1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/notifier.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list. The processor that froze the slab is the one who can
63 * perform list operations on the page. Other processors may put objects
64 * onto the freelist but the processor that froze the slab is the only
65 * one that can retrieve the objects from the page's freelist.
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * Overloading of page flags that are otherwise used for LRU management.
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
119 static inline int kmem_cache_debug(struct kmem_cache
*s
)
121 #ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
128 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
130 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
131 p
+= s
->red_left_pad
;
136 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
138 #ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s
);
146 * Issues still to be resolved:
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
156 /* Enable to log cmpxchg failures */
157 #undef SLUB_DEBUG_CMPXCHG
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
163 #define MIN_PARTIAL 5
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
170 #define MAX_PARTIAL 10
172 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
179 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
188 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
191 #define OO_MASK ((1 << OO_SHIFT) - 1)
192 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
194 /* Internal SLUB flags */
196 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197 /* Use cmpxchg_double */
198 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
201 * Tracking user of a slab.
203 #define TRACK_ADDRS_COUNT 16
205 unsigned long addr
; /* Called from address */
206 #ifdef CONFIG_STACKTRACE
207 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
209 int cpu
; /* Was running on cpu */
210 int pid
; /* Pid context */
211 unsigned long when
; /* When did the operation occur */
214 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
217 static int sysfs_slab_add(struct kmem_cache
*);
218 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
219 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
220 static void sysfs_slab_remove(struct kmem_cache
*s
);
222 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
225 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
226 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
229 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
231 #ifdef CONFIG_SLUB_STATS
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
236 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
240 /********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
249 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
250 unsigned long ptr_addr
)
252 #ifdef CONFIG_SLAB_FREELIST_HARDENED
253 return (void *)((unsigned long)ptr
^ s
->random
^ ptr_addr
);
259 /* Returns the freelist pointer recorded at location ptr_addr. */
260 static inline void *freelist_dereference(const struct kmem_cache
*s
,
263 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
264 (unsigned long)ptr_addr
);
267 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
269 return freelist_dereference(s
, object
+ s
->offset
);
272 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
274 prefetch(object
+ s
->offset
);
277 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
279 unsigned long freepointer_addr
;
282 if (!debug_pagealloc_enabled())
283 return get_freepointer(s
, object
);
285 freepointer_addr
= (unsigned long)object
+ s
->offset
;
286 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
287 return freelist_ptr(s
, p
, freepointer_addr
);
290 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
292 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
294 #ifdef CONFIG_SLAB_FREELIST_HARDENED
295 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
298 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
301 /* Loop over all objects in a slab */
302 #define for_each_object(__p, __s, __addr, __objects) \
303 for (__p = fixup_red_left(__s, __addr); \
304 __p < (__addr) + (__objects) * (__s)->size; \
307 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
308 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
309 __idx <= __objects; \
310 __p += (__s)->size, __idx++)
312 /* Determine object index from a given position */
313 static inline unsigned int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
315 return (p
- addr
) / s
->size
;
318 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
320 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
323 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
326 struct kmem_cache_order_objects x
= {
327 (order
<< OO_SHIFT
) + order_objects(order
, size
)
333 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
335 return x
.x
>> OO_SHIFT
;
338 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
340 return x
.x
& OO_MASK
;
344 * Per slab locking using the pagelock
346 static __always_inline
void slab_lock(struct page
*page
)
348 VM_BUG_ON_PAGE(PageTail(page
), page
);
349 bit_spin_lock(PG_locked
, &page
->flags
);
352 static __always_inline
void slab_unlock(struct page
*page
)
354 VM_BUG_ON_PAGE(PageTail(page
), page
);
355 __bit_spin_unlock(PG_locked
, &page
->flags
);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
360 void *freelist_old
, unsigned long counters_old
,
361 void *freelist_new
, unsigned long counters_new
,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s
->flags
& __CMPXCHG_DOUBLE
) {
368 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
369 freelist_old
, counters_old
,
370 freelist_new
, counters_new
))
376 if (page
->freelist
== freelist_old
&&
377 page
->counters
== counters_old
) {
378 page
->freelist
= freelist_new
;
379 page
->counters
= counters_new
;
387 stat(s
, CMPXCHG_DOUBLE_FAIL
);
389 #ifdef SLUB_DEBUG_CMPXCHG
390 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
396 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
397 void *freelist_old
, unsigned long counters_old
,
398 void *freelist_new
, unsigned long counters_new
,
401 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
402 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
403 if (s
->flags
& __CMPXCHG_DOUBLE
) {
404 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
405 freelist_old
, counters_old
,
406 freelist_new
, counters_new
))
413 local_irq_save(flags
);
415 if (page
->freelist
== freelist_old
&&
416 page
->counters
== counters_old
) {
417 page
->freelist
= freelist_new
;
418 page
->counters
= counters_new
;
420 local_irq_restore(flags
);
424 local_irq_restore(flags
);
428 stat(s
, CMPXCHG_DOUBLE_FAIL
);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
447 void *addr
= page_address(page
);
449 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
450 set_bit(slab_index(p
, s
, addr
), map
);
453 static inline unsigned int size_from_object(struct kmem_cache
*s
)
455 if (s
->flags
& SLAB_RED_ZONE
)
456 return s
->size
- s
->red_left_pad
;
461 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
463 if (s
->flags
& SLAB_RED_ZONE
)
464 p
-= s
->red_left_pad
;
472 #if defined(CONFIG_SLUB_DEBUG_ON)
473 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
475 static slab_flags_t slub_debug
;
478 static char *slub_debug_slabs
;
479 static int disable_higher_order_debug
;
482 * slub is about to manipulate internal object metadata. This memory lies
483 * outside the range of the allocated object, so accessing it would normally
484 * be reported by kasan as a bounds error. metadata_access_enable() is used
485 * to tell kasan that these accesses are OK.
487 static inline void metadata_access_enable(void)
489 kasan_disable_current();
492 static inline void metadata_access_disable(void)
494 kasan_enable_current();
501 /* Verify that a pointer has an address that is valid within a slab page */
502 static inline int check_valid_pointer(struct kmem_cache
*s
,
503 struct page
*page
, void *object
)
510 base
= page_address(page
);
511 object
= restore_red_left(s
, object
);
512 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
513 (object
- base
) % s
->size
) {
520 static void print_section(char *level
, char *text
, u8
*addr
,
523 metadata_access_enable();
524 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
526 metadata_access_disable();
529 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
530 enum track_item alloc
)
535 p
= object
+ s
->offset
+ sizeof(void *);
537 p
= object
+ s
->inuse
;
542 static void set_track(struct kmem_cache
*s
, void *object
,
543 enum track_item alloc
, unsigned long addr
)
545 struct track
*p
= get_track(s
, object
, alloc
);
548 #ifdef CONFIG_STACKTRACE
549 struct stack_trace trace
;
552 trace
.nr_entries
= 0;
553 trace
.max_entries
= TRACK_ADDRS_COUNT
;
554 trace
.entries
= p
->addrs
;
556 metadata_access_enable();
557 save_stack_trace(&trace
);
558 metadata_access_disable();
560 /* See rant in lockdep.c */
561 if (trace
.nr_entries
!= 0 &&
562 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
565 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
569 p
->cpu
= smp_processor_id();
570 p
->pid
= current
->pid
;
573 memset(p
, 0, sizeof(struct track
));
576 static void init_tracking(struct kmem_cache
*s
, void *object
)
578 if (!(s
->flags
& SLAB_STORE_USER
))
581 set_track(s
, object
, TRACK_FREE
, 0UL);
582 set_track(s
, object
, TRACK_ALLOC
, 0UL);
585 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
590 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
591 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
592 #ifdef CONFIG_STACKTRACE
595 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
597 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
604 static void print_tracking(struct kmem_cache
*s
, void *object
)
606 unsigned long pr_time
= jiffies
;
607 if (!(s
->flags
& SLAB_STORE_USER
))
610 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
611 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
614 static void print_page_info(struct page
*page
)
616 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
617 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
621 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
623 struct va_format vaf
;
629 pr_err("=============================================================================\n");
630 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
631 pr_err("-----------------------------------------------------------------------------\n\n");
633 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
637 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
639 struct va_format vaf
;
645 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
649 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
651 unsigned int off
; /* Offset of last byte */
652 u8
*addr
= page_address(page
);
654 print_tracking(s
, p
);
656 print_page_info(page
);
658 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
659 p
, p
- addr
, get_freepointer(s
, p
));
661 if (s
->flags
& SLAB_RED_ZONE
)
662 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
664 else if (p
> addr
+ 16)
665 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
667 print_section(KERN_ERR
, "Object ", p
,
668 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
669 if (s
->flags
& SLAB_RED_ZONE
)
670 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
671 s
->inuse
- s
->object_size
);
674 off
= s
->offset
+ sizeof(void *);
678 if (s
->flags
& SLAB_STORE_USER
)
679 off
+= 2 * sizeof(struct track
);
681 off
+= kasan_metadata_size(s
);
683 if (off
!= size_from_object(s
))
684 /* Beginning of the filler is the free pointer */
685 print_section(KERN_ERR
, "Padding ", p
+ off
,
686 size_from_object(s
) - off
);
691 void object_err(struct kmem_cache
*s
, struct page
*page
,
692 u8
*object
, char *reason
)
694 slab_bug(s
, "%s", reason
);
695 print_trailer(s
, page
, object
);
698 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
699 const char *fmt
, ...)
705 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
707 slab_bug(s
, "%s", buf
);
708 print_page_info(page
);
712 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
716 if (s
->flags
& SLAB_RED_ZONE
)
717 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
719 if (s
->flags
& __OBJECT_POISON
) {
720 memset(p
, POISON_FREE
, s
->object_size
- 1);
721 p
[s
->object_size
- 1] = POISON_END
;
724 if (s
->flags
& SLAB_RED_ZONE
)
725 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
728 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
729 void *from
, void *to
)
731 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
732 memset(from
, data
, to
- from
);
735 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
736 u8
*object
, char *what
,
737 u8
*start
, unsigned int value
, unsigned int bytes
)
742 metadata_access_enable();
743 fault
= memchr_inv(start
, value
, bytes
);
744 metadata_access_disable();
749 while (end
> fault
&& end
[-1] == value
)
752 slab_bug(s
, "%s overwritten", what
);
753 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
754 fault
, end
- 1, fault
[0], value
);
755 print_trailer(s
, page
, object
);
757 restore_bytes(s
, what
, value
, fault
, end
);
765 * Bytes of the object to be managed.
766 * If the freepointer may overlay the object then the free
767 * pointer is the first word of the object.
769 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
772 * object + s->object_size
773 * Padding to reach word boundary. This is also used for Redzoning.
774 * Padding is extended by another word if Redzoning is enabled and
775 * object_size == inuse.
777 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
778 * 0xcc (RED_ACTIVE) for objects in use.
781 * Meta data starts here.
783 * A. Free pointer (if we cannot overwrite object on free)
784 * B. Tracking data for SLAB_STORE_USER
785 * C. Padding to reach required alignment boundary or at mininum
786 * one word if debugging is on to be able to detect writes
787 * before the word boundary.
789 * Padding is done using 0x5a (POISON_INUSE)
792 * Nothing is used beyond s->size.
794 * If slabcaches are merged then the object_size and inuse boundaries are mostly
795 * ignored. And therefore no slab options that rely on these boundaries
796 * may be used with merged slabcaches.
799 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
801 unsigned long off
= s
->inuse
; /* The end of info */
804 /* Freepointer is placed after the object. */
805 off
+= sizeof(void *);
807 if (s
->flags
& SLAB_STORE_USER
)
808 /* We also have user information there */
809 off
+= 2 * sizeof(struct track
);
811 off
+= kasan_metadata_size(s
);
813 if (size_from_object(s
) == off
)
816 return check_bytes_and_report(s
, page
, p
, "Object padding",
817 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
820 /* Check the pad bytes at the end of a slab page */
821 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
830 if (!(s
->flags
& SLAB_POISON
))
833 start
= page_address(page
);
834 length
= PAGE_SIZE
<< compound_order(page
);
835 end
= start
+ length
;
836 remainder
= length
% s
->size
;
840 pad
= end
- remainder
;
841 metadata_access_enable();
842 fault
= memchr_inv(pad
, POISON_INUSE
, remainder
);
843 metadata_access_disable();
846 while (end
> fault
&& end
[-1] == POISON_INUSE
)
849 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
850 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
852 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
856 static int check_object(struct kmem_cache
*s
, struct page
*page
,
857 void *object
, u8 val
)
860 u8
*endobject
= object
+ s
->object_size
;
862 if (s
->flags
& SLAB_RED_ZONE
) {
863 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
864 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
867 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
868 endobject
, val
, s
->inuse
- s
->object_size
))
871 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
872 check_bytes_and_report(s
, page
, p
, "Alignment padding",
873 endobject
, POISON_INUSE
,
874 s
->inuse
- s
->object_size
);
878 if (s
->flags
& SLAB_POISON
) {
879 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
880 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
881 POISON_FREE
, s
->object_size
- 1) ||
882 !check_bytes_and_report(s
, page
, p
, "Poison",
883 p
+ s
->object_size
- 1, POISON_END
, 1)))
886 * check_pad_bytes cleans up on its own.
888 check_pad_bytes(s
, page
, p
);
891 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
893 * Object and freepointer overlap. Cannot check
894 * freepointer while object is allocated.
898 /* Check free pointer validity */
899 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
900 object_err(s
, page
, p
, "Freepointer corrupt");
902 * No choice but to zap it and thus lose the remainder
903 * of the free objects in this slab. May cause
904 * another error because the object count is now wrong.
906 set_freepointer(s
, p
, NULL
);
912 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
916 VM_BUG_ON(!irqs_disabled());
918 if (!PageSlab(page
)) {
919 slab_err(s
, page
, "Not a valid slab page");
923 maxobj
= order_objects(compound_order(page
), s
->size
);
924 if (page
->objects
> maxobj
) {
925 slab_err(s
, page
, "objects %u > max %u",
926 page
->objects
, maxobj
);
929 if (page
->inuse
> page
->objects
) {
930 slab_err(s
, page
, "inuse %u > max %u",
931 page
->inuse
, page
->objects
);
934 /* Slab_pad_check fixes things up after itself */
935 slab_pad_check(s
, page
);
940 * Determine if a certain object on a page is on the freelist. Must hold the
941 * slab lock to guarantee that the chains are in a consistent state.
