1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/kfence.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list except per cpu partial list. The processor that froze the
63 * slab is the one who can perform list operations on the page. Other
64 * processors may put objects onto the freelist but the processor that
65 * froze the slab is the only one that can retrieve the objects from the
68 * The list_lock protects the partial and full list on each node and
69 * the partial slab counter. If taken then no new slabs may be added or
70 * removed from the lists nor make the number of partial slabs be modified.
71 * (Note that the total number of slabs is an atomic value that may be
72 * modified without taking the list lock).
74 * The list_lock is a centralized lock and thus we avoid taking it as
75 * much as possible. As long as SLUB does not have to handle partial
76 * slabs, operations can continue without any centralized lock. F.e.
77 * allocating a long series of objects that fill up slabs does not require
79 * Interrupts are disabled during allocation and deallocation in order to
80 * make the slab allocator safe to use in the context of an irq. In addition
81 * interrupts are disabled to ensure that the processor does not change
82 * while handling per_cpu slabs, due to kernel preemption.
84 * SLUB assigns one slab for allocation to each processor.
85 * Allocations only occur from these slabs called cpu slabs.
87 * Slabs with free elements are kept on a partial list and during regular
88 * operations no list for full slabs is used. If an object in a full slab is
89 * freed then the slab will show up again on the partial lists.
90 * We track full slabs for debugging purposes though because otherwise we
91 * cannot scan all objects.
93 * Slabs are freed when they become empty. Teardown and setup is
94 * minimal so we rely on the page allocators per cpu caches for
95 * fast frees and allocs.
97 * page->frozen The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 #ifdef CONFIG_SLUB_DEBUG
119 #ifdef CONFIG_SLUB_DEBUG_ON
120 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled
);
122 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled
);
126 static inline bool kmem_cache_debug(struct kmem_cache
*s
)
128 return kmem_cache_debug_flags(s
, SLAB_DEBUG_FLAGS
);
131 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
133 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
))
134 p
+= s
->red_left_pad
;
139 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
141 #ifdef CONFIG_SLUB_CPU_PARTIAL
142 return !kmem_cache_debug(s
);
149 * Issues still to be resolved:
151 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
153 * - Variable sizing of the per node arrays
156 /* Enable to test recovery from slab corruption on boot */
157 #undef SLUB_RESILIENCY_TEST
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Minimum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * These debug flags cannot use CMPXCHG because there might be consistency
180 * issues when checking or reading debug information
182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
187 * Debugging flags that require metadata to be stored in the slab. These get
188 * disabled when slub_debug=O is used and a cache's min order increases with
191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
194 #define OO_MASK ((1 << OO_SHIFT) - 1)
195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
197 /* Internal SLUB flags */
199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
200 /* Use cmpxchg_double */
201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
204 * Tracking user of a slab.
206 #define TRACK_ADDRS_COUNT 16
208 unsigned long addr
; /* Called from address */
209 #ifdef CONFIG_STACKTRACE
210 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
212 int cpu
; /* Was running on cpu */
213 int pid
; /* Pid context */
214 unsigned long when
; /* When did the operation occur */
217 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
220 static int sysfs_slab_add(struct kmem_cache
*);
221 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
223 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
224 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
228 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
240 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
241 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
242 * differ during memory hotplug/hotremove operations.
243 * Protected by slab_mutex.
245 static nodemask_t slab_nodes
;
247 /********************************************************************
248 * Core slab cache functions
249 *******************************************************************/
252 * Returns freelist pointer (ptr). With hardening, this is obfuscated
253 * with an XOR of the address where the pointer is held and a per-cache
256 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
257 unsigned long ptr_addr
)
259 #ifdef CONFIG_SLAB_FREELIST_HARDENED
261 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
262 * Normally, this doesn't cause any issues, as both set_freepointer()
263 * and get_freepointer() are called with a pointer with the same tag.
264 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
265 * example, when __free_slub() iterates over objects in a cache, it
266 * passes untagged pointers to check_object(). check_object() in turns
267 * calls get_freepointer() with an untagged pointer, which causes the
268 * freepointer to be restored incorrectly.
270 return (void *)((unsigned long)ptr
^ s
->random
^
271 swab((unsigned long)kasan_reset_tag((void *)ptr_addr
)));
277 /* Returns the freelist pointer recorded at location ptr_addr. */
278 static inline void *freelist_dereference(const struct kmem_cache
*s
,
281 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
282 (unsigned long)ptr_addr
);
285 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
287 object
= kasan_reset_tag(object
);
288 return freelist_dereference(s
, object
+ s
->offset
);
291 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
293 prefetch(object
+ s
->offset
);
296 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
298 unsigned long freepointer_addr
;
301 if (!debug_pagealloc_enabled_static())
302 return get_freepointer(s
, object
);
304 object
= kasan_reset_tag(object
);
305 freepointer_addr
= (unsigned long)object
+ s
->offset
;
306 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
307 return freelist_ptr(s
, p
, freepointer_addr
);
310 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
312 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
314 #ifdef CONFIG_SLAB_FREELIST_HARDENED
315 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
318 freeptr_addr
= (unsigned long)kasan_reset_tag((void *)freeptr_addr
);
319 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
322 /* Loop over all objects in a slab */
323 #define for_each_object(__p, __s, __addr, __objects) \
324 for (__p = fixup_red_left(__s, __addr); \
325 __p < (__addr) + (__objects) * (__s)->size; \
328 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
330 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
333 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
336 struct kmem_cache_order_objects x
= {
337 (order
<< OO_SHIFT
) + order_objects(order
, size
)
343 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
345 return x
.x
>> OO_SHIFT
;
348 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
350 return x
.x
& OO_MASK
;
354 * Per slab locking using the pagelock
356 static __always_inline
void slab_lock(struct page
*page
)
358 VM_BUG_ON_PAGE(PageTail(page
), page
);
359 bit_spin_lock(PG_locked
, &page
->flags
);
362 static __always_inline
void slab_unlock(struct page
*page
)
364 VM_BUG_ON_PAGE(PageTail(page
), page
);
365 __bit_spin_unlock(PG_locked
, &page
->flags
);
368 /* Interrupts must be disabled (for the fallback code to work right) */
369 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
370 void *freelist_old
, unsigned long counters_old
,
371 void *freelist_new
, unsigned long counters_new
,
374 VM_BUG_ON(!irqs_disabled());
375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s
->flags
& __CMPXCHG_DOUBLE
) {
378 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
379 freelist_old
, counters_old
,
380 freelist_new
, counters_new
))
386 if (page
->freelist
== freelist_old
&&
387 page
->counters
== counters_old
) {
388 page
->freelist
= freelist_new
;
389 page
->counters
= counters_new
;
397 stat(s
, CMPXCHG_DOUBLE_FAIL
);
399 #ifdef SLUB_DEBUG_CMPXCHG
400 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
406 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
407 void *freelist_old
, unsigned long counters_old
,
408 void *freelist_new
, unsigned long counters_new
,
411 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
412 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
413 if (s
->flags
& __CMPXCHG_DOUBLE
) {
414 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
415 freelist_old
, counters_old
,
416 freelist_new
, counters_new
))
423 local_irq_save(flags
);
425 if (page
->freelist
== freelist_old
&&
426 page
->counters
== counters_old
) {
427 page
->freelist
= freelist_new
;
428 page
->counters
= counters_new
;
430 local_irq_restore(flags
);
434 local_irq_restore(flags
);
438 stat(s
, CMPXCHG_DOUBLE_FAIL
);
440 #ifdef SLUB_DEBUG_CMPXCHG
441 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
447 #ifdef CONFIG_SLUB_DEBUG
448 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
449 static DEFINE_SPINLOCK(object_map_lock
);
452 * Determine a map of object in use on a page.
454 * Node listlock must be held to guarantee that the page does
455 * not vanish from under us.
457 static unsigned long *get_map(struct kmem_cache
*s
, struct page
*page
)
458 __acquires(&object_map_lock
)
461 void *addr
= page_address(page
);
463 VM_BUG_ON(!irqs_disabled());
465 spin_lock(&object_map_lock
);
467 bitmap_zero(object_map
, page
->objects
);
469 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
470 set_bit(__obj_to_index(s
, addr
, p
), object_map
);
475 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
477 VM_BUG_ON(map
!= object_map
);
478 spin_unlock(&object_map_lock
);
481 static inline unsigned int size_from_object(struct kmem_cache
*s
)
483 if (s
->flags
& SLAB_RED_ZONE
)
484 return s
->size
- s
->red_left_pad
;
489 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
491 if (s
->flags
& SLAB_RED_ZONE
)
492 p
-= s
->red_left_pad
;
500 #if defined(CONFIG_SLUB_DEBUG_ON)
501 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
503 static slab_flags_t slub_debug
;
506 static char *slub_debug_string
;
507 static int disable_higher_order_debug
;
510 * slub is about to manipulate internal object metadata. This memory lies
511 * outside the range of the allocated object, so accessing it would normally
512 * be reported by kasan as a bounds error. metadata_access_enable() is used
513 * to tell kasan that these accesses are OK.
515 static inline void metadata_access_enable(void)
517 kasan_disable_current();
520 static inline void metadata_access_disable(void)
522 kasan_enable_current();
529 /* Verify that a pointer has an address that is valid within a slab page */
530 static inline int check_valid_pointer(struct kmem_cache
*s
,
531 struct page
*page
, void *object
)
538 base
= page_address(page
);
539 object
= kasan_reset_tag(object
);
540 object
= restore_red_left(s
, object
);
541 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
542 (object
- base
) % s
->size
) {
549 static void print_section(char *level
, char *text
, u8
*addr
,
552 metadata_access_enable();
553 print_hex_dump(level
, kasan_reset_tag(text
), DUMP_PREFIX_ADDRESS
,
554 16, 1, addr
, length
, 1);
555 metadata_access_disable();
559 * See comment in calculate_sizes().
561 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
563 return s
->offset
>= s
->inuse
;
567 * Return offset of the end of info block which is inuse + free pointer if
568 * not overlapping with object.
570 static inline unsigned int get_info_end(struct kmem_cache
*s
)
572 if (freeptr_outside_object(s
))
573 return s
->inuse
+ sizeof(void *);
578 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
579 enum track_item alloc
)
583 p
= object
+ get_info_end(s
);
585 return kasan_reset_tag(p
+ alloc
);
588 static void set_track(struct kmem_cache
*s
, void *object
,
589 enum track_item alloc
, unsigned long addr
)
591 struct track
*p
= get_track(s
, object
, alloc
);
594 #ifdef CONFIG_STACKTRACE
595 unsigned int nr_entries
;
597 metadata_access_enable();
598 nr_entries
= stack_trace_save(kasan_reset_tag(p
->addrs
),
599 TRACK_ADDRS_COUNT
, 3);
600 metadata_access_disable();
602 if (nr_entries
< TRACK_ADDRS_COUNT
)
603 p
->addrs
[nr_entries
] = 0;
606 p
->cpu
= smp_processor_id();
607 p
->pid
= current
->pid
;
610 memset(p
, 0, sizeof(struct track
));
614 static void init_tracking(struct kmem_cache
*s
, void *object
)
616 if (!(s
->flags
& SLAB_STORE_USER
))
619 set_track(s
, object
, TRACK_FREE
, 0UL);
620 set_track(s
, object
, TRACK_ALLOC
, 0UL);
623 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
628 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
629 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
630 #ifdef CONFIG_STACKTRACE
633 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
635 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
642 void print_tracking(struct kmem_cache
*s
, void *object
)
644 unsigned long pr_time
= jiffies
;
645 if (!(s
->flags
& SLAB_STORE_USER
))
648 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
649 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
652 static void print_page_info(struct page
*page
)
654 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
655 page
, page
->objects
, page
->inuse
, page
->freelist
,
656 page
->flags
, &page
->flags
);
660 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
662 struct va_format vaf
;
668 pr_err("=============================================================================\n");
669 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
670 pr_err("-----------------------------------------------------------------------------\n\n");
672 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
676 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
678 struct va_format vaf
;
684 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
688 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
689 void **freelist
, void *nextfree
)
691 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
692 !check_valid_pointer(s
, page
, nextfree
) && freelist
) {
693 object_err(s
, page
, *freelist
, "Freechain corrupt");
695 slab_fix(s
, "Isolate corrupted freechain");
702 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
704 unsigned int off
; /* Offset of last byte */
705 u8
*addr
= page_address(page
);
707 print_tracking(s
, p
);
709 print_page_info(page
);
711 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
712 p
, p
- addr
, get_freepointer(s
, p
));
714 if (s
->flags
& SLAB_RED_ZONE
)
715 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
717 else if (p
> addr
+ 16)
718 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
720 print_section(KERN_ERR
, "Object ", p
,
721 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
722 if (s
->flags
& SLAB_RED_ZONE
)
723 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
724 s
->inuse
- s
->object_size
);
726 off
= get_info_end(s
);
728 if (s
->flags
& SLAB_STORE_USER
)
729 off
+= 2 * sizeof(struct track
);
731 off
+= kasan_metadata_size(s
);
733 if (off
!= size_from_object(s
))
734 /* Beginning of the filler is the free pointer */
735 print_section(KERN_ERR
, "Padding ", p
+ off
,
736 size_from_object(s
) - off
);
741 void object_err(struct kmem_cache
*s
, struct page
*page
,
742 u8
*object
, char *reason
)
744 slab_bug(s
, "%s", reason
);
745 print_trailer(s
, page
, object
);
748 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
749 const char *fmt
, ...)
755 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
757 slab_bug(s
, "%s", buf
);
758 print_page_info(page
);
762 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
764 u8
*p
= kasan_reset_tag(object
);
766 if (s
->flags
& SLAB_RED_ZONE
)
767 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
769 if (s
->flags
& __OBJECT_POISON
) {
770 memset(p
, POISON_FREE
, s
->object_size
- 1);
771 p
[s
->object_size
- 1] = POISON_END
;
774 if (s
->flags
& SLAB_RED_ZONE
)
775 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
778 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
779 void *from
, void *to
)
781 slab_fix(s
, "Restoring 0x%p-0x%p=0x%x\n", from
, to
- 1, data
);
782 memset(from
, data
, to
- from
);
785 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
786 u8
*object
, char *what
,
787 u8
*start
, unsigned int value
, unsigned int bytes
)
791 u8
*addr
= page_address(page
);
793 metadata_access_enable();
794 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
795 metadata_access_disable();
800 while (end
> fault
&& end
[-1] == value
)
803 slab_bug(s
, "%s overwritten", what
);
804 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
805 fault
, end
- 1, fault
- addr
,
807 print_trailer(s
, page
, object
);
809 restore_bytes(s
, what
, value
, fault
, end
);
817 * Bytes of the object to be managed.
818 * If the freepointer may overlay the object then the free
819 * pointer is at the middle of the object.
821 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
824 * object + s->object_size
825 * Padding to reach word boundary. This is also used for Redzoning.
826 * Padding is extended by another word if Redzoning is enabled and
827 * object_size == inuse.
829 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
830 * 0xcc (RED_ACTIVE) for objects in use.
833 * Meta data starts here.
835 * A. Free pointer (if we cannot overwrite object on free)
836 * B. Tracking data for SLAB_STORE_USER
837 * C. Padding to reach required alignment boundary or at minimum
838 * one word if debugging is on to be able to detect writes
839 * before the word boundary.
841 * Padding is done using 0x5a (POISON_INUSE)
844 * Nothing is used beyond s->size.
846 * If slabcaches are merged then the object_size and inuse boundaries are mostly
847 * ignored. And therefore no slab options that rely on these boundaries
848 * may be used with merged slabcaches.
