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/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/kfence.h>
32 #include <linux/memory.h>
33 #include <linux/math64.h>
34 #include <linux/fault-inject.h>
35 #include <linux/stacktrace.h>
36 #include <linux/prefetch.h>
37 #include <linux/memcontrol.h>
38 #include <linux/random.h>
39 #include <kunit/test.h>
41 #include <linux/debugfs.h>
42 #include <trace/events/kmem.h>
48 * 1. slab_mutex (Global Mutex)
49 * 2. node->list_lock (Spinlock)
50 * 3. kmem_cache->cpu_slab->lock (Local lock)
51 * 4. slab_lock(page) (Only on some arches or for debugging)
52 * 5. object_map_lock (Only for debugging)
56 * The role of the slab_mutex is to protect the list of all the slabs
57 * and to synchronize major metadata changes to slab cache structures.
58 * Also synchronizes memory hotplug callbacks.
62 * The slab_lock is a wrapper around the page lock, thus it is a bit
65 * The slab_lock is only used for debugging and on arches that do not
66 * have the ability to do a cmpxchg_double. It only protects:
67 * A. page->freelist -> List of object free in a page
68 * B. page->inuse -> Number of objects in use
69 * C. page->objects -> Number of objects in page
70 * D. page->frozen -> frozen state
74 * If a slab is frozen then it is exempt from list management. It is not
75 * on any list except per cpu partial list. The processor that froze the
76 * slab is the one who can perform list operations on the page. Other
77 * processors may put objects onto the freelist but the processor that
78 * froze the slab is the only one that can retrieve the objects from the
83 * The list_lock protects the partial and full list on each node and
84 * the partial slab counter. If taken then no new slabs may be added or
85 * removed from the lists nor make the number of partial slabs be modified.
86 * (Note that the total number of slabs is an atomic value that may be
87 * modified without taking the list lock).
89 * The list_lock is a centralized lock and thus we avoid taking it as
90 * much as possible. As long as SLUB does not have to handle partial
91 * slabs, operations can continue without any centralized lock. F.e.
92 * allocating a long series of objects that fill up slabs does not require
95 * cpu_slab->lock local lock
97 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
98 * except the stat counters. This is a percpu structure manipulated only by
99 * the local cpu, so the lock protects against being preempted or interrupted
100 * by an irq. Fast path operations rely on lockless operations instead.
101 * On PREEMPT_RT, the local lock does not actually disable irqs (and thus
102 * prevent the lockless operations), so fastpath operations also need to take
103 * the lock and are no longer lockless.
107 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
108 * are fully lockless when satisfied from the percpu slab (and when
109 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
110 * They also don't disable preemption or migration or irqs. They rely on
111 * the transaction id (tid) field to detect being preempted or moved to
114 * irq, preemption, migration considerations
116 * Interrupts are disabled as part of list_lock or local_lock operations, or
117 * around the slab_lock operation, in order to make the slab allocator safe
118 * to use in the context of an irq.
120 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
121 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
122 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
123 * doesn't have to be revalidated in each section protected by the local lock.
125 * SLUB assigns one slab for allocation to each processor.
126 * Allocations only occur from these slabs called cpu slabs.
128 * Slabs with free elements are kept on a partial list and during regular
129 * operations no list for full slabs is used. If an object in a full slab is
130 * freed then the slab will show up again on the partial lists.
131 * We track full slabs for debugging purposes though because otherwise we
132 * cannot scan all objects.
134 * Slabs are freed when they become empty. Teardown and setup is
135 * minimal so we rely on the page allocators per cpu caches for
136 * fast frees and allocs.
138 * page->frozen The slab is frozen and exempt from list processing.
139 * This means that the slab is dedicated to a purpose
140 * such as satisfying allocations for a specific
141 * processor. Objects may be freed in the slab while
142 * it is frozen but slab_free will then skip the usual
143 * list operations. It is up to the processor holding
144 * the slab to integrate the slab into the slab lists
145 * when the slab is no longer needed.
147 * One use of this flag is to mark slabs that are
148 * used for allocations. Then such a slab becomes a cpu
149 * slab. The cpu slab may be equipped with an additional
150 * freelist that allows lockless access to
151 * free objects in addition to the regular freelist
152 * that requires the slab lock.
154 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
155 * options set. This moves slab handling out of
156 * the fast path and disables lockless freelists.
160 * We could simply use migrate_disable()/enable() but as long as it's a
161 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
163 #ifndef CONFIG_PREEMPT_RT
164 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
165 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
167 #define slub_get_cpu_ptr(var) \
172 #define slub_put_cpu_ptr(var) \
179 #ifdef CONFIG_SLUB_DEBUG
180 #ifdef CONFIG_SLUB_DEBUG_ON
181 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled
);
183 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled
);
185 #endif /* CONFIG_SLUB_DEBUG */
187 static inline bool kmem_cache_debug(struct kmem_cache
*s
)
189 return kmem_cache_debug_flags(s
, SLAB_DEBUG_FLAGS
);
192 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
194 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
))
195 p
+= s
->red_left_pad
;
200 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
202 #ifdef CONFIG_SLUB_CPU_PARTIAL
203 return !kmem_cache_debug(s
);
210 * Issues still to be resolved:
212 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
214 * - Variable sizing of the per node arrays
217 /* Enable to log cmpxchg failures */
218 #undef SLUB_DEBUG_CMPXCHG
221 * Minimum number of partial slabs. These will be left on the partial
222 * lists even if they are empty. kmem_cache_shrink may reclaim them.
224 #define MIN_PARTIAL 5
227 * Maximum number of desirable partial slabs.
228 * The existence of more partial slabs makes kmem_cache_shrink
229 * sort the partial list by the number of objects in use.
231 #define MAX_PARTIAL 10
233 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
234 SLAB_POISON | SLAB_STORE_USER)
237 * These debug flags cannot use CMPXCHG because there might be consistency
238 * issues when checking or reading debug information
240 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
245 * Debugging flags that require metadata to be stored in the slab. These get
246 * disabled when slub_debug=O is used and a cache's min order increases with
249 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
252 #define OO_MASK ((1 << OO_SHIFT) - 1)
253 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
255 /* Internal SLUB flags */
257 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
258 /* Use cmpxchg_double */
259 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
262 * Tracking user of a slab.
264 #define TRACK_ADDRS_COUNT 16
266 unsigned long addr
; /* Called from address */
267 #ifdef CONFIG_STACKTRACE
268 unsigned long addrs
[TRACK_ADDRS_COUNT
]; /* Called from address */
270 int cpu
; /* Was running on cpu */
271 int pid
; /* Pid context */
272 unsigned long when
; /* When did the operation occur */
275 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
278 static int sysfs_slab_add(struct kmem_cache
*);
279 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
281 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
282 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
286 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
287 static void debugfs_slab_add(struct kmem_cache
*);
289 static inline void debugfs_slab_add(struct kmem_cache
*s
) { }
292 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
294 #ifdef CONFIG_SLUB_STATS
296 * The rmw is racy on a preemptible kernel but this is acceptable, so
297 * avoid this_cpu_add()'s irq-disable overhead.
299 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
304 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
305 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
306 * differ during memory hotplug/hotremove operations.
307 * Protected by slab_mutex.
309 static nodemask_t slab_nodes
;
312 * Workqueue used for flush_cpu_slab().
314 static struct workqueue_struct
*flushwq
;
316 /********************************************************************
317 * Core slab cache functions
318 *******************************************************************/
321 * Returns freelist pointer (ptr). With hardening, this is obfuscated
322 * with an XOR of the address where the pointer is held and a per-cache
325 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
326 unsigned long ptr_addr
)
328 #ifdef CONFIG_SLAB_FREELIST_HARDENED
330 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
331 * Normally, this doesn't cause any issues, as both set_freepointer()
332 * and get_freepointer() are called with a pointer with the same tag.
333 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
334 * example, when __free_slub() iterates over objects in a cache, it
335 * passes untagged pointers to check_object(). check_object() in turns
336 * calls get_freepointer() with an untagged pointer, which causes the
337 * freepointer to be restored incorrectly.
339 return (void *)((unsigned long)ptr
^ s
->random
^
340 swab((unsigned long)kasan_reset_tag((void *)ptr_addr
)));
346 /* Returns the freelist pointer recorded at location ptr_addr. */
347 static inline void *freelist_dereference(const struct kmem_cache
*s
,
350 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
351 (unsigned long)ptr_addr
);
354 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
356 object
= kasan_reset_tag(object
);
357 return freelist_dereference(s
, object
+ s
->offset
);
360 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
362 prefetch(object
+ s
->offset
);
365 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
367 unsigned long freepointer_addr
;
370 if (!debug_pagealloc_enabled_static())
371 return get_freepointer(s
, object
);
373 object
= kasan_reset_tag(object
);
374 freepointer_addr
= (unsigned long)object
+ s
->offset
;
375 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
376 return freelist_ptr(s
, p
, freepointer_addr
);
379 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
381 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
383 #ifdef CONFIG_SLAB_FREELIST_HARDENED
384 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
387 freeptr_addr
= (unsigned long)kasan_reset_tag((void *)freeptr_addr
);
388 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
391 /* Loop over all objects in a slab */
392 #define for_each_object(__p, __s, __addr, __objects) \
393 for (__p = fixup_red_left(__s, __addr); \
394 __p < (__addr) + (__objects) * (__s)->size; \
397 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
399 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
402 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
405 struct kmem_cache_order_objects x
= {
406 (order
<< OO_SHIFT
) + order_objects(order
, size
)
412 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
414 return x
.x
>> OO_SHIFT
;
417 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
419 return x
.x
& OO_MASK
;
423 * Per slab locking using the pagelock
425 static __always_inline
void __slab_lock(struct page
*page
)
427 VM_BUG_ON_PAGE(PageTail(page
), page
);
428 bit_spin_lock(PG_locked
, &page
->flags
);
431 static __always_inline
void __slab_unlock(struct page
*page
)
433 VM_BUG_ON_PAGE(PageTail(page
), page
);
434 __bit_spin_unlock(PG_locked
, &page
->flags
);
437 static __always_inline
void slab_lock(struct page
*page
, unsigned long *flags
)
439 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
440 local_irq_save(*flags
);
444 static __always_inline
void slab_unlock(struct page
*page
, unsigned long *flags
)
447 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
448 local_irq_restore(*flags
);
452 * Interrupts must be disabled (for the fallback code to work right), typically
453 * by an _irqsave() lock variant. Except on PREEMPT_RT where locks are different
454 * so we disable interrupts as part of slab_[un]lock().
456 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
457 void *freelist_old
, unsigned long counters_old
,
458 void *freelist_new
, unsigned long counters_new
,
461 if (!IS_ENABLED(CONFIG_PREEMPT_RT
))
462 lockdep_assert_irqs_disabled();
463 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
464 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
465 if (s
->flags
& __CMPXCHG_DOUBLE
) {
466 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
467 freelist_old
, counters_old
,
468 freelist_new
, counters_new
))
473 /* init to 0 to prevent spurious warnings */
474 unsigned long flags
= 0;
476 slab_lock(page
, &flags
);
477 if (page
->freelist
== freelist_old
&&
478 page
->counters
== counters_old
) {
479 page
->freelist
= freelist_new
;
480 page
->counters
= counters_new
;
481 slab_unlock(page
, &flags
);
484 slab_unlock(page
, &flags
);
488 stat(s
, CMPXCHG_DOUBLE_FAIL
);
490 #ifdef SLUB_DEBUG_CMPXCHG
491 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
497 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct page
*page
,
498 void *freelist_old
, unsigned long counters_old
,
499 void *freelist_new
, unsigned long counters_new
,
502 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
503 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
504 if (s
->flags
& __CMPXCHG_DOUBLE
) {
505 if (cmpxchg_double(&page
->freelist
, &page
->counters
,
506 freelist_old
, counters_old
,
507 freelist_new
, counters_new
))
514 local_irq_save(flags
);
516 if (page
->freelist
== freelist_old
&&
517 page
->counters
== counters_old
) {
518 page
->freelist
= freelist_new
;
519 page
->counters
= counters_new
;
521 local_irq_restore(flags
);
525 local_irq_restore(flags
);
529 stat(s
, CMPXCHG_DOUBLE_FAIL
);
531 #ifdef SLUB_DEBUG_CMPXCHG
532 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
538 #ifdef CONFIG_SLUB_DEBUG
539 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
540 static DEFINE_RAW_SPINLOCK(object_map_lock
);
542 static void __fill_map(unsigned long *obj_map
, struct kmem_cache
*s
,
545 void *addr
= page_address(page
);
548 bitmap_zero(obj_map
, page
->objects
);
550 for (p
= page
->freelist
; p
; p
= get_freepointer(s
, p
))
551 set_bit(__obj_to_index(s
, addr
, p
), obj_map
);
554 #if IS_ENABLED(CONFIG_KUNIT)
555 static bool slab_add_kunit_errors(void)
557 struct kunit_resource
*resource
;
559 if (likely(!current
->kunit_test
))
562 resource
= kunit_find_named_resource(current
->kunit_test
, "slab_errors");
566 (*(int *)resource
->data
)++;
567 kunit_put_resource(resource
);
571 static inline bool slab_add_kunit_errors(void) { return false; }
575 * Determine a map of object in use on a page.
577 * Node listlock must be held to guarantee that the page does
578 * not vanish from under us.
580 static unsigned long *get_map(struct kmem_cache
*s
, struct page
*page
)
581 __acquires(&object_map_lock
)
583 VM_BUG_ON(!irqs_disabled());
585 raw_spin_lock(&object_map_lock
);
587 __fill_map(object_map
, s
, page
);
592 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
594 VM_BUG_ON(map
!= object_map
);
595 raw_spin_unlock(&object_map_lock
);
598 static inline unsigned int size_from_object(struct kmem_cache
*s
)
600 if (s
->flags
& SLAB_RED_ZONE
)
601 return s
->size
- s
->red_left_pad
;
606 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
608 if (s
->flags
& SLAB_RED_ZONE
)
609 p
-= s
->red_left_pad
;
617 #if defined(CONFIG_SLUB_DEBUG_ON)
618 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
620 static slab_flags_t slub_debug
;
623 static char *slub_debug_string
;
624 static int disable_higher_order_debug
;
627 * slub is about to manipulate internal object metadata. This memory lies
628 * outside the range of the allocated object, so accessing it would normally
629 * be reported by kasan as a bounds error. metadata_access_enable() is used
630 * to tell kasan that these accesses are OK.
632 static inline void metadata_access_enable(void)
634 kasan_disable_current();
637 static inline void metadata_access_disable(void)
639 kasan_enable_current();
646 /* Verify that a pointer has an address that is valid within a slab page */
647 static inline int check_valid_pointer(struct kmem_cache
*s
,
648 struct page
*page
, void *object
)
655 base
= page_address(page
);
656 object
= kasan_reset_tag(object
);
657 object
= restore_red_left(s
, object
);
658 if (object
< base
|| object
>= base
+ page
->objects
* s
->size
||
659 (object
- base
) % s
->size
) {
666 static void print_section(char *level
, char *text
, u8
*addr
,
669 metadata_access_enable();
670 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
,
671 16, 1, kasan_reset_tag((void *)addr
), length
, 1);
672 metadata_access_disable();
676 * See comment in calculate_sizes().
678 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
680 return s
->offset
>= s
->inuse
;
684 * Return offset of the end of info block which is inuse + free pointer if
685 * not overlapping with object.
687 static inline unsigned int get_info_end(struct kmem_cache
*s
)
689 if (freeptr_outside_object(s
))
690 return s
->inuse
+ sizeof(void *);
695 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
696 enum track_item alloc
)
700 p
= object
+ get_info_end(s
);
702 return kasan_reset_tag(p
+ alloc
);
705 static void set_track(struct kmem_cache
*s
, void *object
,
706 enum track_item alloc
, unsigned long addr
)
708 struct track
*p
= get_track(s
, object
, alloc
);
711 #ifdef CONFIG_STACKTRACE
712 unsigned int nr_entries
;
714 metadata_access_enable();
715 nr_entries
= stack_trace_save(kasan_reset_tag(p
->addrs
),
716 TRACK_ADDRS_COUNT
, 3);
717 metadata_access_disable();
719 if (nr_entries
< TRACK_ADDRS_COUNT
)
720 p
->addrs
[nr_entries
] = 0;
723 p
->cpu
= smp_processor_id();
724 p
->pid
= current
->pid
;
727 memset(p
, 0, sizeof(struct track
));
731 static void init_tracking(struct kmem_cache
*s
, void *object
)
733 if (!(s
->flags
& SLAB_STORE_USER
))
736 set_track(s
, object
, TRACK_FREE
, 0UL);
737 set_track(s
, object
, TRACK_ALLOC
, 0UL);
740 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
745 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
746 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
747 #ifdef CONFIG_STACKTRACE
750 for (i
= 0; i
< TRACK_ADDRS_COUNT
; i
++)
752 pr_err("\t%pS\n", (void *)t
->addrs
[i
]);
759 void print_tracking(struct kmem_cache
*s
, void *object
)
761 unsigned long pr_time
= jiffies
;
762 if (!(s
->flags
& SLAB_STORE_USER
))
765 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
766 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
769 static void print_page_info(struct page
*page
)
771 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
772 page
, page
->objects
, page
->inuse
, page
->freelist
,
773 page
->flags
, &page
->flags
);
777 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
779 struct va_format vaf
;
785 pr_err("=============================================================================\n");
786 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
787 pr_err("-----------------------------------------------------------------------------\n\n");
792 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
794 struct va_format vaf
;
797 if (slab_add_kunit_errors())
803 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
807 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
808 void **freelist
, void *nextfree
)
810 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
811 !check_valid_pointer(s
, page
, nextfree
) && freelist
) {
812 object_err(s
, page
, *freelist
, "Freechain corrupt");
814 slab_fix(s
, "Isolate corrupted freechain");
821 static void print_trailer(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
823 unsigned int off
; /* Offset of last byte */
824 u8
*addr
= page_address(page
);
826 print_tracking(s
, p
);
828 print_page_info(page
);
830 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
831 p
, p
- addr
, get_freepointer(s
, p
));
833 if (s
->flags
& SLAB_RED_ZONE
)
834 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
836 else if (p
> addr
+ 16)
837 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
839 print_section(KERN_ERR
, "Object ", p
,
840 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
841 if (s
->flags
& SLAB_RED_ZONE
)
842 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
843 s
->inuse
- s
->object_size
);
845 off
= get_info_end(s
);
847 if (s
->flags
& SLAB_STORE_USER
)
848 off
+= 2 * sizeof(struct track
);
850 off
+= kasan_metadata_size(s
);
852 if (off
!= size_from_object(s
))
853 /* Beginning of the filler is the free pointer */
854 print_section(KERN_ERR
, "Padding ", p
+ off
,
855 size_from_object(s
) - off
);
860 void object_err(struct kmem_cache
*s
, struct page
*page
,
861 u8
*object
, char *reason
)
863 if (slab_add_kunit_errors())
866 slab_bug(s
, "%s", reason
);
867 print_trailer(s
, page
, object
);
868 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
871 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct page
*page
,
872 const char *fmt
, ...)
