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 operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/kfence.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list except per cpu partial list. The processor that froze the
63 * slab is the one who can perform list operations on the page. Other
64 * processors may put objects onto the freelist but the processor that
65 * froze the slab is the only one that can retrieve the objects from the
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 * page->frozen The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 #ifdef CONFIG_SLUB_DEBUG
119 #ifdef CONFIG_SLUB_DEBUG_ON
120 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled
);
122 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled
);
126 static inline bool kmem_cache_debug(struct kmem_cache
*s
)
128 return kmem_cache_debug_flags(s
, SLAB_DEBUG_FLAGS
);
131 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
133 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
))
134 p
+= s
->red_left_pad
;
139 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
141 #ifdef CONFIG_SLUB_CPU_PARTIAL
142 return !kmem_cache_debug(s
);
149 * Issues still to be resolved:
151 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
153 * - Variable sizing of the per node arrays
156 /* Enable to test recovery from slab corruption on boot */
157 #undef SLUB_RESILIENCY_TEST
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Minimum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * These debug flags cannot use CMPXCHG because there might be consistency
180 * issues when checking or reading debug information
182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
187 * Debugging flags that require metadata to be stored in the slab. These get
188 * disabled when slub_debug=O is used and a cache's min order increases with
191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
194 #define OO_MASK ((1 << OO_SHIFT) - 1)
195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
197 /* Internal SLUB flags */
199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
200 /* Use cmpxchg_double */
201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
204 * Tracking user of a slab.
206 #define TRACK_ADDRS_COUNT 16
208 unsigned long addr
; /* Called from address */
209 #ifdef CONFIG_STACKTRACE
210 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
212 int cpu
; /* Was running on cpu */
213 int pid
; /* Pid context */
214 unsigned long when
; /* When did the operation occur */
217 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
220 static int sysfs_slab_add(struct kmem_cache
*);
221 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
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
)
228 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
240 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
241 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
242 * differ during memory hotplug/hotremove operations.
243 * Protected by slab_mutex.
245 static nodemask_t slab_nodes
;
247 /********************************************************************
248 * Core slab cache functions
249 *******************************************************************/
252 * Returns freelist pointer (ptr). With hardening, this is obfuscated
253 * with an XOR of the address where the pointer is held and a per-cache
256 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
257 unsigned long ptr_addr
)
259 #ifdef CONFIG_SLAB_FREELIST_HARDENED
261 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
262 * Normally, this doesn't cause any issues, as both set_freepointer()
263 * and get_freepointer() are called with a pointer with the same tag.
264 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
265 * example, when __free_slub() iterates over objects in a cache, it
266 * passes untagged pointers to check_object(). check_object() in turns
267 * calls get_freepointer() with an untagged pointer, which causes the
268 * freepointer to be restored incorrectly.
270 return (void *)((unsigned long)ptr
^ s
->random
^
271 swab((unsigned long)kasan_reset_tag((void *)ptr_addr
)));
277 /* Returns the freelist pointer recorded at location ptr_addr. */
278 static inline void *freelist_dereference(const struct kmem_cache
*s
,
281 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
282 (unsigned long)ptr_addr
);
285 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
287 object
= kasan_reset_tag(object
);
288 return freelist_dereference(s
, object
+ s
->offset
);
291 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
293 prefetch(object
+ s
->offset
);
296 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
298 unsigned long freepointer_addr
;
301 if (!debug_pagealloc_enabled_static())
302 return get_freepointer(s
, object
);
304 freepointer_addr
= (unsigned long)object
+ s
->offset
;
305 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
306 return freelist_ptr(s
, p
, freepointer_addr
);
309 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
311 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
313 #ifdef CONFIG_SLAB_FREELIST_HARDENED
314 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
317 freeptr_addr
= (unsigned long)kasan_reset_tag((void *)freeptr_addr
);
318 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
321 /* Loop over all objects in a slab */
322 #define for_each_object(__p, __s, __addr, __objects) \
323 for (__p = fixup_red_left(__s, __addr); \
324 __p < (__addr) + (__objects) * (__s)->size; \
327 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
329 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
332 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
335 struct kmem_cache_order_objects x
= {
336 (order
<< OO_SHIFT
) + order_objects(order
, size
)
342 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
344 return x
.x
>> OO_SHIFT
;
347 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
349 return x
.x
& OO_MASK
;
353 * Per slab locking using the pagelock
355 static __always_inline
void slab_lock(struct page
*page
)
357 VM_BUG_ON_PAGE(PageTail(page
), page
);
358 bit_spin_lock(PG_locked
, &page
->flags
);
361 static __always_inline
void slab_unlock(struct page
*page
)
363 VM_BUG_ON_PAGE(PageTail(page
), page
);
364 __bit_spin_unlock(PG_locked
, &page
->flags
);
367 /* Interrupts must be disabled (for the fallback code to work right) */
368 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
369 void *freelist_old
, unsigned long counters_old
,
370 void *freelist_new
, unsigned long counters_new
,
373 VM_BUG_ON(!irqs_disabled());
374 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
375 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
376 if (s
->flags
& __CMPXCHG_DOUBLE
) {
377 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
378 freelist_old
, counters_old
,
379 freelist_new
, counters_new
))
385 if (page
->freelist
== freelist_old
&&
386 page
->counters
== counters_old
) {
387 page
->freelist
= freelist_new
;
388 page
->counters
= counters_new
;
396 stat(s
, CMPXCHG_DOUBLE_FAIL
);
398 #ifdef SLUB_DEBUG_CMPXCHG
399 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
405 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
406 void *freelist_old
, unsigned long counters_old
,
407 void *freelist_new
, unsigned long counters_new
,
410 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412 if (s
->flags
& __CMPXCHG_DOUBLE
) {
413 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
414 freelist_old
, counters_old
,
415 freelist_new
, counters_new
))
422 local_irq_save(flags
);
424 if (page
->freelist
== freelist_old
&&
425 page
->counters
== counters_old
) {
426 page
->freelist
= freelist_new
;
427 page
->counters
= counters_new
;
429 local_irq_restore(flags
);
433 local_irq_restore(flags
);
437 stat(s
, CMPXCHG_DOUBLE_FAIL
);
439 #ifdef SLUB_DEBUG_CMPXCHG
440 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
446 #ifdef CONFIG_SLUB_DEBUG
447 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
448 static DEFINE_SPINLOCK(object_map_lock
);
451 * Determine a map of object in use on a page.
453 * Node listlock must be held to guarantee that the page does
454 * not vanish from under us.
456 static unsigned long *get_map(struct kmem_cache
*s
, struct page
*page
)
457 __acquires(&object_map_lock
)
460 void *addr
= page_address(page
);
462 VM_BUG_ON(!irqs_disabled());
464 spin_lock(&object_map_lock
);
466 bitmap_zero(object_map
, page
->objects
);
468 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
469 set_bit(__obj_to_index(s
, addr
, p
), object_map
);
474 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
476 VM_BUG_ON(map
!= object_map
);
477 spin_unlock(&object_map_lock
);
480 static inline unsigned int size_from_object(struct kmem_cache
*s
)
482 if (s
->flags
& SLAB_RED_ZONE
)
483 return s
->size
- s
->red_left_pad
;
488 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
490 if (s
->flags
& SLAB_RED_ZONE
)
491 p
-= s
->red_left_pad
;
499 #if defined(CONFIG_SLUB_DEBUG_ON)
500 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
502 static slab_flags_t slub_debug
;
505 static char *slub_debug_string
;
506 static int disable_higher_order_debug
;
509 * slub is about to manipulate internal object metadata. This memory lies
510 * outside the range of the allocated object, so accessing it would normally
511 * be reported by kasan as a bounds error. metadata_access_enable() is used
512 * to tell kasan that these accesses are OK.
514 static inline void metadata_access_enable(void)
516 kasan_disable_current();
519 static inline void metadata_access_disable(void)
521 kasan_enable_current();
528 /* Verify that a pointer has an address that is valid within a slab page */
529 static inline int check_valid_pointer(struct kmem_cache
*s
,
530 struct page
*page
, void *object
)
537 base
= page_address(page
);
538 object
= kasan_reset_tag(object
);
539 object
= restore_red_left(s
, object
);
540 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
541 (object
- base
) % s
->size
) {
548 static void print_section(char *level
, char *text
, u8
*addr
,
551 metadata_access_enable();
552 print_hex_dump(level
, kasan_reset_tag(text
), DUMP_PREFIX_ADDRESS
,
553 16, 1, addr
, length
, 1);
554 metadata_access_disable();
558 * See comment in calculate_sizes().
560 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
562 return s
->offset
>= s
->inuse
;
566 * Return offset of the end of info block which is inuse + free pointer if
567 * not overlapping with object.
569 static inline unsigned int get_info_end(struct kmem_cache
*s
)
571 if (freeptr_outside_object(s
))
572 return s
->inuse
+ sizeof(void *);
577 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
578 enum track_item alloc
)
582 p
= object
+ get_info_end(s
);
584 return kasan_reset_tag(p
+ alloc
);
587 static void set_track(struct kmem_cache
*s
, void *object
,
588 enum track_item alloc
, unsigned long addr
)
590 struct track
*p
= get_track(s
, object
, alloc
);
593 #ifdef CONFIG_STACKTRACE
594 unsigned int nr_entries
;
596 metadata_access_enable();
597 nr_entries
= stack_trace_save(kasan_reset_tag(p
->addrs
),
598 TRACK_ADDRS_COUNT
, 3);
599 metadata_access_disable();
601 if (nr_entries
< TRACK_ADDRS_COUNT
)
602 p
->addrs
[nr_entries
] = 0;
605 p
->cpu
= smp_processor_id();
606 p
->pid
= current
->pid
;
609 memset(p
, 0, sizeof(struct track
));
613 static void init_tracking(struct kmem_cache
*s
, void *object
)
615 if (!(s
->flags
& SLAB_STORE_USER
))
618 set_track(s
, object
, TRACK_FREE
, 0UL);
619 set_track(s
, object
, TRACK_ALLOC
, 0UL);
622 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
627 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
628 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
629 #ifdef CONFIG_STACKTRACE
632 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
634 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
641 void print_tracking(struct kmem_cache
*s
, void *object
)
643 unsigned long pr_time
= jiffies
;
644 if (!(s
->flags
& SLAB_STORE_USER
))
647 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
648 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
651 static void print_page_info(struct page
*page
)
653 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
654 page
, page
->objects
, page
->inuse
, page
->freelist
,
655 page
->flags
, &page
->flags
);
659 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
661 struct va_format vaf
;
667 pr_err("=============================================================================\n");
668 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
669 pr_err("-----------------------------------------------------------------------------\n\n");
671 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
675 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
677 struct va_format vaf
;
683 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
687 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
688 void **freelist
, void *nextfree
)
690 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
691 !check_valid_pointer(s
, page
, nextfree
) && freelist
) {
692 object_err(s
, page
, *freelist
, "Freechain corrupt");
694 slab_fix(s
, "Isolate corrupted freechain");
701 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
703 unsigned int off
; /* Offset of last byte */
704 u8
*addr
= page_address(page
);
706 print_tracking(s
, p
);
708 print_page_info(page
);
710 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
711 p
, p
- addr
, get_freepointer(s
, p
));
713 if (s
->flags
& SLAB_RED_ZONE
)
714 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
716 else if (p
> addr
+ 16)
717 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
719 print_section(KERN_ERR
, "Object ", p
,
720 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
721 if (s
->flags
& SLAB_RED_ZONE
)
722 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
723 s
->inuse
- s
->object_size
);
725 off
= get_info_end(s
);
727 if (s
->flags
& SLAB_STORE_USER
)
728 off
+= 2 * sizeof(struct track
);
730 off
+= kasan_metadata_size(s
);
732 if (off
!= size_from_object(s
))
733 /* Beginning of the filler is the free pointer */
734 print_section(KERN_ERR
, "Padding ", p
+ off
,
735 size_from_object(s
) - off
);
740 void object_err(struct kmem_cache
*s
, struct page
*page
,
741 u8
*object
, char *reason
)
743 slab_bug(s
, "%s", reason
);
744 print_trailer(s
, page
, object
);
747 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
748 const char *fmt
, ...)
754 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
756 slab_bug(s
, "%s", buf
);
757 print_page_info(page
);
761 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
763 u8
*p
= kasan_reset_tag(object
);
765 if (s
->flags
& SLAB_RED_ZONE
)
766 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
768 if (s
->flags
& __OBJECT_POISON
) {
769 memset(p
, POISON_FREE
, s
->object_size
- 1);
770 p
[s
->object_size
- 1] = POISON_END
;
773 if (s
->flags
& SLAB_RED_ZONE
)
774 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
777 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
778 void *from
, void *to
)
780 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
781 memset(from
, data
, to
- from
);
784 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
785 u8
*object
, char *what
,
786 u8
*start
, unsigned int value
, unsigned int bytes
)
790 u8
*addr
= page_address(page
);
792 metadata_access_enable();
793 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
794 metadata_access_disable();
799 while (end
> fault
&& end
[-1] == value
)
802 slab_bug(s
, "%s overwritten", what
);
803 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
804 fault
, end
- 1, fault
- addr
,
806 print_trailer(s
, page
, object
);
808 restore_bytes(s
, what
, value
, fault
, end
);
816 * Bytes of the object to be managed.
817 * If the freepointer may overlay the object then the free
818 * pointer is at the middle of the object.
820 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
823 * object + s->object_size
824 * Padding to reach word boundary. This is also used for Redzoning.
825 * Padding is extended by another word if Redzoning is enabled and
826 * object_size == inuse.
828 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
829 * 0xcc (RED_ACTIVE) for objects in use.
832 * Meta data starts here.
834 * A. Free pointer (if we cannot overwrite object on free)
835 * B. Tracking data for SLAB_STORE_USER
836 * C. Padding to reach required alignment boundary or at minimum
837 * one word if debugging is on to be able to detect writes
838 * before the word boundary.
840 * Padding is done using 0x5a (POISON_INUSE)
843 * Nothing is used beyond s->size.
845 * If slabcaches are merged then the object_size and inuse boundaries are mostly
846 * ignored. And therefore no slab options that rely on these boundaries
847 * may be used with merged slabcaches.
850 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
852 unsigned long off
= get_info_end(s
); /* The end of info */
854 if (s
->flags
& SLAB_STORE_USER
)
855 /* We also have user information there */
856 off
+= 2 * sizeof(struct track
);
858 off
+= kasan_metadata_size(s
);
860 if (size_from_object(s
) == off
)
863 return check_bytes_and_report(s
, page
, p
, "Object padding",
864 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
867 /* Check the pad bytes at the end of a slab page */
868 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
877 if (!(s
->flags
& SLAB_POISON
))
880 start
= page_address(page
);
881 length
= page_size(page
);
882 end
= start
+ length
;
883 remainder
= length
% s
->size
;
887 pad
= end
- remainder
;
888 metadata_access_enable();
889 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
890 metadata_access_disable();
893 while (end
> fault
&& end
[-1] == POISON_INUSE
)
896 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
897 fault
, end
- 1, fault
- start
);
898 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
900 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
904 static int check_object(struct kmem_cache
*s
, struct page
*page
,
905 void *object
, u8 val
)
908 u8
*endobject
= object
+ s
->object_size
;
910 if (s
->flags
& SLAB_RED_ZONE
) {
911 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
912 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
915 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
916 endobject
, val
, s
->inuse
- s
->object_size
))
919 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
920 check_bytes_and_report(s
, page
, p
, "Alignment padding",
921 endobject
, POISON_INUSE
,
922 s
->inuse
- s
->object_size
);
926 if (s
->flags
& SLAB_POISON
) {
927 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
928 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
929 POISON_FREE
, s
->object_size
- 1) ||
930 !check_bytes_and_report(s
, page
, p
, "Poison",
931 p
+ s
->object_size
- 1, POISON_END
, 1)))
934 * check_pad_bytes cleans up on its own.