851 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
853 unsigned long off
= get_info_end(s
); /* The end of info */
855 if (s
->flags
& SLAB_STORE_USER
)
856 /* We also have user information there */
857 off
+= 2 * sizeof(struct track
);
859 off
+= kasan_metadata_size(s
);
861 if (size_from_object(s
) == off
)
864 return check_bytes_and_report(s
, page
, p
, "Object padding",
865 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
868 /* Check the pad bytes at the end of a slab page */
869 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
878 if (!(s
->flags
& SLAB_POISON
))
881 start
= page_address(page
);
882 length
= page_size(page
);
883 end
= start
+ length
;
884 remainder
= length
% s
->size
;
888 pad
= end
- remainder
;
889 metadata_access_enable();
890 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
891 metadata_access_disable();
894 while (end
> fault
&& end
[-1] == POISON_INUSE
)
897 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
898 fault
, end
- 1, fault
- start
);
899 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
901 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
905 static int check_object(struct kmem_cache
*s
, struct page
*page
,
906 void *object
, u8 val
)
909 u8
*endobject
= object
+ s
->object_size
;
911 if (s
->flags
& SLAB_RED_ZONE
) {
912 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
913 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
916 if (!check_bytes_and_report(s
, page
, object
, "Redzone",
917 endobject
, val
, s
->inuse
- s
->object_size
))
920 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
921 check_bytes_and_report(s
, page
, p
, "Alignment padding",
922 endobject
, POISON_INUSE
,
923 s
->inuse
- s
->object_size
);
927 if (s
->flags
& SLAB_POISON
) {
928 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
929 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
930 POISON_FREE
, s
->object_size
- 1) ||
931 !check_bytes_and_report(s
, page
, p
, "Poison",
932 p
+ s
->object_size
- 1, POISON_END
, 1)))
935 * check_pad_bytes cleans up on its own.
937 check_pad_bytes(s
, page
, p
);
940 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
942 * Object and freepointer overlap. Cannot check
943 * freepointer while object is allocated.
947 /* Check free pointer validity */
948 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
949 object_err(s
, page
, p
, "Freepointer corrupt");
951 * No choice but to zap it and thus lose the remainder
952 * of the free objects in this slab. May cause
953 * another error because the object count is now wrong.
955 set_freepointer(s
, p
, NULL
);
961 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
965 VM_BUG_ON(!irqs_disabled());
967 if (!PageSlab(page
)) {
968 slab_err(s
, page
, "Not a valid slab page");
972 maxobj
= order_objects(compound_order(page
), s
->size
);
973 if (page
->objects
> maxobj
) {
974 slab_err(s
, page
, "objects %u > max %u",
975 page
->objects
, maxobj
);
978 if (page
->inuse
> page
->objects
) {
979 slab_err(s
, page
, "inuse %u > max %u",
980 page
->inuse
, page
->objects
);
983 /* Slab_pad_check fixes things up after itself */
984 slab_pad_check(s
, page
);
989 * Determine if a certain object on a page is on the freelist. Must hold the
990 * slab lock to guarantee that the chains are in a consistent state.
992 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
1000 while (fp
&& nr
<= page
->objects
) {
1003 if (!check_valid_pointer(s
, page
, fp
)) {
1005 object_err(s
, page
, object
,
1006 "Freechain corrupt");
1007 set_freepointer(s
, object
, NULL
);
1009 slab_err(s
, page
, "Freepointer corrupt");
1010 page
->freelist
= NULL
;
1011 page
->inuse
= page
->objects
;
1012 slab_fix(s
, "Freelist cleared");
1018 fp
= get_freepointer(s
, object
);
1022 max_objects
= order_objects(compound_order(page
), s
->size
);
1023 if (max_objects
> MAX_OBJS_PER_PAGE
)
1024 max_objects
= MAX_OBJS_PER_PAGE
;
1026 if (page
->objects
!= max_objects
) {
1027 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1028 page
->objects
, max_objects
);
1029 page
->objects
= max_objects
;
1030 slab_fix(s
, "Number of objects adjusted.");
1032 if (page
->inuse
!= page
->objects
- nr
) {
1033 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1034 page
->inuse
, page
->objects
- nr
);
1035 page
->inuse
= page
->objects
- nr
;
1036 slab_fix(s
, "Object count adjusted.");
1038 return search
== NULL
;
1041 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1044 if (s
->flags
& SLAB_TRACE
) {
1045 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1047 alloc
? "alloc" : "free",
1048 object
, page
->inuse
,
1052 print_section(KERN_INFO
, "Object ", (void *)object
,
1060 * Tracking of fully allocated slabs for debugging purposes.
1062 static void add_full(struct kmem_cache
*s
,
1063 struct kmem_cache_node
*n
, struct page
*page
)
1065 if (!(s
->flags
& SLAB_STORE_USER
))
1068 lockdep_assert_held(&n
->list_lock
);
1069 list_add(&page
->slab_list
, &n
->full
);
1072 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1074 if (!(s
->flags
& SLAB_STORE_USER
))
1077 lockdep_assert_held(&n
->list_lock
);
1078 list_del(&page
->slab_list
);
1081 /* Tracking of the number of slabs for debugging purposes */
1082 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1084 struct kmem_cache_node
*n
= get_node(s
, node
);
1086 return atomic_long_read(&n
->nr_slabs
);
1089 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1091 return atomic_long_read(&n
->nr_slabs
);
1094 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1096 struct kmem_cache_node
*n
= get_node(s
, node
);
1099 * May be called early in order to allocate a slab for the
1100 * kmem_cache_node structure. Solve the chicken-egg
1101 * dilemma by deferring the increment of the count during
1102 * bootstrap (see early_kmem_cache_node_alloc).
1105 atomic_long_inc(&n
->nr_slabs
);
1106 atomic_long_add(objects
, &n
->total_objects
);
1109 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1111 struct kmem_cache_node
*n
= get_node(s
, node
);
1113 atomic_long_dec(&n
->nr_slabs
);
1114 atomic_long_sub(objects
, &n
->total_objects
);
1117 /* Object debug checks for alloc/free paths */
1118 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1121 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1124 init_object(s
, object
, SLUB_RED_INACTIVE
);
1125 init_tracking(s
, object
);
1129 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1131 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1134 metadata_access_enable();
1135 memset(kasan_reset_tag(addr
), POISON_INUSE
, page_size(page
));
1136 metadata_access_disable();
1139 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1140 struct page
*page
, void *object
)
1142 if (!check_slab(s
, page
))
1145 if (!check_valid_pointer(s
, page
, object
)) {
1146 object_err(s
, page
, object
, "Freelist Pointer check fails");
1150 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1156 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1158 void *object
, unsigned long addr
)
1160 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1161 if (!alloc_consistency_checks(s
, page
, object
))
1165 /* Success perform special debug activities for allocs */
1166 if (s
->flags
& SLAB_STORE_USER
)
1167 set_track(s
, object
, TRACK_ALLOC
, addr
);
1168 trace(s
, page
, object
, 1);
1169 init_object(s
, object
, SLUB_RED_ACTIVE
);
1173 if (PageSlab(page
)) {
1175 * If this is a slab page then lets do the best we can
1176 * to avoid issues in the future. Marking all objects
1177 * as used avoids touching the remaining objects.
1179 slab_fix(s
, "Marking all objects used");
1180 page
->inuse
= page
->objects
;
1181 page
->freelist
= NULL
;
1186 static inline int free_consistency_checks(struct kmem_cache
*s
,
1187 struct page
*page
, void *object
, unsigned long addr
)
1189 if (!check_valid_pointer(s
, page
, object
)) {
1190 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1194 if (on_freelist(s
, page
, object
)) {
1195 object_err(s
, page
, object
, "Object already free");
1199 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1202 if (unlikely(s
!= page
->slab_cache
)) {
1203 if (!PageSlab(page
)) {
1204 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1206 } else if (!page
->slab_cache
) {
1207 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1211 object_err(s
, page
, object
,
1212 "page slab pointer corrupt.");
1218 /* Supports checking bulk free of a constructed freelist */
1219 static noinline
int free_debug_processing(
1220 struct kmem_cache
*s
, struct page
*page
,
1221 void *head
, void *tail
, int bulk_cnt
,
1224 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1225 void *object
= head
;
1227 unsigned long flags
;
1230 spin_lock_irqsave(&n
->list_lock
, flags
);
1233 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1234 if (!check_slab(s
, page
))
1241 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1242 if (!free_consistency_checks(s
, page
, object
, addr
))
1246 if (s
->flags
& SLAB_STORE_USER
)
1247 set_track(s
, object
, TRACK_FREE
, addr
);
1248 trace(s
, page
, object
, 0);
1249 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1250 init_object(s
, object
, SLUB_RED_INACTIVE
);
1252 /* Reached end of constructed freelist yet? */
1253 if (object
!= tail
) {
1254 object
= get_freepointer(s
, object
);
1260 if (cnt
!= bulk_cnt
)
1261 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1265 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1267 slab_fix(s
, "Object at 0x%p not freed", object
);
1272 * Parse a block of slub_debug options. Blocks are delimited by ';'
1274 * @str: start of block
1275 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1276 * @slabs: return start of list of slabs, or NULL when there's no list
1277 * @init: assume this is initial parsing and not per-kmem-create parsing
1279 * returns the start of next block if there's any, or NULL
1282 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1284 bool higher_order_disable
= false;
1286 /* Skip any completely empty blocks */
1287 while (*str
&& *str
== ';')
1292 * No options but restriction on slabs. This means full
1293 * debugging for slabs matching a pattern.
1295 *flags
= DEBUG_DEFAULT_FLAGS
;
1300 /* Determine which debug features should be switched on */
1301 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1302 switch (tolower(*str
)) {
1307 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1310 *flags
|= SLAB_RED_ZONE
;
1313 *flags
|= SLAB_POISON
;
1316 *flags
|= SLAB_STORE_USER
;
1319 *flags
|= SLAB_TRACE
;
1322 *flags
|= SLAB_FAILSLAB
;
1326 * Avoid enabling debugging on caches if its minimum
1327 * order would increase as a result.
1329 higher_order_disable
= true;
1333 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1342 /* Skip over the slab list */
1343 while (*str
&& *str
!= ';')
1346 /* Skip any completely empty blocks */
1347 while (*str
&& *str
== ';')
1350 if (init
&& higher_order_disable
)
1351 disable_higher_order_debug
= 1;
1359 static int __init
setup_slub_debug(char *str
)
1364 bool global_slub_debug_changed
= false;
1365 bool slab_list_specified
= false;
1367 slub_debug
= DEBUG_DEFAULT_FLAGS
;
1368 if (*str
++ != '=' || !*str
)
1370 * No options specified. Switch on full debugging.
1376 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1380 global_slub_debug_changed
= true;
1382 slab_list_specified
= true;
1387 * For backwards compatibility, a single list of flags with list of
1388 * slabs means debugging is only enabled for those slabs, so the global
1389 * slub_debug should be 0. We can extended that to multiple lists as
1390 * long as there is no option specifying flags without a slab list.
1392 if (slab_list_specified
) {
1393 if (!global_slub_debug_changed
)
1395 slub_debug_string
= saved_str
;
1398 if (slub_debug
!= 0 || slub_debug_string
)
1399 static_branch_enable(&slub_debug_enabled
);
1400 if ((static_branch_unlikely(&init_on_alloc
) ||
1401 static_branch_unlikely(&init_on_free
)) &&
1402 (slub_debug
& SLAB_POISON
))
1403 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1407 __setup("slub_debug", setup_slub_debug
);
1410 * kmem_cache_flags - apply debugging options to the cache
1411 * @object_size: the size of an object without meta data
1412 * @flags: flags to set
1413 * @name: name of the cache
1415 * Debug option(s) are applied to @flags. In addition to the debug
1416 * option(s), if a slab name (or multiple) is specified i.e.
1417 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1418 * then only the select slabs will receive the debug option(s).
1420 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1421 slab_flags_t flags
, const char *name
)
1426 slab_flags_t block_flags
;
1427 slab_flags_t slub_debug_local
= slub_debug
;
1430 * If the slab cache is for debugging (e.g. kmemleak) then
1431 * don't store user (stack trace) information by default,
1432 * but let the user enable it via the command line below.
1434 if (flags
& SLAB_NOLEAKTRACE
)
1435 slub_debug_local
&= ~SLAB_STORE_USER
;
1438 next_block
= slub_debug_string
;
1439 /* Go through all blocks of debug options, see if any matches our slab's name */
1440 while (next_block
) {
1441 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1444 /* Found a block that has a slab list, search it */
1449 end
= strchrnul(iter
, ',');
1450 if (next_block
&& next_block
< end
)
1451 end
= next_block
- 1;
1453 glob
= strnchr(iter
, end
- iter
, '*');
1455 cmplen
= glob
- iter
;
1457 cmplen
= max_t(size_t, len
, (end
- iter
));
1459 if (!strncmp(name
, iter
, cmplen
)) {
1460 flags
|= block_flags
;
1464 if (!*end
|| *end
== ';')
1470 return flags
| slub_debug_local
;
1472 #else /* !CONFIG_SLUB_DEBUG */
1473 static inline void setup_object_debug(struct kmem_cache
*s
,
1474 struct page
*page
, void *object
) {}
1476 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1478 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1479 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1481 static inline int free_debug_processing(
1482 struct kmem_cache
*s
, struct page
*page
,
1483 void *head
, void *tail
, int bulk_cnt
,
1484 unsigned long addr
) { return 0; }
1486 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1488 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1489 void *object
, u8 val
) { return 1; }
1490 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1491 struct page
*page
) {}
1492 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1493 struct page
*page
) {}
1494 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1495 slab_flags_t flags
, const char *name
)
1499 #define slub_debug 0
1501 #define disable_higher_order_debug 0
1503 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1505 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1507 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1509 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1512 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1513 void **freelist
, void *nextfree
)
1517 #endif /* CONFIG_SLUB_DEBUG */
1520 * Hooks for other subsystems that check memory allocations. In a typical
1521 * production configuration these hooks all should produce no code at all.
1523 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1525 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1526 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1527 kmemleak_alloc(ptr
, size
, 1, flags
);
1531 static __always_inline
void kfree_hook(void *x
)
1534 kasan_kfree_large(x
);
1537 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
,
1540 kmemleak_free_recursive(x
, s
->flags
);
1543 * Trouble is that we may no longer disable interrupts in the fast path
1544 * So in order to make the debug calls that expect irqs to be
1545 * disabled we need to disable interrupts temporarily.
1547 #ifdef CONFIG_LOCKDEP
1549 unsigned long flags
;
1551 local_irq_save(flags
);
1552 debug_check_no_locks_freed(x
, s
->object_size
);
1553 local_irq_restore(flags
);
1556 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1557 debug_check_no_obj_freed(x
, s
->object_size
);
1559 /* Use KCSAN to help debug racy use-after-free. */
1560 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1561 __kcsan_check_access(x
, s
->object_size
,
1562 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1565 * As memory initialization might be integrated into KASAN,
1566 * kasan_slab_free and initialization memset's must be
1567 * kept together to avoid discrepancies in behavior.
1569 * The initialization memset's clear the object and the metadata,
1570 * but don't touch the SLAB redzone.