877 if (slab_add_kunit_errors())
881 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
883 slab_bug(s
, "%s", buf
);
884 print_page_info(page
);
886 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
889 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
891 u8
*p
= kasan_reset_tag(object
);
893 if (s
->flags
& SLAB_RED_ZONE
)
894 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
896 if (s
->flags
& __OBJECT_POISON
) {
897 memset(p
, POISON_FREE
, s
->object_size
- 1);
898 p
[s
->object_size
- 1] = POISON_END
;
901 if (s
->flags
& SLAB_RED_ZONE
)
902 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
905 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
906 void *from
, void *to
)
908 slab_fix(s
, "Restoring %s 0x%p-0x%p=0x%x", message
, from
, to
- 1, data
);
909 memset(from
, data
, to
- from
);
912 static int check_bytes_and_report(struct kmem_cache
*s
, struct page
*page
,
913 u8
*object
, char *what
,
914 u8
*start
, unsigned int value
, unsigned int bytes
)
918 u8
*addr
= page_address(page
);
920 metadata_access_enable();
921 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
922 metadata_access_disable();
927 while (end
> fault
&& end
[-1] == value
)
930 if (slab_add_kunit_errors())
933 slab_bug(s
, "%s overwritten", what
);
934 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
935 fault
, end
- 1, fault
- addr
,
937 print_trailer(s
, page
, object
);
938 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
941 restore_bytes(s
, what
, value
, fault
, end
);
949 * Bytes of the object to be managed.
950 * If the freepointer may overlay the object then the free
951 * pointer is at the middle of the object.
953 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
956 * object + s->object_size
957 * Padding to reach word boundary. This is also used for Redzoning.
958 * Padding is extended by another word if Redzoning is enabled and
959 * object_size == inuse.
961 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
962 * 0xcc (RED_ACTIVE) for objects in use.
965 * Meta data starts here.
967 * A. Free pointer (if we cannot overwrite object on free)
968 * B. Tracking data for SLAB_STORE_USER
969 * C. Padding to reach required alignment boundary or at minimum
970 * one word if debugging is on to be able to detect writes
971 * before the word boundary.
973 * Padding is done using 0x5a (POISON_INUSE)
976 * Nothing is used beyond s->size.
978 * If slabcaches are merged then the object_size and inuse boundaries are mostly
979 * ignored. And therefore no slab options that rely on these boundaries
980 * may be used with merged slabcaches.
983 static int check_pad_bytes(struct kmem_cache
*s
, struct page
*page
, u8
*p
)
985 unsigned long off
= get_info_end(s
); /* The end of info */
987 if (s
->flags
& SLAB_STORE_USER
)
988 /* We also have user information there */
989 off
+= 2 * sizeof(struct track
);
991 off
+= kasan_metadata_size(s
);
993 if (size_from_object(s
) == off
)
996 return check_bytes_and_report(s
, page
, p
, "Object padding",
997 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
1000 /* Check the pad bytes at the end of a slab page */
1001 static int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1010 if (!(s
->flags
& SLAB_POISON
))
1013 start
= page_address(page
);
1014 length
= page_size(page
);
1015 end
= start
+ length
;
1016 remainder
= length
% s
->size
;
1020 pad
= end
- remainder
;
1021 metadata_access_enable();
1022 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
1023 metadata_access_disable();
1026 while (end
> fault
&& end
[-1] == POISON_INUSE
)
1029 slab_err(s
, page
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1030 fault
, end
- 1, fault
- start
);
1031 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
1033 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
1037 static int check_object(struct kmem_cache
*s
, struct page
*page
,
1038 void *object
, u8 val
)
1041 u8
*endobject
= object
+ s
->object_size
;
1043 if (s
->flags
& SLAB_RED_ZONE
) {
1044 if (!check_bytes_and_report(s
, page
, object
, "Left Redzone",
1045 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
1048 if (!check_bytes_and_report(s
, page
, object
, "Right Redzone",
1049 endobject
, val
, s
->inuse
- s
->object_size
))
1052 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
1053 check_bytes_and_report(s
, page
, p
, "Alignment padding",
1054 endobject
, POISON_INUSE
,
1055 s
->inuse
- s
->object_size
);
1059 if (s
->flags
& SLAB_POISON
) {
1060 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
1061 (!check_bytes_and_report(s
, page
, p
, "Poison", p
,
1062 POISON_FREE
, s
->object_size
- 1) ||
1063 !check_bytes_and_report(s
, page
, p
, "End Poison",
1064 p
+ s
->object_size
- 1, POISON_END
, 1)))
1067 * check_pad_bytes cleans up on its own.
1069 check_pad_bytes(s
, page
, p
);
1072 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
1074 * Object and freepointer overlap. Cannot check
1075 * freepointer while object is allocated.
1079 /* Check free pointer validity */
1080 if (!check_valid_pointer(s
, page
, get_freepointer(s
, p
))) {
1081 object_err(s
, page
, p
, "Freepointer corrupt");
1083 * No choice but to zap it and thus lose the remainder
1084 * of the free objects in this slab. May cause
1085 * another error because the object count is now wrong.
1087 set_freepointer(s
, p
, NULL
);
1093 static int check_slab(struct kmem_cache
*s
, struct page
*page
)
1097 if (!PageSlab(page
)) {
1098 slab_err(s
, page
, "Not a valid slab page");
1102 maxobj
= order_objects(compound_order(page
), s
->size
);
1103 if (page
->objects
> maxobj
) {
1104 slab_err(s
, page
, "objects %u > max %u",
1105 page
->objects
, maxobj
);
1108 if (page
->inuse
> page
->objects
) {
1109 slab_err(s
, page
, "inuse %u > max %u",
1110 page
->inuse
, page
->objects
);
1113 /* Slab_pad_check fixes things up after itself */
1114 slab_pad_check(s
, page
);
1119 * Determine if a certain object on a page is on the freelist. Must hold the
1120 * slab lock to guarantee that the chains are in a consistent state.
1122 static int on_freelist(struct kmem_cache
*s
, struct page
*page
, void *search
)
1126 void *object
= NULL
;
1129 fp
= page
->freelist
;
1130 while (fp
&& nr
<= page
->objects
) {
1133 if (!check_valid_pointer(s
, page
, fp
)) {
1135 object_err(s
, page
, object
,
1136 "Freechain corrupt");
1137 set_freepointer(s
, object
, NULL
);
1139 slab_err(s
, page
, "Freepointer corrupt");
1140 page
->freelist
= NULL
;
1141 page
->inuse
= page
->objects
;
1142 slab_fix(s
, "Freelist cleared");
1148 fp
= get_freepointer(s
, object
);
1152 max_objects
= order_objects(compound_order(page
), s
->size
);
1153 if (max_objects
> MAX_OBJS_PER_PAGE
)
1154 max_objects
= MAX_OBJS_PER_PAGE
;
1156 if (page
->objects
!= max_objects
) {
1157 slab_err(s
, page
, "Wrong number of objects. Found %d but should be %d",
1158 page
->objects
, max_objects
);
1159 page
->objects
= max_objects
;
1160 slab_fix(s
, "Number of objects adjusted");
1162 if (page
->inuse
!= page
->objects
- nr
) {
1163 slab_err(s
, page
, "Wrong object count. Counter is %d but counted were %d",
1164 page
->inuse
, page
->objects
- nr
);
1165 page
->inuse
= page
->objects
- nr
;
1166 slab_fix(s
, "Object count adjusted");
1168 return search
== NULL
;
1171 static void trace(struct kmem_cache
*s
, struct page
*page
, void *object
,
1174 if (s
->flags
& SLAB_TRACE
) {
1175 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1177 alloc
? "alloc" : "free",
1178 object
, page
->inuse
,
1182 print_section(KERN_INFO
, "Object ", (void *)object
,
1190 * Tracking of fully allocated slabs for debugging purposes.
1192 static void add_full(struct kmem_cache
*s
,
1193 struct kmem_cache_node
*n
, struct page
*page
)
1195 if (!(s
->flags
& SLAB_STORE_USER
))
1198 lockdep_assert_held(&n
->list_lock
);
1199 list_add(&page
->slab_list
, &n
->full
);
1202 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct page
*page
)
1204 if (!(s
->flags
& SLAB_STORE_USER
))
1207 lockdep_assert_held(&n
->list_lock
);
1208 list_del(&page
->slab_list
);
1211 /* Tracking of the number of slabs for debugging purposes */
1212 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1214 struct kmem_cache_node
*n
= get_node(s
, node
);
1216 return atomic_long_read(&n
->nr_slabs
);
1219 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1221 return atomic_long_read(&n
->nr_slabs
);
1224 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1226 struct kmem_cache_node
*n
= get_node(s
, node
);
1229 * May be called early in order to allocate a slab for the
1230 * kmem_cache_node structure. Solve the chicken-egg
1231 * dilemma by deferring the increment of the count during
1232 * bootstrap (see early_kmem_cache_node_alloc).
1235 atomic_long_inc(&n
->nr_slabs
);
1236 atomic_long_add(objects
, &n
->total_objects
);
1239 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1241 struct kmem_cache_node
*n
= get_node(s
, node
);
1243 atomic_long_dec(&n
->nr_slabs
);
1244 atomic_long_sub(objects
, &n
->total_objects
);
1247 /* Object debug checks for alloc/free paths */
1248 static void setup_object_debug(struct kmem_cache
*s
, struct page
*page
,
1251 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1254 init_object(s
, object
, SLUB_RED_INACTIVE
);
1255 init_tracking(s
, object
);
1259 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
)
1261 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1264 metadata_access_enable();
1265 memset(kasan_reset_tag(addr
), POISON_INUSE
, page_size(page
));
1266 metadata_access_disable();
1269 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1270 struct page
*page
, void *object
)
1272 if (!check_slab(s
, page
))
1275 if (!check_valid_pointer(s
, page
, object
)) {
1276 object_err(s
, page
, object
, "Freelist Pointer check fails");
1280 if (!check_object(s
, page
, object
, SLUB_RED_INACTIVE
))
1286 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1288 void *object
, unsigned long addr
)
1290 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1291 if (!alloc_consistency_checks(s
, page
, object
))
1295 /* Success perform special debug activities for allocs */
1296 if (s
->flags
& SLAB_STORE_USER
)
1297 set_track(s
, object
, TRACK_ALLOC
, addr
);
1298 trace(s
, page
, object
, 1);
1299 init_object(s
, object
, SLUB_RED_ACTIVE
);
1303 if (PageSlab(page
)) {
1305 * If this is a slab page then lets do the best we can
1306 * to avoid issues in the future. Marking all objects
1307 * as used avoids touching the remaining objects.
1309 slab_fix(s
, "Marking all objects used");
1310 page
->inuse
= page
->objects
;
1311 page
->freelist
= NULL
;
1316 static inline int free_consistency_checks(struct kmem_cache
*s
,
1317 struct page
*page
, void *object
, unsigned long addr
)
1319 if (!check_valid_pointer(s
, page
, object
)) {
1320 slab_err(s
, page
, "Invalid object pointer 0x%p", object
);
1324 if (on_freelist(s
, page
, object
)) {
1325 object_err(s
, page
, object
, "Object already free");
1329 if (!check_object(s
, page
, object
, SLUB_RED_ACTIVE
))
1332 if (unlikely(s
!= page
->slab_cache
)) {
1333 if (!PageSlab(page
)) {
1334 slab_err(s
, page
, "Attempt to free object(0x%p) outside of slab",
1336 } else if (!page
->slab_cache
) {
1337 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1341 object_err(s
, page
, object
,
1342 "page slab pointer corrupt.");
1348 /* Supports checking bulk free of a constructed freelist */
1349 static noinline
int free_debug_processing(
1350 struct kmem_cache
*s
, struct page
*page
,
1351 void *head
, void *tail
, int bulk_cnt
,
1354 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
1355 void *object
= head
;
1357 unsigned long flags
, flags2
;
1360 spin_lock_irqsave(&n
->list_lock
, flags
);
1361 slab_lock(page
, &flags2
);
1363 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1364 if (!check_slab(s
, page
))
1371 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1372 if (!free_consistency_checks(s
, page
, object
, addr
))
1376 if (s
->flags
& SLAB_STORE_USER
)
1377 set_track(s
, object
, TRACK_FREE
, addr
);
1378 trace(s
, page
, object
, 0);
1379 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1380 init_object(s
, object
, SLUB_RED_INACTIVE
);
1382 /* Reached end of constructed freelist yet? */
1383 if (object
!= tail
) {
1384 object
= get_freepointer(s
, object
);
1390 if (cnt
!= bulk_cnt
)
1391 slab_err(s
, page
, "Bulk freelist count(%d) invalid(%d)\n",
1394 slab_unlock(page
, &flags2
);
1395 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1397 slab_fix(s
, "Object at 0x%p not freed", object
);
1402 * Parse a block of slub_debug options. Blocks are delimited by ';'
1404 * @str: start of block
1405 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1406 * @slabs: return start of list of slabs, or NULL when there's no list
1407 * @init: assume this is initial parsing and not per-kmem-create parsing
1409 * returns the start of next block if there's any, or NULL
1412 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1414 bool higher_order_disable
= false;
1416 /* Skip any completely empty blocks */
1417 while (*str
&& *str
== ';')
1422 * No options but restriction on slabs. This means full
1423 * debugging for slabs matching a pattern.
1425 *flags
= DEBUG_DEFAULT_FLAGS
;
1430 /* Determine which debug features should be switched on */
1431 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1432 switch (tolower(*str
)) {
1437 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1440 *flags
|= SLAB_RED_ZONE
;
1443 *flags
|= SLAB_POISON
;
1446 *flags
|= SLAB_STORE_USER
;
1449 *flags
|= SLAB_TRACE
;
1452 *flags
|= SLAB_FAILSLAB
;
1456 * Avoid enabling debugging on caches if its minimum
1457 * order would increase as a result.
1459 higher_order_disable
= true;
1463 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1472 /* Skip over the slab list */
1473 while (*str
&& *str
!= ';')
1476 /* Skip any completely empty blocks */
1477 while (*str
&& *str
== ';')
1480 if (init
&& higher_order_disable
)
1481 disable_higher_order_debug
= 1;
1489 static int __init
setup_slub_debug(char *str
)
1492 slab_flags_t global_flags
;
1495 bool global_slub_debug_changed
= false;
1496 bool slab_list_specified
= false;
1498 global_flags
= DEBUG_DEFAULT_FLAGS
;
1499 if (*str
++ != '=' || !*str
)
1501 * No options specified. Switch on full debugging.
1507 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1510 global_flags
= flags
;
1511 global_slub_debug_changed
= true;
1513 slab_list_specified
= true;
1518 * For backwards compatibility, a single list of flags with list of
1519 * slabs means debugging is only changed for those slabs, so the global
1520 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1521 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1522 * long as there is no option specifying flags without a slab list.
1524 if (slab_list_specified
) {
1525 if (!global_slub_debug_changed
)
1526 global_flags
= slub_debug
;
1527 slub_debug_string
= saved_str
;
1530 slub_debug
= global_flags
;
1531 if (slub_debug
!= 0 || slub_debug_string
)
1532 static_branch_enable(&slub_debug_enabled
);
1534 static_branch_disable(&slub_debug_enabled
);
1535 if ((static_branch_unlikely(&init_on_alloc
) ||
1536 static_branch_unlikely(&init_on_free
)) &&
1537 (slub_debug
& SLAB_POISON
))
1538 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1542 __setup("slub_debug", setup_slub_debug
);
1545 * kmem_cache_flags - apply debugging options to the cache
1546 * @object_size: the size of an object without meta data
1547 * @flags: flags to set
1548 * @name: name of the cache
1550 * Debug option(s) are applied to @flags. In addition to the debug
1551 * option(s), if a slab name (or multiple) is specified i.e.
1552 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1553 * then only the select slabs will receive the debug option(s).
1555 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1556 slab_flags_t flags
, const char *name
)
1561 slab_flags_t block_flags
;
1562 slab_flags_t slub_debug_local
= slub_debug
;
1565 * If the slab cache is for debugging (e.g. kmemleak) then
1566 * don't store user (stack trace) information by default,
1567 * but let the user enable it via the command line below.