936 check_pad_bytes(s
, page
, p
);
939 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
941 * Object and freepointer overlap. Cannot check
942 * freepointer while object is allocated.
946 /* Check free pointer validity */
947 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
948 object_err(s
, page
, p
, "Freepointer corrupt");
950 * No choice but to zap it and thus lose the remainder
951 * of the free objects in this slab. May cause
952 * another error because the object count is now wrong.
954 set_freepointer(s
, p
, NULL
);
960 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
964 VM_BUG_ON(!irqs_disabled());
966 if (!PageSlab(page
)) {
967 slab_err(s
, page
, "Not a valid slab page");
971 maxobj
= order_objects(compound_order(page
), s
->size
);
972 if (page
->objects
> maxobj
) {
973 slab_err(s
, page
, "objects %u > max %u",
974 page
->objects
, maxobj
);
977 if (page
->inuse
> page
->objects
) {
978 slab_err(s
, page
, "inuse %u > max %u",
979 page
->inuse
, page
->objects
);
982 /* Slab_pad_check fixes things up after itself */
983 slab_pad_check(s
, page
);
988 * Determine if a certain object on a page is on the freelist. Must hold the
989 * slab lock to guarantee that the chains are in a consistent state.
991 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
999 while (fp
&& nr
<= page
->objects
) {
1002 if (!check_valid_pointer(s
, page
, fp
)) {
1004 object_err(s
, page
, object
,
1005 "Freechain corrupt");
1006 set_freepointer(s
, object
, NULL
);
1008 slab_err(s
, page
, "Freepointer corrupt");
1009 page
->freelist
= NULL
;
1010 page
->inuse
= page
->objects
;
1011 slab_fix(s
, "Freelist cleared");
1017 fp
= get_freepointer(s
, object
);
1021 max_objects
= order_objects(compound_order(page
), s
->size
);
1022 if (max_objects
> MAX_OBJS_PER_PAGE
)
1023 max_objects
= MAX_OBJS_PER_PAGE
;
1025 if (page
->objects
!= max_objects
) {
1026 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1027 page
->objects
, max_objects
);
1028 page
->objects
= max_objects
;
1029 slab_fix(s
, "Number of objects adjusted.");
1031 if (page
->inuse
!= page
->objects
- nr
) {
1032 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1033 page
->inuse
, page
->objects
- nr
);
1034 page
->inuse
= page
->objects
- nr
;
1035 slab_fix(s
, "Object count adjusted.");
1037 return search
== NULL
;
1040 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1043 if (s
->flags
& SLAB_TRACE
) {
1044 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1046 alloc
? "alloc" : "free",
1047 object
, page
->inuse
,
1051 print_section(KERN_INFO
, "Object ", (void *)object
,
1059 * Tracking of fully allocated slabs for debugging purposes.
1061 static void add_full(struct kmem_cache
*s
,
1062 struct kmem_cache_node
*n
, struct page
*page
)
1064 if (!(s
->flags
& SLAB_STORE_USER
))
1067 lockdep_assert_held(&n
->list_lock
);
1068 list_add(&page
->slab_list
, &n
->full
);
1071 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1073 if (!(s
->flags
& SLAB_STORE_USER
))
1076 lockdep_assert_held(&n
->list_lock
);
1077 list_del(&page
->slab_list
);
1080 /* Tracking of the number of slabs for debugging purposes */
1081 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1083 struct kmem_cache_node
*n
= get_node(s
, node
);
1085 return atomic_long_read(&n
->nr_slabs
);
1088 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1090 return atomic_long_read(&n
->nr_slabs
);
1093 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1095 struct kmem_cache_node
*n
= get_node(s
, node
);
1098 * May be called early in order to allocate a slab for the
1099 * kmem_cache_node structure. Solve the chicken-egg
1100 * dilemma by deferring the increment of the count during
1101 * bootstrap (see early_kmem_cache_node_alloc).
1104 atomic_long_inc(&n
->nr_slabs
);
1105 atomic_long_add(objects
, &n
->total_objects
);
1108 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1110 struct kmem_cache_node
*n
= get_node(s
, node
);
1112 atomic_long_dec(&n
->nr_slabs
);
1113 atomic_long_sub(objects
, &n
->total_objects
);
1116 /* Object debug checks for alloc/free paths */
1117 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1120 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1123 init_object(s
, object
, SLUB_RED_INACTIVE
);
1124 init_tracking(s
, object
);
1128 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1130 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1133 metadata_access_enable();
1134 memset(kasan_reset_tag(addr
), POISON_INUSE
, page_size(page
));
1135 metadata_access_disable();
1138 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1139 struct page
*page
, void *object
)
1141 if (!check_slab(s
, page
))
1144 if (!check_valid_pointer(s
, page
, object
)) {
1145 object_err(s
, page
, object
, "Freelist Pointer check fails");
1149 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1155 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1157 void *object
, unsigned long addr
)
1159 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1160 if (!alloc_consistency_checks(s
, page
, object
))
1164 /* Success perform special debug activities for allocs */
1165 if (s
->flags
& SLAB_STORE_USER
)
1166 set_track(s
, object
, TRACK_ALLOC
, addr
);
1167 trace(s
, page
, object
, 1);
1168 init_object(s
, object
, SLUB_RED_ACTIVE
);
1172 if (PageSlab(page
)) {
1174 * If this is a slab page then lets do the best we can
1175 * to avoid issues in the future. Marking all objects
1176 * as used avoids touching the remaining objects.
1178 slab_fix(s
, "Marking all objects used");
1179 page
->inuse
= page
->objects
;
1180 page
->freelist
= NULL
;
1185 static inline int free_consistency_checks(struct kmem_cache
*s
,
1186 struct page
*page
, void *object
, unsigned long addr
)
1188 if (!check_valid_pointer(s
, page
, object
)) {
1189 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1193 if (on_freelist(s
, page
, object
)) {
1194 object_err(s
, page
, object
, "Object already free");
1198 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1201 if (unlikely(s
!= page
->slab_cache
)) {
1202 if (!PageSlab(page
)) {
1203 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1205 } else if (!page
->slab_cache
) {
1206 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1210 object_err(s
, page
, object
,
1211 "page slab pointer corrupt.");
1217 /* Supports checking bulk free of a constructed freelist */
1218 static noinline
int free_debug_processing(
1219 struct kmem_cache
*s
, struct page
*page
,
1220 void *head
, void *tail
, int bulk_cnt
,
1223 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1224 void *object
= head
;
1226 unsigned long flags
;
1229 spin_lock_irqsave(&n
->list_lock
, flags
);
1232 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1233 if (!check_slab(s
, page
))
1240 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1241 if (!free_consistency_checks(s
, page
, object
, addr
))
1245 if (s
->flags
& SLAB_STORE_USER
)
1246 set_track(s
, object
, TRACK_FREE
, addr
);
1247 trace(s
, page
, object
, 0);
1248 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1249 init_object(s
, object
, SLUB_RED_INACTIVE
);
1251 /* Reached end of constructed freelist yet? */
1252 if (object
!= tail
) {
1253 object
= get_freepointer(s
, object
);
1259 if (cnt
!= bulk_cnt
)
1260 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1264 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1266 slab_fix(s
, "Object at 0x%p not freed", object
);
1271 * Parse a block of slub_debug options. Blocks are delimited by ';'
1273 * @str: start of block
1274 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1275 * @slabs: return start of list of slabs, or NULL when there's no list
1276 * @init: assume this is initial parsing and not per-kmem-create parsing
1278 * returns the start of next block if there's any, or NULL
1281 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1283 bool higher_order_disable
= false;
1285 /* Skip any completely empty blocks */
1286 while (*str
&& *str
== ';')
1291 * No options but restriction on slabs. This means full
1292 * debugging for slabs matching a pattern.
1294 *flags
= DEBUG_DEFAULT_FLAGS
;
1299 /* Determine which debug features should be switched on */
1300 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1301 switch (tolower(*str
)) {
1306 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1309 *flags
|= SLAB_RED_ZONE
;
1312 *flags
|= SLAB_POISON
;
1315 *flags
|= SLAB_STORE_USER
;
1318 *flags
|= SLAB_TRACE
;
1321 *flags
|= SLAB_FAILSLAB
;
1325 * Avoid enabling debugging on caches if its minimum
1326 * order would increase as a result.
1328 higher_order_disable
= true;
1332 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1341 /* Skip over the slab list */
1342 while (*str
&& *str
!= ';')
1345 /* Skip any completely empty blocks */
1346 while (*str
&& *str
== ';')
1349 if (init
&& higher_order_disable
)
1350 disable_higher_order_debug
= 1;
1358 static int __init
setup_slub_debug(char *str
)
1363 bool global_slub_debug_changed
= false;
1364 bool slab_list_specified
= false;
1366 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1367 if (*str
++ != '=' || !*str
)
1369 * No options specified. Switch on full debugging.
1375 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1379 global_slub_debug_changed
= true;
1381 slab_list_specified
= true;
1386 * For backwards compatibility, a single list of flags with list of
1387 * slabs means debugging is only enabled for those slabs, so the global
1388 * slub_debug should be 0. We can extended that to multiple lists as
1389 * long as there is no option specifying flags without a slab list.
1391 if (slab_list_specified
) {
1392 if (!global_slub_debug_changed
)
1394 slub_debug_string
= saved_str
;
1397 if (slub_debug
!= 0 || slub_debug_string
)
1398 static_branch_enable(&slub_debug_enabled
);
1399 if ((static_branch_unlikely(&init_on_alloc
) ||
1400 static_branch_unlikely(&init_on_free
)) &&
1401 (slub_debug
& SLAB_POISON
))
1402 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1406 __setup("slub_debug", setup_slub_debug
);
1409 * kmem_cache_flags - apply debugging options to the cache
1410 * @object_size: the size of an object without meta data
1411 * @flags: flags to set
1412 * @name: name of the cache
1414 * Debug option(s) are applied to @flags. In addition to the debug
1415 * option(s), if a slab name (or multiple) is specified i.e.
1416 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1417 * then only the select slabs will receive the debug option(s).
1419 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1420 slab_flags_t flags
, const char *name
)
1425 slab_flags_t block_flags
;
1426 slab_flags_t slub_debug_local
= slub_debug
;
1429 * If the slab cache is for debugging (e.g. kmemleak) then
1430 * don't store user (stack trace) information by default,
1431 * but let the user enable it via the command line below.
1433 if (flags
& SLAB_NOLEAKTRACE
)
1434 slub_debug_local
&= ~SLAB_STORE_USER
;
1437 next_block
= slub_debug_string
;
1438 /* Go through all blocks of debug options, see if any matches our slab's name */
1439 while (next_block
) {
1440 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1443 /* Found a block that has a slab list, search it */
1448 end
= strchrnul(iter
, ',');
1449 if (next_block
&& next_block
< end
)
1450 end
= next_block
- 1;
1452 glob
= strnchr(iter
, end
- iter
, '*');
1454 cmplen
= glob
- iter
;
1456 cmplen
= max_t(size_t, len
, (end
- iter
));
1458 if (!strncmp(name
, iter
, cmplen
)) {
1459 flags
|= block_flags
;
1463 if (!*end
|| *end
== ';')
1469 return flags
| slub_debug_local
;
1471 #else /* !CONFIG_SLUB_DEBUG */
1472 static inline void setup_object_debug(struct kmem_cache
*s
,
1473 struct page
*page
, void *object
) {}
1475 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1477 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1478 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1480 static inline int free_debug_processing(
1481 struct kmem_cache
*s
, struct page
*page
,
1482 void *head
, void *tail
, int bulk_cnt
,
1483 unsigned long addr
) { return 0; }
1485 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1487 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1488 void *object
, u8 val
) { return 1; }
1489 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1490 struct page
*page
) {}
1491 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1492 struct page
*page
) {}
1493 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1494 slab_flags_t flags
, const char *name
)
1498 #define slub_debug 0
1500 #define disable_higher_order_debug 0
1502 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1504 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1506 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1508 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1511 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1512 void **freelist
, void *nextfree
)
1516 #endif /* CONFIG_SLUB_DEBUG */
1519 * Hooks for other subsystems that check memory allocations. In a typical
1520 * production configuration these hooks all should produce no code at all.
1522 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1524 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1525 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1526 kmemleak_alloc(ptr
, size
, 1, flags
);
1530 static __always_inline
void kfree_hook(void *x
)
1533 kasan_kfree_large(x
);
1536 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
,
1539 kmemleak_free_recursive(x
, s
->flags
);
1542 * Trouble is that we may no longer disable interrupts in the fast path
1543 * So in order to make the debug calls that expect irqs to be
1544 * disabled we need to disable interrupts temporarily.
1546 #ifdef CONFIG_LOCKDEP
1548 unsigned long flags
;
1550 local_irq_save(flags
);
1551 debug_check_no_locks_freed(x
, s
->object_size
);
1552 local_irq_restore(flags
);
1555 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1556 debug_check_no_obj_freed(x
, s
->object_size
);
1558 /* Use KCSAN to help debug racy use-after-free. */
1559 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1560 __kcsan_check_access(x
, s
->object_size
,
1561 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1564 * As memory initialization might be integrated into KASAN,
1565 * kasan_slab_free and initialization memset's must be
1566 * kept together to avoid discrepancies in behavior.
1568 * The initialization memset's clear the object and the metadata,
1569 * but don't touch the SLAB redzone.
