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/kmemcheck.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ctype.h>
30 #include <linux/debugobjects.h>
31 #include <linux/kallsyms.h>
32 #include <linux/memory.h>
33 #include <linux/math64.h>
34 #include <linux/fault-inject.h>
35 #include <linux/stacktrace.h>
36 #include <linux/prefetch.h>
37 #include <linux/memcontrol.h>
38 #include <linux/random.h>
40 #include <trace/events/kmem.h>
46 * 1. slab_mutex (Global Mutex)
48 * 3. slab_lock(page) (Only on some arches and for debugging)
52 * The role of the slab_mutex is to protect the list of all the slabs
53 * and to synchronize major metadata changes to slab cache structures.
55 * The slab_lock is only used for debugging and on arches that do not
56 * have the ability to do a cmpxchg_double. It only protects the second
57 * double word in the page struct. Meaning
58 * A. page->freelist -> List of object free in a page
59 * B. page->counters -> Counters of objects
60 * C. page->frozen -> frozen state
62 * If a slab is frozen then it is exempt from list management. It is not
63 * on any list. The processor that froze the slab is the one who can
64 * perform list operations on the page. Other processors may put objects
65 * onto the freelist but the processor that froze the slab is the only
66 * one that can retrieve the objects from the page's freelist.
68 * The list_lock protects the partial and full list on each node and
69 * the partial slab counter. If taken then no new slabs may be added or
70 * removed from the lists nor make the number of partial slabs be modified.
71 * (Note that the total number of slabs is an atomic value that may be
72 * modified without taking the list lock).
74 * The list_lock is a centralized lock and thus we avoid taking it as
75 * much as possible. As long as SLUB does not have to handle partial
76 * slabs, operations can continue without any centralized lock. F.e.
77 * allocating a long series of objects that fill up slabs does not require
79 * Interrupts are disabled during allocation and deallocation in order to
80 * make the slab allocator safe to use in the context of an irq. In addition
81 * interrupts are disabled to ensure that the processor does not change
82 * while handling per_cpu slabs, due to kernel preemption.
84 * SLUB assigns one slab for allocation to each processor.
85 * Allocations only occur from these slabs called cpu slabs.
87 * Slabs with free elements are kept on a partial list and during regular
88 * operations no list for full slabs is used. If an object in a full slab is
89 * freed then the slab will show up again on the partial lists.
90 * We track full slabs for debugging purposes though because otherwise we
91 * cannot scan all objects.
93 * Slabs are freed when they become empty. Teardown and setup is
94 * minimal so we rely on the page allocators per cpu caches for
95 * fast frees and allocs.
97 * Overloading of page flags that are otherwise used for LRU management.
99 * PageActive The slab is frozen and exempt from list processing.
100 * This means that the slab is dedicated to a purpose
101 * such as satisfying allocations for a specific
102 * processor. Objects may be freed in the slab while
103 * it is frozen but slab_free will then skip the usual
104 * list operations. It is up to the processor holding
105 * the slab to integrate the slab into the slab lists
106 * when the slab is no longer needed.
108 * One use of this flag is to mark slabs that are
109 * used for allocations. Then such a slab becomes a cpu
110 * slab. The cpu slab may be equipped with an additional
111 * freelist that allows lockless access to
112 * free objects in addition to the regular freelist
113 * that requires the slab lock.
115 * PageError Slab requires special handling due to debug
116 * options set. This moves slab handling out of
117 * the fast path and disables lockless freelists.
120 static inline int kmem_cache_debug(struct kmem_cache
*s
)
122 #ifdef CONFIG_SLUB_DEBUG
123 return unlikely(s
->flags
& SLAB_DEBUG_FLAGS
);
129 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
131 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
)
132 p
+= s
->red_left_pad
;
137 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
139 #ifdef CONFIG_SLUB_CPU_PARTIAL
140 return !kmem_cache_debug(s
);
147 * Issues still to be resolved:
149 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
151 * - Variable sizing of the per node arrays
154 /* Enable to test recovery from slab corruption on boot */
155 #undef SLUB_RESILIENCY_TEST
157 /* Enable to log cmpxchg failures */
158 #undef SLUB_DEBUG_CMPXCHG
161 * Mininum number of partial slabs. These will be left on the partial
162 * lists even if they are empty. kmem_cache_shrink may reclaim them.
164 #define MIN_PARTIAL 5
167 * Maximum number of desirable partial slabs.
168 * The existence of more partial slabs makes kmem_cache_shrink
169 * sort the partial list by the number of objects in use.
171 #define MAX_PARTIAL 10
173 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_STORE_USER)
177 * These debug flags cannot use CMPXCHG because there might be consistency
178 * issues when checking or reading debug information
180 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
185 * Debugging flags that require metadata to be stored in the slab. These get
186 * disabled when slub_debug=O is used and a cache's min order increases with
189 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
192 #define OO_MASK ((1 << OO_SHIFT) - 1)
193 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
195 /* Internal SLUB flags */
197 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
198 /* Use cmpxchg_double */
199 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
202 * Tracking user of a slab.
204 #define TRACK_ADDRS_COUNT 16
206 unsigned long addr
; /* Called from address */
207 #ifdef CONFIG_STACKTRACE
208 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
210 int cpu
; /* Was running on cpu */
211 int pid
; /* Pid context */
212 unsigned long when
; /* When did the operation occur */
215 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
218 static int sysfs_slab_add(struct kmem_cache
*);
219 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
220 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
);
221 static void sysfs_slab_remove(struct kmem_cache
*s
);
223 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
224 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
226 static inline void memcg_propagate_slab_attrs(struct kmem_cache
*s
) { }
227 static inline void sysfs_slab_remove(struct kmem_cache
*s
) { }
230 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
232 #ifdef CONFIG_SLUB_STATS
234 * The rmw is racy on a preemptible kernel but this is acceptable, so
235 * avoid this_cpu_add()'s irq-disable overhead.
237 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
241 /********************************************************************
242 * Core slab cache functions
243 *******************************************************************/
246 * Returns freelist pointer (ptr). With hardening, this is obfuscated
247 * with an XOR of the address where the pointer is held and a per-cache
250 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
251 unsigned long ptr_addr
)
253 #ifdef CONFIG_SLAB_FREELIST_HARDENED
254 return (void *)((unsigned long)ptr
^ s
->random
^ ptr_addr
);
260 /* Returns the freelist pointer recorded at location ptr_addr. */
261 static inline void *freelist_dereference(const struct kmem_cache
*s
,
264 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
265 (unsigned long)ptr_addr
);
268 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
270 return freelist_dereference(s
, object
+ s
->offset
);
273 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
276 prefetch(freelist_dereference(s
, object
+ s
->offset
));
279 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
281 unsigned long freepointer_addr
;
284 if (!debug_pagealloc_enabled())
285 return get_freepointer(s
, object
);
287 freepointer_addr
= (unsigned long)object
+ s
->offset
;
288 probe_kernel_read(&p
, (void **)freepointer_addr
, sizeof(p
));
289 return freelist_ptr(s
, p
, freepointer_addr
);
292 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
294 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
296 #ifdef CONFIG_SLAB_FREELIST_HARDENED
297 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
300 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
303 /* Loop over all objects in a slab */
304 #define for_each_object(__p, __s, __addr, __objects) \
305 for (__p = fixup_red_left(__s, __addr); \
306 __p < (__addr) + (__objects) * (__s)->size; \
309 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
310 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
311 __idx <= __objects; \
312 __p += (__s)->size, __idx++)
314 /* Determine object index from a given position */
315 static inline int slab_index(void *p
, struct kmem_cache
*s
, void *addr
)
317 return (p
- addr
) / s
->size
;
320 static inline int order_objects(int order
, unsigned long size
, int reserved
)
322 return ((PAGE_SIZE
<< order
) - reserved
) / size
;
325 static inline struct kmem_cache_order_objects
oo_make(int order
,
326 unsigned long size
, int reserved
)
328 struct kmem_cache_order_objects x
= {
329 (order
<< OO_SHIFT
) + order_objects(order
, size
, reserved
)
335 static inline int oo_order(struct kmem_cache_order_objects x
)
337 return x
.x
>> OO_SHIFT
;
340 static inline int oo_objects(struct kmem_cache_order_objects x
)
342 return x
.x
& OO_MASK
;
346 * Per slab locking using the pagelock
348 static __always_inline
void slab_lock(struct page
*page
)
350 VM_BUG_ON_PAGE(PageTail(page
), page
);
351 bit_spin_lock(PG_locked
, &page
->flags
);
354 static __always_inline
void slab_unlock(struct page
*page
)
356 VM_BUG_ON_PAGE(PageTail(page
), page
);
357 __bit_spin_unlock(PG_locked
, &page
->flags
);
360 static inline void set_page_slub_counters(struct page
*page
, unsigned long counters_new
)
363 tmp
.counters
= counters_new
;
365 * page->counters can cover frozen/inuse/objects as well
366 * as page->_refcount. If we assign to ->counters directly
367 * we run the risk of losing updates to page->_refcount, so
368 * be careful and only assign to the fields we need.
370 page
->frozen
= tmp
.frozen
;
371 page
->inuse
= tmp
.inuse
;
372 page
->objects
= tmp
.objects
;
375 /* Interrupts must be disabled (for the fallback code to work right) */
376 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
377 void *freelist_old
, unsigned long counters_old
,
378 void *freelist_new
, unsigned long counters_new
,
381 VM_BUG_ON(!irqs_disabled());
382 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
383 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
384 if (s
->flags
& __CMPXCHG_DOUBLE
) {
385 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
386 freelist_old
, counters_old
,
387 freelist_new
, counters_new
))
393 if (page
->freelist
== freelist_old
&&
394 page
->counters
== counters_old
) {
395 page
->freelist
= freelist_new
;
396 set_page_slub_counters(page
, counters_new
);
404 stat(s
, CMPXCHG_DOUBLE_FAIL
);
406 #ifdef SLUB_DEBUG_CMPXCHG
407 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
413 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
414 void *freelist_old
, unsigned long counters_old
,
415 void *freelist_new
, unsigned long counters_new
,
418 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
419 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
420 if (s
->flags
& __CMPXCHG_DOUBLE
) {
421 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
422 freelist_old
, counters_old
,
423 freelist_new
, counters_new
))
430 local_irq_save(flags
);
432 if (page
->freelist
== freelist_old
&&
433 page
->counters
== counters_old
) {
434 page
->freelist
= freelist_new
;
435 set_page_slub_counters(page
, counters_new
);
437 local_irq_restore(flags
);
441 local_irq_restore(flags
);
445 stat(s
, CMPXCHG_DOUBLE_FAIL
);
447 #ifdef SLUB_DEBUG_CMPXCHG
448 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
454 #ifdef CONFIG_SLUB_DEBUG
456 * Determine a map of object in use on a page.
458 * Node listlock must be held to guarantee that the page does
459 * not vanish from under us.
461 static void get_map(struct kmem_cache
*s
, struct page
*page
, unsigned long *map
)
464 void *addr
= page_address(page
);
466 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
467 set_bit(slab_index(p
, s
, addr
), map
);
470 static inline int size_from_object(struct kmem_cache
*s
)
472 if (s
->flags
& SLAB_RED_ZONE
)
473 return s
->size
- s
->red_left_pad
;
478 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
480 if (s
->flags
& SLAB_RED_ZONE
)
481 p
-= s
->red_left_pad
;
489 #if defined(CONFIG_SLUB_DEBUG_ON)
490 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
492 static slab_flags_t slub_debug
;
495 static char *slub_debug_slabs
;
496 static int disable_higher_order_debug
;
499 * slub is about to manipulate internal object metadata. This memory lies
500 * outside the range of the allocated object, so accessing it would normally
501 * be reported by kasan as a bounds error. metadata_access_enable() is used
502 * to tell kasan that these accesses are OK.
504 static inline void metadata_access_enable(void)
506 kasan_disable_current();
509 static inline void metadata_access_disable(void)
511 kasan_enable_current();
518 /* Verify that a pointer has an address that is valid within a slab page */
519 static inline int check_valid_pointer(struct kmem_cache
*s
,
520 struct page
*page
, void *object
)
527 base
= page_address(page
);
528 object
= restore_red_left(s
, object
);
529 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
530 (object
- base
) % s
->size
) {
537 static void print_section(char *level
, char *text
, u8
*addr
,
540 metadata_access_enable();
541 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
, 16, 1, addr
,
543 metadata_access_disable();
546 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
547 enum track_item alloc
)
552 p
= object
+ s
->offset
+ sizeof(void *);
554 p
= object
+ s
->inuse
;
559 static void set_track(struct kmem_cache
*s
, void *object
,
560 enum track_item alloc
, unsigned long addr
)
562 struct track
*p
= get_track(s
, object
, alloc
);
565 #ifdef CONFIG_STACKTRACE
566 struct stack_trace trace
;
569 trace
.nr_entries
= 0;
570 trace
.max_entries
= TRACK_ADDRS_COUNT
;
571 trace
.entries
= p
->addrs
;
573 metadata_access_enable();
574 save_stack_trace(&trace
);
575 metadata_access_disable();
577 /* See rant in lockdep.c */
578 if (trace
.nr_entries
!= 0 &&
579 trace
.entries
[trace
.nr_entries
- 1] == ULONG_MAX
)
582 for (i
= trace
.nr_entries
; i
< TRACK_ADDRS_COUNT
; i
++)
586 p
->cpu
= smp_processor_id();
587 p
->pid
= current
->pid
;
590 memset(p
, 0, sizeof(struct track
));
593 static void init_tracking(struct kmem_cache
*s
, void *object
)
595 if (!(s
->flags
& SLAB_STORE_USER
))
598 set_track(s
, object
, TRACK_FREE
, 0UL);
599 set_track(s
, object
, TRACK_ALLOC
, 0UL);
602 static void print_track(const char *s
, struct track
*t
)
607 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
608 s
, (void *)t
->addr
, jiffies
- t
->when
, t
->cpu
, t
->pid
);
609 #ifdef CONFIG_STACKTRACE
612 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
614 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
621 static void print_tracking(struct kmem_cache
*s
, void *object
)
623 if (!(s
->flags
& SLAB_STORE_USER
))
626 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
));
627 print_track("Freed", get_track(s
, object
, TRACK_FREE
));
630 static void print_page_info(struct page
*page
)
632 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
633 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
637 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
639 struct va_format vaf
;
645 pr_err("=============================================================================\n");
646 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
647 pr_err("-----------------------------------------------------------------------------\n\n");
649 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
653 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
655 struct va_format vaf
;
661 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
665 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
667 unsigned int off
; /* Offset of last byte */
668 u8
*addr
= page_address(page
);
670 print_tracking(s
, p
);
672 print_page_info(page
);
674 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
675 p
, p
- addr
, get_freepointer(s
, p
));
677 if (s
->flags
& SLAB_RED_ZONE
)
678 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
680 else if (p
> addr
+ 16)
681 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
683 print_section(KERN_ERR
, "Object ", p
,
684 min_t(unsigned long, s
->object_size
, PAGE_SIZE
));
685 if (s
->flags
& SLAB_RED_ZONE
)
686 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
687 s
->inuse
- s
->object_size
);
690 off
= s
->offset
+ sizeof(void *);
694 if (s
->flags
& SLAB_STORE_USER
)
695 off
+= 2 * sizeof(struct track
);
697 off
+= kasan_metadata_size(s
);
699 if (off
!= size_from_object(s
))
700 /* Beginning of the filler is the free pointer */
701 print_section(KERN_ERR
, "Padding ", p
+ off
,
702 size_from_object(s
) - off
);
707 void object_err(struct kmem_cache
*s
, struct page
*page
,
708 u8
*object
, char *reason
)
710 slab_bug(s
, "%s", reason
);
711 print_trailer(s
, page
, object
);
714 static void slab_err(struct kmem_cache
*s
, struct page
*page
,
715 const char *fmt
, ...)