1569 if (flags
& SLAB_NOLEAKTRACE
)
1570 slub_debug_local
&= ~SLAB_STORE_USER
;
1573 next_block
= slub_debug_string
;
1574 /* Go through all blocks of debug options, see if any matches our slab's name */
1575 while (next_block
) {
1576 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1579 /* Found a block that has a slab list, search it */
1584 end
= strchrnul(iter
, ',');
1585 if (next_block
&& next_block
< end
)
1586 end
= next_block
- 1;
1588 glob
= strnchr(iter
, end
- iter
, '*');
1590 cmplen
= glob
- iter
;
1592 cmplen
= max_t(size_t, len
, (end
- iter
));
1594 if (!strncmp(name
, iter
, cmplen
)) {
1595 flags
|= block_flags
;
1599 if (!*end
|| *end
== ';')
1605 return flags
| slub_debug_local
;
1607 #else /* !CONFIG_SLUB_DEBUG */
1608 static inline void setup_object_debug(struct kmem_cache
*s
,
1609 struct page
*page
, void *object
) {}
1611 void setup_page_debug(struct kmem_cache
*s
, struct page
*page
, void *addr
) {}
1613 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1614 struct page
*page
, void *object
, unsigned long addr
) { return 0; }
1616 static inline int free_debug_processing(
1617 struct kmem_cache
*s
, struct page
*page
,
1618 void *head
, void *tail
, int bulk_cnt
,
1619 unsigned long addr
) { return 0; }
1621 static inline int slab_pad_check(struct kmem_cache
*s
, struct page
*page
)
1623 static inline int check_object(struct kmem_cache
*s
, struct page
*page
,
1624 void *object
, u8 val
) { return 1; }
1625 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1626 struct page
*page
) {}
1627 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1628 struct page
*page
) {}
1629 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1630 slab_flags_t flags
, const char *name
)
1634 #define slub_debug 0
1636 #define disable_higher_order_debug 0
1638 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1640 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1642 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1644 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1647 static bool freelist_corrupted(struct kmem_cache
*s
, struct page
*page
,
1648 void **freelist
, void *nextfree
)
1652 #endif /* CONFIG_SLUB_DEBUG */
1655 * Hooks for other subsystems that check memory allocations. In a typical
1656 * production configuration these hooks all should produce no code at all.
1658 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1660 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1661 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1662 kmemleak_alloc(ptr
, size
, 1, flags
);
1666 static __always_inline
void kfree_hook(void *x
)
1669 kasan_kfree_large(x
);
1672 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
,
1675 kmemleak_free_recursive(x
, s
->flags
);
1677 debug_check_no_locks_freed(x
, s
->object_size
);
1679 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1680 debug_check_no_obj_freed(x
, s
->object_size
);
1682 /* Use KCSAN to help debug racy use-after-free. */
1683 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1684 __kcsan_check_access(x
, s
->object_size
,
1685 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1688 * As memory initialization might be integrated into KASAN,
1689 * kasan_slab_free and initialization memset's must be
1690 * kept together to avoid discrepancies in behavior.
1692 * The initialization memset's clear the object and the metadata,
1693 * but don't touch the SLAB redzone.
1698 if (!kasan_has_integrated_init())
1699 memset(kasan_reset_tag(x
), 0, s
->object_size
);
1700 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
: 0;
1701 memset((char *)kasan_reset_tag(x
) + s
->inuse
, 0,
1702 s
->size
- s
->inuse
- rsize
);
1704 /* KASAN might put x into memory quarantine, delaying its reuse. */
1705 return kasan_slab_free(s
, x
, init
);
1708 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1709 void **head
, void **tail
,
1715 void *old_tail
= *tail
? *tail
: *head
;
1717 if (is_kfence_address(next
)) {
1718 slab_free_hook(s
, next
, false);
1722 /* Head and tail of the reconstructed freelist */
1728 next
= get_freepointer(s
, object
);
1730 /* If object's reuse doesn't have to be delayed */
1731 if (!slab_free_hook(s
, object
, slab_want_init_on_free(s
))) {
1732 /* Move object to the new freelist */
1733 set_freepointer(s
, object
, *head
);
1739 * Adjust the reconstructed freelist depth
1740 * accordingly if object's reuse is delayed.
1744 } while (object
!= old_tail
);
1749 return *head
!= NULL
;
1752 static void *setup_object(struct kmem_cache
*s
, struct page
*page
,
1755 setup_object_debug(s
, page
, object
);
1756 object
= kasan_init_slab_obj(s
, object
);
1757 if (unlikely(s
->ctor
)) {
1758 kasan_unpoison_object_data(s
, object
);
1760 kasan_poison_object_data(s
, object
);
1766 * Slab allocation and freeing
1768 static inline struct page
*alloc_slab_page(struct kmem_cache
*s
,
1769 gfp_t flags
, int node
, struct kmem_cache_order_objects oo
)
1772 unsigned int order
= oo_order(oo
);
1774 if (node
== NUMA_NO_NODE
)
1775 page
= alloc_pages(flags
, order
);
1777 page
= __alloc_pages_node(node
, flags
, order
);
1782 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1783 /* Pre-initialize the random sequence cache */
1784 static int init_cache_random_seq(struct kmem_cache
*s
)
1786 unsigned int count
= oo_objects(s
->oo
);
1789 /* Bailout if already initialised */
1793 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1795 pr_err("SLUB: Unable to initialize free list for %s\n",
1800 /* Transform to an offset on the set of pages */
1801 if (s
->random_seq
) {
1804 for (i
= 0; i
< count
; i
++)
1805 s
->random_seq
[i
] *= s
->size
;
1810 /* Initialize each random sequence freelist per cache */
1811 static void __init
init_freelist_randomization(void)
1813 struct kmem_cache
*s
;
1815 mutex_lock(&slab_mutex
);
1817 list_for_each_entry(s
, &slab_caches
, list
)
1818 init_cache_random_seq(s
);
1820 mutex_unlock(&slab_mutex
);
1823 /* Get the next entry on the pre-computed freelist randomized */
1824 static void *next_freelist_entry(struct kmem_cache
*s
, struct page
*page
,
1825 unsigned long *pos
, void *start
,
1826 unsigned long page_limit
,
1827 unsigned long freelist_count
)
1832 * If the target page allocation failed, the number of objects on the
1833 * page might be smaller than the usual size defined by the cache.
1836 idx
= s
->random_seq
[*pos
];
1838 if (*pos
>= freelist_count
)
1840 } while (unlikely(idx
>= page_limit
));
1842 return (char *)start
+ idx
;
1845 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1846 static bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1851 unsigned long idx
, pos
, page_limit
, freelist_count
;
1853 if (page
->objects
< 2 || !s
->random_seq
)
1856 freelist_count
= oo_objects(s
->oo
);
1857 pos
= get_random_int() % freelist_count
;
1859 page_limit
= page
->objects
* s
->size
;
1860 start
= fixup_red_left(s
, page_address(page
));
1862 /* First entry is used as the base of the freelist */
1863 cur
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1865 cur
= setup_object(s
, page
, cur
);
1866 page
->freelist
= cur
;
1868 for (idx
= 1; idx
< page
->objects
; idx
++) {
1869 next
= next_freelist_entry(s
, page
, &pos
, start
, page_limit
,
1871 next
= setup_object(s
, page
, next
);
1872 set_freepointer(s
, cur
, next
);
1875 set_freepointer(s
, cur
, NULL
);
1880 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1884 static inline void init_freelist_randomization(void) { }
1885 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct page
*page
)
1889 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1891 static struct page
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1894 struct kmem_cache_order_objects oo
= s
->oo
;
1896 void *start
, *p
, *next
;
1900 flags
&= gfp_allowed_mask
;
1902 flags
|= s
->allocflags
;
1905 * Let the initial higher-order allocation fail under memory pressure
1906 * so we fall-back to the minimum order allocation.
1908 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1909 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1910 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
1912 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1913 if (unlikely(!page
)) {
1917 * Allocation may have failed due to fragmentation.
1918 * Try a lower order alloc if possible
1920 page
= alloc_slab_page(s
, alloc_gfp
, node
, oo
);
1921 if (unlikely(!page
))
1923 stat(s
, ORDER_FALLBACK
);
1926 page
->objects
= oo_objects(oo
);
1928 account_slab_page(page
, oo_order(oo
), s
, flags
);
1930 page
->slab_cache
= s
;
1931 __SetPageSlab(page
);
1932 if (page_is_pfmemalloc(page
))
1933 SetPageSlabPfmemalloc(page
);
1935 kasan_poison_slab(page
);
1937 start
= page_address(page
);
1939 setup_page_debug(s
, page
, start
);
1941 shuffle
= shuffle_freelist(s
, page
);
1944 start
= fixup_red_left(s
, start
);
1945 start
= setup_object(s
, page
, start
);
1946 page
->freelist
= start
;
1947 for (idx
= 0, p
= start
; idx
< page
->objects
- 1; idx
++) {
1949 next
= setup_object(s
, page
, next
);
1950 set_freepointer(s
, p
, next
);
1953 set_freepointer(s
, p
, NULL
);
1956 page
->inuse
= page
->objects
;
1963 inc_slabs_node(s
, page_to_nid(page
), page
->objects
);
1968 static struct page
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1970 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
1971 flags
= kmalloc_fix_flags(flags
);
1973 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
1975 return allocate_slab(s
,
1976 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
1979 static void __free_slab(struct kmem_cache
*s
, struct page
*page
)
1981 int order
= compound_order(page
);
1982 int pages
= 1 << order
;
1984 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
1987 slab_pad_check(s
, page
);
1988 for_each_object(p
, s
, page_address(page
),
1990 check_object(s
, page
, p
, SLUB_RED_INACTIVE
);
1993 __ClearPageSlabPfmemalloc(page
);
1994 __ClearPageSlab(page
);
1995 /* In union with page->mapping where page allocator expects NULL */
1996 page
->slab_cache
= NULL
;
1997 if (current
->reclaim_state
)
1998 current
->reclaim_state
->reclaimed_slab
+= pages
;
1999 unaccount_slab_page(page
, order
, s
);
2000 __free_pages(page
, order
);
2003 static void rcu_free_slab(struct rcu_head
*h
)
2005 struct page
*page
= container_of(h
, struct page
, rcu_head
);
2007 __free_slab(page
->slab_cache
, page
);
2010 static void free_slab(struct kmem_cache
*s
, struct page
*page
)
2012 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
2013 call_rcu(&page
->rcu_head
, rcu_free_slab
);
2015 __free_slab(s
, page
);
2018 static void discard_slab(struct kmem_cache
*s
, struct page
*page
)
2020 dec_slabs_node(s
, page_to_nid(page
), page
->objects
);
2025 * Management of partially allocated slabs.
2028 __add_partial(struct kmem_cache_node
*n
, struct page
*page
, int tail
)
2031 if (tail
== DEACTIVATE_TO_TAIL
)
2032 list_add_tail(&page
->slab_list
, &n
->partial
);
2034 list_add(&page
->slab_list
, &n
->partial
);
2037 static inline void add_partial(struct kmem_cache_node
*n
,
2038 struct page
*page
, int tail
)
2040 lockdep_assert_held(&n
->list_lock
);
2041 __add_partial(n
, page
, tail
);
2044 static inline void remove_partial(struct kmem_cache_node
*n
,
2047 lockdep_assert_held(&n
->list_lock
);
2048 list_del(&page
->slab_list
);
2053 * Remove slab from the partial list, freeze it and
2054 * return the pointer to the freelist.
2056 * Returns a list of objects or NULL if it fails.
2058 static inline void *acquire_slab(struct kmem_cache
*s
,
2059 struct kmem_cache_node
*n
, struct page
*page
,
2060 int mode
, int *objects
)
2063 unsigned long counters
;
2066 lockdep_assert_held(&n
->list_lock
);
2069 * Zap the freelist and set the frozen bit.
2070 * The old freelist is the list of objects for the
2071 * per cpu allocation list.
2073 freelist
= page
->freelist
;
2074 counters
= page
->counters
;
2075 new.counters
= counters
;
2076 *objects
= new.objects
- new.inuse
;
2078 new.inuse
= page
->objects
;
2079 new.freelist
= NULL
;
2081 new.freelist
= freelist
;
2084 VM_BUG_ON(new.frozen
);
2087 if (!__cmpxchg_double_slab(s
, page
,
2089 new.freelist
, new.counters
,
2093 remove_partial(n
, page
);
2098 #ifdef CONFIG_SLUB_CPU_PARTIAL
2099 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
);
2101 static inline void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
,
2104 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
);
2107 * Try to allocate a partial slab from a specific node.
2109 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
2110 struct page
**ret_page
, gfp_t gfpflags
)
2112 struct page
*page
, *page2
;
2113 void *object
= NULL
;
2114 unsigned int available
= 0;
2115 unsigned long flags
;
2119 * Racy check. If we mistakenly see no partial slabs then we
2120 * just allocate an empty slab. If we mistakenly try to get a
2121 * partial slab and there is none available then get_partial()
2124 if (!n
|| !n
->nr_partial
)
2127 spin_lock_irqsave(&n
->list_lock
, flags
);
2128 list_for_each_entry_safe(page
, page2
, &n
->partial
, slab_list
) {
2131 if (!pfmemalloc_match(page
, gfpflags
))
2134 t
= acquire_slab(s
, n
, page
, object
== NULL
, &objects
);
2138 available
+= objects
;
2141 stat(s
, ALLOC_FROM_PARTIAL
);
2144 put_cpu_partial(s
, page
, 0);
2145 stat(s
, CPU_PARTIAL_NODE
);
2147 if (!kmem_cache_has_cpu_partial(s
)
2148 || available
> slub_cpu_partial(s
) / 2)
2152 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2157 * Get a page from somewhere. Search in increasing NUMA distances.
2159 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2160 struct page
**ret_page
)
2163 struct zonelist
*zonelist
;
2166 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2168 unsigned int cpuset_mems_cookie
;
2171 * The defrag ratio allows a configuration of the tradeoffs between
2172 * inter node defragmentation and node local allocations. A lower
2173 * defrag_ratio increases the tendency to do local allocations
2174 * instead of attempting to obtain partial slabs from other nodes.
2176 * If the defrag_ratio is set to 0 then kmalloc() always
2177 * returns node local objects. If the ratio is higher then kmalloc()
2178 * may return off node objects because partial slabs are obtained
2179 * from other nodes and filled up.
2181 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2182 * (which makes defrag_ratio = 1000) then every (well almost)
2183 * allocation will first attempt to defrag slab caches on other nodes.
2184 * This means scanning over all nodes to look for partial slabs which
2185 * may be expensive if we do it every time we are trying to find a slab
2186 * with available objects.
2188 if (!s
->remote_node_defrag_ratio
||
2189 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2193 cpuset_mems_cookie
= read_mems_allowed_begin();
2194 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2195 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2196 struct kmem_cache_node
*n
;
2198 n
= get_node(s
, zone_to_nid(zone
));
2200 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2201 n
->nr_partial
> s
->min_partial
) {
2202 object
= get_partial_node(s
, n
, ret_page
, flags
);
2205 * Don't check read_mems_allowed_retry()
2206 * here - if mems_allowed was updated in
2207 * parallel, that was a harmless race
2208 * between allocation and the cpuset
2215 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2216 #endif /* CONFIG_NUMA */
2221 * Get a partial page, lock it and return it.
2223 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2224 struct page
**ret_page
)
2227 int searchnode
= node
;
2229 if (node
== NUMA_NO_NODE
)
2230 searchnode
= numa_mem_id();
2232 object
= get_partial_node(s
, get_node(s
, searchnode
), ret_page
, flags
);
2233 if (object
|| node
!= NUMA_NO_NODE
)
2236 return get_any_partial(s
, flags
, ret_page
);
2239 #ifdef CONFIG_PREEMPTION
2241 * Calculate the next globally unique transaction for disambiguation
2242 * during cmpxchg. The transactions start with the cpu number and are then
2243 * incremented by CONFIG_NR_CPUS.
2245 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2248 * No preemption supported therefore also no need to check for
2254 static inline unsigned long next_tid(unsigned long tid
)
2256 return tid
+ TID_STEP
;
2259 #ifdef SLUB_DEBUG_CMPXCHG
2260 static inline unsigned int tid_to_cpu(unsigned long tid
)
2262 return tid
% TID_STEP
;
2265 static inline unsigned long tid_to_event(unsigned long tid
)
2267 return tid
/ TID_STEP
;
2271 static inline unsigned int init_tid(int cpu
)
2276 static inline void note_cmpxchg_failure(const char *n
,
2277 const struct kmem_cache
*s
, unsigned long tid
)
2279 #ifdef SLUB_DEBUG_CMPXCHG
2280 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2282 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2284 #ifdef CONFIG_PREEMPTION
2285 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2286 pr_warn("due to cpu change %d -> %d\n",
2287 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2290 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2291 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2292 tid_to_event(tid
), tid_to_event(actual_tid
));
2294 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2295 actual_tid
, tid
, next_tid(tid
));
2297 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2300 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2303 struct kmem_cache_cpu
*c
;
2305 for_each_possible_cpu(cpu
) {
2306 c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2307 local_lock_init(&c
->lock
);
2308 c
->tid
= init_tid(cpu
);
2313 * Finishes removing the cpu slab. Merges cpu's freelist with page's freelist,
2314 * unfreezes the slabs and puts it on the proper list.
2315 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2318 static void deactivate_slab(struct kmem_cache
*s
, struct page
*page
,
2321 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
};
2322 struct kmem_cache_node
*n
= get_node(s
, page_to_nid(page
));
2323 int lock
= 0, free_delta
= 0;
2324 enum slab_modes l
= M_NONE
, m
= M_NONE
;
2325 void *nextfree
, *freelist_iter
, *freelist_tail
;
2326 int tail
= DEACTIVATE_TO_HEAD
;
2327 unsigned long flags
= 0;
2331 if (page
->freelist
) {
2332 stat(s
, DEACTIVATE_REMOTE_FREES
);
2333 tail
= DEACTIVATE_TO_TAIL
;
2337 * Stage one: Count the objects on cpu's freelist as free_delta and
2338 * remember the last object in freelist_tail for later splicing.
2340 freelist_tail
= NULL
;
2341 freelist_iter
= freelist
;
2342 while (freelist_iter
) {
2343 nextfree
= get_freepointer(s
, freelist_iter
);
2346 * If 'nextfree' is invalid, it is possible that the object at
2347 * 'freelist_iter' is already corrupted. So isolate all objects
2348 * starting at 'freelist_iter' by skipping them.
