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/stackdepot.h>
30 #include <linux/debugobjects.h>
31 #include <linux/kallsyms.h>
32 #include <linux/kfence.h>
33 #include <linux/memory.h>
34 #include <linux/math64.h>
35 #include <linux/fault-inject.h>
36 #include <linux/stacktrace.h>
37 #include <linux/prefetch.h>
38 #include <linux/memcontrol.h>
39 #include <linux/random.h>
40 #include <kunit/test.h>
41 #include <linux/sort.h>
43 #include <linux/debugfs.h>
44 #include <trace/events/kmem.h>
50 * 1. slab_mutex (Global Mutex)
51 * 2. node->list_lock (Spinlock)
52 * 3. kmem_cache->cpu_slab->lock (Local lock)
53 * 4. slab_lock(slab) (Only on some arches or for debugging)
54 * 5. object_map_lock (Only for debugging)
58 * The role of the slab_mutex is to protect the list of all the slabs
59 * and to synchronize major metadata changes to slab cache structures.
60 * Also synchronizes memory hotplug callbacks.
64 * The slab_lock is a wrapper around the page lock, thus it is a bit
67 * The slab_lock is only used for debugging and on arches that do not
68 * have the ability to do a cmpxchg_double. It only protects:
69 * A. slab->freelist -> List of free objects in a slab
70 * B. slab->inuse -> Number of objects in use
71 * C. slab->objects -> Number of objects in slab
72 * D. slab->frozen -> frozen state
76 * If a slab is frozen then it is exempt from list management. It is not
77 * on any list except per cpu partial list. The processor that froze the
78 * slab is the one who can perform list operations on the slab. Other
79 * processors may put objects onto the freelist but the processor that
80 * froze the slab is the only one that can retrieve the objects from the
85 * The list_lock protects the partial and full list on each node and
86 * the partial slab counter. If taken then no new slabs may be added or
87 * removed from the lists nor make the number of partial slabs be modified.
88 * (Note that the total number of slabs is an atomic value that may be
89 * modified without taking the list lock).
91 * The list_lock is a centralized lock and thus we avoid taking it as
92 * much as possible. As long as SLUB does not have to handle partial
93 * slabs, operations can continue without any centralized lock. F.e.
94 * allocating a long series of objects that fill up slabs does not require
97 * cpu_slab->lock local lock
99 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
100 * except the stat counters. This is a percpu structure manipulated only by
101 * the local cpu, so the lock protects against being preempted or interrupted
102 * by an irq. Fast path operations rely on lockless operations instead.
103 * On PREEMPT_RT, the local lock does not actually disable irqs (and thus
104 * prevent the lockless operations), so fastpath operations also need to take
105 * the lock and are no longer lockless.
109 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
110 * are fully lockless when satisfied from the percpu slab (and when
111 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
112 * They also don't disable preemption or migration or irqs. They rely on
113 * the transaction id (tid) field to detect being preempted or moved to
116 * irq, preemption, migration considerations
118 * Interrupts are disabled as part of list_lock or local_lock operations, or
119 * around the slab_lock operation, in order to make the slab allocator safe
120 * to use in the context of an irq.
122 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
123 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
124 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
125 * doesn't have to be revalidated in each section protected by the local lock.
127 * SLUB assigns one slab for allocation to each processor.
128 * Allocations only occur from these slabs called cpu slabs.
130 * Slabs with free elements are kept on a partial list and during regular
131 * operations no list for full slabs is used. If an object in a full slab is
132 * freed then the slab will show up again on the partial lists.
133 * We track full slabs for debugging purposes though because otherwise we
134 * cannot scan all objects.
136 * Slabs are freed when they become empty. Teardown and setup is
137 * minimal so we rely on the page allocators per cpu caches for
138 * fast frees and allocs.
140 * slab->frozen The slab is frozen and exempt from list processing.
141 * This means that the slab is dedicated to a purpose
142 * such as satisfying allocations for a specific
143 * processor. Objects may be freed in the slab while
144 * it is frozen but slab_free will then skip the usual
145 * list operations. It is up to the processor holding
146 * the slab to integrate the slab into the slab lists
147 * when the slab is no longer needed.
149 * One use of this flag is to mark slabs that are
150 * used for allocations. Then such a slab becomes a cpu
151 * slab. The cpu slab may be equipped with an additional
152 * freelist that allows lockless access to
153 * free objects in addition to the regular freelist
154 * that requires the slab lock.
156 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
157 * options set. This moves slab handling out of
158 * the fast path and disables lockless freelists.
162 * We could simply use migrate_disable()/enable() but as long as it's a
163 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
165 #ifndef CONFIG_PREEMPT_RT
166 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
167 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
169 #define slub_get_cpu_ptr(var) \
174 #define slub_put_cpu_ptr(var) \
181 #ifdef CONFIG_SLUB_DEBUG
182 #ifdef CONFIG_SLUB_DEBUG_ON
183 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled
);
185 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled
);
187 #endif /* CONFIG_SLUB_DEBUG */
189 static inline bool kmem_cache_debug(struct kmem_cache
*s
)
191 return kmem_cache_debug_flags(s
, SLAB_DEBUG_FLAGS
);
194 void *fixup_red_left(struct kmem_cache
*s
, void *p
)
196 if (kmem_cache_debug_flags(s
, SLAB_RED_ZONE
))
197 p
+= s
->red_left_pad
;
202 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache
*s
)
204 #ifdef CONFIG_SLUB_CPU_PARTIAL
205 return !kmem_cache_debug(s
);
212 * Issues still to be resolved:
214 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
216 * - Variable sizing of the per node arrays
219 /* Enable to log cmpxchg failures */
220 #undef SLUB_DEBUG_CMPXCHG
223 * Minimum number of partial slabs. These will be left on the partial
224 * lists even if they are empty. kmem_cache_shrink may reclaim them.
226 #define MIN_PARTIAL 5
229 * Maximum number of desirable partial slabs.
230 * The existence of more partial slabs makes kmem_cache_shrink
231 * sort the partial list by the number of objects in use.
233 #define MAX_PARTIAL 10
235 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
236 SLAB_POISON | SLAB_STORE_USER)
239 * These debug flags cannot use CMPXCHG because there might be consistency
240 * issues when checking or reading debug information
242 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
247 * Debugging flags that require metadata to be stored in the slab. These get
248 * disabled when slub_debug=O is used and a cache's min order increases with
251 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
254 #define OO_MASK ((1 << OO_SHIFT) - 1)
255 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
257 /* Internal SLUB flags */
259 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
260 /* Use cmpxchg_double */
261 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
264 * Tracking user of a slab.
266 #define TRACK_ADDRS_COUNT 16
268 unsigned long addr
; /* Called from address */
269 #ifdef CONFIG_STACKDEPOT
270 depot_stack_handle_t handle
;
272 int cpu
; /* Was running on cpu */
273 int pid
; /* Pid context */
274 unsigned long when
; /* When did the operation occur */
277 enum track_item
{ TRACK_ALLOC
, TRACK_FREE
};
280 static int sysfs_slab_add(struct kmem_cache
*);
281 static int sysfs_slab_alias(struct kmem_cache
*, const char *);
283 static inline int sysfs_slab_add(struct kmem_cache
*s
) { return 0; }
284 static inline int sysfs_slab_alias(struct kmem_cache
*s
, const char *p
)
288 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
289 static void debugfs_slab_add(struct kmem_cache
*);
291 static inline void debugfs_slab_add(struct kmem_cache
*s
) { }
294 static inline void stat(const struct kmem_cache
*s
, enum stat_item si
)
296 #ifdef CONFIG_SLUB_STATS
298 * The rmw is racy on a preemptible kernel but this is acceptable, so
299 * avoid this_cpu_add()'s irq-disable overhead.
301 raw_cpu_inc(s
->cpu_slab
->stat
[si
]);
306 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
307 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
308 * differ during memory hotplug/hotremove operations.
309 * Protected by slab_mutex.
311 static nodemask_t slab_nodes
;
313 /********************************************************************
314 * Core slab cache functions
315 *******************************************************************/
318 * Returns freelist pointer (ptr). With hardening, this is obfuscated
319 * with an XOR of the address where the pointer is held and a per-cache
322 static inline void *freelist_ptr(const struct kmem_cache
*s
, void *ptr
,
323 unsigned long ptr_addr
)
325 #ifdef CONFIG_SLAB_FREELIST_HARDENED
327 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
328 * Normally, this doesn't cause any issues, as both set_freepointer()
329 * and get_freepointer() are called with a pointer with the same tag.
330 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
331 * example, when __free_slub() iterates over objects in a cache, it
332 * passes untagged pointers to check_object(). check_object() in turns
333 * calls get_freepointer() with an untagged pointer, which causes the
334 * freepointer to be restored incorrectly.
336 return (void *)((unsigned long)ptr
^ s
->random
^
337 swab((unsigned long)kasan_reset_tag((void *)ptr_addr
)));
343 /* Returns the freelist pointer recorded at location ptr_addr. */
344 static inline void *freelist_dereference(const struct kmem_cache
*s
,
347 return freelist_ptr(s
, (void *)*(unsigned long *)(ptr_addr
),
348 (unsigned long)ptr_addr
);
351 static inline void *get_freepointer(struct kmem_cache
*s
, void *object
)
353 object
= kasan_reset_tag(object
);
354 return freelist_dereference(s
, object
+ s
->offset
);
357 static void prefetch_freepointer(const struct kmem_cache
*s
, void *object
)
359 prefetchw(object
+ s
->offset
);
362 static inline void *get_freepointer_safe(struct kmem_cache
*s
, void *object
)
364 unsigned long freepointer_addr
;
367 if (!debug_pagealloc_enabled_static())
368 return get_freepointer(s
, object
);
370 object
= kasan_reset_tag(object
);
371 freepointer_addr
= (unsigned long)object
+ s
->offset
;
372 copy_from_kernel_nofault(&p
, (void **)freepointer_addr
, sizeof(p
));
373 return freelist_ptr(s
, p
, freepointer_addr
);
376 static inline void set_freepointer(struct kmem_cache
*s
, void *object
, void *fp
)
378 unsigned long freeptr_addr
= (unsigned long)object
+ s
->offset
;
380 #ifdef CONFIG_SLAB_FREELIST_HARDENED
381 BUG_ON(object
== fp
); /* naive detection of double free or corruption */
384 freeptr_addr
= (unsigned long)kasan_reset_tag((void *)freeptr_addr
);
385 *(void **)freeptr_addr
= freelist_ptr(s
, fp
, freeptr_addr
);
388 /* Loop over all objects in a slab */
389 #define for_each_object(__p, __s, __addr, __objects) \
390 for (__p = fixup_red_left(__s, __addr); \
391 __p < (__addr) + (__objects) * (__s)->size; \
394 static inline unsigned int order_objects(unsigned int order
, unsigned int size
)
396 return ((unsigned int)PAGE_SIZE
<< order
) / size
;
399 static inline struct kmem_cache_order_objects
oo_make(unsigned int order
,
402 struct kmem_cache_order_objects x
= {
403 (order
<< OO_SHIFT
) + order_objects(order
, size
)
409 static inline unsigned int oo_order(struct kmem_cache_order_objects x
)
411 return x
.x
>> OO_SHIFT
;
414 static inline unsigned int oo_objects(struct kmem_cache_order_objects x
)
416 return x
.x
& OO_MASK
;
419 #ifdef CONFIG_SLUB_CPU_PARTIAL
420 static void slub_set_cpu_partial(struct kmem_cache
*s
, unsigned int nr_objects
)
422 unsigned int nr_slabs
;
424 s
->cpu_partial
= nr_objects
;
427 * We take the number of objects but actually limit the number of
428 * slabs on the per cpu partial list, in order to limit excessive
429 * growth of the list. For simplicity we assume that the slabs will
432 nr_slabs
= DIV_ROUND_UP(nr_objects
* 2, oo_objects(s
->oo
));
433 s
->cpu_partial_slabs
= nr_slabs
;
437 slub_set_cpu_partial(struct kmem_cache
*s
, unsigned int nr_objects
)
440 #endif /* CONFIG_SLUB_CPU_PARTIAL */
443 * Per slab locking using the pagelock
445 static __always_inline
void __slab_lock(struct slab
*slab
)
447 struct page
*page
= slab_page(slab
);
449 VM_BUG_ON_PAGE(PageTail(page
), page
);
450 bit_spin_lock(PG_locked
, &page
->flags
);
453 static __always_inline
void __slab_unlock(struct slab
*slab
)
455 struct page
*page
= slab_page(slab
);
457 VM_BUG_ON_PAGE(PageTail(page
), page
);
458 __bit_spin_unlock(PG_locked
, &page
->flags
);
461 static __always_inline
void slab_lock(struct slab
*slab
, unsigned long *flags
)
463 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
464 local_irq_save(*flags
);
468 static __always_inline
void slab_unlock(struct slab
*slab
, unsigned long *flags
)
471 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
472 local_irq_restore(*flags
);
476 * Interrupts must be disabled (for the fallback code to work right), typically
477 * by an _irqsave() lock variant. Except on PREEMPT_RT where locks are different
478 * so we disable interrupts as part of slab_[un]lock().
