Setup random freelists for each cache */
4222 init_freelist_randomization();
4224 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4227 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4229 slub_min_order
, slub_max_order
, slub_min_objects
,
4230 nr_cpu_ids
, nr_node_ids
);
4233 void __init
kmem_cache_init_late(void)
4238 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4239 slab_flags_t flags
, void (*ctor
)(void *))
4241 struct kmem_cache
*s
, *c
;
4243 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4248 * Adjust the object sizes so that we clear
4249 * the complete object on kzalloc.
4251 s
->object_size
= max(s
->object_size
, size
);
4252 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4254 for_each_memcg_cache(c
, s
) {
4255 c
->object_size
= s
->object_size
;
4256 c
->inuse
= max(c
->inuse
, ALIGN(size
, sizeof(void *)));
4259 if (sysfs_slab_alias(s
, name
)) {
4268 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4272 err
= kmem_cache_open(s
, flags
);
4276 /* Mutex is not taken during early boot */
4277 if (slab_state
<= UP
)
4280 memcg_propagate_slab_attrs(s
);
4281 err
= sysfs_slab_add(s
);
4283 __kmem_cache_release(s
);
4288 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4290 struct kmem_cache
*s
;
4293 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4294 return kmalloc_large(size
, gfpflags
);
4296 s
= kmalloc_slab(size
, gfpflags
);
4298 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4301 ret
= slab_alloc(s
, gfpflags
, caller
);
4303 /* Honor the call site pointer we received. */
4304 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4310 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4311 int node
, unsigned long caller
)
4313 struct kmem_cache
*s
;
4316 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4317 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4319 trace_kmalloc_node(caller
, ret
,
4320 size
, PAGE_SIZE
<< get_order(size
),
4326 s
= kmalloc_slab(size
, gfpflags
);
4328 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4331 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4333 /* Honor the call site pointer we received. */
4334 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4341 static int count_inuse(struct page
*page
)
4346 static int count_total(struct page
*page
)
4348 return page
->objects
;
4352 #ifdef CONFIG_SLUB_DEBUG
4353 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4357 void *addr
= page_address(page
);
4359 if (!check_slab(s
, page
) ||
4360 !on_freelist(s
, page
, NULL
))
4363 /* Now we know that a valid freelist exists */
4364 bitmap_zero(map
, page
->objects
);
4366 get_map(s
, page
, map
);
4367 for_each_object(p
, s
, addr
, page
->objects
) {
4368 if (test_bit(slab_index(p
, s
, addr
), map
))
4369 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4373 for_each_object(p
, s
, addr
, page
->objects
)
4374 if (!test_bit(slab_index(p
, s
, addr
), map
))
4375 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4380 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4384 validate_slab(s
, page
, map
);
4388 static int validate_slab_node(struct kmem_cache
*s
,
4389 struct kmem_cache_node
*n
, unsigned long *map
)
4391 unsigned long count
= 0;
4393 unsigned long flags
;
4395 spin_lock_irqsave(&n
->list_lock
, flags
);
4397 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4398 validate_slab_slab(s
, page
, map
);
4401 if (count
!= n
->nr_partial
)
4402 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4403 s
->name
, count
, n
->nr_partial
);
4405 if (!(s
->flags
& SLAB_STORE_USER
))
4408 list_for_each_entry(page
, &n
->full
, slab_list
) {
4409 validate_slab_slab(s
, page
, map
);
4412 if (count
!= atomic_long_read(&n
->nr_slabs
))
4413 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4414 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4417 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4421 static long validate_slab_cache(struct kmem_cache
*s
)
4424 unsigned long count
= 0;
4425 struct kmem_cache_node
*n
;
4426 unsigned long *map
= bitmap_alloc(oo_objects(s
->max
), GFP_KERNEL
);
4432 for_each_kmem_cache_node(s
, node
, n
)
4433 count
+= validate_slab_node(s
, n
, map
);
4438 * Generate lists of code addresses where slabcache objects are allocated
4443 unsigned long count
;
4450 DECLARE_BITMAP(cpus
, NR_CPUS
);
4456 unsigned long count
;
4457 struct location
*loc
;
4460 static void free_loc_track(struct loc_track
*t
)
4463 free_pages((unsigned long)t
->loc
,
4464 get_order(sizeof(struct location
) * t
->max
));
4467 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4472 order
= get_order(sizeof(struct location
) * max
);
4474 l
= (void *)__get_free_pages(flags
, order
);
4479 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4487 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4488 const struct track
*track
)
4490 long start
, end
, pos
;
4492 unsigned long caddr
;
4493 unsigned long age
= jiffies
- track
->when
;
4499 pos
= start
+ (end
- start
+ 1) / 2;
4502 * There is nothing at "end". If we end up there
4503 * we need to add something to before end.