2350 if (freelist_corrupted(s
, page
, &freelist_iter
, nextfree
))
2353 freelist_tail
= freelist_iter
;
2356 freelist_iter
= nextfree
;
2360 * Stage two: Unfreeze the page while splicing the per-cpu
2361 * freelist to the head of page's freelist.
2363 * Ensure that the page is unfrozen while the list presence
2364 * reflects the actual number of objects during unfreeze.
2366 * We setup the list membership and then perform a cmpxchg
2367 * with the count. If there is a mismatch then the page
2368 * is not unfrozen but the page is on the wrong list.
2370 * Then we restart the process which may have to remove
2371 * the page from the list that we just put it on again
2372 * because the number of objects in the slab may have
2377 old
.freelist
= READ_ONCE(page
->freelist
);
2378 old
.counters
= READ_ONCE(page
->counters
);
2379 VM_BUG_ON(!old
.frozen
);
2381 /* Determine target state of the slab */
2382 new.counters
= old
.counters
;
2383 if (freelist_tail
) {
2384 new.inuse
-= free_delta
;
2385 set_freepointer(s
, freelist_tail
, old
.freelist
);
2386 new.freelist
= freelist
;
2388 new.freelist
= old
.freelist
;
2392 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
)
2394 else if (new.freelist
) {
2399 * Taking the spinlock removes the possibility
2400 * that acquire_slab() will see a slab page that
2403 spin_lock_irqsave(&n
->list_lock
, flags
);
2407 if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
) && !lock
) {
2410 * This also ensures that the scanning of full
2411 * slabs from diagnostic functions will not see
2414 spin_lock_irqsave(&n
->list_lock
, flags
);
2420 remove_partial(n
, page
);
2421 else if (l
== M_FULL
)
2422 remove_full(s
, n
, page
);
2425 add_partial(n
, page
, tail
);
2426 else if (m
== M_FULL
)
2427 add_full(s
, n
, page
);
2431 if (!cmpxchg_double_slab(s
, page
,
2432 old
.freelist
, old
.counters
,
2433 new.freelist
, new.counters
,
2438 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2442 else if (m
== M_FULL
)
2443 stat(s
, DEACTIVATE_FULL
);
2444 else if (m
== M_FREE
) {
2445 stat(s
, DEACTIVATE_EMPTY
);
2446 discard_slab(s
, page
);
2451 #ifdef CONFIG_SLUB_CPU_PARTIAL
2452 static void __unfreeze_partials(struct kmem_cache
*s
, struct page
*partial_page
)
2454 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2455 struct page
*page
, *discard_page
= NULL
;
2456 unsigned long flags
= 0;
2458 while (partial_page
) {
2462 page
= partial_page
;
2463 partial_page
= page
->next
;
2465 n2
= get_node(s
, page_to_nid(page
));
2468 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2471 spin_lock_irqsave(&n
->list_lock
, flags
);
2476 old
.freelist
= page
->freelist
;
2477 old
.counters
= page
->counters
;
2478 VM_BUG_ON(!old
.frozen
);
2480 new.counters
= old
.counters
;
2481 new.freelist
= old
.freelist
;
2485 } while (!__cmpxchg_double_slab(s
, page
,
2486 old
.freelist
, old
.counters
,
2487 new.freelist
, new.counters
,
2488 "unfreezing slab"));
2490 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2491 page
->next
= discard_page
;
2492 discard_page
= page
;
2494 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
2495 stat(s
, FREE_ADD_PARTIAL
);
2500 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2502 while (discard_page
) {
2503 page
= discard_page
;
2504 discard_page
= discard_page
->next
;
2506 stat(s
, DEACTIVATE_EMPTY
);
2507 discard_slab(s
, page
);
2513 * Unfreeze all the cpu partial slabs.
2515 static void unfreeze_partials(struct kmem_cache
*s
)
2517 struct page
*partial_page
;
2518 unsigned long flags
;
2520 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2521 partial_page
= this_cpu_read(s
->cpu_slab
->partial
);
2522 this_cpu_write(s
->cpu_slab
->partial
, NULL
);
2523 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2526 __unfreeze_partials(s
, partial_page
);
2529 static void unfreeze_partials_cpu(struct kmem_cache
*s
,
2530 struct kmem_cache_cpu
*c
)
2532 struct page
*partial_page
;
2534 partial_page
= slub_percpu_partial(c
);
2538 __unfreeze_partials(s
, partial_page
);
2542 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2543 * partial page slot if available.
2545 * If we did not find a slot then simply move all the partials to the
2546 * per node partial list.
2548 static void put_cpu_partial(struct kmem_cache
*s
, struct page
*page
, int drain
)
2550 struct page
*oldpage
;
2551 struct page
*page_to_unfreeze
= NULL
;
2552 unsigned long flags
;
2556 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2558 oldpage
= this_cpu_read(s
->cpu_slab
->partial
);
2561 if (drain
&& oldpage
->pobjects
> slub_cpu_partial(s
)) {
2563 * Partial array is full. Move the existing set to the
2564 * per node partial list. Postpone the actual unfreezing
2565 * outside of the critical section.
2567 page_to_unfreeze
= oldpage
;
2570 pobjects
= oldpage
->pobjects
;
2571 pages
= oldpage
->pages
;
2576 pobjects
+= page
->objects
- page
->inuse
;
2578 page
->pages
= pages
;
2579 page
->pobjects
= pobjects
;
2580 page
->next
= oldpage
;
2582 this_cpu_write(s
->cpu_slab
->partial
, page
);
2584 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2586 if (page_to_unfreeze
) {
2587 __unfreeze_partials(s
, page_to_unfreeze
);
2588 stat(s
, CPU_PARTIAL_DRAIN
);
2592 #else /* CONFIG_SLUB_CPU_PARTIAL */
2594 static inline void unfreeze_partials(struct kmem_cache
*s
) { }
2595 static inline void unfreeze_partials_cpu(struct kmem_cache
*s
,
2596 struct kmem_cache_cpu
*c
) { }
2598 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2600 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2602 unsigned long flags
;
2606 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2609 freelist
= c
->freelist
;
2613 c
->tid
= next_tid(c
->tid
);
2615 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2618 deactivate_slab(s
, page
, freelist
);
2619 stat(s
, CPUSLAB_FLUSH
);
2623 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2625 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2626 void *freelist
= c
->freelist
;
2627 struct page
*page
= c
->page
;
2631 c
->tid
= next_tid(c
->tid
);
2634 deactivate_slab(s
, page
, freelist
);
2635 stat(s
, CPUSLAB_FLUSH
);
2638 unfreeze_partials_cpu(s
, c
);
2641 struct slub_flush_work
{
2642 struct work_struct work
;
2643 struct kmem_cache
*s
;
2650 * Called from CPU work handler with migration disabled.
2652 static void flush_cpu_slab(struct work_struct
*w
)
2654 struct kmem_cache
*s
;
2655 struct kmem_cache_cpu
*c
;
2656 struct slub_flush_work
*sfw
;
2658 sfw
= container_of(w
, struct slub_flush_work
, work
);
2661 c
= this_cpu_ptr(s
->cpu_slab
);
2666 unfreeze_partials(s
);
2669 static bool has_cpu_slab(int cpu
, struct kmem_cache
*s
)
2671 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2673 return c
->page
|| slub_percpu_partial(c
);
2676 static DEFINE_MUTEX(flush_lock
);
2677 static DEFINE_PER_CPU(struct slub_flush_work
, slub_flush
);
2679 static void flush_all_cpus_locked(struct kmem_cache
*s
)
2681 struct slub_flush_work
*sfw
;
2684 lockdep_assert_cpus_held();
2685 mutex_lock(&flush_lock
);
2687 for_each_online_cpu(cpu
) {
2688 sfw
= &per_cpu(slub_flush
, cpu
);
2689 if (!has_cpu_slab(cpu
, s
)) {
2693 INIT_WORK(&sfw
->work
, flush_cpu_slab
);
2696 queue_work_on(cpu
, flushwq
, &sfw
->work
);
2699 for_each_online_cpu(cpu
) {
2700 sfw
= &per_cpu(slub_flush
, cpu
);
2703 flush_work(&sfw
->work
);
2706 mutex_unlock(&flush_lock
);
2709 static void flush_all(struct kmem_cache
*s
)
2712 flush_all_cpus_locked(s
);
2717 * Use the cpu notifier to insure that the cpu slabs are flushed when
2720 static int slub_cpu_dead(unsigned int cpu
)
2722 struct kmem_cache
*s
;
2724 mutex_lock(&slab_mutex
);
2725 list_for_each_entry(s
, &slab_caches
, list
)
2726 __flush_cpu_slab(s
, cpu
);
2727 mutex_unlock(&slab_mutex
);
2732 * Check if the objects in a per cpu structure fit numa
2733 * locality expectations.
2735 static inline int node_match(struct page
*page
, int node
)
2738 if (node
!= NUMA_NO_NODE
&& page_to_nid(page
) != node
)
2744 #ifdef CONFIG_SLUB_DEBUG
2745 static int count_free(struct page
*page
)
2747 return page
->objects
- page
->inuse
;
2750 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2752 return atomic_long_read(&n
->total_objects
);
2754 #endif /* CONFIG_SLUB_DEBUG */
2756 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2757 static unsigned long count_partial(struct kmem_cache_node
*n
,
2758 int (*get_count
)(struct page
*))
2760 unsigned long flags
;
2761 unsigned long x
= 0;
2764 spin_lock_irqsave(&n
->list_lock
, flags
);
2765 list_for_each_entry(page
, &n
->partial
, slab_list
)
2766 x
+= get_count(page
);
2767 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2770 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2772 static noinline
void
2773 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2775 #ifdef CONFIG_SLUB_DEBUG
2776 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2777 DEFAULT_RATELIMIT_BURST
);
2779 struct kmem_cache_node
*n
;
2781 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2784 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2785 nid
, gfpflags
, &gfpflags
);
2786 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2787 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2790 if (oo_order(s
->min
) > get_order(s
->object_size
))
2791 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2794 for_each_kmem_cache_node(s
, node
, n
) {
2795 unsigned long nr_slabs
;
2796 unsigned long nr_objs
;
2797 unsigned long nr_free
;
2799 nr_free
= count_partial(n
, count_free
);
2800 nr_slabs
= node_nr_slabs(n
);
2801 nr_objs
= node_nr_objs(n
);
2803 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2804 node
, nr_slabs
, nr_objs
, nr_free
);
2809 static inline bool pfmemalloc_match(struct page
*page
, gfp_t gfpflags
)
2811 if (unlikely(PageSlabPfmemalloc(page
)))
2812 return gfp_pfmemalloc_allowed(gfpflags
);
2818 * A variant of pfmemalloc_match() that tests page flags without asserting
2819 * PageSlab. Intended for opportunistic checks before taking a lock and
2820 * rechecking that nobody else freed the page under us.
2822 static inline bool pfmemalloc_match_unsafe(struct page
*page
, gfp_t gfpflags
)
2824 if (unlikely(__PageSlabPfmemalloc(page
)))
2825 return gfp_pfmemalloc_allowed(gfpflags
);
2831 * Check the page->freelist of a page and either transfer the freelist to the
2832 * per cpu freelist or deactivate the page.
2834 * The page is still frozen if the return value is not NULL.
2836 * If this function returns NULL then the page has been unfrozen.
2838 static inline void *get_freelist(struct kmem_cache
*s
, struct page
*page
)
2841 unsigned long counters
;
2844 lockdep_assert_held(this_cpu_ptr(&s
->cpu_slab
->lock
));
2847 freelist
= page
->freelist
;
2848 counters
= page
->counters
;
2850 new.counters
= counters
;
2851 VM_BUG_ON(!new.frozen
);
2853 new.inuse
= page
->objects
;
2854 new.frozen
= freelist
!= NULL
;
2856 } while (!__cmpxchg_double_slab(s
, page
,
2865 * Slow path. The lockless freelist is empty or we need to perform
2868 * Processing is still very fast if new objects have been freed to the
2869 * regular freelist. In that case we simply take over the regular freelist
2870 * as the lockless freelist and zap the regular freelist.
2872 * If that is not working then we fall back to the partial lists. We take the
2873 * first element of the freelist as the object to allocate now and move the
2874 * rest of the freelist to the lockless freelist.
2876 * And if we were unable to get a new slab from the partial slab lists then
2877 * we need to allocate a new slab. This is the slowest path since it involves
2878 * a call to the page allocator and the setup of a new slab.
2880 * Version of __slab_alloc to use when we know that preemption is
2881 * already disabled (which is the case for bulk allocation).
2883 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2884 unsigned long addr
, struct kmem_cache_cpu
*c
)
2888 unsigned long flags
;
2890 stat(s
, ALLOC_SLOWPATH
);
2894 page
= READ_ONCE(c
->page
);
2897 * if the node is not online or has no normal memory, just
2898 * ignore the node constraint
2900 if (unlikely(node
!= NUMA_NO_NODE
&&
2901 !node_isset(node
, slab_nodes
)))
2902 node
= NUMA_NO_NODE
;
2907 if (unlikely(!node_match(page
, node
))) {
2909 * same as above but node_match() being false already
2910 * implies node != NUMA_NO_NODE
2912 if (!node_isset(node
, slab_nodes
)) {
2913 node
= NUMA_NO_NODE
;
2916 stat(s
, ALLOC_NODE_MISMATCH
);
2917 goto deactivate_slab
;
2922 * By rights, we should be searching for a slab page that was
2923 * PFMEMALLOC but right now, we are losing the pfmemalloc
2924 * information when the page leaves the per-cpu allocator
2926 if (unlikely(!pfmemalloc_match_unsafe(page
, gfpflags
)))
2927 goto deactivate_slab
;
2929 /* must check again c->page in case we got preempted and it changed */
2930 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2931 if (unlikely(page
!= c
->page
)) {
2932 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2935 freelist
= c
->freelist
;
2939 freelist
= get_freelist(s
, page
);
2943 c
->tid
= next_tid(c
->tid
);
2944 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2945 stat(s
, DEACTIVATE_BYPASS
);
2949 stat(s
, ALLOC_REFILL
);
2953 lockdep_assert_held(this_cpu_ptr(&s
->cpu_slab
->lock
));
2956 * freelist is pointing to the list of objects to be used.
2957 * page is pointing to the page from which the objects are obtained.
2958 * That page must be frozen for per cpu allocations to work.
2960 VM_BUG_ON(!c
->page
->frozen
);
2961 c
->freelist
= get_freepointer(s
, freelist
);
2962 c
->tid
= next_tid(c
->tid
);
2963 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2968 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2969 if (page
!= c
->page
) {
2970 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2973 freelist
= c
->freelist
;
2976 c
->tid
= next_tid(c
->tid
);
2977 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2978 deactivate_slab(s
, page
, freelist
);
2982 if (slub_percpu_partial(c
)) {
2983 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2984 if (unlikely(c
->page
)) {
2985 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2988 if (unlikely(!slub_percpu_partial(c
))) {
2989 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2990 /* we were preempted and partial list got empty */
2994 page
= c
->page
= slub_percpu_partial(c
);
2995 slub_set_percpu_partial(c
, page
);
2996 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2997 stat(s
, CPU_PARTIAL_ALLOC
);
3003 freelist
= get_partial(s
, gfpflags
, node
, &page
);
3005 goto check_new_page
;
3007 slub_put_cpu_ptr(s
->cpu_slab
);
3008 page
= new_slab(s
, gfpflags
, node
);
3009 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3011 if (unlikely(!page
)) {
3012 slab_out_of_memory(s
, gfpflags
, node
);
3017 * No other reference to the page yet so we can
3018 * muck around with it freely without cmpxchg
3020 freelist
= page
->freelist
;
3021 page
->freelist
= NULL
;
3023 stat(s
, ALLOC_SLAB
);
3027 if (kmem_cache_debug(s
)) {
3028 if (!alloc_debug_processing(s
, page
, freelist
, addr
)) {
3029 /* Slab failed checks. Next slab needed */
3033 * For debug case, we don't load freelist so that all
3034 * allocations go through alloc_debug_processing()
3040 if (unlikely(!pfmemalloc_match(page
, gfpflags
)))
3042 * For !pfmemalloc_match() case we don't load freelist so that
3043 * we don't make further mismatched allocations easier.
3049 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
3050 if (unlikely(c
->page
)) {
3051 void *flush_freelist
= c
->freelist
;
3052 struct page
*flush_page
= c
->page
;
3056 c
->tid
= next_tid(c
->tid
);
3058 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3060 deactivate_slab(s
, flush_page
, flush_freelist
);
3062 stat(s
, CPUSLAB_FLUSH
);
3064 goto retry_load_page
;
3072 deactivate_slab(s
, page
, get_freepointer(s
, freelist
));
3077 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3078 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3081 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
3082 unsigned long addr
, struct kmem_cache_cpu
*c
)
3086 #ifdef CONFIG_PREEMPT_COUNT
3088 * We may have been preempted and rescheduled on a different
3089 * cpu before disabling preemption. Need to reload cpu area
3092 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3095 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
3096 #ifdef CONFIG_PREEMPT_COUNT
3097 slub_put_cpu_ptr(s
->cpu_slab
);
3103 * If the object has been wiped upon free, make sure it's fully initialized by
3104 * zeroing out freelist pointer.
3106 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
3109 if (unlikely(slab_want_init_on_free(s
)) && obj
)
3110 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
3115 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3116 * have the fastpath folded into their functions. So no function call
3117 * overhead for requests that can be satisfied on the fastpath.
3119 * The fastpath works by first checking if the lockless freelist can be used.
3120 * If not then __slab_alloc is called for slow processing.
3122 * Otherwise we can simply pick the next object from the lockless free list.