943 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
951 while (fp
&& nr
<= page
->objects
) {
954 if (!check_valid_pointer(s
, page
, fp
)) {
956 object_err(s
, page
, object
,
957 "Freechain corrupt");
958 set_freepointer(s
, object
, NULL
);
960 slab_err(s
, page
, "Freepointer corrupt");
961 page
->freelist
= NULL
;
962 page
->inuse
= page
->objects
;
963 slab_fix(s
, "Freelist cleared");
969 fp
= get_freepointer(s
, object
);
973 max_objects
= order_objects(compound_order(page
), s
->size
);
974 if (max_objects
> MAX_OBJS_PER_PAGE
)
975 max_objects
= MAX_OBJS_PER_PAGE
;
977 if (page
->objects
!= max_objects
) {
978 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
979 page
->objects
, max_objects
);
980 page
->objects
= max_objects
;
981 slab_fix(s
, "Number of objects adjusted.");
983 if (page
->inuse
!= page
->objects
- nr
) {
984 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
985 page
->inuse
, page
->objects
- nr
);
986 page
->inuse
= page
->objects
- nr
;
987 slab_fix(s
, "Object count adjusted.");
989 return search
== NULL
;
992 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
995 if (s
->flags
& SLAB_TRACE
) {
996 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
998 alloc
? "alloc" : "free",
1003 print_section(KERN_INFO
, "Object ", (void *)object
,
1011 * Tracking of fully allocated slabs for debugging purposes.
1013 static void add_full(struct kmem_cache
*s
,
1014 struct kmem_cache_node
*n
, struct page
*page
)
1016 if (!(s
->flags
& SLAB_STORE_USER
))
1019 lockdep_assert_held(&n
->list_lock
);
1020 list_add(&page
->lru
, &n
->full
);
1023 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1025 if (!(s
->flags
& SLAB_STORE_USER
))
1028 lockdep_assert_held(&n
->list_lock
);
1029 list_del(&page
->lru
);
1032 /* Tracking of the number of slabs for debugging purposes */
1033 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1035 struct kmem_cache_node
*n
= get_node(s
, node
);
1037 return atomic_long_read(&n
->nr_slabs
);
1040 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1042 return atomic_long_read(&n
->nr_slabs
);
1045 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1047 struct kmem_cache_node
*n
= get_node(s
, node
);
1050 * May be called early in order to allocate a slab for the
1051 * kmem_cache_node structure. Solve the chicken-egg
1052 * dilemma by deferring the increment of the count during
1053 * bootstrap (see early_kmem_cache_node_alloc).
1056 atomic_long_inc(&n
->nr_slabs
);
1057 atomic_long_add(objects
, &n
->total_objects
);
1060 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1062 struct kmem_cache_node
*n
= get_node(s
, node
);
1064 atomic_long_dec(&n
->nr_slabs
);
1065 atomic_long_sub(objects
, &n
->total_objects
);
1068 /* Object debug checks for alloc/free paths */
1069 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1072 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1075 init_object(s
, object
, SLUB_RED_INACTIVE
);
1076 init_tracking(s
, object
);
1079 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1081 void *object
, unsigned long addr
)
1083 if (!check_slab(s
, page
))
1086 if (!check_valid_pointer(s
, page
, object
)) {
1087 object_err(s
, page
, object
, "Freelist Pointer check fails");
1091 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1097 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1099 void *object
, unsigned long addr
)
1101 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1102 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1106 /* Success perform special debug activities for allocs */
1107 if (s
->flags
& SLAB_STORE_USER
)
1108 set_track(s
, object
, TRACK_ALLOC
, addr
);
1109 trace(s
, page
, object
, 1);
1110 init_object(s
, object
, SLUB_RED_ACTIVE
);
1114 if (PageSlab(page
)) {
1116 * If this is a slab page then lets do the best we can
1117 * to avoid issues in the future. Marking all objects
1118 * as used avoids touching the remaining objects.
1120 slab_fix(s
, "Marking all objects used");
1121 page
->inuse
= page
->objects
;
1122 page
->freelist
= NULL
;
1127 static inline int free_consistency_checks(struct kmem_cache
*s
,
1128 struct page
*page
, void *object
, unsigned long addr
)
1130 if (!check_valid_pointer(s
, page
, object
)) {
1131 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1135 if (on_freelist(s
, page
, object
)) {
1136 object_err(s
, page
, object
, "Object already free");
1140 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1143 if (unlikely(s
!= page
->slab_cache
)) {
1144 if (!PageSlab(page
)) {
1145 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1147 } else if (!page
->slab_cache
) {
1148 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1152 object_err(s
, page
, object
,
1153 "page slab pointer corrupt.");
1159 /* Supports checking bulk free of a constructed freelist */
1160 static noinline
int free_debug_processing(
1161 struct kmem_cache
*s
, struct page
*page
,
1162 void *head
, void *tail
, int bulk_cnt
,
1165 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1166 void *object
= head
;
1168 unsigned long uninitialized_var(flags
);
1171 spin_lock_irqsave(&n
->list_lock
, flags
);
1174 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1175 if (!check_slab(s
, page
))
1182 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1183 if (!free_consistency_checks(s
, page
, object
, addr
))
1187 if (s
->flags
& SLAB_STORE_USER
)
1188 set_track(s
, object
, TRACK_FREE
, addr
);
1189 trace(s
, page
, object
, 0);
1190 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1191 init_object(s
, object
, SLUB_RED_INACTIVE
);
1193 /* Reached end of constructed freelist yet? */
1194 if (object
!= tail
) {
1195 object
= get_freepointer(s
, object
);
1201 if (cnt
!= bulk_cnt
)
1202 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1206 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1208 slab_fix(s
, "Object at 0x%p not freed", object
);
1212 static int __init
setup_slub_debug(char *str
)
1214 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1215 if (*str
++ != '=' || !*str
)
1217 * No options specified. Switch on full debugging.
1223 * No options but restriction on slabs. This means full
1224 * debugging for slabs matching a pattern.
1231 * Switch off all debugging measures.
1236 * Determine which debug features should be switched on
1238 for (; *str
&& *str
!= ','; str
++) {
1239 switch (tolower(*str
)) {
1241 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1244 slub_debug
|= SLAB_RED_ZONE
;
1247 slub_debug
|= SLAB_POISON
;
1250 slub_debug
|= SLAB_STORE_USER
;
1253 slub_debug
|= SLAB_TRACE
;
1256 slub_debug
|= SLAB_FAILSLAB
;
1260 * Avoid enabling debugging on caches if its minimum
1261 * order would increase as a result.
1263 disable_higher_order_debug
= 1;
1266 pr_err("slub_debug option '%c' unknown. skipped\n",
1273 slub_debug_slabs
= str
+ 1;
1278 __setup("slub_debug", setup_slub_debug
);
1280 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1281 slab_flags_t flags
, const char *name
,
1282 void (*ctor
)(void *))
1285 * Enable debugging if selected on the kernel commandline.
1287 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1288 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1289 flags
|= slub_debug
;
1293 #else /* !CONFIG_SLUB_DEBUG */
1294 static inline void setup_object_debug(struct kmem_cache
*s
,
1295 struct page
*page
, void *object
) {}
1297 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1298 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1300 static inline int free_debug_processing(
1301 struct kmem_cache
*s
, struct page
*page
,
1302 void *head
, void *tail
, int bulk_cnt
,
1303 unsigned long addr
) { return 0; }
1305 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1307 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1308 void *object
, u8 val
) { return 1; }
1309 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1310 struct page
*page
) {}
1311 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1312 struct page
*page
) {}
1313 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1314 slab_flags_t flags
, const char *name
,
1315 void (*ctor
)(void *))
1319 #define slub_debug 0
1321 #define disable_higher_order_debug 0
1323 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1325 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1327 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1329 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1332 #endif /* CONFIG_SLUB_DEBUG */
1335 * Hooks for other subsystems that check memory allocations. In a typical
1336 * production configuration these hooks all should produce no code at all.
1338 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1340 kmemleak_alloc(ptr
, size
, 1, flags
);
1341 kasan_kmalloc_large(ptr
, size
, flags
);
1344 static __always_inline
void kfree_hook(void *x
)
1347 kasan_kfree_large(x
, _RET_IP_
);
1350 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1352 kmemleak_free_recursive(x
, s
->flags
);
1355 * Trouble is that we may no longer disable interrupts in the fast path
1356 * So in order to make the debug calls that expect irqs to be
1357 * disabled we need to disable interrupts temporarily.
1359 #ifdef CONFIG_LOCKDEP
1361 unsigned long flags
;
1363 local_irq_save(flags
);
1364 debug_check_no_locks_freed(x
, s
->object_size
);
1365 local_irq_restore(flags
);
1368 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1369 debug_check_no_obj_freed(x
, s
->object_size
);
1371 /* KASAN might put x into memory quarantine, delaying its reuse */
1372 return kasan_slab_free(s
, x
, _RET_IP_
);
1375 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1376 void **head
, void **tail
)
1379 * Compiler cannot detect this function can be removed if slab_free_hook()
1380 * evaluates to nothing. Thus, catch all relevant config debug options here.
1382 #if defined(CONFIG_LOCKDEP) || \
1383 defined(CONFIG_DEBUG_KMEMLEAK) || \
1384 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1385 defined(CONFIG_KASAN)
1389 void *old_tail
= *tail
? *tail
: *head
;
1391 /* Head and tail of the reconstructed freelist */
1397 next
= get_freepointer(s
, object
);
1398 /* If object's reuse doesn't have to be delayed */
1399 if (!slab_free_hook(s
, object
)) {
1400 /* Move object to the new freelist */
1401 set_freepointer(s
, object
, *head
);
1406 } while (object
!= old_tail
);
1411 return *head
!= NULL
;
1417 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1420 setup_object_debug(s
, page
, object
);
1421 kasan_init_slab_obj(s
, object
);
1422 if (unlikely(s
->ctor
)) {
1423 kasan_unpoison_object_data(s
, object
);
1425 kasan_poison_object_data(s
, object
);
1430 * Slab allocation and freeing
1432 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1433 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1436 unsigned int order
= oo_order(oo
);
1438 if (node
== NUMA_NO_NODE
)
1439 page
= alloc_pages(flags
, order
);
1441 page
= __alloc_pages_node(node
, flags
, order
);
1443 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1444 __free_pages(page
, order
);
1451 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1452 /* Pre-initialize the random sequence cache */
1453 static int init_cache_random_seq(struct kmem_cache
*s
)
1455 unsigned int count
= oo_objects(s
->oo
);
1458 /* Bailout if already initialised */
1462 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1464 pr_err("SLUB: Unable to initialize free list for %s\n",
1469 /* Transform to an offset on the set of pages */
1470 if (s
->random_seq
) {
1473 for (i
= 0; i
< count
; i
++)
1474 s
->random_seq
[i
] *= s
->size
;
1479 /* Initialize each random sequence freelist per cache */
1480 static void __init
init_freelist_randomization(void)
1482 struct kmem_cache
*s
;
1484 mutex_lock(&slab_mutex
);
1486 list_for_each_entry(s
, &slab_caches
, list
)
1487 init_cache_random_seq(s
);
1489 mutex_unlock(&slab_mutex
);
1492 /* Get the next entry on the pre-computed freelist randomized */
1493 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1494 unsigned long *pos
, void *start
,
1495 unsigned long page_limit
,
1496 unsigned long freelist_count
)
1501 * If the target page allocation failed, the number of objects on the
1502 * page might be smaller than the usual size defined by the cache.
1505 idx
= s
->random_seq
[*pos
];
1507 if (*pos
>= freelist_count
)
1509 } while (unlikely(idx
>= page_limit
));
1511 return (char *)start
+ idx
;
1514 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1515 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1520 unsigned long idx
, pos
, page_limit
, freelist_count
;
1522 if (page
->objects
< 2 || !s
->random_seq
)
1525 freelist_count
= oo_objects(s
->oo
);
1526 pos
= get_random_int() % freelist_count
;
1528 page_limit
= page
->objects
* s
->size
;
1529 start
= fixup_red_left(s
, page_address(page
));
1531 /* First entry is used as the base of the freelist */
1532 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1534 page
->freelist
= cur
;
1536 for (idx
= 1; idx
< page
->objects
; idx
++) {
1537 setup_object(s
, page
, cur
);
1538 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1540 set_freepointer(s
, cur
, next
);
1543 setup_object(s
, page
, cur
);
1544 set_freepointer(s
, cur
, NULL
);
1549 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1553 static inline void init_freelist_randomization(void) { }
1554 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1558 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1560 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1563 struct kmem_cache_order_objects oo
= s
->oo
;
1569 flags
&= gfp_allowed_mask
;
1571 if (gfpflags_allow_blocking(flags
))
1574 flags
|= s
->allocflags
;
1577 * Let the initial higher-order allocation fail under memory pressure
1578 * so we fall-back to the minimum order allocation.
1580 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1581 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1582 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1584 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1585 if (unlikely(!page
)) {
1589 * Allocation may have failed due to fragmentation.
1590 * Try a lower order alloc if possible
1592 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1593 if (unlikely(!page
))
1595 stat(s
, ORDER_FALLBACK
);
1598 page
->objects
= oo_objects(oo
);
1600 order
= compound_order(page
);
1601 page
->slab_cache
= s
;
1602 __SetPageSlab(page
);
1603 if (page_is_pfmemalloc(page
))
1604 SetPageSlabPfmemalloc(page
);
1606 start
= page_address(page
);
1608 if (unlikely(s
->flags
& SLAB_POISON
))
1609 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1611 kasan_poison_slab(page
);
1613 shuffle
= shuffle_freelist(s
, page
);
1616 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1617 setup_object(s
, page
, p
);
1618 if (likely(idx
< page
->objects
))
1619 set_freepointer(s
, p
, p
+ s
->size
);
1621 set_freepointer(s
, p
, NULL
);
1623 page
->freelist
= fixup_red_left(s
, start
);
1626 page
->inuse
= page
->objects
;
1630 if (gfpflags_allow_blocking(flags
))
1631 local_irq_disable();
1635 mod_lruvec_page_state(page
,
1636 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1637 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1640 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1645 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1647 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1648 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1649 flags
&= ~GFP_SLAB_BUG_MASK
;
1650 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1651 invalid_mask
, &invalid_mask
, flags
, &flags
);
1655 return allocate_slab(s
,
1656 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1659 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1661 int order
= compound_order(page
);
1662 int pages
= 1 << order
;
1664 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1667 slab_pad_check(s
, page
);
1668 for_each_object(p
, s
, page_address(page
),
1670 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1673 mod_lruvec_page_state(page
,
1674 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1675 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1678 __ClearPageSlabPfmemalloc(page
);
1679 __ClearPageSlab(page
);
1681 page
->mapping
= NULL
;
1682 if (current
->reclaim_state
)
1683 current
->reclaim_state
->reclaimed_slab
+= pages
;
1684 memcg_uncharge_slab(page
, order
, s
);
1685 __free_pages(page
, order
);
1688 static void rcu_free_slab(struct rcu_head
*h
)
1690 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1692 __free_slab(page
->slab_cache
, page
);
1695 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1697 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1698 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1700 __free_slab(s
, page
);
1703 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1705 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1710 * Management of partially allocated slabs.