4508 caddr
= t
->loc
[pos
].addr
;
4509 if (track
->addr
== caddr
) {
4515 if (age
< l
->min_time
)
4517 if (age
> l
->max_time
)
4520 if (track
->pid
< l
->min_pid
)
4521 l
->min_pid
= track
->pid
;
4522 if (track
->pid
> l
->max_pid
)
4523 l
->max_pid
= track
->pid
;
4525 cpumask_set_cpu(track
->cpu
,
4526 to_cpumask(l
->cpus
));
4528 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4532 if (track
->addr
< caddr
)
4539 * Not found. Insert new tracking element.
4541 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4547 (t
->count
- pos
) * sizeof(struct location
));
4550 l
->addr
= track
->addr
;
4554 l
->min_pid
= track
->pid
;
4555 l
->max_pid
= track
->pid
;
4556 cpumask_clear(to_cpumask(l
->cpus
));
4557 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4558 nodes_clear(l
->nodes
);
4559 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4563 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4564 struct page
*page
, enum track_item alloc
,
4567 void *addr
= page_address(page
);
4570 bitmap_zero(map
, page
->objects
);
4571 get_map(s
, page
, map
);
4573 for_each_object(p
, s
, addr
, page
->objects
)
4574 if (!test_bit(slab_index(p
, s
, addr
), map
))
4575 add_location(t
, s
, get_track(s
, p
, alloc
));
4578 static int list_locations(struct kmem_cache
*s
, char *buf
,
4579 enum track_item alloc
)
4583 struct loc_track t
= { 0, 0, NULL
};
4585 struct kmem_cache_node
*n
;
4586 unsigned long *map
= bitmap_alloc(oo_objects(s
->max
), GFP_KERNEL
);
4588 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4591 return sprintf(buf
, "Out of memory\n");
4593 /* Push back cpu slabs */
4596 for_each_kmem_cache_node(s
, node
, n
) {
4597 unsigned long flags
;
4600 if (!atomic_long_read(&n
->nr_slabs
))
4603 spin_lock_irqsave(&n
->list_lock
, flags
);
4604 list_for_each_entry(page
, &n
->partial
, slab_list
)
4605 process_slab(&t
, s
, page
, alloc
, map
);
4606 list_for_each_entry(page
, &n
->full
, slab_list
)
4607 process_slab(&t
, s
, page
, alloc
, map
);
4608 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4611 for (i
= 0; i
< t
.count
; i
++) {
4612 struct location
*l
= &t
.loc
[i
];
4614 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4616 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4619 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4621 len
+= sprintf(buf
+ len
, "<not-available>");
4623 if (l
->sum_time
!= l
->min_time
) {
4624 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4626 (long)div_u64(l
->sum_time
, l
->count
),
4629 len
+= sprintf(buf
+ len
, " age=%ld",
4632 if (l
->min_pid
!= l
->max_pid
)
4633 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4634 l
->min_pid
, l
->max_pid
);
4636 len
+= sprintf(buf
+ len
, " pid=%ld",
4639 if (num_online_cpus() > 1 &&
4640 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4641 len
< PAGE_SIZE
- 60)
4642 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4644 cpumask_pr_args(to_cpumask(l
->cpus
)));
4646 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4647 len
< PAGE_SIZE
- 60)
4648 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4650 nodemask_pr_args(&l
->nodes
));
4652 len
+= sprintf(buf
+ len
, "\n");
4658 len
+= sprintf(buf
, "No data\n");
4661 #endif /* CONFIG_SLUB_DEBUG */
4663 #ifdef SLUB_RESILIENCY_TEST
4664 static void __init
resiliency_test(void)
4667 int type
= KMALLOC_NORMAL
;
4669 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4671 pr_err("SLUB resiliency testing\n");
4672 pr_err("-----------------------\n");
4673 pr_err("A. Corruption after allocation\n");
4675 p
= kzalloc(16, GFP_KERNEL
);
4677 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4680 validate_slab_cache(kmalloc_caches
[type
][4]);
4682 /* Hmmm... The next two are dangerous */
4683 p
= kzalloc(32, GFP_KERNEL
);
4684 p
[32 + sizeof(void *)] = 0x34;
4685 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4687 pr_err("If allocated object is overwritten then not detectable\n\n");
4689 validate_slab_cache(kmalloc_caches
[type
][5]);
4690 p
= kzalloc(64, GFP_KERNEL
);
4691 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4693 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4695 pr_err("If allocated object is overwritten then not detectable\n\n");
4696 validate_slab_cache(kmalloc_caches