721 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
723 slab_bug(s
, "%s", buf
);
724 print_page_info(page
);
728 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
732 if (s
->flags
& SLAB_RED_ZONE
)
733 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
735 if (s
->flags
& __OBJECT_POISON
) {
736 memset(p
, POISON_FREE
, s
->object_size
- 1);
737 p
[s
->object_size
- 1] = POISON_END
;
740 if (s
->flags
& SLAB_RED_ZONE
)
741 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
744 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
745 void *from
, void *to
)
747 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
748 memset(from
, data
, to
- from
);
751 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
752 u8
*object
, char *what
,
753 u8
*start
, unsigned int value
, unsigned int bytes
)
758 metadata_access_enable();
759 fault
= memchr_inv(start
, value
, bytes
);
760 metadata_access_disable();
765 while (end
> fault
&& end
[-1] == value
)
768 slab_bug(s
, "%s overwritten", what
);
769 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
770 fault
, end
- 1, fault
[0], value
);
771 print_trailer(s
, page
, object
);
773 restore_bytes(s
, what
, value
, fault
, end
);
781 * Bytes of the object to be managed.
782 * If the freepointer may overlay the object then the free
783 * pointer is the first word of the object.
785 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
788 * object + s->object_size
789 * Padding to reach word boundary. This is also used for Redzoning.
790 * Padding is extended by another word if Redzoning is enabled and
791 * object_size == inuse.
793 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
794 * 0xcc (RED_ACTIVE) for objects in use.
797 * Meta data starts here.
799 * A. Free pointer (if we cannot overwrite object on free)
800 * B. Tracking data for SLAB_STORE_USER
801 * C. Padding to reach required alignment boundary or at mininum
802 * one word if debugging is on to be able to detect writes
803 * before the word boundary.
805 * Padding is done using 0x5a (POISON_INUSE)
808 * Nothing is used beyond s->size.
810 * If slabcaches are merged then the object_size and inuse boundaries are mostly
811 * ignored. And therefore no slab options that rely on these boundaries
812 * may be used with merged slabcaches.
815 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
817 unsigned long off
= s
->inuse
; /* The end of info */
820 /* Freepointer is placed after the object. */
821 off
+= sizeof(void *);
823 if (s
->flags
& SLAB_STORE_USER
)
824 /* We also have user information there */
825 off
+= 2 * sizeof(struct track
);
827 off
+= kasan_metadata_size(s
);
829 if (size_from_object(s
) == off
)
832 return check_bytes_and_report(s
, page
, p
, "Object padding",
833 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
836 /* Check the pad bytes at the end of a slab page */
837 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
845 if (!(s
->flags
& SLAB_POISON
))
848 start
= page_address(page
);
849 length
= (PAGE_SIZE
<< compound_order(page
)) - s
->reserved
;
850 end
= start
+ length
;
851 remainder
= length
% s
->size
;
855 metadata_access_enable();
856 fault
= memchr_inv(end
- remainder
, POISON_INUSE
, remainder
);
857 metadata_access_disable();
860 while (end
> fault
&& end
[-1] == POISON_INUSE
)
863 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p", fault
, end
- 1);
864 print_section(KERN_ERR
, "Padding ", end
- remainder
, remainder
);
866 restore_bytes(s
, "slab padding", POISON_INUSE
, end
- remainder
, end
);
870 static int check_object(struct kmem_cache
*s
, struct page
*page
,
871 void *object
, u8 val
)
874 u8
*endobject
= object
+ s
->object_size
;
876 if (s
->flags
& SLAB_RED_ZONE
) {
877 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
878 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
881 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
882 endobject
, val
, s
->inuse
- s
->object_size
))
885 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
886 check_bytes_and_report(s
, page
, p
, "Alignment padding",
887 endobject
, POISON_INUSE
,
888 s
->inuse
- s
->object_size
);
892 if (s
->flags
& SLAB_POISON
) {
893 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
894 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
895 POISON_FREE
, s
->object_size
- 1) ||
896 !check_bytes_and_report(s
, page
, p
, "Poison",
897 p
+ s
->object_size
- 1, POISON_END
, 1)))
900 * check_pad_bytes cleans up on its own.
902 check_pad_bytes(s
, page
, p
);
905 if (!s
->offset
&& val
== SLUB_RED_ACTIVE
)
907 * Object and freepointer overlap. Cannot check
908 * freepointer while object is allocated.
912 /* Check free pointer validity */
913 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
914 object_err(s
, page
, p
, "Freepointer corrupt");
916 * No choice but to zap it and thus lose the remainder
917 * of the free objects in this slab. May cause
918 * another error because the object count is now wrong.
920 set_freepointer(s
, p
, NULL
);
926 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
930 VM_BUG_ON(!irqs_disabled());
932 if (!PageSlab(page
)) {
933 slab_err(s
, page
, "Not a valid slab page");
937 maxobj
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
938 if (page
->objects
> maxobj
) {
939 slab_err(s
, page
, "objects %u > max %u",
940 page
->objects
, maxobj
);
943 if (page
->inuse
> page
->objects
) {
944 slab_err(s
, page
, "inuse %u > max %u",
945 page
->inuse
, page
->objects
);
948 /* Slab_pad_check fixes things up after itself */
949 slab_pad_check(s
, page
);
954 * Determine if a certain object on a page is on the freelist. Must hold the
955 * slab lock to guarantee that the chains are in a consistent state.
957 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
965 while (fp
&& nr
<= page
->objects
) {
968 if (!check_valid_pointer(s
, page
, fp
)) {
970 object_err(s
, page
, object
,
971 "Freechain corrupt");
972 set_freepointer(s
, object
, NULL
);
974 slab_err(s
, page
, "Freepointer corrupt");
975 page
->freelist
= NULL
;
976 page
->inuse
= page
->objects
;
977 slab_fix(s
, "Freelist cleared");
983 fp
= get_freepointer(s
, object
);
987 max_objects
= order_objects(compound_order(page
), s
->size
, s
->reserved
);
988 if (max_objects
> MAX_OBJS_PER_PAGE
)
989 max_objects
= MAX_OBJS_PER_PAGE
;
991 if (page
->objects
!= max_objects
) {
992 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
993 page
->objects
, max_objects
);
994 page
->objects
= max_objects
;
995 slab_fix(s
, "Number of objects adjusted.");
997 if (page
->inuse
!= page
->objects
- nr
) {
998 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
999 page
->inuse
, page
->objects
- nr
);
1000 page
->inuse
= page
->objects
- nr
;
1001 slab_fix(s
, "Object count adjusted.");
1003 return search
== NULL
;
1006 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1009 if (s
->flags
& SLAB_TRACE
) {
1010 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1012 alloc
? "alloc" : "free",
1013 object
, page
->inuse
,
1017 print_section(KERN_INFO
, "Object ", (void *)object
,
1025 * Tracking of fully allocated slabs for debugging purposes.
1027 static void add_full(struct kmem_cache
*s
,
1028 struct kmem_cache_node
*n
, struct page
*page
)
1030 if (!(s
->flags
& SLAB_STORE_USER
))
1033 lockdep_assert_held(&n
->list_lock
);
1034 list_add(&page
->lru
, &n
->full
);
1037 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1039 if (!(s
->flags
& SLAB_STORE_USER
))
1042 lockdep_assert_held(&n
->list_lock
);
1043 list_del(&page
->lru
);
1046 /* Tracking of the number of slabs for debugging purposes */
1047 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1049 struct kmem_cache_node
*n
= get_node(s
, node
);
1051 return atomic_long_read(&n
->nr_slabs
);
1054 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1056 return atomic_long_read(&n
->nr_slabs
);
1059 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1061 struct kmem_cache_node
*n
= get_node(s
, node
);
1064 * May be called early in order to allocate a slab for the
1065 * kmem_cache_node structure. Solve the chicken-egg
1066 * dilemma by deferring the increment of the count during
1067 * bootstrap (see early_kmem_cache_node_alloc).
1070 atomic_long_inc(&n
->nr_slabs
);
1071 atomic_long_add(objects
, &n
->total_objects
);
1074 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1076 struct kmem_cache_node
*n
= get_node(s
, node
);
1078 atomic_long_dec(&n
->nr_slabs
);
1079 atomic_long_sub(objects
, &n
->total_objects
);
1082 /* Object debug checks for alloc/free paths */
1083 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1086 if (!(s
->flags
& (SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
)))
1089 init_object(s
, object
, SLUB_RED_INACTIVE
);
1090 init_tracking(s
, object
);
1093 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1095 void *object
, unsigned long addr
)
1097 if (!check_slab(s
, page
))
1100 if (!check_valid_pointer(s
, page
, object
)) {
1101 object_err(s
, page
, object
, "Freelist Pointer check fails");
1105 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1111 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1113 void *object
, unsigned long addr
)
1115 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1116 if (!alloc_consistency_checks(s
, page
, object
, addr
))
1120 /* Success perform special debug activities for allocs */
1121 if (s
->flags
& SLAB_STORE_USER
)
1122 set_track(s
, object
, TRACK_ALLOC
, addr
);
1123 trace(s
, page
, object
, 1);
1124 init_object(s
, object
, SLUB_RED_ACTIVE
);
1128 if (PageSlab(page
)) {
1130 * If this is a slab page then lets do the best we can
1131 * to avoid issues in the future. Marking all objects
1132 * as used avoids touching the remaining objects.
1134 slab_fix(s
, "Marking all objects used");
1135 page
->inuse
= page
->objects
;
1136 page
->freelist
= NULL
;
1141 static inline int free_consistency_checks(struct kmem_cache
*s
,
1142 struct page
*page
, void *object
, unsigned long addr
)
1144 if (!check_valid_pointer(s
, page
, object
)) {
1145 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1149 if (on_freelist(s
, page
, object
)) {
1150 object_err(s
, page
, object
, "Object already free");
1154 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1157 if (unlikely(s
!= page
->slab_cache
)) {
1158 if (!PageSlab(page
)) {
1159 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1161 } else if (!page
->slab_cache
) {
1162 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1166 object_err(s
, page
, object
,
1167 "page slab pointer corrupt.");
1173 /* Supports checking bulk free of a constructed freelist */
1174 static noinline
int free_debug_processing(
1175 struct kmem_cache
*s
, struct page
*page
,
1176 void *head
, void *tail
, int bulk_cnt
,
1179 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1180 void *object
= head
;
1182 unsigned long uninitialized_var(flags
);
1185 spin_lock_irqsave(&n
->list_lock
, flags
);
1188 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1189 if (!check_slab(s
, page
))
1196 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1197 if (!free_consistency_checks(s
, page
, object
, addr
))
1201 if (s
->flags
& SLAB_STORE_USER
)
1202 set_track(s
, object
, TRACK_FREE
, addr
);
1203 trace(s
, page
, object
, 0);
1204 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1205 init_object(s
, object
, SLUB_RED_INACTIVE
);
1207 /* Reached end of constructed freelist yet? */
1208 if (object
!= tail
) {
1209 object
= get_freepointer(s
, object
);
1215 if (cnt
!= bulk_cnt
)
1216 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1220 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1222 slab_fix(s
, "Object at 0x%p not freed", object
);
1226 static int __init
setup_slub_debug(char *str
)
1228 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1229 if (*str
++ != '=' || !*str
)
1231 * No options specified. Switch on full debugging.
1237 * No options but restriction on slabs. This means full
1238 * debugging for slabs matching a pattern.
1245 * Switch off all debugging measures.
1250 * Determine which debug features should be switched on
1252 for (; *str
&& *str
!= ','; str
++) {
1253 switch (tolower(*str
)) {
1255 slub_debug
|= SLAB_CONSISTENCY_CHECKS
;
1258 slub_debug
|= SLAB_RED_ZONE
;
1261 slub_debug
|= SLAB_POISON
;
1264 slub_debug
|= SLAB_STORE_USER
;
1267 slub_debug
|= SLAB_TRACE
;
1270 slub_debug
|= SLAB_FAILSLAB
;
1274 * Avoid enabling debugging on caches if its minimum
1275 * order would increase as a result.
1277 disable_higher_order_debug
= 1;
1280 pr_err("slub_debug option '%c' unknown. skipped\n",
1287 slub_debug_slabs
= str
+ 1;
1292 __setup("slub_debug", setup_slub_debug
);
1294 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1295 slab_flags_t flags
, const char *name
,
1296 void (*ctor
)(void *))
1299 * Enable debugging if selected on the kernel commandline.
1301 if (slub_debug
&& (!slub_debug_slabs
|| (name
&&
1302 !strncmp(slub_debug_slabs
, name
, strlen(slub_debug_slabs
)))))
1303 flags
|= slub_debug
;
1307 #else /* !CONFIG_SLUB_DEBUG */
1308 static inline void setup_object_debug(struct kmem_cache
*s
,
1309 struct page
*page
, void *object
) {}
1311 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1312 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1314 static inline int free_debug_processing(
1315 struct kmem_cache
*s
, struct page
*page
,
1316 void *head
, void *tail
, int bulk_cnt
,
1317 unsigned long addr
) { return 0; }
1319 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1321 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1322 void *object
, u8 val
) { return 1; }
1323 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1324 struct page
*page
) {}
1325 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1326 struct page
*page
) {}
1327 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1328 slab_flags_t flags
, const char *name
,
1329 void (*ctor
)(void *))
1333 #define slub_debug 0
1335 #define disable_higher_order_debug 0
1337 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1339 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1341 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1343 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1346 #endif /* CONFIG_SLUB_DEBUG */
1349 * Hooks for other subsystems that check memory allocations. In a typical
1350 * production configuration these hooks all should produce no code at all.
1352 static inline void kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1354 kmemleak_alloc(ptr
, size
, 1, flags
);
1355 kasan_kmalloc_large(ptr
, size
, flags
);
1358 static inline void kfree_hook(const void *x
)
1361 kasan_kfree_large(x
);
1364 static inline void *slab_free_hook(struct kmem_cache
*s
, void *x
)
1368 kmemleak_free_recursive(x
, s
->flags
);
1371 * Trouble is that we may no longer disable interrupts in the fast path
1372 * So in order to make the debug calls that expect irqs to be
1373 * disabled we need to disable interrupts temporarily.
1375 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1377 unsigned long flags
;
1379 local_irq_save(flags
);
1380 kmemcheck_slab_free(s
, x
, s
->object_size
);
1381 debug_check_no_locks_freed(x
, s
->object_size
);
1382 local_irq_restore(flags
);
1385 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1386 debug_check_no_obj_freed(x
, s
->object_size
);
1388 freeptr
= get_freepointer(s
, x
);
1390 * kasan_slab_free() may put x into memory quarantine, delaying its
1391 * reuse. In this case the object's freelist pointer is changed.
1393 kasan_slab_free(s
, x
);
1397 static inline void slab_free_freelist_hook(struct kmem_cache
*s
,
1398 void *head
, void *tail
)
1401 * Compiler cannot detect this function can be removed if slab_free_hook()
1402 * evaluates to nothing. Thus, catch all relevant config debug options here.
1404 #if defined(CONFIG_KMEMCHECK) || \
1405 defined(CONFIG_LOCKDEP) || \
1406 defined(CONFIG_DEBUG_KMEMLEAK) || \
1407 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1408 defined(CONFIG_KASAN)
1410 void *object
= head
;
1411 void *tail_obj
= tail
? : head
;
1415 freeptr
= slab_free_hook(s
, object
);
1416 } while ((object
!= tail_obj
) && (object
= freeptr
));
1420 static void setup_object(struct kmem_cache
*s
, struct page
*page
,
1423 setup_object_debug(s
, page
, object
);
1424 kasan_init_slab_obj(s
, object
);
1425 if (unlikely(s
->ctor
)) {
1426 kasan_unpoison_object_data(s
, object
);
1428 kasan_poison_object_data(s
, object
);
1433 * Slab allocation and freeing
1435 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1436 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1439 int order
= oo_order(oo
);
1441 flags
|= __GFP_NOTRACK
;
1443 if (node
== NUMA_NO_NODE
)
1444 page
= alloc_pages(flags
, order
);
1446 page
= __alloc_pages_node(node
, flags
, order
);
1448 if (page
&& memcg_charge_slab(page
, flags
, order
, s
)) {
1449 __free_pages(page
, order
);
1456 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1457 /* Pre-initialize the random sequence cache */
1458 static int init_cache_random_seq(struct kmem_cache
*s
)
1461 unsigned long i
, count
= oo_objects(s
->oo
);
1463 /* Bailout if already initialised */
1467 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1469 pr_err("SLUB: Unable to initialize free list for %s\n",
1474 /* Transform to an offset on the set of pages */
1475 if (s
->random_seq
) {
1476 for (i
= 0; i
< count
; i
++)
1477 s
->random_seq
[i
] *= s
->size
;
1482 /* Initialize each random sequence freelist per cache */
1483 static void __init
init_freelist_randomization(void)
1485 struct kmem_cache
*s
;
1487 mutex_lock(&slab_mutex
);
1489 list_for_each_entry(s
, &slab_caches
, list
)
1490 init_cache_random_seq(s
);
1492 mutex_unlock(&slab_mutex
);
1495 /* Get the next entry on the pre-computed freelist randomized */
1496 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1497 unsigned long *pos
, void *start
,
1498 unsigned long page_limit
,
1499 unsigned long freelist_count
)
1504 * If the target page allocation failed, the number of objects on the
1505 * page might be smaller than the usual size defined by the cache.