480 static inline bool __cmpxchg_double_slab(struct kmem_cache
*s
, struct slab
*slab
,
481 void *freelist_old
, unsigned long counters_old
,
482 void *freelist_new
, unsigned long counters_new
,
485 if (!IS_ENABLED(CONFIG_PREEMPT_RT
))
486 lockdep_assert_irqs_disabled();
487 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
488 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
489 if (s
->flags
& __CMPXCHG_DOUBLE
) {
490 if (cmpxchg_double(&slab
->freelist
, &slab
->counters
,
491 freelist_old
, counters_old
,
492 freelist_new
, counters_new
))
497 /* init to 0 to prevent spurious warnings */
498 unsigned long flags
= 0;
500 slab_lock(slab
, &flags
);
501 if (slab
->freelist
== freelist_old
&&
502 slab
->counters
== counters_old
) {
503 slab
->freelist
= freelist_new
;
504 slab
->counters
= counters_new
;
505 slab_unlock(slab
, &flags
);
508 slab_unlock(slab
, &flags
);
512 stat(s
, CMPXCHG_DOUBLE_FAIL
);
514 #ifdef SLUB_DEBUG_CMPXCHG
515 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
521 static inline bool cmpxchg_double_slab(struct kmem_cache
*s
, struct slab
*slab
,
522 void *freelist_old
, unsigned long counters_old
,
523 void *freelist_new
, unsigned long counters_new
,
526 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
527 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
528 if (s
->flags
& __CMPXCHG_DOUBLE
) {
529 if (cmpxchg_double(&slab
->freelist
, &slab
->counters
,
530 freelist_old
, counters_old
,
531 freelist_new
, counters_new
))
538 local_irq_save(flags
);
540 if (slab
->freelist
== freelist_old
&&
541 slab
->counters
== counters_old
) {
542 slab
->freelist
= freelist_new
;
543 slab
->counters
= counters_new
;
545 local_irq_restore(flags
);
549 local_irq_restore(flags
);
553 stat(s
, CMPXCHG_DOUBLE_FAIL
);
555 #ifdef SLUB_DEBUG_CMPXCHG
556 pr_info("%s %s: cmpxchg double redo ", n
, s
->name
);
562 #ifdef CONFIG_SLUB_DEBUG
563 static unsigned long object_map
[BITS_TO_LONGS(MAX_OBJS_PER_PAGE
)];
564 static DEFINE_RAW_SPINLOCK(object_map_lock
);
566 static void __fill_map(unsigned long *obj_map
, struct kmem_cache
*s
,
569 void *addr
= slab_address(slab
);
572 bitmap_zero(obj_map
, slab
->objects
);
574 for (p
= slab
->freelist
; p
; p
= get_freepointer(s
, p
))
575 set_bit(__obj_to_index(s
, addr
, p
), obj_map
);
578 #if IS_ENABLED(CONFIG_KUNIT)
579 static bool slab_add_kunit_errors(void)
581 struct kunit_resource
*resource
;
583 if (likely(!current
->kunit_test
))
586 resource
= kunit_find_named_resource(current
->kunit_test
, "slab_errors");
590 (*(int *)resource
->data
)++;
591 kunit_put_resource(resource
);
595 static inline bool slab_add_kunit_errors(void) { return false; }
599 * Determine a map of objects in use in a slab.
601 * Node listlock must be held to guarantee that the slab does
602 * not vanish from under us.
604 static unsigned long *get_map(struct kmem_cache
*s
, struct slab
*slab
)
605 __acquires(&object_map_lock
)
607 VM_BUG_ON(!irqs_disabled());
609 raw_spin_lock(&object_map_lock
);
611 __fill_map(object_map
, s
, slab
);
616 static void put_map(unsigned long *map
) __releases(&object_map_lock
)
618 VM_BUG_ON(map
!= object_map
);
619 raw_spin_unlock(&object_map_lock
);
622 static inline unsigned int size_from_object(struct kmem_cache
*s
)
624 if (s
->flags
& SLAB_RED_ZONE
)
625 return s
->size
- s
->red_left_pad
;
630 static inline void *restore_red_left(struct kmem_cache
*s
, void *p
)
632 if (s
->flags
& SLAB_RED_ZONE
)
633 p
-= s
->red_left_pad
;
641 #if defined(CONFIG_SLUB_DEBUG_ON)
642 static slab_flags_t slub_debug
= DEBUG_DEFAULT_FLAGS
;
644 static slab_flags_t slub_debug
;
647 static char *slub_debug_string
;
648 static int disable_higher_order_debug
;
651 * slub is about to manipulate internal object metadata. This memory lies
652 * outside the range of the allocated object, so accessing it would normally
653 * be reported by kasan as a bounds error. metadata_access_enable() is used
654 * to tell kasan that these accesses are OK.
656 static inline void metadata_access_enable(void)
658 kasan_disable_current();
661 static inline void metadata_access_disable(void)
663 kasan_enable_current();
670 /* Verify that a pointer has an address that is valid within a slab page */
671 static inline int check_valid_pointer(struct kmem_cache
*s
,
672 struct slab
*slab
, void *object
)
679 base
= slab_address(slab
);
680 object
= kasan_reset_tag(object
);
681 object
= restore_red_left(s
, object
);
682 if (object
< base
|| object
>= base
+ slab
->objects
* s
->size
||
683 (object
- base
) % s
->size
) {
690 static void print_section(char *level
, char *text
, u8
*addr
,
693 metadata_access_enable();
694 print_hex_dump(level
, text
, DUMP_PREFIX_ADDRESS
,
695 16, 1, kasan_reset_tag((void *)addr
), length
, 1);
696 metadata_access_disable();
700 * See comment in calculate_sizes().
702 static inline bool freeptr_outside_object(struct kmem_cache
*s
)
704 return s
->offset
>= s
->inuse
;
708 * Return offset of the end of info block which is inuse + free pointer if
709 * not overlapping with object.
711 static inline unsigned int get_info_end(struct kmem_cache
*s
)
713 if (freeptr_outside_object(s
))
714 return s
->inuse
+ sizeof(void *);
719 static struct track
*get_track(struct kmem_cache
*s
, void *object
,
720 enum track_item alloc
)
724 p
= object
+ get_info_end(s
);
726 return kasan_reset_tag(p
+ alloc
);
729 #ifdef CONFIG_STACKDEPOT
730 static noinline depot_stack_handle_t
set_track_prepare(void)
732 depot_stack_handle_t handle
;
733 unsigned long entries
[TRACK_ADDRS_COUNT
];
734 unsigned int nr_entries
;
736 nr_entries
= stack_trace_save(entries
, ARRAY_SIZE(entries
), 3);
737 handle
= stack_depot_save(entries
, nr_entries
, GFP_NOWAIT
);
742 static inline depot_stack_handle_t
set_track_prepare(void)
748 static void set_track_update(struct kmem_cache
*s
, void *object
,
749 enum track_item alloc
, unsigned long addr
,
750 depot_stack_handle_t handle
)
752 struct track
*p
= get_track(s
, object
, alloc
);
754 #ifdef CONFIG_STACKDEPOT
758 p
->cpu
= smp_processor_id();
759 p
->pid
= current
->pid
;
763 static __always_inline
void set_track(struct kmem_cache
*s
, void *object
,
764 enum track_item alloc
, unsigned long addr
)
766 depot_stack_handle_t handle
= set_track_prepare();
768 set_track_update(s
, object
, alloc
, addr
, handle
);
771 static void init_tracking(struct kmem_cache
*s
, void *object
)
775 if (!(s
->flags
& SLAB_STORE_USER
))
778 p
= get_track(s
, object
, TRACK_ALLOC
);
779 memset(p
, 0, 2*sizeof(struct track
));
782 static void print_track(const char *s
, struct track
*t
, unsigned long pr_time
)
784 depot_stack_handle_t handle __maybe_unused
;
789 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
790 s
, (void *)t
->addr
, pr_time
- t
->when
, t
->cpu
, t
->pid
);
791 #ifdef CONFIG_STACKDEPOT
792 handle
= READ_ONCE(t
->handle
);
794 stack_depot_print(handle
);
796 pr_err("object allocation/free stack trace missing\n");
800 void print_tracking(struct kmem_cache
*s
, void *object
)
802 unsigned long pr_time
= jiffies
;
803 if (!(s
->flags
& SLAB_STORE_USER
))
806 print_track("Allocated", get_track(s
, object
, TRACK_ALLOC
), pr_time
);
807 print_track("Freed", get_track(s
, object
, TRACK_FREE
), pr_time
);
810 static void print_slab_info(const struct slab
*slab
)
812 struct folio
*folio
= (struct folio
*)slab_folio(slab
);
814 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
815 slab
, slab
->objects
, slab
->inuse
, slab
->freelist
,
816 folio_flags(folio
, 0));
819 static void slab_bug(struct kmem_cache
*s
, char *fmt
, ...)
821 struct va_format vaf
;
827 pr_err("=============================================================================\n");
828 pr_err("BUG %s (%s): %pV\n", s
->name
, print_tainted(), &vaf
);
829 pr_err("-----------------------------------------------------------------------------\n\n");
834 static void slab_fix(struct kmem_cache
*s
, char *fmt
, ...)
836 struct va_format vaf
;
839 if (slab_add_kunit_errors())
845 pr_err("FIX %s: %pV\n", s
->name
, &vaf
);
849 static void print_trailer(struct kmem_cache
*s
, struct slab
*slab
, u8
*p
)
851 unsigned int off
; /* Offset of last byte */
852 u8
*addr
= slab_address(slab
);
854 print_tracking(s
, p
);
856 print_slab_info(slab
);
858 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
859 p
, p
- addr
, get_freepointer(s
, p
));
861 if (s
->flags
& SLAB_RED_ZONE
)
862 print_section(KERN_ERR
, "Redzone ", p
- s
->red_left_pad
,
864 else if (p
> addr
+ 16)
865 print_section(KERN_ERR
, "Bytes b4 ", p
- 16, 16);
867 print_section(KERN_ERR
, "Object ", p
,
868 min_t(unsigned int, s
->object_size
, PAGE_SIZE
));
869 if (s
->flags
& SLAB_RED_ZONE
)
870 print_section(KERN_ERR
, "Redzone ", p
+ s
->object_size
,
871 s
->inuse
- s
->object_size
);
873 off
= get_info_end(s
);
875 if (s
->flags
& SLAB_STORE_USER
)
876 off
+= 2 * sizeof(struct track
);
878 off
+= kasan_metadata_size(s
);
880 if (off
!= size_from_object(s
))
881 /* Beginning of the filler is the free pointer */
882 print_section(KERN_ERR
, "Padding ", p
+ off
,
883 size_from_object(s
) - off
);
888 static void object_err(struct kmem_cache
*s
, struct slab
*slab
,
889 u8
*object
, char *reason
)
891 if (slab_add_kunit_errors())
894 slab_bug(s
, "%s", reason
);
895 print_trailer(s
, slab
, object
);
896 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
899 static bool freelist_corrupted(struct kmem_cache
*s
, struct slab
*slab
,
900 void **freelist
, void *nextfree
)
902 if ((s
->flags
& SLAB_CONSISTENCY_CHECKS
) &&
903 !check_valid_pointer(s
, slab
, nextfree
) && freelist
) {
904 object_err(s
, slab
, *freelist
, "Freechain corrupt");
906 slab_fix(s
, "Isolate corrupted freechain");
913 static __printf(3, 4) void slab_err(struct kmem_cache
*s
, struct slab
*slab
,
914 const char *fmt
, ...)
919 if (slab_add_kunit_errors())
923 vsnprintf(buf
, sizeof(buf
), fmt
, args
);
925 slab_bug(s
, "%s", buf
);
926 print_slab_info(slab
);
928 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
931 static void init_object(struct kmem_cache
*s
, void *object
, u8 val
)
933 u8
*p
= kasan_reset_tag(object
);
935 if (s
->flags
& SLAB_RED_ZONE
)
936 memset(p
- s
->red_left_pad
, val
, s
->red_left_pad
);
938 if (s
->flags
& __OBJECT_POISON
) {
939 memset(p
, POISON_FREE
, s
->object_size
- 1);
940 p
[s
->object_size
- 1] = POISON_END
;
943 if (s
->flags
& SLAB_RED_ZONE
)
944 memset(p
+ s
->object_size
, val
, s
->inuse
- s
->object_size
);
947 static void restore_bytes(struct kmem_cache
*s
, char *message
, u8 data
,
948 void *from
, void *to
)
950 slab_fix(s
, "Restoring %s 0x%p-0x%p=0x%x", message
, from
, to
- 1, data
);
951 memset(from
, data
, to
- from
);
954 static int check_bytes_and_report(struct kmem_cache
*s
, struct slab
*slab
,
955 u8
*object
, char *what
,
956 u8
*start
, unsigned int value
, unsigned int bytes
)
960 u8
*addr
= slab_address(slab
);
962 metadata_access_enable();
963 fault
= memchr_inv(kasan_reset_tag(start
), value
, bytes
);
964 metadata_access_disable();
969 while (end
> fault
&& end
[-1] == value
)
972 if (slab_add_kunit_errors())
975 slab_bug(s
, "%s overwritten", what
);
976 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
977 fault
, end
- 1, fault
- addr
,
979 print_trailer(s
, slab
, object
);
980 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
983 restore_bytes(s
, what
, value
, fault
, end
);
991 * Bytes of the object to be managed.
992 * If the freepointer may overlay the object then the free
993 * pointer is at the middle of the object.
995 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
998 * object + s->object_size
999 * Padding to reach word boundary. This is also used for Redzoning.
1000 * Padding is extended by another word if Redzoning is enabled and
1001 * object_size == inuse.
1003 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1004 * 0xcc (RED_ACTIVE) for objects in use.
1007 * Meta data starts here.
1009 * A. Free pointer (if we cannot overwrite object on free)
1010 * B. Tracking data for SLAB_STORE_USER
1011 * C. Padding to reach required alignment boundary or at minimum
1012 * one word if debugging is on to be able to detect writes
1013 * before the word boundary.
1015 * Padding is done using 0x5a (POISON_INUSE)
1018 * Nothing is used beyond s->size.
1020 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1021 * ignored. And therefore no slab options that rely on these boundaries
1022 * may be used with merged slabcaches.