[type
][6]);
4698 pr_err("\nB. Corruption after free\n");
4699 p
= kzalloc(128, GFP_KERNEL
);
4702 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4703 validate_slab_cache(kmalloc_caches
[type
][7]);
4705 p
= kzalloc(256, GFP_KERNEL
);
4708 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4709 validate_slab_cache(kmalloc_caches
[type
][8]);
4711 p
= kzalloc(512, GFP_KERNEL
);
4714 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4715 validate_slab_cache(kmalloc_caches
[type
][9]);
4719 static void resiliency_test(void) {};
4721 #endif /* SLUB_RESILIENCY_TEST */
4724 enum slab_stat_type
{
4725 SL_ALL
, /* All slabs */
4726 SL_PARTIAL
, /* Only partially allocated slabs */
4727 SL_CPU
, /* Only slabs used for cpu caches */
4728 SL_OBJECTS
, /* Determine allocated objects not slabs */
4729 SL_TOTAL
/* Determine object capacity not slabs */
4732 #define SO_ALL (1 << SL_ALL)
4733 #define SO_PARTIAL (1 << SL_PARTIAL)
4734 #define SO_CPU (1 << SL_CPU)
4735 #define SO_OBJECTS (1 << SL_OBJECTS)
4736 #define SO_TOTAL (1 << SL_TOTAL)
4739 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4741 static int __init
setup_slub_memcg_sysfs(char *str
)
4745 if (get_option(&str
, &v
) > 0)
4746 memcg_sysfs_enabled
= v
;
4751 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4754 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4755 char *buf
, unsigned long flags
)
4757 unsigned long total
= 0;
4760 unsigned long *nodes
;
4762 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4766 if (flags
& SO_CPU
) {
4769 for_each_possible_cpu(cpu
) {
4770 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4775 page
= READ_ONCE(c
->page
);
4779 node
= page_to_nid(page
);
4780 if (flags
& SO_TOTAL
)
4782 else if (flags
& SO_OBJECTS
)
4790 page
= slub_percpu_partial_read_once(c
);
4792 node
= page_to_nid(page
);
4793 if (flags
& SO_TOTAL
)
4795 else if (flags
& SO_OBJECTS
)
4806 #ifdef CONFIG_SLUB_DEBUG
4807 if (flags
& SO_ALL
) {
4808 struct kmem_cache_node
*n
;
4810 for_each_kmem_cache_node(s
, node
, n
) {
4812 if (flags
& SO_TOTAL
)
4813 x
= atomic_long_read(&n
->total_objects
);
4814 else if (flags
& SO_OBJECTS
)
4815 x
= atomic_long_read(&n
->total_objects
) -
4816 count_partial(n
, count_free
);
4818 x
= atomic_long_read(&n
->nr_slabs
);
4825 if (flags
& SO_PARTIAL
) {
4826 struct kmem_cache_node
*n
;
4828 for_each_kmem_cache_node(s
, node
, n
) {
4829 if (flags
& SO_TOTAL
)
4830 x
= count_partial(n
, count_total
);
4831 else if (flags
& SO_OBJECTS
)
4832 x
= count_partial(n
, count_inuse
);
4839 x
= sprintf(buf
, "%lu", total
);
4841 for (node
= 0; node
< nr_node_ids
; node
++)
4843 x
+= sprintf(buf
+ x
, " N%d=%lu",
4848 return x
+ sprintf(buf
+ x
, "\n");
4851 #ifdef CONFIG_SLUB_DEBUG
4852 static int any_slab_objects(struct kmem_cache
*s
)
4855 struct kmem_cache_node
*n
;
4857 for_each_kmem_cache_node(s
, node
, n
)
4858 if (atomic_long_read(&n
->total_objects
))
4865 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4866 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4868 struct slab_attribute
{
4869 struct attribute attr
;
4870 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4871 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4874 #define SLAB_ATTR_RO(_name) \
4875 static struct slab_attribute _name##_attr = \
4876 __ATTR(_name, 0400, _name##_show, NULL)
4878 #define SLAB_ATTR(_name) \
4879 static struct slab_attribute _name##_attr = \
4880 __ATTR(_name, 0600, _name##_show, _name##_store)
4882 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4884 return sprintf(buf
, "%u\n", s
->size
);
4886 SLAB_ATTR_RO(slab_size
);
4888 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4890 return sprintf(buf
, "%u\n", s
->align
);
4892 SLAB_ATTR_RO(align
);
4894 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4896 return sprintf(buf
, "%u\n", s
->object_size
);
4898 SLAB_ATTR_RO(object_size
);
4900 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4902 return sprintf(buf
, "%u\n", oo_objects(s
->oo
));
4904 SLAB_ATTR_RO(objs_per_slab
);
4906 static ssize_t
order_store(struct kmem_cache
*s
,
4907 const char *buf
, size_t length
)
4912 err
= kstrtouint(buf
, 10, &order
);
4916 if (order
> slub_max_order
|| order
< slub_min_order
)
4919 calculate_sizes(s