1713 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1716 if (tail
== DEACTIVATE_TO_TAIL
)
1717 list_add_tail(&page
->lru
, &n
->partial
);
1719 list_add(&page
->lru
, &n
->partial
);
1722 static inline void add_partial(struct kmem_cache_node
*n
,
1723 struct page
*page
, int tail
)
1725 lockdep_assert_held(&n
->list_lock
);
1726 __add_partial(n
, page
, tail
);
1729 static inline void remove_partial(struct kmem_cache_node
*n
,
1732 lockdep_assert_held(&n
->list_lock
);
1733 list_del(&page
->lru
);
1738 * Remove slab from the partial list, freeze it and
1739 * return the pointer to the freelist.
1741 * Returns a list of objects or NULL if it fails.
1743 static inline void *acquire_slab(struct kmem_cache
*s
,
1744 struct kmem_cache_node
*n
, struct page
*page
,
1745 int mode
, int *objects
)
1748 unsigned long counters
;
1751 lockdep_assert_held(&n
->list_lock
);
1754 * Zap the freelist and set the frozen bit.
1755 * The old freelist is the list of objects for the
1756 * per cpu allocation list.
1758 freelist
= page
->freelist
;
1759 counters
= page
->counters
;
1760 new.counters
= counters
;
1761 *objects
= new.objects
- new.inuse
;
1763 new.inuse
= page
->objects
;
1764 new.freelist
= NULL
;
1766 new.freelist
= freelist
;
1769 VM_BUG_ON(new.frozen
);
1772 if (!__cmpxchg_double_slab(s
, page
,
1774 new.freelist
, new.counters
,
1778 remove_partial(n
, page
);
1783 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1784 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1787 * Try to allocate a partial slab from a specific node.
1789 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1790 struct kmem_cache_cpu
*c
, gfp_t flags
)
1792 struct page
*page
, *page2
;
1793 void *object
= NULL
;
1794 unsigned int available
= 0;
1798 * Racy check. If we mistakenly see no partial slabs then we
1799 * just allocate an empty slab. If we mistakenly try to get a
1800 * partial slab and there is none available then get_partials()
1803 if (!n
|| !n
->nr_partial
)
1806 spin_lock(&n
->list_lock
);
1807 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1810 if (!pfmemalloc_match(page
, flags
))
1813 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1817 available
+= objects
;
1820 stat(s
, ALLOC_FROM_PARTIAL
);
1823 put_cpu_partial(s
, page
, 0);
1824 stat(s
, CPU_PARTIAL_NODE
);
1826 if (!kmem_cache_has_cpu_partial(s
)
1827 || available
> slub_cpu_partial(s
) / 2)
1831 spin_unlock(&n
->list_lock
);
1836 * Get a page from somewhere. Search in increasing NUMA distances.
1838 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1839 struct kmem_cache_cpu
*c
)
1842 struct zonelist
*zonelist
;
1845 enum zone_type high_zoneidx
= gfp_zone(flags
);
1847 unsigned int cpuset_mems_cookie
;
1850 * The defrag ratio allows a configuration of the tradeoffs between
1851 * inter node defragmentation and node local allocations. A lower
1852 * defrag_ratio increases the tendency to do local allocations
1853 * instead of attempting to obtain partial slabs from other nodes.
1855 * If the defrag_ratio is set to 0 then kmalloc() always
1856 * returns node local objects. If the ratio is higher then kmalloc()
1857 * may return off node objects because partial slabs are obtained
1858 * from other nodes and filled up.
1860 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1861 * (which makes defrag_ratio = 1000) then every (well almost)
1862 * allocation will first attempt to defrag slab caches on other nodes.
1863 * This means scanning over all nodes to look for partial slabs which
1864 * may be expensive if we do it every time we are trying to find a slab
1865 * with available objects.
1867 if (!s
->remote_node_defrag_ratio
||
1868 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1872 cpuset_mems_cookie
= read_mems_allowed_begin();
1873 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1874 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1875 struct kmem_cache_node
*n
;
1877 n
= get_node(s
, zone_to_nid(zone
));
1879 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1880 n
->nr_partial
> s
->min_partial
) {
1881 object
= get_partial_node(s
, n
, c
, flags
);
1884 * Don't check read_mems_allowed_retry()
1885 * here - if mems_allowed was updated in
1886 * parallel, that was a harmless race
1887 * between allocation and the cpuset
1894 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1900 * Get a partial page, lock it and return it.
1902 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1903 struct kmem_cache_cpu
*c
)
1906 int searchnode
= node
;
1908 if (node
== NUMA_NO_NODE
)
1909 searchnode
= numa_mem_id();
1910 else if (!node_present_pages(node
))
1911 searchnode
= node_to_mem_node(node
);
1913 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1914 if (object
|| node
!= NUMA_NO_NODE
)
1917 return get_any_partial(s
, flags
, c
);
1920 #ifdef CONFIG_PREEMPT
1922 * Calculate the next globally unique transaction for disambiguiation
1923 * during cmpxchg. The transactions start with the cpu number and are then
1924 * incremented by CONFIG_NR_CPUS.
1926 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1929 * No preemption supported therefore also no need to check for
1935 static inline unsigned long next_tid(unsigned long tid
)
1937 return tid
+ TID_STEP
;
1940 static inline unsigned int tid_to_cpu(unsigned long tid
)
1942 return tid
% TID_STEP
;
1945 static inline unsigned long tid_to_event(unsigned long tid
)
1947 return tid
/ TID_STEP
;
1950 static inline unsigned int init_tid(int cpu
)
1955 static inline void note_cmpxchg_failure(const char *n
,
1956 const struct kmem_cache
*s
, unsigned long tid
)
1958 #ifdef SLUB_DEBUG_CMPXCHG
1959 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
1961 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
1963 #ifdef CONFIG_PREEMPT
1964 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
1965 pr_warn("due to cpu change %d -> %d\n",
1966 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
1969 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
1970 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1971 tid_to_event(tid
), tid_to_event(actual_tid
));
1973 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1974 actual_tid
, tid
, next_tid(tid
));
1976 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
1979 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
1983 for_each_possible_cpu(cpu
)
1984 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
1988 * Remove the cpu slab
1990 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
1991 void *freelist
, struct kmem_cache_cpu
*c
)
1993 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
1994 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1996 enum slab_modes l
= M_NONE
, m
= M_NONE
;
1998 int tail
= DEACTIVATE_TO_HEAD
;
2002 if (page
->freelist
) {
2003 stat(s
, DEACTIVATE_REMOTE_FREES
);
2004 tail
= DEACTIVATE_TO_TAIL
;
2008 * Stage one: Free all available per cpu objects back
2009 * to the page freelist while it is still frozen. Leave the
2012 * There is no need to take the list->lock because the page
2015 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2017 unsigned long counters
;
2020 prior
= page
->freelist
;
2021 counters
= page
->counters
;
2022 set_freepointer(s
, freelist
, prior
);
2023 new.counters
= counters
;
2025 VM_BUG_ON(!new.frozen
);
2027 } while (!__cmpxchg_double_slab(s
, page
,
2029 freelist
, new.counters
,
2030 "drain percpu freelist"));
2032 freelist
= nextfree
;
2036 * Stage two: Ensure that the page is unfrozen while the
2037 * list presence reflects the actual number of objects
2040 * We setup the list membership and then perform a cmpxchg
2041 * with the count. If there is a mismatch then the page
2042 * is not unfrozen but the page is on the wrong list.
2044 * Then we restart the process which may have to remove
2045 * the page from the list that we just put it on again
2046 * because the number of objects in the slab may have
2051 old
.freelist
= page
->freelist
;
2052 old
.counters
= page
->counters
;
2053 VM_BUG_ON(!old
.frozen
);
2055 /* Determine target state of the slab */
2056 new.counters
= old
.counters
;
2059 set_freepointer(s
, freelist
, old
.freelist
);
2060 new.freelist
= freelist
;
2062 new.freelist
= old
.freelist
;
2066 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2068 else if (new.freelist
) {
2073 * Taking the spinlock removes the possiblity
2074 * that acquire_slab() will see a slab page that
2077 spin_lock(&n
->list_lock
);
2081 if (kmem_cache_debug(s
) && !lock
) {
2084 * This also ensures that the scanning of full
2085 * slabs from diagnostic functions will not see
2088 spin_lock(&n
->list_lock
);
2096 remove_partial(n
, page
);
2098 else if (l
== M_FULL
)
2100 remove_full(s
, n
, page
);
2102 if (m
== M_PARTIAL
) {
2104 add_partial(n
, page
, tail
);
2107 } else if (m
== M_FULL
) {
2109 stat(s
, DEACTIVATE_FULL
);
2110 add_full(s
, n
, page
);
2116 if (!__cmpxchg_double_slab(s
, page
,
2117 old
.freelist
, old
.counters
,
2118 new.freelist
, new.counters
,
2123 spin_unlock(&n
->list_lock
);
2126 stat(s
, DEACTIVATE_EMPTY
);
2127 discard_slab(s
, page
);
2136 * Unfreeze all the cpu partial slabs.
2138 * This function must be called with interrupts disabled
2139 * for the cpu using c (or some other guarantee must be there
2140 * to guarantee no concurrent accesses).
2142 static void unfreeze_partials(struct kmem_cache
*s
,
2143 struct kmem_cache_cpu
*c
)
2145 #ifdef CONFIG_SLUB_CPU_PARTIAL
2146 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2147 struct page
*page
, *discard_page
= NULL
;
2149 while ((page
= c
->partial
)) {
2153 c
->partial
= page
->next
;
2155 n2
= get_node(s
, page_to_nid(page
));
2158 spin_unlock(&n
->list_lock
);
2161 spin_lock(&n
->list_lock
);
2166 old
.freelist
= page
->freelist
;
2167 old
.counters
= page
->counters
;
2168 VM_BUG_ON(!old
.frozen
);
2170 new.counters
= old
.counters
;
2171 new.freelist
= old
.freelist
;
2175 } while (!__cmpxchg_double_slab(s
, page
,
2176 old
.freelist
, old
.counters
,
2177 new.freelist
, new.counters
,
2178 "unfreezing slab"));
2180 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2181 page
->next
= discard_page
;
2182 discard_page
= page
;
2184 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2185 stat(s
, FREE_ADD_PARTIAL
);
2190 spin_unlock(&n
->list_lock
);
2192 while (discard_page
) {
2193 page
= discard_page
;
2194 discard_page
= discard_page
->next
;
2196 stat(s
, DEACTIVATE_EMPTY
);
2197 discard_slab(s
, page
);
2204 * Put a page that was just frozen (in __slab_free) into a partial page
2205 * slot if available.
2207 * If we did not find a slot then simply move all the partials to the
2208 * per node partial list.
2210 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2212 #ifdef CONFIG_SLUB_CPU_PARTIAL
2213 struct page
*oldpage
;
2221 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2224 pobjects
= oldpage
->pobjects
;
2225 pages
= oldpage
->pages
;
2226 if (drain
&& pobjects
> s
->cpu_partial
) {
2227 unsigned long flags
;
2229 * partial array is full. Move the existing
2230 * set to the per node partial list.
2232 local_irq_save(flags
);
2233 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2234 local_irq_restore(flags
);
2238 stat(s
, CPU_PARTIAL_DRAIN
);
2243 pobjects
+= page
->objects
- page
->inuse
;
2245 page
->pages
= pages
;
2246 page
->pobjects
= pobjects
;
2247 page
->next
= oldpage
;
2249 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2251 if (unlikely(!s
->cpu_partial
)) {
2252 unsigned long flags
;
2254 local_irq_save(flags
);
2255 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2256 local_irq_restore(flags
);
2262 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2264 stat(s
, CPUSLAB_FLUSH
);
2265 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2267 c
->tid
= next_tid(c
->tid
);
2273 * Called from IPI handler with interrupts disabled.
2275 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2277 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2283 unfreeze_partials(s
, c
);
2287 static void flush_cpu_slab(void *d
)
2289 struct kmem_cache
*s
= d
;
2291 __flush_cpu_slab(s
, smp_processor_id());
2294 static bool has_cpu_slab(int cpu
, void *info
)
2296 struct kmem_cache
*s
= info
;
2297 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2299 return c
->page
|| slub_percpu_partial(c
);
2302 static void flush_all(struct kmem_cache
*s
)
2304 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2308 * Use the cpu notifier to insure that the cpu slabs are flushed when
2311 static int slub_cpu_dead(unsigned int cpu
)
2313 struct kmem_cache
*s
;
2314 unsigned long flags
;
2316 mutex_lock(&slab_mutex
);
2317 list_for_each_entry(s
, &slab_caches
, list
) {
2318 local_irq_save(flags
);
2319 __flush_cpu_slab(s
, cpu
);
2320 local_irq_restore(flags
);
2322 mutex_unlock(&slab_mutex
);
2327 * Check if the objects in a per cpu structure fit numa
2328 * locality expectations.