1508 idx
= s
->random_seq
[*pos
];
1510 if (*pos
>= freelist_count
)
1512 } while (unlikely(idx
>= page_limit
));
1514 return (char *)start
+ idx
;
1517 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1518 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1523 unsigned long idx
, pos
, page_limit
, freelist_count
;
1525 if (page
->objects
< 2 || !s
->random_seq
)
1528 freelist_count
= oo_objects(s
->oo
);
1529 pos
= get_random_int() % freelist_count
;
1531 page_limit
= page
->objects
* s
->size
;
1532 start
= fixup_red_left(s
, page_address(page
));
1534 /* First entry is used as the base of the freelist */
1535 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1537 page
->freelist
= cur
;
1539 for (idx
= 1; idx
< page
->objects
; idx
++) {
1540 setup_object(s
, page
, cur
);
1541 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1543 set_freepointer(s
, cur
, next
);
1546 setup_object(s
, page
, cur
);
1547 set_freepointer(s
, cur
, NULL
);
1552 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1556 static inline void init_freelist_randomization(void) { }
1557 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1561 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1563 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1566 struct kmem_cache_order_objects oo
= s
->oo
;
1572 flags
&= gfp_allowed_mask
;
1574 if (gfpflags_allow_blocking(flags
))
1577 flags
|= s
->allocflags
;
1580 * Let the initial higher-order allocation fail under memory pressure
1581 * so we fall-back to the minimum order allocation.
1583 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1584 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1585 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1587 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1588 if (unlikely(!page
)) {
1592 * Allocation may have failed due to fragmentation.
1593 * Try a lower order alloc if possible
1595 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1596 if (unlikely(!page
))
1598 stat(s
, ORDER_FALLBACK
);
1601 if (kmemcheck_enabled
&&
1602 !(s
->flags
& (SLAB_NOTRACK
| DEBUG_DEFAULT_FLAGS
))) {
1603 int pages
= 1 << oo_order(oo
);
1605 kmemcheck_alloc_shadow(page
, oo_order(oo
), alloc_gfp
, node
);
1608 * Objects from caches that have a constructor don't get
1609 * cleared when they're allocated, so we need to do it here.
1612 kmemcheck_mark_uninitialized_pages(page
, pages
);
1614 kmemcheck_mark_unallocated_pages(page
, pages
);
1617 page
->objects
= oo_objects(oo
);
1619 order
= compound_order(page
);
1620 page
->slab_cache
= s
;
1621 __SetPageSlab(page
);
1622 if (page_is_pfmemalloc(page
))
1623 SetPageSlabPfmemalloc(page
);
1625 start
= page_address(page
);
1627 if (unlikely(s
->flags
& SLAB_POISON
))
1628 memset(start
, POISON_INUSE
, PAGE_SIZE
<< order
);
1630 kasan_poison_slab(page
);
1632 shuffle
= shuffle_freelist(s
, page
);
1635 for_each_object_idx(p
, idx
, s
, start
, page
->objects
) {
1636 setup_object(s
, page
, p
);
1637 if (likely(idx
< page
->objects
))
1638 set_freepointer(s
, p
, p
+ s
->size
);
1640 set_freepointer(s
, p
, NULL
);
1642 page
->freelist
= fixup_red_left(s
, start
);
1645 page
->inuse
= page
->objects
;
1649 if (gfpflags_allow_blocking(flags
))
1650 local_irq_disable();
1654 mod_lruvec_page_state(page
,
1655 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1656 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1659 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1664 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1666 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
1667 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
1668 flags
&= ~GFP_SLAB_BUG_MASK
;
1669 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1670 invalid_mask
, &invalid_mask
, flags
, &flags
);
1674 return allocate_slab(s
,
1675 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1678 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1680 int order
= compound_order(page
);
1681 int pages
= 1 << order
;
1683 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1686 slab_pad_check(s
, page
);
1687 for_each_object(p
, s
, page_address(page
),
1689 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1692 kmemcheck_free_shadow(page
, compound_order(page
));
1694 mod_lruvec_page_state(page
,
1695 (s
->flags
& SLAB_RECLAIM_ACCOUNT
) ?
1696 NR_SLAB_RECLAIMABLE
: NR_SLAB_UNRECLAIMABLE
,
1699 __ClearPageSlabPfmemalloc(page
);
1700 __ClearPageSlab(page
);
1702 page_mapcount_reset(page
);
1703 if (current
->reclaim_state
)
1704 current
->reclaim_state
->reclaimed_slab
+= pages
;
1705 memcg_uncharge_slab(page
, order
, s
);
1706 __free_pages(page
, order
);
1709 #define need_reserve_slab_rcu \
1710 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1712 static void rcu_free_slab(struct rcu_head
*h
)
1716 if (need_reserve_slab_rcu
)
1717 page
= virt_to_head_page(h
);
1719 page
= container_of((struct list_head
*)h
, struct page
, lru
);
1721 __free_slab(page
->slab_cache
, page
);
1724 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1726 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1727 struct rcu_head
*head
;
1729 if (need_reserve_slab_rcu
) {
1730 int order
= compound_order(page
);
1731 int offset
= (PAGE_SIZE
<< order
) - s
->reserved
;
1733 VM_BUG_ON(s
->reserved
!= sizeof(*head
));
1734 head
= page_address(page
) + offset
;
1736 head
= &page
->rcu_head
;
1739 call_rcu(head
, rcu_free_slab
);
1741 __free_slab(s
, page
);
1744 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1746 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1751 * Management of partially allocated slabs.
1754 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1757 if (tail
== DEACTIVATE_TO_TAIL
)
1758 list_add_tail(&page
->lru
, &n
->partial
);
1760 list_add(&page
->lru
, &n
->partial
);
1763 static inline void add_partial(struct kmem_cache_node
*n
,
1764 struct page
*page
, int tail
)
1766 lockdep_assert_held(&n
->list_lock
);
1767 __add_partial(n
, page
, tail
);
1770 static inline void remove_partial(struct kmem_cache_node
*n
,
1773 lockdep_assert_held(&n
->list_lock
);
1774 list_del(&page
->lru
);
1779 * Remove slab from the partial list, freeze it and
1780 * return the pointer to the freelist.
1782 * Returns a list of objects or NULL if it fails.
1784 static inline void *acquire_slab(struct kmem_cache
*s
,
1785 struct kmem_cache_node
*n
, struct page
*page
,
1786 int mode
, int *objects
)
1789 unsigned long counters
;
1792 lockdep_assert_held(&n
->list_lock
);
1795 * Zap the freelist and set the frozen bit.
1796 * The old freelist is the list of objects for the
1797 * per cpu allocation list.
1799 freelist
= page
->freelist
;
1800 counters
= page
->counters
;
1801 new.counters
= counters
;
1802 *objects
= new.objects
- new.inuse
;
1804 new.inuse
= page
->objects
;
1805 new.freelist
= NULL
;
1807 new.freelist
= freelist
;
1810 VM_BUG_ON(new.frozen
);
1813 if (!__cmpxchg_double_slab(s
, page
,
1815 new.freelist
, new.counters
,
1819 remove_partial(n
, page
);
1824 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1825 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1828 * Try to allocate a partial slab from a specific node.
1830 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1831 struct kmem_cache_cpu
*c
, gfp_t flags
)
1833 struct page
*page
, *page2
;
1834 void *object
= NULL
;
1839 * Racy check. If we mistakenly see no partial slabs then we
1840 * just allocate an empty slab. If we mistakenly try to get a
1841 * partial slab and there is none available then get_partials()
1844 if (!n
|| !n
->nr_partial
)
1847 spin_lock(&n
->list_lock
);
1848 list_for_each_entry_safe(page
, page2
, &n
->partial
, lru
) {
1851 if (!pfmemalloc_match(page
, flags
))
1854 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1858 available
+= objects
;
1861 stat(s
, ALLOC_FROM_PARTIAL
);
1864 put_cpu_partial(s
, page
, 0);
1865 stat(s
, CPU_PARTIAL_NODE
);
1867 if (!kmem_cache_has_cpu_partial(s
)
1868 || available
> slub_cpu_partial(s
) / 2)
1872 spin_unlock(&n
->list_lock
);
1877 * Get a page from somewhere. Search in increasing NUMA distances.
1879 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
1880 struct kmem_cache_cpu
*c
)
1883 struct zonelist
*zonelist
;
1886 enum zone_type high_zoneidx
= gfp_zone(flags
);
1888 unsigned int cpuset_mems_cookie
;
1891 * The defrag ratio allows a configuration of the tradeoffs between
1892 * inter node defragmentation and node local allocations. A lower
1893 * defrag_ratio increases the tendency to do local allocations
1894 * instead of attempting to obtain partial slabs from other nodes.
1896 * If the defrag_ratio is set to 0 then kmalloc() always
1897 * returns node local objects. If the ratio is higher then kmalloc()
1898 * may return off node objects because partial slabs are obtained
1899 * from other nodes and filled up.
1901 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1902 * (which makes defrag_ratio = 1000) then every (well almost)
1903 * allocation will first attempt to defrag slab caches on other nodes.
1904 * This means scanning over all nodes to look for partial slabs which
1905 * may be expensive if we do it every time we are trying to find a slab
1906 * with available objects.
1908 if (!s
->remote_node_defrag_ratio
||
1909 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
1913 cpuset_mems_cookie
= read_mems_allowed_begin();
1914 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
1915 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
1916 struct kmem_cache_node
*n
;
1918 n
= get_node(s
, zone_to_nid(zone
));
1920 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
1921 n
->nr_partial
> s
->min_partial
) {
1922 object
= get_partial_node(s
, n
, c
, flags
);
1925 * Don't check read_mems_allowed_retry()
1926 * here - if mems_allowed was updated in
1927 * parallel, that was a harmless race
1928 * between allocation and the cpuset
1935 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1941 * Get a partial page, lock it and return it.
1943 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
1944 struct kmem_cache_cpu
*c
)
1947 int searchnode
= node
;
1949 if (node
== NUMA_NO_NODE
)
1950 searchnode
= numa_mem_id();
1951 else if (!node_present_pages(node
))
1952 searchnode
= node_to_mem_node(node
);
1954 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
1955 if (object
|| node
!= NUMA_NO_NODE
)
1958 return get_any_partial(s
, flags
, c
);
1961 #ifdef CONFIG_PREEMPT
1963 * Calculate the next globally unique transaction for disambiguiation
1964 * during cmpxchg. The transactions start with the cpu number and are then
1965 * incremented by CONFIG_NR_CPUS.
1967 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1970 * No preemption supported therefore also no need to check for
1976 static inline unsigned long next_tid(unsigned long tid
)
1978 return tid
+ TID_STEP
;
1981 static inline unsigned int tid_to_cpu(unsigned long tid
)
1983 return tid
% TID_STEP
;
1986 static inline unsigned long tid_to_event(unsigned long tid
)
1988 return tid
/ TID_STEP
;
1991 static inline unsigned int init_tid(int cpu
)
1996 static inline void note_cmpxchg_failure(const char *n
,
1997 const struct kmem_cache
*s
, unsigned long tid
)
1999 #ifdef SLUB_DEBUG_CMPXCHG
2000 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2002 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2004 #ifdef CONFIG_PREEMPT
2005 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2006 pr_warn("due to cpu change %d -> %d\n",
2007 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2010 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2011 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2012 tid_to_event(tid
), tid_to_event(actual_tid
));
2014 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2015 actual_tid
, tid
, next_tid(tid
));
2017 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2020 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2024 for_each_possible_cpu(cpu
)
2025 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2029 * Remove the cpu slab
2031 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2032 void *freelist
, struct kmem_cache_cpu
*c
)
2034 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2035 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2037 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2039 int tail
= DEACTIVATE_TO_HEAD
;
2043 if (page
->freelist
) {
2044 stat(s
, DEACTIVATE_REMOTE_FREES
);
2045 tail
= DEACTIVATE_TO_TAIL
;
2049 * Stage one: Free all available per cpu objects back
2050 * to the page freelist while it is still frozen. Leave the
2053 * There is no need to take the list->lock because the page
2056 while (freelist
&& (nextfree
= get_freepointer(s
, freelist
))) {
2058 unsigned long counters
;
2061 prior
= page
->freelist
;
2062 counters
= page
->counters
;
2063 set_freepointer(s
, freelist
, prior
);
2064 new.counters
= counters
;
2066 VM_BUG_ON(!new.frozen
);
2068 } while (!__cmpxchg_double_slab(s
, page
,
2070 freelist
, new.counters
,
2071 "drain percpu freelist"));
2073 freelist
= nextfree
;
2077 * Stage two: Ensure that the page is unfrozen while the
2078 * list presence reflects the actual number of objects
2081 * We setup the list membership and then perform a cmpxchg
2082 * with the count. If there is a mismatch then the page
2083 * is not unfrozen but the page is on the wrong list.
2085 * Then we restart the process which may have to remove
2086 * the page from the list that we just put it on again
2087 * because the number of objects in the slab may have
2092 old
.freelist
= page
->freelist
;
2093 old
.counters
= page
->counters
;
2094 VM_BUG_ON(!old
.frozen
);
2096 /* Determine target state of the slab */
2097 new.counters
= old
.counters
;
2100 set_freepointer(s
, freelist
, old
.freelist
);
2101 new.freelist
= freelist
;
2103 new.freelist
= old
.freelist
;
2107 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2109 else if (new.freelist
) {
2114 * Taking the spinlock removes the possiblity
2115 * that acquire_slab() will see a slab page that
2118 spin_lock(&n
->list_lock
);
2122 if (kmem_cache_debug(s
) && !lock
) {
2125 * This also ensures that the scanning of full
2126 * slabs from diagnostic functions will not see
2129 spin_lock(&n
->list_lock
);
2137 remove_partial(n
, page
);
2139 else if (l
== M_FULL
)
2141 remove_full(s
, n
, page
);
2143 if (m
== M_PARTIAL
) {
2145 add_partial(n
, page
, tail
);
2148 } else if (m
== M_FULL
) {
2150 stat(s
, DEACTIVATE_FULL
);
2151 add_full(s
, n
, page
);
2157 if (!__cmpxchg_double_slab(s
, page
,
2158 old
.freelist
, old
.counters
,
2159 new.freelist
, new.counters
,
2164 spin_unlock(&n
->list_lock
);
2167 stat(s
, DEACTIVATE_EMPTY
);
2168 discard_slab(s
, page
);
2177 * Unfreeze all the cpu partial slabs.
2179 * This function must be called with interrupts disabled
2180 * for the cpu using c (or some other guarantee must be there
2181 * to guarantee no concurrent accesses).
2183 static void unfreeze_partials(struct kmem_cache
*s
,
2184 struct kmem_cache_cpu
*c
)
2186 #ifdef CONFIG_SLUB_CPU_PARTIAL
2187 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2188 struct page
*page
, *discard_page
= NULL
;
2190 while ((page
= c
->partial
)) {
2194 c
->partial
= page
->next
;
2196 n2
= get_node(s
, page_to_nid(page
));
2199 spin_unlock(&n
->list_lock
);
2202 spin_lock(&n
->list_lock
);
2207 old
.freelist
= page
->freelist
;
2208 old
.counters
= page
->counters
;
2209 VM_BUG_ON(!old
.frozen
);
2211 new.counters
= old
.counters
;
2212 new.freelist
= old
.freelist
;
2216 } while (!__cmpxchg_double_slab(s
, page
,
2217 old
.freelist
, old
.counters
,
2218 new.freelist
, new.counters
,
2219 "unfreezing slab"));
2221 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2222 page
->next
= discard_page
;
2223 discard_page
= page
;
2225 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2226 stat(s
, FREE_ADD_PARTIAL
);
2231 spin_unlock(&n
->list_lock
);
2233 while (discard_page
) {
2234 page
= discard_page
;
2235 discard_page
= discard_page
->next
;
2237 stat(s
, DEACTIVATE_EMPTY
);
2238 discard_slab(s
, page
);
2245 * Put a page that was just frozen (in __slab_free) into a partial page
2246 * slot if available. This is done without interrupts disabled and without
2247 * preemption disabled. The cmpxchg is racy and may put the partial page
2248 * onto a random cpus partial slot.