, order
);
4923 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4925 return sprintf(buf
, "%u\n", oo_order(s
->oo
));
4929 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4931 return sprintf(buf
, "%lu\n", s
->min_partial
);
4934 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4940 err
= kstrtoul(buf
, 10, &min
);
4944 set_min_partial(s
, min
);
4947 SLAB_ATTR(min_partial
);
4949 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4951 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4954 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4957 unsigned int objects
;
4960 err
= kstrtouint(buf
, 10, &objects
);
4963 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4966 slub_set_cpu_partial(s
, objects
);
4970 SLAB_ATTR(cpu_partial
);
4972 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4976 return sprintf(buf
, "%pS\n", s
->ctor
);
4980 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4982 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4984 SLAB_ATTR_RO(aliases
);
4986 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4988 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4990 SLAB_ATTR_RO(partial
);
4992 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4994 return show_slab_objects(s
, buf
, SO_CPU
);
4996 SLAB_ATTR_RO(cpu_slabs
);
4998 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5000 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5002 SLAB_ATTR_RO(objects
);
5004 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5006 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5008 SLAB_ATTR_RO(objects_partial
);
5010 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5017 for_each_online_cpu(cpu
) {
5020 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5023 pages
+= page
->pages
;
5024 objects
+= page
->pobjects
;
5028 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5031 for_each_online_cpu(cpu
) {
5034 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5036 if (page
&& len
< PAGE_SIZE
- 20)
5037 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5038 page
->pobjects
, page
->pages
);
5041 return len
+ sprintf(buf
+ len
, "\n");
5043 SLAB_ATTR_RO(slabs_cpu_partial
);
5045 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5047 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5050 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5051 const char *buf
, size_t length
)
5053 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5055 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5058 SLAB_ATTR(reclaim_account
);
5060 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5062 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5064 SLAB_ATTR_RO(hwcache_align
);
5066 #ifdef CONFIG_ZONE_DMA
5067 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5069 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5071 SLAB_ATTR_RO(cache_dma
);
5074 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5076 return sprintf(buf
, "%u\n", s
->usersize
);
5078 SLAB_ATTR_RO(usersize
);
5080 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5082 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5084 SLAB_ATTR_RO(destroy_by_rcu
);
5086 #ifdef CONFIG_SLUB_DEBUG
5087 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5089 return show_slab_objects(s
, buf
, SO_ALL
);
5091 SLAB_ATTR_RO(slabs
);
5093 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5095 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5097 SLAB_ATTR_RO(total_objects
);
5099 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5101 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5104 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5105 const char *buf
, size_t length
)
5107 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5108 if (buf
[0] == '1') {
5109 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5110 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5114 SLAB_ATTR(sanity_checks
);
5116 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5118 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5121 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5125 * Tracing a merged cache is going to give confusing results
5126 * as well as cause other issues like converting a mergeable
5127 * cache into an umergeable one.
