1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/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 * Mininum 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("INFO: %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("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
654 page
, page
->objects
, page
->inuse
, page
->freelist
, page
->flags
);
658 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
660 struct va_format vaf
;
666 pr_err("=============================================================================\n");
667 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
668 pr_err("-----------------------------------------------------------------------------\n\n");
670 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
674 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
676 struct va_format vaf
;
682 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
686 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
687 void **freelist
, void *nextfree
)
689 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
690 !check_valid_pointer(s
, page
, nextfree
) && freelist
) {
691 object_err(s
, page
, *freelist
, "Freechain corrupt");
693 slab_fix(s
, "Isolate corrupted freechain");
700 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
702 unsigned int off
; /* Offset of last byte */
703 u8
*addr
= page_address(page
);
705 print_tracking(s
, p
);
707 print_page_info(page
);
709 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
710 p
, p
- addr
, get_freepointer(s
, p
));
712 if (s
->flags
& SLAB_RED_ZONE
)
713 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
715 else if (p
> addr
+ 16)
716 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
718 print_section(KERN_ERR
, "Object ", p
,
719 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
720 if (s
->flags
& SLAB_RED_ZONE
)
721 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
722 s
->inuse
- s
->object_size
);
724 off
= get_info_end(s
);
726 if (s
->flags
& SLAB_STORE_USER
)
727 off
+= 2 * sizeof(struct track
);
729 off
+= kasan_metadata_size(s
);
731 if (off
!= size_from_object(s
))
732 /* Beginning of the filler is the free pointer */
733 print_section(KERN_ERR
, "Padding ", p
+ off
,
734 size_from_object(s
) - off
);
739 void object_err(struct kmem_cache
*s
, struct page
*page
,
740 u8
*object
, char *reason
)
742 slab_bug(s
, "%s", reason
);
743 print_trailer(s
, page
, object
);
746 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
747 const char *fmt
, ...)
753 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
755 slab_bug(s
, "%s", buf
);
756 print_page_info(page
);
760 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
762 u8
*p
= kasan_reset_tag(object
);
764 if (s
->flags
& SLAB_RED_ZONE
)
765 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
767 if (s
->flags
& __OBJECT_POISON
) {
768 memset(p
, POISON_FREE
, s
->object_size
- 1);
769 p
[s
->object_size
- 1] = POISON_END
;
772 if (s
->flags
& SLAB_RED_ZONE
)
773 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
776 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
777 void *from
, void *to
)
779 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
780 memset(from
, data
, to
- from
);
783 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
784 u8
*object
, char *what
,
785 u8
*start
, unsigned int value
, unsigned int bytes
)
789 u8
*addr
= page_address(page
);
791 metadata_access_enable();
792 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
793 metadata_access_disable();
798 while (end
> fault
&& end
[-1] == value
)
801 slab_bug(s
, "%s overwritten", what
);
802 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
803 fault
, end
- 1, fault
- addr
,
805 print_trailer(s
, page
, object
);
807 restore_bytes(s
, what
, value
, fault
, end
);
815 * Bytes of the object to be managed.
816 * If the freepointer may overlay the object then the free
817 * pointer is at the middle of the object.
819 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
822 * object + s->object_size
823 * Padding to reach word boundary. This is also used for Redzoning.
824 * Padding is extended by another word if Redzoning is enabled and
825 * object_size == inuse.
827 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
828 * 0xcc (RED_ACTIVE) for objects in use.
831 * Meta data starts here.
833 * A. Free pointer (if we cannot overwrite object on free)
834 * B. Tracking data for SLAB_STORE_USER
835 * C. Padding to reach required alignment boundary or at mininum
836 * one word if debugging is on to be able to detect writes
837 * before the word boundary.
839 * Padding is done using 0x5a (POISON_INUSE)
842 * Nothing is used beyond s->size.
844 * If slabcaches are merged then the object_size and inuse boundaries are mostly
845 * ignored. And therefore no slab options that rely on these boundaries
846 * may be used with merged slabcaches.
849 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
851 unsigned long off
= get_info_end(s
); /* The end of info */
853 if (s
->flags
& SLAB_STORE_USER
)
854 /* We also have user information there */
855 off
+= 2 * sizeof(struct track
);
857 off
+= kasan_metadata_size(s
);
859 if (size_from_object(s
) == off
)
862 return check_bytes_and_report(s
, page
, p
, "Object padding",
863 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
866 /* Check the pad bytes at the end of a slab page */
867 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
876 if (!(s
->flags
& SLAB_POISON
))
879 start
= page_address(page
);
880 length
= page_size(page
);
881 end
= start
+ length
;
882 remainder
= length
% s
->size
;
886 pad
= end
- remainder
;
887 metadata_access_enable();
888 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
889 metadata_access_disable();
892 while (end
> fault
&& end
[-1] == POISON_INUSE
)
895 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
896 fault
, end
- 1, fault
- start
);
897 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
899 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
903 static int check_object(struct kmem_cache
*s
, struct page
*page
,
904 void *object
, u8 val
)
907 u8
*endobject
= object
+ s
->object_size
;
909 if (s
->flags
& SLAB_RED_ZONE
) {
910 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
911 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
914 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
915 endobject
, val
, s
->inuse
- s
->object_size
))
918 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
919 check_bytes_and_report(s
, page
, p
, "Alignment padding",
920 endobject
, POISON_INUSE
,
921 s
->inuse
- s
->object_size
);
925 if (s
->flags
& SLAB_POISON
) {
926 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
927 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
928 POISON_FREE
, s
->object_size
- 1) ||
929 !check_bytes_and_report(s
, page
, p
, "Poison",
930 p
+ s
->object_size
- 1, POISON_END
, 1)))
933 * check_pad_bytes cleans up on its own.
935 check_pad_bytes(s
, page
, p
);
938 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
940 * Object and freepointer overlap. Cannot check
941 * freepointer while object is allocated.
945 /* Check free pointer validity */
946 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
947 object_err(s
, page
, p
, "Freepointer corrupt");
949 * No choice but to zap it and thus lose the remainder
950 * of the free objects in this slab. May cause
951 * another error because the object count is now wrong.
953 set_freepointer(s
, p
, NULL
);
959 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
963 VM_BUG_ON(!irqs_disabled());
965 if (!PageSlab(page
)) {
966 slab_err(s
, page
, "Not a valid slab page");
970 maxobj
= order_objects(compound_order(page
), s
->size
);
971 if (page
->objects
> maxobj
) {
972 slab_err(s
, page
, "objects %u > max %u",
973 page
->objects
, maxobj
);
976 if (page
->inuse
> page
->objects
) {
977 slab_err(s
, page
, "inuse %u > max %u",
978 page
->inuse
, page
->objects
);
981 /* Slab_pad_check fixes things up after itself */
982 slab_pad_check(s
, page
);
987 * Determine if a certain object on a page is on the freelist. Must hold the
988 * slab lock to guarantee that the chains are in a consistent state.
990 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
998 while (fp
&& nr
<= page
->objects
) {
1001 if (!check_valid_pointer(s
, page
, fp
)) {
1003 object_err(s
, page
, object
,
1004 "Freechain corrupt");
1005 set_freepointer(s
, object
, NULL
);
1007 slab_err(s
, page
, "Freepointer corrupt");
1008 page
->freelist
= NULL
;
1009 page
->inuse
= page
->objects
;
1010 slab_fix(s
, "Freelist cleared");
1016 fp
= get_freepointer(s
, object
);
1020 max_objects
= order_objects(compound_order(page
), s
->size
);
1021 if (max_objects
> MAX_OBJS_PER_PAGE
)
1022 max_objects
= MAX_OBJS_PER_PAGE
;
1024 if (page
->objects
!= max_objects
) {
1025 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1026 page
->objects
, max_objects
);
1027 page
->objects
= max_objects
;
1028 slab_fix(s
, "Number of objects adjusted.");
1030 if (page
->inuse
!= page
->objects
- nr
) {
1031 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1032 page
->inuse
, page
->objects
- nr
);
1033 page
->inuse
= page
->objects
- nr
;
1034 slab_fix(s
, "Object count adjusted.");
1036 return search
== NULL
;
1039 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1042 if (s
->flags
& SLAB_TRACE
) {
1043 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1045 alloc
? "alloc" : "free",
1046 object
, page
->inuse
,
1050 print_section(KERN_INFO
, "Object ", (void *)object
,
1058 * Tracking of fully allocated slabs for debugging purposes.
1060 static void add_full(struct kmem_cache
*s
,
1061 struct kmem_cache_node
*n
, struct page
*page
)
1063 if (!(s
->flags
& SLAB_STORE_USER
))
1066 lockdep_assert_held(&n
->list_lock
);
1067 list_add(&page
->slab_list
, &n
->full
);
1070 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1072 if (!(s
->flags
& SLAB_STORE_USER
))
1075 lockdep_assert_held(&n
->list_lock
);
1076 list_del(&page
->slab_list
);
1079 /* Tracking of the number of slabs for debugging purposes */
1080 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1082 struct kmem_cache_node
*n
= get_node(s
, node
);
1084 return atomic_long_read(&n
->nr_slabs
);
1087 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1089 return atomic_long_read(&n
->nr_slabs
);
1092 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1094 struct kmem_cache_node
*n
= get_node(s
, node
);
1097 * May be called early in order to allocate a slab for the
1098 * kmem_cache_node structure. Solve the chicken-egg
1099 * dilemma by deferring the increment of the count during
1100 * bootstrap (see early_kmem_cache_node_alloc).
1103 atomic_long_inc(&n
->nr_slabs
);
1104 atomic_long_add(objects
, &n
->total_objects
);
1107 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1109 struct kmem_cache_node
*n
= get_node(s
, node
);
1111 atomic_long_dec(&n
->nr_slabs
);
1112 atomic_long_sub(objects
, &n
->total_objects
);
1115 /* Object debug checks for alloc/free paths */
1116 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1119 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1122 init_object(s
, object
, SLUB_RED_INACTIVE
);
1123 init_tracking(s
, object
);
1127 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1129 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1132 metadata_access_enable();
1133 memset(kasan_reset_tag(addr
), POISON_INUSE
, page_size(page
));
1134 metadata_access_disable();
1137 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1138 struct page
*page
, void *object
)
1140 if (!check_slab(s
, page
))
1143 if (!check_valid_pointer(s
, page
, object
)) {
1144 object_err(s
, page
, object
, "Freelist Pointer check fails");
1148 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1154 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1156 void *object
, unsigned long addr
)
1158 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1159 if (!alloc_consistency_checks(s
, page
, object
))
1163 /* Success perform special debug activities for allocs */
1164 if (s
->flags
& SLAB_STORE_USER
)
1165 set_track(s
, object
, TRACK_ALLOC
, addr
);
1166 trace(s
, page
, object
, 1);
1167 init_object(s
, object
, SLUB_RED_ACTIVE
);
1171 if (PageSlab(page
)) {
1173 * If this is a slab page then lets do the best we can
1174 * to avoid issues in the future. Marking all objects
1175 * as used avoids touching the remaining objects.
1177 slab_fix(s
, "Marking all objects used");
1178 page
->inuse
= page
->objects
;
1179 page
->freelist
= NULL
;
1184 static inline int free_consistency_checks(struct kmem_cache
*s
,
1185 struct page
*page
, void *object
, unsigned long addr
)
1187 if (!check_valid_pointer(s
, page
, object
)) {
1188 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1192 if (on_freelist(s
, page
, object
)) {
1193 object_err(s
, page
, object
, "Object already free");
1197 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1200 if (unlikely(s
!= page
->slab_cache
)) {
1201 if (!PageSlab(page
)) {
1202 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1204 } else if (!page
->slab_cache
) {
1205 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1209 object_err(s
, page
, object
,
1210 "page slab pointer corrupt.");
1216 /* Supports checking bulk free of a constructed freelist */
1217 static noinline
int free_debug_processing(
1218 struct kmem_cache
*s
, struct page
*page
,
1219 void *head
, void *tail
, int bulk_cnt
,
1222 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1223 void *object
= head
;
1225 unsigned long flags
;
1228 spin_lock_irqsave(&n
->list_lock
, flags
);
1231 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1232 if (!check_slab(s
, page
))
1239 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1240 if (!free_consistency_checks(s
, page
, object
, addr
))
1244 if (s
->flags
& SLAB_STORE_USER
)
1245 set_track(s
, object
, TRACK_FREE
, addr
);
1246 trace(s
, page
, object
, 0);
1247 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1248 init_object(s
, object
, SLUB_RED_INACTIVE
);
1250 /* Reached end of constructed freelist yet? */
1251 if (object
!= tail
) {
1252 object
= get_freepointer(s
, object
);
1258 if (cnt
!= bulk_cnt
)
1259 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1263 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1265 slab_fix(s
, "Object at 0x%p not freed", object
);
1270 * Parse a block of slub_debug options. Blocks are delimited by ';'
1272 * @str: start of block
1273 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1274 * @slabs: return start of list of slabs, or NULL when there's no list
1275 * @init: assume this is initial parsing and not per-kmem-create parsing
1277 * returns the start of next block if there's any, or NULL
1280 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1282 bool higher_order_disable
= false;
1284 /* Skip any completely empty blocks */
1285 while (*str
&& *str
== ';')
1290 * No options but restriction on slabs. This means full
1291 * debugging for slabs matching a pattern.
1293 *flags
= DEBUG_DEFAULT_FLAGS
;
1298 /* Determine which debug features should be switched on */
1299 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1300 switch (tolower(*str
)) {
1305 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1308 *flags
|= SLAB_RED_ZONE
;
1311 *flags
|= SLAB_POISON
;
1314 *flags
|= SLAB_STORE_USER
;
1317 *flags
|= SLAB_TRACE
;
1320 *flags
|= SLAB_FAILSLAB
;
1324 * Avoid enabling debugging on caches if its minimum
1325 * order would increase as a result.