2330 static inline int node_match(struct page
*page
, int node
)
2333 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2339 #ifdef CONFIG_SLUB_DEBUG
2340 static int count_free(struct page
*page
)
2342 return page
->objects
- page
->inuse
;
2345 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2347 return atomic_long_read(&n
->total_objects
);
2349 #endif /* CONFIG_SLUB_DEBUG */
2351 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2352 static unsigned long count_partial(struct kmem_cache_node
*n
,
2353 int (*get_count
)(struct page
*))
2355 unsigned long flags
;
2356 unsigned long x
= 0;
2359 spin_lock_irqsave(&n
->list_lock
, flags
);
2360 list_for_each_entry(page
, &n
->partial
, lru
)
2361 x
+= get_count(page
);
2362 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2365 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2367 static noinline
void
2368 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2370 #ifdef CONFIG_SLUB_DEBUG
2371 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2372 DEFAULT_RATELIMIT_BURST
);
2374 struct kmem_cache_node
*n
;
2376 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2379 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2380 nid
, gfpflags
, &gfpflags
);
2381 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2382 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2385 if (oo_order(s
->min
) > get_order(s
->object_size
))
2386 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2389 for_each_kmem_cache_node(s
, node
, n
) {
2390 unsigned long nr_slabs
;
2391 unsigned long nr_objs
;
2392 unsigned long nr_free
;
2394 nr_free
= count_partial(n
, count_free
);
2395 nr_slabs
= node_nr_slabs(n
);
2396 nr_objs
= node_nr_objs(n
);
2398 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2399 node
, nr_slabs
, nr_objs
, nr_free
);
2404 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2405 int node
, struct kmem_cache_cpu
**pc
)
2408 struct kmem_cache_cpu
*c
= *pc
;
2411 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2413 freelist
= get_partial(s
, flags
, node
, c
);
2418 page
= new_slab(s
, flags
, node
);
2420 c
= raw_cpu_ptr(s
->cpu_slab
);
2425 * No other reference to the page yet so we can
2426 * muck around with it freely without cmpxchg
2428 freelist
= page
->freelist
;
2429 page
->freelist
= NULL
;
2431 stat(s
, ALLOC_SLAB
);
2440 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2442 if (unlikely(PageSlabPfmemalloc(page
)))
2443 return gfp_pfmemalloc_allowed(gfpflags
);
2449 * Check the page->freelist of a page and either transfer the freelist to the
2450 * per cpu freelist or deactivate the page.
2452 * The page is still frozen if the return value is not NULL.
2454 * If this function returns NULL then the page has been unfrozen.
2456 * This function must be called with interrupt disabled.
2458 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2461 unsigned long counters
;
2465 freelist
= page
->freelist
;
2466 counters
= page
->counters
;
2468 new.counters
= counters
;
2469 VM_BUG_ON(!new.frozen
);
2471 new.inuse
= page
->objects
;
2472 new.frozen
= freelist
!= NULL
;
2474 } while (!__cmpxchg_double_slab(s
, page
,
2483 * Slow path. The lockless freelist is empty or we need to perform
2486 * Processing is still very fast if new objects have been freed to the
2487 * regular freelist. In that case we simply take over the regular freelist
2488 * as the lockless freelist and zap the regular freelist.
2490 * If that is not working then we fall back to the partial lists. We take the
2491 * first element of the freelist as the object to allocate now and move the
2492 * rest of the freelist to the lockless freelist.
2494 * And if we were unable to get a new slab from the partial slab lists then
2495 * we need to allocate a new slab. This is the slowest path since it involves
2496 * a call to the page allocator and the setup of a new slab.
2498 * Version of __slab_alloc to use when we know that interrupts are
2499 * already disabled (which is the case for bulk allocation).
2501 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2502 unsigned long addr
, struct kmem_cache_cpu
*c
)
2512 if (unlikely(!node_match(page
, node
))) {
2513 int searchnode
= node
;
2515 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2516 searchnode
= node_to_mem_node(node
);
2518 if (unlikely(!node_match(page
, searchnode
))) {
2519 stat(s
, ALLOC_NODE_MISMATCH
);
2520 deactivate_slab(s
, page
, c
->freelist
, c
);
2526 * By rights, we should be searching for a slab page that was
2527 * PFMEMALLOC but right now, we are losing the pfmemalloc
2528 * information when the page leaves the per-cpu allocator
2530 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2531 deactivate_slab(s
, page
, c
->freelist
, c
);
2535 /* must check again c->freelist in case of cpu migration or IRQ */
2536 freelist
= c
->freelist
;
2540 freelist
= get_freelist(s
, page
);
2544 stat(s
, DEACTIVATE_BYPASS
);
2548 stat(s
, ALLOC_REFILL
);
2552 * freelist is pointing to the list of objects to be used.
2553 * page is pointing to the page from which the objects are obtained.
2554 * That page must be frozen for per cpu allocations to work.
2556 VM_BUG_ON(!c
->page
->frozen
);
2557 c
->freelist
= get_freepointer(s
, freelist
);
2558 c
->tid
= next_tid(c
->tid
);
2563 if (slub_percpu_partial(c
)) {
2564 page
= c
->page
= slub_percpu_partial(c
);
2565 slub_set_percpu_partial(c
, page
);
2566 stat(s
, CPU_PARTIAL_ALLOC
);
2570 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2572 if (unlikely(!freelist
)) {
2573 slab_out_of_memory(s
, gfpflags
, node
);
2578 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2581 /* Only entered in the debug case */
2582 if (kmem_cache_debug(s
) &&
2583 !alloc_debug_processing(s
, page
, freelist
, addr
))
2584 goto new_slab
; /* Slab failed checks. Next slab needed */
2586 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2591 * Another one that disabled interrupt and compensates for possible
2592 * cpu changes by refetching the per cpu area pointer.
2594 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2595 unsigned long addr
, struct kmem_cache_cpu
*c
)
2598 unsigned long flags
;
2600 local_irq_save(flags
);
2601 #ifdef CONFIG_PREEMPT
2603 * We may have been preempted and rescheduled on a different
2604 * cpu before disabling interrupts. Need to reload cpu area
2607 c
= this_cpu_ptr(s
->cpu_slab
);
2610 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2611 local_irq_restore(flags
);
2616 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2617 * have the fastpath folded into their functions. So no function call
2618 * overhead for requests that can be satisfied on the fastpath.
2620 * The fastpath works by first checking if the lockless freelist can be used.
2621 * If not then __slab_alloc is called for slow processing.
2623 * Otherwise we can simply pick the next object from the lockless free list.
2625 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2626 gfp_t gfpflags
, int node
, unsigned long addr
)
2629 struct kmem_cache_cpu
*c
;
2633 s
= slab_pre_alloc_hook(s
, gfpflags
);
2638 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2639 * enabled. We may switch back and forth between cpus while
2640 * reading from one cpu area. That does not matter as long
2641 * as we end up on the original cpu again when doing the cmpxchg.
2643 * We should guarantee that tid and kmem_cache are retrieved on
2644 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2645 * to check if it is matched or not.
2648 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2649 c
= raw_cpu_ptr(s
->cpu_slab
);
2650 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2651 unlikely(tid
!= READ_ONCE(c
->tid
)));
2654 * Irqless object alloc/free algorithm used here depends on sequence
2655 * of fetching cpu_slab's data. tid should be fetched before anything
2656 * on c to guarantee that object and page associated with previous tid
2657 * won't be used with current tid. If we fetch tid first, object and
2658 * page could be one associated with next tid and our alloc/free
2659 * request will be failed. In this case, we will retry. So, no problem.
2664 * The transaction ids are globally unique per cpu and per operation on
2665 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2666 * occurs on the right processor and that there was no operation on the
2667 * linked list in between.
2670 object
= c
->freelist
;
2672 if (unlikely(!object
|| !node_match(page
, node
))) {
2673 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2674 stat(s
, ALLOC_SLOWPATH
);
2676 void *next_object
= get_freepointer_safe(s
, object
);
2679 * The cmpxchg will only match if there was no additional
2680 * operation and if we are on the right processor.
2682 * The cmpxchg does the following atomically (without lock
2684 * 1. Relocate first pointer to the current per cpu area.
2685 * 2. Verify that tid and freelist have not been changed
2686 * 3. If they were not changed replace tid and freelist
2688 * Since this is without lock semantics the protection is only
2689 * against code executing on this cpu *not* from access by
2692 if (unlikely(!this_cpu_cmpxchg_double(
2693 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2695 next_object
, next_tid(tid
)))) {
2697 note_cmpxchg_failure("slab_alloc", s
, tid
);
2700 prefetch_freepointer(s
, next_object
);
2701 stat(s
, ALLOC_FASTPATH
);
2704 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2705 memset(object
, 0, s
->object_size
);
2707 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2712 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2713 gfp_t gfpflags
, unsigned long addr
)
2715 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2718 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2720 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2722 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2727 EXPORT_SYMBOL(kmem_cache_alloc
);
2729 #ifdef CONFIG_TRACING
2730 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2732 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2733 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2734 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2737 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2741 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2743 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2745 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2746 s
->object_size
, s
->size
, gfpflags
, node
);
2750 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2752 #ifdef CONFIG_TRACING
2753 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2755 int node
, size_t size
)
2757 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2759 trace_kmalloc_node(_RET_IP_
, ret
,
2760 size
, s
->size
, gfpflags
, node
);
2762 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2765 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2770 * Slow path handling. This may still be called frequently since objects
2771 * have a longer lifetime than the cpu slabs in most processing loads.
2773 * So we still attempt to reduce cache line usage. Just take the slab
2774 * lock and free the item. If there is no additional partial page
2775 * handling required then we can return immediately.
2777 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2778 void *head
, void *tail
, int cnt
,
2785 unsigned long counters
;
2786 struct kmem_cache_node
*n
= NULL
;
2787 unsigned long uninitialized_var(flags
);
2789 stat(s
, FREE_SLOWPATH
);
2791 if (kmem_cache_debug(s
) &&
2792 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2797 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2800 prior
= page
->freelist
;
2801 counters
= page
->counters
;
2802 set_freepointer(s
, tail
, prior
);
2803 new.counters
= counters
;
2804 was_frozen
= new.frozen
;
2806 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2808 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2811 * Slab was on no list before and will be
2813 * We can defer the list move and instead
2818 } else { /* Needs to be taken off a list */
2820 n
= get_node(s
, page_to_nid(page
));
2822 * Speculatively acquire the list_lock.
2823 * If the cmpxchg does not succeed then we may
2824 * drop the list_lock without any processing.
2826 * Otherwise the list_lock will synchronize with
2827 * other processors updating the list of slabs.
2829 spin_lock_irqsave(&n
->list_lock
, flags
);
2834 } while (!cmpxchg_double_slab(s
, page
,
2842 * If we just froze the page then put it onto the
2843 * per cpu partial list.
2845 if (new.frozen
&& !was_frozen
) {
2846 put_cpu_partial(s
, page
, 1);
2847 stat(s
, CPU_PARTIAL_FREE
);
2850 * The list lock was not taken therefore no list
2851 * activity can be necessary.
2854 stat(s
, FREE_FROZEN
);
2858 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2862 * Objects left in the slab. If it was not on the partial list before
2865 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2866 if (kmem_cache_debug(s
))
2867 remove_full(s
, n
, page
);
2868 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2869 stat(s
, FREE_ADD_PARTIAL
);
2871 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2877 * Slab on the partial list.
2879 remove_partial(n
, page
);
2880 stat(s
, FREE_REMOVE_PARTIAL
);
2882 /* Slab must be on the full list */
2883 remove_full(s
, n
, page
);
2886 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2888 discard_slab(s
, page
);
2892 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2893 * can perform fastpath freeing without additional function calls.
2895 * The fastpath is only possible if we are freeing to the current cpu slab
2896 * of this processor. This typically the case if we have just allocated
2899 * If fastpath is not possible then fall back to __slab_free where we deal
2900 * with all sorts of special processing.
2902 * Bulk free of a freelist with several objects (all pointing to the
2903 * same page) possible by specifying head and tail ptr, plus objects
2904 * count (cnt). Bulk free indicated by tail pointer being set.
2906 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2907 struct page
*page
, void *head
, void *tail
,
2908 int cnt
, unsigned long addr
)
2910 void *tail_obj
= tail
? : head
;
2911 struct kmem_cache_cpu
*c
;
2915 * Determine the currently cpus per cpu slab.
2916 * The cpu may change afterward. However that does not matter since
2917 * data is retrieved via this pointer. If we are on the same cpu
2918 * during the cmpxchg then the free will succeed.
2921 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2922 c
= raw_cpu_ptr(s
->cpu_slab
);
2923 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2924 unlikely(tid
!= READ_ONCE(c
->tid
)));
2926 /* Same with comment on barrier() in slab_alloc_node() */
2929 if (likely(page
== c
->page
)) {
2930 set_freepointer(s
, tail_obj
, c
->freelist
);
2932 if (unlikely(!this_cpu_cmpxchg_double(
2933 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2935 head
, next_tid(tid
)))) {
2937 note_cmpxchg_failure("slab_free", s
, tid
);
2940 stat(s
, FREE_FASTPATH
);
2942 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2946 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2947 void *head
, void *tail
, int cnt
,
2951 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2952 * to remove objects, whose reuse must be delayed.
2954 if (slab_free_freelist_hook(s
, &head
, &tail
))
2955 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
2959 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
2961 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
2965 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
2967 s
= cache_from_obj(s
, x
);
2970 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
2971 trace_kmem_cache_free(_RET_IP_
, x
);
2973 EXPORT_SYMBOL(kmem_cache_free
);
2975 struct detached_freelist
{
2980 struct kmem_cache
*s
;
2984 * This function progressively scans the array with free objects (with
2985 * a limited look ahead) and extract objects belonging to the same
2986 * page. It builds a detached freelist directly within the given
2987 * page/objects. This can happen without any need for
2988 * synchronization, because the objects are owned by running process.
2989 * The freelist is build up as a single linked list in the objects.
2990 * The idea is, that this detached freelist can then be bulk
2991 * transferred to the real freelist(s), but only requiring a single
2992 * synchronization primitive. Look ahead in the array is limited due
2993 * to performance reasons.
2996 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
2997 void **p
, struct detached_freelist
*df
)
2999 size_t first_skipped_index
= 0;
3004 /* Always re-init detached_freelist */
3009 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3010 } while (!object
&& size
);
3015 page
= virt_to_head_page(object
);
3017 /* Handle kalloc'ed objects */
3018 if (unlikely(!PageSlab(page
))) {
3019 BUG_ON(!PageCompound(page
));
3021 __free_pages(page
, compound_order(page
));
3022 p
[size
] = NULL
; /* mark object processed */
3025 /* Derive kmem_cache from object */
3026 df
->s
= page
->slab_cache
;
3028 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3031 /* Start new detached freelist */
3033 set_freepointer(df
->s
, object
, NULL
);
3035 df
->freelist
= object
;
3036 p
[size
] = NULL
; /* mark object processed */
3042 continue; /* Skip processed objects */
3044 /* df->page is always set at this point */
3045 if (df
->page
== virt_to_head_page(object
)) {
3046 /* Opportunity build freelist */
3047 set_freepointer(df
->s
, object
, df
->freelist
);
3048 df
->freelist
= object
;
3050 p
[size
] = NULL
; /* mark object processed */
3055 /* Limit look ahead search */
3059 if (!first_skipped_index
)
3060 first_skipped_index
= size
+ 1;
3063 return first_skipped_index
;
3066 /* Note that interrupts must be enabled when calling this function. */
3067 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3073 struct detached_freelist df
;
3075 size
= build_detached_freelist(s
, size
, p
, &df
);
3079 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3080 } while (likely(size
));
3082 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3084 /* Note that interrupts must be enabled when calling this function. */
3085 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3088 struct kmem_cache_cpu
*c
;
3091 /* memcg and kmem_cache debug support */
3092 s
= slab_pre_alloc_hook(s
, flags
);
3096 * Drain objects in the per cpu slab, while disabling local
3097 * IRQs, which protects against PREEMPT and interrupts
3098 * handlers invoking normal fastpath.