5129 if (s
->refcount
> 1)
5132 s
->flags
&= ~SLAB_TRACE
;
5133 if (buf
[0] == '1') {
5134 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5135 s
->flags
|= SLAB_TRACE
;
5141 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5143 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5146 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5147 const char *buf
, size_t length
)
5149 if (any_slab_objects(s
))
5152 s
->flags
&= ~SLAB_RED_ZONE
;
5153 if (buf
[0] == '1') {
5154 s
->flags
|= SLAB_RED_ZONE
;
5156 calculate_sizes(s
, -1);
5159 SLAB_ATTR(red_zone
);
5161 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5163 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5166 static ssize_t
poison_store(struct kmem_cache
*s
,
5167 const char *buf
, size_t length
)
5169 if (any_slab_objects(s
))
5172 s
->flags
&= ~SLAB_POISON
;
5173 if (buf
[0] == '1') {
5174 s
->flags
|= SLAB_POISON
;
5176 calculate_sizes(s
, -1);
5181 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5183 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5186 static ssize_t
store_user_store(struct kmem_cache
*s
,
5187 const char *buf
, size_t length
)
5189 if (any_slab_objects(s
))
5192 s
->flags
&= ~SLAB_STORE_USER
;
5193 if (buf
[0] == '1') {
5194 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5195 s
->flags
|= SLAB_STORE_USER
;
5197 calculate_sizes(s
, -1);
5200 SLAB_ATTR(store_user
);
5202 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5207 static ssize_t
validate_store(struct kmem_cache
*s
,
5208 const char *buf
, size_t length
)
5212 if (buf
[0] == '1') {
5213 ret
= validate_slab_cache(s
);
5219 SLAB_ATTR(validate
);
5221 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5223 if (!(s
->flags
& SLAB_STORE_USER
))
5225 return list_locations(s
, buf
, TRACK_ALLOC
);
5227 SLAB_ATTR_RO(alloc_calls
);
5229 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5231 if (!(s
->flags
& SLAB_STORE_USER
))
5233 return list_locations(s
, buf
, TRACK_FREE
);
5235 SLAB_ATTR_RO(free_calls
);
5236 #endif /* CONFIG_SLUB_DEBUG */
5238 #ifdef CONFIG_FAILSLAB
5239 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5241 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5244 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5247 if (s
->refcount
> 1)
5250 s
->flags
&= ~SLAB_FAILSLAB
;
5252 s
->flags
|= SLAB_FAILSLAB
;
5255 SLAB_ATTR(failslab
);
5258 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5263 static ssize_t
shrink_store(struct kmem_cache
*s
,
5264 const char *buf
, size_t length
)
5267 kmem_cache_shrink(s
);
5275 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5277 return sprintf(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5280 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5281 const char *buf
, size_t length
)
5286 err
= kstrtouint(buf
, 10, &ratio
);
5292 s
->remote_node_defrag_ratio
= ratio
* 10;
5296 SLAB_ATTR(remote_node_defrag_ratio
);
5299 #ifdef CONFIG_SLUB_STATS
5300 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5302 unsigned long sum
= 0;
5305 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5310 for_each_online_cpu(cpu
) {
5311 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5317 len
= sprintf(buf
, "%lu", sum
);
5320 for_each_online_cpu(cpu
) {
5321 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5322 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5326 return len
+ sprintf(buf
+ len
, "\n");
5329 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5333 for_each_online_cpu(cpu
)
5334 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5337 #define STAT_ATTR(si, text) \
5338 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5340 return show_stat(s, buf, si); \
5342 static ssize_t text##_store(struct kmem_cache *s, \
5343 const char *buf, size_t length) \
5345 if (buf[0] != '0') \
5347 clear_stat(s, si); \