3124 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
,
3125 gfp_t gfpflags
, int node
, unsigned long addr
, size_t orig_size
)
3128 struct kmem_cache_cpu
*c
;
3131 struct obj_cgroup
*objcg
= NULL
;
3134 s
= slab_pre_alloc_hook(s
, &objcg
, 1, gfpflags
);
3138 object
= kfence_alloc(s
, orig_size
, gfpflags
);
3139 if (unlikely(object
))
3144 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3145 * enabled. We may switch back and forth between cpus while
3146 * reading from one cpu area. That does not matter as long
3147 * as we end up on the original cpu again when doing the cmpxchg.
3149 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3150 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3151 * the tid. If we are preempted and switched to another cpu between the
3152 * two reads, it's OK as the two are still associated with the same cpu
3153 * and cmpxchg later will validate the cpu.
3155 c
= raw_cpu_ptr(s
->cpu_slab
);
3156 tid
= READ_ONCE(c
->tid
);
3159 * Irqless object alloc/free algorithm used here depends on sequence
3160 * of fetching cpu_slab's data. tid should be fetched before anything
3161 * on c to guarantee that object and page associated with previous tid
3162 * won't be used with current tid. If we fetch tid first, object and
3163 * page could be one associated with next tid and our alloc/free
3164 * request will be failed. In this case, we will retry. So, no problem.
3169 * The transaction ids are globally unique per cpu and per operation on
3170 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3171 * occurs on the right processor and that there was no operation on the
3172 * linked list in between.
3175 object
= c
->freelist
;
3178 * We cannot use the lockless fastpath on PREEMPT_RT because if a
3179 * slowpath has taken the local_lock_irqsave(), it is not protected
3180 * against a fast path operation in an irq handler. So we need to take
3181 * the slow path which uses local_lock. It is still relatively fast if
3182 * there is a suitable cpu freelist.
3184 if (IS_ENABLED(CONFIG_PREEMPT_RT
) ||
3185 unlikely(!object
|| !page
|| !node_match(page
, node
))) {
3186 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
3188 void *next_object
= get_freepointer_safe(s
, object
);
3191 * The cmpxchg will only match if there was no additional
3192 * operation and if we are on the right processor.
3194 * The cmpxchg does the following atomically (without lock
3196 * 1. Relocate first pointer to the current per cpu area.
3197 * 2. Verify that tid and freelist have not been changed
3198 * 3. If they were not changed replace tid and freelist
3200 * Since this is without lock semantics the protection is only
3201 * against code executing on this cpu *not* from access by
3204 if (unlikely(!this_cpu_cmpxchg_double(
3205 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3207 next_object
, next_tid(tid
)))) {
3209 note_cmpxchg_failure("slab_alloc", s
, tid
);
3212 prefetch_freepointer(s
, next_object
);
3213 stat(s
, ALLOC_FASTPATH
);
3216 maybe_wipe_obj_freeptr(s
, object
);
3217 init
= slab_want_init_on_alloc(gfpflags
, s
);
3220 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
, init
);
3225 static __always_inline
void *slab_alloc(struct kmem_cache
*s
,
3226 gfp_t gfpflags
, unsigned long addr
, size_t orig_size
)
3228 return slab_alloc_node(s
, gfpflags
, NUMA_NO_NODE
, addr
, orig_size
);
3231 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
3233 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, s
->object_size
);
3235 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
->object_size
,
3240 EXPORT_SYMBOL(kmem_cache_alloc
);
3242 #ifdef CONFIG_TRACING
3243 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
3245 void *ret
= slab_alloc(s
, gfpflags
, _RET_IP_
, size
);
3246 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, gfpflags
);
3247 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
3250 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3254 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
3256 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, s
->object_size
);
3258 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3259 s
->object_size
, s
->size
, gfpflags
, node
);
3263 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3265 #ifdef CONFIG_TRACING
3266 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
3268 int node
, size_t size
)
3270 void *ret
= slab_alloc_node(s
, gfpflags
, node
, _RET_IP_
, size
);
3272 trace_kmalloc_node(_RET_IP_
, ret
,
3273 size
, s
->size
, gfpflags
, node
);
3275 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
3278 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3280 #endif /* CONFIG_NUMA */
3283 * Slow path handling. This may still be called frequently since objects
3284 * have a longer lifetime than the cpu slabs in most processing loads.
3286 * So we still attempt to reduce cache line usage. Just take the slab
3287 * lock and free the item. If there is no additional partial page
3288 * handling required then we can return immediately.
3290 static void __slab_free(struct kmem_cache
*s
, struct page
*page
,
3291 void *head
, void *tail
, int cnt
,
3298 unsigned long counters
;
3299 struct kmem_cache_node
*n
= NULL
;
3300 unsigned long flags
;
3302 stat(s
, FREE_SLOWPATH
);
3304 if (kfence_free(head
))
3307 if (kmem_cache_debug(s
) &&
3308 !free_debug_processing(s
, page
, head
, tail
, cnt
, addr
))
3313 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3316 prior
= page
->freelist
;
3317 counters
= page
->counters
;
3318 set_freepointer(s
, tail
, prior
);
3319 new.counters
= counters
;
3320 was_frozen
= new.frozen
;
3322 if ((!new.inuse
|| !prior
) && !was_frozen
) {
3324 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
3327 * Slab was on no list before and will be
3329 * We can defer the list move and instead
3334 } else { /* Needs to be taken off a list */
3336 n
= get_node(s
, page_to_nid(page
));
3338 * Speculatively acquire the list_lock.
3339 * If the cmpxchg does not succeed then we may
3340 * drop the list_lock without any processing.
3342 * Otherwise the list_lock will synchronize with
3343 * other processors updating the list of slabs.
3345 spin_lock_irqsave(&n
->list_lock
, flags
);
3350 } while (!cmpxchg_double_slab(s
, page
,
3357 if (likely(was_frozen
)) {
3359 * The list lock was not taken therefore no list
3360 * activity can be necessary.
3362 stat(s
, FREE_FROZEN
);
3363 } else if (new.frozen
) {
3365 * If we just froze the page then put it onto the
3366 * per cpu partial list.
3368 put_cpu_partial(s
, page
, 1);
3369 stat(s
, CPU_PARTIAL_FREE
);
3375 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3379 * Objects left in the slab. If it was not on the partial list before
3382 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3383 remove_full(s
, n
, page
);
3384 add_partial(n
, page
, DEACTIVATE_TO_TAIL
);
3385 stat(s
, FREE_ADD_PARTIAL
);
3387 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3393 * Slab on the partial list.
3395 remove_partial(n
, page
);
3396 stat(s
, FREE_REMOVE_PARTIAL
);
3398 /* Slab must be on the full list */
3399 remove_full(s
, n
, page
);
3402 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3404 discard_slab(s
, page
);
3408 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3409 * can perform fastpath freeing without additional function calls.
3411 * The fastpath is only possible if we are freeing to the current cpu slab
3412 * of this processor. This typically the case if we have just allocated
3415 * If fastpath is not possible then fall back to __slab_free where we deal
3416 * with all sorts of special processing.
3418 * Bulk free of a freelist with several objects (all pointing to the
3419 * same page) possible by specifying head and tail ptr, plus objects
3420 * count (cnt). Bulk free indicated by tail pointer being set.
3422 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3423 struct page
*page
, void *head
, void *tail
,
3424 int cnt
, unsigned long addr
)
3426 void *tail_obj
= tail
? : head
;
3427 struct kmem_cache_cpu
*c
;
3430 /* memcg_slab_free_hook() is already called for bulk free. */
3432 memcg_slab_free_hook(s
, &head
, 1);
3435 * Determine the currently cpus per cpu slab.
3436 * The cpu may change afterward. However that does not matter since
3437 * data is retrieved via this pointer. If we are on the same cpu
3438 * during the cmpxchg then the free will succeed.
3440 c
= raw_cpu_ptr(s
->cpu_slab
);
3441 tid
= READ_ONCE(c
->tid
);
3443 /* Same with comment on barrier() in slab_alloc_node() */
3446 if (likely(page
== c
->page
)) {
3447 #ifndef CONFIG_PREEMPT_RT
3448 void **freelist
= READ_ONCE(c
->freelist
);
3450 set_freepointer(s
, tail_obj
, freelist
);
3452 if (unlikely(!this_cpu_cmpxchg_double(
3453 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3455 head
, next_tid(tid
)))) {
3457 note_cmpxchg_failure("slab_free", s
, tid
);
3460 #else /* CONFIG_PREEMPT_RT */
3462 * We cannot use the lockless fastpath on PREEMPT_RT because if
3463 * a slowpath has taken the local_lock_irqsave(), it is not
3464 * protected against a fast path operation in an irq handler. So
3465 * we need to take the local_lock. We shouldn't simply defer to
3466 * __slab_free() as that wouldn't use the cpu freelist at all.
3470 local_lock(&s
->cpu_slab
->lock
);
3471 c
= this_cpu_ptr(s
->cpu_slab
);
3472 if (unlikely(page
!= c
->page
)) {
3473 local_unlock(&s
->cpu_slab
->lock
);
3477 freelist
= c
->freelist
;
3479 set_freepointer(s
, tail_obj
, freelist
);
3481 c
->tid
= next_tid(tid
);
3483 local_unlock(&s
->cpu_slab
->lock
);
3485 stat(s
, FREE_FASTPATH
);
3487 __slab_free(s
, page
, head
, tail_obj
, cnt
, addr
);
3491 static __always_inline
void slab_free(struct kmem_cache
*s
, struct page
*page
,
3492 void *head
, void *tail
, int cnt
,
3496 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3497 * to remove objects, whose reuse must be delayed.
3499 if (slab_free_freelist_hook(s
, &head
, &tail
, &cnt
))
3500 do_slab_free(s
, page
, head
, tail
, cnt
, addr
);
3503 #ifdef CONFIG_KASAN_GENERIC
3504 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3506 do_slab_free(cache
, virt_to_head_page(x
), x
, NULL
, 1, addr
);
3510 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3512 s
= cache_from_obj(s
, x
);
3515 slab_free(s
, virt_to_head_page(x
), x
, NULL
, 1, _RET_IP_
);
3516 trace_kmem_cache_free(_RET_IP_
, x
, s
->name
);
3518 EXPORT_SYMBOL(kmem_cache_free
);
3520 struct detached_freelist
{
3525 struct kmem_cache
*s
;
3528 static inline void free_nonslab_page(struct page
*page
, void *object
)
3530 unsigned int order
= compound_order(page
);
3532 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
3534 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
, -(PAGE_SIZE
<< order
));
3535 __free_pages(page
, order
);
3539 * This function progressively scans the array with free objects (with
3540 * a limited look ahead) and extract objects belonging to the same
3541 * page. It builds a detached freelist directly within the given
3542 * page/objects. This can happen without any need for
3543 * synchronization, because the objects are owned by running process.
3544 * The freelist is build up as a single linked list in the objects.
3545 * The idea is, that this detached freelist can then be bulk
3546 * transferred to the real freelist(s), but only requiring a single
3547 * synchronization primitive. Look ahead in the array is limited due
3548 * to performance reasons.
3551 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3552 void **p
, struct detached_freelist
*df
)
3554 size_t first_skipped_index
= 0;
3559 /* Always re-init detached_freelist */
3564 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3565 } while (!object
&& size
);
3570 page
= virt_to_head_page(object
);
3572 /* Handle kalloc'ed objects */
3573 if (unlikely(!PageSlab(page
))) {
3574 free_nonslab_page(page
, object
);
3575 p
[size
] = NULL
; /* mark object processed */
3578 /* Derive kmem_cache from object */
3579 df
->s
= page
->slab_cache
;
3581 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3584 if (is_kfence_address(object
)) {
3585 slab_free_hook(df
->s
, object
, false);
3586 __kfence_free(object
);
3587 p
[size
] = NULL
; /* mark object processed */
3591 /* Start new detached freelist */
3593 set_freepointer(df
->s
, object
, NULL
);
3595 df
->freelist
= object
;
3596 p
[size
] = NULL
; /* mark object processed */
3602 continue; /* Skip processed objects */
3604 /* df->page is always set at this point */
3605 if (df
->page
== virt_to_head_page(object
)) {
3606 /* Opportunity build freelist */
3607 set_freepointer(df
->s
, object
, df
->freelist
);
3608 df
->freelist
= object
;
3610 p
[size
] = NULL
; /* mark object processed */
3615 /* Limit look ahead search */
3619 if (!first_skipped_index
)
3620 first_skipped_index
= size
+ 1;
3623 return first_skipped_index
;
3626 /* Note that interrupts must be enabled when calling this function. */
3627 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3632 memcg_slab_free_hook(s
, p
, size
);
3634 struct detached_freelist df
;
3636 size
= build_detached_freelist(s
, size
, p
, &df
);
3640 slab_free(df
.s
, df
.page
, df
.freelist
, df
.tail
, df
.cnt
, _RET_IP_
);
3641 } while (likely(size
));
3643 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3645 /* Note that interrupts must be enabled when calling this function. */
3646 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3649 struct kmem_cache_cpu
*c
;
3651 struct obj_cgroup
*objcg
= NULL
;
3653 /* memcg and kmem_cache debug support */
3654 s
= slab_pre_alloc_hook(s
, &objcg
, size
, flags
);
3658 * Drain objects in the per cpu slab, while disabling local
3659 * IRQs, which protects against PREEMPT and interrupts
3660 * handlers invoking normal fastpath.
3662 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3663 local_lock_irq(&s
->cpu_slab
->lock
);
3665 for (i
= 0; i
< size
; i
++) {
3666 void *object
= kfence_alloc(s
, s
->object_size
, flags
);
3668 if (unlikely(object
)) {
3673 object
= c
->freelist
;
3674 if (unlikely(!object
)) {
3676 * We may have removed an object from c->freelist using
3677 * the fastpath in the previous iteration; in that case,
3678 * c->tid has not been bumped yet.
3679 * Since ___slab_alloc() may reenable interrupts while
3680 * allocating memory, we should bump c->tid now.
3682 c
->tid
= next_tid(c
->tid
);
3684 local_unlock_irq(&s
->cpu_slab
->lock
);
3687 * Invoking slow path likely have side-effect
3688 * of re-populating per CPU c->freelist
3690 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3692 if (unlikely(!p
[i
]))
3695 c
= this_cpu_ptr(s
->cpu_slab
);
3696 maybe_wipe_obj_freeptr(s
, p
[i
]);
3698 local_lock_irq(&s
->cpu_slab
->lock
);
3700 continue; /* goto for-loop */
3702 c
->freelist
= get_freepointer(s
, object
);
3704 maybe_wipe_obj_freeptr(s
, p
[i
]);
3706 c
->tid
= next_tid(c
->tid
);
3707 local_unlock_irq(&s
->cpu_slab
->lock
);
3708 slub_put_cpu_ptr(s
->cpu_slab
);
3711 * memcg and kmem_cache debug support and memory initialization.
3712 * Done outside of the IRQ disabled fastpath loop.
3714 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3715 slab_want_init_on_alloc(flags
, s
));
3718 slub_put_cpu_ptr(s
->cpu_slab
);
3719 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false);
3720 __kmem_cache_free_bulk(s
, i
, p
);
3723 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3727 * Object placement in a slab is made very easy because we always start at
3728 * offset 0. If we tune the size of the object to the alignment then we can
3729 * get the required alignment by putting one properly sized object after
3732 * Notice that the allocation order determines the sizes of the per cpu
3733 * caches. Each processor has always one slab available for allocations.
3734 * Increasing the allocation order reduces the number of times that slabs
3735 * must be moved on and off the partial lists and is therefore a factor in
3740 * Minimum / Maximum order of slab pages. This influences locking overhead
3741 * and slab fragmentation. A higher order reduces the number of partial slabs
3742 * and increases the number of allocations possible without having to
3743 * take the list_lock.
3745 static unsigned int slub_min_order
;
3746 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3747 static unsigned int slub_min_objects
;
3750 * Calculate the order of allocation given an slab object size.
3752 * The order of allocation has significant impact on performance and other
3753 * system components. Generally order 0 allocations should be preferred since
3754 * order 0 does not cause fragmentation in the page allocator. Larger objects
3755 * be problematic to put into order 0 slabs because there may be too much
3756 * unused space left. We go to a higher order if more than 1/16th of the slab
3759 * In order to reach satisfactory performance we must ensure that a minimum
3760 * number of objects is in one slab. Otherwise we may generate too much
3761 * activity on the partial lists which requires taking the list_lock. This is
3762 * less a concern for large slabs though which are rarely used.
3764 * slub_max_order specifies the order where we begin to stop considering the
3765 * number of objects in a slab as critical. If we reach slub_max_order then
3766 * we try to keep the page order as low as possible. So we accept more waste
3767 * of space in favor of a small page order.
3769 * Higher order allocations also allow the placement of more objects in a
3770 * slab and thereby reduce object handling overhead. If the user has
3771 * requested a higher minimum order then we start with that one instead of
3772 * the smallest order which will fit the object.
3774 static inline unsigned int slab_order(unsigned int size
,
3775 unsigned int min_objects
, unsigned int max_order
,
3776 unsigned int fract_leftover
)
3778 unsigned int min_order
= slub_min_order
;
3781 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3782 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3784 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3785 order
<= max_order
; order
++) {
3787 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3790 rem
= slab_size
% size
;
3792 if (rem
<= slab_size
/ fract_leftover
)
3799 static inline int calculate_order(unsigned int size
)
3802 unsigned int min_objects
;
3803 unsigned int max_objects
;
3804 unsigned int nr_cpus
;
3807 * Attempt to find best configuration for a slab. This
3808 * works by first attempting to generate a layout with
3809 * the best configuration and backing off gradually.
3811 * First we increase the acceptable waste in a slab. Then
3812 * we reduce the minimum objects required in a slab.
3814 min_objects
= slub_min_objects
;
3817 * Some architectures will only update present cpus when
3818 * onlining them, so don't trust the number if it's just 1. But
3819 * we also don't want to use nr_cpu_ids always, as on some other
3820 * architectures, there can be many possible cpus, but never
3821 * onlined. Here we compromise between trying to avoid too high
3822 * order on systems that appear larger than they are, and too
3823 * low order on systems that appear smaller than they are.