1574 if (!kasan_has_integrated_init())
1575 memset(kasan_reset_tag(x
), 0, s
->object_size
);
1576 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
: 0;
1577 memset((char *)kasan_reset_tag(x
) + s
->inuse
, 0,
1578 s
->size
- s
->inuse
- rsize
);
1580 /* KASAN might put x into memory quarantine, delaying its reuse. */
1581 return kasan_slab_free(s
, x
, init
);
1584 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1585 void **head
, void **tail
)
1590 void *old_tail
= *tail
? *tail
: *head
;
1592 if (is_kfence_address(next
)) {
1593 slab_free_hook(s
, next
, false);
1597 /* Head and tail of the reconstructed freelist */
1603 next
= get_freepointer(s
, object
);
1605 /* If object's reuse doesn't have to be delayed */
1606 if (!slab_free_hook(s
, object
, slab_want_init_on_free(s
))) {
1607 /* Move object to the new freelist */
1608 set_freepointer(s
, object
, *head
);
1613 } while (object
!= old_tail
);
1618 return *head
!= NULL
;
1621 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1624 setup_object_debug(s
, page
, object
);
1625 object
= kasan_init_slab_obj(s
, object
);
1626 if (unlikely(s
->ctor
)) {
1627 kasan_unpoison_object_data(s
, object
);
1629 kasan_poison_object_data(s
, object
);
1635 * Slab allocation and freeing
1637 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1638 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1641 unsigned int order
= oo_order(oo
);
1643 if (node
== NUMA_NO_NODE
)
1644 page
= alloc_pages(flags
, order
);
1646 page
= __alloc_pages_node(node
, flags
, order
);
1651 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1652 /* Pre-initialize the random sequence cache */
1653 static int init_cache_random_seq(struct kmem_cache
*s
)
1655 unsigned int count
= oo_objects(s
->oo
);
1658 /* Bailout if already initialised */
1662 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1664 pr_err("SLUB: Unable to initialize free list for %s\n",
1669 /* Transform to an offset on the set of pages */
1670 if (s
->random_seq
) {
1673 for (i
= 0; i
< count
; i
++)
1674 s
->random_seq
[i
] *= s
->size
;
1679 /* Initialize each random sequence freelist per cache */
1680 static void __init
init_freelist_randomization(void)
1682 struct kmem_cache
*s
;
1684 mutex_lock(&slab_mutex
);
1686 list_for_each_entry(s
, &slab_caches
, list
)
1687 init_cache_random_seq(s
);
1689 mutex_unlock(&slab_mutex
);
1692 /* Get the next entry on the pre-computed freelist randomized */
1693 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1694 unsigned long *pos
, void *start
,
1695 unsigned long page_limit
,
1696 unsigned long freelist_count
)
1701 * If the target page allocation failed, the number of objects on the
1702 * page might be smaller than the usual size defined by the cache.
1705 idx
= s
->random_seq
[*pos
];
1707 if (*pos
>= freelist_count
)
1709 } while (unlikely(idx
>= page_limit
));
1711 return (char *)start
+ idx
;
1714 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1715 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1720 unsigned long idx
, pos
, page_limit
, freelist_count
;
1722 if (page
->objects
< 2 || !s
->random_seq
)
1725 freelist_count
= oo_objects(s
->oo
);
1726 pos
= get_random_int() % freelist_count
;
1728 page_limit
= page
->objects
* s
->size
;
1729 start
= fixup_red_left(s
, page_address(page
));
1731 /* First entry is used as the base of the freelist */
1732 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1734 cur
= setup_object(s
, page
, cur
);
1735 page
->freelist
= cur
;
1737 for (idx
= 1; idx
< page
->objects
; idx
++) {
1738 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1740 next
= setup_object(s
, page
, next
);
1741 set_freepointer(s
, cur
, next
);
1744 set_freepointer(s
, cur
, NULL
);
1749 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1753 static inline void init_freelist_randomization(void) { }
1754 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1758 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1760 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1763 struct kmem_cache_order_objects oo
= s
->oo
;
1765 void *start
, *p
, *next
;
1769 flags
&= gfp_allowed_mask
;
1771 if (gfpflags_allow_blocking(flags
))
1774 flags
|= s
->allocflags
;
1777 * Let the initial higher-order allocation fail under memory pressure
1778 * so we fall-back to the minimum order allocation.
1780 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1781 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1782 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1784 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1785 if (unlikely(!page
)) {
1789 * Allocation may have failed due to fragmentation.
1790 * Try a lower order alloc if possible
1792 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1793 if (unlikely(!page
))
1795 stat(s
, ORDER_FALLBACK
);
1798 page
->objects
= oo_objects(oo
);
1800 account_slab_page(page
, oo_order(oo
), s
, flags
);
1802 page
->slab_cache
= s
;
1803 __SetPageSlab(page
);
1804 if (page_is_pfmemalloc(page
))
1805 SetPageSlabPfmemalloc(page
);
1807 kasan_poison_slab(page
);
1809 start
= page_address(page
);
1811 setup_page_debug(s
, page
, start
);
1813 shuffle
= shuffle_freelist(s
, page
);
1816 start
= fixup_red_left(s
, start
);
1817 start
= setup_object(s
, page
, start
);
1818 page
->freelist
= start
;
1819 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1821 next
= setup_object(s
, page
, next
);
1822 set_freepointer(s
, p
, next
);
1825 set_freepointer(s
, p
, NULL
);
1828 page
->inuse
= page
->objects
;
1832 if (gfpflags_allow_blocking(flags
))
1833 local_irq_disable();
1837 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1842 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1844 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1845 flags
= kmalloc_fix_flags(flags
);
1847 return allocate_slab(s
,
1848 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1851 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1853 int order
= compound_order(page
);
1854 int pages
= 1 << order
;
1856 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
1859 slab_pad_check(s
, page
);
1860 for_each_object(p
, s
, page_address(page
),
1862 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1865 __ClearPageSlabPfmemalloc(page
);
1866 __ClearPageSlab(page
);
1867 /* In union with page->mapping where page allocator expects NULL */
1868 page
->slab_cache
= NULL
;
1869 if (current
->reclaim_state
)
1870 current
->reclaim_state
->reclaimed_slab
+= pages
;
1871 unaccount_slab_page(page
, order
, s
);
1872 __free_pages(page
, order
);
1875 static void rcu_free_slab(struct rcu_head
*h
)
1877 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1879 __free_slab(page
->slab_cache
, page
);
1882 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1884 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1885 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1887 __free_slab(s
, page
);
1890 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1892 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1897 * Management of partially allocated slabs.
1900 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1903 if (tail
== DEACTIVATE_TO_TAIL
)
1904 list_add_tail(&page
->slab_list
, &n
->partial
);
1906 list_add(&page
->slab_list
, &n
->partial
);
1909 static inline void add_partial(struct kmem_cache_node
*n
,
1910 struct page
*page
, int tail
)
1912 lockdep_assert_held(&n
->list_lock
);
1913 __add_partial(n
, page
, tail
);
1916 static inline void remove_partial(struct kmem_cache_node
*n
,
1919 lockdep_assert_held(&n
->list_lock
);
1920 list_del(&page
->slab_list
);
1925 * Remove slab from the partial list, freeze it and
1926 * return the pointer to the freelist.
1928 * Returns a list of objects or NULL if it fails.
1930 static inline void *acquire_slab(struct kmem_cache
*s
,
1931 struct kmem_cache_node
*n
, struct page
*page
,
1932 int mode
, int *objects
)
1935 unsigned long counters
;
1938 lockdep_assert_held(&n
->list_lock
);
1941 * Zap the freelist and set the frozen bit.
1942 * The old freelist is the list of objects for the
1943 * per cpu allocation list.
1945 freelist
= page
->freelist
;
1946 counters
= page
->counters
;
1947 new.counters
= counters
;
1948 *objects
= new.objects
- new.inuse
;
1950 new.inuse
= page
->objects
;
1951 new.freelist
= NULL
;
1953 new.freelist
= freelist
;
1956 VM_BUG_ON(new.frozen
);
1959 if (!__cmpxchg_double_slab(s
, page
,
1961 new.freelist
, new.counters
,
1965 remove_partial(n
, page
);
1970 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1971 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1974 * Try to allocate a partial slab from a specific node.
1976 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1977 struct kmem_cache_cpu
*c
, gfp_t flags
)
1979 struct page
*page
, *page2
;
1980 void *object
= NULL
;
1981 unsigned int available
= 0;
1985 * Racy check. If we mistakenly see no partial slabs then we
1986 * just allocate an empty slab. If we mistakenly try to get a
1987 * partial slab and there is none available then get_partial()
1990 if (!n
|| !n
->nr_partial
)
1993 spin_lock(&n
->list_lock
);
1994 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1997 if (!pfmemalloc_match(page
, flags
))
2000 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
2004 available
+= objects
;
2007 stat(s
, ALLOC_FROM_PARTIAL
);
2010 put_cpu_partial(s
, page
, 0);
2011 stat(s
, CPU_PARTIAL_NODE
);
2013 if (!kmem_cache_has_cpu_partial(s
)
2014 || available
> slub_cpu_partial(s
) / 2)
2018 spin_unlock(&n
->list_lock
);
2023 * Get a page from somewhere. Search in increasing NUMA distances.
2025 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2026 struct kmem_cache_cpu
*c
)
2029 struct zonelist
*zonelist
;
2032 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2034 unsigned int cpuset_mems_cookie
;
2037 * The defrag ratio allows a configuration of the tradeoffs between
2038 * inter node defragmentation and node local allocations. A lower
2039 * defrag_ratio increases the tendency to do local allocations
2040 * instead of attempting to obtain partial slabs from other nodes.
2042 * If the defrag_ratio is set to 0 then kmalloc() always
2043 * returns node local objects. If the ratio is higher then kmalloc()
2044 * may return off node objects because partial slabs are obtained
2045 * from other nodes and filled up.
2047 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2048 * (which makes defrag_ratio = 1000) then every (well almost)
2049 * allocation will first attempt to defrag slab caches on other nodes.
2050 * This means scanning over all nodes to look for partial slabs which
2051 * may be expensive if we do it every time we are trying to find a slab
2052 * with available objects.
2054 if (!s
->remote_node_defrag_ratio
||
2055 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2059 cpuset_mems_cookie
= read_mems_allowed_begin();
2060 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2061 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2062 struct kmem_cache_node
*n
;
2064 n
= get_node(s
, zone_to_nid(zone
));
2066 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2067 n
->nr_partial
> s
->min_partial
) {
2068 object
= get_partial_node(s
, n
, c
, flags
);
2071 * Don't check read_mems_allowed_retry()
2072 * here - if mems_allowed was updated in
2073 * parallel, that was a harmless race
2074 * between allocation and the cpuset
2081 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2082 #endif /* CONFIG_NUMA */
2087 * Get a partial page, lock it and return it.
2089 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2090 struct kmem_cache_cpu
*c
)
2093 int searchnode
= node
;
2095 if (node
== NUMA_NO_NODE
)
2096 searchnode
= numa_mem_id();
2098 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
2099 if (object
|| node
!= NUMA_NO_NODE
)
2102 return get_any_partial(s
, flags
, c
);
2105 #ifdef CONFIG_PREEMPTION
2107 * Calculate the next globally unique transaction for disambiguation
2108 * during cmpxchg. The transactions start with the cpu number and are then
2109 * incremented by CONFIG_NR_CPUS.
2111 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2114 * No preemption supported therefore also no need to check for
2120 static inline unsigned long next_tid(unsigned long tid
)
2122 return tid
+ TID_STEP
;
2125 #ifdef SLUB_DEBUG_CMPXCHG
2126 static inline unsigned int tid_to_cpu(unsigned long tid
)
2128 return tid
% TID_STEP
;
2131 static inline unsigned long tid_to_event(unsigned long tid
)
2133 return tid
/ TID_STEP
;
2137 static inline unsigned int init_tid(int cpu
)
2142 static inline void note_cmpxchg_failure(const char *n
,
2143 const struct kmem_cache
*s
, unsigned long tid
)
2145 #ifdef SLUB_DEBUG_CMPXCHG
2146 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2148 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2150 #ifdef CONFIG_PREEMPTION
2151 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2152 pr_warn("due to cpu change %d -> %d\n",
2153 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2156 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2157 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2158 tid_to_event(tid
), tid_to_event(actual_tid
));
2160 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2161 actual_tid
, tid
, next_tid(tid
));
2163 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2166 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2170 for_each_possible_cpu(cpu
)
2171 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2175 * Remove the cpu slab
2177 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2178 void *freelist
, struct kmem_cache_cpu
*c
)
2180 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2181 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2182 int lock
= 0, free_delta
= 0;
2183 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2184 void *nextfree
, *freelist_iter
, *freelist_tail
;
2185 int tail
= DEACTIVATE_TO_HEAD
;
2189 if (page
->freelist
) {
2190 stat(s
, DEACTIVATE_REMOTE_FREES
);
2191 tail
= DEACTIVATE_TO_TAIL
;
2195 * Stage one: Count the objects on cpu's freelist as free_delta and
2196 * remember the last object in freelist_tail for later splicing.
2198 freelist_tail
= NULL
;
2199 freelist_iter
= freelist
;
2200 while (freelist_iter
) {
2201 nextfree
= get_freepointer(s
, freelist_iter
);
2204 * If 'nextfree' is invalid, it is possible that the object at
2205 * 'freelist_iter' is already corrupted. So isolate all objects
2206 * starting at 'freelist_iter' by skipping them.
2208 if (freelist_corrupted(s
, page
, &freelist_iter
, nextfree
))
2211 freelist_tail
= freelist_iter
;
2214 freelist_iter
= nextfree
;
2218 * Stage two: Unfreeze the page while splicing the per-cpu
2219 * freelist to the head of page's freelist.
2221 * Ensure that the page is unfrozen while the list presence
2222 * reflects the actual number of objects during unfreeze.
2224 * We setup the list membership and then perform a cmpxchg
2225 * with the count. If there is a mismatch then the page
2226 * is not unfrozen but the page is on the wrong list.
2228 * Then we restart the process which may have to remove
2229 * the page from the list that we just put it on again
2230 * because the number of objects in the slab may have
2235 old
.freelist
= READ_ONCE(page
->freelist
);
2236 old
.counters
= READ_ONCE(page
->counters
);
2237 VM_BUG_ON(!old
.frozen
);
2239 /* Determine target state of the slab */
2240 new.counters
= old
.counters
;
2241 if (freelist_tail
) {
2242 new.inuse
-= free_delta
;
2243 set_freepointer(s
, freelist_tail
, old
.freelist
);
2244 new.freelist
= freelist
;
2246 new.freelist
= old
.freelist
;
2250 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2252 else if (new.freelist
) {
2257 * Taking the spinlock removes the possibility
2258 * that acquire_slab() will see a slab page that
2261 spin_lock(&n
->list_lock
);
2265 if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
) && !lock
) {
2268 * This also ensures that the scanning of full
2269 * slabs from diagnostic functions will not see
2272 spin_lock(&n
->list_lock
);
2278 remove_partial(n
, page
);
2279 else if (l
== M_FULL
)
2280 remove_full(s
, n
, page
);
2283 add_partial(n
, page
, tail
);
2284 else if (m
== M_FULL
)
2285 add_full(s
, n
, page
);
2289 if (!__cmpxchg_double_slab(s
, page
,
2290 old
.freelist
, old
.counters
,
2291 new.freelist
, new.counters
,
2296 spin_unlock(&n
->list_lock
);
2300 else if (m
== M_FULL
)
2301 stat(s
, DEACTIVATE_FULL
);
2302 else if (m
== M_FREE
) {
2303 stat(s
, DEACTIVATE_EMPTY
);
2304 discard_slab(s
, page
);
2313 * Unfreeze all the cpu partial slabs.
2315 * This function must be called with interrupts disabled
2316 * for the cpu using c (or some other guarantee must be there
2317 * to guarantee no concurrent accesses).