1575 if (!kasan_has_integrated_init())
1576 memset(kasan_reset_tag(x
), 0, s
->object_size
);
1577 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
: 0;
1578 memset((char *)kasan_reset_tag(x
) + s
->inuse
, 0,
1579 s
->size
- s
->inuse
- rsize
);
1581 /* KASAN might put x into memory quarantine, delaying its reuse. */
1582 return kasan_slab_free(s
, x
, init
);
1585 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1586 void **head
, void **tail
)
1591 void *old_tail
= *tail
? *tail
: *head
;
1593 if (is_kfence_address(next
)) {
1594 slab_free_hook(s
, next
, false);
1598 /* Head and tail of the reconstructed freelist */
1604 next
= get_freepointer(s
, object
);
1606 /* If object's reuse doesn't have to be delayed */
1607 if (!slab_free_hook(s
, object
, slab_want_init_on_free(s
))) {
1608 /* Move object to the new freelist */
1609 set_freepointer(s
, object
, *head
);
1614 } while (object
!= old_tail
);
1619 return *head
!= NULL
;
1622 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1625 setup_object_debug(s
, page
, object
);
1626 object
= kasan_init_slab_obj(s
, object
);
1627 if (unlikely(s
->ctor
)) {
1628 kasan_unpoison_object_data(s
, object
);
1630 kasan_poison_object_data(s
, object
);
1636 * Slab allocation and freeing
1638 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1639 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1642 unsigned int order
= oo_order(oo
);
1644 if (node
== NUMA_NO_NODE
)
1645 page
= alloc_pages(flags
, order
);
1647 page
= __alloc_pages_node(node
, flags
, order
);
1652 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1653 /* Pre-initialize the random sequence cache */
1654 static int init_cache_random_seq(struct kmem_cache
*s
)
1656 unsigned int count
= oo_objects(s
->oo
);
1659 /* Bailout if already initialised */
1663 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1665 pr_err("SLUB: Unable to initialize free list for %s\n",
1670 /* Transform to an offset on the set of pages */
1671 if (s
->random_seq
) {
1674 for (i
= 0; i
< count
; i
++)
1675 s
->random_seq
[i
] *= s
->size
;
1680 /* Initialize each random sequence freelist per cache */
1681 static void __init
init_freelist_randomization(void)
1683 struct kmem_cache
*s
;
1685 mutex_lock(&slab_mutex
);
1687 list_for_each_entry(s
, &slab_caches
, list
)
1688 init_cache_random_seq(s
);
1690 mutex_unlock(&slab_mutex
);
1693 /* Get the next entry on the pre-computed freelist randomized */
1694 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1695 unsigned long *pos
, void *start
,
1696 unsigned long page_limit
,
1697 unsigned long freelist_count
)
1702 * If the target page allocation failed, the number of objects on the
1703 * page might be smaller than the usual size defined by the cache.
1706 idx
= s
->random_seq
[*pos
];
1708 if (*pos
>= freelist_count
)
1710 } while (unlikely(idx
>= page_limit
));
1712 return (char *)start
+ idx
;
1715 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1716 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1721 unsigned long idx
, pos
, page_limit
, freelist_count
;
1723 if (page
->objects
< 2 || !s
->random_seq
)
1726 freelist_count
= oo_objects(s
->oo
);
1727 pos
= get_random_int() % freelist_count
;
1729 page_limit
= page
->objects
* s
->size
;
1730 start
= fixup_red_left(s
, page_address(page
));
1732 /* First entry is used as the base of the freelist */
1733 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1735 cur
= setup_object(s
, page
, cur
);
1736 page
->freelist
= cur
;
1738 for (idx
= 1; idx
< page
->objects
; idx
++) {
1739 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1741 next
= setup_object(s
, page
, next
);
1742 set_freepointer(s
, cur
, next
);
1745 set_freepointer(s
, cur
, NULL
);
1750 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1754 static inline void init_freelist_randomization(void) { }
1755 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1759 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1761 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1764 struct kmem_cache_order_objects oo
= s
->oo
;
1766 void *start
, *p
, *next
;
1770 flags
&= gfp_allowed_mask
;
1772 if (gfpflags_allow_blocking(flags
))
1775 flags
|= s
->allocflags
;
1778 * Let the initial higher-order allocation fail under memory pressure
1779 * so we fall-back to the minimum order allocation.
1781 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1782 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1783 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1785 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1786 if (unlikely(!page
)) {
1790 * Allocation may have failed due to fragmentation.
1791 * Try a lower order alloc if possible
1793 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1794 if (unlikely(!page
))
1796 stat(s
, ORDER_FALLBACK
);
1799 page
->objects
= oo_objects(oo
);
1801 account_slab_page(page
, oo_order(oo
), s
, flags
);
1803 page
->slab_cache
= s
;
1804 __SetPageSlab(page
);
1805 if (page_is_pfmemalloc(page
))
1806 SetPageSlabPfmemalloc(page
);
1808 kasan_poison_slab(page
);
1810 start
= page_address(page
);
1812 setup_page_debug(s
, page
, start
);
1814 shuffle
= shuffle_freelist(s
, page
);
1817 start
= fixup_red_left(s
, start
);
1818 start
= setup_object(s
, page
, start
);
1819 page
->freelist
= start
;
1820 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1822 next
= setup_object(s
, page
, next
);
1823 set_freepointer(s
, p
, next
);
1826 set_freepointer(s
, p
, NULL
);
1829 page
->inuse
= page
->objects
;
1833 if (gfpflags_allow_blocking(flags
))
1834 local_irq_disable();
1838 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1843 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1845 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1846 flags
= kmalloc_fix_flags(flags
);
1848 return allocate_slab(s
,
1849 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1852 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1854 int order
= compound_order(page
);
1855 int pages
= 1 << order
;
1857 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
1860 slab_pad_check(s
, page
);
1861 for_each_object(p
, s
, page_address(page
),
1863 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1866 __ClearPageSlabPfmemalloc(page
);
1867 __ClearPageSlab(page
);
1868 /* In union with page->mapping where page allocator expects NULL */
1869 page
->slab_cache
= NULL
;
1870 if (current
->reclaim_state
)
1871 current
->reclaim_state
->reclaimed_slab
+= pages
;
1872 unaccount_slab_page(page
, order
, s
);
1873 __free_pages(page
, order
);
1876 static void rcu_free_slab(struct rcu_head
*h
)
1878 struct page
*page
= container_of(h
, struct page
, rcu_head
);
1880 __free_slab(page
->slab_cache
, page
);
1883 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
1885 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
1886 call_rcu(&page
->rcu_head
, rcu_free_slab
);
1888 __free_slab(s
, page
);
1891 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
1893 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
1898 * Management of partially allocated slabs.
1901 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
1904 if (tail
== DEACTIVATE_TO_TAIL
)
1905 list_add_tail(&page
->slab_list
, &n
->partial
);
1907 list_add(&page
->slab_list
, &n
->partial
);
1910 static inline void add_partial(struct kmem_cache_node
*n
,
1911 struct page
*page
, int tail
)
1913 lockdep_assert_held(&n
->list_lock
);
1914 __add_partial(n
, page
, tail
);
1917 static inline void remove_partial(struct kmem_cache_node
*n
,
1920 lockdep_assert_held(&n
->list_lock
);
1921 list_del(&page
->slab_list
);
1926 * Remove slab from the partial list, freeze it and
1927 * return the pointer to the freelist.
1929 * Returns a list of objects or NULL if it fails.
1931 static inline void *acquire_slab(struct kmem_cache
*s
,
1932 struct kmem_cache_node
*n
, struct page
*page
,
1933 int mode
, int *objects
)
1936 unsigned long counters
;
1939 lockdep_assert_held(&n
->list_lock
);
1942 * Zap the freelist and set the frozen bit.
1943 * The old freelist is the list of objects for the
1944 * per cpu allocation list.
1946 freelist
= page
->freelist
;
1947 counters
= page
->counters
;
1948 new.counters
= counters
;
1949 *objects
= new.objects
- new.inuse
;
1951 new.inuse
= page
->objects
;
1952 new.freelist
= NULL
;
1954 new.freelist
= freelist
;
1957 VM_BUG_ON(new.frozen
);
1960 if (!__cmpxchg_double_slab(s
, page
,
1962 new.freelist
, new.counters
,
1966 remove_partial(n
, page
);
1971 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
1972 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
1975 * Try to allocate a partial slab from a specific node.
1977 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1978 struct kmem_cache_cpu
*c
, gfp_t flags
)
1980 struct page
*page
, *page2
;
1981 void *object
= NULL
;
1982 unsigned int available
= 0;
1986 * Racy check. If we mistakenly see no partial slabs then we
1987 * just allocate an empty slab. If we mistakenly try to get a
1988 * partial slab and there is none available then get_partial()
1991 if (!n
|| !n
->nr_partial
)
1994 spin_lock(&n
->list_lock
);
1995 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
1998 if (!pfmemalloc_match(page
, flags
))
2001 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
2005 available
+= objects
;
2008 stat(s
, ALLOC_FROM_PARTIAL
);
2011 put_cpu_partial(s
, page
, 0);
2012 stat(s
, CPU_PARTIAL_NODE
);
2014 if (!kmem_cache_has_cpu_partial(s
)
2015 || available
> slub_cpu_partial(s
) / 2)
2019 spin_unlock(&n
->list_lock
);
2024 * Get a page from somewhere. Search in increasing NUMA distances.
2026 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2027 struct kmem_cache_cpu
*c
)
2030 struct zonelist
*zonelist
;
2033 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2035 unsigned int cpuset_mems_cookie
;
2038 * The defrag ratio allows a configuration of the tradeoffs between
2039 * inter node defragmentation and node local allocations. A lower
2040 * defrag_ratio increases the tendency to do local allocations
2041 * instead of attempting to obtain partial slabs from other nodes.
2043 * If the defrag_ratio is set to 0 then kmalloc() always
2044 * returns node local objects. If the ratio is higher then kmalloc()
2045 * may return off node objects because partial slabs are obtained
2046 * from other nodes and filled up.
2048 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2049 * (which makes defrag_ratio = 1000) then every (well almost)
2050 * allocation will first attempt to defrag slab caches on other nodes.
2051 * This means scanning over all nodes to look for partial slabs which
2052 * may be expensive if we do it every time we are trying to find a slab
2053 * with available objects.
2055 if (!s
->remote_node_defrag_ratio
||
2056 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2060 cpuset_mems_cookie
= read_mems_allowed_begin();
2061 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2062 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2063 struct kmem_cache_node
*n
;
2065 n
= get_node(s
, zone_to_nid(zone
));
2067 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2068 n
->nr_partial
> s
->min_partial
) {
2069 object
= get_partial_node(s
, n
, c
, flags
);
2072 * Don't check read_mems_allowed_retry()
2073 * here - if mems_allowed was updated in
2074 * parallel, that was a harmless race
2075 * between allocation and the cpuset
2082 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2083 #endif /* CONFIG_NUMA */
2088 * Get a partial page, lock it and return it.
2090 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2091 struct kmem_cache_cpu
*c
)
2094 int searchnode
= node
;
2096 if (node
== NUMA_NO_NODE
)
2097 searchnode
= numa_mem_id();
2099 object
= get_partial_node(s
, get_node(s
, searchnode
), c
, flags
);
2100 if (object
|| node
!= NUMA_NO_NODE
)
2103 return get_any_partial(s
, flags
, c
);
2106 #ifdef CONFIG_PREEMPTION
2108 * Calculate the next globally unique transaction for disambiguation
2109 * during cmpxchg. The transactions start with the cpu number and are then
2110 * incremented by CONFIG_NR_CPUS.
2112 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2115 * No preemption supported therefore also no need to check for
2121 static inline unsigned long next_tid(unsigned long tid
)
2123 return tid
+ TID_STEP
;
2126 #ifdef SLUB_DEBUG_CMPXCHG
2127 static inline unsigned int tid_to_cpu(unsigned long tid
)
2129 return tid
% TID_STEP
;
2132 static inline unsigned long tid_to_event(unsigned long tid
)
2134 return tid
/ TID_STEP
;
2138 static inline unsigned int init_tid(int cpu
)
2143 static inline void note_cmpxchg_failure(const char *n
,
2144 const struct kmem_cache
*s
, unsigned long tid
)
2146 #ifdef SLUB_DEBUG_CMPXCHG
2147 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2149 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2151 #ifdef CONFIG_PREEMPTION
2152 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2153 pr_warn("due to cpu change %d -> %d\n",
2154 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2157 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2158 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2159 tid_to_event(tid
), tid_to_event(actual_tid
));
2161 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2162 actual_tid
, tid
, next_tid(tid
));
2164 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2167 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2171 for_each_possible_cpu(cpu
)
2172 per_cpu_ptr(s
->cpu_slab
, cpu
)->tid
= init_tid(cpu
);
2176 * Remove the cpu slab
2178 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2179 void *freelist
, struct kmem_cache_cpu
*c
)
2181 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2182 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2183 int lock
= 0, free_delta
= 0;
2184 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2185 void *nextfree
, *freelist_iter
, *freelist_tail
;
2186 int tail
= DEACTIVATE_TO_HEAD
;
2190 if (page
->freelist
) {
2191 stat(s
, DEACTIVATE_REMOTE_FREES
);
2192 tail
= DEACTIVATE_TO_TAIL
;
2196 * Stage one: Count the objects on cpu's freelist as free_delta and
2197 * remember the last object in freelist_tail for later splicing.
2199 freelist_tail
= NULL
;
2200 freelist_iter
= freelist
;
2201 while (freelist_iter
) {
2202 nextfree
= get_freepointer(s
, freelist_iter
);
2205 * If 'nextfree' is invalid, it is possible that the object at
2206 * 'freelist_iter' is already corrupted. So isolate all objects
2207 * starting at 'freelist_iter' by skipping them.
2209 if (freelist_corrupted(s
, page
, &freelist_iter
, nextfree
))
2212 freelist_tail
= freelist_iter
;
2215 freelist_iter
= nextfree
;
2219 * Stage two: Unfreeze the page while splicing the per-cpu
2220 * freelist to the head of page's freelist.
2222 * Ensure that the page is unfrozen while the list presence
2223 * reflects the actual number of objects during unfreeze.
2225 * We setup the list membership and then perform a cmpxchg
2226 * with the count. If there is a mismatch then the page
2227 * is not unfrozen but the page is on the wrong list.
2229 * Then we restart the process which may have to remove
2230 * the page from the list that we just put it on again
2231 * because the number of objects in the slab may have
2236 old
.freelist
= READ_ONCE(page
->freelist
);
2237 old
.counters
= READ_ONCE(page
->counters
);
2238 VM_BUG_ON(!old
.frozen
);
2240 /* Determine target state of the slab */
2241 new.counters
= old
.counters
;
2242 if (freelist_tail
) {
2243 new.inuse
-= free_delta
;
2244 set_freepointer(s
, freelist_tail
, old
.freelist
);
2245 new.freelist
= freelist
;
2247 new.freelist
= old
.freelist
;
2251 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2253 else if (new.freelist
) {
2258 * Taking the spinlock removes the possibility
2259 * that acquire_slab() will see a slab page that
2262 spin_lock(&n
->list_lock
);
2266 if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
) && !lock
) {
2269 * This also ensures that the scanning of full
2270 * slabs from diagnostic functions will not see
2273 spin_lock(&n
->list_lock
);
2279 remove_partial(n
, page
);
2280 else if (l
== M_FULL
)
2281 remove_full(s
, n
, page
);
2284 add_partial(n
, page
, tail
);
2285 else if (m
== M_FULL
)
2286 add_full(s
, n
, page
);
2290 if (!__cmpxchg_double_slab(s
, page
,
2291 old
.freelist
, old
.counters
,
2292 new.freelist
, new.counters
,
2297 spin_unlock(&n
->list_lock
);
2301 else if (m
== M_FULL
)
2302 stat(s
, DEACTIVATE_FULL
);
2303 else if (m
== M_FREE
) {
2304 stat(s
, DEACTIVATE_EMPTY
);
2305 discard_slab(s
, page
);
2314 * Unfreeze all the cpu partial slabs.