1025 static int check_pad_bytes(struct kmem_cache
*s
, struct slab
*slab
, u8
*p
)
1027 unsigned long off
= get_info_end(s
); /* The end of info */
1029 if (s
->flags
& SLAB_STORE_USER
)
1030 /* We also have user information there */
1031 off
+= 2 * sizeof(struct track
);
1033 off
+= kasan_metadata_size(s
);
1035 if (size_from_object(s
) == off
)
1038 return check_bytes_and_report(s
, slab
, p
, "Object padding",
1039 p
+ off
, POISON_INUSE
, size_from_object(s
) - off
);
1042 /* Check the pad bytes at the end of a slab page */
1043 static void slab_pad_check(struct kmem_cache
*s
, struct slab
*slab
)
1052 if (!(s
->flags
& SLAB_POISON
))
1055 start
= slab_address(slab
);
1056 length
= slab_size(slab
);
1057 end
= start
+ length
;
1058 remainder
= length
% s
->size
;
1062 pad
= end
- remainder
;
1063 metadata_access_enable();
1064 fault
= memchr_inv(kasan_reset_tag(pad
), POISON_INUSE
, remainder
);
1065 metadata_access_disable();
1068 while (end
> fault
&& end
[-1] == POISON_INUSE
)
1071 slab_err(s
, slab
, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1072 fault
, end
- 1, fault
- start
);
1073 print_section(KERN_ERR
, "Padding ", pad
, remainder
);
1075 restore_bytes(s
, "slab padding", POISON_INUSE
, fault
, end
);
1078 static int check_object(struct kmem_cache
*s
, struct slab
*slab
,
1079 void *object
, u8 val
)
1082 u8
*endobject
= object
+ s
->object_size
;
1084 if (s
->flags
& SLAB_RED_ZONE
) {
1085 if (!check_bytes_and_report(s
, slab
, object
, "Left Redzone",
1086 object
- s
->red_left_pad
, val
, s
->red_left_pad
))
1089 if (!check_bytes_and_report(s
, slab
, object
, "Right Redzone",
1090 endobject
, val
, s
->inuse
- s
->object_size
))
1093 if ((s
->flags
& SLAB_POISON
) && s
->object_size
< s
->inuse
) {
1094 check_bytes_and_report(s
, slab
, p
, "Alignment padding",
1095 endobject
, POISON_INUSE
,
1096 s
->inuse
- s
->object_size
);
1100 if (s
->flags
& SLAB_POISON
) {
1101 if (val
!= SLUB_RED_ACTIVE
&& (s
->flags
& __OBJECT_POISON
) &&
1102 (!check_bytes_and_report(s
, slab
, p
, "Poison", p
,
1103 POISON_FREE
, s
->object_size
- 1) ||
1104 !check_bytes_and_report(s
, slab
, p
, "End Poison",
1105 p
+ s
->object_size
- 1, POISON_END
, 1)))
1108 * check_pad_bytes cleans up on its own.
1110 check_pad_bytes(s
, slab
, p
);
1113 if (!freeptr_outside_object(s
) && val
== SLUB_RED_ACTIVE
)
1115 * Object and freepointer overlap. Cannot check
1116 * freepointer while object is allocated.
1120 /* Check free pointer validity */
1121 if (!check_valid_pointer(s
, slab
, get_freepointer(s
, p
))) {
1122 object_err(s
, slab
, p
, "Freepointer corrupt");
1124 * No choice but to zap it and thus lose the remainder
1125 * of the free objects in this slab. May cause
1126 * another error because the object count is now wrong.
1128 set_freepointer(s
, p
, NULL
);
1134 static int check_slab(struct kmem_cache
*s
, struct slab
*slab
)
1138 if (!folio_test_slab(slab_folio(slab
))) {
1139 slab_err(s
, slab
, "Not a valid slab page");
1143 maxobj
= order_objects(slab_order(slab
), s
->size
);
1144 if (slab
->objects
> maxobj
) {
1145 slab_err(s
, slab
, "objects %u > max %u",
1146 slab
->objects
, maxobj
);
1149 if (slab
->inuse
> slab
->objects
) {
1150 slab_err(s
, slab
, "inuse %u > max %u",
1151 slab
->inuse
, slab
->objects
);
1154 /* Slab_pad_check fixes things up after itself */
1155 slab_pad_check(s
, slab
);
1160 * Determine if a certain object in a slab is on the freelist. Must hold the
1161 * slab lock to guarantee that the chains are in a consistent state.
1163 static int on_freelist(struct kmem_cache
*s
, struct slab
*slab
, void *search
)
1167 void *object
= NULL
;
1170 fp
= slab
->freelist
;
1171 while (fp
&& nr
<= slab
->objects
) {
1174 if (!check_valid_pointer(s
, slab
, fp
)) {
1176 object_err(s
, slab
, object
,
1177 "Freechain corrupt");
1178 set_freepointer(s
, object
, NULL
);
1180 slab_err(s
, slab
, "Freepointer corrupt");
1181 slab
->freelist
= NULL
;
1182 slab
->inuse
= slab
->objects
;
1183 slab_fix(s
, "Freelist cleared");
1189 fp
= get_freepointer(s
, object
);
1193 max_objects
= order_objects(slab_order(slab
), s
->size
);
1194 if (max_objects
> MAX_OBJS_PER_PAGE
)
1195 max_objects
= MAX_OBJS_PER_PAGE
;
1197 if (slab
->objects
!= max_objects
) {
1198 slab_err(s
, slab
, "Wrong number of objects. Found %d but should be %d",
1199 slab
->objects
, max_objects
);
1200 slab
->objects
= max_objects
;
1201 slab_fix(s
, "Number of objects adjusted");
1203 if (slab
->inuse
!= slab
->objects
- nr
) {
1204 slab_err(s
, slab
, "Wrong object count. Counter is %d but counted were %d",
1205 slab
->inuse
, slab
->objects
- nr
);
1206 slab
->inuse
= slab
->objects
- nr
;
1207 slab_fix(s
, "Object count adjusted");
1209 return search
== NULL
;
1212 static void trace(struct kmem_cache
*s
, struct slab
*slab
, void *object
,
1215 if (s
->flags
& SLAB_TRACE
) {
1216 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1218 alloc
? "alloc" : "free",
1219 object
, slab
->inuse
,
1223 print_section(KERN_INFO
, "Object ", (void *)object
,
1231 * Tracking of fully allocated slabs for debugging purposes.
1233 static void add_full(struct kmem_cache
*s
,
1234 struct kmem_cache_node
*n
, struct slab
*slab
)
1236 if (!(s
->flags
& SLAB_STORE_USER
))
1239 lockdep_assert_held(&n
->list_lock
);
1240 list_add(&slab
->slab_list
, &n
->full
);
1243 static void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
, struct slab
*slab
)
1245 if (!(s
->flags
& SLAB_STORE_USER
))
1248 lockdep_assert_held(&n
->list_lock
);
1249 list_del(&slab
->slab_list
);
1252 /* Tracking of the number of slabs for debugging purposes */
1253 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1255 struct kmem_cache_node
*n
= get_node(s
, node
);
1257 return atomic_long_read(&n
->nr_slabs
);
1260 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1262 return atomic_long_read(&n
->nr_slabs
);
1265 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1267 struct kmem_cache_node
*n
= get_node(s
, node
);
1270 * May be called early in order to allocate a slab for the
1271 * kmem_cache_node structure. Solve the chicken-egg
1272 * dilemma by deferring the increment of the count during
1273 * bootstrap (see early_kmem_cache_node_alloc).
1276 atomic_long_inc(&n
->nr_slabs
);
1277 atomic_long_add(objects
, &n
->total_objects
);
1280 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
, int objects
)
1282 struct kmem_cache_node
*n
= get_node(s
, node
);
1284 atomic_long_dec(&n
->nr_slabs
);
1285 atomic_long_sub(objects
, &n
->total_objects
);
1288 /* Object debug checks for alloc/free paths */
1289 static void setup_object_debug(struct kmem_cache
*s
, void *object
)
1291 if (!kmem_cache_debug_flags(s
, SLAB_STORE_USER
|SLAB_RED_ZONE
|__OBJECT_POISON
))
1294 init_object(s
, object
, SLUB_RED_INACTIVE
);
1295 init_tracking(s
, object
);
1299 void setup_slab_debug(struct kmem_cache
*s
, struct slab
*slab
, void *addr
)
1301 if (!kmem_cache_debug_flags(s
, SLAB_POISON
))
1304 metadata_access_enable();
1305 memset(kasan_reset_tag(addr
), POISON_INUSE
, slab_size(slab
));
1306 metadata_access_disable();
1309 static inline int alloc_consistency_checks(struct kmem_cache
*s
,
1310 struct slab
*slab
, void *object
)
1312 if (!check_slab(s
, slab
))
1315 if (!check_valid_pointer(s
, slab
, object
)) {
1316 object_err(s
, slab
, object
, "Freelist Pointer check fails");
1320 if (!check_object(s
, slab
, object
, SLUB_RED_INACTIVE
))
1326 static noinline
int alloc_debug_processing(struct kmem_cache
*s
,
1328 void *object
, unsigned long addr
)
1330 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1331 if (!alloc_consistency_checks(s
, slab
, object
))
1335 /* Success perform special debug activities for allocs */
1336 if (s
->flags
& SLAB_STORE_USER
)
1337 set_track(s
, object
, TRACK_ALLOC
, addr
);
1338 trace(s
, slab
, object
, 1);
1339 init_object(s
, object
, SLUB_RED_ACTIVE
);
1343 if (folio_test_slab(slab_folio(slab
))) {
1345 * If this is a slab page then lets do the best we can
1346 * to avoid issues in the future. Marking all objects
1347 * as used avoids touching the remaining objects.
1349 slab_fix(s
, "Marking all objects used");
1350 slab
->inuse
= slab
->objects
;
1351 slab
->freelist
= NULL
;
1356 static inline int free_consistency_checks(struct kmem_cache
*s
,
1357 struct slab
*slab
, void *object
, unsigned long addr
)
1359 if (!check_valid_pointer(s
, slab
, object
)) {
1360 slab_err(s
, slab
, "Invalid object pointer 0x%p", object
);
1364 if (on_freelist(s
, slab
, object
)) {
1365 object_err(s
, slab
, object
, "Object already free");
1369 if (!check_object(s
, slab
, object
, SLUB_RED_ACTIVE
))
1372 if (unlikely(s
!= slab
->slab_cache
)) {
1373 if (!folio_test_slab(slab_folio(slab
))) {
1374 slab_err(s
, slab
, "Attempt to free object(0x%p) outside of slab",
1376 } else if (!slab
->slab_cache
) {
1377 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1381 object_err(s
, slab
, object
,
1382 "page slab pointer corrupt.");
1388 /* Supports checking bulk free of a constructed freelist */
1389 static noinline
int free_debug_processing(
1390 struct kmem_cache
*s
, struct slab
*slab
,
1391 void *head
, void *tail
, int bulk_cnt
,
1394 struct kmem_cache_node
*n
= get_node(s
, slab_nid(slab
));
1395 void *object
= head
;
1397 unsigned long flags
, flags2
;
1399 depot_stack_handle_t handle
= 0;
1401 if (s
->flags
& SLAB_STORE_USER
)
1402 handle
= set_track_prepare();
1404 spin_lock_irqsave(&n
->list_lock
, flags
);
1405 slab_lock(slab
, &flags2
);
1407 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1408 if (!check_slab(s
, slab
))
1415 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
) {
1416 if (!free_consistency_checks(s
, slab
, object
, addr
))
1420 if (s
->flags
& SLAB_STORE_USER
)
1421 set_track_update(s
, object
, TRACK_FREE
, addr
, handle
);
1422 trace(s
, slab
, object
, 0);
1423 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1424 init_object(s
, object
, SLUB_RED_INACTIVE
);
1426 /* Reached end of constructed freelist yet? */
1427 if (object
!= tail
) {
1428 object
= get_freepointer(s
, object
);
1434 if (cnt
!= bulk_cnt
)
1435 slab_err(s
, slab
, "Bulk freelist count(%d) invalid(%d)\n",
1438 slab_unlock(slab
, &flags2
);
1439 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1441 slab_fix(s
, "Object at 0x%p not freed", object
);
1446 * Parse a block of slub_debug options. Blocks are delimited by ';'
1448 * @str: start of block
1449 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1450 * @slabs: return start of list of slabs, or NULL when there's no list
1451 * @init: assume this is initial parsing and not per-kmem-create parsing
1453 * returns the start of next block if there's any, or NULL
1456 parse_slub_debug_flags(char *str
, slab_flags_t
*flags
, char **slabs
, bool init
)
1458 bool higher_order_disable
= false;
1460 /* Skip any completely empty blocks */
1461 while (*str
&& *str
== ';')
1466 * No options but restriction on slabs. This means full
1467 * debugging for slabs matching a pattern.
1469 *flags
= DEBUG_DEFAULT_FLAGS
;
1474 /* Determine which debug features should be switched on */
1475 for (; *str
&& *str
!= ',' && *str
!= ';'; str
++) {
1476 switch (tolower(*str
)) {
1481 *flags
|= SLAB_CONSISTENCY_CHECKS
;
1484 *flags
|= SLAB_RED_ZONE
;
1487 *flags
|= SLAB_POISON
;
1490 *flags
|= SLAB_STORE_USER
;
1493 *flags
|= SLAB_TRACE
;
1496 *flags
|= SLAB_FAILSLAB
;
1500 * Avoid enabling debugging on caches if its minimum
1501 * order would increase as a result.
1503 higher_order_disable
= true;
1507 pr_err("slub_debug option '%c' unknown. skipped\n", *str
);
1516 /* Skip over the slab list */
1517 while (*str
&& *str
!= ';')
1520 /* Skip any completely empty blocks */
1521 while (*str
&& *str
== ';')
1524 if (init
&& higher_order_disable
)
1525 disable_higher_order_debug
= 1;
1533 static int __init
setup_slub_debug(char *str
)
1536 slab_flags_t global_flags
;
1539 bool global_slub_debug_changed
= false;
1540 bool slab_list_specified
= false;
1542 global_flags
= DEBUG_DEFAULT_FLAGS
;
1543 if (*str
++ != '=' || !*str
)
1545 * No options specified. Switch on full debugging.