1327 higher_order_disable
= true;
1331 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1340 /* Skip over the slab list */
1341 while (*str
&& *str
!= ';')
1344 /* Skip any completely empty blocks */
1345 while (*str
&& *str
== ';')
1348 if (init
&& higher_order_disable
)
1349 disable_higher_order_debug
= 1;
1357 static int __init
setup_slub_debug(char *str
)
1362 bool global_slub_debug_changed
= false;
1363 bool slab_list_specified
= false;
1365 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1366 if (*str
++ != '=' || !*str
)
1368 * No options specified. Switch on full debugging.
1374 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1378 global_slub_debug_changed
= true;
1380 slab_list_specified
= true;
1385 * For backwards compatibility, a single list of flags with list of
1386 * slabs means debugging is only enabled for those slabs, so the global
1387 * slub_debug should be 0. We can extended that to multiple lists as
1388 * long as there is no option specifying flags without a slab list.
1390 if (slab_list_specified
) {
1391 if (!global_slub_debug_changed
)
1393 slub_debug_string
= saved_str
;
1396 if (slub_debug
!= 0 || slub_debug_string
)
1397 static_branch_enable(&slub_debug_enabled
);
1398 if ((static_branch_unlikely(&init_on_alloc
) ||
1399 static_branch_unlikely(&init_on_free
)) &&
1400 (slub_debug
& SLAB_POISON
))
1401 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1405 __setup("slub_debug", setup_slub_debug
);
1408 * kmem_cache_flags - apply debugging options to the cache
1409 * @object_size: the size of an object without meta data
1410 * @flags: flags to set
1411 * @name: name of the cache
1413 * Debug option(s) are applied to @flags. In addition to the debug
1414 * option(s), if a slab name (or multiple) is specified i.e.
1415 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1416 * then only the select slabs will receive the debug option(s).
1418 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1419 slab_flags_t flags
, const char *name
)
1424 slab_flags_t block_flags
;
1425 slab_flags_t slub_debug_local
= slub_debug
;
1428 * If the slab cache is for debugging (e.g. kmemleak) then
1429 * don't store user (stack trace) information by default,
1430 * but let the user enable it via the command line below.
1432 if (flags
& SLAB_NOLEAKTRACE
)
1433 slub_debug_local
&= ~SLAB_STORE_USER
;
1436 next_block
= slub_debug_string
;
1437 /* Go through all blocks of debug options, see if any matches our slab's name */
1438 while (next_block
) {
1439 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1442 /* Found a block that has a slab list, search it */
1447 end
= strchrnul(iter
, ',');
1448 if (next_block
&& next_block
< end
)
1449 end
= next_block
- 1;
1451 glob
= strnchr(iter
, end
- iter
, '*');
1453 cmplen
= glob
- iter
;
1455 cmplen
= max_t(size_t, len
, (end
- iter
));
1457 if (!strncmp(name
, iter
, cmplen
)) {
1458 flags
|= block_flags
;
1462 if (!*end
|| *end
== ';')
1468 return flags
| slub_debug_local
;
1470 #else /* !CONFIG_SLUB_DEBUG */
1471 static inline void setup_object_debug(struct kmem_cache
*s
,
1472 struct page
*page
, void *object
) {}
1474 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1476 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1477 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1479 static inline int free_debug_processing(
1480 struct kmem_cache
*s
, struct page
*page
,
1481 void *head
, void *tail
, int bulk_cnt
,
1482 unsigned long addr
) { return 0; }
1484 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1486 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1487 void *object
, u8 val
) { return 1; }
1488 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1489 struct page
*page
) {}
1490 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1491 struct page
*page
) {}
1492 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1493 slab_flags_t flags
, const char *name
)
1497 #define slub_debug 0
1499 #define disable_higher_order_debug 0
1501 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1503 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1505 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1507 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1510 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1511 void **freelist
, void *nextfree
)
1515 #endif /* CONFIG_SLUB_DEBUG */
1518 * Hooks for other subsystems that check memory allocations. In a typical
1519 * production configuration these hooks all should produce no code at all.
1521 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1523 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1524 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1525 kmemleak_alloc(ptr
, size
, 1, flags
);
1529 static __always_inline
void kfree_hook(void *x
)
1532 kasan_kfree_large(x
);
1535 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
, void *x
)
1537 kmemleak_free_recursive(x
, s
->flags
);
1540 * Trouble is that we may no longer disable interrupts in the fast path
1541 * So in order to make the debug calls that expect irqs to be
1542 * disabled we need to disable interrupts temporarily.
1544 #ifdef CONFIG_LOCKDEP
1546 unsigned long flags
;
1548 local_irq_save(flags
);
1549 debug_check_no_locks_freed(x
, s
->object_size
);
1550 local_irq_restore(flags
);
1553 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1554 debug_check_no_obj_freed(x
, s
->object_size
);
1556 /* Use KCSAN to help debug racy use-after-free. */
1557 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1558 __kcsan_check_access(x
, s
->object_size
,
1559 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1561 /* KASAN might put x into memory quarantine, delaying its reuse */
1562 return kasan_slab_free(s
, x
);
1565 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1566 void **head
, void **tail
)
1571 void *old_tail
= *tail
? *tail
: *head
;
1574 if (is_kfence_address(next
)) {
1575 slab_free_hook(s
, next
);
1579 /* Head and tail of the reconstructed freelist */
1585 next
= get_freepointer(s
, object
);
1587 if (slab_want_init_on_free(s
)) {
1589 * Clear the object and the metadata, but don't touch
1592 memset(kasan_reset_tag(object
), 0, s
->object_size
);
1593 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
1595 memset((char *)kasan_reset_tag(object
) + s
->inuse
, 0,
1596 s
->size
- s
->inuse
- rsize
);
1599 /* If object's reuse doesn't have to be delayed */
1600 if (!slab_free_hook(s
, object
)) {
1601 /* Move object to the new freelist */
1602 set_freepointer(s
, object
, *head
);
1607 } while (object
!= old_tail
);
1612 return *head
!= NULL
;
1615 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1618 setup_object_debug(s
, page
, object
);
1619 object
= kasan_init_slab_obj(s
, object
);
1620 if (unlikely(s
->ctor
)) {
1621 kasan_unpoison_object_data(s
, object
);
1623 kasan_poison_object_data(s
, object
);
1629 * Slab allocation and freeing
1631 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1632 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1635 unsigned int order
= oo_order(oo
);
1637 if (node
== NUMA_NO_NODE
)
1638 page
= alloc_pages(flags
, order
);
1640 page
= __alloc_pages_node(node
, flags
, order
);
1645 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1646 /* Pre-initialize the random sequence cache */
1647 static int init_cache_random_seq(struct kmem_cache
*s
)
1649 unsigned int count
= oo_objects(s
->oo
);
1652 /* Bailout if already initialised */
1656 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1658 pr_err("SLUB: Unable to initialize free list for %s\n",
1663 /* Transform to an offset on the set of pages */
1664 if (s
->random_seq
) {
1667 for (i
= 0; i
< count
; i
++)
1668 s
->random_seq
[i
] *= s
->size
;
1673 /* Initialize each random sequence freelist per cache */
1674 static void __init
init_freelist_randomization(void)
1676 struct kmem_cache
*s
;
1678 mutex_lock(&slab_mutex
);
1680 list_for_each_entry(s
, &slab_caches
, list
)
1681 init_cache_random_seq(s
);
1683 mutex_unlock(&slab_mutex
);
1686 /* Get the next entry on the pre-computed freelist randomized */
1687 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1688 unsigned long *pos
, void *start
,
1689 unsigned long page_limit
,
1690 unsigned long freelist_count
)
1695 * If the target page allocation failed, the number of objects on the
1696 * page might be smaller than the usual size defined by the cache.
1699 idx
= s
->random_seq
[*pos
];
1701 if (*pos
>= freelist_count
)
1703 } while (unlikely(idx
>= page_limit
));
1705 return (char *)start
+ idx
;
1708 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1709 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1714 unsigned long idx
, pos
, page_limit
, freelist_count
;
1716 if (page
->objects
< 2 || !s
->random_seq
)
1719 freelist_count
= oo_objects(s
->oo
);
1720 pos
= get_random_int() % freelist_count
;
1722 page_limit
= page
->objects
* s
->size
;
1723 start
= fixup_red_left(s
, page_address(page
));
1725 /* First entry is used as the base of the freelist */
1726 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1728 cur
= setup_object(s
, page
, cur
);
1729 page
->freelist
= cur
;
1731 for (idx
= 1; idx
< page
->objects
; idx
++) {
1732 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1734 next
= setup_object(s
, page
, next
);
1735 set_freepointer(s
, cur
, next
);
1738 set_freepointer(s
, cur
, NULL
);
1743 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1747 static inline void init_freelist_randomization(void) { }
1748 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1752 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1754 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1757 struct kmem_cache_order_objects oo
= s
->oo
;
1759 void *start
, *p
, *next
;
1763 flags
&= gfp_allowed_mask
;
1765 if (gfpflags_allow_blocking(flags
))
1768 flags
|= s
->allocflags
;
1771 * Let the initial higher-order allocation fail under memory pressure
1772 * so we fall-back to the minimum order allocation.
1774 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1775 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1776 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1778 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1779 if (unlikely(!page
)) {
1783 * Allocation may have failed due to fragmentation.
1784 * Try a lower order alloc if possible
1786 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1787 if (unlikely(!page
))
1789 stat(s
, ORDER_FALLBACK
);
1792 page
->objects
= oo_objects(oo
);
1794 account_slab_page(page
, oo_order(oo
), s
, flags
);
1796 page
->slab_cache
= s
;
1797 __SetPageSlab(page
);
1798 if (page_is_pfmemalloc(page
))
1799 SetPageSlabPfmemalloc(page
);
1801 kasan_poison_slab(page
);
1803 start
= page_address(page
);
1805 setup_page_debug(s
, page
, start
);
1807 shuffle
= shuffle_freelist(s
, page
);
1810 start
= fixup_red_left(s
, start
);
1811 start
= setup_object(s
, page
, start
);
1812 page
->freelist
= start
;
1813 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1815 next
= setup_object(s
, page
, next
);
1816 set_freepointer(s
, p
, next
);
1819 set_freepointer(s
, p
, NULL
);
1822 page
->inuse
= page
->objects
;
1826 if (gfpflags_allow_blocking(flags
))
1827 local_irq_disable();
1831 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1836 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1838 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1839 flags
= kmalloc_fix_flags(flags
);
1841 return allocate_slab(s
,
1842 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1845 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1847 int order
= compound_order(page
);
1848 int pages
= 1 << order
;
1850 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
1853 slab_pad_check(s
, page
);
1854 for_each_object(p
, s
, page_address(page
),
1856 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1859 __ClearPageSlabPfmemalloc(page
);
1860 __ClearPageSlab(page
);
1861 /* In union with page->mapping where page allocator expects NULL */
1862 page
->slab_cache
= NULL
;
1863 if (current
->reclaim_state
)
1864 current
->reclaim_state
->reclaimed_slab
+= pages
;
1865 unaccount_slab_page(page
, order
, s
);
1866 __free_pages(page
, order
);
1869 static void rcu_free_slab(struct rcu_head
*h
)
1871 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1873 __free_slab(page
->slab_cache
, page
);
1876 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1878 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1879 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1881 __free_slab(s
, page
);
1884 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1886 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1891 * Management of partially allocated slabs.
1894 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1897 if (tail
== DEACTIVATE_TO_TAIL
)
1898 list_add_tail(&page
->slab_list
, &n
->partial
);
1900 list_add(&page
->slab_list
, &n
->partial
);
1903 static inline void add_partial(struct kmem_cache_node
*n
,
1904 struct page
*page
, int tail
)
1906 lockdep_assert_held(&n
->list_lock
);
1907 __add_partial(n
, page
, tail
);
1910 static inline void remove_partial(struct kmem_cache_node
*n
,
1913 lockdep_assert_held(&n
->list_lock
);
1914 list_del(&page
->slab_list
);
1919 * Remove slab from the partial list, freeze it and
1920 * return the pointer to the freelist.
1922 * Returns a list of objects or NULL if it fails.
1924 static inline void *acquire_slab(struct kmem_cache
*s
,
1925 struct kmem_cache_node
*n
, struct page
*page
,
1926 int mode
, int *objects
)
1929 unsigned long counters
;
1932 lockdep_assert_held(&n
->list_lock
);
1935 * Zap the freelist and set the frozen bit.
1936 * The old freelist is the list of objects for the
1937 * per cpu allocation list.
1939 freelist
= page
->freelist
;
1940 counters
= page
->counters
;
1941 new.counters
= counters
;
1942 *objects
= new.objects
- new.inuse
;
1944 new.inuse
= page
->objects
;
1945 new.freelist
= NULL
;
1947 new.freelist
= freelist
;
1950 VM_BUG_ON(new.frozen
);
1953 if (!__cmpxchg_double_slab(s
, page
,
1955 new.freelist
, new.counters
,
1959 remove_partial(n
, page
);
1964 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1965 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1968 * Try to allocate a partial slab from a specific node.
1970 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1971 struct kmem_cache_cpu
*c
, gfp_t flags
)
1973 struct page
*page
, *page2
;
1974 void *object
= NULL
;
1975 unsigned int available
= 0;
1979 * Racy check. If we mistakenly see no partial slabs then we
1980 * just allocate an empty slab. If we mistakenly try to get a
1981 * partial slab and there is none available then get_partial()
1984 if (!n
|| !n
->nr_partial
)
1987 spin_lock(&n
->list_lock
);
1988 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1991 if (!pfmemalloc_match(page
, flags
))
1994 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
1996 continue; /* cmpxchg raced */
1998 available
+= objects
;
2001 stat(s
, ALLOC_FROM_PARTIAL
);
2004 put_cpu_partial(s
, page
, 0);
2005 stat(s
, CPU_PARTIAL_NODE
);
2007 if (!kmem_cache_has_cpu_partial(s
)
2008 || available
> slub_cpu_partial(s
) / 2)
2012 spin_unlock(&n
->list_lock
);
2017 * Get a page from somewhere. Search in increasing NUMA distances.