2316 * This function must be called with interrupts disabled
2317 * for the cpu using c (or some other guarantee must be there
2318 * to guarantee no concurrent accesses).
2320 static void unfreeze_partials(struct kmem_cache
*s
,
2321 struct kmem_cache_cpu
*c
)
2323 #ifdef CONFIG_SLUB_CPU_PARTIAL
2324 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2325 struct page
*page
, *discard_page
= NULL
;
2327 while ((page
= slub_percpu_partial(c
))) {
2331 slub_set_percpu_partial(c
, page
);
2333 n2
= get_node(s
, page_to_nid(page
));
2336 spin_unlock(&n
->list_lock
);
2339 spin_lock(&n
->list_lock
);
2344 old
.freelist
= page
->freelist
;
2345 old
.counters
= page
->counters
;
2346 VM_BUG_ON(!old
.frozen
);
2348 new.counters
= old
.counters
;
2349 new.freelist
= old
.freelist
;
2353 } while (!__cmpxchg_double_slab(s
, page
,
2354 old
.freelist
, old
.counters
,
2355 new.freelist
, new.counters
,
2356 "unfreezing slab"));
2358 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2359 page
->next
= discard_page
;
2360 discard_page
= page
;
2362 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2363 stat(s
, FREE_ADD_PARTIAL
);
2368 spin_unlock(&n
->list_lock
);
2370 while (discard_page
) {
2371 page
= discard_page
;
2372 discard_page
= discard_page
->next
;
2374 stat(s
, DEACTIVATE_EMPTY
);
2375 discard_slab(s
, page
);
2378 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2382 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2383 * partial page slot if available.
2385 * If we did not find a slot then simply move all the partials to the
2386 * per node partial list.
2388 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2390 #ifdef CONFIG_SLUB_CPU_PARTIAL
2391 struct page
*oldpage
;
2399 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2402 pobjects
= oldpage
->pobjects
;
2403 pages
= oldpage
->pages
;
2404 if (drain
&& pobjects
> slub_cpu_partial(s
)) {
2405 unsigned long flags
;
2407 * partial array is full. Move the existing
2408 * set to the per node partial list.
2410 local_irq_save(flags
);
2411 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2412 local_irq_restore(flags
);
2416 stat(s
, CPU_PARTIAL_DRAIN
);
2421 pobjects
+= page
->objects
- page
->inuse
;
2423 page
->pages
= pages
;
2424 page
->pobjects
= pobjects
;
2425 page
->next
= oldpage
;
2427 } while (this_cpu_cmpxchg(s
->cpu_slab
->partial
, oldpage
, page
)
2429 if (unlikely(!slub_cpu_partial(s
))) {
2430 unsigned long flags
;
2432 local_irq_save(flags
);
2433 unfreeze_partials(s
, this_cpu_ptr(s
->cpu_slab
));
2434 local_irq_restore(flags
);
2437 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2440 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2442 stat(s
, CPUSLAB_FLUSH
);
2443 deactivate_slab(s
, c
->page
, c
->freelist
, c
);
2445 c
->tid
= next_tid(c
->tid
);
2451 * Called from IPI handler with interrupts disabled.
2453 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2455 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2460 unfreeze_partials(s
, c
);
2463 static void flush_cpu_slab(void *d
)
2465 struct kmem_cache
*s
= d
;
2467 __flush_cpu_slab(s
, smp_processor_id());
2470 static bool has_cpu_slab(int cpu
, void *info
)
2472 struct kmem_cache
*s
= info
;
2473 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2475 return c
->page
|| slub_percpu_partial(c
);
2478 static void flush_all(struct kmem_cache
*s
)
2480 on_each_cpu_cond(has_cpu_slab
, flush_cpu_slab
, s
, 1);
2484 * Use the cpu notifier to insure that the cpu slabs are flushed when
2487 static int slub_cpu_dead(unsigned int cpu
)
2489 struct kmem_cache
*s
;
2490 unsigned long flags
;
2492 mutex_lock(&slab_mutex
);
2493 list_for_each_entry(s
, &slab_caches
, list
) {
2494 local_irq_save(flags
);
2495 __flush_cpu_slab(s
, cpu
);
2496 local_irq_restore(flags
);
2498 mutex_unlock(&slab_mutex
);
2503 * Check if the objects in a per cpu structure fit numa
2504 * locality expectations.
2506 static inline int node_match(struct page
*page
, int node
)
2509 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2515 #ifdef CONFIG_SLUB_DEBUG
2516 static int count_free(struct page
*page
)
2518 return page
->objects
- page
->inuse
;
2521 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2523 return atomic_long_read(&n
->total_objects
);
2525 #endif /* CONFIG_SLUB_DEBUG */
2527 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2528 static unsigned long count_partial(struct kmem_cache_node
*n
,
2529 int (*get_count
)(struct page
*))
2531 unsigned long flags
;
2532 unsigned long x
= 0;
2535 spin_lock_irqsave(&n
->list_lock
, flags
);
2536 list_for_each_entry(page
, &n
->partial
, slab_list
)
2537 x
+= get_count(page
);
2538 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2541 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2543 static noinline
void
2544 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2546 #ifdef CONFIG_SLUB_DEBUG
2547 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2548 DEFAULT_RATELIMIT_BURST
);
2550 struct kmem_cache_node
*n
;
2552 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2555 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2556 nid
, gfpflags
, &gfpflags
);
2557 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2558 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2561 if (oo_order(s
->min
) > get_order(s
->object_size
))
2562 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2565 for_each_kmem_cache_node(s
, node
, n
) {
2566 unsigned long nr_slabs
;
2567 unsigned long nr_objs
;
2568 unsigned long nr_free
;
2570 nr_free
= count_partial(n
, count_free
);
2571 nr_slabs
= node_nr_slabs(n
);
2572 nr_objs
= node_nr_objs(n
);
2574 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2575 node
, nr_slabs
, nr_objs
, nr_free
);
2580 static inline void *new_slab_objects(struct kmem_cache
*s
, gfp_t flags
,
2581 int node
, struct kmem_cache_cpu
**pc
)
2584 struct kmem_cache_cpu
*c
= *pc
;
2587 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2589 freelist
= get_partial(s
, flags
, node
, c
);
2594 page
= new_slab(s
, flags
, node
);
2596 c
= raw_cpu_ptr(s
->cpu_slab
);
2601 * No other reference to the page yet so we can
2602 * muck around with it freely without cmpxchg
2604 freelist
= page
->freelist
;
2605 page
->freelist
= NULL
;
2607 stat(s
, ALLOC_SLAB
);
2615 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2617 if (unlikely(PageSlabPfmemalloc(page
)))
2618 return gfp_pfmemalloc_allowed(gfpflags
);
2624 * Check the page->freelist of a page and either transfer the freelist to the
2625 * per cpu freelist or deactivate the page.
2627 * The page is still frozen if the return value is not NULL.
2629 * If this function returns NULL then the page has been unfrozen.
2631 * This function must be called with interrupt disabled.
2633 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2636 unsigned long counters
;
2640 freelist
= page
->freelist
;
2641 counters
= page
->counters
;
2643 new.counters
= counters
;
2644 VM_BUG_ON(!new.frozen
);
2646 new.inuse
= page
->objects
;
2647 new.frozen
= freelist
!= NULL
;
2649 } while (!__cmpxchg_double_slab(s
, page
,
2658 * Slow path. The lockless freelist is empty or we need to perform
2661 * Processing is still very fast if new objects have been freed to the
2662 * regular freelist. In that case we simply take over the regular freelist
2663 * as the lockless freelist and zap the regular freelist.
2665 * If that is not working then we fall back to the partial lists. We take the
2666 * first element of the freelist as the object to allocate now and move the
2667 * rest of the freelist to the lockless freelist.
2669 * And if we were unable to get a new slab from the partial slab lists then
2670 * we need to allocate a new slab. This is the slowest path since it involves
2671 * a call to the page allocator and the setup of a new slab.
2673 * Version of __slab_alloc to use when we know that interrupts are
2674 * already disabled (which is the case for bulk allocation).
2676 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2677 unsigned long addr
, struct kmem_cache_cpu
*c
)
2682 stat(s
, ALLOC_SLOWPATH
);
2687 * if the node is not online or has no normal memory, just
2688 * ignore the node constraint
2690 if (unlikely(node
!= NUMA_NO_NODE
&&
2691 !node_isset(node
, slab_nodes
)))
2692 node
= NUMA_NO_NODE
;
2697 if (unlikely(!node_match(page
, node
))) {
2699 * same as above but node_match() being false already
2700 * implies node != NUMA_NO_NODE
2702 if (!node_isset(node
, slab_nodes
)) {
2703 node
= NUMA_NO_NODE
;
2706 stat(s
, ALLOC_NODE_MISMATCH
);
2707 deactivate_slab(s
, page
, c
->freelist
, c
);
2713 * By rights, we should be searching for a slab page that was
2714 * PFMEMALLOC but right now, we are losing the pfmemalloc
2715 * information when the page leaves the per-cpu allocator
2717 if (unlikely(!pfmemalloc_match(page
, gfpflags
))) {
2718 deactivate_slab(s
, page
, c
->freelist
, c
);
2722 /* must check again c->freelist in case of cpu migration or IRQ */
2723 freelist
= c
->freelist
;
2727 freelist
= get_freelist(s
, page
);
2731 stat(s
, DEACTIVATE_BYPASS
);
2735 stat(s
, ALLOC_REFILL
);
2739 * freelist is pointing to the list of objects to be used.
2740 * page is pointing to the page from which the objects are obtained.
2741 * That page must be frozen for per cpu allocations to work.
2743 VM_BUG_ON(!c
->page
->frozen
);
2744 c
->freelist
= get_freepointer(s
, freelist
);
2745 c
->tid
= next_tid(c
->tid
);
2750 if (slub_percpu_partial(c
)) {
2751 page
= c
->page
= slub_percpu_partial(c
);
2752 slub_set_percpu_partial(c
, page
);
2753 stat(s
, CPU_PARTIAL_ALLOC
);
2757 freelist
= new_slab_objects(s
, gfpflags
, node
, &c
);
2759 if (unlikely(!freelist
)) {
2760 slab_out_of_memory(s
, gfpflags
, node
);
2765 if (likely(!kmem_cache_debug(s
) && pfmemalloc_match(page
, gfpflags
)))
2768 /* Only entered in the debug case */
2769 if (kmem_cache_debug(s
) &&
2770 !alloc_debug_processing(s
, page
, freelist
, addr
))
2771 goto new_slab
; /* Slab failed checks. Next slab needed */
2773 deactivate_slab(s
, page
, get_freepointer(s
, freelist
), c
);
2778 * Another one that disabled interrupt and compensates for possible
2779 * cpu changes by refetching the per cpu area pointer.
2781 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2782 unsigned long addr
, struct kmem_cache_cpu
*c
)
2785 unsigned long flags
;
2787 local_irq_save(flags
);
2788 #ifdef CONFIG_PREEMPTION
2790 * We may have been preempted and rescheduled on a different
2791 * cpu before disabling interrupts. Need to reload cpu area
2794 c
= this_cpu_ptr(s
->cpu_slab
);
2797 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
2798 local_irq_restore(flags
);
2803 * If the object has been wiped upon free, make sure it's fully initialized by
2804 * zeroing out freelist pointer.
2806 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
2809 if (unlikely(slab_want_init_on_free(s
)) && obj
)
2810 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
2815 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2816 * have the fastpath folded into their functions. So no function call
2817 * overhead for requests that can be satisfied on the fastpath.
2819 * The fastpath works by first checking if the lockless freelist can be used.
2820 * If not then __slab_alloc is called for slow processing.
2822 * Otherwise we can simply pick the next object from the lockless free list.
2824 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
2825 gfp_t gfpflags
, int node
, unsigned long addr
, size_t orig_size
)
2828 struct kmem_cache_cpu
*c
;
2831 struct obj_cgroup
*objcg
= NULL
;
2834 s
= slab_pre_alloc_hook(s
, &objcg
, 1, gfpflags
);
2838 object
= kfence_alloc(s
, orig_size
, gfpflags
);
2839 if (unlikely(object
))
2844 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2845 * enabled. We may switch back and forth between cpus while
2846 * reading from one cpu area. That does not matter as long
2847 * as we end up on the original cpu again when doing the cmpxchg.
2849 * We should guarantee that tid and kmem_cache are retrieved on
2850 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2851 * to check if it is matched or not.
2854 tid
= this_cpu_read(s
->cpu_slab
->tid
);
2855 c
= raw_cpu_ptr(s
->cpu_slab
);
2856 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
2857 unlikely(tid
!= READ_ONCE(c
->tid
)));
2860 * Irqless object alloc/free algorithm used here depends on sequence
2861 * of fetching cpu_slab's data. tid should be fetched before anything
2862 * on c to guarantee that object and page associated with previous tid
2863 * won't be used with current tid. If we fetch tid first, object and
2864 * page could be one associated with next tid and our alloc/free
2865 * request will be failed. In this case, we will retry. So, no problem.
2870 * The transaction ids are globally unique per cpu and per operation on
2871 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2872 * occurs on the right processor and that there was no operation on the
2873 * linked list in between.
2876 object
= c
->freelist
;
2878 if (unlikely(!object
|| !page
|| !node_match(page
, node
))) {
2879 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
2881 void *next_object
= get_freepointer_safe(s
, object
);
2884 * The cmpxchg will only match if there was no additional
2885 * operation and if we are on the right processor.
2887 * The cmpxchg does the following atomically (without lock
2889 * 1. Relocate first pointer to the current per cpu area.
2890 * 2. Verify that tid and freelist have not been changed
2891 * 3. If they were not changed replace tid and freelist
2893 * Since this is without lock semantics the protection is only
2894 * against code executing on this cpu *not* from access by
2897 if (unlikely(!this_cpu_cmpxchg_double(
2898 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
2900 next_object
, next_tid(tid
)))) {
2902 note_cmpxchg_failure("slab_alloc", s
, tid
);
2905 prefetch_freepointer(s
, next_object
);
2906 stat(s
, ALLOC_FASTPATH
);
2909 maybe_wipe_obj_freeptr(s
, object
);
2910 init
= slab_want_init_on_alloc(gfpflags
, s
);
2913 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
, init
);
2918 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
2919 gfp_t gfpflags
, unsigned long addr
, size_t orig_size
)
2921 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
, orig_size
);
2924 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
2926 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, s
->object_size
);
2928 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
2933 EXPORT_SYMBOL(kmem_cache_alloc
);
2935 #ifdef CONFIG_TRACING
2936 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
2938 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, size
);
2939 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
2940 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2943 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
2947 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
2949 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, s
->object_size
);
2951 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
2952 s
->object_size
, s
->size
, gfpflags
, node
);
2956 EXPORT_SYMBOL(kmem_cache_alloc_node
);
2958 #ifdef CONFIG_TRACING
2959 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
2961 int node
, size_t size
)
2963 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, size
);
2965 trace_kmalloc_node(_RET_IP_
, ret
,
2966 size
, s
->size
, gfpflags
, node
);
2968 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
2971 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
2973 #endif /* CONFIG_NUMA */
2976 * Slow path handling. This may still be called frequently since objects
2977 * have a longer lifetime than the cpu slabs in most processing loads.
2979 * So we still attempt to reduce cache line usage. Just take the slab
2980 * lock and free the item. If there is no additional partial page
2981 * handling required then we can return immediately.