1551 str
= parse_slub_debug_flags(str
, &flags
, &slab_list
, true);
1554 global_flags
= flags
;
1555 global_slub_debug_changed
= true;
1557 slab_list_specified
= true;
1558 if (flags
& SLAB_STORE_USER
)
1559 stack_depot_want_early_init();
1564 * For backwards compatibility, a single list of flags with list of
1565 * slabs means debugging is only changed for those slabs, so the global
1566 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1567 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1568 * long as there is no option specifying flags without a slab list.
1570 if (slab_list_specified
) {
1571 if (!global_slub_debug_changed
)
1572 global_flags
= slub_debug
;
1573 slub_debug_string
= saved_str
;
1576 slub_debug
= global_flags
;
1577 if (slub_debug
& SLAB_STORE_USER
)
1578 stack_depot_want_early_init();
1579 if (slub_debug
!= 0 || slub_debug_string
)
1580 static_branch_enable(&slub_debug_enabled
);
1582 static_branch_disable(&slub_debug_enabled
);
1583 if ((static_branch_unlikely(&init_on_alloc
) ||
1584 static_branch_unlikely(&init_on_free
)) &&
1585 (slub_debug
& SLAB_POISON
))
1586 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1590 __setup("slub_debug", setup_slub_debug
);
1593 * kmem_cache_flags - apply debugging options to the cache
1594 * @object_size: the size of an object without meta data
1595 * @flags: flags to set
1596 * @name: name of the cache
1598 * Debug option(s) are applied to @flags. In addition to the debug
1599 * option(s), if a slab name (or multiple) is specified i.e.
1600 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1601 * then only the select slabs will receive the debug option(s).
1603 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1604 slab_flags_t flags
, const char *name
)
1609 slab_flags_t block_flags
;
1610 slab_flags_t slub_debug_local
= slub_debug
;
1612 if (flags
& SLAB_NO_USER_FLAGS
)
1616 * If the slab cache is for debugging (e.g. kmemleak) then
1617 * don't store user (stack trace) information by default,
1618 * but let the user enable it via the command line below.
1620 if (flags
& SLAB_NOLEAKTRACE
)
1621 slub_debug_local
&= ~SLAB_STORE_USER
;
1624 next_block
= slub_debug_string
;
1625 /* Go through all blocks of debug options, see if any matches our slab's name */
1626 while (next_block
) {
1627 next_block
= parse_slub_debug_flags(next_block
, &block_flags
, &iter
, false);
1630 /* Found a block that has a slab list, search it */
1635 end
= strchrnul(iter
, ',');
1636 if (next_block
&& next_block
< end
)
1637 end
= next_block
- 1;
1639 glob
= strnchr(iter
, end
- iter
, '*');
1641 cmplen
= glob
- iter
;
1643 cmplen
= max_t(size_t, len
, (end
- iter
));
1645 if (!strncmp(name
, iter
, cmplen
)) {
1646 flags
|= block_flags
;
1650 if (!*end
|| *end
== ';')
1656 return flags
| slub_debug_local
;
1658 #else /* !CONFIG_SLUB_DEBUG */
1659 static inline void setup_object_debug(struct kmem_cache
*s
, void *object
) {}
1661 void setup_slab_debug(struct kmem_cache
*s
, struct slab
*slab
, void *addr
) {}
1663 static inline int alloc_debug_processing(struct kmem_cache
*s
,
1664 struct slab
*slab
, void *object
, unsigned long addr
) { return 0; }
1666 static inline int free_debug_processing(
1667 struct kmem_cache
*s
, struct slab
*slab
,
1668 void *head
, void *tail
, int bulk_cnt
,
1669 unsigned long addr
) { return 0; }
1671 static inline void slab_pad_check(struct kmem_cache
*s
, struct slab
*slab
) {}
1672 static inline int check_object(struct kmem_cache
*s
, struct slab
*slab
,
1673 void *object
, u8 val
) { return 1; }
1674 static inline void add_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1675 struct slab
*slab
) {}
1676 static inline void remove_full(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
1677 struct slab
*slab
) {}
1678 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1679 slab_flags_t flags
, const char *name
)
1683 #define slub_debug 0
1685 #define disable_higher_order_debug 0
1687 static inline unsigned long slabs_node(struct kmem_cache
*s
, int node
)
1689 static inline unsigned long node_nr_slabs(struct kmem_cache_node
*n
)
1691 static inline void inc_slabs_node(struct kmem_cache
*s
, int node
,
1693 static inline void dec_slabs_node(struct kmem_cache
*s
, int node
,
1696 static bool freelist_corrupted(struct kmem_cache
*s
, struct slab
*slab
,
1697 void **freelist
, void *nextfree
)
1701 #endif /* CONFIG_SLUB_DEBUG */
1704 * Hooks for other subsystems that check memory allocations. In a typical
1705 * production configuration these hooks all should produce no code at all.
1707 static inline void *kmalloc_large_node_hook(void *ptr
, size_t size
, gfp_t flags
)
1709 ptr
= kasan_kmalloc_large(ptr
, size
, flags
);
1710 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1711 kmemleak_alloc(ptr
, size
, 1, flags
);
1715 static __always_inline
void kfree_hook(void *x
)
1718 kasan_kfree_large(x
);
1721 static __always_inline
bool slab_free_hook(struct kmem_cache
*s
,
1724 kmemleak_free_recursive(x
, s
->flags
);
1726 debug_check_no_locks_freed(x
, s
->object_size
);
1728 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
1729 debug_check_no_obj_freed(x
, s
->object_size
);
1731 /* Use KCSAN to help debug racy use-after-free. */
1732 if (!(s
->flags
& SLAB_TYPESAFE_BY_RCU
))
1733 __kcsan_check_access(x
, s
->object_size
,
1734 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
1737 * As memory initialization might be integrated into KASAN,
1738 * kasan_slab_free and initialization memset's must be
1739 * kept together to avoid discrepancies in behavior.
1741 * The initialization memset's clear the object and the metadata,
1742 * but don't touch the SLAB redzone.
1747 if (!kasan_has_integrated_init())
1748 memset(kasan_reset_tag(x
), 0, s
->object_size
);
1749 rsize
= (s
->flags
& SLAB_RED_ZONE
) ? s
->red_left_pad
: 0;
1750 memset((char *)kasan_reset_tag(x
) + s
->inuse
, 0,
1751 s
->size
- s
->inuse
- rsize
);
1753 /* KASAN might put x into memory quarantine, delaying its reuse. */
1754 return kasan_slab_free(s
, x
, init
);
1757 static inline bool slab_free_freelist_hook(struct kmem_cache
*s
,
1758 void **head
, void **tail
,
1764 void *old_tail
= *tail
? *tail
: *head
;
1766 if (is_kfence_address(next
)) {
1767 slab_free_hook(s
, next
, false);
1771 /* Head and tail of the reconstructed freelist */
1777 next
= get_freepointer(s
, object
);
1779 /* If object's reuse doesn't have to be delayed */
1780 if (!slab_free_hook(s
, object
, slab_want_init_on_free(s
))) {
1781 /* Move object to the new freelist */
1782 set_freepointer(s
, object
, *head
);
1788 * Adjust the reconstructed freelist depth
1789 * accordingly if object's reuse is delayed.
1793 } while (object
!= old_tail
);
1798 return *head
!= NULL
;
1801 static void *setup_object(struct kmem_cache
*s
, void *object
)
1803 setup_object_debug(s
, object
);
1804 object
= kasan_init_slab_obj(s
, object
);
1805 if (unlikely(s
->ctor
)) {
1806 kasan_unpoison_object_data(s
, object
);
1808 kasan_poison_object_data(s
, object
);
1814 * Slab allocation and freeing
1816 static inline struct slab
*alloc_slab_page(gfp_t flags
, int node
,
1817 struct kmem_cache_order_objects oo
)
1819 struct folio
*folio
;
1821 unsigned int order
= oo_order(oo
);
1823 if (node
== NUMA_NO_NODE
)
1824 folio
= (struct folio
*)alloc_pages(flags
, order
);
1826 folio
= (struct folio
*)__alloc_pages_node(node
, flags
, order
);
1831 slab
= folio_slab(folio
);
1832 __folio_set_slab(folio
);
1833 if (page_is_pfmemalloc(folio_page(folio
, 0)))
1834 slab_set_pfmemalloc(slab
);
1839 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1840 /* Pre-initialize the random sequence cache */
1841 static int init_cache_random_seq(struct kmem_cache
*s
)
1843 unsigned int count
= oo_objects(s
->oo
);
1846 /* Bailout if already initialised */
1850 err
= cache_random_seq_create(s
, count
, GFP_KERNEL
);
1852 pr_err("SLUB: Unable to initialize free list for %s\n",
1857 /* Transform to an offset on the set of pages */
1858 if (s
->random_seq
) {
1861 for (i
= 0; i
< count
; i
++)
1862 s
->random_seq
[i
] *= s
->size
;
1867 /* Initialize each random sequence freelist per cache */
1868 static void __init
init_freelist_randomization(void)
1870 struct kmem_cache
*s
;
1872 mutex_lock(&slab_mutex
);
1874 list_for_each_entry(s
, &slab_caches
, list
)
1875 init_cache_random_seq(s
);
1877 mutex_unlock(&slab_mutex
);
1880 /* Get the next entry on the pre-computed freelist randomized */
1881 static void *next_freelist_entry(struct kmem_cache
*s
, struct slab
*slab
,
1882 unsigned long *pos
, void *start
,
1883 unsigned long page_limit
,
1884 unsigned long freelist_count
)
1889 * If the target page allocation failed, the number of objects on the
1890 * page might be smaller than the usual size defined by the cache.
1893 idx
= s
->random_seq
[*pos
];
1895 if (*pos
>= freelist_count
)
1897 } while (unlikely(idx
>= page_limit
));
1899 return (char *)start
+ idx
;
1902 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1903 static bool shuffle_freelist(struct kmem_cache
*s
, struct slab
*slab
)
1908 unsigned long idx
, pos
, page_limit
, freelist_count
;
1910 if (slab
->objects
< 2 || !s
->random_seq
)
1913 freelist_count
= oo_objects(s
->oo
);
1914 pos
= get_random_int() % freelist_count
;
1916 page_limit
= slab
->objects
* s
->size
;
1917 start
= fixup_red_left(s
, slab_address(slab
));
1919 /* First entry is used as the base of the freelist */
1920 cur
= next_freelist_entry(s
, slab
, &pos
, start
, page_limit
,
1922 cur
= setup_object(s
, cur
);
1923 slab
->freelist
= cur
;
1925 for (idx
= 1; idx
< slab
->objects
; idx
++) {
1926 next
= next_freelist_entry(s
, slab
, &pos
, start
, page_limit
,
1928 next
= setup_object(s
, next
);
1929 set_freepointer(s
, cur
, next
);
1932 set_freepointer(s
, cur
, NULL
);
1937 static inline int init_cache_random_seq(struct kmem_cache
*s
)
1941 static inline void init_freelist_randomization(void) { }
1942 static inline bool shuffle_freelist(struct kmem_cache
*s
, struct slab
*slab
)
1946 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1948 static struct slab
*allocate_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
1951 struct kmem_cache_order_objects oo
= s
->oo
;
1953 void *start
, *p
, *next
;
1957 flags
&= gfp_allowed_mask
;
1959 flags
|= s
->allocflags
;
1962 * Let the initial higher-order allocation fail under memory pressure
1963 * so we fall-back to the minimum order allocation.
1965 alloc_gfp
= (flags
| __GFP_NOWARN
| __GFP_NORETRY
) & ~__GFP_NOFAIL
;
1966 if ((alloc_gfp
& __GFP_DIRECT_RECLAIM
) && oo_order(oo
) > oo_order(s
->min
))
1967 alloc_gfp
= (alloc_gfp
| __GFP_NOMEMALLOC
) & ~__GFP_RECLAIM
;
1969 slab
= alloc_slab_page(alloc_gfp
, node
, oo
);
1970 if (unlikely(!slab
)) {
1974 * Allocation may have failed due to fragmentation.
1975 * Try a lower order alloc if possible
1977 slab
= alloc_slab_page(alloc_gfp
, node
, oo
);
1978 if (unlikely(!slab
))
1980 stat(s
, ORDER_FALLBACK
);
1983 slab
->objects
= oo_objects(oo
);
1985 account_slab(slab
, oo_order(oo
), s
, flags
);
1987 slab
->slab_cache
= s
;
1989 kasan_poison_slab(slab
);
1991 start
= slab_address(slab
);
1993 setup_slab_debug(s
, slab
, start
);
1995 shuffle
= shuffle_freelist(s
, slab
);
1998 start
= fixup_red_left(s
, start
);
1999 start
= setup_object(s
, start
);
2000 slab
->freelist
= start
;
2001 for (idx
= 0, p
= start
; idx
< slab
->objects
- 1; idx
++) {
2003 next
= setup_object(s
, next
);
2004 set_freepointer(s
, p
, next
);
2007 set_freepointer(s
, p
, NULL
);
2010 slab
->inuse
= slab
->objects
;
2017 inc_slabs_node(s
, slab_nid(slab
), slab
->objects
);
2022 static struct slab
*new_slab(struct kmem_cache
*s
, gfp_t flags
, int node
)
2024 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
2025 flags
= kmalloc_fix_flags(flags
);
2027 WARN_ON_ONCE(s
->ctor
&& (flags
& __GFP_ZERO
));
2029 return allocate_slab(s
,
2030 flags
& (GFP_RECLAIM_MASK
| GFP_CONSTRAINT_MASK
), node
);
2033 static void __free_slab(struct kmem_cache
*s
, struct slab
*slab
)
2035 struct folio
*folio
= slab_folio(slab
);
2036 int order
= folio_order(folio
);
2037 int pages
= 1 << order
;
2039 if (kmem_cache_debug_flags(s
, SLAB_CONSISTENCY_CHECKS
)) {
2042 slab_pad_check(s
, slab
);
2043 for_each_object(p
, s
, slab_address(slab
), slab
->objects
)
2044 check_object(s
, slab
, p
, SLUB_RED_INACTIVE
);
2047 __slab_clear_pfmemalloc(slab
);
2048 __folio_clear_slab(folio
);
2049 folio
->mapping
= NULL
;
2050 if (current
->reclaim_state
)
2051 current
->reclaim_state
->reclaimed_slab
+= pages
;
2052 unaccount_slab(slab
, order
, s
);
2053 __free_pages(folio_page(folio
, 0), order
);
2056 static void rcu_free_slab(struct rcu_head
*h
)
2058 struct slab
*slab
= container_of(h
, struct slab
, rcu_head
);
2060 __free_slab(slab
->slab_cache
, slab
);
2063 static void free_slab(struct kmem_cache
*s
, struct slab
*slab
)
2065 if (unlikely(s
->flags
& SLAB_TYPESAFE_BY_RCU
)) {
2066 call_rcu(&slab
->rcu_head
, rcu_free_slab
);
2068 __free_slab(s
, slab
);
2071 static void discard_slab(struct kmem_cache
*s
, struct slab
*slab
)
2073 dec_slabs_node(s
, slab_nid(slab
), slab
->objects
);
2078 * Management of partially allocated slabs.