2250 * If we did not find a slot then simply move all the partials to the
2251 * per node partial list.
2253 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2255 #ifdef CONFIG_SLUB_CPU_PARTIAL
2256 struct page
*oldpage
;
2264 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2267 pobjects
= oldpage
->pobjects
;
2268 pages
= oldpage
->pages
;
2269 if (drain
&& pobjects
> s
->cpu_partial
) {
2270 unsigned long flags
;
2272 * partial array is full. Move the existing
2273 * set to the per node partial list.
2275 local_irq_save(flags
);
2276 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2277 local_irq_restore(flags
);
2281 stat(s
, CPU_PARTIAL_DRAIN
);
2286 pobjects
+= page
->objects
- page
->inuse
;
2288 page
->pages
= pages
;
2289 page
->pobjects
= pobjects
;
2290 page
->next
= oldpage
;
2292 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2294 if (unlikely(!s
->cpu_partial
)) {
2295 unsigned long flags
;
2297 local_irq_save(flags
);
2298 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2299 local_irq_restore(flags
);
2305 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2307 stat(s
, CPUSLAB_FLUSH
);
2308 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2310 c
->tid
= next_tid(c
->tid
);
2316 * Called from IPI handler with interrupts disabled.
2318 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2320 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2326 unfreeze_partials(s
, c
);
2330 static void flush_cpu_slab(void *d
)
2332 struct kmem_cache
*s
= d
;
2334 __flush_cpu_slab(s
, smp_processor_id());
2337 static bool has_cpu_slab(int cpu
, void *info
)
2339 struct kmem_cache
*s
= info
;
2340 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2342 return c
->page
|| slub_percpu_partial(c
);
2345 static void flush_all(struct kmem_cache
*s
)
2347 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1, GFP_ATOMIC
);
2351 * Use the cpu notifier to insure that the cpu slabs are flushed when
2354 static int slub_cpu_dead(unsigned int cpu
)
2356 struct kmem_cache
*s
;
2357 unsigned long flags
;
2359 mutex_lock(&slab_mutex
);
2360 list_for_each_entry(s
, &slab_caches
, list
) {
2361 local_irq_save(flags
);
2362 __flush_cpu_slab(s
, cpu
);
2363 local_irq_restore(flags
);
2365 mutex_unlock(&slab_mutex
);
2370 * Check if the objects in a per cpu structure fit numa
2371 * locality expectations.
2373 static inline int node_match(struct page
*page
, int node
)
2376 if (!page
|| (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
))
2382 #ifdef CONFIG_SLUB_DEBUG
2383 static int count_free(struct page
*page
)
2385 return page
->objects
- page
->inuse
;
2388 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2390 return atomic_long_read(&n
->total_objects
);
2392 #endif /* CONFIG_SLUB_DEBUG */
2394 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2395 static unsigned long count_partial(struct kmem_cache_node
*n
,
2396 int (*get_count
)(struct page
*))
2398 unsigned long flags
;
2399 unsigned long x
= 0;
2402 spin_lock_irqsave(&n
->list_lock
, flags
);
2403 list_for_each_entry(page
, &n
->partial
, lru
)
2404 x
+= get_count(page
);
2405 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2408 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2410 static noinline
void
2411 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2413 #ifdef CONFIG_SLUB_DEBUG
2414 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2415 DEFAULT_RATELIMIT_BURST
);
2417 struct kmem_cache_node
*n
;
2419 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2422 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2423 nid
, gfpflags
, &gfpflags
);
2424 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2425 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2428 if (oo_order(s
->min
) > get_order(s
->object_size
))
2429 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2432 for_each_kmem_cache_node(s
, node
, n
) {
2433 unsigned long nr_slabs
;
2434 unsigned long nr_objs
;
2435 unsigned long nr_free
;
2437 nr_free
= count_partial(n
, count_free
);
2438 nr_slabs
= node_nr_slabs(n
);
2439 nr_objs
= node_nr_objs(n
);
2441 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2442 node
, nr_slabs
, nr_objs
, nr_free
);
2447 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2448 int node
, struct kmem_cache_cpu
**pc
)
2451 struct kmem_cache_cpu
*c
= *pc
;
2454 freelist
= get_partial(s
, flags
, node
, c
);
2459 page
= new_slab(s
, flags
, node
);
2461 c
= raw_cpu_ptr(s
->cpu_slab
);
2466 * No other reference to the page yet so we can
2467 * muck around with it freely without cmpxchg
2469 freelist
= page
->freelist
;
2470 page
->freelist
= NULL
;
2472 stat(s
, ALLOC_SLAB
);
2481 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2483 if (unlikely(PageSlabPfmemalloc(page
)))
2484 return gfp_pfmemalloc_allowed(gfpflags
);
2490 * Check the page->freelist of a page and either transfer the freelist to the
2491 * per cpu freelist or deactivate the page.
2493 * The page is still frozen if the return value is not NULL.
2495 * If this function returns NULL then the page has been unfrozen.
2497 * This function must be called with interrupt disabled.
2499 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2502 unsigned long counters
;
2506 freelist
= page
->freelist
;
2507 counters
= page
->counters
;
2509 new.counters
= counters
;
2510 VM_BUG_ON(!new.frozen
);
2512 new.inuse
= page
->objects
;
2513 new.frozen
= freelist
!= NULL
;
2515 } while (!__cmpxchg_double_slab(s
, page
,
2524 * Slow path. The lockless freelist is empty or we need to perform
2527 * Processing is still very fast if new objects have been freed to the
2528 * regular freelist. In that case we simply take over the regular freelist
2529 * as the lockless freelist and zap the regular freelist.
2531 * If that is not working then we fall back to the partial lists. We take the
2532 * first element of the freelist as the object to allocate now and move the
2533 * rest of the freelist to the lockless freelist.
2535 * And if we were unable to get a new slab from the partial slab lists then
2536 * we need to allocate a new slab. This is the slowest path since it involves
2537 * a call to the page allocator and the setup of a new slab.
2539 * Version of __slab_alloc to use when we know that interrupts are
2540 * already disabled (which is the case for bulk allocation).
2542 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2543 unsigned long addr
, struct kmem_cache_cpu
*c
)
2553 if (unlikely(!node_match(page
, node
))) {
2554 int searchnode
= node
;
2556 if (node
!= NUMA_NO_NODE
&& !node_present_pages(node
))
2557 searchnode
= node_to_mem_node(node
);
2559 if (unlikely(!node_match(page
, searchnode
))) {
2560 stat(s
, ALLOC_NODE_MISMATCH
);
2561 deactivate_slab(s
, page
, c
->freelist
, c
);
2567 * By rights, we should be searching for a slab page that was
2568 * PFMEMALLOC but right now, we are losing the pfmemalloc
2569 * information when the page leaves the per-cpu allocator
2571 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2572 deactivate_slab(s
, page
, c
->freelist
, c
);
2576 /* must check again c->freelist in case of cpu migration or IRQ */
2577 freelist
= c
->freelist
;
2581 freelist
= get_freelist(s
, page
);
2585 stat(s
, DEACTIVATE_BYPASS
);
2589 stat(s
, ALLOC_REFILL
);
2593 * freelist is pointing to the list of objects to be used.
2594 * page is pointing to the page from which the objects are obtained.
2595 * That page must be frozen for per cpu allocations to work.
2597 VM_BUG_ON(!c
->page
->frozen
);
2598 c
->freelist
= get_freepointer(s
, freelist
);
2599 c
->tid
= next_tid(c
->tid
);
2604 if (slub_percpu_partial(c
)) {
2605 page
= c
->page
= slub_percpu_partial(c
);
2606 slub_set_percpu_partial(c
, page
);
2607 stat(s
, CPU_PARTIAL_ALLOC
);
2611 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2613 if (unlikely(!freelist
)) {
2614 slab_out_of_memory(s
, gfpflags
, node
);
2619 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2622 /* Only entered in the debug case */
2623 if (kmem_cache_debug(s
) &&
2624 !alloc_debug_processing(s
, page
, freelist
, addr
))
2625 goto new_slab
; /* Slab failed checks. Next slab needed */
2627 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2632 * Another one that disabled interrupt and compensates for possible
2633 * cpu changes by refetching the per cpu area pointer.
2635 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2636 unsigned long addr
, struct kmem_cache_cpu
*c
)
2639 unsigned long flags
;
2641 local_irq_save(flags
);
2642 #ifdef CONFIG_PREEMPT
2644 * We may have been preempted and rescheduled on a different
2645 * cpu before disabling interrupts. Need to reload cpu area
2648 c
= this_cpu_ptr(s
->cpu_slab
);
2651 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2652 local_irq_restore(flags
);
2657 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2658 * have the fastpath folded into their functions. So no function call
2659 * overhead for requests that can be satisfied on the fastpath.
2661 * The fastpath works by first checking if the lockless freelist can be used.
2662 * If not then __slab_alloc is called for slow processing.
2664 * Otherwise we can simply pick the next object from the lockless free list.
2666 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2667 gfp_t gfpflags
, int node
, unsigned long addr
)
2670 struct kmem_cache_cpu
*c
;
2674 s
= slab_pre_alloc_hook(s
, gfpflags
);
2679 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2680 * enabled. We may switch back and forth between cpus while
2681 * reading from one cpu area. That does not matter as long
2682 * as we end up on the original cpu again when doing the cmpxchg.
2684 * We should guarantee that tid and kmem_cache are retrieved on
2685 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2686 * to check if it is matched or not.
2689 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2690 c
= raw_cpu_ptr(s
->cpu_slab
);
2691 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2692 unlikely(tid
!= READ_ONCE(c
->tid
)));
2695 * Irqless object alloc/free algorithm used here depends on sequence
2696 * of fetching cpu_slab's data. tid should be fetched before anything
2697 * on c to guarantee that object and page associated with previous tid
2698 * won't be used with current tid. If we fetch tid first, object and
2699 * page could be one associated with next tid and our alloc/free
2700 * request will be failed. In this case, we will retry. So, no problem.
2705 * The transaction ids are globally unique per cpu and per operation on
2706 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2707 * occurs on the right processor and that there was no operation on the
2708 * linked list in between.
2711 object
= c
->freelist
;
2713 if (unlikely(!object
|| !node_match(page
, node
))) {
2714 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2715 stat(s
, ALLOC_SLOWPATH
);
2717 void *next_object
= get_freepointer_safe(s
, object
);
2720 * The cmpxchg will only match if there was no additional
2721 * operation and if we are on the right processor.
2723 * The cmpxchg does the following atomically (without lock
2725 * 1. Relocate first pointer to the current per cpu area.
2726 * 2. Verify that tid and freelist have not been changed
2727 * 3. If they were not changed replace tid and freelist
2729 * Since this is without lock semantics the protection is only
2730 * against code executing on this cpu *not* from access by
2733 if (unlikely(!this_cpu_cmpxchg_double(
2734 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2736 next_object
, next_tid(tid
)))) {
2738 note_cmpxchg_failure("slab_alloc", s
, tid
);
2741 prefetch_freepointer(s
, next_object
);
2742 stat(s
, ALLOC_FASTPATH
);
2745 if (unlikely(gfpflags
& __GFP_ZERO
) && object
)
2746 memset(object
, 0, s
->object_size
);
2748 slab_post_alloc_hook(s
, gfpflags
, 1, &object
);
2753 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2754 gfp_t gfpflags
, unsigned long addr
)
2756 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
);
2759 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2761 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2763 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2768 EXPORT_SYMBOL(kmem_cache_alloc
);
2770 #ifdef CONFIG_TRACING
2771 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2773 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
);
2774 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2775 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2778 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2782 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2784 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2786 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2787 s
->object_size
, s
->size
, gfpflags
, node
);
2791 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2793 #ifdef CONFIG_TRACING
2794 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2796 int node
, size_t size
)
2798 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
);
2800 trace_kmalloc_node(_RET_IP_
, ret
,
2801 size
, s
->size
, gfpflags
, node
);
2803 kasan_kmalloc(s
, ret
, size
, gfpflags
);
2806 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2811 * Slow path handling. This may still be called frequently since objects
2812 * have a longer lifetime than the cpu slabs in most processing loads.
2814 * So we still attempt to reduce cache line usage. Just take the slab
2815 * lock and free the item. If there is no additional partial page
2816 * handling required then we can return immediately.
2818 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2819 void *head
, void *tail
, int cnt
,
2826 unsigned long counters
;
2827 struct kmem_cache_node
*n
= NULL
;
2828 unsigned long uninitialized_var(flags
);
2830 stat(s
, FREE_SLOWPATH
);
2832 if (kmem_cache_debug(s
) &&
2833 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
2838 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2841 prior
= page
->freelist
;
2842 counters
= page
->counters
;
2843 set_freepointer(s
, tail
, prior
);
2844 new.counters
= counters
;
2845 was_frozen
= new.frozen
;
2847 if ((!new.inuse
|| !prior
) && !was_frozen
) {
2849 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
2852 * Slab was on no list before and will be
2854 * We can defer the list move and instead
2859 } else { /* Needs to be taken off a list */
2861 n
= get_node(s
, page_to_nid(page
));
2863 * Speculatively acquire the list_lock.
2864 * If the cmpxchg does not succeed then we may
2865 * drop the list_lock without any processing.
2867 * Otherwise the list_lock will synchronize with
2868 * other processors updating the list of slabs.
2870 spin_lock_irqsave(&n
->list_lock
, flags
);
2875 } while (!cmpxchg_double_slab(s
, page
,
2883 * If we just froze the page then put it onto the
2884 * per cpu partial list.
2886 if (new.frozen
&& !was_frozen
) {
2887 put_cpu_partial(s
, page
, 1);
2888 stat(s
, CPU_PARTIAL_FREE
);
2891 * The list lock was not taken therefore no list
2892 * activity can be necessary.
2895 stat(s
, FREE_FROZEN
);
2899 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
2903 * Objects left in the slab. If it was not on the partial list before
2906 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
2907 if (kmem_cache_debug(s
))
2908 remove_full(s
, n
, page
);
2909 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2910 stat(s
, FREE_ADD_PARTIAL
);
2912 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2918 * Slab on the partial list.
2920 remove_partial(n
, page
);
2921 stat(s
, FREE_REMOVE_PARTIAL
);
2923 /* Slab must be on the full list */
2924 remove_full(s
, n
, page
);
2927 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2929 discard_slab(s
, page
);
2933 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2934 * can perform fastpath freeing without additional function calls.
2936 * The fastpath is only possible if we are freeing to the current cpu slab
2937 * of this processor. This typically the case if we have just allocated
2940 * If fastpath is not possible then fall back to __slab_free where we deal
2941 * with all sorts of special processing.
2943 * Bulk free of a freelist with several objects (all pointing to the
2944 * same page) possible by specifying head and tail ptr, plus objects
2945 * count (cnt). Bulk free indicated by tail pointer being set.
2947 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
2948 struct page
*page
, void *head
, void *tail
,
2949 int cnt
, unsigned long addr
)
2951 void *tail_obj
= tail
? : head
;
2952 struct kmem_cache_cpu
*c
;
2956 * Determine the currently cpus per cpu slab.
2957 * The cpu may change afterward. However that does not matter since
2958 * data is retrieved via this pointer. If we are on the same cpu
2959 * during the cmpxchg then the free will succeed.