2983 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
2984 void *head
, void *tail
, int cnt
,
2991 unsigned long counters
;
2992 struct kmem_cache_node
*n
= NULL
;
2993 unsigned long flags
;
2995 stat(s
, FREE_SLOWPATH
);
2997 if (kfence_free(head
))
3000 if (kmem_cache_debug(s
) &&
3001 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
3006 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3009 prior
= page
->freelist
;
3010 counters
= page
->counters
;
3011 set_freepointer(s
, tail
, prior
);
3012 new.counters
= counters
;
3013 was_frozen
= new.frozen
;
3015 if ((!new.inuse
|| !prior
) && !was_frozen
) {
3017 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
3020 * Slab was on no list before and will be
3022 * We can defer the list move and instead
3027 } else { /* Needs to be taken off a list */
3029 n
= get_node(s
, page_to_nid(page
));
3031 * Speculatively acquire the list_lock.
3032 * If the cmpxchg does not succeed then we may
3033 * drop the list_lock without any processing.
3035 * Otherwise the list_lock will synchronize with
3036 * other processors updating the list of slabs.
3038 spin_lock_irqsave(&n
->list_lock
, flags
);
3043 } while (!cmpxchg_double_slab(s
, page
,
3050 if (likely(was_frozen
)) {
3052 * The list lock was not taken therefore no list
3053 * activity can be necessary.
3055 stat(s
, FREE_FROZEN
);
3056 } else if (new.frozen
) {
3058 * If we just froze the page then put it onto the
3059 * per cpu partial list.
3061 put_cpu_partial(s
, page
, 1);
3062 stat(s
, CPU_PARTIAL_FREE
);
3068 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3072 * Objects left in the slab. If it was not on the partial list before
3075 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3076 remove_full(s
, n
, page
);
3077 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
3078 stat(s
, FREE_ADD_PARTIAL
);
3080 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3086 * Slab on the partial list.
3088 remove_partial(n
, page
);
3089 stat(s
, FREE_REMOVE_PARTIAL
);
3091 /* Slab must be on the full list */
3092 remove_full(s
, n
, page
);
3095 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3097 discard_slab(s
, page
);
3101 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3102 * can perform fastpath freeing without additional function calls.
3104 * The fastpath is only possible if we are freeing to the current cpu slab
3105 * of this processor. This typically the case if we have just allocated
3108 * If fastpath is not possible then fall back to __slab_free where we deal
3109 * with all sorts of special processing.
3111 * Bulk free of a freelist with several objects (all pointing to the
3112 * same page) possible by specifying head and tail ptr, plus objects
3113 * count (cnt). Bulk free indicated by tail pointer being set.
3115 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3116 struct page
*page
, void *head
, void *tail
,
3117 int cnt
, unsigned long addr
)
3119 void *tail_obj
= tail
? : head
;
3120 struct kmem_cache_cpu
*c
;
3123 memcg_slab_free_hook(s
, &head
, 1);
3126 * Determine the currently cpus per cpu slab.
3127 * The cpu may change afterward. However that does not matter since
3128 * data is retrieved via this pointer. If we are on the same cpu
3129 * during the cmpxchg then the free will succeed.
3132 tid
= this_cpu_read(s
->cpu_slab
->tid
);
3133 c
= raw_cpu_ptr(s
->cpu_slab
);
3134 } while (IS_ENABLED(CONFIG_PREEMPTION
) &&
3135 unlikely(tid
!= READ_ONCE(c
->tid
)));
3137 /* Same with comment on barrier() in slab_alloc_node() */
3140 if (likely(page
== c
->page
)) {
3141 void **freelist
= READ_ONCE(c
->freelist
);
3143 set_freepointer(s
, tail_obj
, freelist
);
3145 if (unlikely(!this_cpu_cmpxchg_double(
3146 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3148 head
, next_tid(tid
)))) {
3150 note_cmpxchg_failure("slab_free", s
, tid
);
3153 stat(s
, FREE_FASTPATH
);
3155 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3159 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3160 void *head
, void *tail
, int cnt
,
3164 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3165 * to remove objects, whose reuse must be delayed.
3167 if (slab_free_freelist_hook(s
, &head
, &tail
))
3168 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3171 #ifdef CONFIG_KASAN_GENERIC
3172 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3174 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3178 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3180 s
= cache_from_obj(s
, x
);
3183 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3184 trace_kmem_cache_free(_RET_IP_
, x
, s
->name
);
3186 EXPORT_SYMBOL(kmem_cache_free
);
3188 struct detached_freelist
{
3193 struct kmem_cache
*s
;
3197 * This function progressively scans the array with free objects (with
3198 * a limited look ahead) and extract objects belonging to the same
3199 * page. It builds a detached freelist directly within the given
3200 * page/objects. This can happen without any need for
3201 * synchronization, because the objects are owned by running process.
3202 * The freelist is build up as a single linked list in the objects.
3203 * The idea is, that this detached freelist can then be bulk
3204 * transferred to the real freelist(s), but only requiring a single
3205 * synchronization primitive. Look ahead in the array is limited due
3206 * to performance reasons.
3209 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3210 void **p
, struct detached_freelist
*df
)
3212 size_t first_skipped_index
= 0;
3217 /* Always re-init detached_freelist */
3222 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3223 } while (!object
&& size
);
3228 page
= virt_to_head_page(object
);
3230 /* Handle kalloc'ed objects */
3231 if (unlikely(!PageSlab(page
))) {
3232 BUG_ON(!PageCompound(page
));
3234 __free_pages(page
, compound_order(page
));
3235 p
[size
] = NULL
; /* mark object processed */
3238 /* Derive kmem_cache from object */
3239 df
->s
= page
->slab_cache
;
3241 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3244 if (is_kfence_address(object
)) {
3245 slab_free_hook(df
->s
, object
, false);
3246 __kfence_free(object
);
3247 p
[size
] = NULL
; /* mark object processed */
3251 /* Start new detached freelist */
3253 set_freepointer(df
->s
, object
, NULL
);
3255 df
->freelist
= object
;
3256 p
[size
] = NULL
; /* mark object processed */
3262 continue; /* Skip processed objects */
3264 /* df->page is always set at this point */
3265 if (df
->page
== virt_to_head_page(object
)) {
3266 /* Opportunity build freelist */
3267 set_freepointer(df
->s
, object
, df
->freelist
);
3268 df
->freelist
= object
;
3270 p
[size
] = NULL
; /* mark object processed */
3275 /* Limit look ahead search */
3279 if (!first_skipped_index
)
3280 first_skipped_index
= size
+ 1;
3283 return first_skipped_index
;
3286 /* Note that interrupts must be enabled when calling this function. */
3287 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3292 memcg_slab_free_hook(s
, p
, size
);
3294 struct detached_freelist df
;
3296 size
= build_detached_freelist(s
, size
, p
, &df
);
3300 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
, _RET_IP_
);
3301 } while (likely(size
));
3303 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3305 /* Note that interrupts must be enabled when calling this function. */
3306 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3309 struct kmem_cache_cpu
*c
;
3311 struct obj_cgroup
*objcg
= NULL
;
3313 /* memcg and kmem_cache debug support */
3314 s
= slab_pre_alloc_hook(s
, &objcg
, size
, flags
);
3318 * Drain objects in the per cpu slab, while disabling local
3319 * IRQs, which protects against PREEMPT and interrupts
3320 * handlers invoking normal fastpath.
3322 local_irq_disable();
3323 c
= this_cpu_ptr(s
->cpu_slab
);
3325 for (i
= 0; i
< size
; i
++) {
3326 void *object
= kfence_alloc(s
, s
->object_size
, flags
);
3328 if (unlikely(object
)) {
3333 object
= c
->freelist
;
3334 if (unlikely(!object
)) {
3336 * We may have removed an object from c->freelist using
3337 * the fastpath in the previous iteration; in that case,
3338 * c->tid has not been bumped yet.
3339 * Since ___slab_alloc() may reenable interrupts while
3340 * allocating memory, we should bump c->tid now.
3342 c
->tid
= next_tid(c
->tid
);
3345 * Invoking slow path likely have side-effect
3346 * of re-populating per CPU c->freelist
3348 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3350 if (unlikely(!p
[i
]))
3353 c
= this_cpu_ptr(s
->cpu_slab
);
3354 maybe_wipe_obj_freeptr(s
, p
[i
]);
3356 continue; /* goto for-loop */
3358 c
->freelist
= get_freepointer(s
, object
);
3360 maybe_wipe_obj_freeptr(s
, p
[i
]);
3362 c
->tid
= next_tid(c
->tid
);
3366 * memcg and kmem_cache debug support and memory initialization.
3367 * Done outside of the IRQ disabled fastpath loop.
3369 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3370 slab_want_init_on_alloc(flags
, s
));
3374 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false);
3375 __kmem_cache_free_bulk(s
, i
, p
);
3378 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3382 * Object placement in a slab is made very easy because we always start at
3383 * offset 0. If we tune the size of the object to the alignment then we can
3384 * get the required alignment by putting one properly sized object after
3387 * Notice that the allocation order determines the sizes of the per cpu
3388 * caches. Each processor has always one slab available for allocations.
3389 * Increasing the allocation order reduces the number of times that slabs
3390 * must be moved on and off the partial lists and is therefore a factor in
3395 * Minimum / Maximum order of slab pages. This influences locking overhead
3396 * and slab fragmentation. A higher order reduces the number of partial slabs
3397 * and increases the number of allocations possible without having to
3398 * take the list_lock.
3400 static unsigned int slub_min_order
;
3401 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3402 static unsigned int slub_min_objects
;
3405 * Calculate the order of allocation given an slab object size.
3407 * The order of allocation has significant impact on performance and other
3408 * system components. Generally order 0 allocations should be preferred since
3409 * order 0 does not cause fragmentation in the page allocator. Larger objects
3410 * be problematic to put into order 0 slabs because there may be too much
3411 * unused space left. We go to a higher order if more than 1/16th of the slab
3414 * In order to reach satisfactory performance we must ensure that a minimum
3415 * number of objects is in one slab. Otherwise we may generate too much
3416 * activity on the partial lists which requires taking the list_lock. This is
3417 * less a concern for large slabs though which are rarely used.
3419 * slub_max_order specifies the order where we begin to stop considering the
3420 * number of objects in a slab as critical. If we reach slub_max_order then
3421 * we try to keep the page order as low as possible. So we accept more waste
3422 * of space in favor of a small page order.
3424 * Higher order allocations also allow the placement of more objects in a
3425 * slab and thereby reduce object handling overhead. If the user has
3426 * requested a higher minimum order then we start with that one instead of
3427 * the smallest order which will fit the object.
3429 static inline unsigned int slab_order(unsigned int size
,
3430 unsigned int min_objects
, unsigned int max_order
,
3431 unsigned int fract_leftover
)
3433 unsigned int min_order
= slub_min_order
;
3436 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3437 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3439 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3440 order
<= max_order
; order
++) {
3442 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3445 rem
= slab_size
% size
;
3447 if (rem
<= slab_size
/ fract_leftover
)
3454 static inline int calculate_order(unsigned int size
)
3457 unsigned int min_objects
;
3458 unsigned int max_objects
;
3459 unsigned int nr_cpus
;
3462 * Attempt to find best configuration for a slab. This
3463 * works by first attempting to generate a layout with
3464 * the best configuration and backing off gradually.
3466 * First we increase the acceptable waste in a slab. Then
3467 * we reduce the minimum objects required in a slab.
3469 min_objects
= slub_min_objects
;
3472 * Some architectures will only update present cpus when
3473 * onlining them, so don't trust the number if it's just 1. But
3474 * we also don't want to use nr_cpu_ids always, as on some other
3475 * architectures, there can be many possible cpus, but never
3476 * onlined. Here we compromise between trying to avoid too high
3477 * order on systems that appear larger than they are, and too
3478 * low order on systems that appear smaller than they are.
3480 nr_cpus
= num_present_cpus();
3482 nr_cpus
= nr_cpu_ids
;
3483 min_objects
= 4 * (fls(nr_cpus
) + 1);
3485 max_objects
= order_objects(slub_max_order
, size
);
3486 min_objects
= min(min_objects
, max_objects
);
3488 while (min_objects
> 1) {
3489 unsigned int fraction
;
3492 while (fraction
>= 4) {
3493 order
= slab_order(size
, min_objects
,
3494 slub_max_order
, fraction
);
3495 if (order
<= slub_max_order
)
3503 * We were unable to place multiple objects in a slab. Now
3504 * lets see if we can place a single object there.
3506 order
= slab_order(size
, 1, slub_max_order
, 1);
3507 if (order
<= slub_max_order
)
3511 * Doh this slab cannot be placed using slub_max_order.
3513 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3514 if (order
< MAX_ORDER
)
3520 init_kmem_cache_node(struct kmem_cache_node
*n
)
3523 spin_lock_init(&n
->list_lock
);
3524 INIT_LIST_HEAD(&n
->partial
);
3525 #ifdef CONFIG_SLUB_DEBUG
3526 atomic_long_set(&n
->nr_slabs
, 0);
3527 atomic_long_set(&n
->total_objects
, 0);
3528 INIT_LIST_HEAD(&n
->full
);
3532 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3534 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3535 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3538 * Must align to double word boundary for the double cmpxchg
3539 * instructions to work; see __pcpu_double_call_return_bool().
3541 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3542 2 * sizeof(void *));
3547 init_kmem_cache_cpus(s
);
3552 static struct kmem_cache
*kmem_cache_node
;
3555 * No kmalloc_node yet so do it by hand. We know that this is the first
3556 * slab on the node for this slabcache. There are no concurrent accesses
3559 * Note that this function only works on the kmem_cache_node
3560 * when allocating for the kmem_cache_node. This is used for bootstrapping
3561 * memory on a fresh node that has no slab structures yet.
3563 static void early_kmem_cache_node_alloc(int node
)
3566 struct kmem_cache_node
*n
;
3568 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3570 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3573 if (page_to_nid(page
) != node
) {
3574 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3575 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3580 #ifdef CONFIG_SLUB_DEBUG
3581 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3582 init_tracking(kmem_cache_node
, n
);
3584 n
= kasan_slab_alloc(kmem_cache_node
, n
, GFP_KERNEL
, false);
3585 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3588 kmem_cache_node
->node
[node
] = n
;
3589 init_kmem_cache_node(n
);
3590 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3593 * No locks need to be taken here as it has just been
3594 * initialized and there is no concurrent access.
3596 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3599 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3602 struct kmem_cache_node
*n
;
3604 for_each_kmem_cache_node(s
, node
, n
) {
3605 s
->node
[node
] = NULL
;
3606 kmem_cache_free(kmem_cache_node
, n
);
3610 void __kmem_cache_release(struct kmem_cache
*s
)
3612 cache_random_seq_destroy(s
);
3613 free_percpu(s
->cpu_slab
);
3614 free_kmem_cache_nodes(s
);
3617 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3621 for_each_node_mask(node
, slab_nodes
) {
3622 struct kmem_cache_node
*n
;
3624 if (slab_state
== DOWN
) {
3625 early_kmem_cache_node_alloc(node
);
3628 n
= kmem_cache_alloc_node(kmem_cache_node
,
3632 free_kmem_cache_nodes(s
);
3636 init_kmem_cache_node(n
);
3642 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3644 if (min
< MIN_PARTIAL
)
3646 else if (min
> MAX_PARTIAL
)
3648 s
->min_partial
= min
;
3651 static void set_cpu_partial(struct kmem_cache
*s
)
3653 #ifdef CONFIG_SLUB_CPU_PARTIAL
3655 * cpu_partial determined the maximum number of objects kept in the
3656 * per cpu partial lists of a processor.