2319 static void unfreeze_partials(struct kmem_cache
*s
,
2320 struct kmem_cache_cpu
*c
)
2322 #ifdef CONFIG_SLUB_CPU_PARTIAL
2323 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2324 struct page
*page
, *discard_page
= NULL
;
2326 while ((page
= slub_percpu_partial(c
))) {
2330 slub_set_percpu_partial(c
, page
);
2332 n2
= get_node(s
, page_to_nid(page
));
2335 spin_unlock(&n
->list_lock
);
2338 spin_lock(&n
->list_lock
);
2343 old
.freelist
= page
->freelist
;
2344 old
.counters
= page
->counters
;
2345 VM_BUG_ON(!old
.frozen
);
2347 new.counters
= old
.counters
;
2348 new.freelist
= old
.freelist
;
2352 } while (!__cmpxchg_double_slab(s
, page
,
2353 old
.freelist
, old
.counters
,
2354 new.freelist
, new.counters
,
2355 "unfreezing slab"));
2357 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2358 page
->next
= discard_page
;
2359 discard_page
= page
;
2361 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2362 stat(s
, FREE_ADD_PARTIAL
);
2367 spin_unlock(&n
->list_lock
);
2369 while (discard_page
) {
2370 page
= discard_page
;
2371 discard_page
= discard_page
->next
;
2373 stat(s
, DEACTIVATE_EMPTY
);
2374 discard_slab(s
, page
);
2377 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2381 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2382 * partial page slot if available.
2384 * If we did not find a slot then simply move all the partials to the
2385 * per node partial list.
2387 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2389 #ifdef CONFIG_SLUB_CPU_PARTIAL
2390 struct page
*oldpage
;
2398 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2401 pobjects
= oldpage
->pobjects
;
2402 pages
= oldpage
->pages
;
2403 if (drain
&& pobjects
> slub_cpu_partial(s
)) {
2404 unsigned long flags
;
2406 * partial array is full. Move the existing
2407 * set to the per node partial list.
2409 local_irq_save(flags
);
2410 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2411 local_irq_restore(flags
);
2415 stat(s
, CPU_PARTIAL_DRAIN
);
2420 pobjects
+= page
->objects
- page
->inuse
;
2422 page
->pages
= pages
;
2423 page
->pobjects
= pobjects
;
2424 page
->next
= oldpage
;
2426 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2428 if (unlikely(!slub_cpu_partial(s
))) {
2429 unsigned long flags
;
2431 local_irq_save(flags
);
2432 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2433 local_irq_restore(flags
);
2436 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2439 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2441 stat(s
, CPUSLAB_FLUSH
);
2442 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2444 c
->tid
= next_tid(c
->tid
);
2450 * Called from IPI handler with interrupts disabled.
2452 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2454 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2459 unfreeze_partials(s
, c
);
2462 static void flush_cpu_slab(void *d
)
2464 struct kmem_cache
*s
= d
;
2466 __flush_cpu_slab(s
, smp_processor_id());
2469 static bool has_cpu_slab(int cpu
, void *info
)
2471 struct kmem_cache
*s
= info
;
2472 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2474 return c
->page
|| slub_percpu_partial(c
);
2477 static void flush_all(struct kmem_cache
*s
)
2479 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1);
2483 * Use the cpu notifier to insure that the cpu slabs are flushed when
2486 static int slub_cpu_dead(unsigned int cpu
)
2488 struct kmem_cache
*s
;
2489 unsigned long flags
;
2491 mutex_lock(&slab_mutex
);
2492 list_for_each_entry(s
, &slab_caches
, list
) {
2493 local_irq_save(flags
);
2494 __flush_cpu_slab(s
, cpu
);
2495 local_irq_restore(flags
);
2497 mutex_unlock(&slab_mutex
);
2502 * Check if the objects in a per cpu structure fit numa
2503 * locality expectations.
2505 static inline int node_match(struct page
*page
, int node
)
2508 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2514 #ifdef CONFIG_SLUB_DEBUG
2515 static int count_free(struct page
*page
)
2517 return page
->objects
- page
->inuse
;
2520 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2522 return atomic_long_read(&n
->total_objects
);
2524 #endif /* CONFIG_SLUB_DEBUG */
2526 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2527 static unsigned long count_partial(struct kmem_cache_node
*n
,
2528 int (*get_count
)(struct page
*))
2530 unsigned long flags
;
2531 unsigned long x
= 0;
2534 spin_lock_irqsave(&n
->list_lock
, flags
);
2535 list_for_each_entry(page
, &n
->partial
, slab_list
)
2536 x
+= get_count(page
);
2537 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2540 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2542 static noinline
void
2543 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2545 #ifdef CONFIG_SLUB_DEBUG
2546 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2547 DEFAULT_RATELIMIT_BURST
);
2549 struct kmem_cache_node
*n
;
2551 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2554 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2555 nid
, gfpflags
, &gfpflags
);
2556 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2557 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2560 if (oo_order(s
->min
) > get_order(s
->object_size
))
2561 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2564 for_each_kmem_cache_node(s
, node
, n
) {
2565 unsigned long nr_slabs
;
2566 unsigned long nr_objs
;
2567 unsigned long nr_free
;
2569 nr_free
= count_partial(n
, count_free
);
2570 nr_slabs
= node_nr_slabs(n
);
2571 nr_objs
= node_nr_objs(n
);
2573 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2574 node
, nr_slabs
, nr_objs
, nr_free
);
2579 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2580 int node
, struct kmem_cache_cpu
**pc
)
2583 struct kmem_cache_cpu
*c
= *pc
;
2586 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2588 freelist
= get_partial(s
, flags
, node
, c
);
2593 page
= new_slab(s
, flags
, node
);
2595 c
= raw_cpu_ptr(s
->cpu_slab
);
2600 * No other reference to the page yet so we can
2601 * muck around with it freely without cmpxchg
2603 freelist
= page
->freelist
;
2604 page
->freelist
= NULL
;
2606 stat(s
, ALLOC_SLAB
);
2614 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2616 if (unlikely(PageSlabPfmemalloc(page
)))
2617 return gfp_pfmemalloc_allowed(gfpflags
);
2623 * Check the page->freelist of a page and either transfer the freelist to the
2624 * per cpu freelist or deactivate the page.
2626 * The page is still frozen if the return value is not NULL.
2628 * If this function returns NULL then the page has been unfrozen.
2630 * This function must be called with interrupt disabled.
2632 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2635 unsigned long counters
;
2639 freelist
= page
->freelist
;
2640 counters
= page
->counters
;
2642 new.counters
= counters
;
2643 VM_BUG_ON(!new.frozen
);
2645 new.inuse
= page
->objects
;
2646 new.frozen
= freelist
!= NULL
;
2648 } while (!__cmpxchg_double_slab(s
, page
,
2657 * Slow path. The lockless freelist is empty or we need to perform
2660 * Processing is still very fast if new objects have been freed to the
2661 * regular freelist. In that case we simply take over the regular freelist
2662 * as the lockless freelist and zap the regular freelist.
2664 * If that is not working then we fall back to the partial lists. We take the
2665 * first element of the freelist as the object to allocate now and move the
2666 * rest of the freelist to the lockless freelist.
2668 * And if we were unable to get a new slab from the partial slab lists then
2669 * we need to allocate a new slab. This is the slowest path since it involves
2670 * a call to the page allocator and the setup of a new slab.
2672 * Version of __slab_alloc to use when we know that interrupts are
2673 * already disabled (which is the case for bulk allocation).
2675 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2676 unsigned long addr
, struct kmem_cache_cpu
*c
)
2681 stat(s
, ALLOC_SLOWPATH
);
2686 * if the node is not online or has no normal memory, just
2687 * ignore the node constraint
2689 if (unlikely(node
!= NUMA_NO_NODE
&&
2690 !node_isset(node
, slab_nodes
)))
2691 node
= NUMA_NO_NODE
;
2696 if (unlikely(!node_match(page
, node
))) {
2698 * same as above but node_match() being false already
2699 * implies node != NUMA_NO_NODE
2701 if (!node_isset(node
, slab_nodes
)) {
2702 node
= NUMA_NO_NODE
;
2705 stat(s
, ALLOC_NODE_MISMATCH
);
2706 deactivate_slab(s
, page
, c
->freelist
, c
);
2712 * By rights, we should be searching for a slab page that was
2713 * PFMEMALLOC but right now, we are losing the pfmemalloc
2714 * information when the page leaves the per-cpu allocator
2716 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2717 deactivate_slab(s
, page
, c
->freelist
, c
);
2721 /* must check again c->freelist in case of cpu migration or IRQ */
2722 freelist
= c
->freelist
;
2726 freelist
= get_freelist(s
, page
);
2730 stat(s
, DEACTIVATE_BYPASS
);
2734 stat(s
, ALLOC_REFILL
);
2738 * freelist is pointing to the list of objects to be used.
2739 * page is pointing to the page from which the objects are obtained.
2740 * That page must be frozen for per cpu allocations to work.
2742 VM_BUG_ON(!c
->page
->frozen
);
2743 c
->freelist
= get_freepointer(s
, freelist
);
2744 c
->tid
= next_tid(c
->tid
);
2749 if (slub_percpu_partial(c
)) {
2750 page
= c
->page
= slub_percpu_partial(c
);
2751 slub_set_percpu_partial(c
, page
);
2752 stat(s
, CPU_PARTIAL_ALLOC
);
2756 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2758 if (unlikely(!freelist
)) {
2759 slab_out_of_memory(s
, gfpflags
, node
);
2764 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2767 /* Only entered in the debug case */
2768 if (kmem_cache_debug(s
) &&
2769 !alloc_debug_processing(s
, page
, freelist
, addr
))
2770 goto new_slab
; /* Slab failed checks. Next slab needed */
2772 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2777 * Another one that disabled interrupt and compensates for possible
2778 * cpu changes by refetching the per cpu area pointer.
2780 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2781 unsigned long addr
, struct kmem_cache_cpu
*c
)
2784 unsigned long flags
;
2786 local_irq_save(flags
);
2787 #ifdef CONFIG_PREEMPTION
2789 * We may have been preempted and rescheduled on a different
2790 * cpu before disabling interrupts. Need to reload cpu area
2793 c
= this_cpu_ptr(s
->cpu_slab
);
2796 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2797 local_irq_restore(flags
);
2802 * If the object has been wiped upon free, make sure it's fully initialized by
2803 * zeroing out freelist pointer.
2805 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2808 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2809 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
2814 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2815 * have the fastpath folded into their functions. So no function call
2816 * overhead for requests that can be satisfied on the fastpath.
2818 * The fastpath works by first checking if the lockless freelist can be used.
2819 * If not then __slab_alloc is called for slow processing.
2821 * Otherwise we can simply pick the next object from the lockless free list.
2823 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2824 gfp_t gfpflags
, int node
, unsigned long addr
, size_t orig_size
)
2827 struct kmem_cache_cpu
*c
;
2830 struct obj_cgroup
*objcg
= NULL
;
2833 s
= slab_pre_alloc_hook(s
, &objcg
, 1, gfpflags
);
2837 object
= kfence_alloc(s
, orig_size
, gfpflags
);
2838 if (unlikely(object
))
2843 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2844 * enabled. We may switch back and forth between cpus while
2845 * reading from one cpu area. That does not matter as long
2846 * as we end up on the original cpu again when doing the cmpxchg.
2848 * We should guarantee that tid and kmem_cache are retrieved on
2849 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2850 * to check if it is matched or not.
2853 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2854 c
= raw_cpu_ptr(s
->cpu_slab
);
2855 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
2856 unlikely(tid
!= READ_ONCE(c
->tid
)));
2859 * Irqless object alloc/free algorithm used here depends on sequence
2860 * of fetching cpu_slab's data. tid should be fetched before anything
2861 * on c to guarantee that object and page associated with previous tid
2862 * won't be used with current tid. If we fetch tid first, object and
2863 * page could be one associated with next tid and our alloc/free
2864 * request will be failed. In this case, we will retry. So, no problem.
2869 * The transaction ids are globally unique per cpu and per operation on
2870 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2871 * occurs on the right processor and that there was no operation on the
2872 * linked list in between.
2875 object
= c
->freelist
;
2877 if (unlikely(!object
|| !page
|| !node_match(page
, node
))) {
2878 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2880 void *next_object
= get_freepointer_safe(s
, object
);
2883 * The cmpxchg will only match if there was no additional
2884 * operation and if we are on the right processor.
2886 * The cmpxchg does the following atomically (without lock
2888 * 1. Relocate first pointer to the current per cpu area.
2889 * 2. Verify that tid and freelist have not been changed
2890 * 3. If they were not changed replace tid and freelist
2892 * Since this is without lock semantics the protection is only
2893 * against code executing on this cpu *not* from access by
2896 if (unlikely(!this_cpu_cmpxchg_double(
2897 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2899 next_object
, next_tid(tid
)))) {
2901 note_cmpxchg_failure("slab_alloc", s
, tid
);
2904 prefetch_freepointer(s
, next_object
);
2905 stat(s
, ALLOC_FASTPATH
);
2908 maybe_wipe_obj_freeptr(s
, object
);
2909 init
= slab_want_init_on_alloc(gfpflags
, s
);
2912 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
, init
);
2917 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2918 gfp_t gfpflags
, unsigned long addr
, size_t orig_size
)
2920 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
, orig_size
);
2923 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2925 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, s
->object_size
);
2927 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2932 EXPORT_SYMBOL(kmem_cache_alloc
);
2934 #ifdef CONFIG_TRACING
2935 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2937 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, size
);
2938 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2939 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2942 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2946 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2948 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, s
->object_size
);
2950 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2951 s
->object_size
, s
->size
, gfpflags
, node
);
2955 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2957 #ifdef CONFIG_TRACING
2958 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2960 int node
, size_t size
)
2962 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, size
);
2964 trace_kmalloc_node(_RET_IP_
, ret
,
2965 size
, s
->size
, gfpflags
, node
);
2967 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2970 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2972 #endif /* CONFIG_NUMA */
2975 * Slow path handling. This may still be called frequently since objects
2976 * have a longer lifetime than the cpu slabs in most processing loads.
2978 * So we still attempt to reduce cache line usage. Just take the slab
2979 * lock and free the item. If there is no additional partial page
2980 * handling required then we can return immediately.
2982 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2983 void *head
, void *tail
, int cnt
,
2990 unsigned long counters
;
2991 struct kmem_cache_node
*n
= NULL
;
2992 unsigned long flags
;
2994 stat(s
, FREE_SLOWPATH
);
2996 if (kfence_free(head
))
2999 if (kmem_cache_debug(s
) &&
3000 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
3005 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3008 prior
= page
->freelist
;
3009 counters
= page
->counters
;
3010 set_freepointer(s
, tail
, prior
);
3011 new.counters
= counters
;
3012 was_frozen
= new.frozen
;
3014 if ((!new.inuse
|| !prior
) && !was_frozen
) {
3016 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
3019 * Slab was on no list before and will be
3021 * We can defer the list move and instead
3026 } else { /* Needs to be taken off a list */
3028 n
= get_node(s
, page_to_nid(page
));
3030 * Speculatively acquire the list_lock.
3031 * If the cmpxchg does not succeed then we may
3032 * drop the list_lock without any processing.
3034 * Otherwise the list_lock will synchronize with
3035 * other processors updating the list of slabs.