3100 local_irq_disable();
3101 c
= this_cpu_ptr(s
->cpu_slab
);
3103 for (i
= 0; i
< size
; i
++) {
3104 void *object
= c
->freelist
;
3106 if (unlikely(!object
)) {
3108 * Invoking slow path likely have side-effect
3109 * of re-populating per CPU c->freelist
3111 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3113 if (unlikely(!p
[i
]))
3116 c
= this_cpu_ptr(s
->cpu_slab
);
3117 continue; /* goto for-loop */
3119 c
->freelist
= get_freepointer(s
, object
);
3122 c
->tid
= next_tid(c
->tid
);
3125 /* Clear memory outside IRQ disabled fastpath loop */
3126 if (unlikely(flags
& __GFP_ZERO
)) {
3129 for (j
= 0; j
< i
; j
++)
3130 memset(p
[j
], 0, s
->object_size
);
3133 /* memcg and kmem_cache debug support */
3134 slab_post_alloc_hook(s
, flags
, size
, p
);
3138 slab_post_alloc_hook(s
, flags
, i
, p
);
3139 __kmem_cache_free_bulk(s
, i
, p
);
3142 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3146 * Object placement in a slab is made very easy because we always start at
3147 * offset 0. If we tune the size of the object to the alignment then we can
3148 * get the required alignment by putting one properly sized object after
3151 * Notice that the allocation order determines the sizes of the per cpu
3152 * caches. Each processor has always one slab available for allocations.
3153 * Increasing the allocation order reduces the number of times that slabs
3154 * must be moved on and off the partial lists and is therefore a factor in
3159 * Mininum / Maximum order of slab pages. This influences locking overhead
3160 * and slab fragmentation. A higher order reduces the number of partial slabs
3161 * and increases the number of allocations possible without having to
3162 * take the list_lock.
3164 static unsigned int slub_min_order
;
3165 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3166 static unsigned int slub_min_objects
;
3169 * Calculate the order of allocation given an slab object size.
3171 * The order of allocation has significant impact on performance and other
3172 * system components. Generally order 0 allocations should be preferred since
3173 * order 0 does not cause fragmentation in the page allocator. Larger objects
3174 * be problematic to put into order 0 slabs because there may be too much
3175 * unused space left. We go to a higher order if more than 1/16th of the slab
3178 * In order to reach satisfactory performance we must ensure that a minimum
3179 * number of objects is in one slab. Otherwise we may generate too much
3180 * activity on the partial lists which requires taking the list_lock. This is
3181 * less a concern for large slabs though which are rarely used.
3183 * slub_max_order specifies the order where we begin to stop considering the
3184 * number of objects in a slab as critical. If we reach slub_max_order then
3185 * we try to keep the page order as low as possible. So we accept more waste
3186 * of space in favor of a small page order.
3188 * Higher order allocations also allow the placement of more objects in a
3189 * slab and thereby reduce object handling overhead. If the user has
3190 * requested a higher mininum order then we start with that one instead of
3191 * the smallest order which will fit the object.
3193 static inline unsigned int slab_order(unsigned int size
,
3194 unsigned int min_objects
, unsigned int max_order
,
3195 unsigned int fract_leftover
)
3197 unsigned int min_order
= slub_min_order
;
3200 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3201 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3203 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3204 order
<= max_order
; order
++) {
3206 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3209 rem
= slab_size
% size
;
3211 if (rem
<= slab_size
/ fract_leftover
)
3218 static inline int calculate_order(unsigned int size
)
3221 unsigned int min_objects
;
3222 unsigned int max_objects
;
3225 * Attempt to find best configuration for a slab. This
3226 * works by first attempting to generate a layout with
3227 * the best configuration and backing off gradually.
3229 * First we increase the acceptable waste in a slab. Then
3230 * we reduce the minimum objects required in a slab.
3232 min_objects
= slub_min_objects
;
3234 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3235 max_objects
= order_objects(slub_max_order
, size
);
3236 min_objects
= min(min_objects
, max_objects
);
3238 while (min_objects
> 1) {
3239 unsigned int fraction
;
3242 while (fraction
>= 4) {
3243 order
= slab_order(size
, min_objects
,
3244 slub_max_order
, fraction
);
3245 if (order
<= slub_max_order
)
3253 * We were unable to place multiple objects in a slab. Now
3254 * lets see if we can place a single object there.
3256 order
= slab_order(size
, 1, slub_max_order
, 1);
3257 if (order
<= slub_max_order
)
3261 * Doh this slab cannot be placed using slub_max_order.
3263 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3264 if (order
< MAX_ORDER
)
3270 init_kmem_cache_node(struct kmem_cache_node
*n
)
3273 spin_lock_init(&n
->list_lock
);
3274 INIT_LIST_HEAD(&n
->partial
);
3275 #ifdef CONFIG_SLUB_DEBUG
3276 atomic_long_set(&n
->nr_slabs
, 0);
3277 atomic_long_set(&n
->total_objects
, 0);
3278 INIT_LIST_HEAD(&n
->full
);
3282 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3284 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3285 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3288 * Must align to double word boundary for the double cmpxchg
3289 * instructions to work; see __pcpu_double_call_return_bool().
3291 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3292 2 * sizeof(void *));
3297 init_kmem_cache_cpus(s
);
3302 static struct kmem_cache
*kmem_cache_node
;
3305 * No kmalloc_node yet so do it by hand. We know that this is the first
3306 * slab on the node for this slabcache. There are no concurrent accesses
3309 * Note that this function only works on the kmem_cache_node
3310 * when allocating for the kmem_cache_node. This is used for bootstrapping
3311 * memory on a fresh node that has no slab structures yet.
3313 static void early_kmem_cache_node_alloc(int node
)
3316 struct kmem_cache_node
*n
;
3318 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3320 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3323 if (page_to_nid(page
) != node
) {
3324 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3325 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3330 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3333 kmem_cache_node
->node
[node
] = n
;
3334 #ifdef CONFIG_SLUB_DEBUG
3335 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3336 init_tracking(kmem_cache_node
, n
);
3338 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3340 init_kmem_cache_node(n
);
3341 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3344 * No locks need to be taken here as it has just been
3345 * initialized and there is no concurrent access.
3347 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3350 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3353 struct kmem_cache_node
*n
;
3355 for_each_kmem_cache_node(s
, node
, n
) {
3356 s
->node
[node
] = NULL
;
3357 kmem_cache_free(kmem_cache_node
, n
);
3361 void __kmem_cache_release(struct kmem_cache
*s
)
3363 cache_random_seq_destroy(s
);
3364 free_percpu(s
->cpu_slab
);
3365 free_kmem_cache_nodes(s
);
3368 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3372 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3373 struct kmem_cache_node
*n
;
3375 if (slab_state
== DOWN
) {
3376 early_kmem_cache_node_alloc(node
);
3379 n
= kmem_cache_alloc_node(kmem_cache_node
,
3383 free_kmem_cache_nodes(s
);
3387 init_kmem_cache_node(n
);
3393 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3395 if (min
< MIN_PARTIAL
)
3397 else if (min
> MAX_PARTIAL
)
3399 s
->min_partial
= min
;
3402 static void set_cpu_partial(struct kmem_cache
*s
)
3404 #ifdef CONFIG_SLUB_CPU_PARTIAL
3406 * cpu_partial determined the maximum number of objects kept in the
3407 * per cpu partial lists of a processor.
3409 * Per cpu partial lists mainly contain slabs that just have one
3410 * object freed. If they are used for allocation then they can be
3411 * filled up again with minimal effort. The slab will never hit the
3412 * per node partial lists and therefore no locking will be required.
3414 * This setting also determines
3416 * A) The number of objects from per cpu partial slabs dumped to the
3417 * per node list when we reach the limit.
3418 * B) The number of objects in cpu partial slabs to extract from the
3419 * per node list when we run out of per cpu objects. We only fetch
3420 * 50% to keep some capacity around for frees.
3422 if (!kmem_cache_has_cpu_partial(s
))
3424 else if (s
->size
>= PAGE_SIZE
)
3426 else if (s
->size
>= 1024)
3428 else if (s
->size
>= 256)
3429 s
->cpu_partial
= 13;
3431 s
->cpu_partial
= 30;
3436 * calculate_sizes() determines the order and the distribution of data within
3439 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3441 slab_flags_t flags
= s
->flags
;
3442 unsigned int size
= s
->object_size
;
3446 * Round up object size to the next word boundary. We can only
3447 * place the free pointer at word boundaries and this determines
3448 * the possible location of the free pointer.
3450 size
= ALIGN(size
, sizeof(void *));
3452 #ifdef CONFIG_SLUB_DEBUG
3454 * Determine if we can poison the object itself. If the user of
3455 * the slab may touch the object after free or before allocation
3456 * then we should never poison the object itself.
3458 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3460 s
->flags
|= __OBJECT_POISON
;
3462 s
->flags
&= ~__OBJECT_POISON
;
3466 * If we are Redzoning then check if there is some space between the
3467 * end of the object and the free pointer. If not then add an
3468 * additional word to have some bytes to store Redzone information.
3470 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3471 size
+= sizeof(void *);
3475 * With that we have determined the number of bytes in actual use
3476 * by the object. This is the potential offset to the free pointer.
3480 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3483 * Relocate free pointer after the object if it is not
3484 * permitted to overwrite the first word of the object on
3487 * This is the case if we do RCU, have a constructor or
3488 * destructor or are poisoning the objects.
3491 size
+= sizeof(void *);
3494 #ifdef CONFIG_SLUB_DEBUG
3495 if (flags
& SLAB_STORE_USER
)
3497 * Need to store information about allocs and frees after
3500 size
+= 2 * sizeof(struct track
);
3503 kasan_cache_create(s
, &size
, &s
->flags
);
3504 #ifdef CONFIG_SLUB_DEBUG
3505 if (flags
& SLAB_RED_ZONE
) {
3507 * Add some empty padding so that we can catch
3508 * overwrites from earlier objects rather than let
3509 * tracking information or the free pointer be
3510 * corrupted if a user writes before the start
3513 size
+= sizeof(void *);
3515 s
->red_left_pad
= sizeof(void *);
3516 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3517 size
+= s
->red_left_pad
;
3522 * SLUB stores one object immediately after another beginning from
3523 * offset 0. In order to align the objects we have to simply size
3524 * each object to conform to the alignment.
3526 size
= ALIGN(size
, s
->align
);
3528 if (forced_order
>= 0)
3529 order
= forced_order
;
3531 order
= calculate_order(size
);
3538 s
->allocflags
|= __GFP_COMP
;
3540 if (s
->flags
& SLAB_CACHE_DMA
)
3541 s
->allocflags
|= GFP_DMA
;
3543 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3544 s
->allocflags
|= __GFP_RECLAIMABLE
;
3547 * Determine the number of objects per slab
3549 s
->oo
= oo_make(order
, size
);
3550 s
->min
= oo_make(get_order(size
), size
);
3551 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3554 return !!oo_objects(s
->oo
);
3557 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3559 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3560 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3561 s
->random
= get_random_long();
3564 if (!calculate_sizes(s
, -1))
3566 if (disable_higher_order_debug
) {
3568 * Disable debugging flags that store metadata if the min slab
3571 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3572 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3574 if (!calculate_sizes(s
, -1))
3579 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3580 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3581 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3582 /* Enable fast mode */
3583 s
->flags
|= __CMPXCHG_DOUBLE
;
3587 * The larger the object size is, the more pages we want on the partial
3588 * list to avoid pounding the page allocator excessively.
3590 set_min_partial(s
, ilog2(s
->size
) / 2);
3595 s
->remote_node_defrag_ratio
= 1000;
3598 /* Initialize the pre-computed randomized freelist if slab is up */
3599 if (slab_state
>= UP
) {
3600 if (init_cache_random_seq(s
))
3604 if (!init_kmem_cache_nodes(s
))
3607 if (alloc_kmem_cache_cpus(s
))
3610 free_kmem_cache_nodes(s
);
3612 if (flags
& SLAB_PANIC
)
3613 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3614 s
->name
, s
->size
, s
->size
,
3615 oo_order(s
->oo
), s
->offset
, (unsigned long)flags
);
3619 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3622 #ifdef CONFIG_SLUB_DEBUG
3623 void *addr
= page_address(page
);
3625 unsigned long *map
= kcalloc(BITS_TO_LONGS(page
->objects
),
3630 slab_err(s
, page
, text
, s
->name
);
3633 get_map(s
, page
, map
);
3634 for_each_object(p
, s
, addr
, page
->objects
) {
3636 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3637 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3638 print_tracking(s
, p
);
3647 * Attempt to free all partial slabs on a node.
3648 * This is called from __kmem_cache_shutdown(). We must take list_lock
3649 * because sysfs file might still access partial list after the shutdowning.
3651 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3654 struct page
*page
, *h
;
3656 BUG_ON(irqs_disabled());
3657 spin_lock_irq(&n
->list_lock
);
3658 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3660 remove_partial(n
, page
);
3661 list_add(&page
->lru
, &discard
);
3663 list_slab_objects(s
, page
,
3664 "Objects remaining in %s on __kmem_cache_shutdown()");
3667 spin_unlock_irq(&n
->list_lock
);
3669 list_for_each_entry_safe(page
, h
, &discard
, lru
)
3670 discard_slab(s
, page
);
3673 bool __kmem_cache_empty(struct kmem_cache
*s
)
3676 struct kmem_cache_node
*n
;
3678 for_each_kmem_cache_node(s
, node
, n
)
3679 if (n
->nr_partial
|| slabs_node(s
, node
))
3685 * Release all resources used by a slab cache.