5352 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5353 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5354 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5355 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5356 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5357 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5358 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5359 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5360 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5361 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5362 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5363 STAT_ATTR(FREE_SLAB
, free_slab
);
5364 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5365 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5366 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5367 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5368 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5369 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5370 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5371 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5372 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5373 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5374 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5375 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5376 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5377 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5378 #endif /* CONFIG_SLUB_STATS */
5380 static struct attribute
*slab_attrs
[] = {
5381 &slab_size_attr
.attr
,
5382 &object_size_attr
.attr
,
5383 &objs_per_slab_attr
.attr
,
5385 &min_partial_attr
.attr
,
5386 &cpu_partial_attr
.attr
,
5388 &objects_partial_attr
.attr
,
5390 &cpu_slabs_attr
.attr
,
5394 &hwcache_align_attr
.attr
,
5395 &reclaim_account_attr
.attr
,
5396 &destroy_by_rcu_attr
.attr
,
5398 &slabs_cpu_partial_attr
.attr
,
5399 #ifdef CONFIG_SLUB_DEBUG
5400 &total_objects_attr
.attr
,
5402 &sanity_checks_attr
.attr
,
5404 &red_zone_attr
.attr
,
5406 &store_user_attr
.attr
,
5407 &validate_attr
.attr
,
5408 &alloc_calls_attr
.attr
,
5409 &free_calls_attr
.attr
,
5411 #ifdef CONFIG_ZONE_DMA
5412 &cache_dma_attr
.attr
,
5415 &remote_node_defrag_ratio_attr
.attr
,
5417 #ifdef CONFIG_SLUB_STATS
5418 &alloc_fastpath_attr
.attr
,
5419 &alloc_slowpath_attr
.attr
,
5420 &free_fastpath_attr
.attr
,
5421 &free_slowpath_attr
.attr
,
5422 &free_frozen_attr
.attr
,
5423 &free_add_partial_attr
.attr
,
5424 &free_remove_partial_attr
.attr
,
5425 &alloc_from_partial_attr
.attr
,
5426 &alloc_slab_attr
.attr
,
5427 &alloc_refill_attr
.attr
,
5428 &alloc_node_mismatch_attr
.attr
,
5429 &free_slab_attr
.attr
,
5430 &cpuslab_flush_attr
.attr
,
5431 &deactivate_full_attr
.attr
,
5432 &deactivate_empty_attr
.attr
,
5433 &deactivate_to_head_attr
.attr
,
5434 &deactivate_to_tail_attr
.attr
,
5435 &deactivate_remote_frees_attr
.attr
,
5436 &deactivate_bypass_attr
.attr
,
5437 &order_fallback_attr
.attr
,
5438 &cmpxchg_double_fail_attr
.attr
,
5439 &cmpxchg_double_cpu_fail_attr
.attr
,
5440 &cpu_partial_alloc_attr
.attr
,
5441 &cpu_partial_free_attr
.attr
,
5442 &cpu_partial_node_attr
.attr
,
5443 &cpu_partial_drain_attr
.attr
,
5445 #ifdef CONFIG_FAILSLAB
5446 &failslab_attr
.attr
,
5448 &usersize_attr
.attr
,
5453 static const struct attribute_group slab_attr_group
= {
5454 .attrs
= slab_attrs
,
5457 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5458 struct attribute
*attr
,
5461 struct slab_attribute
*attribute
;
5462 struct kmem_cache
*s
;
5465 attribute
= to_slab_attr(attr
);
5468 if (!attribute
->show
)
5471 err
= attribute
->show(s
, buf
);
5476 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5477 struct attribute
*attr
,
5478 const char *buf
, size_t len
)
5480 struct slab_attribute
*attribute
;
5481 struct kmem_cache
*s
;
5484 attribute
= to_slab_attr(attr
);
5487 if (!attribute
->store
)
5490 err
= attribute
->store(s
, buf
, len
);
5492 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5493 struct kmem_cache
*c
;
5495 mutex_lock(&slab_mutex
);
5496 if (s
->max_attr_size
< len
)
5497 s
->max_attr_size
= len
;
5500 * This is a best effort propagation, so this function's return
5501 * value will be determined by the parent cache only. This is
5502 * basically because not all attributes will have a well
5503 * defined semantics for rollbacks - most of the actions will
5504 * have permanent effects.