2962 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2963 c
= raw_cpu_ptr(s
->cpu_slab
);
2964 } while (IS_ENABLED(CONFIG_PREEMPT
) &&
2965 unlikely(tid
!= READ_ONCE(c
->tid
)));
2967 /* Same with comment on barrier() in slab_alloc_node() */
2970 if (likely(page
== c
->page
)) {
2971 set_freepointer(s
, tail_obj
, c
->freelist
);
2973 if (unlikely(!this_cpu_cmpxchg_double(
2974 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2976 head
, next_tid(tid
)))) {
2978 note_cmpxchg_failure("slab_free", s
, tid
);
2981 stat(s
, FREE_FASTPATH
);
2983 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
2987 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
2988 void *head
, void *tail
, int cnt
,
2991 slab_free_freelist_hook(s
, head
, tail
);
2993 * slab_free_freelist_hook() could have put the items into quarantine.
2994 * If so, no need to free them.
2996 if (s
->flags
& SLAB_KASAN
&& !(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
2998 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3002 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3004 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3008 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3010 s
= cache_from_obj(s
, x
);
3013 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3014 trace_kmem_cache_free(_RET_IP_
, x
);
3016 EXPORT_SYMBOL(kmem_cache_free
);
3018 struct detached_freelist
{
3023 struct kmem_cache
*s
;
3027 * This function progressively scans the array with free objects (with
3028 * a limited look ahead) and extract objects belonging to the same
3029 * page. It builds a detached freelist directly within the given
3030 * page/objects. This can happen without any need for
3031 * synchronization, because the objects are owned by running process.
3032 * The freelist is build up as a single linked list in the objects.
3033 * The idea is, that this detached freelist can then be bulk
3034 * transferred to the real freelist(s), but only requiring a single
3035 * synchronization primitive. Look ahead in the array is limited due
3036 * to performance reasons.
3039 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3040 void **p
, struct detached_freelist
*df
)
3042 size_t first_skipped_index
= 0;
3047 /* Always re-init detached_freelist */
3052 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3053 } while (!object
&& size
);
3058 page
= virt_to_head_page(object
);
3060 /* Handle kalloc'ed objects */
3061 if (unlikely(!PageSlab(page
))) {
3062 BUG_ON(!PageCompound(page
));
3064 __free_pages(page
, compound_order(page
));
3065 p
[size
] = NULL
; /* mark object processed */
3068 /* Derive kmem_cache from object */
3069 df
->s
= page
->slab_cache
;
3071 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3074 /* Start new detached freelist */
3076 set_freepointer(df
->s
, object
, NULL
);
3078 df
->freelist
= object
;
3079 p
[size
] = NULL
; /* mark object processed */
3085 continue; /* Skip processed objects */
3087 /* df->page is always set at this point */
3088 if (df
->page
== virt_to_head_page(object
)) {
3089 /* Opportunity build freelist */
3090 set_freepointer(df
->s
, object
, df
->freelist
);
3091 df
->freelist
= object
;
3093 p
[size
] = NULL
; /* mark object processed */
3098 /* Limit look ahead search */
3102 if (!first_skipped_index
)
3103 first_skipped_index
= size
+ 1;
3106 return first_skipped_index
;
3109 /* Note that interrupts must be enabled when calling this function. */
3110 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3116 struct detached_freelist df
;
3118 size
= build_detached_freelist(s
, size
, p
, &df
);
3122 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
,_RET_IP_
);
3123 } while (likely(size
));
3125 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3127 /* Note that interrupts must be enabled when calling this function. */
3128 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3131 struct kmem_cache_cpu
*c
;
3134 /* memcg and kmem_cache debug support */
3135 s
= slab_pre_alloc_hook(s
, flags
);
3139 * Drain objects in the per cpu slab, while disabling local
3140 * IRQs, which protects against PREEMPT and interrupts
3141 * handlers invoking normal fastpath.
3143 local_irq_disable();
3144 c
= this_cpu_ptr(s
->cpu_slab
);
3146 for (i
= 0; i
< size
; i
++) {
3147 void *object
= c
->freelist
;
3149 if (unlikely(!object
)) {
3151 * Invoking slow path likely have side-effect
3152 * of re-populating per CPU c->freelist
3154 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3156 if (unlikely(!p
[i
]))
3159 c
= this_cpu_ptr(s
->cpu_slab
);
3160 continue; /* goto for-loop */
3162 c
->freelist
= get_freepointer(s
, object
);
3165 c
->tid
= next_tid(c
->tid
);
3168 /* Clear memory outside IRQ disabled fastpath loop */
3169 if (unlikely(flags
& __GFP_ZERO
)) {
3172 for (j
= 0; j
< i
; j
++)
3173 memset(p
[j
], 0, s
->object_size
);
3176 /* memcg and kmem_cache debug support */
3177 slab_post_alloc_hook(s
, flags
, size
, p
);
3181 slab_post_alloc_hook(s
, flags
, i
, p
);
3182 __kmem_cache_free_bulk(s
, i
, p
);
3185 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3189 * Object placement in a slab is made very easy because we always start at
3190 * offset 0. If we tune the size of the object to the alignment then we can
3191 * get the required alignment by putting one properly sized object after
3194 * Notice that the allocation order determines the sizes of the per cpu
3195 * caches. Each processor has always one slab available for allocations.
3196 * Increasing the allocation order reduces the number of times that slabs
3197 * must be moved on and off the partial lists and is therefore a factor in
3202 * Mininum / Maximum order of slab pages. This influences locking overhead
3203 * and slab fragmentation. A higher order reduces the number of partial slabs
3204 * and increases the number of allocations possible without having to
3205 * take the list_lock.
3207 static int slub_min_order
;
3208 static int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3209 static int slub_min_objects
;
3212 * Calculate the order of allocation given an slab object size.
3214 * The order of allocation has significant impact on performance and other
3215 * system components. Generally order 0 allocations should be preferred since
3216 * order 0 does not cause fragmentation in the page allocator. Larger objects
3217 * be problematic to put into order 0 slabs because there may be too much
3218 * unused space left. We go to a higher order if more than 1/16th of the slab
3221 * In order to reach satisfactory performance we must ensure that a minimum
3222 * number of objects is in one slab. Otherwise we may generate too much
3223 * activity on the partial lists which requires taking the list_lock. This is
3224 * less a concern for large slabs though which are rarely used.
3226 * slub_max_order specifies the order where we begin to stop considering the
3227 * number of objects in a slab as critical. If we reach slub_max_order then
3228 * we try to keep the page order as low as possible. So we accept more waste
3229 * of space in favor of a small page order.
3231 * Higher order allocations also allow the placement of more objects in a
3232 * slab and thereby reduce object handling overhead. If the user has
3233 * requested a higher mininum order then we start with that one instead of
3234 * the smallest order which will fit the object.
3236 static inline int slab_order(int size
, int min_objects
,
3237 int max_order
, int fract_leftover
, int reserved
)
3241 int min_order
= slub_min_order
;
3243 if (order_objects(min_order
, size
, reserved
) > MAX_OBJS_PER_PAGE
)
3244 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3246 for (order
= max(min_order
, get_order(min_objects
* size
+ reserved
));
3247 order
<= max_order
; order
++) {
3249 unsigned long slab_size
= PAGE_SIZE
<< order
;
3251 rem
= (slab_size
- reserved
) % size
;
3253 if (rem
<= slab_size
/ fract_leftover
)
3260 static inline int calculate_order(int size
, int reserved
)
3268 * Attempt to find best configuration for a slab. This
3269 * works by first attempting to generate a layout with
3270 * the best configuration and backing off gradually.
3272 * First we increase the acceptable waste in a slab. Then
3273 * we reduce the minimum objects required in a slab.
3275 min_objects
= slub_min_objects
;
3277 min_objects
= 4 * (fls(nr_cpu_ids
) + 1);
3278 max_objects
= order_objects(slub_max_order
, size
, reserved
);
3279 min_objects
= min(min_objects
, max_objects
);
3281 while (min_objects
> 1) {
3283 while (fraction
>= 4) {
3284 order
= slab_order(size
, min_objects
,
3285 slub_max_order
, fraction
, reserved
);
3286 if (order
<= slub_max_order
)
3294 * We were unable to place multiple objects in a slab. Now
3295 * lets see if we can place a single object there.
3297 order
= slab_order(size
, 1, slub_max_order
, 1, reserved
);
3298 if (order
<= slub_max_order
)
3302 * Doh this slab cannot be placed using slub_max_order.
3304 order
= slab_order(size
, 1, MAX_ORDER
, 1, reserved
);
3305 if (order
< MAX_ORDER
)
3311 init_kmem_cache_node(struct kmem_cache_node
*n
)
3314 spin_lock_init(&n
->list_lock
);
3315 INIT_LIST_HEAD(&n
->partial
);
3316 #ifdef CONFIG_SLUB_DEBUG
3317 atomic_long_set(&n
->nr_slabs
, 0);
3318 atomic_long_set(&n
->total_objects
, 0);
3319 INIT_LIST_HEAD(&n
->full
);
3323 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3325 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3326 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3329 * Must align to double word boundary for the double cmpxchg
3330 * instructions to work; see __pcpu_double_call_return_bool().
3332 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3333 2 * sizeof(void *));
3338 init_kmem_cache_cpus(s
);
3343 static struct kmem_cache
*kmem_cache_node
;
3346 * No kmalloc_node yet so do it by hand. We know that this is the first
3347 * slab on the node for this slabcache. There are no concurrent accesses
3350 * Note that this function only works on the kmem_cache_node
3351 * when allocating for the kmem_cache_node. This is used for bootstrapping
3352 * memory on a fresh node that has no slab structures yet.
3354 static void early_kmem_cache_node_alloc(int node
)
3357 struct kmem_cache_node
*n
;
3359 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3361 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3364 if (page_to_nid(page
) != node
) {
3365 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3366 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3371 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3374 kmem_cache_node
->node
[node
] = n
;
3375 #ifdef CONFIG_SLUB_DEBUG
3376 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3377 init_tracking(kmem_cache_node
, n
);
3379 kasan_kmalloc(kmem_cache_node
, n
, sizeof(struct kmem_cache_node
),
3381 init_kmem_cache_node(n
);
3382 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3385 * No locks need to be taken here as it has just been
3386 * initialized and there is no concurrent access.
3388 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3391 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3394 struct kmem_cache_node
*n
;
3396 for_each_kmem_cache_node(s
, node
, n
) {
3397 s
->node
[node
] = NULL
;
3398 kmem_cache_free(kmem_cache_node
, n
);
3402 void __kmem_cache_release(struct kmem_cache
*s
)
3404 cache_random_seq_destroy(s
);
3405 free_percpu(s
->cpu_slab
);
3406 free_kmem_cache_nodes(s
);
3409 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3413 for_each_node_state(node
, N_NORMAL_MEMORY
) {
3414 struct kmem_cache_node
*n
;
3416 if (slab_state
== DOWN
) {
3417 early_kmem_cache_node_alloc(node
);
3420 n
= kmem_cache_alloc_node(kmem_cache_node
,
3424 free_kmem_cache_nodes(s
);
3428 init_kmem_cache_node(n
);
3434 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3436 if (min
< MIN_PARTIAL
)
3438 else if (min
> MAX_PARTIAL
)
3440 s
->min_partial
= min
;
3443 static void set_cpu_partial(struct kmem_cache
*s
)
3445 #ifdef CONFIG_SLUB_CPU_PARTIAL
3447 * cpu_partial determined the maximum number of objects kept in the
3448 * per cpu partial lists of a processor.
3450 * Per cpu partial lists mainly contain slabs that just have one
3451 * object freed. If they are used for allocation then they can be
3452 * filled up again with minimal effort. The slab will never hit the
3453 * per node partial lists and therefore no locking will be required.
3455 * This setting also determines
3457 * A) The number of objects from per cpu partial slabs dumped to the
3458 * per node list when we reach the limit.
3459 * B) The number of objects in cpu partial slabs to extract from the
3460 * per node list when we run out of per cpu objects. We only fetch
3461 * 50% to keep some capacity around for frees.
3463 if (!kmem_cache_has_cpu_partial(s
))
3465 else if (s
->size
>= PAGE_SIZE
)
3467 else if (s
->size
>= 1024)
3469 else if (s
->size
>= 256)
3470 s
->cpu_partial
= 13;
3472 s
->cpu_partial
= 30;
3477 * calculate_sizes() determines the order and the distribution of data within
3480 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3482 slab_flags_t flags
= s
->flags
;
3483 size_t size
= s
->object_size
;
3487 * Round up object size to the next word boundary. We can only
3488 * place the free pointer at word boundaries and this determines
3489 * the possible location of the free pointer.
3491 size
= ALIGN(size
, sizeof(void *));
3493 #ifdef CONFIG_SLUB_DEBUG
3495 * Determine if we can poison the object itself. If the user of
3496 * the slab may touch the object after free or before allocation
3497 * then we should never poison the object itself.
3499 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3501 s
->flags
|= __OBJECT_POISON
;
3503 s
->flags
&= ~__OBJECT_POISON
;
3507 * If we are Redzoning then check if there is some space between the
3508 * end of the object and the free pointer. If not then add an
3509 * additional word to have some bytes to store Redzone information.
3511 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3512 size
+= sizeof(void *);
3516 * With that we have determined the number of bytes in actual use
3517 * by the object. This is the potential offset to the free pointer.
3521 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3524 * Relocate free pointer after the object if it is not
3525 * permitted to overwrite the first word of the object on
3528 * This is the case if we do RCU, have a constructor or
3529 * destructor or are poisoning the objects.
3532 size
+= sizeof(void *);
3535 #ifdef CONFIG_SLUB_DEBUG
3536 if (flags
& SLAB_STORE_USER
)
3538 * Need to store information about allocs and frees after
3541 size
+= 2 * sizeof(struct track
);
3544 kasan_cache_create(s
, &size
, &s
->flags
);
3545 #ifdef CONFIG_SLUB_DEBUG
3546 if (flags
& SLAB_RED_ZONE
) {
3548 * Add some empty padding so that we can catch
3549 * overwrites from earlier objects rather than let
3550 * tracking information or the free pointer be
3551 * corrupted if a user writes before the start
3554 size
+= sizeof(void *);
3556 s
->red_left_pad
= sizeof(void *);
3557 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3558 size
+= s
->red_left_pad
;
3563 * SLUB stores one object immediately after another beginning from
3564 * offset 0. In order to align the objects we have to simply size
3565 * each object to conform to the alignment.
3567 size
= ALIGN(size
, s
->align
);
3569 if (forced_order
>= 0)
3570 order
= forced_order
;
3572 order
= calculate_order(size
, s
->reserved
);
3579 s
->allocflags
|= __GFP_COMP
;
3581 if (s
->flags
& SLAB_CACHE_DMA
)
3582 s
->allocflags
|= GFP_DMA
;
3584 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3585 s
->allocflags
|= __GFP_RECLAIMABLE
;
3588 * Determine the number of objects per slab
3590 s
->oo
= oo_make(order
, size
, s
->reserved
);
3591 s
->min
= oo_make(get_order(size
), size
, s
->reserved
);
3592 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3595 return !!oo_objects(s
->oo
);
3598 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3600 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
, s
->ctor
);
3602 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3603 s
->random
= get_random_long();
3606 if (need_reserve_slab_rcu
&& (s
->flags
& SLAB_TYPESAFE_BY_RCU
))
3607 s
->reserved
= sizeof(struct rcu_head
);
3609 if (!calculate_sizes(s
, -1))
3611 if (disable_higher_order_debug
) {
3613 * Disable debugging flags that store metadata if the min slab
3616 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3617 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3619 if (!calculate_sizes(s
, -1))
3624 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3625 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3626 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3627 /* Enable fast mode */
3628 s
->flags
|= __CMPXCHG_DOUBLE
;
3632 * The larger the object size is, the more pages we want on the partial
3633 * list to avoid pounding the page allocator excessively.
3635 set_min_partial(s
, ilog2(s
->size
) / 2);
3640 s
->remote_node_defrag_ratio
= 1000;
3643 /* Initialize the pre-computed randomized freelist if slab is up */
3644 if (slab_state
>= UP
) {
3645 if (init_cache_random_seq(s
))
3649 if (!init_kmem_cache_nodes(s
))
3652 if (alloc_kmem_cache_cpus(s
))
3655 free_kmem_cache_nodes(s
);
3657 if (flags
& SLAB_PANIC
)
3658 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3659 s
->name
, (unsigned long)s
->size
, s
->size
,
3660 oo_order(s
->oo
), s
->offset
, (unsigned long)flags
);
3664 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3667 #ifdef CONFIG_SLUB_DEBUG
3668 void *addr
= page_address(page
);
3670 unsigned long *map
= kzalloc(BITS_TO_LONGS(page
->objects
) *
3671 sizeof(long), GFP_ATOMIC
);
3674 slab_err(s
, page
, text
, s
->name
);
3677 get_map(s
, page
, map
);
3678 for_each_object(p
, s
, addr
, page
->objects
) {
3680 if (!test_bit(slab_index(p
, s
, addr
), map
)) {
3681 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3682 print_tracking(s
, p
);
3691 * Attempt to free all partial slabs on a node.