2081 __add_partial(struct kmem_cache_node
*n
, struct slab
*slab
, int tail
)
2084 if (tail
== DEACTIVATE_TO_TAIL
)
2085 list_add_tail(&slab
->slab_list
, &n
->partial
);
2087 list_add(&slab
->slab_list
, &n
->partial
);
2090 static inline void add_partial(struct kmem_cache_node
*n
,
2091 struct slab
*slab
, int tail
)
2093 lockdep_assert_held(&n
->list_lock
);
2094 __add_partial(n
, slab
, tail
);
2097 static inline void remove_partial(struct kmem_cache_node
*n
,
2100 lockdep_assert_held(&n
->list_lock
);
2101 list_del(&slab
->slab_list
);
2106 * Remove slab from the partial list, freeze it and
2107 * return the pointer to the freelist.
2109 * Returns a list of objects or NULL if it fails.
2111 static inline void *acquire_slab(struct kmem_cache
*s
,
2112 struct kmem_cache_node
*n
, struct slab
*slab
,
2116 unsigned long counters
;
2119 lockdep_assert_held(&n
->list_lock
);
2122 * Zap the freelist and set the frozen bit.
2123 * The old freelist is the list of objects for the
2124 * per cpu allocation list.
2126 freelist
= slab
->freelist
;
2127 counters
= slab
->counters
;
2128 new.counters
= counters
;
2130 new.inuse
= slab
->objects
;
2131 new.freelist
= NULL
;
2133 new.freelist
= freelist
;
2136 VM_BUG_ON(new.frozen
);
2139 if (!__cmpxchg_double_slab(s
, slab
,
2141 new.freelist
, new.counters
,
2145 remove_partial(n
, slab
);
2150 #ifdef CONFIG_SLUB_CPU_PARTIAL
2151 static void put_cpu_partial(struct kmem_cache
*s
, struct slab
*slab
, int drain
);
2153 static inline void put_cpu_partial(struct kmem_cache
*s
, struct slab
*slab
,
2156 static inline bool pfmemalloc_match(struct slab
*slab
, gfp_t gfpflags
);
2159 * Try to allocate a partial slab from a specific node.
2161 static void *get_partial_node(struct kmem_cache
*s
, struct kmem_cache_node
*n
,
2162 struct slab
**ret_slab
, gfp_t gfpflags
)
2164 struct slab
*slab
, *slab2
;
2165 void *object
= NULL
;
2166 unsigned long flags
;
2167 unsigned int partial_slabs
= 0;
2170 * Racy check. If we mistakenly see no partial slabs then we
2171 * just allocate an empty slab. If we mistakenly try to get a
2172 * partial slab and there is none available then get_partial()
2175 if (!n
|| !n
->nr_partial
)
2178 spin_lock_irqsave(&n
->list_lock
, flags
);
2179 list_for_each_entry_safe(slab
, slab2
, &n
->partial
, slab_list
) {
2182 if (!pfmemalloc_match(slab
, gfpflags
))
2185 t
= acquire_slab(s
, n
, slab
, object
== NULL
);
2191 stat(s
, ALLOC_FROM_PARTIAL
);
2194 put_cpu_partial(s
, slab
, 0);
2195 stat(s
, CPU_PARTIAL_NODE
);
2198 #ifdef CONFIG_SLUB_CPU_PARTIAL
2199 if (!kmem_cache_has_cpu_partial(s
)
2200 || partial_slabs
> s
->cpu_partial_slabs
/ 2)
2207 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2212 * Get a slab from somewhere. Search in increasing NUMA distances.
2214 static void *get_any_partial(struct kmem_cache
*s
, gfp_t flags
,
2215 struct slab
**ret_slab
)
2218 struct zonelist
*zonelist
;
2221 enum zone_type highest_zoneidx
= gfp_zone(flags
);
2223 unsigned int cpuset_mems_cookie
;
2226 * The defrag ratio allows a configuration of the tradeoffs between
2227 * inter node defragmentation and node local allocations. A lower
2228 * defrag_ratio increases the tendency to do local allocations
2229 * instead of attempting to obtain partial slabs from other nodes.
2231 * If the defrag_ratio is set to 0 then kmalloc() always
2232 * returns node local objects. If the ratio is higher then kmalloc()
2233 * may return off node objects because partial slabs are obtained
2234 * from other nodes and filled up.
2236 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2237 * (which makes defrag_ratio = 1000) then every (well almost)
2238 * allocation will first attempt to defrag slab caches on other nodes.
2239 * This means scanning over all nodes to look for partial slabs which
2240 * may be expensive if we do it every time we are trying to find a slab
2241 * with available objects.
2243 if (!s
->remote_node_defrag_ratio
||
2244 get_cycles() % 1024 > s
->remote_node_defrag_ratio
)
2248 cpuset_mems_cookie
= read_mems_allowed_begin();
2249 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
2250 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
2251 struct kmem_cache_node
*n
;
2253 n
= get_node(s
, zone_to_nid(zone
));
2255 if (n
&& cpuset_zone_allowed(zone
, flags
) &&
2256 n
->nr_partial
> s
->min_partial
) {
2257 object
= get_partial_node(s
, n
, ret_slab
, flags
);
2260 * Don't check read_mems_allowed_retry()
2261 * here - if mems_allowed was updated in
2262 * parallel, that was a harmless race
2263 * between allocation and the cpuset
2270 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
2271 #endif /* CONFIG_NUMA */
2276 * Get a partial slab, lock it and return it.
2278 static void *get_partial(struct kmem_cache
*s
, gfp_t flags
, int node
,
2279 struct slab
**ret_slab
)
2282 int searchnode
= node
;
2284 if (node
== NUMA_NO_NODE
)
2285 searchnode
= numa_mem_id();
2287 object
= get_partial_node(s
, get_node(s
, searchnode
), ret_slab
, flags
);
2288 if (object
|| node
!= NUMA_NO_NODE
)
2291 return get_any_partial(s
, flags
, ret_slab
);
2294 #ifdef CONFIG_PREEMPTION
2296 * Calculate the next globally unique transaction for disambiguation
2297 * during cmpxchg. The transactions start with the cpu number and are then
2298 * incremented by CONFIG_NR_CPUS.
2300 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2303 * No preemption supported therefore also no need to check for
2309 static inline unsigned long next_tid(unsigned long tid
)
2311 return tid
+ TID_STEP
;
2314 #ifdef SLUB_DEBUG_CMPXCHG
2315 static inline unsigned int tid_to_cpu(unsigned long tid
)
2317 return tid
% TID_STEP
;
2320 static inline unsigned long tid_to_event(unsigned long tid
)
2322 return tid
/ TID_STEP
;
2326 static inline unsigned int init_tid(int cpu
)
2331 static inline void note_cmpxchg_failure(const char *n
,
2332 const struct kmem_cache
*s
, unsigned long tid
)
2334 #ifdef SLUB_DEBUG_CMPXCHG
2335 unsigned long actual_tid
= __this_cpu_read(s
->cpu_slab
->tid
);
2337 pr_info("%s %s: cmpxchg redo ", n
, s
->name
);
2339 #ifdef CONFIG_PREEMPTION
2340 if (tid_to_cpu(tid
) != tid_to_cpu(actual_tid
))
2341 pr_warn("due to cpu change %d -> %d\n",
2342 tid_to_cpu(tid
), tid_to_cpu(actual_tid
));
2345 if (tid_to_event(tid
) != tid_to_event(actual_tid
))
2346 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2347 tid_to_event(tid
), tid_to_event(actual_tid
));
2349 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2350 actual_tid
, tid
, next_tid(tid
));
2352 stat(s
, CMPXCHG_DOUBLE_CPU_FAIL
);
2355 static void init_kmem_cache_cpus(struct kmem_cache
*s
)
2358 struct kmem_cache_cpu
*c
;
2360 for_each_possible_cpu(cpu
) {
2361 c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2362 local_lock_init(&c
->lock
);
2363 c
->tid
= init_tid(cpu
);
2368 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2369 * unfreezes the slabs and puts it on the proper list.
2370 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2373 static void deactivate_slab(struct kmem_cache
*s
, struct slab
*slab
,
2376 enum slab_modes
{ M_NONE
, M_PARTIAL
, M_FULL
, M_FREE
, M_FULL_NOLIST
};
2377 struct kmem_cache_node
*n
= get_node(s
, slab_nid(slab
));
2379 enum slab_modes mode
= M_NONE
;
2380 void *nextfree
, *freelist_iter
, *freelist_tail
;
2381 int tail
= DEACTIVATE_TO_HEAD
;
2382 unsigned long flags
= 0;
2386 if (slab
->freelist
) {
2387 stat(s
, DEACTIVATE_REMOTE_FREES
);
2388 tail
= DEACTIVATE_TO_TAIL
;
2392 * Stage one: Count the objects on cpu's freelist as free_delta and
2393 * remember the last object in freelist_tail for later splicing.
2395 freelist_tail
= NULL
;
2396 freelist_iter
= freelist
;
2397 while (freelist_iter
) {
2398 nextfree
= get_freepointer(s
, freelist_iter
);
2401 * If 'nextfree' is invalid, it is possible that the object at
2402 * 'freelist_iter' is already corrupted. So isolate all objects
2403 * starting at 'freelist_iter' by skipping them.
2405 if (freelist_corrupted(s
, slab
, &freelist_iter
, nextfree
))
2408 freelist_tail
= freelist_iter
;
2411 freelist_iter
= nextfree
;
2415 * Stage two: Unfreeze the slab while splicing the per-cpu
2416 * freelist to the head of slab's freelist.
2418 * Ensure that the slab is unfrozen while the list presence
2419 * reflects the actual number of objects during unfreeze.
2421 * We first perform cmpxchg holding lock and insert to list
2422 * when it succeed. If there is mismatch then the slab is not
2423 * unfrozen and number of objects in the slab may have changed.
2424 * Then release lock and retry cmpxchg again.
2428 old
.freelist
= READ_ONCE(slab
->freelist
);
2429 old
.counters
= READ_ONCE(slab
->counters
);
2430 VM_BUG_ON(!old
.frozen
);
2432 /* Determine target state of the slab */
2433 new.counters
= old
.counters
;
2434 if (freelist_tail
) {
2435 new.inuse
-= free_delta
;
2436 set_freepointer(s
, freelist_tail
, old
.freelist
);
2437 new.freelist
= freelist
;
2439 new.freelist
= old
.freelist
;
2443 if (!new.inuse
&& n
->nr_partial
>= s
->min_partial
) {
2445 } else if (new.freelist
) {
2448 * Taking the spinlock removes the possibility that
2449 * acquire_slab() will see a slab that is frozen
2451 spin_lock_irqsave(&n
->list_lock
, flags
);
2452 } else if (kmem_cache_debug_flags(s
, SLAB_STORE_USER
)) {
2455 * This also ensures that the scanning of full
2456 * slabs from diagnostic functions will not see
2459 spin_lock_irqsave(&n
->list_lock
, flags
);
2461 mode
= M_FULL_NOLIST
;
2465 if (!cmpxchg_double_slab(s
, slab
,
2466 old
.freelist
, old
.counters
,
2467 new.freelist
, new.counters
,
2468 "unfreezing slab")) {
2469 if (mode
== M_PARTIAL
|| mode
== M_FULL
)
2470 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2475 if (mode
== M_PARTIAL
) {
2476 add_partial(n
, slab
, tail
);
2477 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2479 } else if (mode
== M_FREE
) {
2480 stat(s
, DEACTIVATE_EMPTY
);
2481 discard_slab(s
, slab
);
2483 } else if (mode
== M_FULL
) {
2484 add_full(s
, n
, slab
);
2485 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2486 stat(s
, DEACTIVATE_FULL
);
2487 } else if (mode
== M_FULL_NOLIST
) {
2488 stat(s
, DEACTIVATE_FULL
);
2492 #ifdef CONFIG_SLUB_CPU_PARTIAL
2493 static void __unfreeze_partials(struct kmem_cache
*s
, struct slab
*partial_slab
)
2495 struct kmem_cache_node
*n
= NULL
, *n2
= NULL
;
2496 struct slab
*slab
, *slab_to_discard
= NULL
;
2497 unsigned long flags
= 0;
2499 while (partial_slab
) {
2503 slab
= partial_slab
;
2504 partial_slab
= slab
->next
;
2506 n2
= get_node(s
, slab_nid(slab
));
2509 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2512 spin_lock_irqsave(&n
->list_lock
, flags
);
2517 old
.freelist
= slab
->freelist
;
2518 old
.counters
= slab
->counters
;
2519 VM_BUG_ON(!old
.frozen
);
2521 new.counters
= old
.counters
;
2522 new.freelist
= old
.freelist
;
2526 } while (!__cmpxchg_double_slab(s
, slab
,
2527 old
.freelist
, old
.counters
,
2528 new.freelist
, new.counters
,
2529 "unfreezing slab"));
2531 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
)) {
2532 slab
->next
= slab_to_discard
;
2533 slab_to_discard
= slab
;
2535 add_partial(n
, slab
, DEACTIVATE_TO_TAIL
);
2536 stat(s
, FREE_ADD_PARTIAL
);
2541 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2543 while (slab_to_discard
) {
2544 slab
= slab_to_discard
;
2545 slab_to_discard
= slab_to_discard
->next
;
2547 stat(s
, DEACTIVATE_EMPTY
);
2548 discard_slab(s
, slab
);
2554 * Unfreeze all the cpu partial slabs.