3687 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3690 struct kmem_cache_node
*n
;
3693 /* Attempt to free all objects */
3694 for_each_kmem_cache_node(s
, node
, n
) {
3696 if (n
->nr_partial
|| slabs_node(s
, node
))
3699 sysfs_slab_remove(s
);
3703 /********************************************************************
3705 *******************************************************************/
3707 static int __init
setup_slub_min_order(char *str
)
3709 get_option(&str
, (int *)&slub_min_order
);
3714 __setup("slub_min_order=", setup_slub_min_order
);
3716 static int __init
setup_slub_max_order(char *str
)
3718 get_option(&str
, (int *)&slub_max_order
);
3719 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
3724 __setup("slub_max_order=", setup_slub_max_order
);
3726 static int __init
setup_slub_min_objects(char *str
)
3728 get_option(&str
, (int *)&slub_min_objects
);
3733 __setup("slub_min_objects=", setup_slub_min_objects
);
3735 void *__kmalloc(size_t size
, gfp_t flags
)
3737 struct kmem_cache
*s
;
3740 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3741 return kmalloc_large(size
, flags
);
3743 s
= kmalloc_slab(size
, flags
);
3745 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3748 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3750 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3752 kasan_kmalloc(s
, ret
, size
, flags
);
3756 EXPORT_SYMBOL(__kmalloc
);
3759 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3764 flags
|= __GFP_COMP
;
3765 page
= alloc_pages_node(node
, flags
, get_order(size
));
3767 ptr
= page_address(page
);
3769 kmalloc_large_node_hook(ptr
, size
, flags
);
3773 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3775 struct kmem_cache
*s
;
3778 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3779 ret
= kmalloc_large_node(size
, flags
, node
);
3781 trace_kmalloc_node(_RET_IP_
, ret
,
3782 size
, PAGE_SIZE
<< get_order(size
),
3788 s
= kmalloc_slab(size
, flags
);
3790 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3793 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3795 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3797 kasan_kmalloc(s
, ret
, size
, flags
);
3801 EXPORT_SYMBOL(__kmalloc_node
);
3804 #ifdef CONFIG_HARDENED_USERCOPY
3806 * Rejects incorrectly sized objects and objects that are to be copied
3807 * to/from userspace but do not fall entirely within the containing slab
3808 * cache's usercopy region.
3810 * Returns NULL if check passes, otherwise const char * to name of cache
3811 * to indicate an error.
3813 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
3816 struct kmem_cache
*s
;
3817 unsigned int offset
;
3820 /* Find object and usable object size. */
3821 s
= page
->slab_cache
;
3823 /* Reject impossible pointers. */
3824 if (ptr
< page_address(page
))
3825 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
3828 /* Find offset within object. */
3829 offset
= (ptr
- page_address(page
)) % s
->size
;
3831 /* Adjust for redzone and reject if within the redzone. */
3832 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3833 if (offset
< s
->red_left_pad
)
3834 usercopy_abort("SLUB object in left red zone",
3835 s
->name
, to_user
, offset
, n
);
3836 offset
-= s
->red_left_pad
;
3839 /* Allow address range falling entirely within usercopy region. */
3840 if (offset
>= s
->useroffset
&&
3841 offset
- s
->useroffset
<= s
->usersize
&&
3842 n
<= s
->useroffset
- offset
+ s
->usersize
)
3846 * If the copy is still within the allocated object, produce
3847 * a warning instead of rejecting the copy. This is intended
3848 * to be a temporary method to find any missing usercopy
3851 object_size
= slab_ksize(s
);
3852 if (usercopy_fallback
&&
3853 offset
<= object_size
&& n
<= object_size
- offset
) {
3854 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
3858 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
3860 #endif /* CONFIG_HARDENED_USERCOPY */
3862 static size_t __ksize(const void *object
)
3866 if (unlikely(object
== ZERO_SIZE_PTR
))
3869 page
= virt_to_head_page(object
);
3871 if (unlikely(!PageSlab(page
))) {
3872 WARN_ON(!PageCompound(page
));
3873 return PAGE_SIZE
<< compound_order(page
);
3876 return slab_ksize(page
->slab_cache
);
3879 size_t ksize(const void *object
)
3881 size_t size
= __ksize(object
);
3882 /* We assume that ksize callers could use whole allocated area,
3883 * so we need to unpoison this area.
3885 kasan_unpoison_shadow(object
, size
);
3888 EXPORT_SYMBOL(ksize
);
3890 void kfree(const void *x
)
3893 void *object
= (void *)x
;
3895 trace_kfree(_RET_IP_
, x
);
3897 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3900 page
= virt_to_head_page(x
);
3901 if (unlikely(!PageSlab(page
))) {
3902 BUG_ON(!PageCompound(page
));
3904 __free_pages(page
, compound_order(page
));
3907 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3909 EXPORT_SYMBOL(kfree
);
3911 #define SHRINK_PROMOTE_MAX 32
3914 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3915 * up most to the head of the partial lists. New allocations will then
3916 * fill those up and thus they can be removed from the partial lists.
3918 * The slabs with the least items are placed last. This results in them
3919 * being allocated from last increasing the chance that the last objects
3920 * are freed in them.
3922 int __kmem_cache_shrink(struct kmem_cache
*s
)
3926 struct kmem_cache_node
*n
;
3929 struct list_head discard
;
3930 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3931 unsigned long flags
;
3935 for_each_kmem_cache_node(s
, node
, n
) {
3936 INIT_LIST_HEAD(&discard
);
3937 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3938 INIT_LIST_HEAD(promote
+ i
);
3940 spin_lock_irqsave(&n
->list_lock
, flags
);
3943 * Build lists of slabs to discard or promote.
3945 * Note that concurrent frees may occur while we hold the
3946 * list_lock. page->inuse here is the upper limit.
3948 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3949 int free
= page
->objects
- page
->inuse
;
3951 /* Do not reread page->inuse */
3954 /* We do not keep full slabs on the list */
3957 if (free
== page
->objects
) {
3958 list_move(&page
->lru
, &discard
);
3960 } else if (free
<= SHRINK_PROMOTE_MAX
)
3961 list_move(&page
->lru
, promote
+ free
- 1);
3965 * Promote the slabs filled up most to the head of the
3968 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3969 list_splice(promote
+ i
, &n
->partial
);
3971 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3973 /* Release empty slabs */
3974 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3975 discard_slab(s
, page
);
3977 if (slabs_node(s
, node
))
3985 static void kmemcg_cache_deact_after_rcu(struct kmem_cache
*s
)
3988 * Called with all the locks held after a sched RCU grace period.
3989 * Even if @s becomes empty after shrinking, we can't know that @s
3990 * doesn't have allocations already in-flight and thus can't
3991 * destroy @s until the associated memcg is released.
3993 * However, let's remove the sysfs files for empty caches here.
3994 * Each cache has a lot of interface files which aren't
3995 * particularly useful for empty draining caches; otherwise, we can
3996 * easily end up with millions of unnecessary sysfs files on
3997 * systems which have a lot of memory and transient cgroups.
3999 if (!__kmem_cache_shrink(s
))
4000 sysfs_slab_remove(s
);
4003 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4006 * Disable empty slabs caching. Used to avoid pinning offline
4007 * memory cgroups by kmem pages that can be freed.
4009 slub_set_cpu_partial(s
, 0);
4013 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4014 * we have to make sure the change is visible before shrinking.
4016 slab_deactivate_memcg_cache_rcu_sched(s
, kmemcg_cache_deact_after_rcu
);
4020 static int slab_mem_going_offline_callback(void *arg
)
4022 struct kmem_cache
*s
;
4024 mutex_lock(&slab_mutex
);
4025 list_for_each_entry(s
, &slab_caches
, list
)
4026 __kmem_cache_shrink(s
);
4027 mutex_unlock(&slab_mutex
);
4032 static void slab_mem_offline_callback(void *arg
)
4034 struct kmem_cache_node
*n
;
4035 struct kmem_cache
*s
;
4036 struct memory_notify
*marg
= arg
;
4039 offline_node
= marg
->status_change_nid_normal
;
4042 * If the node still has available memory. we need kmem_cache_node
4045 if (offline_node
< 0)
4048 mutex_lock(&slab_mutex
);
4049 list_for_each_entry(s
, &slab_caches
, list
) {
4050 n
= get_node(s
, offline_node
);
4053 * if n->nr_slabs > 0, slabs still exist on the node
4054 * that is going down. We were unable to free them,
4055 * and offline_pages() function shouldn't call this
4056 * callback. So, we must fail.
4058 BUG_ON(slabs_node(s
, offline_node
));
4060 s
->node
[offline_node
] = NULL
;
4061 kmem_cache_free(kmem_cache_node
, n
);
4064 mutex_unlock(&slab_mutex
);
4067 static int slab_mem_going_online_callback(void *arg
)
4069 struct kmem_cache_node
*n
;
4070 struct kmem_cache
*s
;
4071 struct memory_notify
*marg
= arg
;
4072 int nid
= marg
->status_change_nid_normal
;
4076 * If the node's memory is already available, then kmem_cache_node is
4077 * already created. Nothing to do.
4083 * We are bringing a node online. No memory is available yet. We must
4084 * allocate a kmem_cache_node structure in order to bring the node
4087 mutex_lock(&slab_mutex
);
4088 list_for_each_entry(s
, &slab_caches
, list
) {
4090 * XXX: kmem_cache_alloc_node will fallback to other nodes
4091 * since memory is not yet available from the node that
4094 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4099 init_kmem_cache_node(n
);
4103 mutex_unlock(&slab_mutex
);
4107 static int slab_memory_callback(struct notifier_block
*self
,
4108 unsigned long action
, void *arg
)
4113 case MEM_GOING_ONLINE
:
4114 ret
= slab_mem_going_online_callback(arg
);
4116 case MEM_GOING_OFFLINE
:
4117 ret
= slab_mem_going_offline_callback(arg
);
4120 case MEM_CANCEL_ONLINE
:
4121 slab_mem_offline_callback(arg
);
4124 case MEM_CANCEL_OFFLINE
:
4128 ret
= notifier_from_errno(ret
);
4134 static struct notifier_block slab_memory_callback_nb
= {
4135 .notifier_call
= slab_memory_callback
,
4136 .priority
= SLAB_CALLBACK_PRI
,
4139 /********************************************************************
4140 * Basic setup of slabs
4141 *******************************************************************/
4144 * Used for early kmem_cache structures that were allocated using
4145 * the page allocator. Allocate them properly then fix up the pointers
4146 * that may be pointing to the wrong kmem_cache structure.
4149 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4152 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4153 struct kmem_cache_node
*n
;
4155 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4158 * This runs very early, and only the boot processor is supposed to be
4159 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4162 __flush_cpu_slab(s
, smp_processor_id());
4163 for_each_kmem_cache_node(s
, node
, n
) {
4166 list_for_each_entry(p
, &n
->partial
, lru
)
4169 #ifdef CONFIG_SLUB_DEBUG
4170 list_for_each_entry(p
, &n
->full
, lru
)
4174 slab_init_memcg_params(s
);
4175 list_add(&s
->list
, &slab_caches
);
4176 memcg_link_cache(s
);
4180 void __init
kmem_cache_init(void)
4182 static __initdata
struct kmem_cache boot_kmem_cache
,
4183 boot_kmem_cache_node
;
4185 if (debug_guardpage_minorder())
4188 kmem_cache_node
= &boot_kmem_cache_node
;
4189 kmem_cache
= &boot_kmem_cache
;
4191 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4192 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4194 register_hotmemory_notifier(&slab_memory_callback_nb
);
4196 /* Able to allocate the per node structures */
4197 slab_state
= PARTIAL
;
4199 create_boot_cache(kmem_cache
, "kmem_cache",
4200 offsetof(struct kmem_cache
, node
) +
4201 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4202 SLAB_HWCACHE_ALIGN
, 0, 0);
4204 kmem_cache
= bootstrap(&boot_kmem_cache
);
4205 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4207 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4208 setup_kmalloc_cache_index_table();
4209 create_kmalloc_caches(0);
4211 /* Setup random freelists for each cache */
4212 init_freelist_randomization();
4214 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4217 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n",
4219 slub_min_order
, slub_max_order
, slub_min_objects
,
4220 nr_cpu_ids
, nr_node_ids
);
4223 void __init
kmem_cache_init_late(void)
4228 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4229 slab_flags_t flags
, void (*ctor
)(void *))
4231 struct kmem_cache
*s
, *c
;
4233 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4238 * Adjust the object sizes so that we clear
4239 * the complete object on kzalloc.
4241 s
->object_size
= max(s
->object_size
, size
);
4242 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4244 for_each_memcg_cache(c
, s
) {
4245 c
->object_size
= s
->object_size
;
4246 c
->inuse
= max(c
->inuse
, ALIGN(size
, sizeof(void *)));
4249 if (sysfs_slab_alias(s
, name
)) {
4258 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4262 err
= kmem_cache_open(s
, flags
);
4266 /* Mutex is not taken during early boot */
4267 if (slab_state
<= UP
)
4270 memcg_propagate_slab_attrs(s
);
4271 err
= sysfs_slab_add(s
);
4273 __kmem_cache_release(s
);
4278 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4280 struct kmem_cache
*s
;
4283 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4284 return kmalloc_large(size
, gfpflags
);
4286 s
= kmalloc_slab(size
, gfpflags
);
4288 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4291 ret
= slab_alloc(s
, gfpflags
, caller
);
4293 /* Honor the call site pointer we received. */
4294 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4300 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4301 int node
, unsigned long caller
)
4303 struct kmem_cache
*s
;
4306 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4307 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4309 trace_kmalloc_node(caller
, ret
,
4310 size
, PAGE_SIZE
<< get_order(size
),
4316 s
= kmalloc_slab(size
, gfpflags
);
4318 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4321 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4323 /* Honor the call site pointer we received. */
4324 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4331 static int count_inuse(struct page
*page
)
4336 static int count_total(struct page
*page
)
4338 return page
->objects
;
4342 #ifdef CONFIG_SLUB_DEBUG
4343 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4347 void *addr
= page_address(page
);
4349 if (!check_slab(s
, page
) ||
4350 !on_freelist(s
, page
, NULL
))
4353 /* Now we know that a valid freelist exists */
4354 bitmap_zero(map
, page
->objects
);
4356 get_map(s
, page
, map
);
4357 for_each_object(p
, s
, addr
, page
->objects
) {
4358 if (test_bit(slab_index(p
, s
, addr
), map
))
4359 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4363 for_each_object(p
, s
, addr
, page
->objects
)
4364 if (!test_bit(slab_index(p
, s
, addr
), map
))
4365 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4370 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4374 validate_slab(s
, page
, map
);
4378 static int validate_slab_node(struct kmem_cache
*s
,
4379 struct kmem_cache_node
*n
, unsigned long *map
)
4381 unsigned long count
= 0;
4383 unsigned long flags
;
4385 spin_lock_irqsave(&n
->list_lock
, flags
);
4387 list_for_each_entry(page
, &n
->partial
, lru
) {
4388 validate_slab_slab(s
, page
, map
);
4391 if (count
!= n
->nr_partial
)
4392 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4393 s
->name
, count
, n
->nr_partial
);
4395 if (!(s
->flags
& SLAB_STORE_USER
))
4398 list_for_each_entry(page
, &n
->full
, lru
) {
4399 validate_slab_slab(s
, page
, map
);
4402 if (count
!= atomic_long_read(&n
->nr_slabs
))
4403 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4404 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4407 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4411 static long validate_slab_cache(struct kmem_cache
*s
)
4414 unsigned long count
= 0;
4415 unsigned long *map
= kmalloc_array(BITS_TO_LONGS(oo_objects(s
->max
)),
4416 sizeof(unsigned long),
4418 struct kmem_cache_node
*n
;
4424 for_each_kmem_cache_node(s
, node
, n
)
4425 count
+= validate_slab_node(s
, n
, map
);
4430 * Generate lists of code addresses where slabcache objects are allocated
4435 unsigned long count
;
4442 DECLARE_BITMAP(cpus
, NR_CPUS
);
4448 unsigned long count
;
4449 struct location
*loc
;
4452 static void free_loc_track(struct loc_track
*t
)
4455 free_pages((unsigned long)t
->loc
,
4456 get_order(sizeof(struct location
) * t
->max
));
4459 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4464 order
= get_order(sizeof(struct location
) * max
);
4466 l
= (void *)__get_free_pages(flags
, order
);
4471 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4479 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4480 const struct track
*track
)
4482 long start
, end
, pos
;
4484 unsigned long caddr
;
4485 unsigned long age
= jiffies
- track
->when
;
4491 pos
= start
+ (end
- start
+ 1) / 2;
4494 * There is nothing at "end". If we end up there
4495 * we need to add something to before end.