3825 nr_cpus
= num_present_cpus();
3827 nr_cpus
= nr_cpu_ids
;
3828 min_objects
= 4 * (fls(nr_cpus
) + 1);
3830 max_objects
= order_objects(slub_max_order
, size
);
3831 min_objects
= min(min_objects
, max_objects
);
3833 while (min_objects
> 1) {
3834 unsigned int fraction
;
3837 while (fraction
>= 4) {
3838 order
= slab_order(size
, min_objects
,
3839 slub_max_order
, fraction
);
3840 if (order
<= slub_max_order
)
3848 * We were unable to place multiple objects in a slab. Now
3849 * lets see if we can place a single object there.
3851 order
= slab_order(size
, 1, slub_max_order
, 1);
3852 if (order
<= slub_max_order
)
3856 * Doh this slab cannot be placed using slub_max_order.
3858 order
= slab_order(size
, 1, MAX_ORDER
, 1);
3859 if (order
< MAX_ORDER
)
3865 init_kmem_cache_node(struct kmem_cache_node
*n
)
3868 spin_lock_init(&n
->list_lock
);
3869 INIT_LIST_HEAD(&n
->partial
);
3870 #ifdef CONFIG_SLUB_DEBUG
3871 atomic_long_set(&n
->nr_slabs
, 0);
3872 atomic_long_set(&n
->total_objects
, 0);
3873 INIT_LIST_HEAD(&n
->full
);
3877 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3879 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3880 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3883 * Must align to double word boundary for the double cmpxchg
3884 * instructions to work; see __pcpu_double_call_return_bool().
3886 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3887 2 * sizeof(void *));
3892 init_kmem_cache_cpus(s
);
3897 static struct kmem_cache
*kmem_cache_node
;
3900 * No kmalloc_node yet so do it by hand. We know that this is the first
3901 * slab on the node for this slabcache. There are no concurrent accesses
3904 * Note that this function only works on the kmem_cache_node
3905 * when allocating for the kmem_cache_node. This is used for bootstrapping
3906 * memory on a fresh node that has no slab structures yet.
3908 static void early_kmem_cache_node_alloc(int node
)
3911 struct kmem_cache_node
*n
;
3913 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3915 page
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3918 if (page_to_nid(page
) != node
) {
3919 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3920 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3925 #ifdef CONFIG_SLUB_DEBUG
3926 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3927 init_tracking(kmem_cache_node
, n
);
3929 n
= kasan_slab_alloc(kmem_cache_node
, n
, GFP_KERNEL
, false);
3930 page
->freelist
= get_freepointer(kmem_cache_node
, n
);
3933 kmem_cache_node
->node
[node
] = n
;
3934 init_kmem_cache_node(n
);
3935 inc_slabs_node(kmem_cache_node
, node
, page
->objects
);
3938 * No locks need to be taken here as it has just been
3939 * initialized and there is no concurrent access.
3941 __add_partial(n
, page
, DEACTIVATE_TO_HEAD
);
3944 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3947 struct kmem_cache_node
*n
;
3949 for_each_kmem_cache_node(s
, node
, n
) {
3950 s
->node
[node
] = NULL
;
3951 kmem_cache_free(kmem_cache_node
, n
);
3955 void __kmem_cache_release(struct kmem_cache
*s
)
3957 cache_random_seq_destroy(s
);
3958 free_percpu(s
->cpu_slab
);
3959 free_kmem_cache_nodes(s
);
3962 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3966 for_each_node_mask(node
, slab_nodes
) {
3967 struct kmem_cache_node
*n
;
3969 if (slab_state
== DOWN
) {
3970 early_kmem_cache_node_alloc(node
);
3973 n
= kmem_cache_alloc_node(kmem_cache_node
,
3977 free_kmem_cache_nodes(s
);
3981 init_kmem_cache_node(n
);
3987 static void set_min_partial(struct kmem_cache
*s
, unsigned long min
)
3989 if (min
< MIN_PARTIAL
)
3991 else if (min
> MAX_PARTIAL
)
3993 s
->min_partial
= min
;
3996 static void set_cpu_partial(struct kmem_cache
*s
)
3998 #ifdef CONFIG_SLUB_CPU_PARTIAL
4000 * cpu_partial determined the maximum number of objects kept in the
4001 * per cpu partial lists of a processor.
4003 * Per cpu partial lists mainly contain slabs that just have one
4004 * object freed. If they are used for allocation then they can be
4005 * filled up again with minimal effort. The slab will never hit the
4006 * per node partial lists and therefore no locking will be required.
4008 * This setting also determines
4010 * A) The number of objects from per cpu partial slabs dumped to the
4011 * per node list when we reach the limit.
4012 * B) The number of objects in cpu partial slabs to extract from the
4013 * per node list when we run out of per cpu objects. We only fetch
4014 * 50% to keep some capacity around for frees.
4016 if (!kmem_cache_has_cpu_partial(s
))
4017 slub_set_cpu_partial(s
, 0);
4018 else if (s
->size
>= PAGE_SIZE
)
4019 slub_set_cpu_partial(s
, 2);
4020 else if (s
->size
>= 1024)
4021 slub_set_cpu_partial(s
, 6);
4022 else if (s
->size
>= 256)
4023 slub_set_cpu_partial(s
, 13);
4025 slub_set_cpu_partial(s
, 30);
4030 * calculate_sizes() determines the order and the distribution of data within
4033 static int calculate_sizes(struct kmem_cache
*s
, int forced_order
)
4035 slab_flags_t flags
= s
->flags
;
4036 unsigned int size
= s
->object_size
;
4040 * Round up object size to the next word boundary. We can only
4041 * place the free pointer at word boundaries and this determines
4042 * the possible location of the free pointer.
4044 size
= ALIGN(size
, sizeof(void *));
4046 #ifdef CONFIG_SLUB_DEBUG
4048 * Determine if we can poison the object itself. If the user of
4049 * the slab may touch the object after free or before allocation
4050 * then we should never poison the object itself.
4052 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
4054 s
->flags
|= __OBJECT_POISON
;
4056 s
->flags
&= ~__OBJECT_POISON
;
4060 * If we are Redzoning then check if there is some space between the
4061 * end of the object and the free pointer. If not then add an
4062 * additional word to have some bytes to store Redzone information.
4064 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
4065 size
+= sizeof(void *);
4069 * With that we have determined the number of bytes in actual use
4070 * by the object and redzoning.
4074 if ((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
4075 ((flags
& SLAB_RED_ZONE
) && s
->object_size
< sizeof(void *)) ||
4078 * Relocate free pointer after the object if it is not
4079 * permitted to overwrite the first word of the object on
4082 * This is the case if we do RCU, have a constructor or
4083 * destructor, are poisoning the objects, or are
4084 * redzoning an object smaller than sizeof(void *).
4086 * The assumption that s->offset >= s->inuse means free
4087 * pointer is outside of the object is used in the
4088 * freeptr_outside_object() function. If that is no
4089 * longer true, the function needs to be modified.
4092 size
+= sizeof(void *);
4095 * Store freelist pointer near middle of object to keep
4096 * it away from the edges of the object to avoid small
4097 * sized over/underflows from neighboring allocations.
4099 s
->offset
= ALIGN_DOWN(s
->object_size
/ 2, sizeof(void *));
4102 #ifdef CONFIG_SLUB_DEBUG
4103 if (flags
& SLAB_STORE_USER
)
4105 * Need to store information about allocs and frees after
4108 size
+= 2 * sizeof(struct track
);
4111 kasan_cache_create(s
, &size
, &s
->flags
);
4112 #ifdef CONFIG_SLUB_DEBUG
4113 if (flags
& SLAB_RED_ZONE
) {
4115 * Add some empty padding so that we can catch
4116 * overwrites from earlier objects rather than let
4117 * tracking information or the free pointer be
4118 * corrupted if a user writes before the start
4121 size
+= sizeof(void *);
4123 s
->red_left_pad
= sizeof(void *);
4124 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
4125 size
+= s
->red_left_pad
;
4130 * SLUB stores one object immediately after another beginning from
4131 * offset 0. In order to align the objects we have to simply size
4132 * each object to conform to the alignment.
4134 size
= ALIGN(size
, s
->align
);
4136 s
->reciprocal_size
= reciprocal_value(size
);
4137 if (forced_order
>= 0)
4138 order
= forced_order
;
4140 order
= calculate_order(size
);
4147 s
->allocflags
|= __GFP_COMP
;
4149 if (s
->flags
& SLAB_CACHE_DMA
)
4150 s
->allocflags
|= GFP_DMA
;
4152 if (s
->flags
& SLAB_CACHE_DMA32
)
4153 s
->allocflags
|= GFP_DMA32
;
4155 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
4156 s
->allocflags
|= __GFP_RECLAIMABLE
;
4159 * Determine the number of objects per slab
4161 s
->oo
= oo_make(order
, size
);
4162 s
->min
= oo_make(get_order(size
), size
);
4163 if (oo_objects(s
->oo
) > oo_objects(s
->max
))
4166 return !!oo_objects(s
->oo
);
4169 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
4171 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
);
4172 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4173 s
->random
= get_random_long();
4176 if (!calculate_sizes(s
, -1))
4178 if (disable_higher_order_debug
) {
4180 * Disable debugging flags that store metadata if the min slab
4183 if (get_order(s
->size
) > get_order(s
->object_size
)) {
4184 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
4186 if (!calculate_sizes(s
, -1))
4191 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4192 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4193 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
4194 /* Enable fast mode */
4195 s
->flags
|= __CMPXCHG_DOUBLE
;
4199 * The larger the object size is, the more pages we want on the partial
4200 * list to avoid pounding the page allocator excessively.
4202 set_min_partial(s
, ilog2(s
->size
) / 2);
4207 s
->remote_node_defrag_ratio
= 1000;
4210 /* Initialize the pre-computed randomized freelist if slab is up */
4211 if (slab_state
>= UP
) {
4212 if (init_cache_random_seq(s
))
4216 if (!init_kmem_cache_nodes(s
))
4219 if (alloc_kmem_cache_cpus(s
))
4223 __kmem_cache_release(s
);
4227 static void list_slab_objects(struct kmem_cache
*s
, struct page
*page
,
4230 #ifdef CONFIG_SLUB_DEBUG
4231 void *addr
= page_address(page
);
4232 unsigned long flags
;
4236 slab_err(s
, page
, text
, s
->name
);
4237 slab_lock(page
, &flags
);
4239 map
= get_map(s
, page
);
4240 for_each_object(p
, s
, addr
, page
->objects
) {
4242 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
4243 pr_err("Object 0x%p @offset=%tu\n", p
, p
- addr
);
4244 print_tracking(s
, p
);
4248 slab_unlock(page
, &flags
);
4253 * Attempt to free all partial slabs on a node.
4254 * This is called from __kmem_cache_shutdown(). We must take list_lock
4255 * because sysfs file might still access partial list after the shutdowning.
4257 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
4260 struct page
*page
, *h
;
4262 BUG_ON(irqs_disabled());
4263 spin_lock_irq(&n
->list_lock
);
4264 list_for_each_entry_safe(page
, h
, &n
->partial
, slab_list
) {
4266 remove_partial(n
, page
);
4267 list_add(&page
->slab_list
, &discard
);
4269 list_slab_objects(s
, page
,
4270 "Objects remaining in %s on __kmem_cache_shutdown()");
4273 spin_unlock_irq(&n
->list_lock
);
4275 list_for_each_entry_safe(page
, h
, &discard
, slab_list
)
4276 discard_slab(s
, page
);
4279 bool __kmem_cache_empty(struct kmem_cache
*s
)
4282 struct kmem_cache_node
*n
;
4284 for_each_kmem_cache_node(s
, node
, n
)
4285 if (n
->nr_partial
|| slabs_node(s
, node
))
4291 * Release all resources used by a slab cache.
4293 int __kmem_cache_shutdown(struct kmem_cache
*s
)
4296 struct kmem_cache_node
*n
;
4298 flush_all_cpus_locked(s
);
4299 /* Attempt to free all objects */
4300 for_each_kmem_cache_node(s
, node
, n
) {
4302 if (n
->nr_partial
|| slabs_node(s
, node
))
4308 #ifdef CONFIG_PRINTK
4309 void __kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct page
*page
)
4312 int __maybe_unused i
;
4316 struct kmem_cache
*s
= page
->slab_cache
;
4317 struct track __maybe_unused
*trackp
;
4319 kpp
->kp_ptr
= object
;
4320 kpp
->kp_page
= page
;
4321 kpp
->kp_slab_cache
= s
;
4322 base
= page_address(page
);
4323 objp0
= kasan_reset_tag(object
);
4324 #ifdef CONFIG_SLUB_DEBUG
4325 objp
= restore_red_left(s
, objp0
);
4329 objnr
= obj_to_index(s
, page
, objp
);
4330 kpp
->kp_data_offset
= (unsigned long)((char *)objp0
- (char *)objp
);
4331 objp
= base
+ s
->size
* objnr
;
4332 kpp
->kp_objp
= objp
;
4333 if (WARN_ON_ONCE(objp
< base
|| objp
>= base
+ page
->objects
* s
->size
|| (objp
- base
) % s
->size
) ||
4334 !(s
->flags
& SLAB_STORE_USER
))
4336 #ifdef CONFIG_SLUB_DEBUG
4337 objp
= fixup_red_left(s
, objp
);
4338 trackp
= get_track(s
, objp
, TRACK_ALLOC
);
4339 kpp
->kp_ret
= (void *)trackp
->addr
;
4340 #ifdef CONFIG_STACKTRACE
4341 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< TRACK_ADDRS_COUNT
; i
++) {
4342 kpp
->kp_stack
[i
] = (void *)trackp
->addrs
[i
];
4343 if (!kpp
->kp_stack
[i
])
4347 trackp
= get_track(s
, objp
, TRACK_FREE
);
4348 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< TRACK_ADDRS_COUNT
; i
++) {
4349 kpp
->kp_free_stack
[i
] = (void *)trackp
->addrs
[i
];
4350 if (!kpp
->kp_free_stack
[i
])
4358 /********************************************************************
4360 *******************************************************************/
4362 static int __init
setup_slub_min_order(char *str
)
4364 get_option(&str
, (int *)&slub_min_order
);
4369 __setup("slub_min_order=", setup_slub_min_order
);
4371 static int __init
setup_slub_max_order(char *str
)
4373 get_option(&str
, (int *)&slub_max_order
);
4374 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
4379 __setup("slub_max_order=", setup_slub_max_order
);
4381 static int __init
setup_slub_min_objects(char *str
)
4383 get_option(&str
, (int *)&slub_min_objects
);
4388 __setup("slub_min_objects=", setup_slub_min_objects
);
4390 void *__kmalloc(size_t size
, gfp_t flags
)
4392 struct kmem_cache
*s
;
4395 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4396 return kmalloc_large(size
, flags
);
4398 s
= kmalloc_slab(size
, flags
);
4400 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4403 ret
= slab_alloc(s
, flags
, _RET_IP_
, size
);
4405 trace_kmalloc(_RET_IP_
, ret
, size
, s
->size
, flags
);
4407 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4411 EXPORT_SYMBOL(__kmalloc
);
4414 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
4418 unsigned int order
= get_order(size
);
4420 flags
|= __GFP_COMP
;
4421 page
= alloc_pages_node(node
, flags
, order
);
4423 ptr
= page_address(page
);
4424 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4425 PAGE_SIZE
<< order
);
4428 return kmalloc_large_node_hook(ptr
, size
, flags
);
4431 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4433 struct kmem_cache
*s
;
4436 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4437 ret
= kmalloc_large_node(size
, flags
, node
);
4439 trace_kmalloc_node(_RET_IP_
, ret
,
4440 size
, PAGE_SIZE
<< get_order(size
),
4446 s
= kmalloc_slab(size
, flags
);
4448 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4451 ret
= slab_alloc_node(s
, flags
, node
, _RET_IP_
, size
);
4453 trace_kmalloc_node(_RET_IP_
, ret
, size
, s
->size
, flags
, node
);
4455 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4459 EXPORT_SYMBOL(__kmalloc_node
);
4460 #endif /* CONFIG_NUMA */
4462 #ifdef CONFIG_HARDENED_USERCOPY
4464 * Rejects incorrectly sized objects and objects that are to be copied
4465 * to/from userspace but do not fall entirely within the containing slab
4466 * cache's usercopy region.
4468 * Returns NULL if check passes, otherwise const char * to name of cache
4469 * to indicate an error.
4471 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4474 struct kmem_cache
*s
;
4475 unsigned int offset
;
4477 bool is_kfence
= is_kfence_address(ptr
);
4479 ptr
= kasan_reset_tag(ptr
);
4481 /* Find object and usable object size. */
4482 s
= page
->slab_cache
;
4484 /* Reject impossible pointers. */
4485 if (ptr
< page_address(page
))
4486 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4489 /* Find offset within object. */
4491 offset
= ptr
- kfence_object_start(ptr
);
4493 offset
= (ptr
- page_address(page
)) % s
->size
;
4495 /* Adjust for redzone and reject if within the redzone. */
4496 if (!is_kfence
&& kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4497 if (offset
< s
->red_left_pad
)
4498 usercopy_abort("SLUB object in left red zone",
4499 s
->name
, to_user
, offset
, n
);
4500 offset
-= s
->red_left_pad
;
4503 /* Allow address range falling entirely within usercopy region. */
4504 if (offset
>= s
->useroffset
&&
4505 offset
- s
->useroffset
<= s
->usersize
&&
4506 n
<= s
->useroffset
- offset
+ s
->usersize
)
4510 * If the copy is still within the allocated object, produce
4511 * a warning instead of rejecting the copy. This is intended
4512 * to be a temporary method to find any missing usercopy
4515 object_size
= slab_ksize(s
);
4516 if (usercopy_fallback
&&
4517 offset
<= object_size
&& n
<= object_size
- offset
) {
4518 usercopy_warn("SLUB object", s
->name
, to_user
, offset
, n
);
4522 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4524 #endif /* CONFIG_HARDENED_USERCOPY */
4526 size_t __ksize(const void *object
)
4530 if (unlikely(object
== ZERO_SIZE_PTR
))
4533 page
= virt_to_head_page(object
);
4535 if (unlikely(!PageSlab(page
))) {
4536 WARN_ON(!PageCompound(page
));
4537 return page_size(page
);
4540 return slab_ksize(page
->slab_cache
);
4542 EXPORT_SYMBOL(__ksize
);
4544 void kfree(const void *x
)
4547 void *object
= (void *)x
;
4549 trace_kfree(_RET_IP_
, x
);
4551 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4554 page
= virt_to_head_page(x
);
4555 if (unlikely(!PageSlab(page
))) {
4556 free_nonslab_page(page
, object
);
4559 slab_free(page
->slab_cache
, page
, object
, NULL
, 1, _RET_IP_
);
4561 EXPORT_SYMBOL(kfree
);
4563 #define SHRINK_PROMOTE_MAX 32
4566 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4567 * up most to the head of the partial lists. New allocations will then
4568 * fill those up and thus they can be removed from the partial lists.