3658 * Per cpu partial lists mainly contain slabs that just have one
3659 * object freed. If they are used for allocation then they can be
3660 * filled up again with minimal effort. The slab will never hit the
3661 * per node partial lists and therefore no locking will be required.
3663 * This setting also determines
3665 * A) The number of objects from per cpu partial slabs dumped to the
3666 * per node list when we reach the limit.
3667 * B) The number of objects in cpu partial slabs to extract from the
3668 * per node list when we run out of per cpu objects. We only fetch
3669 * 50% to keep some capacity around for frees.
3671 if (!kmem_cache_has_cpu_partial(s
))
3672 slub_set_cpu_partial(s
, 0);
3673 else if (s
->size
>= PAGE_SIZE
)
3674 slub_set_cpu_partial(s
, 2);
3675 else if (s
->size
>= 1024)
3676 slub_set_cpu_partial(s
, 6);
3677 else if (s
->size
>= 256)
3678 slub_set_cpu_partial(s
, 13);
3680 slub_set_cpu_partial(s
, 30);
3685 * calculate_sizes() determines the order and the distribution of data within
3688 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
3690 slab_flags_t flags
= s
->flags
;
3691 unsigned int size
= s
->object_size
;
3692 unsigned int freepointer_area
;
3696 * Round up object size to the next word boundary. We can only
3697 * place the free pointer at word boundaries and this determines
3698 * the possible location of the free pointer.
3700 size
= ALIGN(size
, sizeof(void *));
3702 * This is the area of the object where a freepointer can be
3703 * safely written. If redzoning adds more to the inuse size, we
3704 * can't use that portion for writing the freepointer, so
3705 * s->offset must be limited within this for the general case.
3707 freepointer_area
= size
;
3709 #ifdef CONFIG_SLUB_DEBUG
3711 * Determine if we can poison the object itself. If the user of
3712 * the slab may touch the object after free or before allocation
3713 * then we should never poison the object itself.
3715 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
3717 s
->flags
|= __OBJECT_POISON
;
3719 s
->flags
&= ~__OBJECT_POISON
;
3723 * If we are Redzoning then check if there is some space between the
3724 * end of the object and the free pointer. If not then add an
3725 * additional word to have some bytes to store Redzone information.
3727 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
3728 size
+= sizeof(void *);
3732 * With that we have determined the number of bytes in actual use
3733 * by the object. This is the potential offset to the free pointer.
3737 if (((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
3740 * Relocate free pointer after the object if it is not
3741 * permitted to overwrite the first word of the object on
3744 * This is the case if we do RCU, have a constructor or
3745 * destructor or are poisoning the objects.
3747 * The assumption that s->offset >= s->inuse means free
3748 * pointer is outside of the object is used in the
3749 * freeptr_outside_object() function. If that is no
3750 * longer true, the function needs to be modified.
3753 size
+= sizeof(void *);
3754 } else if (freepointer_area
> sizeof(void *)) {
3756 * Store freelist pointer near middle of object to keep
3757 * it away from the edges of the object to avoid small
3758 * sized over/underflows from neighboring allocations.
3760 s
->offset
= ALIGN(freepointer_area
/ 2, sizeof(void *));
3763 #ifdef CONFIG_SLUB_DEBUG
3764 if (flags
& SLAB_STORE_USER
)
3766 * Need to store information about allocs and frees after
3769 size
+= 2 * sizeof(struct track
);
3772 kasan_cache_create(s
, &size
, &s
->flags
);
3773 #ifdef CONFIG_SLUB_DEBUG
3774 if (flags
& SLAB_RED_ZONE
) {
3776 * Add some empty padding so that we can catch
3777 * overwrites from earlier objects rather than let
3778 * tracking information or the free pointer be
3779 * corrupted if a user writes before the start
3782 size
+= sizeof(void *);
3784 s
->red_left_pad
= sizeof(void *);
3785 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
3786 size
+= s
->red_left_pad
;
3791 * SLUB stores one object immediately after another beginning from
3792 * offset 0. In order to align the objects we have to simply size
3793 * each object to conform to the alignment.
3795 size
= ALIGN(size
, s
->align
);
3797 s
->reciprocal_size
= reciprocal_value(size
);
3798 if (forced_order
>= 0)
3799 order
= forced_order
;
3801 order
= calculate_order(size
);
3808 s
->allocflags
|= __GFP_COMP
;
3810 if (s
->flags
& SLAB_CACHE_DMA
)
3811 s
->allocflags
|= GFP_DMA
;
3813 if (s
->flags
& SLAB_CACHE_DMA32
)
3814 s
->allocflags
|= GFP_DMA32
;
3816 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
3817 s
->allocflags
|= __GFP_RECLAIMABLE
;
3820 * Determine the number of objects per slab
3822 s
->oo
= oo_make(order
, size
);
3823 s
->min
= oo_make(get_order(size
), size
);
3824 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
3827 return !!oo_objects(s
->oo
);
3830 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
3832 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
);
3833 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3834 s
->random
= get_random_long();
3837 if (!calculate_sizes(s
, -1))
3839 if (disable_higher_order_debug
) {
3841 * Disable debugging flags that store metadata if the min slab
3844 if (get_order(s
->size
) > get_order(s
->object_size
)) {
3845 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
3847 if (!calculate_sizes(s
, -1))
3852 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3853 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3854 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
3855 /* Enable fast mode */
3856 s
->flags
|= __CMPXCHG_DOUBLE
;
3860 * The larger the object size is, the more pages we want on the partial
3861 * list to avoid pounding the page allocator excessively.
3863 set_min_partial(s
, ilog2(s
->size
) / 2);
3868 s
->remote_node_defrag_ratio
= 1000;
3871 /* Initialize the pre-computed randomized freelist if slab is up */
3872 if (slab_state
>= UP
) {
3873 if (init_cache_random_seq(s
))
3877 if (!init_kmem_cache_nodes(s
))
3880 if (alloc_kmem_cache_cpus(s
))
3883 free_kmem_cache_nodes(s
);
3888 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
3891 #ifdef CONFIG_SLUB_DEBUG
3892 void *addr
= page_address(page
);
3896 slab_err(s
, page
, text
, s
->name
);
3899 map
= get_map(s
, page
);
3900 for_each_object(p
, s
, addr
, page
->objects
) {
3902 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
3903 pr_err("Object 0x%p @offset=%tu\n", p
, p
- addr
);
3904 print_tracking(s
, p
);
3913 * Attempt to free all partial slabs on a node.
3914 * This is called from __kmem_cache_shutdown(). We must take list_lock
3915 * because sysfs file might still access partial list after the shutdowning.
3917 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
3920 struct page
*page
, *h
;
3922 BUG_ON(irqs_disabled());
3923 spin_lock_irq(&n
->list_lock
);
3924 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
3926 remove_partial(n
, page
);
3927 list_add(&page
->slab_list
, &discard
);
3929 list_slab_objects(s
, page
,
3930 "Objects remaining in %s on __kmem_cache_shutdown()");
3933 spin_unlock_irq(&n
->list_lock
);
3935 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
3936 discard_slab(s
, page
);
3939 bool __kmem_cache_empty(struct kmem_cache
*s
)
3942 struct kmem_cache_node
*n
;
3944 for_each_kmem_cache_node(s
, node
, n
)
3945 if (n
->nr_partial
|| slabs_node(s
, node
))
3951 * Release all resources used by a slab cache.
3953 int __kmem_cache_shutdown(struct kmem_cache
*s
)
3956 struct kmem_cache_node
*n
;
3959 /* Attempt to free all objects */
3960 for_each_kmem_cache_node(s
, node
, n
) {
3962 if (n
->nr_partial
|| slabs_node(s
, node
))
3968 #ifdef CONFIG_PRINTK
3969 void kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct page
*page
)
3972 int __maybe_unused i
;
3976 struct kmem_cache
*s
= page
->slab_cache
;
3977 struct track __maybe_unused
*trackp
;
3979 kpp
->kp_ptr
= object
;
3980 kpp
->kp_page
= page
;
3981 kpp
->kp_slab_cache
= s
;
3982 base
= page_address(page
);
3983 objp0
= kasan_reset_tag(object
);
3984 #ifdef CONFIG_SLUB_DEBUG
3985 objp
= restore_red_left(s
, objp0
);
3989 objnr
= obj_to_index(s
, page
, objp
);
3990 kpp
->kp_data_offset
= (unsigned long)((char *)objp0
- (char *)objp
);
3991 objp
= base
+ s
->size
* objnr
;
3992 kpp
->kp_objp
= objp
;
3993 if (WARN_ON_ONCE(objp
< base
|| objp
>= base
+ page
->objects
* s
->size
|| (objp
- base
) % s
->size
) ||
3994 !(s
->flags
& SLAB_STORE_USER
))
3996 #ifdef CONFIG_SLUB_DEBUG
3997 trackp
= get_track(s
, objp
, TRACK_ALLOC
);
3998 kpp
->kp_ret
= (void *)trackp
->addr
;
3999 #ifdef CONFIG_STACKTRACE
4000 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< TRACK_ADDRS_COUNT
; i
++) {
4001 kpp
->kp_stack
[i
] = (void *)trackp
->addrs
[i
];
4002 if (!kpp
->kp_stack
[i
])
4010 /********************************************************************
4012 *******************************************************************/
4014 static int __init
setup_slub_min_order(char *str
)
4016 get_option(&str
, (int *)&slub_min_order
);
4021 __setup("slub_min_order=", setup_slub_min_order
);
4023 static int __init
setup_slub_max_order(char *str
)
4025 get_option(&str
, (int *)&slub_max_order
);
4026 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
4031 __setup("slub_max_order=", setup_slub_max_order
);
4033 static int __init
setup_slub_min_objects(char *str
)
4035 get_option(&str
, (int *)&slub_min_objects
);
4040 __setup("slub_min_objects=", setup_slub_min_objects
);
4042 void *__kmalloc(size_t size
, gfp_t flags
)
4044 struct kmem_cache
*s
;
4047 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4048 return kmalloc_large(size
, flags
);
4050 s
= kmalloc_slab(size
, flags
);
4052 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4055 ret
= slab_alloc(s
, flags
, _RET_IP_
, size
);
4057 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
4059 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4063 EXPORT_SYMBOL(__kmalloc
);
4066 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
4070 unsigned int order
= get_order(size
);
4072 flags
|= __GFP_COMP
;
4073 page
= alloc_pages_node(node
, flags
, order
);
4075 ptr
= page_address(page
);
4076 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4077 PAGE_SIZE
<< order
);
4080 return kmalloc_large_node_hook(ptr
, size
, flags
);
4083 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4085 struct kmem_cache
*s
;
4088 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4089 ret
= kmalloc_large_node(size
, flags
, node
);
4091 trace_kmalloc_node(_RET_IP_
, ret
,
4092 size
, PAGE_SIZE
<< get_order(size
),
4098 s
= kmalloc_slab(size
, flags
);
4100 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4103 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
, size
);
4105 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
4107 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4111 EXPORT_SYMBOL(__kmalloc_node
);
4112 #endif /* CONFIG_NUMA */
4114 #ifdef CONFIG_HARDENED_USERCOPY
4116 * Rejects incorrectly sized objects and objects that are to be copied
4117 * to/from userspace but do not fall entirely within the containing slab
4118 * cache's usercopy region.
4120 * Returns NULL if check passes, otherwise const char * to name of cache
4121 * to indicate an error.
4123 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4126 struct kmem_cache
*s
;
4127 unsigned int offset
;
4129 bool is_kfence
= is_kfence_address(ptr
);
4131 ptr
= kasan_reset_tag(ptr
);
4133 /* Find object and usable object size. */
4134 s
= page
->slab_cache
;
4136 /* Reject impossible pointers. */
4137 if (ptr
< page_address(page
))
4138 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4141 /* Find offset within object. */
4143 offset
= ptr
- kfence_object_start(ptr
);
4145 offset
= (ptr
- page_address(page
)) % s
->size
;
4147 /* Adjust for redzone and reject if within the redzone. */
4148 if (!is_kfence
&& kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4149 if (offset
< s
->red_left_pad
)
4150 usercopy_abort("SLUB object in left red zone",
4151 s
->name
, to_user
, offset
, n
);
4152 offset
-= s
->red_left_pad
;
4155 /* Allow address range falling entirely within usercopy region. */
4156 if (offset
>= s
->useroffset
&&
4157 offset
- s
->useroffset
<= s
->usersize
&&
4158 n
<= s
->useroffset
- offset
+ s
->usersize
)
4162 * If the copy is still within the allocated object, produce
4163 * a warning instead of rejecting the copy. This is intended
4164 * to be a temporary method to find any missing usercopy
4167 object_size
= slab_ksize(s
);
4168 if (usercopy_fallback
&&
4169 offset
<= object_size
&& n
<= object_size
- offset
) {
4170 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4174 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4176 #endif /* CONFIG_HARDENED_USERCOPY */
4178 size_t __ksize(const void *object
)
4182 if (unlikely(object
== ZERO_SIZE_PTR
))
4185 page
= virt_to_head_page(object
);
4187 if (unlikely(!PageSlab(page
))) {
4188 WARN_ON(!PageCompound(page
));
4189 return page_size(page
);
4192 return slab_ksize(page
->slab_cache
);
4194 EXPORT_SYMBOL(__ksize
);
4196 void kfree(const void *x
)
4199 void *object
= (void *)x
;
4201 trace_kfree(_RET_IP_
, x
);
4203 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4206 page
= virt_to_head_page(x
);
4207 if (unlikely(!PageSlab(page
))) {
4208 unsigned int order
= compound_order(page
);
4210 BUG_ON(!PageCompound(page
));
4212 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4213 -(PAGE_SIZE
<< order
));
4214 __free_pages(page
, order
);
4217 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4219 EXPORT_SYMBOL(kfree
);
4221 #define SHRINK_PROMOTE_MAX 32
4224 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4225 * up most to the head of the partial lists. New allocations will then
4226 * fill those up and thus they can be removed from the partial lists.
4228 * The slabs with the least items are placed last. This results in them
4229 * being allocated from last increasing the chance that the last objects
4230 * are freed in them.
4232 int __kmem_cache_shrink(struct kmem_cache
*s
)
4236 struct kmem_cache_node
*n
;
4239 struct list_head discard
;
4240 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4241 unsigned long flags
;
4245 for_each_kmem_cache_node(s
, node
, n
) {
4246 INIT_LIST_HEAD(&discard
);
4247 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4248 INIT_LIST_HEAD(promote
+ i
);
4250 spin_lock_irqsave(&n
->list_lock
, flags
);
4253 * Build lists of slabs to discard or promote.
4255 * Note that concurrent frees may occur while we hold the
4256 * list_lock. page->inuse here is the upper limit.