4500 caddr
= t
->loc
[pos
].addr
;
4501 if (track
->addr
== caddr
) {
4507 if (age
< l
->min_time
)
4509 if (age
> l
->max_time
)
4512 if (track
->pid
< l
->min_pid
)
4513 l
->min_pid
= track
->pid
;
4514 if (track
->pid
> l
->max_pid
)
4515 l
->max_pid
= track
->pid
;
4517 cpumask_set_cpu(track
->cpu
,
4518 to_cpumask(l
->cpus
));
4520 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4524 if (track
->addr
< caddr
)
4531 * Not found. Insert new tracking element.
4533 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4539 (t
->count
- pos
) * sizeof(struct location
));
4542 l
->addr
= track
->addr
;
4546 l
->min_pid
= track
->pid
;
4547 l
->max_pid
= track
->pid
;
4548 cpumask_clear(to_cpumask(l
->cpus
));
4549 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4550 nodes_clear(l
->nodes
);
4551 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4555 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4556 struct page
*page
, enum track_item alloc
,
4559 void *addr
= page_address(page
);
4562 bitmap_zero(map
, page
->objects
);
4563 get_map(s
, page
, map
);
4565 for_each_object(p
, s
, addr
, page
->objects
)
4566 if (!test_bit(slab_index(p
, s
, addr
), map
))
4567 add_location(t
, s
, get_track(s
, p
, alloc
));
4570 static int list_locations(struct kmem_cache
*s
, char *buf
,
4571 enum track_item alloc
)
4575 struct loc_track t
= { 0, 0, NULL
};
4577 unsigned long *map
= kmalloc_array(BITS_TO_LONGS(oo_objects(s
->max
)),
4578 sizeof(unsigned long),
4580 struct kmem_cache_node
*n
;
4582 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4585 return sprintf(buf
, "Out of memory\n");
4587 /* Push back cpu slabs */
4590 for_each_kmem_cache_node(s
, node
, n
) {
4591 unsigned long flags
;
4594 if (!atomic_long_read(&n
->nr_slabs
))
4597 spin_lock_irqsave(&n
->list_lock
, flags
);
4598 list_for_each_entry(page
, &n
->partial
, lru
)
4599 process_slab(&t
, s
, page
, alloc
, map
);
4600 list_for_each_entry(page
, &n
->full
, lru
)
4601 process_slab(&t
, s
, page
, alloc
, map
);
4602 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4605 for (i
= 0; i
< t
.count
; i
++) {
4606 struct location
*l
= &t
.loc
[i
];
4608 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4610 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4613 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4615 len
+= sprintf(buf
+ len
, "<not-available>");
4617 if (l
->sum_time
!= l
->min_time
) {
4618 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4620 (long)div_u64(l
->sum_time
, l
->count
),
4623 len
+= sprintf(buf
+ len
, " age=%ld",
4626 if (l
->min_pid
!= l
->max_pid
)
4627 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4628 l
->min_pid
, l
->max_pid
);
4630 len
+= sprintf(buf
+ len
, " pid=%ld",
4633 if (num_online_cpus() > 1 &&
4634 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4635 len
< PAGE_SIZE
- 60)
4636 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4638 cpumask_pr_args(to_cpumask(l
->cpus
)));
4640 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4641 len
< PAGE_SIZE
- 60)
4642 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4644 nodemask_pr_args(&l
->nodes
));
4646 len
+= sprintf(buf
+ len
, "\n");
4652 len
+= sprintf(buf
, "No data\n");
4657 #ifdef SLUB_RESILIENCY_TEST
4658 static void __init
resiliency_test(void)
4662 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4664 pr_err("SLUB resiliency testing\n");
4665 pr_err("-----------------------\n");
4666 pr_err("A. Corruption after allocation\n");
4668 p
= kzalloc(16, GFP_KERNEL
);
4670 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4673 validate_slab_cache(kmalloc_caches
[4]);
4675 /* Hmmm... The next two are dangerous */
4676 p
= kzalloc(32, GFP_KERNEL
);
4677 p
[32 + sizeof(void *)] = 0x34;
4678 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4680 pr_err("If allocated object is overwritten then not detectable\n\n");
4682 validate_slab_cache(kmalloc_caches
[5]);
4683 p
= kzalloc(64, GFP_KERNEL
);
4684 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4686 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4688 pr_err("If allocated object is overwritten then not detectable\n\n");
4689 validate_slab_cache(kmalloc_caches
[6]);
4691 pr_err("\nB. Corruption after free\n");
4692 p
= kzalloc(128, GFP_KERNEL
);
4695 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4696 validate_slab_cache(kmalloc_caches
[7]);
4698 p
= kzalloc(256, GFP_KERNEL
);
4701 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4702 validate_slab_cache(kmalloc_caches
[8]);
4704 p
= kzalloc(512, GFP_KERNEL
);
4707 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4708 validate_slab_cache(kmalloc_caches
[9]);
4712 static void resiliency_test(void) {};
4717 enum slab_stat_type
{
4718 SL_ALL
, /* All slabs */
4719 SL_PARTIAL
, /* Only partially allocated slabs */
4720 SL_CPU
, /* Only slabs used for cpu caches */
4721 SL_OBJECTS
, /* Determine allocated objects not slabs */
4722 SL_TOTAL
/* Determine object capacity not slabs */
4725 #define SO_ALL (1 << SL_ALL)
4726 #define SO_PARTIAL (1 << SL_PARTIAL)
4727 #define SO_CPU (1 << SL_CPU)
4728 #define SO_OBJECTS (1 << SL_OBJECTS)
4729 #define SO_TOTAL (1 << SL_TOTAL)
4732 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4734 static int __init
setup_slub_memcg_sysfs(char *str
)
4738 if (get_option(&str
, &v
) > 0)
4739 memcg_sysfs_enabled
= v
;
4744 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4747 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4748 char *buf
, unsigned long flags
)
4750 unsigned long total
= 0;
4753 unsigned long *nodes
;
4755 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4759 if (flags
& SO_CPU
) {
4762 for_each_possible_cpu(cpu
) {
4763 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4768 page
= READ_ONCE(c
->page
);
4772 node
= page_to_nid(page
);
4773 if (flags
& SO_TOTAL
)
4775 else if (flags
& SO_OBJECTS
)
4783 page
= slub_percpu_partial_read_once(c
);
4785 node
= page_to_nid(page
);
4786 if (flags
& SO_TOTAL
)
4788 else if (flags
& SO_OBJECTS
)
4799 #ifdef CONFIG_SLUB_DEBUG
4800 if (flags
& SO_ALL
) {
4801 struct kmem_cache_node
*n
;
4803 for_each_kmem_cache_node(s
, node
, n
) {
4805 if (flags
& SO_TOTAL
)
4806 x
= atomic_long_read(&n
->total_objects
);
4807 else if (flags
& SO_OBJECTS
)
4808 x
= atomic_long_read(&n
->total_objects
) -
4809 count_partial(n
, count_free
);
4811 x
= atomic_long_read(&n
->nr_slabs
);
4818 if (flags
& SO_PARTIAL
) {
4819 struct kmem_cache_node
*n
;
4821 for_each_kmem_cache_node(s
, node
, n
) {
4822 if (flags
& SO_TOTAL
)
4823 x
= count_partial(n
, count_total
);
4824 else if (flags
& SO_OBJECTS
)
4825 x
= count_partial(n
, count_inuse
);
4832 x
= sprintf(buf
, "%lu", total
);
4834 for (node
= 0; node
< nr_node_ids
; node
++)
4836 x
+= sprintf(buf
+ x
, " N%d=%lu",
4841 return x
+ sprintf(buf
+ x
, "\n");
4844 #ifdef CONFIG_SLUB_DEBUG
4845 static int any_slab_objects(struct kmem_cache
*s
)
4848 struct kmem_cache_node
*n
;
4850 for_each_kmem_cache_node(s
, node
, n
)
4851 if (atomic_long_read(&n
->total_objects
))
4858 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4859 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4861 struct slab_attribute
{
4862 struct attribute attr
;
4863 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4864 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4867 #define SLAB_ATTR_RO(_name) \
4868 static struct slab_attribute _name##_attr = \
4869 __ATTR(_name, 0400, _name##_show, NULL)
4871 #define SLAB_ATTR(_name) \
4872 static struct slab_attribute _name##_attr = \
4873 __ATTR(_name, 0600, _name##_show, _name##_store)
4875 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4877 return sprintf(buf
, "%u\n", s
->size
);
4879 SLAB_ATTR_RO(slab_size
);
4881 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4883 return sprintf(buf
, "%u\n", s
->align
);
4885 SLAB_ATTR_RO(align
);
4887 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4889 return sprintf(buf
, "%u\n", s
->object_size
);
4891 SLAB_ATTR_RO(object_size
);
4893 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4895 return sprintf(buf
, "%u\n", oo_objects(s
->oo
));
4897 SLAB_ATTR_RO(objs_per_slab
);
4899 static ssize_t
order_store(struct kmem_cache
*s
,
4900 const char *buf
, size_t length
)
4905 err
= kstrtouint(buf
, 10, &order
);
4909 if (order
> slub_max_order
|| order
< slub_min_order
)
4912 calculate_sizes(s
, order
);
4916 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4918 return sprintf(buf
, "%u\n", oo_order(s
->oo
));
4922 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4924 return sprintf(buf
, "%lu\n", s
->min_partial
);
4927 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4933 err
= kstrtoul(buf
, 10, &min
);
4937 set_min_partial(s
, min
);
4940 SLAB_ATTR(min_partial
);
4942 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4944 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4947 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4950 unsigned int objects
;
4953 err
= kstrtouint(buf
, 10, &objects
);
4956 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4959 slub_set_cpu_partial(s
, objects
);
4963 SLAB_ATTR(cpu_partial
);
4965 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4969 return sprintf(buf
, "%pS\n", s
->ctor
);
4973 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4975 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4977 SLAB_ATTR_RO(aliases
);
4979 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
4981 return show_slab_objects(s
, buf
, SO_PARTIAL
);
4983 SLAB_ATTR_RO(partial
);
4985 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
4987 return show_slab_objects(s
, buf
, SO_CPU
);
4989 SLAB_ATTR_RO(cpu_slabs
);
4991 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
4993 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
4995 SLAB_ATTR_RO(objects
);
4997 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
4999 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5001 SLAB_ATTR_RO(objects_partial
);
5003 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5010 for_each_online_cpu(cpu
) {
5013 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5016 pages
+= page
->pages
;
5017 objects
+= page
->pobjects
;
5021 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5024 for_each_online_cpu(cpu
) {
5027 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5029 if (page
&& len
< PAGE_SIZE
- 20)
5030 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5031 page
->pobjects
, page
->pages
);
5034 return len
+ sprintf(buf
+ len
, "\n");
5036 SLAB_ATTR_RO(slabs_cpu_partial
);
5038 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5040 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5043 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5044 const char *buf
, size_t length
)
5046 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5048 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5051 SLAB_ATTR(reclaim_account
);
5053 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5055 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5057 SLAB_ATTR_RO(hwcache_align
);
5059 #ifdef CONFIG_ZONE_DMA
5060 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5062 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5064 SLAB_ATTR_RO(cache_dma
);
5067 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5069 return sprintf(buf
, "%u\n", s
->usersize
);
5071 SLAB_ATTR_RO(usersize
);
5073 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5075 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5077 SLAB_ATTR_RO(destroy_by_rcu
);
5079 #ifdef CONFIG_SLUB_DEBUG
5080 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5082 return show_slab_objects(s
, buf
, SO_ALL
);
5084 SLAB_ATTR_RO(slabs
);
5086 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5088 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5090 SLAB_ATTR_RO(total_objects
);
5092 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5094 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5097 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5098 const char *buf
, size_t length
)
5100 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5101 if (buf
[0] == '1') {
5102 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5103 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5107 SLAB_ATTR(sanity_checks
);
5109 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5111 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5114 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5118 * Tracing a merged cache is going to give confusing results
5119 * as well as cause other issues like converting a mergeable
5120 * cache into an umergeable one.