2556 static void unfreeze_partials(struct kmem_cache
*s
)
2558 struct slab
*partial_slab
;
2559 unsigned long flags
;
2561 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2562 partial_slab
= this_cpu_read(s
->cpu_slab
->partial
);
2563 this_cpu_write(s
->cpu_slab
->partial
, NULL
);
2564 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2567 __unfreeze_partials(s
, partial_slab
);
2570 static void unfreeze_partials_cpu(struct kmem_cache
*s
,
2571 struct kmem_cache_cpu
*c
)
2573 struct slab
*partial_slab
;
2575 partial_slab
= slub_percpu_partial(c
);
2579 __unfreeze_partials(s
, partial_slab
);
2583 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2584 * partial slab slot if available.
2586 * If we did not find a slot then simply move all the partials to the
2587 * per node partial list.
2589 static void put_cpu_partial(struct kmem_cache
*s
, struct slab
*slab
, int drain
)
2591 struct slab
*oldslab
;
2592 struct slab
*slab_to_unfreeze
= NULL
;
2593 unsigned long flags
;
2596 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2598 oldslab
= this_cpu_read(s
->cpu_slab
->partial
);
2601 if (drain
&& oldslab
->slabs
>= s
->cpu_partial_slabs
) {
2603 * Partial array is full. Move the existing set to the
2604 * per node partial list. Postpone the actual unfreezing
2605 * outside of the critical section.
2607 slab_to_unfreeze
= oldslab
;
2610 slabs
= oldslab
->slabs
;
2616 slab
->slabs
= slabs
;
2617 slab
->next
= oldslab
;
2619 this_cpu_write(s
->cpu_slab
->partial
, slab
);
2621 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2623 if (slab_to_unfreeze
) {
2624 __unfreeze_partials(s
, slab_to_unfreeze
);
2625 stat(s
, CPU_PARTIAL_DRAIN
);
2629 #else /* CONFIG_SLUB_CPU_PARTIAL */
2631 static inline void unfreeze_partials(struct kmem_cache
*s
) { }
2632 static inline void unfreeze_partials_cpu(struct kmem_cache
*s
,
2633 struct kmem_cache_cpu
*c
) { }
2635 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2637 static inline void flush_slab(struct kmem_cache
*s
, struct kmem_cache_cpu
*c
)
2639 unsigned long flags
;
2643 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2646 freelist
= c
->freelist
;
2650 c
->tid
= next_tid(c
->tid
);
2652 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2655 deactivate_slab(s
, slab
, freelist
);
2656 stat(s
, CPUSLAB_FLUSH
);
2660 static inline void __flush_cpu_slab(struct kmem_cache
*s
, int cpu
)
2662 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2663 void *freelist
= c
->freelist
;
2664 struct slab
*slab
= c
->slab
;
2668 c
->tid
= next_tid(c
->tid
);
2671 deactivate_slab(s
, slab
, freelist
);
2672 stat(s
, CPUSLAB_FLUSH
);
2675 unfreeze_partials_cpu(s
, c
);
2678 struct slub_flush_work
{
2679 struct work_struct work
;
2680 struct kmem_cache
*s
;
2687 * Called from CPU work handler with migration disabled.
2689 static void flush_cpu_slab(struct work_struct
*w
)
2691 struct kmem_cache
*s
;
2692 struct kmem_cache_cpu
*c
;
2693 struct slub_flush_work
*sfw
;
2695 sfw
= container_of(w
, struct slub_flush_work
, work
);
2698 c
= this_cpu_ptr(s
->cpu_slab
);
2703 unfreeze_partials(s
);
2706 static bool has_cpu_slab(int cpu
, struct kmem_cache
*s
)
2708 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
, cpu
);
2710 return c
->slab
|| slub_percpu_partial(c
);
2713 static DEFINE_MUTEX(flush_lock
);
2714 static DEFINE_PER_CPU(struct slub_flush_work
, slub_flush
);
2716 static void flush_all_cpus_locked(struct kmem_cache
*s
)
2718 struct slub_flush_work
*sfw
;
2721 lockdep_assert_cpus_held();
2722 mutex_lock(&flush_lock
);
2724 for_each_online_cpu(cpu
) {
2725 sfw
= &per_cpu(slub_flush
, cpu
);
2726 if (!has_cpu_slab(cpu
, s
)) {
2730 INIT_WORK(&sfw
->work
, flush_cpu_slab
);
2733 schedule_work_on(cpu
, &sfw
->work
);
2736 for_each_online_cpu(cpu
) {
2737 sfw
= &per_cpu(slub_flush
, cpu
);
2740 flush_work(&sfw
->work
);
2743 mutex_unlock(&flush_lock
);
2746 static void flush_all(struct kmem_cache
*s
)
2749 flush_all_cpus_locked(s
);
2754 * Use the cpu notifier to insure that the cpu slabs are flushed when
2757 static int slub_cpu_dead(unsigned int cpu
)
2759 struct kmem_cache
*s
;
2761 mutex_lock(&slab_mutex
);
2762 list_for_each_entry(s
, &slab_caches
, list
)
2763 __flush_cpu_slab(s
, cpu
);
2764 mutex_unlock(&slab_mutex
);
2769 * Check if the objects in a per cpu structure fit numa
2770 * locality expectations.
2772 static inline int node_match(struct slab
*slab
, int node
)
2775 if (node
!= NUMA_NO_NODE
&& slab_nid(slab
) != node
)
2781 #ifdef CONFIG_SLUB_DEBUG
2782 static int count_free(struct slab
*slab
)
2784 return slab
->objects
- slab
->inuse
;
2787 static inline unsigned long node_nr_objs(struct kmem_cache_node
*n
)
2789 return atomic_long_read(&n
->total_objects
);
2791 #endif /* CONFIG_SLUB_DEBUG */
2793 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2794 static unsigned long count_partial(struct kmem_cache_node
*n
,
2795 int (*get_count
)(struct slab
*))
2797 unsigned long flags
;
2798 unsigned long x
= 0;
2801 spin_lock_irqsave(&n
->list_lock
, flags
);
2802 list_for_each_entry(slab
, &n
->partial
, slab_list
)
2803 x
+= get_count(slab
);
2804 spin_unlock_irqrestore(&n
->list_lock
, flags
);
2807 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2809 static noinline
void
2810 slab_out_of_memory(struct kmem_cache
*s
, gfp_t gfpflags
, int nid
)
2812 #ifdef CONFIG_SLUB_DEBUG
2813 static DEFINE_RATELIMIT_STATE(slub_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
2814 DEFAULT_RATELIMIT_BURST
);
2816 struct kmem_cache_node
*n
;
2818 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slub_oom_rs
))
2821 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2822 nid
, gfpflags
, &gfpflags
);
2823 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2824 s
->name
, s
->object_size
, s
->size
, oo_order(s
->oo
),
2827 if (oo_order(s
->min
) > get_order(s
->object_size
))
2828 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2831 for_each_kmem_cache_node(s
, node
, n
) {
2832 unsigned long nr_slabs
;
2833 unsigned long nr_objs
;
2834 unsigned long nr_free
;
2836 nr_free
= count_partial(n
, count_free
);
2837 nr_slabs
= node_nr_slabs(n
);
2838 nr_objs
= node_nr_objs(n
);
2840 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2841 node
, nr_slabs
, nr_objs
, nr_free
);
2846 static inline bool pfmemalloc_match(struct slab
*slab
, gfp_t gfpflags
)
2848 if (unlikely(slab_test_pfmemalloc(slab
)))
2849 return gfp_pfmemalloc_allowed(gfpflags
);
2855 * Check the slab->freelist and either transfer the freelist to the
2856 * per cpu freelist or deactivate the slab.
2858 * The slab is still frozen if the return value is not NULL.
2860 * If this function returns NULL then the slab has been unfrozen.
2862 static inline void *get_freelist(struct kmem_cache
*s
, struct slab
*slab
)
2865 unsigned long counters
;
2868 lockdep_assert_held(this_cpu_ptr(&s
->cpu_slab
->lock
));
2871 freelist
= slab
->freelist
;
2872 counters
= slab
->counters
;
2874 new.counters
= counters
;
2875 VM_BUG_ON(!new.frozen
);
2877 new.inuse
= slab
->objects
;
2878 new.frozen
= freelist
!= NULL
;
2880 } while (!__cmpxchg_double_slab(s
, slab
,
2889 * Slow path. The lockless freelist is empty or we need to perform
2892 * Processing is still very fast if new objects have been freed to the
2893 * regular freelist. In that case we simply take over the regular freelist
2894 * as the lockless freelist and zap the regular freelist.
2896 * If that is not working then we fall back to the partial lists. We take the
2897 * first element of the freelist as the object to allocate now and move the
2898 * rest of the freelist to the lockless freelist.
2900 * And if we were unable to get a new slab from the partial slab lists then
2901 * we need to allocate a new slab. This is the slowest path since it involves
2902 * a call to the page allocator and the setup of a new slab.
2904 * Version of __slab_alloc to use when we know that preemption is
2905 * already disabled (which is the case for bulk allocation).
2907 static void *___slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
2908 unsigned long addr
, struct kmem_cache_cpu
*c
)
2912 unsigned long flags
;
2914 stat(s
, ALLOC_SLOWPATH
);
2918 slab
= READ_ONCE(c
->slab
);
2921 * if the node is not online or has no normal memory, just
2922 * ignore the node constraint
2924 if (unlikely(node
!= NUMA_NO_NODE
&&
2925 !node_isset(node
, slab_nodes
)))
2926 node
= NUMA_NO_NODE
;
2931 if (unlikely(!node_match(slab
, node
))) {
2933 * same as above but node_match() being false already
2934 * implies node != NUMA_NO_NODE
2936 if (!node_isset(node
, slab_nodes
)) {
2937 node
= NUMA_NO_NODE
;
2939 stat(s
, ALLOC_NODE_MISMATCH
);
2940 goto deactivate_slab
;
2945 * By rights, we should be searching for a slab page that was
2946 * PFMEMALLOC but right now, we are losing the pfmemalloc
2947 * information when the page leaves the per-cpu allocator
2949 if (unlikely(!pfmemalloc_match(slab
, gfpflags
)))
2950 goto deactivate_slab
;
2952 /* must check again c->slab in case we got preempted and it changed */
2953 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2954 if (unlikely(slab
!= c
->slab
)) {
2955 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2958 freelist
= c
->freelist
;
2962 freelist
= get_freelist(s
, slab
);
2966 c
->tid
= next_tid(c
->tid
);
2967 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2968 stat(s
, DEACTIVATE_BYPASS
);
2972 stat(s
, ALLOC_REFILL
);
2976 lockdep_assert_held(this_cpu_ptr(&s
->cpu_slab
->lock
));
2979 * freelist is pointing to the list of objects to be used.
2980 * slab is pointing to the slab from which the objects are obtained.
2981 * That slab must be frozen for per cpu allocations to work.
2983 VM_BUG_ON(!c
->slab
->frozen
);
2984 c
->freelist
= get_freepointer(s
, freelist
);
2985 c
->tid
= next_tid(c
->tid
);
2986 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2991 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
2992 if (slab
!= c
->slab
) {
2993 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
2996 freelist
= c
->freelist
;
2999 c
->tid
= next_tid(c
->tid
);
3000 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3001 deactivate_slab(s
, slab
, freelist
);
3005 if (slub_percpu_partial(c
)) {
3006 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
3007 if (unlikely(c
->slab
)) {
3008 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3011 if (unlikely(!slub_percpu_partial(c
))) {
3012 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3013 /* we were preempted and partial list got empty */
3017 slab
= c
->slab
= slub_percpu_partial(c
);
3018 slub_set_percpu_partial(c
, slab
);
3019 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3020 stat(s
, CPU_PARTIAL_ALLOC
);
3026 freelist
= get_partial(s
, gfpflags
, node
, &slab
);
3028 goto check_new_slab
;
3030 slub_put_cpu_ptr(s
->cpu_slab
);
3031 slab
= new_slab(s
, gfpflags
, node
);
3032 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3034 if (unlikely(!slab
)) {
3035 slab_out_of_memory(s
, gfpflags
, node
);
3040 * No other reference to the slab yet so we can
3041 * muck around with it freely without cmpxchg
3043 freelist
= slab
->freelist
;
3044 slab
->freelist
= NULL
;
3046 stat(s
, ALLOC_SLAB
);
3050 if (kmem_cache_debug(s
)) {
3051 if (!alloc_debug_processing(s
, slab
, freelist
, addr
)) {
3052 /* Slab failed checks. Next slab needed */
3056 * For debug case, we don't load freelist so that all
3057 * allocations go through alloc_debug_processing()
3063 if (unlikely(!pfmemalloc_match(slab
, gfpflags
)))
3065 * For !pfmemalloc_match() case we don't load freelist so that
3066 * we don't make further mismatched allocations easier.