4258 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4259 int free
= page
->objects
- page
->inuse
;
4261 /* Do not reread page->inuse */
4264 /* We do not keep full slabs on the list */
4267 if (free
== page
->objects
) {
4268 list_move(&page
->slab_list
, &discard
);
4270 } else if (free
<= SHRINK_PROMOTE_MAX
)
4271 list_move(&page
->slab_list
, promote
+ free
- 1);
4275 * Promote the slabs filled up most to the head of the
4278 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4279 list_splice(promote
+ i
, &n
->partial
);
4281 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4283 /* Release empty slabs */
4284 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4285 discard_slab(s
, page
);
4287 if (slabs_node(s
, node
))
4294 static int slab_mem_going_offline_callback(void *arg
)
4296 struct kmem_cache
*s
;
4298 mutex_lock(&slab_mutex
);
4299 list_for_each_entry(s
, &slab_caches
, list
)
4300 __kmem_cache_shrink(s
);
4301 mutex_unlock(&slab_mutex
);
4306 static void slab_mem_offline_callback(void *arg
)
4308 struct memory_notify
*marg
= arg
;
4311 offline_node
= marg
->status_change_nid_normal
;
4314 * If the node still has available memory. we need kmem_cache_node
4317 if (offline_node
< 0)
4320 mutex_lock(&slab_mutex
);
4321 node_clear(offline_node
, slab_nodes
);
4323 * We no longer free kmem_cache_node structures here, as it would be
4324 * racy with all get_node() users, and infeasible to protect them with
4327 mutex_unlock(&slab_mutex
);
4330 static int slab_mem_going_online_callback(void *arg
)
4332 struct kmem_cache_node
*n
;
4333 struct kmem_cache
*s
;
4334 struct memory_notify
*marg
= arg
;
4335 int nid
= marg
->status_change_nid_normal
;
4339 * If the node's memory is already available, then kmem_cache_node is
4340 * already created. Nothing to do.
4346 * We are bringing a node online. No memory is available yet. We must
4347 * allocate a kmem_cache_node structure in order to bring the node
4350 mutex_lock(&slab_mutex
);
4351 list_for_each_entry(s
, &slab_caches
, list
) {
4353 * The structure may already exist if the node was previously
4354 * onlined and offlined.
4356 if (get_node(s
, nid
))
4359 * XXX: kmem_cache_alloc_node will fallback to other nodes
4360 * since memory is not yet available from the node that
4363 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4368 init_kmem_cache_node(n
);
4372 * Any cache created after this point will also have kmem_cache_node
4373 * initialized for the new node.
4375 node_set(nid
, slab_nodes
);
4377 mutex_unlock(&slab_mutex
);
4381 static int slab_memory_callback(struct notifier_block
*self
,
4382 unsigned long action
, void *arg
)
4387 case MEM_GOING_ONLINE
:
4388 ret
= slab_mem_going_online_callback(arg
);
4390 case MEM_GOING_OFFLINE
:
4391 ret
= slab_mem_going_offline_callback(arg
);
4394 case MEM_CANCEL_ONLINE
:
4395 slab_mem_offline_callback(arg
);
4398 case MEM_CANCEL_OFFLINE
:
4402 ret
= notifier_from_errno(ret
);
4408 static struct notifier_block slab_memory_callback_nb
= {
4409 .notifier_call
= slab_memory_callback
,
4410 .priority
= SLAB_CALLBACK_PRI
,
4413 /********************************************************************
4414 * Basic setup of slabs
4415 *******************************************************************/
4418 * Used for early kmem_cache structures that were allocated using
4419 * the page allocator. Allocate them properly then fix up the pointers
4420 * that may be pointing to the wrong kmem_cache structure.
4423 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4426 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4427 struct kmem_cache_node
*n
;
4429 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4432 * This runs very early, and only the boot processor is supposed to be
4433 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4436 __flush_cpu_slab(s
, smp_processor_id());
4437 for_each_kmem_cache_node(s
, node
, n
) {
4440 list_for_each_entry(p
, &n
->partial
, slab_list
)
4443 #ifdef CONFIG_SLUB_DEBUG
4444 list_for_each_entry(p
, &n
->full
, slab_list
)
4448 list_add(&s
->list
, &slab_caches
);
4452 void __init
kmem_cache_init(void)
4454 static __initdata
struct kmem_cache boot_kmem_cache
,
4455 boot_kmem_cache_node
;
4458 if (debug_guardpage_minorder())
4461 kmem_cache_node
= &boot_kmem_cache_node
;
4462 kmem_cache
= &boot_kmem_cache
;
4465 * Initialize the nodemask for which we will allocate per node
4466 * structures. Here we don't need taking slab_mutex yet.
4468 for_each_node_state(node
, N_NORMAL_MEMORY
)
4469 node_set(node
, slab_nodes
);
4471 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4472 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4474 register_hotmemory_notifier(&slab_memory_callback_nb
);
4476 /* Able to allocate the per node structures */
4477 slab_state
= PARTIAL
;
4479 create_boot_cache(kmem_cache
, "kmem_cache",
4480 offsetof(struct kmem_cache
, node
) +
4481 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4482 SLAB_HWCACHE_ALIGN
, 0, 0);
4484 kmem_cache
= bootstrap(&boot_kmem_cache
);
4485 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4487 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4488 setup_kmalloc_cache_index_table();
4489 create_kmalloc_caches(0);
4491 /* Setup random freelists for each cache */
4492 init_freelist_randomization();
4494 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4497 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4499 slub_min_order
, slub_max_order
, slub_min_objects
,
4500 nr_cpu_ids
, nr_node_ids
);
4503 void __init
kmem_cache_init_late(void)
4508 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4509 slab_flags_t flags
, void (*ctor
)(void *))
4511 struct kmem_cache
*s
;
4513 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4518 * Adjust the object sizes so that we clear
4519 * the complete object on kzalloc.
4521 s
->object_size
= max(s
->object_size
, size
);
4522 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4524 if (sysfs_slab_alias(s
, name
)) {
4533 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4537 err
= kmem_cache_open(s
, flags
);
4541 /* Mutex is not taken during early boot */
4542 if (slab_state
<= UP
)
4545 err
= sysfs_slab_add(s
);
4547 __kmem_cache_release(s
);
4552 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4554 struct kmem_cache
*s
;
4557 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4558 return kmalloc_large(size
, gfpflags
);
4560 s
= kmalloc_slab(size
, gfpflags
);
4562 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4565 ret
= slab_alloc(s
, gfpflags
, caller
, size
);
4567 /* Honor the call site pointer we received. */
4568 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4572 EXPORT_SYMBOL(__kmalloc_track_caller
);
4575 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4576 int node
, unsigned long caller
)
4578 struct kmem_cache
*s
;
4581 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4582 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4584 trace_kmalloc_node(caller
, ret
,
4585 size
, PAGE_SIZE
<< get_order(size
),
4591 s
= kmalloc_slab(size
, gfpflags
);
4593 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4596 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
, size
);
4598 /* Honor the call site pointer we received. */
4599 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4603 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4607 static int count_inuse(struct page
*page
)
4612 static int count_total(struct page
*page
)
4614 return page
->objects
;
4618 #ifdef CONFIG_SLUB_DEBUG
4619 static void validate_slab(struct kmem_cache
*s
, struct page
*page
)
4622 void *addr
= page_address(page
);
4627 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4630 /* Now we know that a valid freelist exists */
4631 map
= get_map(s
, page
);
4632 for_each_object(p
, s
, addr
, page
->objects
) {
4633 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), map
) ?
4634 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4636 if (!check_object(s
, page
, p
, val
))
4644 static int validate_slab_node(struct kmem_cache
*s
,
4645 struct kmem_cache_node
*n
)
4647 unsigned long count
= 0;
4649 unsigned long flags
;
4651 spin_lock_irqsave(&n
->list_lock
, flags
);
4653 list_for_each_entry(page
, &n
->partial
, slab_list
) {
4654 validate_slab(s
, page
);
4657 if (count
!= n
->nr_partial
)
4658 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4659 s
->name
, count
, n
->nr_partial
);
4661 if (!(s
->flags
& SLAB_STORE_USER
))
4664 list_for_each_entry(page
, &n
->full
, slab_list
) {
4665 validate_slab(s
, page
);
4668 if (count
!= atomic_long_read(&n
->nr_slabs
))
4669 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4670 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
4673 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4677 static long validate_slab_cache(struct kmem_cache
*s
)
4680 unsigned long count
= 0;
4681 struct kmem_cache_node
*n
;
4684 for_each_kmem_cache_node(s
, node
, n
)
4685 count
+= validate_slab_node(s
, n
);
4690 * Generate lists of code addresses where slabcache objects are allocated
4695 unsigned long count
;
4702 DECLARE_BITMAP(cpus
, NR_CPUS
);
4708 unsigned long count
;
4709 struct location
*loc
;
4712 static void free_loc_track(struct loc_track
*t
)
4715 free_pages((unsigned long)t
->loc
,
4716 get_order(sizeof(struct location
) * t
->max
));
4719 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
4724 order
= get_order(sizeof(struct location
) * max
);
4726 l
= (void *)__get_free_pages(flags
, order
);
4731 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
4739 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
4740 const struct track
*track
)
4742 long start
, end
, pos
;
4744 unsigned long caddr
;
4745 unsigned long age
= jiffies
- track
->when
;
4751 pos
= start
+ (end
- start
+ 1) / 2;
4754 * There is nothing at "end". If we end up there
4755 * we need to add something to before end.
4760 caddr
= t
->loc
[pos
].addr
;
4761 if (track
->addr
== caddr
) {
4767 if (age
< l
->min_time
)
4769 if (age
> l
->max_time
)
4772 if (track
->pid
< l
->min_pid
)
4773 l
->min_pid
= track
->pid
;
4774 if (track
->pid
> l
->max_pid
)
4775 l
->max_pid
= track
->pid
;
4777 cpumask_set_cpu(track
->cpu
,
4778 to_cpumask(l
->cpus
));
4780 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4784 if (track
->addr
< caddr
)
4791 * Not found. Insert new tracking element.
4793 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
4799 (t
->count
- pos
) * sizeof(struct location
));
4802 l
->addr
= track
->addr
;
4806 l
->min_pid
= track
->pid
;
4807 l
->max_pid
= track
->pid
;
4808 cpumask_clear(to_cpumask(l
->cpus
));
4809 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
4810 nodes_clear(l
->nodes
);
4811 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
4815 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
4816 struct page
*page
, enum track_item alloc
)
4818 void *addr
= page_address(page
);
4822 map
= get_map(s
, page
);
4823 for_each_object(p
, s
, addr
, page
->objects
)
4824 if (!test_bit(__obj_to_index(s
, addr
, p
), map
))
4825 add_location(t
, s
, get_track(s
, p
, alloc
));
4829 static int list_locations(struct kmem_cache
*s
, char *buf
,
4830 enum track_item alloc
)
4834 struct loc_track t
= { 0, 0, NULL
};
4836 struct kmem_cache_node
*n
;
4838 if (!alloc_loc_track(&t
, PAGE_SIZE
/ sizeof(struct location
),
4840 return sysfs_emit(buf
, "Out of memory\n");
4842 /* Push back cpu slabs */
4845 for_each_kmem_cache_node(s
, node
, n
) {
4846 unsigned long flags
;
4849 if (!atomic_long_read(&n
->nr_slabs
))
4852 spin_lock_irqsave(&n
->list_lock
, flags
);
4853 list_for_each_entry(page
, &n
->partial
, slab_list
)
4854 process_slab(&t
, s
, page
, alloc
);
4855 list_for_each_entry(page
, &n
->full
, slab_list
)
4856 process_slab(&t
, s
, page
, alloc
);
4857 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4860 for (i
= 0; i
< t
.count
; i
++) {
4861 struct location
*l
= &t
.loc
[i
];
4863 len
+= sysfs_emit_at(buf
, len
, "%7ld ", l
->count
);
4866 len
+= sysfs_emit_at(buf
, len
, "%pS", (void *)l
->addr
);
4868 len
+= sysfs_emit_at(buf
, len
, "<not-available>");
4870 if (l
->sum_time
!= l
->min_time
)
4871 len
+= sysfs_emit_at(buf
, len
, " age=%ld/%ld/%ld",
4873 (long)div_u64(l
->sum_time
,
4877 len
+= sysfs_emit_at(buf
, len
, " age=%ld", l
->min_time
);
4879 if (l
->min_pid
!= l
->max_pid
)
4880 len
+= sysfs_emit_at(buf
, len
, " pid=%ld-%ld",
4881 l
->min_pid
, l
->max_pid
);
4883 len
+= sysfs_emit_at(buf
, len
, " pid=%ld",
4886 if (num_online_cpus() > 1 &&
4887 !cpumask_empty(to_cpumask(l
->cpus
)))
4888 len
+= sysfs_emit_at(buf
, len
, " cpus=%*pbl",
4889 cpumask_pr_args(to_cpumask(l
->cpus
)));
4891 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
4892 len
+= sysfs_emit_at(buf
, len
, " nodes=%*pbl",
4893 nodemask_pr_args(&l
->nodes
));
4895 len
+= sysfs_emit_at(buf
, len
, "\n");
4900 len
+= sysfs_emit_at(buf
, len
, "No data\n");
4904 #endif /* CONFIG_SLUB_DEBUG */
4906 #ifdef SLUB_RESILIENCY_TEST
4907 static void __init
resiliency_test(void)
4910 int type
= KMALLOC_NORMAL
;
4912 BUILD_BUG_ON(KMALLOC_MIN_SIZE
> 16 || KMALLOC_SHIFT_HIGH
< 10);
4914 pr_err("SLUB resiliency testing\n");
4915 pr_err("-----------------------\n");
4916 pr_err("A. Corruption after allocation\n");
4918 p
= kzalloc(16, GFP_KERNEL
);
4920 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4923 validate_slab_cache(kmalloc_caches
[type
][4]);
4925 /* Hmmm... The next two are dangerous */
4926 p
= kzalloc(32, GFP_KERNEL
);
4927 p
[32 + sizeof(void *)] = 0x34;
4928 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4930 pr_err("If allocated object is overwritten then not detectable\n\n");
4932 validate_slab_cache(kmalloc_caches
[type
][5]);
4933 p
= kzalloc(64, GFP_KERNEL
);
4934 p
+= 64 + (get_cycles() & 0xff) * sizeof(void *);
4936 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4938 pr_err("If allocated object is overwritten then not detectable\n\n");
4939 validate_slab_cache(kmalloc_caches
[type
][6]);
4941 pr_err("\nB. Corruption after free\n");
4942 p
= kzalloc(128, GFP_KERNEL
);
4945 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p
);
4946 validate_slab_cache(kmalloc_caches
[type
][7]);
4948 p
= kzalloc(256, GFP_KERNEL
);
4951 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p
);
4952 validate_slab_cache(kmalloc_caches
[type
][8]);
4954 p
= kzalloc(512, GFP_KERNEL
);
4957 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p
);
4958 validate_slab_cache(kmalloc_caches
[type
][9]);
4962 static void resiliency_test(void) {};
4964 #endif /* SLUB_RESILIENCY_TEST */
4967 enum slab_stat_type
{
4968 SL_ALL
, /* All slabs */
4969 SL_PARTIAL
, /* Only partially allocated slabs */
4970 SL_CPU
, /* Only slabs used for cpu caches */
4971 SL_OBJECTS
, /* Determine allocated objects not slabs */
4972 SL_TOTAL
/* Determine object capacity not slabs */
4975 #define SO_ALL (1 << SL_ALL)
4976 #define SO_PARTIAL (1 << SL_PARTIAL)
4977 #define SO_CPU (1 << SL_CPU)
4978 #define SO_OBJECTS (1 << SL_OBJECTS)
4979 #define SO_TOTAL (1 << SL_TOTAL)
4981 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
4982 char *buf
, unsigned long flags
)
4984 unsigned long total
= 0;
4987 unsigned long *nodes
;
4990 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
4994 if (flags
& SO_CPU
) {
4997 for_each_possible_cpu(cpu
) {
4998 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
5003 page
= READ_ONCE(c
->page
);
5007 node
= page_to_nid(page
);
5008 if (flags
& SO_TOTAL
)
5010 else if (flags
& SO_OBJECTS
)
5018 page
= slub_percpu_partial_read_once(c
);
5020 node
= page_to_nid(page
);
5021 if (flags
& SO_TOTAL
)
5023 else if (flags
& SO_OBJECTS
)
5034 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5035 * already held which will conflict with an existing lock order:
5037 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5039 * We don't really need mem_hotplug_lock (to hold off
5040 * slab_mem_going_offline_callback) here because slab's memory hot
5041 * unplug code doesn't destroy the kmem_cache->node[] data.