5122 if (s
->refcount
> 1)
5125 s
->flags
&= ~SLAB_TRACE
;
5126 if (buf
[0] == '1') {
5127 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5128 s
->flags
|= SLAB_TRACE
;
5134 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5136 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5139 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5140 const char *buf
, size_t length
)
5142 if (any_slab_objects(s
))
5145 s
->flags
&= ~SLAB_RED_ZONE
;
5146 if (buf
[0] == '1') {
5147 s
->flags
|= SLAB_RED_ZONE
;
5149 calculate_sizes(s
, -1);
5152 SLAB_ATTR(red_zone
);
5154 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5156 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5159 static ssize_t
poison_store(struct kmem_cache
*s
,
5160 const char *buf
, size_t length
)
5162 if (any_slab_objects(s
))
5165 s
->flags
&= ~SLAB_POISON
;
5166 if (buf
[0] == '1') {
5167 s
->flags
|= SLAB_POISON
;
5169 calculate_sizes(s
, -1);
5174 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5176 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5179 static ssize_t
store_user_store(struct kmem_cache
*s
,
5180 const char *buf
, size_t length
)
5182 if (any_slab_objects(s
))
5185 s
->flags
&= ~SLAB_STORE_USER
;
5186 if (buf
[0] == '1') {
5187 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5188 s
->flags
|= SLAB_STORE_USER
;
5190 calculate_sizes(s
, -1);
5193 SLAB_ATTR(store_user
);
5195 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5200 static ssize_t
validate_store(struct kmem_cache
*s
,
5201 const char *buf
, size_t length
)
5205 if (buf
[0] == '1') {
5206 ret
= validate_slab_cache(s
);
5212 SLAB_ATTR(validate
);
5214 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5216 if (!(s
->flags
& SLAB_STORE_USER
))
5218 return list_locations(s
, buf
, TRACK_ALLOC
);
5220 SLAB_ATTR_RO(alloc_calls
);
5222 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5224 if (!(s
->flags
& SLAB_STORE_USER
))
5226 return list_locations(s
, buf
, TRACK_FREE
);
5228 SLAB_ATTR_RO(free_calls
);
5229 #endif /* CONFIG_SLUB_DEBUG */
5231 #ifdef CONFIG_FAILSLAB
5232 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5234 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5237 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5240 if (s
->refcount
> 1)
5243 s
->flags
&= ~SLAB_FAILSLAB
;
5245 s
->flags
|= SLAB_FAILSLAB
;
5248 SLAB_ATTR(failslab
);
5251 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5256 static ssize_t
shrink_store(struct kmem_cache
*s
,
5257 const char *buf
, size_t length
)
5260 kmem_cache_shrink(s
);
5268 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5270 return sprintf(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5273 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5274 const char *buf
, size_t length
)
5279 err
= kstrtouint(buf
, 10, &ratio
);
5285 s
->remote_node_defrag_ratio
= ratio
* 10;
5289 SLAB_ATTR(remote_node_defrag_ratio
);
5292 #ifdef CONFIG_SLUB_STATS
5293 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5295 unsigned long sum
= 0;
5298 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5303 for_each_online_cpu(cpu
) {
5304 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5310 len
= sprintf(buf
, "%lu", sum
);
5313 for_each_online_cpu(cpu
) {
5314 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5315 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5319 return len
+ sprintf(buf
+ len
, "\n");
5322 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5326 for_each_online_cpu(cpu
)
5327 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5330 #define STAT_ATTR(si, text) \
5331 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5333 return show_stat(s, buf, si); \
5335 static ssize_t text##_store(struct kmem_cache *s, \
5336 const char *buf, size_t length) \
5338 if (buf[0] != '0') \
5340 clear_stat(s, si); \
5345 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5346 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5347 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5348 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5349 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5350 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5351 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5352 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5353 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5354 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5355 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5356 STAT_ATTR(FREE_SLAB
, free_slab
);
5357 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5358 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5359 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5360 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5361 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5362 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5363 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5364 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5365 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5366 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5367 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5368 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5369 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5370 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5373 static struct attribute
*slab_attrs
[] = {
5374 &slab_size_attr
.attr
,
5375 &object_size_attr
.attr
,
5376 &objs_per_slab_attr
.attr
,
5378 &min_partial_attr
.attr
,
5379 &cpu_partial_attr
.attr
,
5381 &objects_partial_attr
.attr
,
5383 &cpu_slabs_attr
.attr
,
5387 &hwcache_align_attr
.attr
,
5388 &reclaim_account_attr
.attr
,
5389 &destroy_by_rcu_attr
.attr
,
5391 &slabs_cpu_partial_attr
.attr
,
5392 #ifdef CONFIG_SLUB_DEBUG
5393 &total_objects_attr
.attr
,
5395 &sanity_checks_attr
.attr
,
5397 &red_zone_attr
.attr
,
5399 &store_user_attr
.attr
,
5400 &validate_attr
.attr
,
5401 &alloc_calls_attr
.attr
,
5402 &free_calls_attr
.attr
,
5404 #ifdef CONFIG_ZONE_DMA
5405 &cache_dma_attr
.attr
,
5408 &remote_node_defrag_ratio_attr
.attr
,
5410 #ifdef CONFIG_SLUB_STATS
5411 &alloc_fastpath_attr
.attr
,
5412 &alloc_slowpath_attr
.attr
,
5413 &free_fastpath_attr
.attr
,
5414 &free_slowpath_attr
.attr
,
5415 &free_frozen_attr
.attr
,
5416 &free_add_partial_attr
.attr
,
5417 &free_remove_partial_attr
.attr
,
5418 &alloc_from_partial_attr
.attr
,
5419 &alloc_slab_attr
.attr
,
5420 &alloc_refill_attr
.attr
,
5421 &alloc_node_mismatch_attr
.attr
,
5422 &free_slab_attr
.attr
,
5423 &cpuslab_flush_attr
.attr
,
5424 &deactivate_full_attr
.attr
,
5425 &deactivate_empty_attr
.attr
,
5426 &deactivate_to_head_attr
.attr
,
5427 &deactivate_to_tail_attr
.attr
,
5428 &deactivate_remote_frees_attr
.attr
,
5429 &deactivate_bypass_attr
.attr
,
5430 &order_fallback_attr
.attr
,
5431 &cmpxchg_double_fail_attr
.attr
,
5432 &cmpxchg_double_cpu_fail_attr
.attr
,
5433 &cpu_partial_alloc_attr
.attr
,
5434 &cpu_partial_free_attr
.attr
,
5435 &cpu_partial_node_attr
.attr
,
5436 &cpu_partial_drain_attr
.attr
,
5438 #ifdef CONFIG_FAILSLAB
5439 &failslab_attr
.attr
,
5441 &usersize_attr
.attr
,
5446 static const struct attribute_group slab_attr_group
= {
5447 .attrs
= slab_attrs
,
5450 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5451 struct attribute
*attr
,
5454 struct slab_attribute
*attribute
;
5455 struct kmem_cache
*s
;
5458 attribute
= to_slab_attr(attr
);
5461 if (!attribute
->show
)
5464 err
= attribute
->show(s
, buf
);
5469 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5470 struct attribute
*attr
,
5471 const char *buf
, size_t len
)
5473 struct slab_attribute
*attribute
;
5474 struct kmem_cache
*s
;
5477 attribute
= to_slab_attr(attr
);
5480 if (!attribute
->store
)
5483 err
= attribute
->store(s
, buf
, len
);
5485 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5486 struct kmem_cache
*c
;
5488 mutex_lock(&slab_mutex
);
5489 if (s
->max_attr_size
< len
)
5490 s
->max_attr_size
= len
;
5493 * This is a best effort propagation, so this function's return
5494 * value will be determined by the parent cache only. This is
5495 * basically because not all attributes will have a well
5496 * defined semantics for rollbacks - most of the actions will
5497 * have permanent effects.
5499 * Returning the error value of any of the children that fail
5500 * is not 100 % defined, in the sense that users seeing the
5501 * error code won't be able to know anything about the state of
5504 * Only returning the error code for the parent cache at least
5505 * has well defined semantics. The cache being written to
5506 * directly either failed or succeeded, in which case we loop
5507 * through the descendants with best-effort propagation.
5509 for_each_memcg_cache(c
, s
)
5510 attribute
->store(c
, buf
, len
);
5511 mutex_unlock(&slab_mutex
);
5517 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5521 char *buffer
= NULL
;
5522 struct kmem_cache
*root_cache
;
5524 if (is_root_cache(s
))
5527 root_cache
= s
->memcg_params
.root_cache
;
5530 * This mean this cache had no attribute written. Therefore, no point
5531 * in copying default values around
5533 if (!root_cache
->max_attr_size
)
5536 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5539 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5542 if (!attr
|| !attr
->store
|| !attr
->show
)
5546 * It is really bad that we have to allocate here, so we will
5547 * do it only as a fallback. If we actually allocate, though,
5548 * we can just use the allocated buffer until the end.
5550 * Most of the slub attributes will tend to be very small in
5551 * size, but sysfs allows buffers up to a page, so they can
5552 * theoretically happen.
5556 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5559 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5560 if (WARN_ON(!buffer
))
5565 len
= attr
->show(root_cache
, buf
);
5567 attr
->store(s
, buf
, len
);
5571 free_page((unsigned long)buffer
);
5575 static void kmem_cache_release(struct kobject
*k
)
5577 slab_kmem_cache_release(to_slab(k
));
5580 static const struct sysfs_ops slab_sysfs_ops
= {
5581 .show
= slab_attr_show
,
5582 .store
= slab_attr_store
,
5585 static struct kobj_type slab_ktype
= {
5586 .sysfs_ops
= &slab_sysfs_ops
,
5587 .release
= kmem_cache_release
,
5590 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5592 struct kobj_type
*ktype
= get_ktype(kobj
);
5594 if (ktype
== &slab_ktype
)
5599 static const struct kset_uevent_ops slab_uevent_ops
= {
5600 .filter
= uevent_filter
,
5603 static struct kset
*slab_kset
;
5605 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5608 if (!is_root_cache(s
))
5609 return s
->memcg_params
.root_cache
->memcg_kset
;
5614 #define ID_STR_LENGTH 64
5616 /* Create a unique string id for a slab cache:
5618 * Format :[flags-]size
5620 static char *create_unique_id(struct kmem_cache
*s
)
5622 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5629 * First flags affecting slabcache operations. We will only
5630 * get here for aliasable slabs so we do not need to support
5631 * too many flags. The flags here must cover all flags that
5632 * are matched during merging to guarantee that the id is
5635 if (s
->flags
& SLAB_CACHE_DMA
)
5637 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5639 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5641 if (s
->flags
& SLAB_ACCOUNT
)
5645 p
+= sprintf(p
, "%07u", s
->size
);
5647 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5651 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5653 struct kmem_cache
*s
=
5654 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5656 if (!s
->kobj
.state_in_sysfs
)
5658 * For a memcg cache, this may be called during
5659 * deactivation and again on shutdown. Remove only once.
5660 * A cache is never shut down before deactivation is
5661 * complete, so no need to worry about synchronization.
5666 kset_unregister(s
->memcg_kset
);
5668 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5670 kobject_put(&s
->kobj
);
5673 static int sysfs_slab_add(struct kmem_cache
*s
)
5677 struct kset
*kset
= cache_kset(s
);
5678 int unmergeable
= slab_unmergeable(s
);
5680 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5683 kobject_init(&s
->kobj
, &slab_ktype
);
5687 if (!unmergeable
&& disable_higher_order_debug
&&
5688 (slub_debug
& DEBUG_METADATA_FLAGS
))
5693 * Slabcache can never be merged so we can use the name proper.
5694 * This is typically the case for debug situations. In that
5695 * case we can catch duplicate names easily.
5697 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5701 * Create a unique name for the slab as a target
5704 name
= create_unique_id(s
);
5707 s
->kobj
.kset
= kset
;
5708 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5712 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5717 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5718 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5719 if (!s
->memcg_kset
) {
5726 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5728 /* Setup first alias */
5729 sysfs_slab_alias(s
, s
->name
);
5736 kobject_del(&s
->kobj
);
5740 static void sysfs_slab_remove(struct kmem_cache
*s
)
5742 if (slab_state
< FULL
)
5744 * Sysfs has not been setup yet so no need to remove the
5749 kobject_get(&s
->kobj
);
5750 schedule_work(&s
->kobj_remove_work
);
5753 void sysfs_slab_unlink(struct kmem_cache
*s
)
5755 if (slab_state
>= FULL
)
5756 kobject_del(&s
->kobj
);
5759 void sysfs_slab_release(struct kmem_cache
*s
)
5761 if (slab_state
>= FULL
)
5762 kobject_put(&s
->kobj
);
5766 * Need to buffer aliases during bootup until sysfs becomes
5767 * available lest we lose that information.
5769 struct saved_alias
{
5770 struct kmem_cache
*s
;
5772 struct saved_alias
*next
;
5775 static struct saved_alias
*alias_list
;
5777 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5779 struct saved_alias
*al
;
5781 if (slab_state
== FULL
) {
5783 * If we have a leftover link then remove it.
5785 sysfs_remove_link(&slab_kset
->kobj
, name
);
5786 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5789 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5795 al
->next
= alias_list
;
5800 static int __init
slab_sysfs_init(void)
5802 struct kmem_cache
*s
;
5805 mutex_lock(&slab_mutex
);
5807 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5809 mutex_unlock(&slab_mutex
);
5810 pr_err("Cannot register slab subsystem.\n");
5816 list_for_each_entry(s
, &slab_caches
, list
) {
5817 err
= sysfs_slab_add(s
);
5819 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5823 while (alias_list
) {
5824 struct saved_alias
*al
= alias_list
;
5826 alias_list
= alias_list
->next
;
5827 err
= sysfs_slab_alias(al
->s
, al
->name
);
5829 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5834 mutex_unlock(&slab_mutex
);
5839 __initcall(slab_sysfs_init
);
5840 #endif /* CONFIG_SYSFS */
5843 * The /proc/slabinfo ABI
5845 #ifdef CONFIG_SLUB_DEBUG
5846 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5848 unsigned long nr_slabs
= 0;
5849 unsigned long nr_objs
= 0;
5850 unsigned long nr_free
= 0;
5852 struct kmem_cache_node
*n
;
5854 for_each_kmem_cache_node(s
, node
, n
) {
5855 nr_slabs
+= node_nr_slabs(n
);
5856 nr_objs
+= node_nr_objs(n
);
5857 nr_free
+= count_partial(n
, count_free
);
5860 sinfo
->active_objs
= nr_objs
- nr_free
;
5861 sinfo
->num_objs
= nr_objs
;
5862 sinfo
->active_slabs
= nr_slabs
;
5863 sinfo
->num_slabs
= nr_slabs
;
5864 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5865 sinfo
->cache_order
= oo_order(s
->oo
);
5868 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5872 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5873 size_t count
, loff_t
*ppos
)
5877 #endif /* CONFIG_SLUB_DEBUG */