2019 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2020 struct kmem_cache_cpu
*c
)
2023 struct zonelist
*zonelist
;
2026 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2028 unsigned int cpuset_mems_cookie
;
2031 * The defrag ratio allows a configuration of the tradeoffs between
2032 * inter node defragmentation and node local allocations. A lower
2033 * defrag_ratio increases the tendency to do local allocations
2034 * instead of attempting to obtain partial slabs from other nodes.
2036 * If the defrag_ratio is set to 0 then kmalloc() always
2037 * returns node local objects. If the ratio is higher then kmalloc()
2038 * may return off node objects because partial slabs are obtained
2039 * from other nodes and filled up.
2041 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2042 * (which makes defrag_ratio = 1000) then every (well almost)
2043 * allocation will first attempt to defrag slab caches on other nodes.
2044 * This means scanning over all nodes to look for partial slabs which
2045 * may be expensive if we do it every time we are trying to find a slab
2046 * with available objects.
2048 if (!s
->remote_node_defrag_ratio
||
2049 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2053 cpuset_mems_cookie
= read_mems_allowed_begin();
2054 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2055 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2056 struct kmem_cache_node
*n
;
2058 n
= get_node(s
, zone_to_nid(zone
));
2060 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2061 n
->nr_partial
> s
->min_partial
) {
2062 object
= get_partial_node(s
, n
, c
, flags
);
2065 * Don't check read_mems_allowed_retry()
2066 * here - if mems_allowed was updated in
2067 * parallel, that was a harmless race
2068 * between allocation and the cpuset
2075 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2076 #endif /* CONFIG_NUMA */
2081 * Get a partial page, lock it and return it.
2083 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2084 struct kmem_cache_cpu
*c
)
2087 int searchnode
= node
;
2089 if (node
== NUMA_NO_NODE
)
2090 searchnode
= numa_mem_id();
2092 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
2093 if (object
|| node
!= NUMA_NO_NODE
)
2096 return get_any_partial(s
, flags
, c
);
2099 #ifdef CONFIG_PREEMPTION
2101 * Calculate the next globally unique transaction for disambiguation
2102 * during cmpxchg. The transactions start with the cpu number and are then
2103 * incremented by CONFIG_NR_CPUS.
2105 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2108 * No preemption supported therefore also no need to check for
2114 static inline unsigned long next_tid(unsigned long tid
)
2116 return tid
+ TID_STEP
;
2119 #ifdef SLUB_DEBUG_CMPXCHG
2120 static inline unsigned int tid_to_cpu(unsigned long tid
)
2122 return tid
% TID_STEP
;
2125 static inline unsigned long tid_to_event(unsigned long tid
)
2127 return tid
/ TID_STEP
;
2131 static inline unsigned int init_tid(int cpu
)
2136 static inline void note_cmpxchg_failure(const char *n
,
2137 const struct kmem_cache
*s
, unsigned long tid
)
2139 #ifdef SLUB_DEBUG_CMPXCHG
2140 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2142 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2144 #ifdef CONFIG_PREEMPTION
2145 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2146 pr_warn("due to cpu change %d -> %d\n",
2147 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2150 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2151 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2152 tid_to_event(tid
), tid_to_event(actual_tid
));
2154 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2155 actual_tid
, tid
, next_tid(tid
));
2157 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2160 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2164 for_each_possible_cpu(cpu
)
2165 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2169 * Remove the cpu slab
2171 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2172 void *freelist
, struct kmem_cache_cpu
*c
)
2174 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2175 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2176 int lock
= 0, free_delta
= 0;
2177 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2178 void *nextfree
, *freelist_iter
, *freelist_tail
;
2179 int tail
= DEACTIVATE_TO_HEAD
;
2183 if (page
->freelist
) {
2184 stat(s
, DEACTIVATE_REMOTE_FREES
);
2185 tail
= DEACTIVATE_TO_TAIL
;
2189 * Stage one: Count the objects on cpu's freelist as free_delta and
2190 * remember the last object in freelist_tail for later splicing.
2192 freelist_tail
= NULL
;
2193 freelist_iter
= freelist
;
2194 while (freelist_iter
) {
2195 nextfree
= get_freepointer(s
, freelist_iter
);
2198 * If 'nextfree' is invalid, it is possible that the object at
2199 * 'freelist_iter' is already corrupted. So isolate all objects
2200 * starting at 'freelist_iter' by skipping them.
2202 if (freelist_corrupted(s
, page
, &freelist_iter
, nextfree
))
2205 freelist_tail
= freelist_iter
;
2208 freelist_iter
= nextfree
;
2212 * Stage two: Unfreeze the page while splicing the per-cpu
2213 * freelist to the head of page's freelist.
2215 * Ensure that the page is unfrozen while the list presence
2216 * reflects the actual number of objects during unfreeze.
2218 * We setup the list membership and then perform a cmpxchg
2219 * with the count. If there is a mismatch then the page
2220 * is not unfrozen but the page is on the wrong list.
2222 * Then we restart the process which may have to remove
2223 * the page from the list that we just put it on again
2224 * because the number of objects in the slab may have
2229 old
.freelist
= READ_ONCE(page
->freelist
);
2230 old
.counters
= READ_ONCE(page
->counters
);
2231 VM_BUG_ON(!old
.frozen
);
2233 /* Determine target state of the slab */
2234 new.counters
= old
.counters
;
2235 if (freelist_tail
) {
2236 new.inuse
-= free_delta
;
2237 set_freepointer(s
, freelist_tail
, old
.freelist
);
2238 new.freelist
= freelist
;
2240 new.freelist
= old
.freelist
;
2244 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2246 else if (new.freelist
) {
2251 * Taking the spinlock removes the possibility
2252 * that acquire_slab() will see a slab page that
2255 spin_lock(&n
->list_lock
);
2259 if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
) && !lock
) {
2262 * This also ensures that the scanning of full
2263 * slabs from diagnostic functions will not see
2266 spin_lock(&n
->list_lock
);
2272 remove_partial(n
, page
);
2273 else if (l
== M_FULL
)
2274 remove_full(s
, n
, page
);
2277 add_partial(n
, page
, tail
);
2278 else if (m
== M_FULL
)
2279 add_full(s
, n
, page
);
2283 if (!__cmpxchg_double_slab(s
, page
,
2284 old
.freelist
, old
.counters
,
2285 new.freelist
, new.counters
,
2290 spin_unlock(&n
->list_lock
);
2294 else if (m
== M_FULL
)
2295 stat(s
, DEACTIVATE_FULL
);
2296 else if (m
== M_FREE
) {
2297 stat(s
, DEACTIVATE_EMPTY
);
2298 discard_slab(s
, page
);
2307 * Unfreeze all the cpu partial slabs.
2309 * This function must be called with interrupts disabled
2310 * for the cpu using c (or some other guarantee must be there
2311 * to guarantee no concurrent accesses).
2313 static void unfreeze_partials(struct kmem_cache
*s
,
2314 struct kmem_cache_cpu
*c
)
2316 #ifdef CONFIG_SLUB_CPU_PARTIAL
2317 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2318 struct page
*page
, *discard_page
= NULL
;
2320 while ((page
= slub_percpu_partial(c
))) {
2324 slub_set_percpu_partial(c
, page
);
2326 n2
= get_node(s
, page_to_nid(page
));
2329 spin_unlock(&n
->list_lock
);
2332 spin_lock(&n
->list_lock
);
2337 old
.freelist
= page
->freelist
;
2338 old
.counters
= page
->counters
;
2339 VM_BUG_ON(!old
.frozen
);
2341 new.counters
= old
.counters
;
2342 new.freelist
= old
.freelist
;
2346 } while (!__cmpxchg_double_slab(s
, page
,
2347 old
.freelist
, old
.counters
,
2348 new.freelist
, new.counters
,
2349 "unfreezing slab"));
2351 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2352 page
->next
= discard_page
;
2353 discard_page
= page
;
2355 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2356 stat(s
, FREE_ADD_PARTIAL
);
2361 spin_unlock(&n
->list_lock
);
2363 while (discard_page
) {
2364 page
= discard_page
;
2365 discard_page
= discard_page
->next
;
2367 stat(s
, DEACTIVATE_EMPTY
);
2368 discard_slab(s
, page
);
2371 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2375 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2376 * partial page slot if available.
2378 * If we did not find a slot then simply move all the partials to the
2379 * per node partial list.
2381 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2383 #ifdef CONFIG_SLUB_CPU_PARTIAL
2384 struct page
*oldpage
;
2392 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2395 pobjects
= oldpage
->pobjects
;
2396 pages
= oldpage
->pages
;
2397 if (drain
&& pobjects
> slub_cpu_partial(s
)) {
2398 unsigned long flags
;
2400 * partial array is full. Move the existing
2401 * set to the per node partial list.
2403 local_irq_save(flags
);
2404 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2405 local_irq_restore(flags
);
2409 stat(s
, CPU_PARTIAL_DRAIN
);
2414 pobjects
+= page
->objects
- page
->inuse
;
2416 page
->pages
= pages
;
2417 page
->pobjects
= pobjects
;
2418 page
->next
= oldpage
;
2420 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2422 if (unlikely(!slub_cpu_partial(s
))) {
2423 unsigned long flags
;
2425 local_irq_save(flags
);
2426 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2427 local_irq_restore(flags
);
2430 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2433 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2435 stat(s
, CPUSLAB_FLUSH
);
2436 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2438 c
->tid
= next_tid(c
->tid
);
2444 * Called from IPI handler with interrupts disabled.
2446 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2448 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2453 unfreeze_partials(s
, c
);
2456 static void flush_cpu_slab(void *d
)
2458 struct kmem_cache
*s
= d
;
2460 __flush_cpu_slab(s
, smp_processor_id());
2463 static bool has_cpu_slab(int cpu
, void *info
)
2465 struct kmem_cache
*s
= info
;
2466 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2468 return c
->page
|| slub_percpu_partial(c
);
2471 static void flush_all(struct kmem_cache
*s
)
2473 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1);
2477 * Use the cpu notifier to insure that the cpu slabs are flushed when
2480 static int slub_cpu_dead(unsigned int cpu
)
2482 struct kmem_cache
*s
;
2483 unsigned long flags
;
2485 mutex_lock(&slab_mutex
);
2486 list_for_each_entry(s
, &slab_caches
, list
) {
2487 local_irq_save(flags
);
2488 __flush_cpu_slab(s
, cpu
);
2489 local_irq_restore(flags
);
2491 mutex_unlock(&slab_mutex
);
2496 * Check if the objects in a per cpu structure fit numa
2497 * locality expectations.
2499 static inline int node_match(struct page
*page
, int node
)
2502 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2508 #ifdef CONFIG_SLUB_DEBUG
2509 static int count_free(struct page
*page
)
2511 return page
->objects
- page
->inuse
;
2514 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2516 return atomic_long_read(&n
->total_objects
);
2518 #endif /* CONFIG_SLUB_DEBUG */
2520 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2521 static unsigned long count_partial(struct kmem_cache_node
*n
,
2522 int (*get_count
)(struct page
*))
2524 unsigned long flags
;
2525 unsigned long x
= 0;
2528 spin_lock_irqsave(&n
->list_lock
, flags
);
2529 list_for_each_entry(page
, &n
->partial
, slab_list
)
2530 x
+= get_count(page
);
2531 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2534 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2536 static noinline
void
2537 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2539 #ifdef CONFIG_SLUB_DEBUG
2540 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2541 DEFAULT_RATELIMIT_BURST
);
2543 struct kmem_cache_node
*n
;
2545 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2548 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2549 nid
, gfpflags
, &gfpflags
);
2550 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2551 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2554 if (oo_order(s
->min
) > get_order(s
->object_size
))
2555 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2558 for_each_kmem_cache_node(s
, node
, n
) {
2559 unsigned long nr_slabs
;
2560 unsigned long nr_objs
;
2561 unsigned long nr_free
;
2563 nr_free
= count_partial(n
, count_free
);
2564 nr_slabs
= node_nr_slabs(n
);
2565 nr_objs
= node_nr_objs(n
);
2567 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2568 node
, nr_slabs
, nr_objs
, nr_free
);
2573 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2574 int node
, struct kmem_cache_cpu
**pc
)
2577 struct kmem_cache_cpu
*c
= *pc
;
2580 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2582 freelist
= get_partial(s
, flags
, node
, c
);
2587 page
= new_slab(s
, flags
, node
);
2589 c
= raw_cpu_ptr(s
->cpu_slab
);
2594 * No other reference to the page yet so we can
2595 * muck around with it freely without cmpxchg
2597 freelist
= page
->freelist
;
2598 page
->freelist
= NULL
;
2600 stat(s
, ALLOC_SLAB
);
2608 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2610 if (unlikely(PageSlabPfmemalloc(page
)))
2611 return gfp_pfmemalloc_allowed(gfpflags
);
2617 * Check the page->freelist of a page and either transfer the freelist to the
2618 * per cpu freelist or deactivate the page.
2620 * The page is still frozen if the return value is not NULL.
2622 * If this function returns NULL then the page has been unfrozen.
2624 * This function must be called with interrupt disabled.
2626 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2629 unsigned long counters
;
2633 freelist
= page
->freelist
;
2634 counters
= page
->counters
;
2636 new.counters
= counters
;
2637 VM_BUG_ON(!new.frozen
);
2639 new.inuse
= page
->objects
;
2640 new.frozen
= freelist
!= NULL
;
2642 } while (!__cmpxchg_double_slab(s
, page
,
2651 * Slow path. The lockless freelist is empty or we need to perform
2654 * Processing is still very fast if new objects have been freed to the
2655 * regular freelist. In that case we simply take over the regular freelist
2656 * as the lockless freelist and zap the regular freelist.