5044 #ifdef CONFIG_SLUB_DEBUG
5045 if (flags
& SO_ALL
) {
5046 struct kmem_cache_node
*n
;
5048 for_each_kmem_cache_node(s
, node
, n
) {
5050 if (flags
& SO_TOTAL
)
5051 x
= atomic_long_read(&n
->total_objects
);
5052 else if (flags
& SO_OBJECTS
)
5053 x
= atomic_long_read(&n
->total_objects
) -
5054 count_partial(n
, count_free
);
5056 x
= atomic_long_read(&n
->nr_slabs
);
5063 if (flags
& SO_PARTIAL
) {
5064 struct kmem_cache_node
*n
;
5066 for_each_kmem_cache_node(s
, node
, n
) {
5067 if (flags
& SO_TOTAL
)
5068 x
= count_partial(n
, count_total
);
5069 else if (flags
& SO_OBJECTS
)
5070 x
= count_partial(n
, count_inuse
);
5078 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5080 for (node
= 0; node
< nr_node_ids
; node
++) {
5082 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5086 len
+= sysfs_emit_at(buf
, len
, "\n");
5092 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5093 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5095 struct slab_attribute
{
5096 struct attribute attr
;
5097 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5098 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5101 #define SLAB_ATTR_RO(_name) \
5102 static struct slab_attribute _name##_attr = \
5103 __ATTR(_name, 0400, _name##_show, NULL)
5105 #define SLAB_ATTR(_name) \
5106 static struct slab_attribute _name##_attr = \
5107 __ATTR(_name, 0600, _name##_show, _name##_store)
5109 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5111 return sysfs_emit(buf
, "%u\n", s
->size
);
5113 SLAB_ATTR_RO(slab_size
);
5115 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5117 return sysfs_emit(buf
, "%u\n", s
->align
);
5119 SLAB_ATTR_RO(align
);
5121 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5123 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5125 SLAB_ATTR_RO(object_size
);
5127 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5129 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5131 SLAB_ATTR_RO(objs_per_slab
);
5133 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5135 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5137 SLAB_ATTR_RO(order
);
5139 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5141 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5144 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5150 err
= kstrtoul(buf
, 10, &min
);
5154 set_min_partial(s
, min
);
5157 SLAB_ATTR(min_partial
);
5159 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5161 return sysfs_emit(buf
, "%u\n", slub_cpu_partial(s
));
5164 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5167 unsigned int objects
;
5170 err
= kstrtouint(buf
, 10, &objects
);
5173 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5176 slub_set_cpu_partial(s
, objects
);
5180 SLAB_ATTR(cpu_partial
);
5182 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5186 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5190 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5192 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5194 SLAB_ATTR_RO(aliases
);
5196 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5198 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5200 SLAB_ATTR_RO(partial
);
5202 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5204 return show_slab_objects(s
, buf
, SO_CPU
);
5206 SLAB_ATTR_RO(cpu_slabs
);
5208 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5210 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5212 SLAB_ATTR_RO(objects
);
5214 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5216 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5218 SLAB_ATTR_RO(objects_partial
);
5220 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5227 for_each_online_cpu(cpu
) {
5230 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5233 pages
+= page
->pages
;
5234 objects
+= page
->pobjects
;
5238 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, pages
);
5241 for_each_online_cpu(cpu
) {
5244 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5246 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5247 cpu
, page
->pobjects
, page
->pages
);
5250 len
+= sysfs_emit_at(buf
, len
, "\n");
5254 SLAB_ATTR_RO(slabs_cpu_partial
);
5256 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5258 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5260 SLAB_ATTR_RO(reclaim_account
);
5262 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5264 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5266 SLAB_ATTR_RO(hwcache_align
);
5268 #ifdef CONFIG_ZONE_DMA
5269 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5271 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5273 SLAB_ATTR_RO(cache_dma
);
5276 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5278 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5280 SLAB_ATTR_RO(usersize
);
5282 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5284 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5286 SLAB_ATTR_RO(destroy_by_rcu
);
5288 #ifdef CONFIG_SLUB_DEBUG
5289 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5291 return show_slab_objects(s
, buf
, SO_ALL
);
5293 SLAB_ATTR_RO(slabs
);
5295 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5297 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5299 SLAB_ATTR_RO(total_objects
);
5301 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5303 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5305 SLAB_ATTR_RO(sanity_checks
);
5307 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5309 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5311 SLAB_ATTR_RO(trace
);
5313 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5315 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5318 SLAB_ATTR_RO(red_zone
);
5320 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5322 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5325 SLAB_ATTR_RO(poison
);
5327 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5329 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5332 SLAB_ATTR_RO(store_user
);
5334 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5339 static ssize_t
validate_store(struct kmem_cache
*s
,
5340 const char *buf
, size_t length
)
5344 if (buf
[0] == '1') {
5345 ret
= validate_slab_cache(s
);
5351 SLAB_ATTR(validate
);
5353 static ssize_t
alloc_calls_show(struct kmem_cache
*s
, char *buf
)
5355 if (!(s
->flags
& SLAB_STORE_USER
))
5357 return list_locations(s
, buf
, TRACK_ALLOC
);
5359 SLAB_ATTR_RO(alloc_calls
);
5361 static ssize_t
free_calls_show(struct kmem_cache
*s
, char *buf
)
5363 if (!(s
->flags
& SLAB_STORE_USER
))
5365 return list_locations(s
, buf
, TRACK_FREE
);
5367 SLAB_ATTR_RO(free_calls
);
5368 #endif /* CONFIG_SLUB_DEBUG */
5370 #ifdef CONFIG_FAILSLAB
5371 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5373 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5375 SLAB_ATTR_RO(failslab
);
5378 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5383 static ssize_t
shrink_store(struct kmem_cache
*s
,
5384 const char *buf
, size_t length
)
5387 kmem_cache_shrink(s
);
5395 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5397 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5400 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5401 const char *buf
, size_t length
)
5406 err
= kstrtouint(buf
, 10, &ratio
);
5412 s
->remote_node_defrag_ratio
= ratio
* 10;
5416 SLAB_ATTR(remote_node_defrag_ratio
);
5419 #ifdef CONFIG_SLUB_STATS
5420 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5422 unsigned long sum
= 0;
5425 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5430 for_each_online_cpu(cpu
) {
5431 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5437 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5440 for_each_online_cpu(cpu
) {
5442 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5447 len
+= sysfs_emit_at(buf
, len
, "\n");
5452 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5456 for_each_online_cpu(cpu
)
5457 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5460 #define STAT_ATTR(si, text) \
5461 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5463 return show_stat(s, buf, si); \
5465 static ssize_t text##_store(struct kmem_cache *s, \
5466 const char *buf, size_t length) \
5468 if (buf[0] != '0') \
5470 clear_stat(s, si); \
5475 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5476 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5477 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5478 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5479 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5480 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5481 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5482 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5483 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5484 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5485 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5486 STAT_ATTR(FREE_SLAB
, free_slab
);
5487 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5488 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5489 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5490 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5491 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5492 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5493 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5494 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5495 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5496 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5497 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5498 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5499 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5500 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5501 #endif /* CONFIG_SLUB_STATS */
5503 static struct attribute
*slab_attrs
[] = {
5504 &slab_size_attr
.attr
,
5505 &object_size_attr
.attr
,
5506 &objs_per_slab_attr
.attr
,
5508 &min_partial_attr
.attr
,
5509 &cpu_partial_attr
.attr
,
5511 &objects_partial_attr
.attr
,
5513 &cpu_slabs_attr
.attr
,
5517 &hwcache_align_attr
.attr
,
5518 &reclaim_account_attr
.attr
,
5519 &destroy_by_rcu_attr
.attr
,
5521 &slabs_cpu_partial_attr
.attr
,
5522 #ifdef CONFIG_SLUB_DEBUG
5523 &total_objects_attr
.attr
,
5525 &sanity_checks_attr
.attr
,
5527 &red_zone_attr
.attr
,
5529 &store_user_attr
.attr
,
5530 &validate_attr
.attr
,
5531 &alloc_calls_attr
.attr
,
5532 &free_calls_attr
.attr
,
5534 #ifdef CONFIG_ZONE_DMA
5535 &cache_dma_attr
.attr
,
5538 &remote_node_defrag_ratio_attr
.attr
,
5540 #ifdef CONFIG_SLUB_STATS
5541 &alloc_fastpath_attr
.attr
,
5542 &alloc_slowpath_attr
.attr
,
5543 &free_fastpath_attr
.attr
,
5544 &free_slowpath_attr
.attr
,
5545 &free_frozen_attr
.attr
,
5546 &free_add_partial_attr
.attr
,
5547 &free_remove_partial_attr
.attr
,
5548 &alloc_from_partial_attr
.attr
,
5549 &alloc_slab_attr
.attr
,
5550 &alloc_refill_attr
.attr
,
5551 &alloc_node_mismatch_attr
.attr
,
5552 &free_slab_attr
.attr
,
5553 &cpuslab_flush_attr
.attr
,
5554 &deactivate_full_attr
.attr
,
5555 &deactivate_empty_attr
.attr
,
5556 &deactivate_to_head_attr
.attr
,
5557 &deactivate_to_tail_attr
.attr
,
5558 &deactivate_remote_frees_attr
.attr
,
5559 &deactivate_bypass_attr
.attr
,
5560 &order_fallback_attr
.attr
,
5561 &cmpxchg_double_fail_attr
.attr
,
5562 &cmpxchg_double_cpu_fail_attr
.attr
,
5563 &cpu_partial_alloc_attr
.attr
,
5564 &cpu_partial_free_attr
.attr
,
5565 &cpu_partial_node_attr
.attr
,
5566 &cpu_partial_drain_attr
.attr
,
5568 #ifdef CONFIG_FAILSLAB
5569 &failslab_attr
.attr
,
5571 &usersize_attr
.attr
,
5576 static const struct attribute_group slab_attr_group
= {
5577 .attrs
= slab_attrs
,
5580 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5581 struct attribute
*attr
,
5584 struct slab_attribute
*attribute
;
5585 struct kmem_cache
*s
;
5588 attribute
= to_slab_attr(attr
);
5591 if (!attribute
->show
)
5594 err
= attribute
->show(s
, buf
);
5599 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5600 struct attribute
*attr
,
5601 const char *buf
, size_t len
)
5603 struct slab_attribute
*attribute
;
5604 struct kmem_cache
*s
;
5607 attribute
= to_slab_attr(attr
);
5610 if (!attribute
->store
)
5613 err
= attribute
->store(s
, buf
, len
);
5617 static void kmem_cache_release(struct kobject
*k
)
5619 slab_kmem_cache_release(to_slab(k
));
5622 static const struct sysfs_ops slab_sysfs_ops
= {
5623 .show
= slab_attr_show
,
5624 .store
= slab_attr_store
,
5627 static struct kobj_type slab_ktype
= {
5628 .sysfs_ops
= &slab_sysfs_ops
,
5629 .release
= kmem_cache_release
,
5632 static struct kset
*slab_kset
;
5634 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5639 #define ID_STR_LENGTH 64
5641 /* Create a unique string id for a slab cache:
5643 * Format :[flags-]size
5645 static char *create_unique_id(struct kmem_cache
*s
)
5647 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5654 * First flags affecting slabcache operations. We will only
5655 * get here for aliasable slabs so we do not need to support
5656 * too many flags. The flags here must cover all flags that
5657 * are matched during merging to guarantee that the id is
5660 if (s
->flags
& SLAB_CACHE_DMA
)
5662 if (s
->flags
& SLAB_CACHE_DMA32
)
5664 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5666 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5668 if (s
->flags
& SLAB_ACCOUNT
)
5672 p
+= sprintf(p
, "%07u", s
->size
);
5674 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5678 static int sysfs_slab_add(struct kmem_cache
*s
)
5682 struct kset
*kset
= cache_kset(s
);
5683 int unmergeable
= slab_unmergeable(s
);
5686 kobject_init(&s
->kobj
, &slab_ktype
);
5690 if (!unmergeable
&& disable_higher_order_debug
&&
5691 (slub_debug
& DEBUG_METADATA_FLAGS
))
5696 * Slabcache can never be merged so we can use the name proper.
5697 * This is typically the case for debug situations. In that
5698 * case we can catch duplicate names easily.
5700 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5704 * Create a unique name for the slab as a target
5707 name
= create_unique_id(s
);
5710 s
->kobj
.kset
= kset
;
5711 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5715 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5720 /* Setup first alias */
5721 sysfs_slab_alias(s
, s
->name
);
5728 kobject_del(&s
->kobj
);
5732 void sysfs_slab_unlink(struct kmem_cache
*s
)
5734 if (slab_state
>= FULL
)
5735 kobject_del(&s
->kobj
);
5738 void sysfs_slab_release(struct kmem_cache
*s
)
5740 if (slab_state
>= FULL
)
5741 kobject_put(&s
->kobj
);
5745 * Need to buffer aliases during bootup until sysfs becomes
5746 * available lest we lose that information.
5748 struct saved_alias
{
5749 struct kmem_cache
*s
;
5751 struct saved_alias
*next
;
5754 static struct saved_alias
*alias_list
;
5756 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5758 struct saved_alias
*al
;
5760 if (slab_state
== FULL
) {
5762 * If we have a leftover link then remove it.
5764 sysfs_remove_link(&slab_kset
->kobj
, name
);
5765 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
5768 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
5774 al
->next
= alias_list
;
5779 static int __init
slab_sysfs_init(void)
5781 struct kmem_cache
*s
;
5784 mutex_lock(&slab_mutex
);
5786 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
5788 mutex_unlock(&slab_mutex
);
5789 pr_err("Cannot register slab subsystem.\n");
5795 list_for_each_entry(s
, &slab_caches
, list
) {
5796 err
= sysfs_slab_add(s
);
5798 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5802 while (alias_list
) {
5803 struct saved_alias
*al
= alias_list
;
5805 alias_list
= alias_list
->next
;
5806 err
= sysfs_slab_alias(al
->s
, al
->name
);
5808 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5813 mutex_unlock(&slab_mutex
);
5818 __initcall(slab_sysfs_init
);
5819 #endif /* CONFIG_SYSFS */
5822 * The /proc/slabinfo ABI
5824 #ifdef CONFIG_SLUB_DEBUG
5825 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
5827 unsigned long nr_slabs
= 0;
5828 unsigned long nr_objs
= 0;
5829 unsigned long nr_free
= 0;
5831 struct kmem_cache_node
*n
;
5833 for_each_kmem_cache_node(s
, node
, n
) {
5834 nr_slabs
+= node_nr_slabs(n
);
5835 nr_objs
+= node_nr_objs(n
);
5836 nr_free
+= count_partial(n
, count_free
);
5839 sinfo
->active_objs
= nr_objs
- nr_free
;
5840 sinfo
->num_objs
= nr_objs
;
5841 sinfo
->active_slabs
= nr_slabs
;
5842 sinfo
->num_slabs
= nr_slabs
;
5843 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
5844 sinfo
->cache_order
= oo_order(s
->oo
);
5847 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
5851 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
5852 size_t count
, loff_t
*ppos
)
5856 #endif /* CONFIG_SLUB_DEBUG */