3037 spin_lock_irqsave(&n
->list_lock
, flags
);
3042 } while (!cmpxchg_double_slab(s
, page
,
3049 if (likely(was_frozen
)) {
3051 * The list lock was not taken therefore no list
3052 * activity can be necessary.
3054 stat(s
, FREE_FROZEN
);
3055 } else if (new.frozen
) {
3057 * If we just froze the page then put it onto the
3058 * per cpu partial list.
3060 put_cpu_partial(s
, page
, 1);
3061 stat(s
, CPU_PARTIAL_FREE
);
3067 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3071 * Objects left in the slab. If it was not on the partial list before
3074 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3075 remove_full(s
, n
, page
);
3076 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
3077 stat(s
, FREE_ADD_PARTIAL
);
3079 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3085 * Slab on the partial list.
3087 remove_partial(n
, page
);
3088 stat(s
, FREE_REMOVE_PARTIAL
);
3090 /* Slab must be on the full list */
3091 remove_full(s
, n
, page
);
3094 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3096 discard_slab(s
, page
);
3100 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3101 * can perform fastpath freeing without additional function calls.
3103 * The fastpath is only possible if we are freeing to the current cpu slab
3104 * of this processor. This typically the case if we have just allocated
3107 * If fastpath is not possible then fall back to __slab_free where we deal
3108 * with all sorts of special processing.
3110 * Bulk free of a freelist with several objects (all pointing to the
3111 * same page) possible by specifying head and tail ptr, plus objects
3112 * count (cnt). Bulk free indicated by tail pointer being set.
3114 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3115 struct page
*page
, void *head
, void *tail
,
3116 int cnt
, unsigned long addr
)
3118 void *tail_obj
= tail
? : head
;
3119 struct kmem_cache_cpu
*c
;
3122 memcg_slab_free_hook(s
, &head
, 1);
3125 * Determine the currently cpus per cpu slab.
3126 * The cpu may change afterward. However that does not matter since
3127 * data is retrieved via this pointer. If we are on the same cpu
3128 * during the cmpxchg then the free will succeed.
3131 tid
= this_cpu_read(s
->cpu_slab
->tid
);
3132 c
= raw_cpu_ptr(s
->cpu_slab
);
3133 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
3134 unlikely(tid
!= READ_ONCE(c
->tid
)));
3136 /* Same with comment on barrier() in slab_alloc_node() */
3139 if (likely(page
== c
->page
)) {
3140 void **freelist
= READ_ONCE(c
->freelist
);
3142 set_freepointer(s
, tail_obj
, freelist
);
3144 if (unlikely(!this_cpu_cmpxchg_double(
3145 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3147 head
, next_tid(tid
)))) {
3149 note_cmpxchg_failure("slab_free", s
, tid
);
3152 stat(s
, FREE_FASTPATH
);
3154 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3158 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3159 void *head
, void *tail
, int cnt
,
3163 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3164 * to remove objects, whose reuse must be delayed.
3166 if (slab_free_freelist_hook(s
, &head
, &tail
))
3167 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3170 #ifdef CONFIG_KASAN_GENERIC
3171 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3173 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3177 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3179 s
= cache_from_obj(s
, x
);
3182 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3183 trace_kmem_cache_free(_RET_IP_
, x
, s
->name
);
3185 EXPORT_SYMBOL(kmem_cache_free
);
3187 struct detached_freelist
{
3192 struct kmem_cache
*s
;
3196 * This function progressively scans the array with free objects (with
3197 * a limited look ahead) and extract objects belonging to the same
3198 * page. It builds a detached freelist directly within the given
3199 * page/objects. This can happen without any need for
3200 * synchronization, because the objects are owned by running process.
3201 * The freelist is build up as a single linked list in the objects.
3202 * The idea is, that this detached freelist can then be bulk
3203 * transferred to the real freelist(s), but only requiring a single
3204 * synchronization primitive. Look ahead in the array is limited due
3205 * to performance reasons.
3208 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3209 void **p
, struct detached_freelist
*df
)
3211 size_t first_skipped_index
= 0;
3216 /* Always re-init detached_freelist */
3221 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3222 } while (!object
&& size
);
3227 page
= virt_to_head_page(object
);
3229 /* Handle kalloc'ed objects */
3230 if (unlikely(!PageSlab(page
))) {
3231 BUG_ON(!PageCompound(page
));
3233 __free_pages(page
, compound_order(page
));
3234 p
[size
] = NULL
; /* mark object processed */
3237 /* Derive kmem_cache from object */
3238 df
->s
= page
->slab_cache
;
3240 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3243 if (is_kfence_address(object
)) {
3244 slab_free_hook(df
->s
, object
, false);
3245 __kfence_free(object
);
3246 p
[size
] = NULL
; /* mark object processed */
3250 /* Start new detached freelist */
3252 set_freepointer(df
->s
, object
, NULL
);
3254 df
->freelist
= object
;
3255 p
[size
] = NULL
; /* mark object processed */
3261 continue; /* Skip processed objects */
3263 /* df->page is always set at this point */
3264 if (df
->page
== virt_to_head_page(object
)) {
3265 /* Opportunity build freelist */
3266 set_freepointer(df
->s
, object
, df
->freelist
);
3267 df
->freelist
= object
;
3269 p
[size
] = NULL
; /* mark object processed */
3274 /* Limit look ahead search */
3278 if (!first_skipped_index
)
3279 first_skipped_index
= size
+ 1;
3282 return first_skipped_index
;
3285 /* Note that interrupts must be enabled when calling this function. */
3286 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3291 memcg_slab_free_hook(s
, p
, size
);
3293 struct detached_freelist df
;
3295 size
= build_detached_freelist(s
, size
, p
, &df
);
3299 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
, _RET_IP_
);
3300 } while (likely(size
));
3302 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3304 /* Note that interrupts must be enabled when calling this function. */
3305 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3308 struct kmem_cache_cpu
*c
;
3310 struct obj_cgroup
*objcg
= NULL
;
3312 /* memcg and kmem_cache debug support */
3313 s
= slab_pre_alloc_hook(s
, &objcg
, size
, flags
);
3317 * Drain objects in the per cpu slab, while disabling local
3318 * IRQs, which protects against PREEMPT and interrupts
3319 * handlers invoking normal fastpath.
3321 local_irq_disable();
3322 c
= this_cpu_ptr(s
->cpu_slab
);
3324 for (i
= 0; i
< size
; i
++) {
3325 void *object
= kfence_alloc(s
, s
->object_size
, flags
);
3327 if (unlikely(object
)) {
3332 object
= c
->freelist
;
3333 if (unlikely(!object
)) {
3335 * We may have removed an object from c->freelist using
3336 * the fastpath in the previous iteration; in that case,
3337 * c->tid has not been bumped yet.
3338 * Since ___slab_alloc() may reenable interrupts while
3339 * allocating memory, we should bump c->tid now.
3341 c
->tid
= next_tid(c
->tid
);
3344 * Invoking slow path likely have side-effect
3345 * of re-populating per CPU c->freelist
3347 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3349 if (unlikely(!p
[i
]))
3352 c
= this_cpu_ptr(s
->cpu_slab
);
3353 maybe_wipe_obj_freeptr(s
, p
[i
]);
3355 continue; /* goto for-loop */
3357 c
->freelist
= get_freepointer(s
, object
);
3359 maybe_wipe_obj_freeptr(s
, p
[i
]);
3361 c
->tid
= next_tid(c
->tid
);
3365 * memcg and kmem_cache debug support and memory initialization.
3366 * Done outside of the IRQ disabled fastpath loop.
3368 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3369 slab_want_init_on_alloc(flags
, s
));
3373 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false);
3374 __kmem_cache_free_bulk(s
, i
, p
);
3377 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3381 * Object placement in a slab is made very easy because we always start at
3382 * offset 0. If we tune the size of the object to the alignment then we can
3383 * get the required alignment by putting one properly sized object after
3386 * Notice that the allocation order determines the sizes of the per cpu
3387 * caches. Each processor has always one slab available for allocations.
3388 * Increasing the allocation order reduces the number of times that slabs
3389 * must be moved on and off the partial lists and is therefore a factor in
3394 * Mininum / Maximum order of slab pages. This influences locking overhead
3395 * and slab fragmentation. A higher order reduces the number of partial slabs
3396 * and increases the number of allocations possible without having to
3397 * take the list_lock.
3399 static unsigned int slub_min_order
;
3400 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3401 static unsigned int slub_min_objects
;
3404 * Calculate the order of allocation given an slab object size.
3406 * The order of allocation has significant impact on performance and other
3407 * system components. Generally order 0 allocations should be preferred since
3408 * order 0 does not cause fragmentation in the page allocator. Larger objects
3409 * be problematic to put into order 0 slabs because there may be too much
3410 * unused space left. We go to a higher order if more than 1/16th of the slab
3413 * In order to reach satisfactory performance we must ensure that a minimum
3414 * number of objects is in one slab. Otherwise we may generate too much
3415 * activity on the partial lists which requires taking the list_lock. This is
3416 * less a concern for large slabs though which are rarely used.
3418 * slub_max_order specifies the order where we begin to stop considering the
3419 * number of objects in a slab as critical. If we reach slub_max_order then
3420 * we try to keep the page order as low as possible. So we accept more waste
3421 * of space in favor of a small page order.
3423 * Higher order allocations also allow the placement of more objects in a
3424 * slab and thereby reduce object handling overhead. If the user has
3425 * requested a higher minimum order then we start with that one instead of
3426 * the smallest order which will fit the object.
3428 static inline unsigned int slab_order(unsigned int size
,
3429 unsigned int min_objects
, unsigned int max_order
,
3430 unsigned int fract_leftover
)
3432 unsigned int min_order
= slub_min_order
;
3435 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3436 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3438 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3439 order
<= max_order
; order
++) {
3441 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3444 rem
= slab_size
% size
;
3446 if (rem
<= slab_size
/ fract_leftover
)
3453 static inline int calculate_order(unsigned int size
)
3456 unsigned int min_objects
;
3457 unsigned int max_objects
;
3458 unsigned int nr_cpus
;
3461 * Attempt to find best configuration for a slab. This
3462 * works by first attempting to generate a layout with
3463 * the best configuration and backing off gradually.
3465 * First we increase the acceptable waste in a slab. Then
3466 * we reduce the minimum objects required in a slab.
3468 min_objects
= slub_min_objects
;
3471 * Some architectures will only update present cpus when
3472 * onlining them, so don't trust the number if it's just 1. But
3473 * we also don't want to use nr_cpu_ids always, as on some other
3474 * architectures, there can be many possible cpus, but never
3475 * onlined. Here we compromise between trying to avoid too high
3476 * order on systems that appear larger than they are, and too
3477 * low order on systems that appear smaller than they are.
3479 nr_cpus
= num_present_cpus();
3481 nr_cpus
= nr_cpu_ids
;
3482 min_objects
= 4 * (fls(nr_cpus
) + 1);
3484 max_objects
= order_objects(slub_max_order
, size
);
3485 min_objects
= min(min_objects
, max_objects
);
3487 while (min_objects
> 1) {
3488 unsigned int fraction
;
3491 while (fraction
>= 4) {
3492 order
= slab_order(size
, min_objects
,
3493 slub_max_order
, fraction
);
3494 if (order
<= slub_max_order
)
3502 * We were unable to place multiple objects in a slab. Now
3503 * lets see if we can place a single object there.
3505 order
= slab_order(size
, 1, slub_max_order
, 1);
3506 if (order
<= slub_max_order
)
3510 * Doh this slab cannot be placed using slub_max_order.
3512 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3513 if (order
< MAX_ORDER
)
3519 init_kmem_cache_node(struct kmem_cache_node
*n
)
3522 spin_lock_init(&n
->list_lock
);
3523 INIT_LIST_HEAD(&n
->partial
);
3524 #ifdef CONFIG_SLUB_DEBUG
3525 atomic_long_set(&n
->nr_slabs
, 0);
3526 atomic_long_set(&n
->total_objects
, 0);
3527 INIT_LIST_HEAD(&n
->full
);
3531 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3533 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3534 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3537 * Must align to double word boundary for the double cmpxchg
3538 * instructions to work; see __pcpu_double_call_return_bool().
3540 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3541 2 * sizeof(void *));
3546 init_kmem_cache_cpus(s
);
3551 static struct kmem_cache
*kmem_cache_node
;
3554 * No kmalloc_node yet so do it by hand. We know that this is the first
3555 * slab on the node for this slabcache. There are no concurrent accesses
3558 * Note that this function only works on the kmem_cache_node
3559 * when allocating for the kmem_cache_node. This is used for bootstrapping
3560 * memory on a fresh node that has no slab structures yet.
3562 static void early_kmem_cache_node_alloc(int node
)
3565 struct kmem_cache_node
*n
;
3567 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3569 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3572 if (page_to_nid(page
) != node
) {
3573 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3574 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3579 #ifdef CONFIG_SLUB_DEBUG
3580 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3581 init_tracking(kmem_cache_node
, n
);
3583 n
= kasan_slab_alloc(kmem_cache_node
, n
, GFP_KERNEL
, false);
3584 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3587 kmem_cache_node
->node
[node
] = n
;
3588 init_kmem_cache_node(n
);
3589 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3592 * No locks need to be taken here as it has just been
3593 * initialized and there is no concurrent access.
3595 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3598 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3601 struct kmem_cache_node
*n
;
3603 for_each_kmem_cache_node(s
, node
, n
) {
3604 s
->node
[node
] = NULL
;
3605 kmem_cache_free(kmem_cache_node
, n
);
3609 void __kmem_cache_release(struct kmem_cache
*s
)
3611 cache_random_seq_destroy(s
);
3612 free_percpu(s
->cpu_slab
);
3613 free_kmem_cache_nodes(s
);
3616 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3620 for_each_node_mask(node
, slab_nodes
) {
3621 struct kmem_cache_node
*n
;
3623 if (slab_state
== DOWN
) {
3624 early_kmem_cache_node_alloc(node
);
3627 n
= kmem_cache_alloc_node(kmem_cache_node
,
3631 free_kmem_cache_nodes(s
);
3635 init_kmem_cache_node(n
);
3641 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3643 if (min
< MIN_PARTIAL
)
3645 else if (min
> MAX_PARTIAL
)
3647 s
->min_partial
= min
;
3650 static void set_cpu_partial(struct kmem_cache
*s
)
3652 #ifdef CONFIG_SLUB_CPU_PARTIAL
3654 * cpu_partial determined the maximum number of objects kept in the
3655 * per cpu partial lists of a processor.
3657 * Per cpu partial lists mainly contain slabs that just have one
3658 * object freed. If they are used for allocation then they can be
3659 * filled up again with minimal effort. The slab will never hit the
3660 * per node partial lists and therefore no locking will be required.
3662 * This setting also determines
3664 * A) The number of objects from per cpu partial slabs dumped to the
3665 * per node list when we reach the limit.
3666 * B) The number of objects in cpu partial slabs to extract from the
3667 * per node list when we run out of per cpu objects. We only fetch
3668 * 50% to keep some capacity around for frees.