2658 * If that is not working then we fall back to the partial lists. We take the
2659 * first element of the freelist as the object to allocate now and move the
2660 * rest of the freelist to the lockless freelist.
2662 * And if we were unable to get a new slab from the partial slab lists then
2663 * we need to allocate a new slab. This is the slowest path since it involves
2664 * a call to the page allocator and the setup of a new slab.
2666 * Version of __slab_alloc to use when we know that interrupts are
2667 * already disabled (which is the case for bulk allocation).
2669 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2670 unsigned long addr
, struct kmem_cache_cpu
*c
)
2675 stat(s
, ALLOC_SLOWPATH
);
2680 * if the node is not online or has no normal memory, just
2681 * ignore the node constraint
2683 if (unlikely(node
!= NUMA_NO_NODE
&&
2684 !node_isset(node
, slab_nodes
)))
2685 node
= NUMA_NO_NODE
;
2690 if (unlikely(!node_match(page
, node
))) {
2692 * same as above but node_match() being false already
2693 * implies node != NUMA_NO_NODE
2695 if (!node_isset(node
, slab_nodes
)) {
2696 node
= NUMA_NO_NODE
;
2699 stat(s
, ALLOC_NODE_MISMATCH
);
2700 deactivate_slab(s
, page
, c
->freelist
, c
);
2706 * By rights, we should be searching for a slab page that was
2707 * PFMEMALLOC but right now, we are losing the pfmemalloc
2708 * information when the page leaves the per-cpu allocator
2710 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2711 deactivate_slab(s
, page
, c
->freelist
, c
);
2715 /* must check again c->freelist in case of cpu migration or IRQ */
2716 freelist
= c
->freelist
;
2720 freelist
= get_freelist(s
, page
);
2724 stat(s
, DEACTIVATE_BYPASS
);
2728 stat(s
, ALLOC_REFILL
);
2732 * freelist is pointing to the list of objects to be used.
2733 * page is pointing to the page from which the objects are obtained.
2734 * That page must be frozen for per cpu allocations to work.
2736 VM_BUG_ON(!c
->page
->frozen
);
2737 c
->freelist
= get_freepointer(s
, freelist
);
2738 c
->tid
= next_tid(c
->tid
);
2743 if (slub_percpu_partial(c
)) {
2744 page
= c
->page
= slub_percpu_partial(c
);
2745 slub_set_percpu_partial(c
, page
);
2746 stat(s
, CPU_PARTIAL_ALLOC
);
2750 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2752 if (unlikely(!freelist
)) {
2753 slab_out_of_memory(s
, gfpflags
, node
);
2758 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2761 /* Only entered in the debug case */
2762 if (kmem_cache_debug(s
) &&
2763 !alloc_debug_processing(s
, page
, freelist
, addr
))
2764 goto new_slab
; /* Slab failed checks. Next slab needed */
2766 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2771 * Another one that disabled interrupt and compensates for possible
2772 * cpu changes by refetching the per cpu area pointer.
2774 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2775 unsigned long addr
, struct kmem_cache_cpu
*c
)
2778 unsigned long flags
;
2780 local_irq_save(flags
);
2781 #ifdef CONFIG_PREEMPTION
2783 * We may have been preempted and rescheduled on a different
2784 * cpu before disabling interrupts. Need to reload cpu area
2787 c
= this_cpu_ptr(s
->cpu_slab
);
2790 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2791 local_irq_restore(flags
);
2796 * If the object has been wiped upon free, make sure it's fully initialized by
2797 * zeroing out freelist pointer.
2799 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2802 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2803 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
2808 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2809 * have the fastpath folded into their functions. So no function call
2810 * overhead for requests that can be satisfied on the fastpath.
2812 * The fastpath works by first checking if the lockless freelist can be used.
2813 * If not then __slab_alloc is called for slow processing.
2815 * Otherwise we can simply pick the next object from the lockless free list.
2817 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2818 gfp_t gfpflags
, int node
, unsigned long addr
, size_t orig_size
)
2821 struct kmem_cache_cpu
*c
;
2824 struct obj_cgroup
*objcg
= NULL
;
2826 s
= slab_pre_alloc_hook(s
, &objcg
, 1, gfpflags
);
2830 object
= kfence_alloc(s
, orig_size
, gfpflags
);
2831 if (unlikely(object
))
2836 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2837 * enabled. We may switch back and forth between cpus while
2838 * reading from one cpu area. That does not matter as long
2839 * as we end up on the original cpu again when doing the cmpxchg.
2841 * We should guarantee that tid and kmem_cache are retrieved on
2842 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2843 * to check if it is matched or not.
2846 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2847 c
= raw_cpu_ptr(s
->cpu_slab
);
2848 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
2849 unlikely(tid
!= READ_ONCE(c
->tid
)));
2852 * Irqless object alloc/free algorithm used here depends on sequence
2853 * of fetching cpu_slab's data. tid should be fetched before anything
2854 * on c to guarantee that object and page associated with previous tid
2855 * won't be used with current tid. If we fetch tid first, object and
2856 * page could be one associated with next tid and our alloc/free
2857 * request will be failed. In this case, we will retry. So, no problem.
2862 * The transaction ids are globally unique per cpu and per operation on
2863 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2864 * occurs on the right processor and that there was no operation on the
2865 * linked list in between.
2868 object
= c
->freelist
;
2870 if (unlikely(!object
|| !page
|| !node_match(page
, node
))) {
2871 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2873 void *next_object
= get_freepointer_safe(s
, object
);
2876 * The cmpxchg will only match if there was no additional
2877 * operation and if we are on the right processor.
2879 * The cmpxchg does the following atomically (without lock
2881 * 1. Relocate first pointer to the current per cpu area.
2882 * 2. Verify that tid and freelist have not been changed
2883 * 3. If they were not changed replace tid and freelist
2885 * Since this is without lock semantics the protection is only
2886 * against code executing on this cpu *not* from access by
2889 if (unlikely(!this_cpu_cmpxchg_double(
2890 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2892 next_object
, next_tid(tid
)))) {
2894 note_cmpxchg_failure("slab_alloc", s
, tid
);
2897 prefetch_freepointer(s
, next_object
);
2898 stat(s
, ALLOC_FASTPATH
);
2901 maybe_wipe_obj_freeptr(s
, object
);
2903 if (unlikely(slab_want_init_on_alloc(gfpflags
, s
)) && object
)
2904 memset(kasan_reset_tag(object
), 0, s
->object_size
);
2907 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
);
2912 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2913 gfp_t gfpflags
, unsigned long addr
, size_t orig_size
)
2915 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
, orig_size
);
2918 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2920 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, s
->object_size
);
2922 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2927 EXPORT_SYMBOL(kmem_cache_alloc
);
2929 #ifdef CONFIG_TRACING
2930 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2932 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, size
);
2933 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2934 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2937 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2941 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2943 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, s
->object_size
);
2945 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2946 s
->object_size
, s
->size
, gfpflags
, node
);
2950 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2952 #ifdef CONFIG_TRACING
2953 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2955 int node
, size_t size
)
2957 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, size
);
2959 trace_kmalloc_node(_RET_IP_
, ret
,
2960 size
, s
->size
, gfpflags
, node
);
2962 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2965 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2967 #endif /* CONFIG_NUMA */
2970 * Slow path handling. This may still be called frequently since objects
2971 * have a longer lifetime than the cpu slabs in most processing loads.
2973 * So we still attempt to reduce cache line usage. Just take the slab
2974 * lock and free the item. If there is no additional partial page
2975 * handling required then we can return immediately.
2977 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2978 void *head
, void *tail
, int cnt
,
2985 unsigned long counters
;
2986 struct kmem_cache_node
*n
= NULL
;
2987 unsigned long flags
;
2989 stat(s
, FREE_SLOWPATH
);
2991 if (kfence_free(head
))
2994 if (kmem_cache_debug(s
) &&
2995 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
3000 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3003 prior
= page
->freelist
;
3004 counters
= page
->counters
;
3005 set_freepointer(s
, tail
, prior
);
3006 new.counters
= counters
;
3007 was_frozen
= new.frozen
;
3009 if ((!new.inuse
|| !prior
) && !was_frozen
) {
3011 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
3014 * Slab was on no list before and will be
3016 * We can defer the list move and instead
3021 } else { /* Needs to be taken off a list */
3023 n
= get_node(s
, page_to_nid(page
));
3025 * Speculatively acquire the list_lock.
3026 * If the cmpxchg does not succeed then we may
3027 * drop the list_lock without any processing.
3029 * Otherwise the list_lock will synchronize with
3030 * other processors updating the list of slabs.
3032 spin_lock_irqsave(&n
->list_lock
, flags
);
3037 } while (!cmpxchg_double_slab(s
, page
,
3044 if (likely(was_frozen
)) {
3046 * The list lock was not taken therefore no list
3047 * activity can be necessary.
3049 stat(s
, FREE_FROZEN
);
3050 } else if (new.frozen
) {
3052 * If we just froze the page then put it onto the
3053 * per cpu partial list.
3055 put_cpu_partial(s
, page
, 1);
3056 stat(s
, CPU_PARTIAL_FREE
);
3062 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3066 * Objects left in the slab. If it was not on the partial list before
3069 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3070 remove_full(s
, n
, page
);
3071 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
3072 stat(s
, FREE_ADD_PARTIAL
);
3074 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3080 * Slab on the partial list.
3082 remove_partial(n
, page
);
3083 stat(s
, FREE_REMOVE_PARTIAL
);
3085 /* Slab must be on the full list */
3086 remove_full(s
, n
, page
);
3089 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3091 discard_slab(s
, page
);
3095 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3096 * can perform fastpath freeing without additional function calls.
3098 * The fastpath is only possible if we are freeing to the current cpu slab
3099 * of this processor. This typically the case if we have just allocated
3102 * If fastpath is not possible then fall back to __slab_free where we deal
3103 * with all sorts of special processing.
3105 * Bulk free of a freelist with several objects (all pointing to the
3106 * same page) possible by specifying head and tail ptr, plus objects
3107 * count (cnt). Bulk free indicated by tail pointer being set.
3109 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3110 struct page
*page
, void *head
, void *tail
,
3111 int cnt
, unsigned long addr
)
3113 void *tail_obj
= tail
? : head
;
3114 struct kmem_cache_cpu
*c
;
3117 memcg_slab_free_hook(s
, &head
, 1);
3120 * Determine the currently cpus per cpu slab.
3121 * The cpu may change afterward. However that does not matter since
3122 * data is retrieved via this pointer. If we are on the same cpu
3123 * during the cmpxchg then the free will succeed.
3126 tid
= this_cpu_read(s
->cpu_slab
->tid
);
3127 c
= raw_cpu_ptr(s
->cpu_slab
);
3128 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
3129 unlikely(tid
!= READ_ONCE(c
->tid
)));
3131 /* Same with comment on barrier() in slab_alloc_node() */
3134 if (likely(page
== c
->page
)) {
3135 void **freelist
= READ_ONCE(c
->freelist
);
3137 set_freepointer(s
, tail_obj
, freelist
);
3139 if (unlikely(!this_cpu_cmpxchg_double(
3140 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3142 head
, next_tid(tid
)))) {
3144 note_cmpxchg_failure("slab_free", s
, tid
);
3147 stat(s
, FREE_FASTPATH
);
3149 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3153 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3154 void *head
, void *tail
, int cnt
,
3158 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3159 * to remove objects, whose reuse must be delayed.
3161 if (slab_free_freelist_hook(s
, &head
, &tail
))
3162 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3165 #ifdef CONFIG_KASAN_GENERIC
3166 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3168 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3172 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3174 s
= cache_from_obj(s
, x
);
3177 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3178 trace_kmem_cache_free(_RET_IP_
, x
, s
->name
);
3180 EXPORT_SYMBOL(kmem_cache_free
);
3182 struct detached_freelist
{
3187 struct kmem_cache
*s
;
3191 * This function progressively scans the array with free objects (with
3192 * a limited look ahead) and extract objects belonging to the same
3193 * page. It builds a detached freelist directly within the given
3194 * page/objects. This can happen without any need for
3195 * synchronization, because the objects are owned by running process.
3196 * The freelist is build up as a single linked list in the objects.
3197 * The idea is, that this detached freelist can then be bulk
3198 * transferred to the real freelist(s), but only requiring a single
3199 * synchronization primitive. Look ahead in the array is limited due
3200 * to performance reasons.
3203 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3204 void **p
, struct detached_freelist
*df
)
3206 size_t first_skipped_index
= 0;
3211 /* Always re-init detached_freelist */
3216 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3217 } while (!object
&& size
);
3222 page
= virt_to_head_page(object
);
3224 /* Handle kalloc'ed objects */
3225 if (unlikely(!PageSlab(page
))) {
3226 BUG_ON(!PageCompound(page
));
3228 __free_pages(page
, compound_order(page
));
3229 p
[size
] = NULL
; /* mark object processed */
3232 /* Derive kmem_cache from object */
3233 df
->s
= page
->slab_cache
;
3235 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3238 if (is_kfence_address(object
)) {
3239 slab_free_hook(df
->s
, object
);
3240 __kfence_free(object
);
3241 p
[size
] = NULL
; /* mark object processed */
3245 /* Start new detached freelist */
3247 set_freepointer(df
->s
, object
, NULL
);
3249 df
->freelist
= object
;
3250 p
[size
] = NULL
; /* mark object processed */
3256 continue; /* Skip processed objects */
3258 /* df->page is always set at this point */
3259 if (df
->page
== virt_to_head_page(object
)) {
3260 /* Opportunity build freelist */
3261 set_freepointer(df
->s
, object
, df
->freelist
);
3262 df
->freelist
= object
;
3264 p
[size
] = NULL
; /* mark object processed */
3269 /* Limit look ahead search */
3273 if (!first_skipped_index
)
3274 first_skipped_index
= size
+ 1;
3277 return first_skipped_index
;
3280 /* Note that interrupts must be enabled when calling this function. */
3281 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3286 memcg_slab_free_hook(s
, p
, size
);
3288 struct detached_freelist df
;
3290 size
= build_detached_freelist(s
, size
, p
, &df
);
3294 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
, _RET_IP_
);
3295 } while (likely(size
));
3297 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3299 /* Note that interrupts must be enabled when calling this function. */
3300 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3303 struct kmem_cache_cpu
*c
;
3305 struct obj_cgroup
*objcg
= NULL
;
3307 /* memcg and kmem_cache debug support */
3308 s
= slab_pre_alloc_hook(s
, &objcg
, size
, flags
);
3312 * Drain objects in the per cpu slab, while disabling local
3313 * IRQs, which protects against PREEMPT and interrupts
3314 * handlers invoking normal fastpath.