5506 * Returning the error value of any of the children that fail
5507 * is not 100 % defined, in the sense that users seeing the
5508 * error code won't be able to know anything about the state of
5511 * Only returning the error code for the parent cache at least
5512 * has well defined semantics. The cache being written to
5513 * directly either failed or succeeded, in which case we loop
5514 * through the descendants with best-effort propagation.
5516 for_each_memcg_cache(c
, s
)
5517 attribute
->store(c
, buf
, len
);
5518 mutex_unlock(&slab_mutex
);
5524 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5528 char *buffer
= NULL
;
5529 struct kmem_cache
*root_cache
;
5531 if (is_root_cache(s
))
5534 root_cache
= s
->memcg_params
.root_cache
;
5537 * This mean this cache had no attribute written. Therefore, no point
5538 * in copying default values around
5540 if (!root_cache
->max_attr_size
)
5543 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5546 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5549 if (!attr
|| !attr
->store
|| !attr
->show
)
5553 * It is really bad that we have to allocate here, so we will
5554 * do it only as a fallback. If we actually allocate, though,
5555 * we can just use the allocated buffer until the end.
5557 * Most of the slub attributes will tend to be very small in
5558 * size, but sysfs allows buffers up to a page, so they can
5559 * theoretically happen.
5563 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5566 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5567 if (WARN_ON(!buffer
))
5572 len
= attr
->show(root_cache
, buf
);
5574 attr
->store(s
, buf
, len
);
5578 free_page((unsigned long)buffer
);
5579 #endif /* CONFIG_MEMCG */
5582 static void kmem_cache_release(struct kobject
*k
)
5584 slab_kmem_cache_release(to_slab(k
));
5587 static const struct sysfs_ops slab_sysfs_ops
= {
5588 .show
= slab_attr_show
,
5589 .store
= slab_attr_store
,
5592 static struct kobj_type slab_ktype
= {
5593 .sysfs_ops
= &slab_sysfs_ops
,
5594 .release
= kmem_cache_release
,
5597 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5599 struct kobj_type
*ktype
= get_ktype(kobj
);
5601 if (ktype
== &slab_ktype
)
5606 static const struct kset_uevent_ops slab_uevent_ops
= {
5607 .filter
= uevent_filter
,
5610 static struct kset
*slab_kset
;
5612 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5615 if (!is_root_cache(s
))
5616 return s
->memcg_params
.root_cache
->memcg_kset
;
5621 #define ID_STR_LENGTH 64
5623 /* Create a unique string id for a slab cache:
5625 * Format :[flags-]size
5627 static char *create_unique_id(struct kmem_cache
*s
)
5629 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5636 * First flags affecting slabcache operations. We will only
5637 * get here for aliasable slabs so we do not need to support
5638 * too many flags. The flags here must cover all flags that
5639 * are matched during merging to guarantee that the id is
5642 if (s
->flags
& SLAB_CACHE_DMA
)
5644 if (s
->flags
& SLAB_CACHE_DMA32
)
5646 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5648 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5650 if (s
->flags
& SLAB_ACCOUNT
)
5654 p
+= sprintf(p
, "%07u", s
->size
);
5656 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5660 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5662 struct kmem_cache
*s
=
5663 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5665 if (!s
->kobj
.state_in_sysfs
)
5667 * For a memcg cache, this may be called during
5668 * deactivation and again on shutdown. Remove only once.