3072 local_lock_irqsave(&s
->cpu_slab
->lock
, flags
);
3073 if (unlikely(c
->slab
)) {
3074 void *flush_freelist
= c
->freelist
;
3075 struct slab
*flush_slab
= c
->slab
;
3079 c
->tid
= next_tid(c
->tid
);
3081 local_unlock_irqrestore(&s
->cpu_slab
->lock
, flags
);
3083 deactivate_slab(s
, flush_slab
, flush_freelist
);
3085 stat(s
, CPUSLAB_FLUSH
);
3087 goto retry_load_slab
;
3095 deactivate_slab(s
, slab
, get_freepointer(s
, freelist
));
3100 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3101 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3104 static void *__slab_alloc(struct kmem_cache
*s
, gfp_t gfpflags
, int node
,
3105 unsigned long addr
, struct kmem_cache_cpu
*c
)
3109 #ifdef CONFIG_PREEMPT_COUNT
3111 * We may have been preempted and rescheduled on a different
3112 * cpu before disabling preemption. Need to reload cpu area
3115 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3118 p
= ___slab_alloc(s
, gfpflags
, node
, addr
, c
);
3119 #ifdef CONFIG_PREEMPT_COUNT
3120 slub_put_cpu_ptr(s
->cpu_slab
);
3126 * If the object has been wiped upon free, make sure it's fully initialized by
3127 * zeroing out freelist pointer.
3129 static __always_inline
void maybe_wipe_obj_freeptr(struct kmem_cache
*s
,
3132 if (unlikely(slab_want_init_on_free(s
)) && obj
)
3133 memset((void *)((char *)kasan_reset_tag(obj
) + s
->offset
),
3138 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3139 * have the fastpath folded into their functions. So no function call
3140 * overhead for requests that can be satisfied on the fastpath.
3142 * The fastpath works by first checking if the lockless freelist can be used.
3143 * If not then __slab_alloc is called for slow processing.
3145 * Otherwise we can simply pick the next object from the lockless free list.
3147 static __always_inline
void *slab_alloc_node(struct kmem_cache
*s
, struct list_lru
*lru
,
3148 gfp_t gfpflags
, int node
, unsigned long addr
, size_t orig_size
)
3151 struct kmem_cache_cpu
*c
;
3154 struct obj_cgroup
*objcg
= NULL
;
3157 s
= slab_pre_alloc_hook(s
, lru
, &objcg
, 1, gfpflags
);
3161 object
= kfence_alloc(s
, orig_size
, gfpflags
);
3162 if (unlikely(object
))
3167 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3168 * enabled. We may switch back and forth between cpus while
3169 * reading from one cpu area. That does not matter as long
3170 * as we end up on the original cpu again when doing the cmpxchg.
3172 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3173 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3174 * the tid. If we are preempted and switched to another cpu between the
3175 * two reads, it's OK as the two are still associated with the same cpu
3176 * and cmpxchg later will validate the cpu.
3178 c
= raw_cpu_ptr(s
->cpu_slab
);
3179 tid
= READ_ONCE(c
->tid
);
3182 * Irqless object alloc/free algorithm used here depends on sequence
3183 * of fetching cpu_slab's data. tid should be fetched before anything
3184 * on c to guarantee that object and slab associated with previous tid
3185 * won't be used with current tid. If we fetch tid first, object and
3186 * slab could be one associated with next tid and our alloc/free
3187 * request will be failed. In this case, we will retry. So, no problem.
3192 * The transaction ids are globally unique per cpu and per operation on
3193 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3194 * occurs on the right processor and that there was no operation on the
3195 * linked list in between.
3198 object
= c
->freelist
;
3201 * We cannot use the lockless fastpath on PREEMPT_RT because if a
3202 * slowpath has taken the local_lock_irqsave(), it is not protected
3203 * against a fast path operation in an irq handler. So we need to take
3204 * the slow path which uses local_lock. It is still relatively fast if
3205 * there is a suitable cpu freelist.
3207 if (IS_ENABLED(CONFIG_PREEMPT_RT
) ||
3208 unlikely(!object
|| !slab
|| !node_match(slab
, node
))) {
3209 object
= __slab_alloc(s
, gfpflags
, node
, addr
, c
);
3211 void *next_object
= get_freepointer_safe(s
, object
);
3214 * The cmpxchg will only match if there was no additional
3215 * operation and if we are on the right processor.
3217 * The cmpxchg does the following atomically (without lock
3219 * 1. Relocate first pointer to the current per cpu area.
3220 * 2. Verify that tid and freelist have not been changed
3221 * 3. If they were not changed replace tid and freelist
3223 * Since this is without lock semantics the protection is only
3224 * against code executing on this cpu *not* from access by
3227 if (unlikely(!this_cpu_cmpxchg_double(
3228 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3230 next_object
, next_tid(tid
)))) {
3232 note_cmpxchg_failure("slab_alloc", s
, tid
);
3235 prefetch_freepointer(s
, next_object
);
3236 stat(s
, ALLOC_FASTPATH
);
3239 maybe_wipe_obj_freeptr(s
, object
);
3240 init
= slab_want_init_on_alloc(gfpflags
, s
);
3243 slab_post_alloc_hook(s
, objcg
, gfpflags
, 1, &object
, init
);
3248 static __always_inline
void *slab_alloc(struct kmem_cache
*s
, struct list_lru
*lru
,
3249 gfp_t gfpflags
, unsigned long addr
, size_t orig_size
)
3251 return slab_alloc_node(s
, lru
, gfpflags
, NUMA_NO_NODE
, addr
, orig_size
);
3254 static __always_inline
3255 void *__kmem_cache_alloc_lru(struct kmem_cache
*s
, struct list_lru
*lru
,
3258 void *ret
= slab_alloc(s
, lru
, gfpflags
, _RET_IP_
, s
->object_size
);
3260 trace_kmem_cache_alloc(_RET_IP_
, ret
, s
, s
->object_size
,
3266 void *kmem_cache_alloc(struct kmem_cache
*s
, gfp_t gfpflags
)
3268 return __kmem_cache_alloc_lru(s
, NULL
, gfpflags
);
3270 EXPORT_SYMBOL(kmem_cache_alloc
);
3272 void *kmem_cache_alloc_lru(struct kmem_cache
*s
, struct list_lru
*lru
,
3275 return __kmem_cache_alloc_lru(s
, lru
, gfpflags
);
3277 EXPORT_SYMBOL(kmem_cache_alloc_lru
);
3279 #ifdef CONFIG_TRACING
3280 void *kmem_cache_alloc_trace(struct kmem_cache
*s
, gfp_t gfpflags
, size_t size
)
3282 void *ret
= slab_alloc(s
, NULL
, gfpflags
, _RET_IP_
, size
);
3283 trace_kmalloc(_RET_IP_
, ret
, s
, size
, s
->size
, gfpflags
);
3284 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
3287 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3291 void *kmem_cache_alloc_node(struct kmem_cache
*s
, gfp_t gfpflags
, int node
)
3293 void *ret
= slab_alloc_node(s
, NULL
, gfpflags
, node
, _RET_IP_
, s
->object_size
);
3295 trace_kmem_cache_alloc_node(_RET_IP_
, ret
, s
,
3296 s
->object_size
, s
->size
, gfpflags
, node
);
3300 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3302 #ifdef CONFIG_TRACING
3303 void *kmem_cache_alloc_node_trace(struct kmem_cache
*s
,
3305 int node
, size_t size
)
3307 void *ret
= slab_alloc_node(s
, NULL
, gfpflags
, node
, _RET_IP_
, size
);
3309 trace_kmalloc_node(_RET_IP_
, ret
, s
,
3310 size
, s
->size
, gfpflags
, node
);
3312 ret
= kasan_kmalloc(s
, ret
, size
, gfpflags
);
3315 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3317 #endif /* CONFIG_NUMA */
3320 * Slow path handling. This may still be called frequently since objects
3321 * have a longer lifetime than the cpu slabs in most processing loads.
3323 * So we still attempt to reduce cache line usage. Just take the slab
3324 * lock and free the item. If there is no additional partial slab
3325 * handling required then we can return immediately.
3327 static void __slab_free(struct kmem_cache
*s
, struct slab
*slab
,
3328 void *head
, void *tail
, int cnt
,
3335 unsigned long counters
;
3336 struct kmem_cache_node
*n
= NULL
;
3337 unsigned long flags
;
3339 stat(s
, FREE_SLOWPATH
);
3341 if (kfence_free(head
))
3344 if (kmem_cache_debug(s
) &&
3345 !free_debug_processing(s
, slab
, head
, tail
, cnt
, addr
))
3350 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3353 prior
= slab
->freelist
;
3354 counters
= slab
->counters
;
3355 set_freepointer(s
, tail
, prior
);
3356 new.counters
= counters
;
3357 was_frozen
= new.frozen
;
3359 if ((!new.inuse
|| !prior
) && !was_frozen
) {
3361 if (kmem_cache_has_cpu_partial(s
) && !prior
) {
3364 * Slab was on no list before and will be
3366 * We can defer the list move and instead
3371 } else { /* Needs to be taken off a list */
3373 n
= get_node(s
, slab_nid(slab
));
3375 * Speculatively acquire the list_lock.
3376 * If the cmpxchg does not succeed then we may
3377 * drop the list_lock without any processing.
3379 * Otherwise the list_lock will synchronize with
3380 * other processors updating the list of slabs.
3382 spin_lock_irqsave(&n
->list_lock
, flags
);
3387 } while (!cmpxchg_double_slab(s
, slab
,
3394 if (likely(was_frozen
)) {
3396 * The list lock was not taken therefore no list
3397 * activity can be necessary.
3399 stat(s
, FREE_FROZEN
);
3400 } else if (new.frozen
) {
3402 * If we just froze the slab then put it onto the
3403 * per cpu partial list.
3405 put_cpu_partial(s
, slab
, 1);
3406 stat(s
, CPU_PARTIAL_FREE
);
3412 if (unlikely(!new.inuse
&& n
->nr_partial
>= s
->min_partial
))
3416 * Objects left in the slab. If it was not on the partial list before
3419 if (!kmem_cache_has_cpu_partial(s
) && unlikely(!prior
)) {
3420 remove_full(s
, n
, slab
);
3421 add_partial(n
, slab
, DEACTIVATE_TO_TAIL
);
3422 stat(s
, FREE_ADD_PARTIAL
);
3424 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3430 * Slab on the partial list.
3432 remove_partial(n
, slab
);
3433 stat(s
, FREE_REMOVE_PARTIAL
);
3435 /* Slab must be on the full list */
3436 remove_full(s
, n
, slab
);
3439 spin_unlock_irqrestore(&n
->list_lock
, flags
);
3441 discard_slab(s
, slab
);
3445 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3446 * can perform fastpath freeing without additional function calls.
3448 * The fastpath is only possible if we are freeing to the current cpu slab
3449 * of this processor. This typically the case if we have just allocated
3452 * If fastpath is not possible then fall back to __slab_free where we deal
3453 * with all sorts of special processing.
3455 * Bulk free of a freelist with several objects (all pointing to the
3456 * same slab) possible by specifying head and tail ptr, plus objects
3457 * count (cnt). Bulk free indicated by tail pointer being set.
3459 static __always_inline
void do_slab_free(struct kmem_cache
*s
,
3460 struct slab
*slab
, void *head
, void *tail
,
3461 int cnt
, unsigned long addr
)
3463 void *tail_obj
= tail
? : head
;
3464 struct kmem_cache_cpu
*c
;
3469 * Determine the currently cpus per cpu slab.
3470 * The cpu may change afterward. However that does not matter since
3471 * data is retrieved via this pointer. If we are on the same cpu
3472 * during the cmpxchg then the free will succeed.
3474 c
= raw_cpu_ptr(s
->cpu_slab
);
3475 tid
= READ_ONCE(c
->tid
);
3477 /* Same with comment on barrier() in slab_alloc_node() */
3480 if (likely(slab
== c
->slab
)) {
3481 #ifndef CONFIG_PREEMPT_RT
3482 void **freelist
= READ_ONCE(c
->freelist
);
3484 set_freepointer(s
, tail_obj
, freelist
);
3486 if (unlikely(!this_cpu_cmpxchg_double(
3487 s
->cpu_slab
->freelist
, s
->cpu_slab
->tid
,
3489 head
, next_tid(tid
)))) {
3491 note_cmpxchg_failure("slab_free", s
, tid
);
3494 #else /* CONFIG_PREEMPT_RT */
3496 * We cannot use the lockless fastpath on PREEMPT_RT because if
3497 * a slowpath has taken the local_lock_irqsave(), it is not
3498 * protected against a fast path operation in an irq handler. So
3499 * we need to take the local_lock. We shouldn't simply defer to
3500 * __slab_free() as that wouldn't use the cpu freelist at all.
3504 local_lock(&s
->cpu_slab
->lock
);
3505 c
= this_cpu_ptr(s
->cpu_slab
);
3506 if (unlikely(slab
!= c
->slab
)) {
3507 local_unlock(&s
->cpu_slab
->lock
);
3511 freelist
= c
->freelist
;
3513 set_freepointer(s
, tail_obj
, freelist
);
3515 c
->tid
= next_tid(tid
);
3517 local_unlock(&s
->cpu_slab
->lock
);
3519 stat(s
, FREE_FASTPATH
);
3521 __slab_free(s
, slab
, head
, tail_obj
, cnt
, addr
);
3525 static __always_inline
void slab_free(struct kmem_cache
*s
, struct slab
*slab
,
3526 void *head
, void *tail
, void **p
, int cnt
,
3529 memcg_slab_free_hook(s
, slab
, p
, cnt
);
3531 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3532 * to remove objects, whose reuse must be delayed.