3316 local_irq_disable();
3317 c
= this_cpu_ptr(s
->cpu_slab
);
3319 for (i
= 0; i
< size
; i
++) {
3320 void *object
= kfence_alloc(s
, s
->object_size
, flags
);
3322 if (unlikely(object
)) {
3327 object
= c
->freelist
;
3328 if (unlikely(!object
)) {
3330 * We may have removed an object from c->freelist using
3331 * the fastpath in the previous iteration; in that case,
3332 * c->tid has not been bumped yet.
3333 * Since ___slab_alloc() may reenable interrupts while
3334 * allocating memory, we should bump c->tid now.
3336 c
->tid
= next_tid(c
->tid
);
3339 * Invoking slow path likely have side-effect
3340 * of re-populating per CPU c->freelist
3342 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3344 if (unlikely(!p
[i
]))
3347 c
= this_cpu_ptr(s
->cpu_slab
);
3348 maybe_wipe_obj_freeptr(s
, p
[i
]);
3350 continue; /* goto for-loop */
3352 c
->freelist
= get_freepointer(s
, object
);
3354 maybe_wipe_obj_freeptr(s
, p
[i
]);
3356 c
->tid
= next_tid(c
->tid
);
3359 /* Clear memory outside IRQ disabled fastpath loop */
3360 if (unlikely(slab_want_init_on_alloc(flags
, s
))) {
3363 for (j
= 0; j
< i
; j
++)
3364 memset(kasan_reset_tag(p
[j
]), 0, s
->object_size
);
3367 /* memcg and kmem_cache debug support */
3368 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
);
3372 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
);
3373 __kmem_cache_free_bulk(s
, i
, p
);
3376 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3380 * Object placement in a slab is made very easy because we always start at
3381 * offset 0. If we tune the size of the object to the alignment then we can
3382 * get the required alignment by putting one properly sized object after
3385 * Notice that the allocation order determines the sizes of the per cpu
3386 * caches. Each processor has always one slab available for allocations.
3387 * Increasing the allocation order reduces the number of times that slabs
3388 * must be moved on and off the partial lists and is therefore a factor in
3393 * Mininum / Maximum order of slab pages. This influences locking overhead
3394 * and slab fragmentation. A higher order reduces the number of partial slabs
3395 * and increases the number of allocations possible without having to
3396 * take the list_lock.
3398 static unsigned int slub_min_order
;
3399 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3400 static unsigned int slub_min_objects
;
3403 * Calculate the order of allocation given an slab object size.
3405 * The order of allocation has significant impact on performance and other
3406 * system components. Generally order 0 allocations should be preferred since
3407 * order 0 does not cause fragmentation in the page allocator. Larger objects
3408 * be problematic to put into order 0 slabs because there may be too much
3409 * unused space left. We go to a higher order if more than 1/16th of the slab
3412 * In order to reach satisfactory performance we must ensure that a minimum
3413 * number of objects is in one slab. Otherwise we may generate too much
3414 * activity on the partial lists which requires taking the list_lock. This is
3415 * less a concern for large slabs though which are rarely used.
3417 * slub_max_order specifies the order where we begin to stop considering the
3418 * number of objects in a slab as critical. If we reach slub_max_order then
3419 * we try to keep the page order as low as possible. So we accept more waste
3420 * of space in favor of a small page order.
3422 * Higher order allocations also allow the placement of more objects in a
3423 * slab and thereby reduce object handling overhead. If the user has
3424 * requested a higher mininum order then we start with that one instead of
3425 * the smallest order which will fit the object.
3427 static inline unsigned int slab_order(unsigned int size
,
3428 unsigned int min_objects
, unsigned int max_order
,
3429 unsigned int fract_leftover
)
3431 unsigned int min_order
= slub_min_order
;
3434 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3435 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3437 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3438 order
<= max_order
; order
++) {
3440 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3443 rem
= slab_size
% size
;
3445 if (rem
<= slab_size
/ fract_leftover
)
3452 static inline int calculate_order(unsigned int size
)
3455 unsigned int min_objects
;
3456 unsigned int max_objects
;
3457 unsigned int nr_cpus
;
3460 * Attempt to find best configuration for a slab. This
3461 * works by first attempting to generate a layout with
3462 * the best configuration and backing off gradually.
3464 * First we increase the acceptable waste in a slab. Then
3465 * we reduce the minimum objects required in a slab.
3467 min_objects
= slub_min_objects
;
3470 * Some architectures will only update present cpus when
3471 * onlining them, so don't trust the number if it's just 1. But
3472 * we also don't want to use nr_cpu_ids always, as on some other
3473 * architectures, there can be many possible cpus, but never
3474 * onlined. Here we compromise between trying to avoid too high
3475 * order on systems that appear larger than they are, and too
3476 * low order on systems that appear smaller than they are.
3478 nr_cpus
= num_present_cpus();
3480 nr_cpus
= nr_cpu_ids
;
3481 min_objects
= 4 * (fls(nr_cpus
) + 1);
3483 max_objects
= order_objects(slub_max_order
, size
);
3484 min_objects
= min(min_objects
, max_objects
);
3486 while (min_objects
> 1) {
3487 unsigned int fraction
;
3490 while (fraction
>= 4) {
3491 order
= slab_order(size
, min_objects
,
3492 slub_max_order
, fraction
);
3493 if (order
<= slub_max_order
)
3501 * We were unable to place multiple objects in a slab. Now
3502 * lets see if we can place a single object there.
3504 order
= slab_order(size
, 1, slub_max_order
, 1);
3505 if (order
<= slub_max_order
)
3509 * Doh this slab cannot be placed using slub_max_order.
3511 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3512 if (order
< MAX_ORDER
)
3518 init_kmem_cache_node(struct kmem_cache_node
*n
)
3521 spin_lock_init(&n
->list_lock
);
3522 INIT_LIST_HEAD(&n
->partial
);
3523 #ifdef CONFIG_SLUB_DEBUG
3524 atomic_long_set(&n
->nr_slabs
, 0);
3525 atomic_long_set(&n
->total_objects
, 0);
3526 INIT_LIST_HEAD(&n
->full
);
3530 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3532 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3533 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3536 * Must align to double word boundary for the double cmpxchg
3537 * instructions to work; see __pcpu_double_call_return_bool().
3539 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3540 2 * sizeof(void *));
3545 init_kmem_cache_cpus(s
);
3550 static struct kmem_cache
*kmem_cache_node
;
3553 * No kmalloc_node yet so do it by hand. We know that this is the first
3554 * slab on the node for this slabcache. There are no concurrent accesses
3557 * Note that this function only works on the kmem_cache_node
3558 * when allocating for the kmem_cache_node. This is used for bootstrapping
3559 * memory on a fresh node that has no slab structures yet.
3561 static void early_kmem_cache_node_alloc(int node
)
3564 struct kmem_cache_node
*n
;
3566 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3568 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3571 if (page_to_nid(page
) != node
) {
3572 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3573 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3578 #ifdef CONFIG_SLUB_DEBUG
3579 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3580 init_tracking(kmem_cache_node
, n
);
3582 n
= kasan_slab_alloc(kmem_cache_node
, n
, GFP_KERNEL
);
3583 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3586 kmem_cache_node
->node
[node
] = n
;
3587 init_kmem_cache_node(n
);
3588 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3591 * No locks need to be taken here as it has just been
3592 * initialized and there is no concurrent access.
3594 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3597 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3600 struct kmem_cache_node
*n
;
3602 for_each_kmem_cache_node(s
, node
, n
) {
3603 s
->node
[node
] = NULL
;
3604 kmem_cache_free(kmem_cache_node
, n
);
3608 void __kmem_cache_release(struct kmem_cache
*s
)
3610 cache_random_seq_destroy(s
);
3611 free_percpu(s
->cpu_slab
);
3612 free_kmem_cache_nodes(s
);
3615 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3619 for_each_node_mask(node
, slab_nodes
) {
3620 struct kmem_cache_node
*n
;
3622 if (slab_state
== DOWN
) {
3623 early_kmem_cache_node_alloc(node
);
3626 n
= kmem_cache_alloc_node(kmem_cache_node
,
3630 free_kmem_cache_nodes(s
);
3634 init_kmem_cache_node(n
);
3640 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3642 if (min
< MIN_PARTIAL
)
3644 else if (min
> MAX_PARTIAL
)
3646 s
->min_partial
= min
;
3649 static void set_cpu_partial(struct kmem_cache
*s
)
3651 #ifdef CONFIG_SLUB_CPU_PARTIAL
3653 * cpu_partial determined the maximum number of objects kept in the
3654 * per cpu partial lists of a processor.
3656 * Per cpu partial lists mainly contain slabs that just have one
3657 * object freed. If they are used for allocation then they can be
3658 * filled up again with minimal effort. The slab will never hit the
3659 * per node partial lists and therefore no locking will be required.
3661 * This setting also determines
3663 * A) The number of objects from per cpu partial slabs dumped to the
3664 * per node list when we reach the limit.
3665 * B) The number of objects in cpu partial slabs to extract from the
3666 * per node list when we run out of per cpu objects. We only fetch
3667 * 50% to keep some capacity around for frees.
3669 if (!kmem_cache_has_cpu_partial(s
))
3670 slub_set_cpu_partial(s
, 0);
3671 else if (s
->size
>= PAGE_SIZE
)
3672 slub_set_cpu_partial(s
, 2);
3673 else if (s
->size
>= 1024)
3674 slub_set_cpu_partial(s
, 6);
3675 else if (s
->size
>= 256)
3676 slub_set_cpu_partial(s
, 13);
3678 slub_set_cpu_partial(s
, 30);
3683 * calculate_sizes() determines the order and the distribution of data within
3686 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3688 slab_flags_t flags
= s
->flags
;
3689 unsigned int size
= s
->object_size
;
3690 unsigned int freepointer_area
;
3694 * Round up object size to the next word boundary. We can only
3695 * place the free pointer at word boundaries and this determines
3696 * the possible location of the free pointer.
3698 size
= ALIGN(size
, sizeof(void *));
3700 * This is the area of the object where a freepointer can be
3701 * safely written. If redzoning adds more to the inuse size, we
3702 * can't use that portion for writing the freepointer, so
3703 * s->offset must be limited within this for the general case.
3705 freepointer_area
= size
;
3707 #ifdef CONFIG_SLUB_DEBUG
3709 * Determine if we can poison the object itself. If the user of
3710 * the slab may touch the object after free or before allocation
3711 * then we should never poison the object itself.
3713 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3715 s
->flags
|= __OBJECT_POISON
;
3717 s
->flags
&= ~__OBJECT_POISON
;
3721 * If we are Redzoning then check if there is some space between the
3722 * end of the object and the free pointer. If not then add an
3723 * additional word to have some bytes to store Redzone information.
3725 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3726 size
+= sizeof(void *);
3730 * With that we have determined the number of bytes in actual use
3731 * by the object. This is the potential offset to the free pointer.
3735 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3738 * Relocate free pointer after the object if it is not
3739 * permitted to overwrite the first word of the object on
3742 * This is the case if we do RCU, have a constructor or
3743 * destructor or are poisoning the objects.
3745 * The assumption that s->offset >= s->inuse means free
3746 * pointer is outside of the object is used in the
3747 * freeptr_outside_object() function. If that is no
3748 * longer true, the function needs to be modified.
3751 size
+= sizeof(void *);
3752 } else if (freepointer_area
> sizeof(void *)) {
3754 * Store freelist pointer near middle of object to keep
3755 * it away from the edges of the object to avoid small
3756 * sized over/underflows from neighboring allocations.
3758 s
->offset
= ALIGN(freepointer_area
/ 2, sizeof(void *));
3761 #ifdef CONFIG_SLUB_DEBUG
3762 if (flags
& SLAB_STORE_USER
)
3764 * Need to store information about allocs and frees after
3767 size
+= 2 * sizeof(struct track
);
3770 kasan_cache_create(s
, &size
, &s
->flags
);
3771 #ifdef CONFIG_SLUB_DEBUG
3772 if (flags
& SLAB_RED_ZONE
) {
3774 * Add some empty padding so that we can catch
3775 * overwrites from earlier objects rather than let
3776 * tracking information or the free pointer be
3777 * corrupted if a user writes before the start
3780 size
+= sizeof(void *);
3782 s
->red_left_pad
= sizeof(void *);
3783 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3784 size
+= s
->red_left_pad
;
3789 * SLUB stores one object immediately after another beginning from
3790 * offset 0. In order to align the objects we have to simply size
3791 * each object to conform to the alignment.