4570 * The slabs with the least items are placed last. This results in them
4571 * being allocated from last increasing the chance that the last objects
4572 * are freed in them.
4574 static int __kmem_cache_do_shrink(struct kmem_cache
*s
)
4578 struct kmem_cache_node
*n
;
4581 struct list_head discard
;
4582 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4583 unsigned long flags
;
4586 for_each_kmem_cache_node(s
, node
, n
) {
4587 INIT_LIST_HEAD(&discard
);
4588 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4589 INIT_LIST_HEAD(promote
+ i
);
4591 spin_lock_irqsave(&n
->list_lock
, flags
);
4594 * Build lists of slabs to discard or promote.
4596 * Note that concurrent frees may occur while we hold the
4597 * list_lock. page->inuse here is the upper limit.
4599 list_for_each_entry_safe(page
, t
, &n
->partial
, slab_list
) {
4600 int free
= page
->objects
- page
->inuse
;
4602 /* Do not reread page->inuse */
4605 /* We do not keep full slabs on the list */
4608 if (free
== page
->objects
) {
4609 list_move(&page
->slab_list
, &discard
);
4611 } else if (free
<= SHRINK_PROMOTE_MAX
)
4612 list_move(&page
->slab_list
, promote
+ free
- 1);
4616 * Promote the slabs filled up most to the head of the
4619 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4620 list_splice(promote
+ i
, &n
->partial
);
4622 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4624 /* Release empty slabs */
4625 list_for_each_entry_safe(page
, t
, &discard
, slab_list
)
4626 discard_slab(s
, page
);
4628 if (slabs_node(s
, node
))
4635 int __kmem_cache_shrink(struct kmem_cache
*s
)
4638 return __kmem_cache_do_shrink(s
);
4641 static int slab_mem_going_offline_callback(void *arg
)
4643 struct kmem_cache
*s
;
4645 mutex_lock(&slab_mutex
);
4646 list_for_each_entry(s
, &slab_caches
, list
) {
4647 flush_all_cpus_locked(s
);
4648 __kmem_cache_do_shrink(s
);
4650 mutex_unlock(&slab_mutex
);
4655 static void slab_mem_offline_callback(void *arg
)
4657 struct memory_notify
*marg
= arg
;
4660 offline_node
= marg
->status_change_nid_normal
;
4663 * If the node still has available memory. we need kmem_cache_node
4666 if (offline_node
< 0)
4669 mutex_lock(&slab_mutex
);
4670 node_clear(offline_node
, slab_nodes
);
4672 * We no longer free kmem_cache_node structures here, as it would be
4673 * racy with all get_node() users, and infeasible to protect them with
4676 mutex_unlock(&slab_mutex
);
4679 static int slab_mem_going_online_callback(void *arg
)
4681 struct kmem_cache_node
*n
;
4682 struct kmem_cache
*s
;
4683 struct memory_notify
*marg
= arg
;
4684 int nid
= marg
->status_change_nid_normal
;
4688 * If the node's memory is already available, then kmem_cache_node is
4689 * already created. Nothing to do.
4695 * We are bringing a node online. No memory is available yet. We must
4696 * allocate a kmem_cache_node structure in order to bring the node
4699 mutex_lock(&slab_mutex
);
4700 list_for_each_entry(s
, &slab_caches
, list
) {
4702 * The structure may already exist if the node was previously
4703 * onlined and offlined.
4705 if (get_node(s
, nid
))
4708 * XXX: kmem_cache_alloc_node will fallback to other nodes
4709 * since memory is not yet available from the node that
4712 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4717 init_kmem_cache_node(n
);
4721 * Any cache created after this point will also have kmem_cache_node
4722 * initialized for the new node.
4724 node_set(nid
, slab_nodes
);
4726 mutex_unlock(&slab_mutex
);
4730 static int slab_memory_callback(struct notifier_block
*self
,
4731 unsigned long action
, void *arg
)
4736 case MEM_GOING_ONLINE
:
4737 ret
= slab_mem_going_online_callback(arg
);
4739 case MEM_GOING_OFFLINE
:
4740 ret
= slab_mem_going_offline_callback(arg
);
4743 case MEM_CANCEL_ONLINE
:
4744 slab_mem_offline_callback(arg
);
4747 case MEM_CANCEL_OFFLINE
:
4751 ret
= notifier_from_errno(ret
);
4757 static struct notifier_block slab_memory_callback_nb
= {
4758 .notifier_call
= slab_memory_callback
,
4759 .priority
= SLAB_CALLBACK_PRI
,
4762 /********************************************************************
4763 * Basic setup of slabs
4764 *******************************************************************/
4767 * Used for early kmem_cache structures that were allocated using
4768 * the page allocator. Allocate them properly then fix up the pointers
4769 * that may be pointing to the wrong kmem_cache structure.
4772 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4775 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4776 struct kmem_cache_node
*n
;
4778 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4781 * This runs very early, and only the boot processor is supposed to be
4782 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4785 __flush_cpu_slab(s
, smp_processor_id());
4786 for_each_kmem_cache_node(s
, node
, n
) {
4789 list_for_each_entry(p
, &n
->partial
, slab_list
)
4792 #ifdef CONFIG_SLUB_DEBUG
4793 list_for_each_entry(p
, &n
->full
, slab_list
)
4797 list_add(&s
->list
, &slab_caches
);
4801 void __init
kmem_cache_init(void)
4803 static __initdata
struct kmem_cache boot_kmem_cache
,
4804 boot_kmem_cache_node
;
4807 if (debug_guardpage_minorder())
4810 /* Print slub debugging pointers without hashing */
4811 if (__slub_debug_enabled())
4812 no_hash_pointers_enable(NULL
);
4814 kmem_cache_node
= &boot_kmem_cache_node
;
4815 kmem_cache
= &boot_kmem_cache
;
4818 * Initialize the nodemask for which we will allocate per node
4819 * structures. Here we don't need taking slab_mutex yet.
4821 for_each_node_state(node
, N_NORMAL_MEMORY
)
4822 node_set(node
, slab_nodes
);
4824 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4825 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4827 register_hotmemory_notifier(&slab_memory_callback_nb
);
4829 /* Able to allocate the per node structures */
4830 slab_state
= PARTIAL
;
4832 create_boot_cache(kmem_cache
, "kmem_cache",
4833 offsetof(struct kmem_cache
, node
) +
4834 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4835 SLAB_HWCACHE_ALIGN
, 0, 0);
4837 kmem_cache
= bootstrap(&boot_kmem_cache
);
4838 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4840 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4841 setup_kmalloc_cache_index_table();
4842 create_kmalloc_caches(0);
4844 /* Setup random freelists for each cache */
4845 init_freelist_randomization();
4847 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4850 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4852 slub_min_order
, slub_max_order
, slub_min_objects
,
4853 nr_cpu_ids
, nr_node_ids
);
4856 void __init
kmem_cache_init_late(void)
4858 flushwq
= alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM
, 0);
4863 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4864 slab_flags_t flags
, void (*ctor
)(void *))
4866 struct kmem_cache
*s
;
4868 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4873 * Adjust the object sizes so that we clear
4874 * the complete object on kzalloc.
4876 s
->object_size
= max(s
->object_size
, size
);
4877 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4879 if (sysfs_slab_alias(s
, name
)) {
4888 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4892 err
= kmem_cache_open(s
, flags
);
4896 /* Mutex is not taken during early boot */
4897 if (slab_state
<= UP
)
4900 err
= sysfs_slab_add(s
);
4902 __kmem_cache_release(s
);
4906 if (s
->flags
& SLAB_STORE_USER
)
4907 debugfs_slab_add(s
);
4912 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4914 struct kmem_cache
*s
;
4917 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4918 return kmalloc_large(size
, gfpflags
);
4920 s
= kmalloc_slab(size
, gfpflags
);
4922 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4925 ret
= slab_alloc(s
, gfpflags
, caller
, size
);
4927 /* Honor the call site pointer we received. */
4928 trace_kmalloc(caller
, ret
, size
, s
->size
, gfpflags
);
4930 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
4934 EXPORT_SYMBOL(__kmalloc_track_caller
);
4937 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4938 int node
, unsigned long caller
)
4940 struct kmem_cache
*s
;
4943 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4944 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4946 trace_kmalloc_node(caller
, ret
,
4947 size
, PAGE_SIZE
<< get_order(size
),
4953 s
= kmalloc_slab(size
, gfpflags
);
4955 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4958 ret
= slab_alloc_node(s
, gfpflags
, node
, caller
, size
);
4960 /* Honor the call site pointer we received. */
4961 trace_kmalloc_node(caller
, ret
, size
, s
->size
, gfpflags
, node
);
4963 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
4967 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4971 static int count_inuse(struct page
*page
)
4976 static int count_total(struct page
*page
)
4978 return page
->objects
;
4982 #ifdef CONFIG_SLUB_DEBUG
4983 static void validate_slab(struct kmem_cache
*s
, struct page
*page
,
4984 unsigned long *obj_map
)
4987 void *addr
= page_address(page
);
4988 unsigned long flags
;
4990 slab_lock(page
, &flags
);
4992 if (!check_slab(s
, page
) || !on_freelist(s
, page
, NULL
))
4995 /* Now we know that a valid freelist exists */
4996 __fill_map(obj_map
, s
, page
);
4997 for_each_object(p
, s
, addr
, page
->objects
) {
4998 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), obj_map
) ?
4999 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
5001 if (!check_object(s
, page
, p
, val
))
5005 slab_unlock(page
, &flags
);
5008 static int validate_slab_node(struct kmem_cache
*s
,
5009 struct kmem_cache_node
*n
, unsigned long *obj_map
)
5011 unsigned long count
= 0;
5013 unsigned long flags
;
5015 spin_lock_irqsave(&n
->list_lock
, flags
);
5017 list_for_each_entry(page
, &n
->partial
, slab_list
) {
5018 validate_slab(s
, page
, obj_map
);
5021 if (count
!= n
->nr_partial
) {
5022 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5023 s
->name
, count
, n
->nr_partial
);
5024 slab_add_kunit_errors();
5027 if (!(s
->flags
& SLAB_STORE_USER
))
5030 list_for_each_entry(page
, &n
->full
, slab_list
) {
5031 validate_slab(s
, page
, obj_map
);
5034 if (count
!= atomic_long_read(&n
->nr_slabs
)) {
5035 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5036 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
5037 slab_add_kunit_errors();
5041 spin_unlock_irqrestore(&n
->list_lock
, flags
);
5045 long validate_slab_cache(struct kmem_cache
*s
)
5048 unsigned long count
= 0;
5049 struct kmem_cache_node
*n
;
5050 unsigned long *obj_map
;
5052 obj_map
= bitmap_alloc(oo_objects(s
->oo
), GFP_KERNEL
);
5057 for_each_kmem_cache_node(s
, node
, n
)
5058 count
+= validate_slab_node(s
, n
, obj_map
);
5060 bitmap_free(obj_map
);
5064 EXPORT_SYMBOL(validate_slab_cache
);
5066 #ifdef CONFIG_DEBUG_FS
5068 * Generate lists of code addresses where slabcache objects are allocated
5073 unsigned long count
;
5080 DECLARE_BITMAP(cpus
, NR_CPUS
);
5086 unsigned long count
;
5087 struct location
*loc
;
5091 static struct dentry
*slab_debugfs_root
;
5093 static void free_loc_track(struct loc_track
*t
)
5096 free_pages((unsigned long)t
->loc
,
5097 get_order(sizeof(struct location
) * t
->max
));
5100 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
5105 order
= get_order(sizeof(struct location
) * max
);
5107 l
= (void *)__get_free_pages(flags
, order
);
5112 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
5120 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
5121 const struct track
*track
)
5123 long start
, end
, pos
;
5125 unsigned long caddr
;
5126 unsigned long age
= jiffies
- track
->when
;
5132 pos
= start
+ (end
- start
+ 1) / 2;
5135 * There is nothing at "end". If we end up there
5136 * we need to add something to before end.
5141 caddr
= t
->loc
[pos
].addr
;
5142 if (track
->addr
== caddr
) {
5148 if (age
< l
->min_time
)
5150 if (age
> l
->max_time
)
5153 if (track
->pid
< l
->min_pid
)
5154 l
->min_pid
= track
->pid
;
5155 if (track
->pid
> l
->max_pid
)
5156 l
->max_pid
= track
->pid
;
5158 cpumask_set_cpu(track
->cpu
,
5159 to_cpumask(l
->cpus
));
5161 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
5165 if (track
->addr
< caddr
)
5172 * Not found. Insert new tracking element.
5174 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
5180 (t
->count
- pos
) * sizeof(struct location
));
5183 l
->addr
= track
->addr
;
5187 l
->min_pid
= track
->pid
;
5188 l
->max_pid
= track
->pid
;
5189 cpumask_clear(to_cpumask(l
->cpus
));
5190 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
5191 nodes_clear(l
->nodes
);
5192 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
5196 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
5197 struct page
*page
, enum track_item alloc
,
5198 unsigned long *obj_map
)
5200 void *addr
= page_address(page
);
5203 __fill_map(obj_map
, s
, page
);
5205 for_each_object(p
, s
, addr
, page
->objects
)
5206 if (!test_bit(__obj_to_index(s
, addr
, p
), obj_map
))
5207 add_location(t
, s
, get_track(s
, p
, alloc
));
5209 #endif /* CONFIG_DEBUG_FS */
5210 #endif /* CONFIG_SLUB_DEBUG */
5213 enum slab_stat_type
{
5214 SL_ALL
, /* All slabs */
5215 SL_PARTIAL
, /* Only partially allocated slabs */
5216 SL_CPU
, /* Only slabs used for cpu caches */
5217 SL_OBJECTS
, /* Determine allocated objects not slabs */
5218 SL_TOTAL
/* Determine object capacity not slabs */
5221 #define SO_ALL (1 << SL_ALL)
5222 #define SO_PARTIAL (1 << SL_PARTIAL)
5223 #define SO_CPU (1 << SL_CPU)
5224 #define SO_OBJECTS (1 << SL_OBJECTS)
5225 #define SO_TOTAL (1 << SL_TOTAL)
5227 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
5228 char *buf
, unsigned long flags
)
5230 unsigned long total
= 0;
5233 unsigned long *nodes
;
5236 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
5240 if (flags
& SO_CPU
) {
5243 for_each_possible_cpu(cpu
) {
5244 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
5249 page
= READ_ONCE(c
->page
);
5253 node
= page_to_nid(page
);
5254 if (flags
& SO_TOTAL
)
5256 else if (flags
& SO_OBJECTS
)
5264 page
= slub_percpu_partial_read_once(c
);
5266 node
= page_to_nid(page
);
5267 if (flags
& SO_TOTAL
)
5269 else if (flags
& SO_OBJECTS
)
5280 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5281 * already held which will conflict with an existing lock order:
5283 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5285 * We don't really need mem_hotplug_lock (to hold off
5286 * slab_mem_going_offline_callback) here because slab's memory hot
5287 * unplug code doesn't destroy the kmem_cache->node[] data.