3670 if (!kmem_cache_has_cpu_partial(s
))
3671 slub_set_cpu_partial(s
, 0);
3672 else if (s
->size
>= PAGE_SIZE
)
3673 slub_set_cpu_partial(s
, 2);
3674 else if (s
->size
>= 1024)
3675 slub_set_cpu_partial(s
, 6);
3676 else if (s
->size
>= 256)
3677 slub_set_cpu_partial(s
, 13);
3679 slub_set_cpu_partial(s
, 30);
3684 * calculate_sizes() determines the order and the distribution of data within
3687 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3689 slab_flags_t flags
= s
->flags
;
3690 unsigned int size
= s
->object_size
;
3691 unsigned int freepointer_area
;
3695 * Round up object size to the next word boundary. We can only
3696 * place the free pointer at word boundaries and this determines
3697 * the possible location of the free pointer.
3699 size
= ALIGN(size
, sizeof(void *));
3701 * This is the area of the object where a freepointer can be
3702 * safely written. If redzoning adds more to the inuse size, we
3703 * can't use that portion for writing the freepointer, so
3704 * s->offset must be limited within this for the general case.
3706 freepointer_area
= size
;
3708 #ifdef CONFIG_SLUB_DEBUG
3710 * Determine if we can poison the object itself. If the user of
3711 * the slab may touch the object after free or before allocation
3712 * then we should never poison the object itself.
3714 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3716 s
->flags
|= __OBJECT_POISON
;
3718 s
->flags
&= ~__OBJECT_POISON
;
3722 * If we are Redzoning then check if there is some space between the
3723 * end of the object and the free pointer. If not then add an
3724 * additional word to have some bytes to store Redzone information.
3726 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3727 size
+= sizeof(void *);
3731 * With that we have determined the number of bytes in actual use
3732 * by the object. This is the potential offset to the free pointer.
3736 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3739 * Relocate free pointer after the object if it is not
3740 * permitted to overwrite the first word of the object on
3743 * This is the case if we do RCU, have a constructor or
3744 * destructor or are poisoning the objects.
3746 * The assumption that s->offset >= s->inuse means free
3747 * pointer is outside of the object is used in the
3748 * freeptr_outside_object() function. If that is no
3749 * longer true, the function needs to be modified.
3752 size
+= sizeof(void *);
3753 } else if (freepointer_area
> sizeof(void *)) {
3755 * Store freelist pointer near middle of object to keep
3756 * it away from the edges of the object to avoid small
3757 * sized over/underflows from neighboring allocations.
3759 s
->offset
= ALIGN(freepointer_area
/ 2, sizeof(void *));
3762 #ifdef CONFIG_SLUB_DEBUG
3763 if (flags
& SLAB_STORE_USER
)
3765 * Need to store information about allocs and frees after
3768 size
+= 2 * sizeof(struct track
);
3771 kasan_cache_create(s
, &size
, &s
->flags
);
3772 #ifdef CONFIG_SLUB_DEBUG
3773 if (flags
& SLAB_RED_ZONE
) {
3775 * Add some empty padding so that we can catch
3776 * overwrites from earlier objects rather than let
3777 * tracking information or the free pointer be
3778 * corrupted if a user writes before the start
3781 size
+= sizeof(void *);
3783 s
->red_left_pad
= sizeof(void *);
3784 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3785 size
+= s
->red_left_pad
;
3790 * SLUB stores one object immediately after another beginning from
3791 * offset 0. In order to align the objects we have to simply size
3792 * each object to conform to the alignment.
3794 size
= ALIGN(size
, s
->align
);
3796 s
->reciprocal_size
= reciprocal_value(size
);
3797 if (forced_order
>= 0)
3798 order
= forced_order
;
3800 order
= calculate_order(size
);
3807 s
->allocflags
|= __GFP_COMP
;
3809 if (s
->flags
& SLAB_CACHE_DMA
)
3810 s
->allocflags
|= GFP_DMA
;
3812 if (s
->flags
& SLAB_CACHE_DMA32
)
3813 s
->allocflags
|= GFP_DMA32
;
3815 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3816 s
->allocflags
|= __GFP_RECLAIMABLE
;
3819 * Determine the number of objects per slab
3821 s
->oo
= oo_make(order
, size
);
3822 s
->min
= oo_make(get_order(size
), size
);
3823 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3826 return !!oo_objects(s
->oo
);
3829 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3831 #ifdef CONFIG_SLUB_DEBUG
3833 * If no slub_debug was enabled globally, the static key is not yet
3834 * enabled by setup_slub_debug(). Enable it if the cache is being
3835 * created with any of the debugging flags passed explicitly.
3837 if (flags
& SLAB_DEBUG_FLAGS
)
3838 static_branch_enable(&slub_debug_enabled
);
3840 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
);
3841 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3842 s
->random
= get_random_long();
3845 if (!calculate_sizes(s
, -1))
3847 if (disable_higher_order_debug
) {
3849 * Disable debugging flags that store metadata if the min slab
3852 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3853 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3855 if (!calculate_sizes(s
, -1))
3860 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3861 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3862 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3863 /* Enable fast mode */
3864 s
->flags
|= __CMPXCHG_DOUBLE
;
3868 * The larger the object size is, the more pages we want on the partial
3869 * list to avoid pounding the page allocator excessively.
3871 set_min_partial(s
, ilog2(s
->size
) / 2);
3876 s
->remote_node_defrag_ratio
= 1000;
3879 /* Initialize the pre-computed randomized freelist if slab is up */
3880 if (slab_state
>= UP
) {
3881 if (init_cache_random_seq(s
))
3885 if (!init_kmem_cache_nodes(s
))
3888 if (alloc_kmem_cache_cpus(s
))
3891 free_kmem_cache_nodes(s
);
3896 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3899 #ifdef CONFIG_SLUB_DEBUG
3900 void *addr
= page_address(page
);
3904 slab_err(s
, page
, text
, s
->name
);
3907 map
= get_map(s
, page
);
3908 for_each_object(p
, s
, addr
, page
->objects
) {
3910 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
3911 pr_err("Object 0x%p @offset=%tu\n", p
, p
- addr
);
3912 print_tracking(s
, p
);
3921 * Attempt to free all partial slabs on a node.
3922 * This is called from __kmem_cache_shutdown(). We must take list_lock
3923 * because sysfs file might still access partial list after the shutdowning.
3925 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3928 struct page
*page
, *h
;
3930 BUG_ON(irqs_disabled());
3931 spin_lock_irq(&n
->list_lock
);
3932 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3934 remove_partial(n
, page
);
3935 list_add(&page
->slab_list
, &discard
);
3937 list_slab_objects(s
, page
,
3938 "Objects remaining in %s on __kmem_cache_shutdown()");
3941 spin_unlock_irq(&n
->list_lock
);
3943 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3944 discard_slab(s
, page
);
3947 bool __kmem_cache_empty(struct kmem_cache
*s
)
3950 struct kmem_cache_node
*n
;
3952 for_each_kmem_cache_node(s
, node
, n
)
3953 if (n
->nr_partial
|| slabs_node(s
, node
))
3959 * Release all resources used by a slab cache.
3961 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3964 struct kmem_cache_node
*n
;
3967 /* Attempt to free all objects */
3968 for_each_kmem_cache_node(s
, node
, n
) {
3970 if (n
->nr_partial
|| slabs_node(s
, node
))
3976 #ifdef CONFIG_PRINTK
3977 void kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct page
*page
)
3980 int __maybe_unused i
;
3984 struct kmem_cache
*s
= page
->slab_cache
;
3985 struct track __maybe_unused
*trackp
;
3987 kpp
->kp_ptr
= object
;
3988 kpp
->kp_page
= page
;
3989 kpp
->kp_slab_cache
= s
;
3990 base
= page_address(page
);
3991 objp0
= kasan_reset_tag(object
);
3992 #ifdef CONFIG_SLUB_DEBUG
3993 objp
= restore_red_left(s
, objp0
);
3997 objnr
= obj_to_index(s
, page
, objp
);
3998 kpp
->kp_data_offset
= (unsigned long)((char *)objp0
- (char *)objp
);
3999 objp
= base
+ s
->size
* objnr
;
4000 kpp
->kp_objp
= objp
;
4001 if (WARN_ON_ONCE(objp
< base
|| objp
>= base
+ page
->objects
* s
->size
|| (objp
- base
) % s
->size
) ||
4002 !(s
->flags
& SLAB_STORE_USER
))
4004 #ifdef CONFIG_SLUB_DEBUG
4005 trackp
= get_track(s
, objp
, TRACK_ALLOC
);
4006 kpp
->kp_ret
= (void *)trackp
->addr
;
4007 #ifdef CONFIG_STACKTRACE
4008 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< TRACK_ADDRS_COUNT
; i
++) {
4009 kpp
->kp_stack
[i
] = (void *)trackp
->addrs
[i
];
4010 if (!kpp
->kp_stack
[i
])
4018 /********************************************************************
4020 *******************************************************************/
4022 static int __init
setup_slub_min_order(char *str
)
4024 get_option(&str
, (int *)&slub_min_order
);
4029 __setup("slub_min_order=", setup_slub_min_order
);
4031 static int __init
setup_slub_max_order(char *str
)
4033 get_option(&str
, (int *)&slub_max_order
);
4034 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
4039 __setup("slub_max_order=", setup_slub_max_order
);
4041 static int __init
setup_slub_min_objects(char *str
)
4043 get_option(&str
, (int *)&slub_min_objects
);
4048 __setup("slub_min_objects=", setup_slub_min_objects
);
4050 void *__kmalloc(size_t size
, gfp_t flags
)
4052 struct kmem_cache
*s
;
4055 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4056 return kmalloc_large(size
, flags
);
4058 s
= kmalloc_slab(size
, flags
);
4060 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4063 ret
= slab_alloc(s
, flags
, _RET_IP_
, size
);
4065 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
4067 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4071 EXPORT_SYMBOL(__kmalloc
);
4074 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
4078 unsigned int order
= get_order(size
);
4080 flags
|= __GFP_COMP
;
4081 page
= alloc_pages_node(node
, flags
, order
);
4083 ptr
= page_address(page
);
4084 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4085 PAGE_SIZE
<< order
);
4088 return kmalloc_large_node_hook(ptr
, size
, flags
);
4091 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4093 struct kmem_cache
*s
;
4096 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4097 ret
= kmalloc_large_node(size
, flags
, node
);
4099 trace_kmalloc_node(_RET_IP_
, ret
,
4100 size
, PAGE_SIZE
<< get_order(size
),
4106 s
= kmalloc_slab(size
, flags
);
4108 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4111 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
, size
);
4113 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
4115 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4119 EXPORT_SYMBOL(__kmalloc_node
);
4120 #endif /* CONFIG_NUMA */
4122 #ifdef CONFIG_HARDENED_USERCOPY
4124 * Rejects incorrectly sized objects and objects that are to be copied
4125 * to/from userspace but do not fall entirely within the containing slab
4126 * cache's usercopy region.
4128 * Returns NULL if check passes, otherwise const char * to name of cache
4129 * to indicate an error.
4131 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4134 struct kmem_cache
*s
;
4135 unsigned int offset
;
4137 bool is_kfence
= is_kfence_address(ptr
);
4139 ptr
= kasan_reset_tag(ptr
);
4141 /* Find object and usable object size. */
4142 s
= page
->slab_cache
;
4144 /* Reject impossible pointers. */
4145 if (ptr
< page_address(page
))
4146 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4149 /* Find offset within object. */
4151 offset
= ptr
- kfence_object_start(ptr
);
4153 offset
= (ptr
- page_address(page
)) % s
->size
;
4155 /* Adjust for redzone and reject if within the redzone. */
4156 if (!is_kfence
&& kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4157 if (offset
< s
->red_left_pad
)
4158 usercopy_abort("SLUB object in left red zone",
4159 s
->name
, to_user
, offset
, n
);
4160 offset
-= s
->red_left_pad
;
4163 /* Allow address range falling entirely within usercopy region. */
4164 if (offset
>= s
->useroffset
&&
4165 offset
- s
->useroffset
<= s
->usersize
&&
4166 n
<= s
->useroffset
- offset
+ s
->usersize
)
4170 * If the copy is still within the allocated object, produce
4171 * a warning instead of rejecting the copy. This is intended
4172 * to be a temporary method to find any missing usercopy
4175 object_size
= slab_ksize(s
);
4176 if (usercopy_fallback
&&
4177 offset
<= object_size
&& n
<= object_size
- offset
) {
4178 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4182 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4184 #endif /* CONFIG_HARDENED_USERCOPY */
4186 size_t __ksize(const void *object
)
4190 if (unlikely(object
== ZERO_SIZE_PTR
))
4193 page
= virt_to_head_page(object
);
4195 if (unlikely(!PageSlab(page
))) {
4196 WARN_ON(!PageCompound(page
));
4197 return page_size(page
);
4200 return slab_ksize(page
->slab_cache
);
4202 EXPORT_SYMBOL(__ksize
);
4204 void kfree(const void *x
)
4207 void *object
= (void *)x
;
4209 trace_kfree(_RET_IP_
, x
);
4211 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4214 page
= virt_to_head_page(x
);
4215 if (unlikely(!PageSlab(page
))) {
4216 unsigned int order
= compound_order(page
);
4218 BUG_ON(!PageCompound(page
));
4220 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4221 -(PAGE_SIZE
<< order
));
4222 __free_pages(page
, order
);
4225 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4227 EXPORT_SYMBOL(kfree
);
4229 #define SHRINK_PROMOTE_MAX 32
4232 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4233 * up most to the head of the partial lists. New allocations will then
4234 * fill those up and thus they can be removed from the partial lists.
4236 * The slabs with the least items are placed last. This results in them
4237 * being allocated from last increasing the chance that the last objects
4238 * are freed in them.
4240 int __kmem_cache_shrink(struct kmem_cache
*s
)
4244 struct kmem_cache_node
*n
;
4247 struct list_head discard
;
4248 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4249 unsigned long flags
;
4253 for_each_kmem_cache_node(s
, node
, n
) {
4254 INIT_LIST_HEAD(&discard
);
4255 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4256 INIT_LIST_HEAD(promote
+ i
);
4258 spin_lock_irqsave(&n
->list_lock
, flags
);
4261 * Build lists of slabs to discard or promote.
4263 * Note that concurrent frees may occur while we hold the
4264 * list_lock. page->inuse here is the upper limit.
4266 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4267 int free
= page
->objects
- page
->inuse
;
4269 /* Do not reread page->inuse */
4272 /* We do not keep full slabs on the list */
4275 if (free
== page
->objects
) {
4276 list_move(&page
->slab_list
, &discard
);
4278 } else if (free
<= SHRINK_PROMOTE_MAX
)
4279 list_move(&page
->slab_list
, promote
+ free
- 1);
4283 * Promote the slabs filled up most to the head of the
4286 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4287 list_splice(promote
+ i
, &n
->partial
);
4289 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4291 /* Release empty slabs */
4292 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4293 discard_slab(s
, page
);
4295 if (slabs_node(s
, node
))
4302 static int slab_mem_going_offline_callback(void *arg
)
4304 struct kmem_cache
*s
;
4306 mutex_lock(&slab_mutex
);
4307 list_for_each_entry(s
, &slab_caches
, list
)
4308 __kmem_cache_shrink(s
);
4309 mutex_unlock(&slab_mutex
);
4314 static void slab_mem_offline_callback(void *arg
)
4316 struct memory_notify
*marg
= arg
;
4319 offline_node
= marg
->status_change_nid_normal
;
4322 * If the node still has available memory. we need kmem_cache_node
4325 if (offline_node
< 0)
4328 mutex_lock(&slab_mutex
);
4329 node_clear(offline_node
, slab_nodes
);
4331 * We no longer free kmem_cache_node structures here, as it would be
4332 * racy with all get_node() users, and infeasible to protect them with
4335 mutex_unlock(&slab_mutex
);
4338 static int slab_mem_going_online_callback(void *arg
)
4340 struct kmem_cache_node
*n
;
4341 struct kmem_cache
*s
;
4342 struct memory_notify
*marg
= arg
;
4343 int nid
= marg
->status_change_nid_normal
;
4347 * If the node's memory is already available, then kmem_cache_node is
4348 * already created. Nothing to do.