3793 size
= ALIGN(size
, s
->align
);
3795 s
->reciprocal_size
= reciprocal_value(size
);
3796 if (forced_order
>= 0)
3797 order
= forced_order
;
3799 order
= calculate_order(size
);
3806 s
->allocflags
|= __GFP_COMP
;
3808 if (s
->flags
& SLAB_CACHE_DMA
)
3809 s
->allocflags
|= GFP_DMA
;
3811 if (s
->flags
& SLAB_CACHE_DMA32
)
3812 s
->allocflags
|= GFP_DMA32
;
3814 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3815 s
->allocflags
|= __GFP_RECLAIMABLE
;
3818 * Determine the number of objects per slab
3820 s
->oo
= oo_make(order
, size
);
3821 s
->min
= oo_make(get_order(size
), size
);
3822 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3825 return !!oo_objects(s
->oo
);
3828 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3830 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
);
3831 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3832 s
->random
= get_random_long();
3835 if (!calculate_sizes(s
, -1))
3837 if (disable_higher_order_debug
) {
3839 * Disable debugging flags that store metadata if the min slab
3842 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3843 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3845 if (!calculate_sizes(s
, -1))
3850 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3851 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3852 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3853 /* Enable fast mode */
3854 s
->flags
|= __CMPXCHG_DOUBLE
;
3858 * The larger the object size is, the more pages we want on the partial
3859 * list to avoid pounding the page allocator excessively.
3861 set_min_partial(s
, ilog2(s
->size
) / 2);
3866 s
->remote_node_defrag_ratio
= 1000;
3869 /* Initialize the pre-computed randomized freelist if slab is up */
3870 if (slab_state
>= UP
) {
3871 if (init_cache_random_seq(s
))
3875 if (!init_kmem_cache_nodes(s
))
3878 if (alloc_kmem_cache_cpus(s
))
3881 free_kmem_cache_nodes(s
);
3886 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3889 #ifdef CONFIG_SLUB_DEBUG
3890 void *addr
= page_address(page
);
3894 slab_err(s
, page
, text
, s
->name
);
3897 map
= get_map(s
, page
);
3898 for_each_object(p
, s
, addr
, page
->objects
) {
3900 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
3901 pr_err("INFO: Object 0x%p @offset=%tu\n", p
, p
- addr
);
3902 print_tracking(s
, p
);
3911 * Attempt to free all partial slabs on a node.
3912 * This is called from __kmem_cache_shutdown(). We must take list_lock
3913 * because sysfs file might still access partial list after the shutdowning.
3915 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3918 struct page
*page
, *h
;
3920 BUG_ON(irqs_disabled());
3921 spin_lock_irq(&n
->list_lock
);
3922 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3924 remove_partial(n
, page
);
3925 list_add(&page
->slab_list
, &discard
);
3927 list_slab_objects(s
, page
,
3928 "Objects remaining in %s on __kmem_cache_shutdown()");
3931 spin_unlock_irq(&n
->list_lock
);
3933 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3934 discard_slab(s
, page
);
3937 bool __kmem_cache_empty(struct kmem_cache
*s
)
3940 struct kmem_cache_node
*n
;
3942 for_each_kmem_cache_node(s
, node
, n
)
3943 if (n
->nr_partial
|| slabs_node(s
, node
))
3949 * Release all resources used by a slab cache.
3951 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3954 struct kmem_cache_node
*n
;
3957 /* Attempt to free all objects */
3958 for_each_kmem_cache_node(s
, node
, n
) {
3960 if (n
->nr_partial
|| slabs_node(s
, node
))
3966 void kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct page
*page
)
3969 int __maybe_unused i
;
3973 struct kmem_cache
*s
= page
->slab_cache
;
3974 struct track __maybe_unused
*trackp
;
3976 kpp
->kp_ptr
= object
;
3977 kpp
->kp_page
= page
;
3978 kpp
->kp_slab_cache
= s
;
3979 base
= page_address(page
);
3980 objp0
= kasan_reset_tag(object
);
3981 #ifdef CONFIG_SLUB_DEBUG
3982 objp
= restore_red_left(s
, objp0
);
3986 objnr
= obj_to_index(s
, page
, objp
);
3987 kpp
->kp_data_offset
= (unsigned long)((char *)objp0
- (char *)objp
);
3988 objp
= base
+ s
->size
* objnr
;
3989 kpp
->kp_objp
= objp
;
3990 if (WARN_ON_ONCE(objp
< base
|| objp
>= base
+ page
->objects
* s
->size
|| (objp
- base
) % s
->size
) ||
3991 !(s
->flags
& SLAB_STORE_USER
))
3993 #ifdef CONFIG_SLUB_DEBUG
3994 trackp
= get_track(s
, objp
, TRACK_ALLOC
);
3995 kpp
->kp_ret
= (void *)trackp
->addr
;
3996 #ifdef CONFIG_STACKTRACE
3997 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< TRACK_ADDRS_COUNT
; i
++) {
3998 kpp
->kp_stack
[i
] = (void *)trackp
->addrs
[i
];
3999 if (!kpp
->kp_stack
[i
])
4006 /********************************************************************
4008 *******************************************************************/
4010 static int __init
setup_slub_min_order(char *str
)
4012 get_option(&str
, (int *)&slub_min_order
);
4017 __setup("slub_min_order=", setup_slub_min_order
);
4019 static int __init
setup_slub_max_order(char *str
)
4021 get_option(&str
, (int *)&slub_max_order
);
4022 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
4027 __setup("slub_max_order=", setup_slub_max_order
);
4029 static int __init
setup_slub_min_objects(char *str
)
4031 get_option(&str
, (int *)&slub_min_objects
);
4036 __setup("slub_min_objects=", setup_slub_min_objects
);
4038 void *__kmalloc(size_t size
, gfp_t flags
)
4040 struct kmem_cache
*s
;
4043 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4044 return kmalloc_large(size
, flags
);
4046 s
= kmalloc_slab(size
, flags
);
4048 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4051 ret
= slab_alloc(s
, flags
, _RET_IP_
, size
);
4053 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
4055 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4059 EXPORT_SYMBOL(__kmalloc
);
4062 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
4066 unsigned int order
= get_order(size
);
4068 flags
|= __GFP_COMP
;
4069 page
= alloc_pages_node(node
, flags
, order
);
4071 ptr
= page_address(page
);
4072 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4073 PAGE_SIZE
<< order
);
4076 return kmalloc_large_node_hook(ptr
, size
, flags
);
4079 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4081 struct kmem_cache
*s
;
4084 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4085 ret
= kmalloc_large_node(size
, flags
, node
);
4087 trace_kmalloc_node(_RET_IP_
, ret
,
4088 size
, PAGE_SIZE
<< get_order(size
),
4094 s
= kmalloc_slab(size
, flags
);
4096 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4099 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
, size
);
4101 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
4103 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4107 EXPORT_SYMBOL(__kmalloc_node
);
4108 #endif /* CONFIG_NUMA */
4110 #ifdef CONFIG_HARDENED_USERCOPY
4112 * Rejects incorrectly sized objects and objects that are to be copied
4113 * to/from userspace but do not fall entirely within the containing slab
4114 * cache's usercopy region.
4116 * Returns NULL if check passes, otherwise const char * to name of cache
4117 * to indicate an error.
4119 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4122 struct kmem_cache
*s
;
4123 unsigned int offset
;
4125 bool is_kfence
= is_kfence_address(ptr
);
4127 ptr
= kasan_reset_tag(ptr
);
4129 /* Find object and usable object size. */
4130 s
= page
->slab_cache
;
4132 /* Reject impossible pointers. */
4133 if (ptr
< page_address(page
))
4134 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4137 /* Find offset within object. */
4139 offset
= ptr
- kfence_object_start(ptr
);
4141 offset
= (ptr
- page_address(page
)) % s
->size
;
4143 /* Adjust for redzone and reject if within the redzone. */
4144 if (!is_kfence
&& kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4145 if (offset
< s
->red_left_pad
)
4146 usercopy_abort("SLUB object in left red zone",
4147 s
->name
, to_user
, offset
, n
);
4148 offset
-= s
->red_left_pad
;
4151 /* Allow address range falling entirely within usercopy region. */
4152 if (offset
>= s
->useroffset
&&
4153 offset
- s
->useroffset
<= s
->usersize
&&
4154 n
<= s
->useroffset
- offset
+ s
->usersize
)
4158 * If the copy is still within the allocated object, produce
4159 * a warning instead of rejecting the copy. This is intended
4160 * to be a temporary method to find any missing usercopy
4163 object_size
= slab_ksize(s
);
4164 if (usercopy_fallback
&&
4165 offset
<= object_size
&& n
<= object_size
- offset
) {
4166 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4170 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4172 #endif /* CONFIG_HARDENED_USERCOPY */
4174 size_t __ksize(const void *object
)
4178 if (unlikely(object
== ZERO_SIZE_PTR
))
4181 page
= virt_to_head_page(object
);
4183 if (unlikely(!PageSlab(page
))) {
4184 WARN_ON(!PageCompound(page
));
4185 return page_size(page
);
4188 return slab_ksize(page
->slab_cache
);
4190 EXPORT_SYMBOL(__ksize
);
4192 void kfree(const void *x
)
4195 void *object
= (void *)x
;
4197 trace_kfree(_RET_IP_
, x
);
4199 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4202 page
= virt_to_head_page(x
);
4203 if (unlikely(!PageSlab(page
))) {
4204 unsigned int order
= compound_order(page
);
4206 BUG_ON(!PageCompound(page
));
4208 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4209 -(PAGE_SIZE
<< order
));
4210 __free_pages(page
, order
);
4213 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4215 EXPORT_SYMBOL(kfree
);
4217 #define SHRINK_PROMOTE_MAX 32
4220 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4221 * up most to the head of the partial lists. New allocations will then
4222 * fill those up and thus they can be removed from the partial lists.
4224 * The slabs with the least items are placed last. This results in them
4225 * being allocated from last increasing the chance that the last objects
4226 * are freed in them.
4228 int __kmem_cache_shrink(struct kmem_cache
*s
)
4232 struct kmem_cache_node
*n
;
4235 struct list_head discard
;
4236 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4237 unsigned long flags
;
4241 for_each_kmem_cache_node(s
, node
, n
) {
4242 INIT_LIST_HEAD(&discard
);
4243 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4244 INIT_LIST_HEAD(promote
+ i
);
4246 spin_lock_irqsave(&n
->list_lock
, flags
);
4249 * Build lists of slabs to discard or promote.
4251 * Note that concurrent frees may occur while we hold the
4252 * list_lock. page->inuse here is the upper limit.
4254 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4255 int free
= page
->objects
- page
->inuse
;
4257 /* Do not reread page->inuse */
4260 /* We do not keep full slabs on the list */
4263 if (free
== page
->objects
) {
4264 list_move(&page
->slab_list
, &discard
);
4266 } else if (free
<= SHRINK_PROMOTE_MAX
)
4267 list_move(&page
->slab_list
, promote
+ free
- 1);
4271 * Promote the slabs filled up most to the head of the
4274 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4275 list_splice(promote
+ i
, &n
->partial
);
4277 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4279 /* Release empty slabs */
4280 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4281 discard_slab(s
, page
);
4283 if (slabs_node(s
, node
))
4290 static int slab_mem_going_offline_callback(void *arg
)
4292 struct kmem_cache
*s
;
4294 mutex_lock(&slab_mutex
);
4295 list_for_each_entry(s
, &slab_caches
, list
)
4296 __kmem_cache_shrink(s
);
4297 mutex_unlock(&slab_mutex
);
4302 static void slab_mem_offline_callback(void *arg
)
4304 struct memory_notify
*marg
= arg
;
4307 offline_node
= marg
->status_change_nid_normal
;
4310 * If the node still has available memory. we need kmem_cache_node
4313 if (offline_node
< 0)
4316 mutex_lock(&slab_mutex
);
4317 node_clear(offline_node
, slab_nodes
);
4319 * We no longer free kmem_cache_node structures here, as it would be
4320 * racy with all get_node() users, and infeasible to protect them with
4323 mutex_unlock(&slab_mutex
);
4326 static int slab_mem_going_online_callback(void *arg
)
4328 struct kmem_cache_node
*n
;
4329 struct kmem_cache
*s
;
4330 struct memory_notify
*marg
= arg
;
4331 int nid
= marg
->status_change_nid_normal
;
4335 * If the node's memory is already available, then kmem_cache_node is
4336 * already created. Nothing to do.
4342 * We are bringing a node online. No memory is available yet. We must
4343 * allocate a kmem_cache_node structure in order to bring the node
4346 mutex_lock(&slab_mutex
);
4347 list_for_each_entry(s
, &slab_caches
, list
) {
4349 * The structure may already exist if the node was previously
4350 * onlined and offlined.
4352 if (get_node(s
, nid
))
4355 * XXX: kmem_cache_alloc_node will fallback to other nodes
4356 * since memory is not yet available from the node that
4359 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4364 init_kmem_cache_node(n
);
4368 * Any cache created after this point will also have kmem_cache_node
4369 * initialized for the new node.
4371 node_set(nid
, slab_nodes
);
4373 mutex_unlock(&slab_mutex
);
4377 static int slab_memory_callback(struct notifier_block
*self
,
4378 unsigned long action
, void *arg
)
4383 case MEM_GOING_ONLINE
:
4384 ret
= slab_mem_going_online_callback(arg
);
4386 case MEM_GOING_OFFLINE
:
4387 ret
= slab_mem_going_offline_callback(arg
);
4390 case MEM_CANCEL_ONLINE
:
4391 slab_mem_offline_callback(arg
);
4394 case MEM_CANCEL_OFFLINE
:
4398 ret
= notifier_from_errno(ret
);
4404 static struct notifier_block slab_memory_callback_nb
= {
4405 .notifier_call
= slab_memory_callback
,
4406 .priority
= SLAB_CALLBACK_PRI
,
4409 /********************************************************************
4410 * Basic setup of slabs
4411 *******************************************************************/
4414 * Used for early kmem_cache structures that were allocated using
4415 * the page allocator. Allocate them properly then fix up the pointers
4416 * that may be pointing to the wrong kmem_cache structure.