3692 * This is called from __kmem_cache_shutdown(). We must take list_lock
3693 * because sysfs file might still access partial list after the shutdowning.
3695 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3698 struct page
*page
, *h
;
3700 BUG_ON(irqs_disabled());
3701 spin_lock_irq(&n
->list_lock
);
3702 list_for_each_entry_safe(page
, h
, &n
->partial
, lru
) {
3704 remove_partial(n
, page
);
3705 list_add(&page
->lru
, &discard
);
3707 list_slab_objects(s
, page
,
3708 "Objects remaining in %s on __kmem_cache_shutdown()");
3711 spin_unlock_irq(&n
->list_lock
);
3713 list_for_each_entry_safe(page
, h
, &discard
, lru
)
3714 discard_slab(s
, page
);
3718 * Release all resources used by a slab cache.
3720 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3723 struct kmem_cache_node
*n
;
3726 /* Attempt to free all objects */
3727 for_each_kmem_cache_node(s
, node
, n
) {
3729 if (n
->nr_partial
|| slabs_node(s
, node
))
3732 sysfs_slab_remove(s
);
3736 /********************************************************************
3738 *******************************************************************/
3740 static int __init
setup_slub_min_order(char *str
)
3742 get_option(&str
, &slub_min_order
);
3747 __setup("slub_min_order=", setup_slub_min_order
);
3749 static int __init
setup_slub_max_order(char *str
)
3751 get_option(&str
, &slub_max_order
);
3752 slub_max_order
= min(slub_max_order
, MAX_ORDER
- 1);
3757 __setup("slub_max_order=", setup_slub_max_order
);
3759 static int __init
setup_slub_min_objects(char *str
)
3761 get_option(&str
, &slub_min_objects
);
3766 __setup("slub_min_objects=", setup_slub_min_objects
);
3768 void *__kmalloc(size_t size
, gfp_t flags
)
3770 struct kmem_cache
*s
;
3773 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3774 return kmalloc_large(size
, flags
);
3776 s
= kmalloc_slab(size
, flags
);
3778 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3781 ret
= slab_alloc(s
, flags
, _RET_IP_
);
3783 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
3785 kasan_kmalloc(s
, ret
, size
, flags
);
3789 EXPORT_SYMBOL(__kmalloc
);
3792 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
3797 flags
|= __GFP_COMP
| __GFP_NOTRACK
;
3798 page
= alloc_pages_node(node
, flags
, get_order(size
));
3800 ptr
= page_address(page
);
3802 kmalloc_large_node_hook(ptr
, size
, flags
);
3806 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3808 struct kmem_cache
*s
;
3811 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
3812 ret
= kmalloc_large_node(size
, flags
, node
);
3814 trace_kmalloc_node(_RET_IP_
, ret
,
3815 size
, PAGE_SIZE
<< get_order(size
),
3821 s
= kmalloc_slab(size
, flags
);
3823 if (unlikely(ZERO_OR_NULL_PTR(s
)))
3826 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
);
3828 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
3830 kasan_kmalloc(s
, ret
, size
, flags
);
3834 EXPORT_SYMBOL(__kmalloc_node
);
3837 #ifdef CONFIG_HARDENED_USERCOPY
3839 * Rejects objects that are incorrectly sized.
3841 * Returns NULL if check passes, otherwise const char * to name of cache
3842 * to indicate an error.
3844 const char *__check_heap_object(const void *ptr
, unsigned long n
,
3847 struct kmem_cache
*s
;
3848 unsigned long offset
;
3851 /* Find object and usable object size. */
3852 s
= page
->slab_cache
;
3853 object_size
= slab_ksize(s
);
3855 /* Reject impossible pointers. */
3856 if (ptr
< page_address(page
))
3859 /* Find offset within object. */
3860 offset
= (ptr
- page_address(page
)) % s
->size
;
3862 /* Adjust for redzone and reject if within the redzone. */
3863 if (kmem_cache_debug(s
) && s
->flags
& SLAB_RED_ZONE
) {
3864 if (offset
< s
->red_left_pad
)
3866 offset
-= s
->red_left_pad
;
3869 /* Allow address range falling entirely within object size. */
3870 if (offset
<= object_size
&& n
<= object_size
- offset
)
3875 #endif /* CONFIG_HARDENED_USERCOPY */
3877 static size_t __ksize(const void *object
)
3881 if (unlikely(object
== ZERO_SIZE_PTR
))
3884 page
= virt_to_head_page(object
);
3886 if (unlikely(!PageSlab(page
))) {
3887 WARN_ON(!PageCompound(page
));
3888 return PAGE_SIZE
<< compound_order(page
);
3891 return slab_ksize(page
->slab_cache
);
3894 size_t ksize(const void *object
)
3896 size_t size
= __ksize(object
);
3897 /* We assume that ksize callers could use whole allocated area,
3898 * so we need to unpoison this area.
3900 kasan_unpoison_shadow(object
, size
);
3903 EXPORT_SYMBOL(ksize
);
3905 void kfree(const void *x
)
3908 void *object
= (void *)x
;
3910 trace_kfree(_RET_IP_
, x
);
3912 if (unlikely(ZERO_OR_NULL_PTR(x
)))
3915 page
= virt_to_head_page(x
);
3916 if (unlikely(!PageSlab(page
))) {
3917 BUG_ON(!PageCompound(page
));
3919 __free_pages(page
, compound_order(page
));
3922 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
3924 EXPORT_SYMBOL(kfree
);
3926 #define SHRINK_PROMOTE_MAX 32
3929 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3930 * up most to the head of the partial lists. New allocations will then
3931 * fill those up and thus they can be removed from the partial lists.
3933 * The slabs with the least items are placed last. This results in them
3934 * being allocated from last increasing the chance that the last objects
3935 * are freed in them.
3937 int __kmem_cache_shrink(struct kmem_cache
*s
)
3941 struct kmem_cache_node
*n
;
3944 struct list_head discard
;
3945 struct list_head promote
[SHRINK_PROMOTE_MAX
];
3946 unsigned long flags
;
3950 for_each_kmem_cache_node(s
, node
, n
) {
3951 INIT_LIST_HEAD(&discard
);
3952 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
3953 INIT_LIST_HEAD(promote
+ i
);
3955 spin_lock_irqsave(&n
->list_lock
, flags
);
3958 * Build lists of slabs to discard or promote.
3960 * Note that concurrent frees may occur while we hold the
3961 * list_lock. page->inuse here is the upper limit.
3963 list_for_each_entry_safe(page
, t
, &n
->partial
, lru
) {
3964 int free
= page
->objects
- page
->inuse
;
3966 /* Do not reread page->inuse */
3969 /* We do not keep full slabs on the list */
3972 if (free
== page
->objects
) {
3973 list_move(&page
->lru
, &discard
);
3975 } else if (free
<= SHRINK_PROMOTE_MAX
)
3976 list_move(&page
->lru
, promote
+ free
- 1);
3980 * Promote the slabs filled up most to the head of the
3983 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
3984 list_splice(promote
+ i
, &n
->partial
);
3986 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3988 /* Release empty slabs */
3989 list_for_each_entry_safe(page
, t
, &discard
, lru
)
3990 discard_slab(s
, page
);
3992 if (slabs_node(s
, node
))
4000 static void kmemcg_cache_deact_after_rcu(struct kmem_cache
*s
)
4003 * Called with all the locks held after a sched RCU grace period.
4004 * Even if @s becomes empty after shrinking, we can't know that @s
4005 * doesn't have allocations already in-flight and thus can't
4006 * destroy @s until the associated memcg is released.
4008 * However, let's remove the sysfs files for empty caches here.
4009 * Each cache has a lot of interface files which aren't
4010 * particularly useful for empty draining caches; otherwise, we can
4011 * easily end up with millions of unnecessary sysfs files on
4012 * systems which have a lot of memory and transient cgroups.
4014 if (!__kmem_cache_shrink(s
))
4015 sysfs_slab_remove(s
);
4018 void __kmemcg_cache_deactivate(struct kmem_cache
*s
)
4021 * Disable empty slabs caching. Used to avoid pinning offline
4022 * memory cgroups by kmem pages that can be freed.
4024 slub_set_cpu_partial(s
, 0);
4028 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4029 * we have to make sure the change is visible before shrinking.
4031 slab_deactivate_memcg_cache_rcu_sched(s
, kmemcg_cache_deact_after_rcu
);
4035 static int slab_mem_going_offline_callback(void *arg
)
4037 struct kmem_cache
*s
;
4039 mutex_lock(&slab_mutex
);
4040 list_for_each_entry(s
, &slab_caches
, list
)
4041 __kmem_cache_shrink(s
);
4042 mutex_unlock(&slab_mutex
);
4047 static void slab_mem_offline_callback(void *arg
)
4049 struct kmem_cache_node
*n
;
4050 struct kmem_cache
*s
;
4051 struct memory_notify
*marg
= arg
;
4054 offline_node
= marg
->status_change_nid_normal
;
4057 * If the node still has available memory. we need kmem_cache_node
4060 if (offline_node
< 0)
4063 mutex_lock(&slab_mutex
);
4064 list_for_each_entry(s
, &slab_caches
, list
) {
4065 n
= get_node(s
, offline_node
);
4068 * if n->nr_slabs > 0, slabs still exist on the node
4069 * that is going down. We were unable to free them,
4070 * and offline_pages() function shouldn't call this
4071 * callback. So, we must fail.
4073 BUG_ON(slabs_node(s
, offline_node
));
4075 s
->node
[offline_node
] = NULL
;
4076 kmem_cache_free(kmem_cache_node
, n
);
4079 mutex_unlock(&slab_mutex
);
4082 static int slab_mem_going_online_callback(void *arg
)
4084 struct kmem_cache_node
*n
;
4085 struct kmem_cache
*s
;
4086 struct memory_notify
*marg
= arg
;
4087 int nid
= marg
->status_change_nid_normal
;
4091 * If the node's memory is already available, then kmem_cache_node is
4092 * already created. Nothing to do.
4098 * We are bringing a node online. No memory is available yet. We must
4099 * allocate a kmem_cache_node structure in order to bring the node
4102 mutex_lock(&slab_mutex
);
4103 list_for_each_entry(s
, &slab_caches
, list
) {
4105 * XXX: kmem_cache_alloc_node will fallback to other nodes
4106 * since memory is not yet available from the node that
4109 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4114 init_kmem_cache_node(n
);
4118 mutex_unlock(&slab_mutex
);
4122 static int slab_memory_callback(struct notifier_block
*self
,
4123 unsigned long action
, void *arg
)
4128 case MEM_GOING_ONLINE
:
4129 ret
= slab_mem_going_online_callback(arg
);
4131 case MEM_GOING_OFFLINE
:
4132 ret
= slab_mem_going_offline_callback(arg
);
4135 case MEM_CANCEL_ONLINE
:
4136 slab_mem_offline_callback(arg
);
4139 case MEM_CANCEL_OFFLINE
:
4143 ret
= notifier_from_errno(ret
);
4149 static struct notifier_block slab_memory_callback_nb
= {
4150 .notifier_call
= slab_memory_callback
,
4151 .priority
= SLAB_CALLBACK_PRI
,
4154 /********************************************************************
4155 * Basic setup of slabs
4156 *******************************************************************/
4159 * Used for early kmem_cache structures that were allocated using
4160 * the page allocator. Allocate them properly then fix up the pointers
4161 * that may be pointing to the wrong kmem_cache structure.
4164 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4167 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4168 struct kmem_cache_node
*n
;
4170 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4173 * This runs very early, and only the boot processor is supposed to be
4174 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4177 __flush_cpu_slab(s
, smp_processor_id());
4178 for_each_kmem_cache_node(s
, node
, n
) {
4181 list_for_each_entry(p
, &n
->partial
, lru
)
4184 #ifdef CONFIG_SLUB_DEBUG
4185 list_for_each_entry(p
, &n
->full
, lru
)
4189 slab_init_memcg_params(s
);
4190 list_add(&s
->list
, &slab_caches
);
4191 memcg_link_cache(s
);
4195 void __init
kmem_cache_init(void)
4197 static __initdata
struct kmem_cache boot_kmem_cache
,
4198 boot_kmem_cache_node
;
4200 if (debug_guardpage_minorder())
4203 kmem_cache_node
= &boot_kmem_cache_node
;
4204 kmem_cache
= &boot_kmem_cache
;
4206 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4207 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
);
4209 register_hotmemory_notifier(&slab_memory_callback_nb
);
4211 /* Able to allocate the per node structures */
4212 slab_state
= PARTIAL
;
4214 create_boot_cache(kmem_cache
, "kmem_cache",
4215 offsetof(struct kmem_cache
, node
) +
4216 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4217 SLAB_HWCACHE_ALIGN
);
4219 kmem_cache
= bootstrap(&boot_kmem_cache
);
4222 * Allocate kmem_cache_node properly from the kmem_cache slab.
4223 * kmem_cache_node is separately allocated so no need to
4224 * update any list pointers.
4226 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4228 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4229 setup_kmalloc_cache_index_table();
4230 create_kmalloc_caches(0);
4232 /* Setup random freelists for each cache */
4233 init_freelist_randomization();
4235 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4238 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%u, Nodes=%d\n",
4240 slub_min_order
, slub_max_order
, slub_min_objects
,
4241 nr_cpu_ids
, nr_node_ids
);
4244 void __init
kmem_cache_init_late(void)
4249 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
4250 slab_flags_t flags
, void (*ctor
)(void *))
4252 struct kmem_cache
*s
, *c
;
4254 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4259 * Adjust the object sizes so that we clear
4260 * the complete object on kzalloc.