4354 * We are bringing a node online. No memory is available yet. We must
4355 * allocate a kmem_cache_node structure in order to bring the node
4358 mutex_lock(&slab_mutex
);
4359 list_for_each_entry(s
, &slab_caches
, list
) {
4361 * The structure may already exist if the node was previously
4362 * onlined and offlined.
4364 if (get_node(s
, nid
))
4367 * XXX: kmem_cache_alloc_node will fallback to other nodes
4368 * since memory is not yet available from the node that
4371 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4376 init_kmem_cache_node(n
);
4380 * Any cache created after this point will also have kmem_cache_node
4381 * initialized for the new node.
4383 node_set(nid
, slab_nodes
);
4385 mutex_unlock(&slab_mutex
);
4389 static int slab_memory_callback(struct notifier_block
*self
,
4390 unsigned long action
, void *arg
)
4395 case MEM_GOING_ONLINE
:
4396 ret
= slab_mem_going_online_callback(arg
);
4398 case MEM_GOING_OFFLINE
:
4399 ret
= slab_mem_going_offline_callback(arg
);
4402 case MEM_CANCEL_ONLINE
:
4403 slab_mem_offline_callback(arg
);
4406 case MEM_CANCEL_OFFLINE
:
4410 ret
= notifier_from_errno(ret
);
4416 static struct notifier_block slab_memory_callback_nb
= {
4417 .notifier_call
= slab_memory_callback
,
4418 .priority
= SLAB_CALLBACK_PRI
,
4421 /********************************************************************
4422 * Basic setup of slabs
4423 *******************************************************************/
4426 * Used for early kmem_cache structures that were allocated using
4427 * the page allocator. Allocate them properly then fix up the pointers
4428 * that may be pointing to the wrong kmem_cache structure.
4431 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4434 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4435 struct kmem_cache_node
*n
;
4437 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4440 * This runs very early, and only the boot processor is supposed to be
4441 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4444 __flush_cpu_slab(s
, smp_processor_id());
4445 for_each_kmem_cache_node(s
, node
, n
) {
4448 list_for_each_entry(p
, &n
->partial
, slab_list
)
4451 #ifdef CONFIG_SLUB_DEBUG
4452 list_for_each_entry(p
, &n
->full
, slab_list
)
4456 list_add(&s
->list
, &slab_caches
);
4460 void __init
kmem_cache_init(void)
4462 static __initdata
struct kmem_cache boot_kmem_cache
,
4463 boot_kmem_cache_node
;
4466 if (debug_guardpage_minorder())
4469 kmem_cache_node
= &boot_kmem_cache_node
;
4470 kmem_cache
= &boot_kmem_cache
;
4473 * Initialize the nodemask for which we will allocate per node
4474 * structures. Here we don't need taking slab_mutex yet.
4476 for_each_node_state(node
, N_NORMAL_MEMORY
)
4477 node_set(node
, slab_nodes
);
4479 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4480 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4482 register_hotmemory_notifier(&slab_memory_callback_nb
);
4484 /* Able to allocate the per node structures */
4485 slab_state
= PARTIAL
;
4487 create_boot_cache(kmem_cache
, "kmem_cache",
4488 offsetof(struct kmem_cache
, node
) +
4489 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4490 SLAB_HWCACHE_ALIGN
, 0, 0);
4492 kmem_cache
= bootstrap(&boot_kmem_cache
);
4493 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4495 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4496 setup_kmalloc_cache_index_table();
4497 create_kmalloc_caches(0);
4499 /* Setup random freelists for each cache */
4500 init_freelist_randomization();
4502 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4505 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4507 slub_min_order
, slub_max_order
, slub_min_objects
,
4508 nr_cpu_ids
, nr_node_ids
);
4511 void __init
kmem_cache_init_late(void)
4516 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4517 slab_flags_t flags
, void (*ctor
)(void *))
4519 struct kmem_cache
*s
;
4521 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4526 * Adjust the object sizes so that we clear
4527 * the complete object on kzalloc.
4529 s
->object_size
= max(s
->object_size
, size
);
4530 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4532 if (sysfs_slab_alias(s
, name
)) {
4541 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4545 err
= kmem_cache_open(s
, flags
);
4549 /* Mutex is not taken during early boot */
4550 if (slab_state
<= UP
)
4553 err
= sysfs_slab_add(s
);
4555 __kmem_cache_release(s
);
4560 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4562 struct kmem_cache
*s
;
4565 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4566 return kmalloc_large(size
, gfpflags
);
4568 s
= kmalloc_slab(size
, gfpflags
);
4570 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4573 ret
= slab_alloc(s
, gfpflags
, caller
, size
);
4575 /* Honor the call site pointer we received. */
4576 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4580 EXPORT_SYMBOL(__kmalloc_track_caller
);
4583 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4584 int node
, unsigned long caller
)
4586 struct kmem_cache
*s
;
4589 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4590 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4592 trace_kmalloc_node(caller
, ret
,
4593 size
, PAGE_SIZE
<< get_order(size
),
4599 s
= kmalloc_slab(size
, gfpflags
);
4601 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4604 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
, size
);
4606 /* Honor the call site pointer we received. */
4607 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4611 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4615 static int count_inuse(struct page
*page
)
4620 static int count_total(struct page
*page
)
4622 return page
->objects
;
4626 #ifdef CONFIG_SLUB_DEBUG
4627 static void validate_slab(struct kmem_cache
*s
, struct page
*page
)
4630 void *addr
= page_address(page
);
4635 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4638 /* Now we know that a valid freelist exists */
4639 map
= get_map(s
, page
);
4640 for_each_object(p
, s
, addr
, page
->objects
) {
4641 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), map
) ?
4642 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4644 if (!check_object(s
, page
, p
, val
))
4652 static int validate_slab_node(struct kmem_cache
*s
,
4653 struct kmem_cache_node
*n
)
4655 unsigned long count
= 0;
4657 unsigned long flags
;
4659 spin_lock_irqsave(&n
->list_lock
, flags
);
4661 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4662 validate_slab(s
, page
);
4665 if (count
!= n
->nr_partial
)
4666 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4667 s
->name
, count
, n
->nr_partial
);
4669 if (!(s
->flags
& SLAB_STORE_USER
))
4672 list_for_each_entry(page
, &n
->full
, slab_list
) {
4673 validate_slab(s
, page
);
4676 if (count
!= atomic_long_read(&n
->nr_slabs
))
4677 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4678 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4681 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4685 static long validate_slab_cache(struct kmem_cache
*s
)
4688 unsigned long count
= 0;
4689 struct kmem_cache_node
*n
;
4692 for_each_kmem_cache_node(s
, node
, n
)
4693 count
+= validate_slab_node(s
, n
);
4698 * Generate lists of code addresses where slabcache objects are allocated
4703 unsigned long count
;
4710 DECLARE_BITMAP(cpus
, NR_CPUS
);
4716 unsigned long count
;
4717 struct location
*loc
;
4720 static void free_loc_track(struct loc_track
*t
)
4723 free_pages((unsigned long)t
->loc
,
4724 get_order(sizeof(struct location
) * t
->max
));
4727 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4732 order
= get_order(sizeof(struct location
) * max
);
4734 l
= (void *)__get_free_pages(flags
, order
);
4739 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4747 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4748 const struct track
*track
)
4750 long start
, end
, pos
;
4752 unsigned long caddr
;
4753 unsigned long age
= jiffies
- track
->when
;
4759 pos
= start
+ (end
- start
+ 1) / 2;
4762 * There is nothing at "end". If we end up there
4763 * we need to add something to before end.
4768 caddr
= t
->loc
[pos
].addr
;
4769 if (track
->addr
== caddr
) {
4775 if (age
< l
->min_time
)
4777 if (age
> l
->max_time
)
4780 if (track
->pid
< l
->min_pid
)
4781 l
->min_pid
= track
->pid
;
4782 if (track
->pid
> l
->max_pid
)
4783 l
->max_pid
= track
->pid
;
4785 cpumask_set_cpu(track
->cpu
,
4786 to_cpumask(l
->cpus
));
4788 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4792 if (track
->addr
< caddr
)
4799 * Not found. Insert new tracking element.
4801 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4807 (t
->count
- pos
) * sizeof(struct location
));
4810 l
->addr
= track
->addr
;
4814 l
->min_pid
= track
->pid
;
4815 l
->max_pid
= track
->pid
;
4816 cpumask_clear(to_cpumask(l
->cpus
));
4817 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4818 nodes_clear(l
->nodes
);
4819 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4823 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4824 struct page
*page
, enum track_item alloc
)
4826 void *addr
= page_address(page
);
4830 map
= get_map(s
, page
);
4831 for_each_object(p
, s
, addr
, page
->objects
)
4832 if (!test_bit(__obj_to_index(s
, addr
, p
), map
))
4833 add_location(t
, s
, get_track(s
, p
, alloc
));
4837 static int list_locations(struct kmem_cache
*s
, char *buf
,
4838 enum track_item alloc
)
4842 struct loc_track t
= { 0, 0, NULL
};
4844 struct kmem_cache_node
*n
;
4846 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4848 return sysfs_emit(buf
, "Out of memory\n");
4850 /* Push back cpu slabs */
4853 for_each_kmem_cache_node(s
, node
, n
) {
4854 unsigned long flags
;
4857 if (!atomic_long_read(&n
->nr_slabs
))
4860 spin_lock_irqsave(&n
->list_lock
, flags
);
4861 list_for_each_entry(page
, &n
->partial
, slab_list
)
4862 process_slab(&t
, s
, page
, alloc
);
4863 list_for_each_entry(page
, &n
->full
, slab_list
)
4864 process_slab(&t
, s
, page
, alloc
);
4865 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4868 for (i
= 0; i
< t
.count
; i
++) {
4869 struct location
*l
= &t
.loc
[i
];
4871 len
+= sysfs_emit_at(buf
, len
, "%7ld ", l
->count
);
4874 len
+= sysfs_emit_at(buf
, len
, "%pS", (void *)l
->addr
);
4876 len
+= sysfs_emit_at(buf
, len
, "<not-available>");
4878 if (l
->sum_time
!= l
->min_time
)
4879 len
+= sysfs_emit_at(buf
, len
, " age=%ld/%ld/%ld",
4881 (long)div_u64(l
->sum_time
,
4885 len
+= sysfs_emit_at(buf
, len
, " age=%ld", l
->min_time
);
4887 if (l
->min_pid
!= l
->max_pid
)
4888 len
+= sysfs_emit_at(buf
, len
, " pid=%ld-%ld",
4889 l
->min_pid
, l
->max_pid
);
4891 len
+= sysfs_emit_at(buf
, len
, " pid=%ld",
4894 if (num_online_cpus() > 1 &&
4895 !cpumask_empty(to_cpumask(l
->cpus
)))
4896 len
+= sysfs_emit_at(buf
, len
, " cpus=%*pbl",
4897 cpumask_pr_args(to_cpumask(l
->cpus
)));
4899 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
4900 len
+= sysfs_emit_at(buf
, len
, " nodes=%*pbl",
4901 nodemask_pr_args(&l
->nodes
));
4903 len
+= sysfs_emit_at(buf
, len
, "\n");
4908 len
+= sysfs_emit_at(buf
, len
, "No data\n");
4912 #endif /* CONFIG_SLUB_DEBUG */
4914 #ifdef SLUB_RESILIENCY_TEST
4915 static void __init
resiliency_test(void)
4918 int type
= KMALLOC_NORMAL
;
4920 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4922 pr_err("SLUB resiliency testing\n");
4923 pr_err("-----------------------\n");
4924 pr_err("A. Corruption after allocation\n");
4926 p
= kzalloc(16, GFP_KERNEL
);
4928 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4931 validate_slab_cache(kmalloc_caches
[type
][4]);
4933 /* Hmmm... The next two are dangerous */
4934 p
= kzalloc(32, GFP_KERNEL
);
4935 p
[32 + sizeof(void *)] = 0x34;
4936 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4938 pr_err("If allocated object is overwritten then not detectable\n\n");
4940 validate_slab_cache(kmalloc_caches
[type
][5]);
4941 p
= kzalloc(64, GFP_KERNEL
);
4942 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4944 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4946 pr_err("If allocated object is overwritten then not detectable\n\n");
4947 validate_slab_cache(kmalloc_caches
[type
][6]);
4949 pr_err("\nB. Corruption after free\n");
4950 p
= kzalloc(128, GFP_KERNEL
);
4953 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4954 validate_slab_cache(kmalloc_caches
[type
][7]);
4956 p
= kzalloc(256, GFP_KERNEL
);
4959 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4960 validate_slab_cache(kmalloc_caches
[type
][8]);
4962 p
= kzalloc(512, GFP_KERNEL
);
4965 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4966 validate_slab_cache(kmalloc_caches
[type
][9]);
4970 static void resiliency_test(void) {};
4972 #endif /* SLUB_RESILIENCY_TEST */
4975 enum slab_stat_type
{
4976 SL_ALL
, /* All slabs */
4977 SL_PARTIAL
, /* Only partially allocated slabs */
4978 SL_CPU
, /* Only slabs used for cpu caches */
4979 SL_OBJECTS
, /* Determine allocated objects not slabs */
4980 SL_TOTAL
/* Determine object capacity not slabs */
4983 #define SO_ALL (1 << SL_ALL)
4984 #define SO_PARTIAL (1 << SL_PARTIAL)
4985 #define SO_CPU (1 << SL_CPU)
4986 #define SO_OBJECTS (1 << SL_OBJECTS)
4987 #define SO_TOTAL (1 << SL_TOTAL)
4989 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4990 char *buf
, unsigned long flags
)
4992 unsigned long total
= 0;
4995 unsigned long *nodes
;
4998 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
5002 if (flags
& SO_CPU
) {
5005 for_each_possible_cpu(cpu
) {
5006 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
5011 page
= READ_ONCE(c
->page
);
5015 node
= page_to_nid(page
);
5016 if (flags
& SO_TOTAL
)
5018 else if (flags
& SO_OBJECTS
)
5026 page
= slub_percpu_partial_read_once(c
);
5028 node
= page_to_nid(page
);
5029 if (flags
& SO_TOTAL
)
5031 else if (flags
& SO_OBJECTS
)
5042 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5043 * already held which will conflict with an existing lock order:
5045 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5047 * We don't really need mem_hotplug_lock (to hold off
5048 * slab_mem_going_offline_callback) here because slab's memory hot
5049 * unplug code doesn't destroy the kmem_cache->node[] data.