4419 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4422 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4423 struct kmem_cache_node
*n
;
4425 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4428 * This runs very early, and only the boot processor is supposed to be
4429 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4432 __flush_cpu_slab(s
, smp_processor_id());
4433 for_each_kmem_cache_node(s
, node
, n
) {
4436 list_for_each_entry(p
, &n
->partial
, slab_list
)
4439 #ifdef CONFIG_SLUB_DEBUG
4440 list_for_each_entry(p
, &n
->full
, slab_list
)
4444 list_add(&s
->list
, &slab_caches
);
4448 void __init
kmem_cache_init(void)
4450 static __initdata
struct kmem_cache boot_kmem_cache
,
4451 boot_kmem_cache_node
;
4454 if (debug_guardpage_minorder())
4457 kmem_cache_node
= &boot_kmem_cache_node
;
4458 kmem_cache
= &boot_kmem_cache
;
4461 * Initialize the nodemask for which we will allocate per node
4462 * structures. Here we don't need taking slab_mutex yet.
4464 for_each_node_state(node
, N_NORMAL_MEMORY
)
4465 node_set(node
, slab_nodes
);
4467 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4468 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4470 register_hotmemory_notifier(&slab_memory_callback_nb
);
4472 /* Able to allocate the per node structures */
4473 slab_state
= PARTIAL
;
4475 create_boot_cache(kmem_cache
, "kmem_cache",
4476 offsetof(struct kmem_cache
, node
) +
4477 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4478 SLAB_HWCACHE_ALIGN
, 0, 0);
4480 kmem_cache
= bootstrap(&boot_kmem_cache
);
4481 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4483 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4484 setup_kmalloc_cache_index_table();
4485 create_kmalloc_caches(0);
4487 /* Setup random freelists for each cache */
4488 init_freelist_randomization();
4490 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4493 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4495 slub_min_order
, slub_max_order
, slub_min_objects
,
4496 nr_cpu_ids
, nr_node_ids
);
4499 void __init
kmem_cache_init_late(void)
4504 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4505 slab_flags_t flags
, void (*ctor
)(void *))
4507 struct kmem_cache
*s
;
4509 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4514 * Adjust the object sizes so that we clear
4515 * the complete object on kzalloc.
4517 s
->object_size
= max(s
->object_size
, size
);
4518 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4520 if (sysfs_slab_alias(s
, name
)) {
4529 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4533 err
= kmem_cache_open(s
, flags
);
4537 /* Mutex is not taken during early boot */
4538 if (slab_state
<= UP
)
4541 err
= sysfs_slab_add(s
);
4543 __kmem_cache_release(s
);
4548 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4550 struct kmem_cache
*s
;
4553 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4554 return kmalloc_large(size
, gfpflags
);
4556 s
= kmalloc_slab(size
, gfpflags
);
4558 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4561 ret
= slab_alloc(s
, gfpflags
, caller
, size
);
4563 /* Honor the call site pointer we received. */
4564 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4568 EXPORT_SYMBOL(__kmalloc_track_caller
);
4571 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4572 int node
, unsigned long caller
)
4574 struct kmem_cache
*s
;
4577 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4578 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4580 trace_kmalloc_node(caller
, ret
,
4581 size
, PAGE_SIZE
<< get_order(size
),
4587 s
= kmalloc_slab(size
, gfpflags
);
4589 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4592 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
, size
);
4594 /* Honor the call site pointer we received. */
4595 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4599 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4603 static int count_inuse(struct page
*page
)
4608 static int count_total(struct page
*page
)
4610 return page
->objects
;
4614 #ifdef CONFIG_SLUB_DEBUG
4615 static void validate_slab(struct kmem_cache
*s
, struct page
*page
)
4618 void *addr
= page_address(page
);
4623 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4626 /* Now we know that a valid freelist exists */
4627 map
= get_map(s
, page
);
4628 for_each_object(p
, s
, addr
, page
->objects
) {
4629 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), map
) ?
4630 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4632 if (!check_object(s
, page
, p
, val
))
4640 static int validate_slab_node(struct kmem_cache
*s
,
4641 struct kmem_cache_node
*n
)
4643 unsigned long count
= 0;
4645 unsigned long flags
;
4647 spin_lock_irqsave(&n
->list_lock
, flags
);
4649 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4650 validate_slab(s
, page
);
4653 if (count
!= n
->nr_partial
)
4654 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4655 s
->name
, count
, n
->nr_partial
);
4657 if (!(s
->flags
& SLAB_STORE_USER
))
4660 list_for_each_entry(page
, &n
->full
, slab_list
) {
4661 validate_slab(s
, page
);
4664 if (count
!= atomic_long_read(&n
->nr_slabs
))
4665 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4666 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4669 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4673 static long validate_slab_cache(struct kmem_cache
*s
)
4676 unsigned long count
= 0;
4677 struct kmem_cache_node
*n
;
4680 for_each_kmem_cache_node(s
, node
, n
)
4681 count
+= validate_slab_node(s
, n
);
4686 * Generate lists of code addresses where slabcache objects are allocated
4691 unsigned long count
;
4698 DECLARE_BITMAP(cpus
, NR_CPUS
);
4704 unsigned long count
;
4705 struct location
*loc
;
4708 static void free_loc_track(struct loc_track
*t
)
4711 free_pages((unsigned long)t
->loc
,
4712 get_order(sizeof(struct location
) * t
->max
));
4715 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4720 order
= get_order(sizeof(struct location
) * max
);
4722 l
= (void *)__get_free_pages(flags
, order
);
4727 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4735 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4736 const struct track
*track
)
4738 long start
, end
, pos
;
4740 unsigned long caddr
;
4741 unsigned long age
= jiffies
- track
->when
;
4747 pos
= start
+ (end
- start
+ 1) / 2;
4750 * There is nothing at "end". If we end up there
4751 * we need to add something to before end.
4756 caddr
= t
->loc
[pos
].addr
;
4757 if (track
->addr
== caddr
) {
4763 if (age
< l
->min_time
)
4765 if (age
> l
->max_time
)
4768 if (track
->pid
< l
->min_pid
)
4769 l
->min_pid
= track
->pid
;
4770 if (track
->pid
> l
->max_pid
)
4771 l
->max_pid
= track
->pid
;
4773 cpumask_set_cpu(track
->cpu
,
4774 to_cpumask(l
->cpus
));
4776 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4780 if (track
->addr
< caddr
)
4787 * Not found. Insert new tracking element.
4789 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4795 (t
->count
- pos
) * sizeof(struct location
));
4798 l
->addr
= track
->addr
;
4802 l
->min_pid
= track
->pid
;
4803 l
->max_pid
= track
->pid
;
4804 cpumask_clear(to_cpumask(l
->cpus
));
4805 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4806 nodes_clear(l
->nodes
);
4807 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4811 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4812 struct page
*page
, enum track_item alloc
)
4814 void *addr
= page_address(page
);
4818 map
= get_map(s
, page
);
4819 for_each_object(p
, s
, addr
, page
->objects
)
4820 if (!test_bit(__obj_to_index(s
, addr
, p
), map
))
4821 add_location(t
, s
, get_track(s
, p
, alloc
));
4825 static int list_locations(struct kmem_cache
*s
, char *buf
,
4826 enum track_item alloc
)
4830 struct loc_track t
= { 0, 0, NULL
};
4832 struct kmem_cache_node
*n
;
4834 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4836 return sysfs_emit(buf
, "Out of memory\n");
4838 /* Push back cpu slabs */
4841 for_each_kmem_cache_node(s
, node
, n
) {
4842 unsigned long flags
;
4845 if (!atomic_long_read(&n
->nr_slabs
))
4848 spin_lock_irqsave(&n
->list_lock
, flags
);
4849 list_for_each_entry(page
, &n
->partial
, slab_list
)
4850 process_slab(&t
, s
, page
, alloc
);
4851 list_for_each_entry(page
, &n
->full
, slab_list
)
4852 process_slab(&t
, s
, page
, alloc
);
4853 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4856 for (i
= 0; i
< t
.count
; i
++) {
4857 struct location
*l
= &t
.loc
[i
];
4859 len
+= sysfs_emit_at(buf
, len
, "%7ld ", l
->count
);
4862 len
+= sysfs_emit_at(buf
, len
, "%pS", (void *)l
->addr
);
4864 len
+= sysfs_emit_at(buf
, len
, "<not-available>");
4866 if (l
->sum_time
!= l
->min_time
)
4867 len
+= sysfs_emit_at(buf
, len
, " age=%ld/%ld/%ld",
4869 (long)div_u64(l
->sum_time
,
4873 len
+= sysfs_emit_at(buf
, len
, " age=%ld", l
->min_time
);
4875 if (l
->min_pid
!= l
->max_pid
)
4876 len
+= sysfs_emit_at(buf
, len
, " pid=%ld-%ld",
4877 l
->min_pid
, l
->max_pid
);
4879 len
+= sysfs_emit_at(buf
, len
, " pid=%ld",
4882 if (num_online_cpus() > 1 &&
4883 !cpumask_empty(to_cpumask(l
->cpus
)))
4884 len
+= sysfs_emit_at(buf
, len
, " cpus=%*pbl",
4885 cpumask_pr_args(to_cpumask(l
->cpus
)));
4887 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
4888 len
+= sysfs_emit_at(buf
, len
, " nodes=%*pbl",
4889 nodemask_pr_args(&l
->nodes
));
4891 len
+= sysfs_emit_at(buf
, len
, "\n");
4896 len
+= sysfs_emit_at(buf
, len
, "No data\n");
4900 #endif /* CONFIG_SLUB_DEBUG */
4902 #ifdef SLUB_RESILIENCY_TEST
4903 static void __init
resiliency_test(void)
4906 int type
= KMALLOC_NORMAL
;
4908 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4910 pr_err("SLUB resiliency testing\n");
4911 pr_err("-----------------------\n");
4912 pr_err("A. Corruption after allocation\n");
4914 p
= kzalloc(16, GFP_KERNEL
);
4916 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4919 validate_slab_cache(kmalloc_caches
[type
][4]);
4921 /* Hmmm... The next two are dangerous */
4922 p
= kzalloc(32, GFP_KERNEL
);
4923 p
[32 + sizeof(void *)] = 0x34;
4924 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4926 pr_err("If allocated object is overwritten then not detectable\n\n");
4928 validate_slab_cache(kmalloc_caches
[type
][5]);
4929 p
= kzalloc(64, GFP_KERNEL
);
4930 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4932 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4934 pr_err("If allocated object is overwritten then not detectable\n\n");
4935 validate_slab_cache(kmalloc_caches
[type
][6]);
4937 pr_err("\nB. Corruption after free\n");
4938 p
= kzalloc(128, GFP_KERNEL
);
4941 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4942 validate_slab_cache(kmalloc_caches
[type
][7]);
4944 p
= kzalloc(256, GFP_KERNEL
);
4947 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4948 validate_slab_cache(kmalloc_caches
[type
][8]);
4950 p
= kzalloc(512, GFP_KERNEL
);
4953 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4954 validate_slab_cache(kmalloc_caches
[type
][9]);
4958 static void resiliency_test(void) {};
4960 #endif /* SLUB_RESILIENCY_TEST */
4963 enum slab_stat_type
{
4964 SL_ALL
, /* All slabs */
4965 SL_PARTIAL
, /* Only partially allocated slabs */
4966 SL_CPU
, /* Only slabs used for cpu caches */
4967 SL_OBJECTS
, /* Determine allocated objects not slabs */
4968 SL_TOTAL
/* Determine object capacity not slabs */
4971 #define SO_ALL (1 << SL_ALL)
4972 #define SO_PARTIAL (1 << SL_PARTIAL)
4973 #define SO_CPU (1 << SL_CPU)
4974 #define SO_OBJECTS (1 << SL_OBJECTS)
4975 #define SO_TOTAL (1 << SL_TOTAL)
4977 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4978 char *buf
, unsigned long flags
)
4980 unsigned long total
= 0;
4983 unsigned long *nodes
;
4986 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4990 if (flags
& SO_CPU
) {
4993 for_each_possible_cpu(cpu
) {
4994 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
4999 page
= READ_ONCE(c
->page
);
5003 node
= page_to_nid(page
);
5004 if (flags
& SO_TOTAL
)
5006 else if (flags
& SO_OBJECTS
)
5014 page
= slub_percpu_partial_read_once(c
);
5016 node
= page_to_nid(page
);
5017 if (flags
& SO_TOTAL
)
5019 else if (flags
& SO_OBJECTS
)
5030 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5031 * already held which will conflict with an existing lock order:
5033 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5035 * We don't really need mem_hotplug_lock (to hold off
5036 * slab_mem_going_offline_callback) here because slab's memory hot
5037 * unplug code doesn't destroy the kmem_cache->node[] data.