5669 * A cache is never shut down before deactivation is
5670 * complete, so no need to worry about synchronization.
5675 kset_unregister(s
->memcg_kset
);
5677 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5679 kobject_put(&s
->kobj
);
5682 static int sysfs_slab_add(struct kmem_cache
*s
)
5686 struct kset
*kset
= cache_kset(s
);
5687 int unmergeable
= slab_unmergeable(s
);
5689 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5692 kobject_init(&s
->kobj
, &slab_ktype
);
5696 if (!unmergeable
&& disable_higher_order_debug
&&
5697 (slub_debug
& DEBUG_METADATA_FLAGS
))
5702 * Slabcache can never be merged so we can use the name proper.
5703 * This is typically the case for debug situations. In that
5704 * case we can catch duplicate names easily.
5706 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5710 * Create a unique name for the slab as a target
5713 name
= create_unique_id(s
);
5716 s
->kobj
.kset
= kset
;
5717 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5721 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5726 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5727 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5728 if (!s
->memcg_kset
) {
5735 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5737 /* Setup first alias */
5738 sysfs_slab_alias(s
, s
->name
);
5745 kobject_del(&s
->kobj
);
5749 static void sysfs_slab_remove(struct kmem_cache
*s
)
5751 if (slab_state
< FULL
)
5753 * Sysfs has not been setup yet so no need to remove the
5758 kobject_get(&s
->kobj
);
5759 schedule_work(&s
->kobj_remove_work
);
5762 void sysfs_slab_unlink(struct kmem_cache
*s
)
5764 if (slab_state
>= FULL
)
5765 kobject_del(&s
->kobj
);
5768 void sysfs_slab_release(struct kmem_cache
*s
)
5770 if (slab_state
>= FULL
)
5771 kobject_put(&s
->kobj
);
5775 * Need to buffer aliases during bootup until sysfs becomes
5776 * available lest we lose that information.
5778 struct saved_alias
{
5779 struct kmem_cache
*s
;
5781 struct saved_alias
*next
;
5784 static struct saved_alias
*alias_list
;
5786 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5788 struct saved_alias
*al
;
5790 if (slab_state
== FULL
) {
5792 * If we have a leftover link then remove it.
5794 sysfs_remove_link(&slab_kset
->kobj
, name
);
5795 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5798 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5804 al
->next
= alias_list
;
5809 static int __init
slab_sysfs_init(void)
5811 struct kmem_cache
*s
;
5814 mutex_lock(&slab_mutex
);
5816 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5818 mutex_unlock(&slab_mutex
);
5819 pr_err("Cannot register slab subsystem.\n");
5825 list_for_each_entry(s
, &slab_caches
, list
) {
5826 err
= sysfs_slab_add(s
);
5828 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5832 while (alias_list
) {
5833 struct saved_alias
*al
= alias_list
;
5835 alias_list
= alias_list
->next
;
5836 err
= sysfs_slab_alias(al
->s
, al
->name
);
5838 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5843 mutex_unlock(&slab_mutex
);
5848 __initcall(slab_sysfs_init
);
5849 #endif /* CONFIG_SYSFS */
5852 * The /proc/slabinfo ABI
5854 #ifdef CONFIG_SLUB_DEBUG
5855 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5857 unsigned long nr_slabs
= 0;
5858 unsigned long nr_objs
= 0;
5859 unsigned long nr_free
= 0;
5861 struct kmem_cache_node
*n
;
5863 for_each_kmem_cache_node(s
, node
, n
) {
5864 nr_slabs
+= node_nr_slabs(n
);
5865 nr_objs
+= node_nr_objs(n
);
5866 nr_free
+= count_partial(n
, count_free
);
5869 sinfo
->active_objs
= nr_objs
- nr_free
;
5870 sinfo
->num_objs
= nr_objs
;
5871 sinfo
->active_slabs
= nr_slabs
;
5872 sinfo
->num_slabs
= nr_slabs
;
5873 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5874 sinfo
->cache_order
= oo_order(s
->oo
);
5877 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5881 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5882 size_t count
, loff_t
*ppos
)
5886 #endif /* CONFIG_SLUB_DEBUG */