5052 #ifdef CONFIG_SLUB_DEBUG
5053 if (flags
& SO_ALL
) {
5054 struct kmem_cache_node
*n
;
5056 for_each_kmem_cache_node(s
, node
, n
) {
5058 if (flags
& SO_TOTAL
)
5059 x
= atomic_long_read(&n
->total_objects
);
5060 else if (flags
& SO_OBJECTS
)
5061 x
= atomic_long_read(&n
->total_objects
) -
5062 count_partial(n
, count_free
);
5064 x
= atomic_long_read(&n
->nr_slabs
);
5071 if (flags
& SO_PARTIAL
) {
5072 struct kmem_cache_node
*n
;
5074 for_each_kmem_cache_node(s
, node
, n
) {
5075 if (flags
& SO_TOTAL
)
5076 x
= count_partial(n
, count_total
);
5077 else if (flags
& SO_OBJECTS
)
5078 x
= count_partial(n
, count_inuse
);
5086 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5088 for (node
= 0; node
< nr_node_ids
; node
++) {
5090 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5094 len
+= sysfs_emit_at(buf
, len
, "\n");
5100 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5101 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5103 struct slab_attribute
{
5104 struct attribute attr
;
5105 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5106 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5109 #define SLAB_ATTR_RO(_name) \
5110 static struct slab_attribute _name##_attr = \
5111 __ATTR(_name, 0400, _name##_show, NULL)
5113 #define SLAB_ATTR(_name) \
5114 static struct slab_attribute _name##_attr = \
5115 __ATTR(_name, 0600, _name##_show, _name##_store)
5117 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5119 return sysfs_emit(buf
, "%u\n", s
->size
);
5121 SLAB_ATTR_RO(slab_size
);
5123 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5125 return sysfs_emit(buf
, "%u\n", s
->align
);
5127 SLAB_ATTR_RO(align
);
5129 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5131 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5133 SLAB_ATTR_RO(object_size
);
5135 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5137 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5139 SLAB_ATTR_RO(objs_per_slab
);
5141 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5143 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5145 SLAB_ATTR_RO(order
);
5147 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5149 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5152 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5158 err
= kstrtoul(buf
, 10, &min
);
5162 set_min_partial(s
, min
);
5165 SLAB_ATTR(min_partial
);
5167 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5169 return sysfs_emit(buf
, "%u\n", slub_cpu_partial(s
));
5172 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5175 unsigned int objects
;
5178 err
= kstrtouint(buf
, 10, &objects
);
5181 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5184 slub_set_cpu_partial(s
, objects
);
5188 SLAB_ATTR(cpu_partial
);
5190 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5194 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5198 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5200 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5202 SLAB_ATTR_RO(aliases
);
5204 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5206 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5208 SLAB_ATTR_RO(partial
);
5210 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5212 return show_slab_objects(s
, buf
, SO_CPU
);
5214 SLAB_ATTR_RO(cpu_slabs
);
5216 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5218 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5220 SLAB_ATTR_RO(objects
);
5222 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5224 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5226 SLAB_ATTR_RO(objects_partial
);
5228 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5235 for_each_online_cpu(cpu
) {
5238 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5241 pages
+= page
->pages
;
5242 objects
+= page
->pobjects
;
5246 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, pages
);
5249 for_each_online_cpu(cpu
) {
5252 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5254 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5255 cpu
, page
->pobjects
, page
->pages
);
5258 len
+= sysfs_emit_at(buf
, len
, "\n");
5262 SLAB_ATTR_RO(slabs_cpu_partial
);
5264 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5266 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5268 SLAB_ATTR_RO(reclaim_account
);
5270 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5272 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5274 SLAB_ATTR_RO(hwcache_align
);
5276 #ifdef CONFIG_ZONE_DMA
5277 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5279 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5281 SLAB_ATTR_RO(cache_dma
);
5284 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5286 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5288 SLAB_ATTR_RO(usersize
);
5290 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5292 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5294 SLAB_ATTR_RO(destroy_by_rcu
);
5296 #ifdef CONFIG_SLUB_DEBUG
5297 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5299 return show_slab_objects(s
, buf
, SO_ALL
);
5301 SLAB_ATTR_RO(slabs
);
5303 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5305 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5307 SLAB_ATTR_RO(total_objects
);
5309 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5311 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5313 SLAB_ATTR_RO(sanity_checks
);
5315 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5317 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5319 SLAB_ATTR_RO(trace
);
5321 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5323 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5326 SLAB_ATTR_RO(red_zone
);
5328 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5330 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5333 SLAB_ATTR_RO(poison
);
5335 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5337 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5340 SLAB_ATTR_RO(store_user
);
5342 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5347 static ssize_t
validate_store(struct kmem_cache
*s
,
5348 const char *buf
, size_t length
)
5352 if (buf
[0] == '1') {
5353 ret
= validate_slab_cache(s
);
5359 SLAB_ATTR(validate
);
5361 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5363 if (!(s
->flags
& SLAB_STORE_USER
))
5365 return list_locations(s
, buf
, TRACK_ALLOC
);
5367 SLAB_ATTR_RO(alloc_calls
);
5369 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5371 if (!(s
->flags
& SLAB_STORE_USER
))
5373 return list_locations(s
, buf
, TRACK_FREE
);
5375 SLAB_ATTR_RO(free_calls
);
5376 #endif /* CONFIG_SLUB_DEBUG */
5378 #ifdef CONFIG_FAILSLAB
5379 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5381 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5383 SLAB_ATTR_RO(failslab
);
5386 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5391 static ssize_t
shrink_store(struct kmem_cache
*s
,
5392 const char *buf
, size_t length
)
5395 kmem_cache_shrink(s
);
5403 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5405 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5408 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5409 const char *buf
, size_t length
)
5414 err
= kstrtouint(buf
, 10, &ratio
);
5420 s
->remote_node_defrag_ratio
= ratio
* 10;
5424 SLAB_ATTR(remote_node_defrag_ratio
);
5427 #ifdef CONFIG_SLUB_STATS
5428 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5430 unsigned long sum
= 0;
5433 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5438 for_each_online_cpu(cpu
) {
5439 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5445 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5448 for_each_online_cpu(cpu
) {
5450 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5455 len
+= sysfs_emit_at(buf
, len
, "\n");
5460 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5464 for_each_online_cpu(cpu
)
5465 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5468 #define STAT_ATTR(si, text) \
5469 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5471 return show_stat(s, buf, si); \
5473 static ssize_t text##_store(struct kmem_cache *s, \
5474 const char *buf, size_t length) \
5476 if (buf[0] != '0') \
5478 clear_stat(s, si); \
5483 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5484 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5485 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5486 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5487 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5488 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5489 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5490 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5491 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5492 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5493 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5494 STAT_ATTR(FREE_SLAB
, free_slab
);
5495 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5496 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5497 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5498 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5499 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5500 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5501 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5502 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5503 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5504 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5505 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5506 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5507 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5508 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5509 #endif /* CONFIG_SLUB_STATS */
5511 static struct attribute
*slab_attrs
[] = {
5512 &slab_size_attr
.attr
,
5513 &object_size_attr
.attr
,
5514 &objs_per_slab_attr
.attr
,
5516 &min_partial_attr
.attr
,
5517 &cpu_partial_attr
.attr
,
5519 &objects_partial_attr
.attr
,
5521 &cpu_slabs_attr
.attr
,
5525 &hwcache_align_attr
.attr
,
5526 &reclaim_account_attr
.attr
,
5527 &destroy_by_rcu_attr
.attr
,
5529 &slabs_cpu_partial_attr
.attr
,
5530 #ifdef CONFIG_SLUB_DEBUG
5531 &total_objects_attr
.attr
,
5533 &sanity_checks_attr
.attr
,
5535 &red_zone_attr
.attr
,
5537 &store_user_attr
.attr
,
5538 &validate_attr
.attr
,
5539 &alloc_calls_attr
.attr
,
5540 &free_calls_attr
.attr
,
5542 #ifdef CONFIG_ZONE_DMA
5543 &cache_dma_attr
.attr
,
5546 &remote_node_defrag_ratio_attr
.attr
,
5548 #ifdef CONFIG_SLUB_STATS
5549 &alloc_fastpath_attr
.attr
,
5550 &alloc_slowpath_attr
.attr
,
5551 &free_fastpath_attr
.attr
,
5552 &free_slowpath_attr
.attr
,
5553 &free_frozen_attr
.attr
,
5554 &free_add_partial_attr
.attr
,
5555 &free_remove_partial_attr
.attr
,
5556 &alloc_from_partial_attr
.attr
,
5557 &alloc_slab_attr
.attr
,
5558 &alloc_refill_attr
.attr
,
5559 &alloc_node_mismatch_attr
.attr
,
5560 &free_slab_attr
.attr
,
5561 &cpuslab_flush_attr
.attr
,
5562 &deactivate_full_attr
.attr
,
5563 &deactivate_empty_attr
.attr
,
5564 &deactivate_to_head_attr
.attr
,
5565 &deactivate_to_tail_attr
.attr
,
5566 &deactivate_remote_frees_attr
.attr
,
5567 &deactivate_bypass_attr
.attr
,
5568 &order_fallback_attr
.attr
,
5569 &cmpxchg_double_fail_attr
.attr
,
5570 &cmpxchg_double_cpu_fail_attr
.attr
,
5571 &cpu_partial_alloc_attr
.attr
,
5572 &cpu_partial_free_attr
.attr
,
5573 &cpu_partial_node_attr
.attr
,
5574 &cpu_partial_drain_attr
.attr
,
5576 #ifdef CONFIG_FAILSLAB
5577 &failslab_attr
.attr
,
5579 &usersize_attr
.attr
,
5584 static const struct attribute_group slab_attr_group
= {
5585 .attrs
= slab_attrs
,
5588 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5589 struct attribute
*attr
,
5592 struct slab_attribute
*attribute
;
5593 struct kmem_cache
*s
;
5596 attribute
= to_slab_attr(attr
);
5599 if (!attribute
->show
)
5602 err
= attribute
->show(s
, buf
);
5607 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5608 struct attribute
*attr
,
5609 const char *buf
, size_t len
)
5611 struct slab_attribute
*attribute
;
5612 struct kmem_cache
*s
;
5615 attribute
= to_slab_attr(attr
);
5618 if (!attribute
->store
)
5621 err
= attribute
->store(s
, buf
, len
);
5625 static void kmem_cache_release(struct kobject
*k
)
5627 slab_kmem_cache_release(to_slab(k
));
5630 static const struct sysfs_ops slab_sysfs_ops
= {
5631 .show
= slab_attr_show
,
5632 .store
= slab_attr_store
,
5635 static struct kobj_type slab_ktype
= {
5636 .sysfs_ops
= &slab_sysfs_ops
,
5637 .release
= kmem_cache_release
,
5640 static struct kset
*slab_kset
;
5642 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5647 #define ID_STR_LENGTH 64
5649 /* Create a unique string id for a slab cache:
5651 * Format :[flags-]size
5653 static char *create_unique_id(struct kmem_cache
*s
)
5655 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5662 * First flags affecting slabcache operations. We will only
5663 * get here for aliasable slabs so we do not need to support
5664 * too many flags. The flags here must cover all flags that
5665 * are matched during merging to guarantee that the id is
5668 if (s
->flags
& SLAB_CACHE_DMA
)
5670 if (s
->flags
& SLAB_CACHE_DMA32
)
5672 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5674 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5676 if (s
->flags
& SLAB_ACCOUNT
)
5680 p
+= sprintf(p
, "%07u", s
->size
);
5682 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5686 static int sysfs_slab_add(struct kmem_cache
*s
)
5690 struct kset
*kset
= cache_kset(s
);
5691 int unmergeable
= slab_unmergeable(s
);
5694 kobject_init(&s
->kobj
, &slab_ktype
);
5698 if (!unmergeable
&& disable_higher_order_debug
&&
5699 (slub_debug
& DEBUG_METADATA_FLAGS
))
5704 * Slabcache can never be merged so we can use the name proper.
5705 * This is typically the case for debug situations. In that
5706 * case we can catch duplicate names easily.
5708 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5712 * Create a unique name for the slab as a target
5715 name
= create_unique_id(s
);
5718 s
->kobj
.kset
= kset
;
5719 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5723 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5728 /* Setup first alias */
5729 sysfs_slab_alias(s
, s
->name
);
5736 kobject_del(&s
->kobj
);
5740 void sysfs_slab_unlink(struct kmem_cache
*s
)
5742 if (slab_state
>= FULL
)
5743 kobject_del(&s
->kobj
);
5746 void sysfs_slab_release(struct kmem_cache
*s
)
5748 if (slab_state
>= FULL
)
5749 kobject_put(&s
->kobj
);
5753 * Need to buffer aliases during bootup until sysfs becomes
5754 * available lest we lose that information.
5756 struct saved_alias
{
5757 struct kmem_cache
*s
;
5759 struct saved_alias
*next
;
5762 static struct saved_alias
*alias_list
;
5764 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5766 struct saved_alias
*al
;
5768 if (slab_state
== FULL
) {
5770 * If we have a leftover link then remove it.
5772 sysfs_remove_link(&slab_kset
->kobj
, name
);
5773 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5776 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5782 al
->next
= alias_list
;
5787 static int __init
slab_sysfs_init(void)
5789 struct kmem_cache
*s
;
5792 mutex_lock(&slab_mutex
);
5794 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
5796 mutex_unlock(&slab_mutex
);
5797 pr_err("Cannot register slab subsystem.\n");
5803 list_for_each_entry(s
, &slab_caches
, list
) {
5804 err
= sysfs_slab_add(s
);
5806 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5810 while (alias_list
) {
5811 struct saved_alias
*al
= alias_list
;
5813 alias_list
= alias_list
->next
;
5814 err
= sysfs_slab_alias(al
->s
, al
->name
);
5816 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5821 mutex_unlock(&slab_mutex
);
5826 __initcall(slab_sysfs_init
);
5827 #endif /* CONFIG_SYSFS */
5830 * The /proc/slabinfo ABI
5832 #ifdef CONFIG_SLUB_DEBUG
5833 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5835 unsigned long nr_slabs
= 0;
5836 unsigned long nr_objs
= 0;
5837 unsigned long nr_free
= 0;
5839 struct kmem_cache_node
*n
;
5841 for_each_kmem_cache_node(s
, node
, n
) {
5842 nr_slabs
+= node_nr_slabs(n
);
5843 nr_objs
+= node_nr_objs(n
);
5844 nr_free
+= count_partial(n
, count_free
);
5847 sinfo
->active_objs
= nr_objs
- nr_free
;
5848 sinfo
->num_objs
= nr_objs
;
5849 sinfo
->active_slabs
= nr_slabs
;
5850 sinfo
->num_slabs
= nr_slabs
;
5851 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5852 sinfo
->cache_order
= oo_order(s
->oo
);
5855 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5859 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5860 size_t count
, loff_t
*ppos
)
5864 #endif /* CONFIG_SLUB_DEBUG */