5040 #ifdef CONFIG_SLUB_DEBUG
5041 if (flags
& SO_ALL
) {
5042 struct kmem_cache_node
*n
;
5044 for_each_kmem_cache_node(s
, node
, n
) {
5046 if (flags
& SO_TOTAL
)
5047 x
= atomic_long_read(&n
->total_objects
);
5048 else if (flags
& SO_OBJECTS
)
5049 x
= atomic_long_read(&n
->total_objects
) -
5050 count_partial(n
, count_free
);
5052 x
= atomic_long_read(&n
->nr_slabs
);
5059 if (flags
& SO_PARTIAL
) {
5060 struct kmem_cache_node
*n
;
5062 for_each_kmem_cache_node(s
, node
, n
) {
5063 if (flags
& SO_TOTAL
)
5064 x
= count_partial(n
, count_total
);
5065 else if (flags
& SO_OBJECTS
)
5066 x
= count_partial(n
, count_inuse
);
5074 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5076 for (node
= 0; node
< nr_node_ids
; node
++) {
5078 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5082 len
+= sysfs_emit_at(buf
, len
, "\n");
5088 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5089 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5091 struct slab_attribute
{
5092 struct attribute attr
;
5093 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5094 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5097 #define SLAB_ATTR_RO(_name) \
5098 static struct slab_attribute _name##_attr = \
5099 __ATTR(_name, 0400, _name##_show, NULL)
5101 #define SLAB_ATTR(_name) \
5102 static struct slab_attribute _name##_attr = \
5103 __ATTR(_name, 0600, _name##_show, _name##_store)
5105 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5107 return sysfs_emit(buf
, "%u\n", s
->size
);
5109 SLAB_ATTR_RO(slab_size
);
5111 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5113 return sysfs_emit(buf
, "%u\n", s
->align
);
5115 SLAB_ATTR_RO(align
);
5117 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5119 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5121 SLAB_ATTR_RO(object_size
);
5123 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5125 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5127 SLAB_ATTR_RO(objs_per_slab
);
5129 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5131 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5133 SLAB_ATTR_RO(order
);
5135 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5137 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5140 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5146 err
= kstrtoul(buf
, 10, &min
);
5150 set_min_partial(s
, min
);
5153 SLAB_ATTR(min_partial
);
5155 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5157 return sysfs_emit(buf
, "%u\n", slub_cpu_partial(s
));
5160 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5163 unsigned int objects
;
5166 err
= kstrtouint(buf
, 10, &objects
);
5169 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5172 slub_set_cpu_partial(s
, objects
);
5176 SLAB_ATTR(cpu_partial
);
5178 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5182 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5186 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5188 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5190 SLAB_ATTR_RO(aliases
);
5192 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5194 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5196 SLAB_ATTR_RO(partial
);
5198 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5200 return show_slab_objects(s
, buf
, SO_CPU
);
5202 SLAB_ATTR_RO(cpu_slabs
);
5204 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5206 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5208 SLAB_ATTR_RO(objects
);
5210 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5212 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5214 SLAB_ATTR_RO(objects_partial
);
5216 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5223 for_each_online_cpu(cpu
) {
5226 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5229 pages
+= page
->pages
;
5230 objects
+= page
->pobjects
;
5234 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, pages
);
5237 for_each_online_cpu(cpu
) {
5240 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5242 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5243 cpu
, page
->pobjects
, page
->pages
);
5246 len
+= sysfs_emit_at(buf
, len
, "\n");
5250 SLAB_ATTR_RO(slabs_cpu_partial
);
5252 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5254 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5256 SLAB_ATTR_RO(reclaim_account
);
5258 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5260 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5262 SLAB_ATTR_RO(hwcache_align
);
5264 #ifdef CONFIG_ZONE_DMA
5265 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5267 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5269 SLAB_ATTR_RO(cache_dma
);
5272 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5274 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5276 SLAB_ATTR_RO(usersize
);
5278 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5280 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5282 SLAB_ATTR_RO(destroy_by_rcu
);
5284 #ifdef CONFIG_SLUB_DEBUG
5285 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5287 return show_slab_objects(s
, buf
, SO_ALL
);
5289 SLAB_ATTR_RO(slabs
);
5291 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5293 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5295 SLAB_ATTR_RO(total_objects
);
5297 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5299 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5301 SLAB_ATTR_RO(sanity_checks
);
5303 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5305 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5307 SLAB_ATTR_RO(trace
);
5309 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5311 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5314 SLAB_ATTR_RO(red_zone
);
5316 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5318 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5321 SLAB_ATTR_RO(poison
);
5323 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5325 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5328 SLAB_ATTR_RO(store_user
);
5330 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5335 static ssize_t
validate_store(struct kmem_cache
*s
,
5336 const char *buf
, size_t length
)
5340 if (buf
[0] == '1') {
5341 ret
= validate_slab_cache(s
);
5347 SLAB_ATTR(validate
);
5349 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5351 if (!(s
->flags
& SLAB_STORE_USER
))
5353 return list_locations(s
, buf
, TRACK_ALLOC
);
5355 SLAB_ATTR_RO(alloc_calls
);
5357 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5359 if (!(s
->flags
& SLAB_STORE_USER
))
5361 return list_locations(s
, buf
, TRACK_FREE
);
5363 SLAB_ATTR_RO(free_calls
);
5364 #endif /* CONFIG_SLUB_DEBUG */
5366 #ifdef CONFIG_FAILSLAB
5367 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5369 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5371 SLAB_ATTR_RO(failslab
);
5374 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5379 static ssize_t
shrink_store(struct kmem_cache
*s
,
5380 const char *buf
, size_t length
)
5383 kmem_cache_shrink(s
);
5391 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5393 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5396 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5397 const char *buf
, size_t length
)
5402 err
= kstrtouint(buf
, 10, &ratio
);
5408 s
->remote_node_defrag_ratio
= ratio
* 10;
5412 SLAB_ATTR(remote_node_defrag_ratio
);
5415 #ifdef CONFIG_SLUB_STATS
5416 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5418 unsigned long sum
= 0;
5421 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5426 for_each_online_cpu(cpu
) {
5427 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5433 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5436 for_each_online_cpu(cpu
) {
5438 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5443 len
+= sysfs_emit_at(buf
, len
, "\n");
5448 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5452 for_each_online_cpu(cpu
)
5453 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5456 #define STAT_ATTR(si, text) \
5457 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5459 return show_stat(s, buf, si); \
5461 static ssize_t text##_store(struct kmem_cache *s, \
5462 const char *buf, size_t length) \
5464 if (buf[0] != '0') \
5466 clear_stat(s, si); \
5471 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5472 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5473 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5474 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5475 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5476 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5477 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5478 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5479 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5480 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5481 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5482 STAT_ATTR(FREE_SLAB
, free_slab
);
5483 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5484 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5485 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5486 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5487 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5488 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5489 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5490 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5491 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5492 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5493 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5494 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5495 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5496 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5497 #endif /* CONFIG_SLUB_STATS */
5499 static struct attribute
*slab_attrs
[] = {
5500 &slab_size_attr
.attr
,
5501 &object_size_attr
.attr
,
5502 &objs_per_slab_attr
.attr
,
5504 &min_partial_attr
.attr
,
5505 &cpu_partial_attr
.attr
,
5507 &objects_partial_attr
.attr
,
5509 &cpu_slabs_attr
.attr
,
5513 &hwcache_align_attr
.attr
,
5514 &reclaim_account_attr
.attr
,
5515 &destroy_by_rcu_attr
.attr
,
5517 &slabs_cpu_partial_attr
.attr
,
5518 #ifdef CONFIG_SLUB_DEBUG
5519 &total_objects_attr
.attr
,
5521 &sanity_checks_attr
.attr
,
5523 &red_zone_attr
.attr
,
5525 &store_user_attr
.attr
,
5526 &validate_attr
.attr
,
5527 &alloc_calls_attr
.attr
,
5528 &free_calls_attr
.attr
,
5530 #ifdef CONFIG_ZONE_DMA
5531 &cache_dma_attr
.attr
,
5534 &remote_node_defrag_ratio_attr
.attr
,
5536 #ifdef CONFIG_SLUB_STATS
5537 &alloc_fastpath_attr
.attr
,
5538 &alloc_slowpath_attr
.attr
,
5539 &free_fastpath_attr
.attr
,
5540 &free_slowpath_attr
.attr
,
5541 &free_frozen_attr
.attr
,
5542 &free_add_partial_attr
.attr
,
5543 &free_remove_partial_attr
.attr
,
5544 &alloc_from_partial_attr
.attr
,
5545 &alloc_slab_attr
.attr
,
5546 &alloc_refill_attr
.attr
,
5547 &alloc_node_mismatch_attr
.attr
,
5548 &free_slab_attr
.attr
,
5549 &cpuslab_flush_attr
.attr
,
5550 &deactivate_full_attr
.attr
,
5551 &deactivate_empty_attr
.attr
,
5552 &deactivate_to_head_attr
.attr
,
5553 &deactivate_to_tail_attr
.attr
,
5554 &deactivate_remote_frees_attr
.attr
,
5555 &deactivate_bypass_attr
.attr
,
5556 &order_fallback_attr
.attr
,
5557 &cmpxchg_double_fail_attr
.attr
,
5558 &cmpxchg_double_cpu_fail_attr
.attr
,
5559 &cpu_partial_alloc_attr
.attr
,
5560 &cpu_partial_free_attr
.attr
,
5561 &cpu_partial_node_attr
.attr
,
5562 &cpu_partial_drain_attr
.attr
,
5564 #ifdef CONFIG_FAILSLAB
5565 &failslab_attr
.attr
,
5567 &usersize_attr
.attr
,
5572 static const struct attribute_group slab_attr_group
= {
5573 .attrs
= slab_attrs
,
5576 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5577 struct attribute
*attr
,
5580 struct slab_attribute
*attribute
;
5581 struct kmem_cache
*s
;
5584 attribute
= to_slab_attr(attr
);
5587 if (!attribute
->show
)
5590 err
= attribute
->show(s
, buf
);
5595 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5596 struct attribute
*attr
,
5597 const char *buf
, size_t len
)
5599 struct slab_attribute
*attribute
;
5600 struct kmem_cache
*s
;
5603 attribute
= to_slab_attr(attr
);
5606 if (!attribute
->store
)
5609 err
= attribute
->store(s
, buf
, len
);
5613 static void kmem_cache_release(struct kobject
*k
)
5615 slab_kmem_cache_release(to_slab(k
));
5618 static const struct sysfs_ops slab_sysfs_ops
= {
5619 .show
= slab_attr_show
,
5620 .store
= slab_attr_store
,
5623 static struct kobj_type slab_ktype
= {
5624 .sysfs_ops
= &slab_sysfs_ops
,
5625 .release
= kmem_cache_release
,
5628 static struct kset
*slab_kset
;
5630 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5635 #define ID_STR_LENGTH 64
5637 /* Create a unique string id for a slab cache:
5639 * Format :[flags-]size
5641 static char *create_unique_id(struct kmem_cache
*s
)
5643 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5650 * First flags affecting slabcache operations. We will only
5651 * get here for aliasable slabs so we do not need to support
5652 * too many flags. The flags here must cover all flags that
5653 * are matched during merging to guarantee that the id is
5656 if (s
->flags
& SLAB_CACHE_DMA
)
5658 if (s
->flags
& SLAB_CACHE_DMA32
)
5660 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5662 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5664 if (s
->flags
& SLAB_ACCOUNT
)
5668 p
+= sprintf(p
, "%07u", s
->size
);
5670 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5674 static int sysfs_slab_add(struct kmem_cache
*s
)
5678 struct kset
*kset
= cache_kset(s
);
5679 int unmergeable
= slab_unmergeable(s
);
5682 kobject_init(&s
->kobj
, &slab_ktype
);
5686 if (!unmergeable
&& disable_higher_order_debug
&&
5687 (slub_debug
& DEBUG_METADATA_FLAGS
))
5692 * Slabcache can never be merged so we can use the name proper.
5693 * This is typically the case for debug situations. In that
5694 * case we can catch duplicate names easily.
5696 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5700 * Create a unique name for the slab as a target
5703 name
= create_unique_id(s
);
5706 s
->kobj
.kset
= kset
;
5707 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5711 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5716 /* Setup first alias */
5717 sysfs_slab_alias(s
, s
->name
);
5724 kobject_del(&s
->kobj
);
5728 void sysfs_slab_unlink(struct kmem_cache
*s
)
5730 if (slab_state
>= FULL
)
5731 kobject_del(&s
->kobj
);
5734 void sysfs_slab_release(struct kmem_cache
*s
)
5736 if (slab_state
>= FULL
)
5737 kobject_put(&s
->kobj
);
5741 * Need to buffer aliases during bootup until sysfs becomes
5742 * available lest we lose that information.
5744 struct saved_alias
{
5745 struct kmem_cache
*s
;
5747 struct saved_alias
*next
;
5750 static struct saved_alias
*alias_list
;
5752 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5754 struct saved_alias
*al
;
5756 if (slab_state
== FULL
) {
5758 * If we have a leftover link then remove it.
5760 sysfs_remove_link(&slab_kset
->kobj
, name
);
5761 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5764 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5770 al
->next
= alias_list
;
5775 static int __init
slab_sysfs_init(void)
5777 struct kmem_cache
*s
;
5780 mutex_lock(&slab_mutex
);
5782 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
5784 mutex_unlock(&slab_mutex
);
5785 pr_err("Cannot register slab subsystem.\n");
5791 list_for_each_entry(s
, &slab_caches
, list
) {
5792 err
= sysfs_slab_add(s
);
5794 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5798 while (alias_list
) {
5799 struct saved_alias
*al
= alias_list
;
5801 alias_list
= alias_list
->next
;
5802 err
= sysfs_slab_alias(al
->s
, al
->name
);
5804 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5809 mutex_unlock(&slab_mutex
);
5814 __initcall(slab_sysfs_init
);
5815 #endif /* CONFIG_SYSFS */
5818 * The /proc/slabinfo ABI
5820 #ifdef CONFIG_SLUB_DEBUG
5821 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5823 unsigned long nr_slabs
= 0;
5824 unsigned long nr_objs
= 0;
5825 unsigned long nr_free
= 0;
5827 struct kmem_cache_node
*n
;
5829 for_each_kmem_cache_node(s
, node
, n
) {
5830 nr_slabs
+= node_nr_slabs(n
);
5831 nr_objs
+= node_nr_objs(n
);
5832 nr_free
+= count_partial(n
, count_free
);
5835 sinfo
->active_objs
= nr_objs
- nr_free
;
5836 sinfo
->num_objs
= nr_objs
;
5837 sinfo
->active_slabs
= nr_slabs
;
5838 sinfo
->num_slabs
= nr_slabs
;
5839 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5840 sinfo
->cache_order
= oo_order(s
->oo
);
5843 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5847 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5848 size_t count
, loff_t
*ppos
)
5852 #endif /* CONFIG_SLUB_DEBUG */