4262 s
->object_size
= max(s
->object_size
, (int)size
);
4263 s
->inuse
= max_t(int, s
->inuse
, ALIGN(size
, sizeof(void *)));
4265 for_each_memcg_cache(c
, s
) {
4266 c
->object_size
= s
->object_size
;
4267 c
->inuse
= max_t(int, c
->inuse
,
4268 ALIGN(size
, sizeof(void *)));
4271 if (sysfs_slab_alias(s
, name
)) {
4280 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4284 err
= kmem_cache_open(s
, flags
);
4288 /* Mutex is not taken during early boot */
4289 if (slab_state
<= UP
)
4292 memcg_propagate_slab_attrs(s
);
4293 err
= sysfs_slab_add(s
);
4295 __kmem_cache_release(s
);
4300 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4302 struct kmem_cache
*s
;
4305 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4306 return kmalloc_large(size
, gfpflags
);
4308 s
= kmalloc_slab(size
, gfpflags
);
4310 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4313 ret
= slab_alloc(s
, gfpflags
, caller
);
4315 /* Honor the call site pointer we received. */
4316 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4322 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4323 int node
, unsigned long caller
)
4325 struct kmem_cache
*s
;
4328 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4329 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4331 trace_kmalloc_node(caller
, ret
,
4332 size
, PAGE_SIZE
<< get_order(size
),
4338 s
= kmalloc_slab(size
, gfpflags
);
4340 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4343 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
);
4345 /* Honor the call site pointer we received. */
4346 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4353 static int count_inuse(struct page
*page
)
4358 static int count_total(struct page
*page
)
4360 return page
->objects
;
4364 #ifdef CONFIG_SLUB_DEBUG
4365 static int validate_slab(struct kmem_cache
*s
, struct page
*page
,
4369 void *addr
= page_address(page
);
4371 if (!check_slab(s
, page
) ||
4372 !on_freelist(s
, page
, NULL
))
4375 /* Now we know that a valid freelist exists */
4376 bitmap_zero(map
, page
->objects
);
4378 get_map(s
, page
, map
);
4379 for_each_object(p
, s
, addr
, page
->objects
) {
4380 if (test_bit(slab_index(p
, s
, addr
), map
))
4381 if (!check_object(s
, page
, p
, SLUB_RED_INACTIVE
))
4385 for_each_object(p
, s
, addr
, page
->objects
)
4386 if (!test_bit(slab_index(p
, s
, addr
), map
))
4387 if (!check_object(s
, page
, p
, SLUB_RED_ACTIVE
))
4392 static void validate_slab_slab(struct kmem_cache
*s
, struct page
*page
,
4396 validate_slab(s
, page
, map
);
4400 static int validate_slab_node(struct kmem_cache
*s
,
4401 struct kmem_cache_node
*n
, unsigned long *map
)
4403 unsigned long count
= 0;
4405 unsigned long flags
;
4407 spin_lock_irqsave(&n
->list_lock
, flags
);
4409 list_for_each_entry(page
, &n
->partial
, lru
) {
4410 validate_slab_slab(s
, page
, map
);
4413 if (count
!= n
->nr_partial
)
4414 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4415 s
->name
, count
, n
->nr_partial
);
4417 if (!(s
->flags
& SLAB_STORE_USER
))
4420 list_for_each_entry(page
, &n
->full
, lru
) {
4421 validate_slab_slab(s
, page
, map
);
4424 if (count
!= atomic_long_read(&n
->nr_slabs
))
4425 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4426 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4429 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4433 static long validate_slab_cache(struct kmem_cache
*s
)
4436 unsigned long count
= 0;
4437 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4438 sizeof(unsigned long), GFP_KERNEL
);
4439 struct kmem_cache_node
*n
;
4445 for_each_kmem_cache_node(s
, node
, n
)
4446 count
+= validate_slab_node(s
, n
, map
);
4451 * Generate lists of code addresses where slabcache objects are allocated
4456 unsigned long count
;
4463 DECLARE_BITMAP(cpus
, NR_CPUS
);
4469 unsigned long count
;
4470 struct location
*loc
;
4473 static void free_loc_track(struct loc_track
*t
)
4476 free_pages((unsigned long)t
->loc
,
4477 get_order(sizeof(struct location
) * t
->max
));
4480 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4485 order
= get_order(sizeof(struct location
) * max
);
4487 l
= (void *)__get_free_pages(flags
, order
);
4492 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4500 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4501 const struct track
*track
)
4503 long start
, end
, pos
;
4505 unsigned long caddr
;
4506 unsigned long age
= jiffies
- track
->when
;
4512 pos
= start
+ (end
- start
+ 1) / 2;
4515 * There is nothing at "end". If we end up there
4516 * we need to add something to before end.
4521 caddr
= t
->loc
[pos
].addr
;
4522 if (track
->addr
== caddr
) {
4528 if (age
< l
->min_time
)
4530 if (age
> l
->max_time
)
4533 if (track
->pid
< l
->min_pid
)
4534 l
->min_pid
= track
->pid
;
4535 if (track
->pid
> l
->max_pid
)
4536 l
->max_pid
= track
->pid
;
4538 cpumask_set_cpu(track
->cpu
,
4539 to_cpumask(l
->cpus
));
4541 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4545 if (track
->addr
< caddr
)
4552 * Not found. Insert new tracking element.
4554 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4560 (t
->count
- pos
) * sizeof(struct location
));
4563 l
->addr
= track
->addr
;
4567 l
->min_pid
= track
->pid
;
4568 l
->max_pid
= track
->pid
;
4569 cpumask_clear(to_cpumask(l
->cpus
));
4570 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4571 nodes_clear(l
->nodes
);
4572 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4576 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4577 struct page
*page
, enum track_item alloc
,
4580 void *addr
= page_address(page
);
4583 bitmap_zero(map
, page
->objects
);
4584 get_map(s
, page
, map
);
4586 for_each_object(p
, s
, addr
, page
->objects
)
4587 if (!test_bit(slab_index(p
, s
, addr
), map
))
4588 add_location(t
, s
, get_track(s
, p
, alloc
));
4591 static int list_locations(struct kmem_cache
*s
, char *buf
,
4592 enum track_item alloc
)
4596 struct loc_track t
= { 0, 0, NULL
};
4598 unsigned long *map
= kmalloc(BITS_TO_LONGS(oo_objects(s
->max
)) *
4599 sizeof(unsigned long), GFP_KERNEL
);
4600 struct kmem_cache_node
*n
;
4602 if (!map
|| !alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4605 return sprintf(buf
, "Out of memory\n");
4607 /* Push back cpu slabs */
4610 for_each_kmem_cache_node(s
, node
, n
) {
4611 unsigned long flags
;
4614 if (!atomic_long_read(&n
->nr_slabs
))
4617 spin_lock_irqsave(&n
->list_lock
, flags
);
4618 list_for_each_entry(page
, &n
->partial
, lru
)
4619 process_slab(&t
, s
, page
, alloc
, map
);
4620 list_for_each_entry(page
, &n
->full
, lru
)
4621 process_slab(&t
, s
, page
, alloc
, map
);
4622 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4625 for (i
= 0; i
< t
.count
; i
++) {
4626 struct location
*l
= &t
.loc
[i
];
4628 if (len
> PAGE_SIZE
- KSYM_SYMBOL_LEN
- 100)
4630 len
+= sprintf(buf
+ len
, "%7ld ", l
->count
);
4633 len
+= sprintf(buf
+ len
, "%pS", (void *)l
->addr
);
4635 len
+= sprintf(buf
+ len
, "<not-available>");
4637 if (l
->sum_time
!= l
->min_time
) {
4638 len
+= sprintf(buf
+ len
, " age=%ld/%ld/%ld",
4640 (long)div_u64(l
->sum_time
, l
->count
),
4643 len
+= sprintf(buf
+ len
, " age=%ld",
4646 if (l
->min_pid
!= l
->max_pid
)
4647 len
+= sprintf(buf
+ len
, " pid=%ld-%ld",
4648 l
->min_pid
, l
->max_pid
);
4650 len
+= sprintf(buf
+ len
, " pid=%ld",
4653 if (num_online_cpus() > 1 &&
4654 !cpumask_empty(to_cpumask(l
->cpus
)) &&
4655 len
< PAGE_SIZE
- 60)
4656 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4658 cpumask_pr_args(to_cpumask(l
->cpus
)));
4660 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
) &&
4661 len
< PAGE_SIZE
- 60)
4662 len
+= scnprintf(buf
+ len
, PAGE_SIZE
- len
- 50,
4664 nodemask_pr_args(&l
->nodes
));
4666 len
+= sprintf(buf
+ len
, "\n");
4672 len
+= sprintf(buf
, "No data\n");
4677 #ifdef SLUB_RESILIENCY_TEST
4678 static void __init
resiliency_test(void)
4682 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4684 pr_err("SLUB resiliency testing\n");
4685 pr_err("-----------------------\n");
4686 pr_err("A. Corruption after allocation\n");
4688 p
= kzalloc(16, GFP_KERNEL
);
4690 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4693 validate_slab_cache(kmalloc_caches
[4]);
4695 /* Hmmm... The next two are dangerous */
4696 p
= kzalloc(32, GFP_KERNEL
);
4697 p
[32 + sizeof(void *)] = 0x34;
4698 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4700 pr_err("If allocated object is overwritten then not detectable\n\n");
4702 validate_slab_cache(kmalloc_caches
[5]);
4703 p
= kzalloc(64, GFP_KERNEL
);
4704 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4706 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4708 pr_err("If allocated object is overwritten then not detectable\n\n");
4709 validate_slab_cache(kmalloc_caches
[6]);
4711 pr_err("\nB. Corruption after free\n");
4712 p
= kzalloc(128, GFP_KERNEL
);
4715 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4716 validate_slab_cache(kmalloc_caches
[7]);
4718 p
= kzalloc(256, GFP_KERNEL
);
4721 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4722 validate_slab_cache(kmalloc_caches
[8]);
4724 p
= kzalloc(512, GFP_KERNEL
);
4727 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4728 validate_slab_cache(kmalloc_caches
[9]);
4732 static void resiliency_test(void) {};
4737 enum slab_stat_type
{
4738 SL_ALL
, /* All slabs */
4739 SL_PARTIAL
, /* Only partially allocated slabs */
4740 SL_CPU
, /* Only slabs used for cpu caches */
4741 SL_OBJECTS
, /* Determine allocated objects not slabs */
4742 SL_TOTAL
/* Determine object capacity not slabs */
4745 #define SO_ALL (1 << SL_ALL)
4746 #define SO_PARTIAL (1 << SL_PARTIAL)
4747 #define SO_CPU (1 << SL_CPU)
4748 #define SO_OBJECTS (1 << SL_OBJECTS)
4749 #define SO_TOTAL (1 << SL_TOTAL)
4752 static bool memcg_sysfs_enabled
= IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON
);
4754 static int __init
setup_slub_memcg_sysfs(char *str
)
4758 if (get_option(&str
, &v
) > 0)
4759 memcg_sysfs_enabled
= v
;
4764 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs
);
4767 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4768 char *buf
, unsigned long flags
)
4770 unsigned long total
= 0;
4773 unsigned long *nodes
;
4775 nodes
= kzalloc(sizeof(unsigned long) * nr_node_ids
, GFP_KERNEL
);
4779 if (flags
& SO_CPU
) {
4782 for_each_possible_cpu(cpu
) {
4783 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4788 page
= READ_ONCE(c
->page
);
4792 node
= page_to_nid(page
);
4793 if (flags
& SO_TOTAL
)
4795 else if (flags
& SO_OBJECTS
)
4803 page
= slub_percpu_partial_read_once(c
);
4805 node
= page_to_nid(page
);
4806 if (flags
& SO_TOTAL
)
4808 else if (flags
& SO_OBJECTS
)
4819 #ifdef CONFIG_SLUB_DEBUG
4820 if (flags
& SO_ALL
) {
4821 struct kmem_cache_node
*n
;
4823 for_each_kmem_cache_node(s
, node
, n
) {
4825 if (flags
& SO_TOTAL
)
4826 x
= atomic_long_read(&n
->total_objects
);
4827 else if (flags
& SO_OBJECTS
)
4828 x
= atomic_long_read(&n
->total_objects
) -
4829 count_partial(n
, count_free
);
4831 x
= atomic_long_read(&n
->nr_slabs
);
4838 if (flags
& SO_PARTIAL
) {
4839 struct kmem_cache_node
*n
;
4841 for_each_kmem_cache_node(s
, node
, n
) {
4842 if (flags
& SO_TOTAL
)
4843 x
= count_partial(n
, count_total
);
4844 else if (flags
& SO_OBJECTS
)
4845 x
= count_partial(n
, count_inuse
);
4852 x
= sprintf(buf
, "%lu", total
);
4854 for (node
= 0; node
< nr_node_ids
; node
++)
4856 x
+= sprintf(buf
+ x
, " N%d=%lu",
4861 return x
+ sprintf(buf
+ x
, "\n");
4864 #ifdef CONFIG_SLUB_DEBUG
4865 static int any_slab_objects(struct kmem_cache
*s
)
4868 struct kmem_cache_node
*n
;
4870 for_each_kmem_cache_node(s
, node
, n
)
4871 if (atomic_long_read(&n
->total_objects
))
4878 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4879 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4881 struct slab_attribute
{
4882 struct attribute attr
;
4883 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
4884 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
4887 #define SLAB_ATTR_RO(_name) \
4888 static struct slab_attribute _name##_attr = \
4889 __ATTR(_name, 0400, _name##_show, NULL)
4891 #define SLAB_ATTR(_name) \
4892 static struct slab_attribute _name##_attr = \
4893 __ATTR(_name, 0600, _name##_show, _name##_store)
4895 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
4897 return sprintf(buf
, "%d\n", s
->size
);
4899 SLAB_ATTR_RO(slab_size
);
4901 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
4903 return sprintf(buf
, "%d\n", s
->align
);
4905 SLAB_ATTR_RO(align
);
4907 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
4909 return sprintf(buf
, "%d\n", s
->object_size
);
4911 SLAB_ATTR_RO(object_size
);
4913 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
4915 return sprintf(buf
, "%d\n", oo_objects(s
->oo
));
4917 SLAB_ATTR_RO(objs_per_slab
);
4919 static ssize_t
order_store(struct kmem_cache
*s
,
4920 const char *buf
, size_t length
)
4922 unsigned long order
;
4925 err
= kstrtoul(buf
, 10, &order
);
4929 if (order
> slub_max_order
|| order
< slub_min_order
)
4932 calculate_sizes(s
, order
);
4936 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
4938 return sprintf(buf
, "%d\n", oo_order(s
->oo
));
4942 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
4944 return sprintf(buf
, "%lu\n", s
->min_partial
);
4947 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
4953 err
= kstrtoul(buf
, 10, &min
);
4957 set_min_partial(s
, min
);
4960 SLAB_ATTR(min_partial
);
4962 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
4964 return sprintf(buf
, "%u\n", slub_cpu_partial(s
));
4967 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
4970 unsigned long objects
;
4973 err
= kstrtoul(buf
, 10, &objects
);
4976 if (objects
&& !kmem_cache_has_cpu_partial(s
))
4979 slub_set_cpu_partial(s
, objects
);
4983 SLAB_ATTR(cpu_partial
);
4985 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
4989 return sprintf(buf
, "%pS\n", s
->ctor
);
4993 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
4995 return sprintf(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
4997 SLAB_ATTR_RO(aliases
);
4999 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5001 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5003 SLAB_ATTR_RO(partial
);
5005 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5007 return show_slab_objects(s
, buf
, SO_CPU
);
5009 SLAB_ATTR_RO(cpu_slabs
);
5011 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5013 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5015 SLAB_ATTR_RO(objects
);
5017 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5019 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5021 SLAB_ATTR_RO(objects_partial
);
5023 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5030 for_each_online_cpu(cpu
) {
5033 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5036 pages
+= page
->pages
;
5037 objects
+= page
->pobjects
;
5041 len
= sprintf(buf
, "%d(%d)", objects
, pages
);
5044 for_each_online_cpu(cpu
) {
5047 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5049 if (page
&& len
< PAGE_SIZE
- 20)
5050 len
+= sprintf(buf
+ len
, " C%d=%d(%d)", cpu
,
5051 page
->pobjects
, page
->pages
);
5054 return len
+ sprintf(buf
+ len
, "\n");
5056 SLAB_ATTR_RO(slabs_cpu_partial
);
5058 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5060 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5063 static ssize_t
reclaim_account_store(struct kmem_cache
*s
,
5064 const char *buf
, size_t length
)
5066 s
->flags
&= ~SLAB_RECLAIM_ACCOUNT
;
5068 s
->flags
|= SLAB_RECLAIM_ACCOUNT
;
5071 SLAB_ATTR(reclaim_account
);
5073 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5075 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5077 SLAB_ATTR_RO(hwcache_align
);
5079 #ifdef CONFIG_ZONE_DMA
5080 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5082 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5084 SLAB_ATTR_RO(cache_dma
);
5087 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5089 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5091 SLAB_ATTR_RO(destroy_by_rcu
);
5093 static ssize_t
reserved_show(struct kmem_cache
*s
, char *buf
)
5095 return sprintf(buf
, "%d\n", s
->reserved
);
5097 SLAB_ATTR_RO(reserved
);
5099 #ifdef CONFIG_SLUB_DEBUG
5100 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5102 return show_slab_objects(s
, buf
, SO_ALL
);
5104 SLAB_ATTR_RO(slabs
);
5106 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5108 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5110 SLAB_ATTR_RO(total_objects
);
5112 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5114 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5117 static ssize_t
sanity_checks_store(struct kmem_cache
*s
,
5118 const char *buf
, size_t length
)
5120 s
->flags
&= ~SLAB_CONSISTENCY_CHECKS
;
5121 if (buf
[0] == '1') {
5122 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5123 s
->flags
|= SLAB_CONSISTENCY_CHECKS
;
5127 SLAB_ATTR(sanity_checks
);
5129 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5131 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5134 static ssize_t
trace_store(struct kmem_cache
*s
, const char *buf
,
5138 * Tracing a merged cache is going to give confusing results
5139 * as well as cause other issues like converting a mergeable
5140 * cache into an umergeable one.