5290 #ifdef CONFIG_SLUB_DEBUG
5291 if (flags
& SO_ALL
) {
5292 struct kmem_cache_node
*n
;
5294 for_each_kmem_cache_node(s
, node
, n
) {
5296 if (flags
& SO_TOTAL
)
5297 x
= atomic_long_read(&n
->total_objects
);
5298 else if (flags
& SO_OBJECTS
)
5299 x
= atomic_long_read(&n
->total_objects
) -
5300 count_partial(n
, count_free
);
5302 x
= atomic_long_read(&n
->nr_slabs
);
5309 if (flags
& SO_PARTIAL
) {
5310 struct kmem_cache_node
*n
;
5312 for_each_kmem_cache_node(s
, node
, n
) {
5313 if (flags
& SO_TOTAL
)
5314 x
= count_partial(n
, count_total
);
5315 else if (flags
& SO_OBJECTS
)
5316 x
= count_partial(n
, count_inuse
);
5324 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5326 for (node
= 0; node
< nr_node_ids
; node
++) {
5328 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5332 len
+= sysfs_emit_at(buf
, len
, "\n");
5338 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5339 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5341 struct slab_attribute
{
5342 struct attribute attr
;
5343 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5344 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5347 #define SLAB_ATTR_RO(_name) \
5348 static struct slab_attribute _name##_attr = \
5349 __ATTR(_name, 0400, _name##_show, NULL)
5351 #define SLAB_ATTR(_name) \
5352 static struct slab_attribute _name##_attr = \
5353 __ATTR(_name, 0600, _name##_show, _name##_store)
5355 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5357 return sysfs_emit(buf
, "%u\n", s
->size
);
5359 SLAB_ATTR_RO(slab_size
);
5361 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5363 return sysfs_emit(buf
, "%u\n", s
->align
);
5365 SLAB_ATTR_RO(align
);
5367 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5369 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5371 SLAB_ATTR_RO(object_size
);
5373 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5375 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5377 SLAB_ATTR_RO(objs_per_slab
);
5379 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5381 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5383 SLAB_ATTR_RO(order
);
5385 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5387 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5390 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5396 err
= kstrtoul(buf
, 10, &min
);
5400 set_min_partial(s
, min
);
5403 SLAB_ATTR(min_partial
);
5405 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5407 return sysfs_emit(buf
, "%u\n", slub_cpu_partial(s
));
5410 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5413 unsigned int objects
;
5416 err
= kstrtouint(buf
, 10, &objects
);
5419 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5422 slub_set_cpu_partial(s
, objects
);
5426 SLAB_ATTR(cpu_partial
);
5428 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5432 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5436 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5438 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5440 SLAB_ATTR_RO(aliases
);
5442 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5444 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5446 SLAB_ATTR_RO(partial
);
5448 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5450 return show_slab_objects(s
, buf
, SO_CPU
);
5452 SLAB_ATTR_RO(cpu_slabs
);
5454 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5456 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5458 SLAB_ATTR_RO(objects
);
5460 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5462 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5464 SLAB_ATTR_RO(objects_partial
);
5466 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5473 for_each_online_cpu(cpu
) {
5476 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5479 pages
+= page
->pages
;
5480 objects
+= page
->pobjects
;
5484 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, pages
);
5487 for_each_online_cpu(cpu
) {
5490 page
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5492 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5493 cpu
, page
->pobjects
, page
->pages
);
5496 len
+= sysfs_emit_at(buf
, len
, "\n");
5500 SLAB_ATTR_RO(slabs_cpu_partial
);
5502 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5504 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5506 SLAB_ATTR_RO(reclaim_account
);
5508 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5510 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5512 SLAB_ATTR_RO(hwcache_align
);
5514 #ifdef CONFIG_ZONE_DMA
5515 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5517 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5519 SLAB_ATTR_RO(cache_dma
);
5522 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5524 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5526 SLAB_ATTR_RO(usersize
);
5528 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5530 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5532 SLAB_ATTR_RO(destroy_by_rcu
);
5534 #ifdef CONFIG_SLUB_DEBUG
5535 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5537 return show_slab_objects(s
, buf
, SO_ALL
);
5539 SLAB_ATTR_RO(slabs
);
5541 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5543 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5545 SLAB_ATTR_RO(total_objects
);
5547 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5549 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5551 SLAB_ATTR_RO(sanity_checks
);
5553 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5555 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5557 SLAB_ATTR_RO(trace
);
5559 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5561 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5564 SLAB_ATTR_RO(red_zone
);
5566 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5568 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5571 SLAB_ATTR_RO(poison
);
5573 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5575 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5578 SLAB_ATTR_RO(store_user
);
5580 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5585 static ssize_t
validate_store(struct kmem_cache
*s
,
5586 const char *buf
, size_t length
)
5590 if (buf
[0] == '1') {
5591 ret
= validate_slab_cache(s
);
5597 SLAB_ATTR(validate
);
5599 #endif /* CONFIG_SLUB_DEBUG */
5601 #ifdef CONFIG_FAILSLAB
5602 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5604 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5606 SLAB_ATTR_RO(failslab
);
5609 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5614 static ssize_t
shrink_store(struct kmem_cache
*s
,
5615 const char *buf
, size_t length
)
5618 kmem_cache_shrink(s
);
5626 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5628 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5631 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5632 const char *buf
, size_t length
)
5637 err
= kstrtouint(buf
, 10, &ratio
);
5643 s
->remote_node_defrag_ratio
= ratio
* 10;
5647 SLAB_ATTR(remote_node_defrag_ratio
);
5650 #ifdef CONFIG_SLUB_STATS
5651 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5653 unsigned long sum
= 0;
5656 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5661 for_each_online_cpu(cpu
) {
5662 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5668 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5671 for_each_online_cpu(cpu
) {
5673 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5678 len
+= sysfs_emit_at(buf
, len
, "\n");
5683 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5687 for_each_online_cpu(cpu
)
5688 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5691 #define STAT_ATTR(si, text) \
5692 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5694 return show_stat(s, buf, si); \
5696 static ssize_t text##_store(struct kmem_cache *s, \
5697 const char *buf, size_t length) \
5699 if (buf[0] != '0') \
5701 clear_stat(s, si); \
5706 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5707 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5708 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5709 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5710 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5711 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5712 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5713 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5714 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5715 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5716 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5717 STAT_ATTR(FREE_SLAB
, free_slab
);
5718 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5719 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5720 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5721 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5722 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5723 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5724 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5725 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5726 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5727 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5728 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5729 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5730 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5731 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5732 #endif /* CONFIG_SLUB_STATS */
5734 static struct attribute
*slab_attrs
[] = {
5735 &slab_size_attr
.attr
,
5736 &object_size_attr
.attr
,
5737 &objs_per_slab_attr
.attr
,
5739 &min_partial_attr
.attr
,
5740 &cpu_partial_attr
.attr
,
5742 &objects_partial_attr
.attr
,
5744 &cpu_slabs_attr
.attr
,
5748 &hwcache_align_attr
.attr
,
5749 &reclaim_account_attr
.attr
,
5750 &destroy_by_rcu_attr
.attr
,
5752 &slabs_cpu_partial_attr
.attr
,
5753 #ifdef CONFIG_SLUB_DEBUG
5754 &total_objects_attr
.attr
,
5756 &sanity_checks_attr
.attr
,
5758 &red_zone_attr
.attr
,
5760 &store_user_attr
.attr
,
5761 &validate_attr
.attr
,
5763 #ifdef CONFIG_ZONE_DMA
5764 &cache_dma_attr
.attr
,
5767 &remote_node_defrag_ratio_attr
.attr
,
5769 #ifdef CONFIG_SLUB_STATS
5770 &alloc_fastpath_attr
.attr
,
5771 &alloc_slowpath_attr
.attr
,
5772 &free_fastpath_attr
.attr
,
5773 &free_slowpath_attr
.attr
,
5774 &free_frozen_attr
.attr
,
5775 &free_add_partial_attr
.attr
,
5776 &free_remove_partial_attr
.attr
,
5777 &alloc_from_partial_attr
.attr
,
5778 &alloc_slab_attr
.attr
,
5779 &alloc_refill_attr
.attr
,
5780 &alloc_node_mismatch_attr
.attr
,
5781 &free_slab_attr
.attr
,
5782 &cpuslab_flush_attr
.attr
,
5783 &deactivate_full_attr
.attr
,
5784 &deactivate_empty_attr
.attr
,
5785 &deactivate_to_head_attr
.attr
,
5786 &deactivate_to_tail_attr
.attr
,
5787 &deactivate_remote_frees_attr
.attr
,
5788 &deactivate_bypass_attr
.attr
,
5789 &order_fallback_attr
.attr
,
5790 &cmpxchg_double_fail_attr
.attr
,
5791 &cmpxchg_double_cpu_fail_attr
.attr
,
5792 &cpu_partial_alloc_attr
.attr
,
5793 &cpu_partial_free_attr
.attr
,
5794 &cpu_partial_node_attr
.attr
,
5795 &cpu_partial_drain_attr
.attr
,
5797 #ifdef CONFIG_FAILSLAB
5798 &failslab_attr
.attr
,
5800 &usersize_attr
.attr
,
5805 static const struct attribute_group slab_attr_group
= {
5806 .attrs
= slab_attrs
,
5809 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5810 struct attribute
*attr
,
5813 struct slab_attribute
*attribute
;
5814 struct kmem_cache
*s
;
5817 attribute
= to_slab_attr(attr
);
5820 if (!attribute
->show
)
5823 err
= attribute
->show(s
, buf
);
5828 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5829 struct attribute
*attr
,
5830 const char *buf
, size_t len
)
5832 struct slab_attribute
*attribute
;
5833 struct kmem_cache
*s
;
5836 attribute
= to_slab_attr(attr
);
5839 if (!attribute
->store
)
5842 err
= attribute
->store(s
, buf
, len
);
5846 static void kmem_cache_release(struct kobject
*k
)
5848 slab_kmem_cache_release(to_slab(k
));
5851 static const struct sysfs_ops slab_sysfs_ops
= {
5852 .show
= slab_attr_show
,
5853 .store
= slab_attr_store
,
5856 static struct kobj_type slab_ktype
= {
5857 .sysfs_ops
= &slab_sysfs_ops
,
5858 .release
= kmem_cache_release
,
5861 static struct kset
*slab_kset
;
5863 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5868 #define ID_STR_LENGTH 64
5870 /* Create a unique string id for a slab cache:
5872 * Format :[flags-]size
5874 static char *create_unique_id(struct kmem_cache
*s
)
5876 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5880 return ERR_PTR(-ENOMEM
);
5884 * First flags affecting slabcache operations. We will only
5885 * get here for aliasable slabs so we do not need to support
5886 * too many flags. The flags here must cover all flags that
5887 * are matched during merging to guarantee that the id is
5890 if (s
->flags
& SLAB_CACHE_DMA
)
5892 if (s
->flags
& SLAB_CACHE_DMA32
)
5894 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5896 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5898 if (s
->flags
& SLAB_ACCOUNT
)
5902 p
+= sprintf(p
, "%07u", s
->size
);
5904 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5908 static int sysfs_slab_add(struct kmem_cache
*s
)
5912 struct kset
*kset
= cache_kset(s
);
5913 int unmergeable
= slab_unmergeable(s
);
5916 kobject_init(&s
->kobj
, &slab_ktype
);
5920 if (!unmergeable
&& disable_higher_order_debug
&&
5921 (slub_debug
& DEBUG_METADATA_FLAGS
))
5926 * Slabcache can never be merged so we can use the name proper.
5927 * This is typically the case for debug situations. In that
5928 * case we can catch duplicate names easily.
5930 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5934 * Create a unique name for the slab as a target
5937 name
= create_unique_id(s
);
5939 return PTR_ERR(name
);
5942 s
->kobj
.kset
= kset
;
5943 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5947 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5952 /* Setup first alias */
5953 sysfs_slab_alias(s
, s
->name
);
5960 kobject_del(&s
->kobj
);
5964 void sysfs_slab_unlink(struct kmem_cache
*s
)
5966 if (slab_state
>= FULL
)
5967 kobject_del(&s
->kobj
);
5970 void sysfs_slab_release(struct kmem_cache
*s
)
5972 if (slab_state
>= FULL
)
5973 kobject_put(&s
->kobj
);
5977 * Need to buffer aliases during bootup until sysfs becomes
5978 * available lest we lose that information.
5980 struct saved_alias
{
5981 struct kmem_cache
*s
;
5983 struct saved_alias
*next
;
5986 static struct saved_alias
*alias_list
;
5988 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
5990 struct saved_alias
*al
;
5992 if (slab_state
== FULL
) {
5994 * If we have a leftover link then remove it.
5996 sysfs_remove_link(&slab_kset
->kobj
, name
);
5997 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
6000 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
6006 al
->next
= alias_list
;
6011 static int __init
slab_sysfs_init(void)
6013 struct kmem_cache
*s
;
6016 mutex_lock(&slab_mutex
);
6018 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
6020 mutex_unlock(&slab_mutex
);
6021 pr_err("Cannot register slab subsystem.\n");
6027 list_for_each_entry(s
, &slab_caches
, list
) {
6028 err
= sysfs_slab_add(s
);
6030 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6034 while (alias_list
) {
6035 struct saved_alias
*al
= alias_list
;
6037 alias_list
= alias_list
->next
;
6038 err
= sysfs_slab_alias(al
->s
, al
->name
);
6040 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6045 mutex_unlock(&slab_mutex
);
6049 __initcall(slab_sysfs_init
);
6050 #endif /* CONFIG_SYSFS */
6052 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6053 static int slab_debugfs_show(struct seq_file
*seq
, void *v
)
6055 struct loc_track
*t
= seq
->private;
6059 idx
= (unsigned long) t
->idx
;
6060 if (idx
< t
->count
) {
6063 seq_printf(seq
, "%7ld ", l
->count
);
6066 seq_printf(seq
, "%pS", (void *)l
->addr
);
6068 seq_puts(seq
, "<not-available>");
6070 if (l
->sum_time
!= l
->min_time
) {
6071 seq_printf(seq
, " age=%ld/%llu/%ld",
6072 l
->min_time
, div_u64(l
->sum_time
, l
->count
),
6075 seq_printf(seq
, " age=%ld", l
->min_time
);
6077 if (l
->min_pid
!= l
->max_pid
)
6078 seq_printf(seq
, " pid=%ld-%ld", l
->min_pid
, l
->max_pid
);
6080 seq_printf(seq
, " pid=%ld",
6083 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l
->cpus
)))
6084 seq_printf(seq
, " cpus=%*pbl",
6085 cpumask_pr_args(to_cpumask(l
->cpus
)));
6087 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
6088 seq_printf(seq
, " nodes=%*pbl",
6089 nodemask_pr_args(&l
->nodes
));
6091 seq_puts(seq
, "\n");
6094 if (!idx
&& !t
->count
)
6095 seq_puts(seq
, "No data\n");
6100 static void slab_debugfs_stop(struct seq_file
*seq
, void *v
)
6104 static void *slab_debugfs_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
6106 struct loc_track
*t
= seq
->private;
6109 if (*ppos
<= t
->count
)
6115 static void *slab_debugfs_start(struct seq_file
*seq
, loff_t
*ppos
)
6117 struct loc_track
*t
= seq
->private;
6123 static const struct seq_operations slab_debugfs_sops
= {
6124 .start
= slab_debugfs_start
,
6125 .next
= slab_debugfs_next
,
6126 .stop
= slab_debugfs_stop
,
6127 .show
= slab_debugfs_show
,
6130 static int slab_debug_trace_open(struct inode
*inode
, struct file
*filep
)
6133 struct kmem_cache_node
*n
;
6134 enum track_item alloc
;
6136 struct loc_track
*t
= __seq_open_private(filep
, &slab_debugfs_sops
,
6137 sizeof(struct loc_track
));
6138 struct kmem_cache
*s
= file_inode(filep
)->i_private
;
6139 unsigned long *obj_map
;
6144 obj_map
= bitmap_alloc(oo_objects(s
->oo
), GFP_KERNEL
);
6146 seq_release_private(inode
, filep
);
6150 if (strcmp(filep
->f_path
.dentry
->d_name
.name
, "alloc_traces") == 0)
6151 alloc
= TRACK_ALLOC
;
6155 if (!alloc_loc_track(t
, PAGE_SIZE
/ sizeof(struct location
), GFP_KERNEL
)) {
6156 bitmap_free(obj_map
);
6157 seq_release_private(inode
, filep
);
6161 for_each_kmem_cache_node(s
, node
, n
) {
6162 unsigned long flags
;
6165 if (!atomic_long_read(&n
->nr_slabs
))
6168 spin_lock_irqsave(&n
->list_lock
, flags
);
6169 list_for_each_entry(page
, &n
->partial
, slab_list
)
6170 process_slab(t
, s
, page
, alloc
, obj_map
);
6171 list_for_each_entry(page
, &n
->full
, slab_list
)
6172 process_slab(t
, s
, page
, alloc
, obj_map
);
6173 spin_unlock_irqrestore(&n
->list_lock
, flags
);
6176 bitmap_free(obj_map
);
6180 static int slab_debug_trace_release(struct inode
*inode
, struct file
*file
)
6182 struct seq_file
*seq
= file
->private_data
;
6183 struct loc_track
*t
= seq
->private;
6186 return seq_release_private(inode
, file
);
6189 static const struct file_operations slab_debugfs_fops
= {
6190 .open
= slab_debug_trace_open
,
6192 .llseek
= seq_lseek
,
6193 .release
= slab_debug_trace_release
,
6196 static void debugfs_slab_add(struct kmem_cache
*s
)
6198 struct dentry
*slab_cache_dir
;
6200 if (unlikely(!slab_debugfs_root
))
6203 slab_cache_dir
= debugfs_create_dir(s
->name
, slab_debugfs_root
);
6205 debugfs_create_file("alloc_traces", 0400,
6206 slab_cache_dir
, s
, &slab_debugfs_fops
);
6208 debugfs_create_file("free_traces", 0400,
6209 slab_cache_dir
, s
, &slab_debugfs_fops
);
6212 void debugfs_slab_release(struct kmem_cache
*s
)
6214 debugfs_remove_recursive(debugfs_lookup(s
->name
, slab_debugfs_root
));
6217 static int __init
slab_debugfs_init(void)
6219 struct kmem_cache
*s
;
6221 slab_debugfs_root
= debugfs_create_dir("slab", NULL
);
6223 list_for_each_entry(s
, &slab_caches
, list
)
6224 if (s
->flags
& SLAB_STORE_USER
)
6225 debugfs_slab_add(s
);
6230 __initcall(slab_debugfs_init
);
6233 * The /proc/slabinfo ABI
6235 #ifdef CONFIG_SLUB_DEBUG
6236 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
6238 unsigned long nr_slabs
= 0;
6239 unsigned long nr_objs
= 0;
6240 unsigned long nr_free
= 0;
6242 struct kmem_cache_node
*n
;
6244 for_each_kmem_cache_node(s
, node
, n
) {
6245 nr_slabs
+= node_nr_slabs(n
);
6246 nr_objs
+= node_nr_objs(n
);
6247 nr_free
+= count_partial(n
, count_free
);
6250 sinfo
->active_objs
= nr_objs
- nr_free
;
6251 sinfo
->num_objs
= nr_objs
;
6252 sinfo
->active_slabs
= nr_slabs
;
6253 sinfo
->num_slabs
= nr_slabs
;
6254 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
6255 sinfo
->cache_order
= oo_order(s
->oo
);
6258 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
6262 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
6263 size_t count
, loff_t
*ppos
)
6267 #endif /* CONFIG_SLUB_DEBUG */