5142 if (s
->refcount
> 1)
5145 s
->flags
&= ~SLAB_TRACE
;
5146 if (buf
[0] == '1') {
5147 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5148 s
->flags
|= SLAB_TRACE
;
5154 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5156 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5159 static ssize_t
red_zone_store(struct kmem_cache
*s
,
5160 const char *buf
, size_t length
)
5162 if (any_slab_objects(s
))
5165 s
->flags
&= ~SLAB_RED_ZONE
;
5166 if (buf
[0] == '1') {
5167 s
->flags
|= SLAB_RED_ZONE
;
5169 calculate_sizes(s
, -1);
5172 SLAB_ATTR(red_zone
);
5174 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5176 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5179 static ssize_t
poison_store(struct kmem_cache
*s
,
5180 const char *buf
, size_t length
)
5182 if (any_slab_objects(s
))
5185 s
->flags
&= ~SLAB_POISON
;
5186 if (buf
[0] == '1') {
5187 s
->flags
|= SLAB_POISON
;
5189 calculate_sizes(s
, -1);
5194 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5196 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5199 static ssize_t
store_user_store(struct kmem_cache
*s
,
5200 const char *buf
, size_t length
)
5202 if (any_slab_objects(s
))
5205 s
->flags
&= ~SLAB_STORE_USER
;
5206 if (buf
[0] == '1') {
5207 s
->flags
&= ~__CMPXCHG_DOUBLE
;
5208 s
->flags
|= SLAB_STORE_USER
;
5210 calculate_sizes(s
, -1);
5213 SLAB_ATTR(store_user
);
5215 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5220 static ssize_t
validate_store(struct kmem_cache
*s
,
5221 const char *buf
, size_t length
)
5225 if (buf
[0] == '1') {
5226 ret
= validate_slab_cache(s
);
5232 SLAB_ATTR(validate
);
5234 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5236 if (!(s
->flags
& SLAB_STORE_USER
))
5238 return list_locations(s
, buf
, TRACK_ALLOC
);
5240 SLAB_ATTR_RO(alloc_calls
);
5242 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5244 if (!(s
->flags
& SLAB_STORE_USER
))
5246 return list_locations(s
, buf
, TRACK_FREE
);
5248 SLAB_ATTR_RO(free_calls
);
5249 #endif /* CONFIG_SLUB_DEBUG */
5251 #ifdef CONFIG_FAILSLAB
5252 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5254 return sprintf(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5257 static ssize_t
failslab_store(struct kmem_cache
*s
, const char *buf
,
5260 if (s
->refcount
> 1)
5263 s
->flags
&= ~SLAB_FAILSLAB
;
5265 s
->flags
|= SLAB_FAILSLAB
;
5268 SLAB_ATTR(failslab
);
5271 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5276 static ssize_t
shrink_store(struct kmem_cache
*s
,
5277 const char *buf
, size_t length
)
5280 kmem_cache_shrink(s
);
5288 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5290 return sprintf(buf
, "%d\n", s
->remote_node_defrag_ratio
/ 10);
5293 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5294 const char *buf
, size_t length
)
5296 unsigned long ratio
;
5299 err
= kstrtoul(buf
, 10, &ratio
);
5304 s
->remote_node_defrag_ratio
= ratio
* 10;
5308 SLAB_ATTR(remote_node_defrag_ratio
);
5311 #ifdef CONFIG_SLUB_STATS
5312 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5314 unsigned long sum
= 0;
5317 int *data
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
5322 for_each_online_cpu(cpu
) {
5323 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5329 len
= sprintf(buf
, "%lu", sum
);
5332 for_each_online_cpu(cpu
) {
5333 if (data
[cpu
] && len
< PAGE_SIZE
- 20)
5334 len
+= sprintf(buf
+ len
, " C%d=%u", cpu
, data
[cpu
]);
5338 return len
+ sprintf(buf
+ len
, "\n");
5341 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5345 for_each_online_cpu(cpu
)
5346 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5349 #define STAT_ATTR(si, text) \
5350 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5352 return show_stat(s, buf, si); \
5354 static ssize_t text##_store(struct kmem_cache *s, \
5355 const char *buf, size_t length) \
5357 if (buf[0] != '0') \
5359 clear_stat(s, si); \
5364 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5365 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5366 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5367 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5368 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5369 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5370 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5371 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5372 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5373 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5374 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5375 STAT_ATTR(FREE_SLAB
, free_slab
);
5376 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5377 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5378 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5379 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5380 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5381 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5382 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5383 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5384 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5385 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5386 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5387 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5388 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5389 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5392 static struct attribute
*slab_attrs
[] = {
5393 &slab_size_attr
.attr
,
5394 &object_size_attr
.attr
,
5395 &objs_per_slab_attr
.attr
,
5397 &min_partial_attr
.attr
,
5398 &cpu_partial_attr
.attr
,
5400 &objects_partial_attr
.attr
,
5402 &cpu_slabs_attr
.attr
,
5406 &hwcache_align_attr
.attr
,
5407 &reclaim_account_attr
.attr
,
5408 &destroy_by_rcu_attr
.attr
,
5410 &reserved_attr
.attr
,
5411 &slabs_cpu_partial_attr
.attr
,
5412 #ifdef CONFIG_SLUB_DEBUG
5413 &total_objects_attr
.attr
,
5415 &sanity_checks_attr
.attr
,
5417 &red_zone_attr
.attr
,
5419 &store_user_attr
.attr
,
5420 &validate_attr
.attr
,
5421 &alloc_calls_attr
.attr
,
5422 &free_calls_attr
.attr
,
5424 #ifdef CONFIG_ZONE_DMA
5425 &cache_dma_attr
.attr
,
5428 &remote_node_defrag_ratio_attr
.attr
,
5430 #ifdef CONFIG_SLUB_STATS
5431 &alloc_fastpath_attr
.attr
,
5432 &alloc_slowpath_attr
.attr
,
5433 &free_fastpath_attr
.attr
,
5434 &free_slowpath_attr
.attr
,
5435 &free_frozen_attr
.attr
,
5436 &free_add_partial_attr
.attr
,
5437 &free_remove_partial_attr
.attr
,
5438 &alloc_from_partial_attr
.attr
,
5439 &alloc_slab_attr
.attr
,
5440 &alloc_refill_attr
.attr
,
5441 &alloc_node_mismatch_attr
.attr
,
5442 &free_slab_attr
.attr
,
5443 &cpuslab_flush_attr
.attr
,
5444 &deactivate_full_attr
.attr
,
5445 &deactivate_empty_attr
.attr
,
5446 &deactivate_to_head_attr
.attr
,
5447 &deactivate_to_tail_attr
.attr
,
5448 &deactivate_remote_frees_attr
.attr
,
5449 &deactivate_bypass_attr
.attr
,
5450 &order_fallback_attr
.attr
,
5451 &cmpxchg_double_fail_attr
.attr
,
5452 &cmpxchg_double_cpu_fail_attr
.attr
,
5453 &cpu_partial_alloc_attr
.attr
,
5454 &cpu_partial_free_attr
.attr
,
5455 &cpu_partial_node_attr
.attr
,
5456 &cpu_partial_drain_attr
.attr
,
5458 #ifdef CONFIG_FAILSLAB
5459 &failslab_attr
.attr
,
5465 static const struct attribute_group slab_attr_group
= {
5466 .attrs
= slab_attrs
,
5469 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5470 struct attribute
*attr
,
5473 struct slab_attribute
*attribute
;
5474 struct kmem_cache
*s
;
5477 attribute
= to_slab_attr(attr
);
5480 if (!attribute
->show
)
5483 err
= attribute
->show(s
, buf
);
5488 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5489 struct attribute
*attr
,
5490 const char *buf
, size_t len
)
5492 struct slab_attribute
*attribute
;
5493 struct kmem_cache
*s
;
5496 attribute
= to_slab_attr(attr
);
5499 if (!attribute
->store
)
5502 err
= attribute
->store(s
, buf
, len
);
5504 if (slab_state
>= FULL
&& err
>= 0 && is_root_cache(s
)) {
5505 struct kmem_cache
*c
;
5507 mutex_lock(&slab_mutex
);
5508 if (s
->max_attr_size
< len
)
5509 s
->max_attr_size
= len
;
5512 * This is a best effort propagation, so this function's return
5513 * value will be determined by the parent cache only. This is
5514 * basically because not all attributes will have a well
5515 * defined semantics for rollbacks - most of the actions will
5516 * have permanent effects.
5518 * Returning the error value of any of the children that fail
5519 * is not 100 % defined, in the sense that users seeing the
5520 * error code won't be able to know anything about the state of
5523 * Only returning the error code for the parent cache at least
5524 * has well defined semantics. The cache being written to
5525 * directly either failed or succeeded, in which case we loop
5526 * through the descendants with best-effort propagation.
5528 for_each_memcg_cache(c
, s
)
5529 attribute
->store(c
, buf
, len
);
5530 mutex_unlock(&slab_mutex
);
5536 static void memcg_propagate_slab_attrs(struct kmem_cache
*s
)
5540 char *buffer
= NULL
;
5541 struct kmem_cache
*root_cache
;
5543 if (is_root_cache(s
))
5546 root_cache
= s
->memcg_params
.root_cache
;
5549 * This mean this cache had no attribute written. Therefore, no point
5550 * in copying default values around
5552 if (!root_cache
->max_attr_size
)
5555 for (i
= 0; i
< ARRAY_SIZE(slab_attrs
); i
++) {
5558 struct slab_attribute
*attr
= to_slab_attr(slab_attrs
[i
]);
5561 if (!attr
|| !attr
->store
|| !attr
->show
)
5565 * It is really bad that we have to allocate here, so we will
5566 * do it only as a fallback. If we actually allocate, though,
5567 * we can just use the allocated buffer until the end.
5569 * Most of the slub attributes will tend to be very small in
5570 * size, but sysfs allows buffers up to a page, so they can
5571 * theoretically happen.
5575 else if (root_cache
->max_attr_size
< ARRAY_SIZE(mbuf
))
5578 buffer
= (char *) get_zeroed_page(GFP_KERNEL
);
5579 if (WARN_ON(!buffer
))
5584 len
= attr
->show(root_cache
, buf
);
5586 attr
->store(s
, buf
, len
);
5590 free_page((unsigned long)buffer
);
5594 static void kmem_cache_release(struct kobject
*k
)
5596 slab_kmem_cache_release(to_slab(k
));
5599 static const struct sysfs_ops slab_sysfs_ops
= {
5600 .show
= slab_attr_show
,
5601 .store
= slab_attr_store
,
5604 static struct kobj_type slab_ktype
= {
5605 .sysfs_ops
= &slab_sysfs_ops
,
5606 .release
= kmem_cache_release
,
5609 static int uevent_filter(struct kset
*kset
, struct kobject
*kobj
)
5611 struct kobj_type
*ktype
= get_ktype(kobj
);
5613 if (ktype
== &slab_ktype
)
5618 static const struct kset_uevent_ops slab_uevent_ops
= {
5619 .filter
= uevent_filter
,
5622 static struct kset
*slab_kset
;
5624 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5627 if (!is_root_cache(s
))
5628 return s
->memcg_params
.root_cache
->memcg_kset
;
5633 #define ID_STR_LENGTH 64
5635 /* Create a unique string id for a slab cache:
5637 * Format :[flags-]size
5639 static char *create_unique_id(struct kmem_cache
*s
)
5641 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5648 * First flags affecting slabcache operations. We will only
5649 * get here for aliasable slabs so we do not need to support
5650 * too many flags. The flags here must cover all flags that
5651 * are matched during merging to guarantee that the id is
5654 if (s
->flags
& SLAB_CACHE_DMA
)
5656 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5658 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5660 if (!(s
->flags
& SLAB_NOTRACK
))
5662 if (s
->flags
& SLAB_ACCOUNT
)
5666 p
+= sprintf(p
, "%07d", s
->size
);
5668 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5672 static void sysfs_slab_remove_workfn(struct work_struct
*work
)
5674 struct kmem_cache
*s
=
5675 container_of(work
, struct kmem_cache
, kobj_remove_work
);
5677 if (!s
->kobj
.state_in_sysfs
)
5679 * For a memcg cache, this may be called during
5680 * deactivation and again on shutdown. Remove only once.
5681 * A cache is never shut down before deactivation is
5682 * complete, so no need to worry about synchronization.
5687 kset_unregister(s
->memcg_kset
);
5689 kobject_uevent(&s
->kobj
, KOBJ_REMOVE
);
5690 kobject_del(&s
->kobj
);
5692 kobject_put(&s
->kobj
);
5695 static int sysfs_slab_add(struct kmem_cache
*s
)
5699 struct kset
*kset
= cache_kset(s
);
5700 int unmergeable
= slab_unmergeable(s
);
5702 INIT_WORK(&s
->kobj_remove_work
, sysfs_slab_remove_workfn
);
5705 kobject_init(&s
->kobj
, &slab_ktype
);
5711 * Slabcache can never be merged so we can use the name proper.
5712 * This is typically the case for debug situations. In that
5713 * case we can catch duplicate names easily.
5715 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5719 * Create a unique name for the slab as a target
5722 name
= create_unique_id(s
);
5725 s
->kobj
.kset
= kset
;
5726 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5730 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5735 if (is_root_cache(s
) && memcg_sysfs_enabled
) {
5736 s
->memcg_kset
= kset_create_and_add("cgroup", NULL
, &s
->kobj
);
5737 if (!s
->memcg_kset
) {
5744 kobject_uevent(&s
->kobj
, KOBJ_ADD
);
5746 /* Setup first alias */
5747 sysfs_slab_alias(s
, s
->name
);
5754 kobject_del(&s
->kobj
);
5758 static void sysfs_slab_remove(struct kmem_cache
*s
)
5760 if (slab_state
< FULL
)
5762 * Sysfs has not been setup yet so no need to remove the
5767 kobject_get(&s
->kobj
);
5768 schedule_work(&s
->kobj_remove_work
);
5771 void sysfs_slab_release(struct kmem_cache
*s
)
5773 if (slab_state
>= FULL
)
5774 kobject_put(&s
->kobj
);
5778 * Need to buffer aliases during bootup until sysfs becomes
5779 * available lest we lose that information.
5781 struct saved_alias
{
5782 struct kmem_cache
*s
;
5784 struct saved_alias
*next
;
5787 static struct saved_alias
*alias_list
;
5789 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5791 struct saved_alias
*al
;
5793 if (slab_state
== FULL
) {
5795 * If we have a leftover link then remove it.
5797 sysfs_remove_link(&slab_kset
->kobj
, name
);
5798 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5801 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5807 al
->next
= alias_list
;
5812 static int __init
slab_sysfs_init(void)
5814 struct kmem_cache
*s
;
5817 mutex_lock(&slab_mutex
);
5819 slab_kset
= kset_create_and_add("slab", &slab_uevent_ops
, kernel_kobj
);
5821 mutex_unlock(&slab_mutex
);
5822 pr_err("Cannot register slab subsystem.\n");
5828 list_for_each_entry(s
, &slab_caches
, list
) {
5829 err
= sysfs_slab_add(s
);
5831 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5835 while (alias_list
) {
5836 struct saved_alias
*al
= alias_list
;
5838 alias_list
= alias_list
->next
;
5839 err
= sysfs_slab_alias(al
->s
, al
->name
);
5841 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5846 mutex_unlock(&slab_mutex
);
5851 __initcall(slab_sysfs_init
);
5852 #endif /* CONFIG_SYSFS */
5855 * The /proc/slabinfo ABI
5857 #ifdef CONFIG_SLUB_DEBUG
5858 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5860 unsigned long nr_slabs
= 0;
5861 unsigned long nr_objs
= 0;
5862 unsigned long nr_free
= 0;
5864 struct kmem_cache_node
*n
;
5866 for_each_kmem_cache_node(s
, node
, n
) {
5867 nr_slabs
+= node_nr_slabs(n
);
5868 nr_objs
+= node_nr_objs(n
);
5869 nr_free
+= count_partial(n
, count_free
);
5872 sinfo
->active_objs
= nr_objs
- nr_free
;
5873 sinfo
->num_objs
= nr_objs
;
5874 sinfo
->active_slabs
= nr_slabs
;
5875 sinfo
->num_slabs
= nr_slabs
;
5876 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5877 sinfo
->cache_order
= oo_order(s
->oo
);
5880 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5884 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5885 size_t count
, loff_t
*ppos
)
5889 #endif /* CONFIG_SLUB_DEBUG */