3534 if (slab_free_freelist_hook(s
, &head
, &tail
, &cnt
))
3535 do_slab_free(s
, slab
, head
, tail
, cnt
, addr
);
3538 #ifdef CONFIG_KASAN_GENERIC
3539 void ___cache_free(struct kmem_cache
*cache
, void *x
, unsigned long addr
)
3541 do_slab_free(cache
, virt_to_slab(x
), x
, NULL
, 1, addr
);
3545 void kmem_cache_free(struct kmem_cache
*s
, void *x
)
3547 s
= cache_from_obj(s
, x
);
3550 trace_kmem_cache_free(_RET_IP_
, x
, s
->name
);
3551 slab_free(s
, virt_to_slab(x
), x
, NULL
, &x
, 1, _RET_IP_
);
3553 EXPORT_SYMBOL(kmem_cache_free
);
3555 struct detached_freelist
{
3560 struct kmem_cache
*s
;
3563 static inline void free_large_kmalloc(struct folio
*folio
, void *object
)
3565 unsigned int order
= folio_order(folio
);
3567 if (WARN_ON_ONCE(order
== 0))
3568 pr_warn_once("object pointer: 0x%p\n", object
);
3571 mod_lruvec_page_state(folio_page(folio
, 0), NR_SLAB_UNRECLAIMABLE_B
,
3572 -(PAGE_SIZE
<< order
));
3573 __free_pages(folio_page(folio
, 0), order
);
3577 * This function progressively scans the array with free objects (with
3578 * a limited look ahead) and extract objects belonging to the same
3579 * slab. It builds a detached freelist directly within the given
3580 * slab/objects. This can happen without any need for
3581 * synchronization, because the objects are owned by running process.
3582 * The freelist is build up as a single linked list in the objects.
3583 * The idea is, that this detached freelist can then be bulk
3584 * transferred to the real freelist(s), but only requiring a single
3585 * synchronization primitive. Look ahead in the array is limited due
3586 * to performance reasons.
3589 int build_detached_freelist(struct kmem_cache
*s
, size_t size
,
3590 void **p
, struct detached_freelist
*df
)
3594 struct folio
*folio
;
3598 folio
= virt_to_folio(object
);
3600 /* Handle kalloc'ed objects */
3601 if (unlikely(!folio_test_slab(folio
))) {
3602 free_large_kmalloc(folio
, object
);
3606 /* Derive kmem_cache from object */
3607 df
->slab
= folio_slab(folio
);
3608 df
->s
= df
->slab
->slab_cache
;
3610 df
->slab
= folio_slab(folio
);
3611 df
->s
= cache_from_obj(s
, object
); /* Support for memcg */
3614 /* Start new detached freelist */
3616 df
->freelist
= object
;
3619 if (is_kfence_address(object
))
3622 set_freepointer(df
->s
, object
, NULL
);
3627 /* df->slab is always set at this point */
3628 if (df
->slab
== virt_to_slab(object
)) {
3629 /* Opportunity build freelist */
3630 set_freepointer(df
->s
, object
, df
->freelist
);
3631 df
->freelist
= object
;
3635 swap(p
[size
], p
[same
]);
3639 /* Limit look ahead search */
3647 /* Note that interrupts must be enabled when calling this function. */
3648 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3654 struct detached_freelist df
;
3656 size
= build_detached_freelist(s
, size
, p
, &df
);
3660 slab_free(df
.s
, df
.slab
, df
.freelist
, df
.tail
, &p
[size
], df
.cnt
,
3662 } while (likely(size
));
3664 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3666 /* Note that interrupts must be enabled when calling this function. */
3667 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3670 struct kmem_cache_cpu
*c
;
3672 struct obj_cgroup
*objcg
= NULL
;
3674 /* memcg and kmem_cache debug support */
3675 s
= slab_pre_alloc_hook(s
, NULL
, &objcg
, size
, flags
);
3679 * Drain objects in the per cpu slab, while disabling local
3680 * IRQs, which protects against PREEMPT and interrupts
3681 * handlers invoking normal fastpath.
3683 c
= slub_get_cpu_ptr(s
->cpu_slab
);
3684 local_lock_irq(&s
->cpu_slab
->lock
);
3686 for (i
= 0; i
< size
; i
++) {
3687 void *object
= kfence_alloc(s
, s
->object_size
, flags
);
3689 if (unlikely(object
)) {
3694 object
= c
->freelist
;
3695 if (unlikely(!object
)) {
3697 * We may have removed an object from c->freelist using
3698 * the fastpath in the previous iteration; in that case,
3699 * c->tid has not been bumped yet.
3700 * Since ___slab_alloc() may reenable interrupts while
3701 * allocating memory, we should bump c->tid now.
3703 c
->tid
= next_tid(c
->tid
);
3705 local_unlock_irq(&s
->cpu_slab
->lock
);
3708 * Invoking slow path likely have side-effect
3709 * of re-populating per CPU c->freelist
3711 p
[i
] = ___slab_alloc(s
, flags
, NUMA_NO_NODE
,
3713 if (unlikely(!p
[i
]))
3716 c
= this_cpu_ptr(s
->cpu_slab
);
3717 maybe_wipe_obj_freeptr(s
, p
[i
]);
3719 local_lock_irq(&s
->cpu_slab
->lock
);
3721 continue; /* goto for-loop */
3723 c
->freelist
= get_freepointer(s
, object
);
3725 maybe_wipe_obj_freeptr(s
, p
[i
]);
3727 c
->tid
= next_tid(c
->tid
);
3728 local_unlock_irq(&s
->cpu_slab
->lock
);
3729 slub_put_cpu_ptr(s
->cpu_slab
);
3732 * memcg and kmem_cache debug support and memory initialization.
3733 * Done outside of the IRQ disabled fastpath loop.
3735 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3736 slab_want_init_on_alloc(flags
, s
));
3739 slub_put_cpu_ptr(s
->cpu_slab
);
3740 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false);
3741 kmem_cache_free_bulk(s
, i
, p
);
3744 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3748 * Object placement in a slab is made very easy because we always start at
3749 * offset 0. If we tune the size of the object to the alignment then we can
3750 * get the required alignment by putting one properly sized object after
3753 * Notice that the allocation order determines the sizes of the per cpu
3754 * caches. Each processor has always one slab available for allocations.
3755 * Increasing the allocation order reduces the number of times that slabs
3756 * must be moved on and off the partial lists and is therefore a factor in
3761 * Minimum / Maximum order of slab pages. This influences locking overhead
3762 * and slab fragmentation. A higher order reduces the number of partial slabs
3763 * and increases the number of allocations possible without having to
3764 * take the list_lock.
3766 static unsigned int slub_min_order
;
3767 static unsigned int slub_max_order
= PAGE_ALLOC_COSTLY_ORDER
;
3768 static unsigned int slub_min_objects
;
3771 * Calculate the order of allocation given an slab object size.
3773 * The order of allocation has significant impact on performance and other
3774 * system components. Generally order 0 allocations should be preferred since
3775 * order 0 does not cause fragmentation in the page allocator. Larger objects
3776 * be problematic to put into order 0 slabs because there may be too much
3777 * unused space left. We go to a higher order if more than 1/16th of the slab
3780 * In order to reach satisfactory performance we must ensure that a minimum
3781 * number of objects is in one slab. Otherwise we may generate too much
3782 * activity on the partial lists which requires taking the list_lock. This is
3783 * less a concern for large slabs though which are rarely used.
3785 * slub_max_order specifies the order where we begin to stop considering the
3786 * number of objects in a slab as critical. If we reach slub_max_order then
3787 * we try to keep the page order as low as possible. So we accept more waste
3788 * of space in favor of a small page order.
3790 * Higher order allocations also allow the placement of more objects in a
3791 * slab and thereby reduce object handling overhead. If the user has
3792 * requested a higher minimum order then we start with that one instead of
3793 * the smallest order which will fit the object.
3795 static inline unsigned int calc_slab_order(unsigned int size
,
3796 unsigned int min_objects
, unsigned int max_order
,
3797 unsigned int fract_leftover
)
3799 unsigned int min_order
= slub_min_order
;
3802 if (order_objects(min_order
, size
) > MAX_OBJS_PER_PAGE
)
3803 return get_order(size
* MAX_OBJS_PER_PAGE
) - 1;
3805 for (order
= max(min_order
, (unsigned int)get_order(min_objects
* size
));
3806 order
<= max_order
; order
++) {
3808 unsigned int slab_size
= (unsigned int)PAGE_SIZE
<< order
;
3811 rem
= slab_size
% size
;
3813 if (rem
<= slab_size
/ fract_leftover
)
3820 static inline int calculate_order(unsigned int size
)
3823 unsigned int min_objects
;
3824 unsigned int max_objects
;
3825 unsigned int nr_cpus
;
3828 * Attempt to find best configuration for a slab. This
3829 * works by first attempting to generate a layout with
3830 * the best configuration and backing off gradually.
3832 * First we increase the acceptable waste in a slab. Then
3833 * we reduce the minimum objects required in a slab.
3835 min_objects
= slub_min_objects
;
3838 * Some architectures will only update present cpus when
3839 * onlining them, so don't trust the number if it's just 1. But
3840 * we also don't want to use nr_cpu_ids always, as on some other
3841 * architectures, there can be many possible cpus, but never
3842 * onlined. Here we compromise between trying to avoid too high
3843 * order on systems that appear larger than they are, and too
3844 * low order on systems that appear smaller than they are.
3846 nr_cpus
= num_present_cpus();
3848 nr_cpus
= nr_cpu_ids
;
3849 min_objects
= 4 * (fls(nr_cpus
) + 1);
3851 max_objects
= order_objects(slub_max_order
, size
);
3852 min_objects
= min(min_objects
, max_objects
);
3854 while (min_objects
> 1) {
3855 unsigned int fraction
;
3858 while (fraction
>= 4) {
3859 order
= calc_slab_order(size
, min_objects
,
3860 slub_max_order
, fraction
);
3861 if (order
<= slub_max_order
)
3869 * We were unable to place multiple objects in a slab. Now
3870 * lets see if we can place a single object there.
3872 order
= calc_slab_order(size
, 1, slub_max_order
, 1);
3873 if (order
<= slub_max_order
)
3877 * Doh this slab cannot be placed using slub_max_order.
3879 order
= calc_slab_order(size
, 1, MAX_ORDER
, 1);
3880 if (order
< MAX_ORDER
)
3886 init_kmem_cache_node(struct kmem_cache_node
*n
)
3889 spin_lock_init(&n
->list_lock
);
3890 INIT_LIST_HEAD(&n
->partial
);
3891 #ifdef CONFIG_SLUB_DEBUG
3892 atomic_long_set(&n
->nr_slabs
, 0);
3893 atomic_long_set(&n
->total_objects
, 0);
3894 INIT_LIST_HEAD(&n
->full
);
3898 static inline int alloc_kmem_cache_cpus(struct kmem_cache
*s
)
3900 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE
<
3901 KMALLOC_SHIFT_HIGH
* sizeof(struct kmem_cache_cpu
));
3904 * Must align to double word boundary for the double cmpxchg
3905 * instructions to work; see __pcpu_double_call_return_bool().
3907 s
->cpu_slab
= __alloc_percpu(sizeof(struct kmem_cache_cpu
),
3908 2 * sizeof(void *));
3913 init_kmem_cache_cpus(s
);
3918 static struct kmem_cache
*kmem_cache_node
;
3921 * No kmalloc_node yet so do it by hand. We know that this is the first
3922 * slab on the node for this slabcache. There are no concurrent accesses
3925 * Note that this function only works on the kmem_cache_node
3926 * when allocating for the kmem_cache_node. This is used for bootstrapping
3927 * memory on a fresh node that has no slab structures yet.
3929 static void early_kmem_cache_node_alloc(int node
)
3932 struct kmem_cache_node
*n
;
3934 BUG_ON(kmem_cache_node
->size
< sizeof(struct kmem_cache_node
));
3936 slab
= new_slab(kmem_cache_node
, GFP_NOWAIT
, node
);
3939 if (slab_nid(slab
) != node
) {
3940 pr_err("SLUB: Unable to allocate memory from node %d\n", node
);
3941 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3946 #ifdef CONFIG_SLUB_DEBUG
3947 init_object(kmem_cache_node
, n
, SLUB_RED_ACTIVE
);
3948 init_tracking(kmem_cache_node
, n
);
3950 n
= kasan_slab_alloc(kmem_cache_node
, n
, GFP_KERNEL
, false);
3951 slab
->freelist
= get_freepointer(kmem_cache_node
, n
);
3954 kmem_cache_node
->node
[node
] = n
;
3955 init_kmem_cache_node(n
);
3956 inc_slabs_node(kmem_cache_node
, node
, slab
->objects
);
3959 * No locks need to be taken here as it has just been
3960 * initialized and there is no concurrent access.
3962 __add_partial(n
, slab
, DEACTIVATE_TO_HEAD
);
3965 static void free_kmem_cache_nodes(struct kmem_cache
*s
)
3968 struct kmem_cache_node
*n
;
3970 for_each_kmem_cache_node(s
, node
, n
) {
3971 s
->node
[node
] = NULL
;
3972 kmem_cache_free(kmem_cache_node
, n
);
3976 void __kmem_cache_release(struct kmem_cache
*s
)
3978 cache_random_seq_destroy(s
);
3979 free_percpu(s
->cpu_slab
);
3980 free_kmem_cache_nodes(s
);
3983 static int init_kmem_cache_nodes(struct kmem_cache
*s
)
3987 for_each_node_mask(node
, slab_nodes
) {
3988 struct kmem_cache_node
*n
;
3990 if (slab_state
== DOWN
) {
3991 early_kmem_cache_node_alloc(node
);
3994 n
= kmem_cache_alloc_node(kmem_cache_node
,
3998 free_kmem_cache_nodes(s
);
4002 init_kmem_cache_node(n
);
4008 static void set_cpu_partial(struct kmem_cache
*s
)
4010 #ifdef CONFIG_SLUB_CPU_PARTIAL
4011 unsigned int nr_objects
;
4014 * cpu_partial determined the maximum number of objects kept in the
4015 * per cpu partial lists of a processor.
4017 * Per cpu partial lists mainly contain slabs that just have one
4018 * object freed. If they are used for allocation then they can be
4019 * filled up again with minimal effort. The slab will never hit the
4020 * per node partial lists and therefore no locking will be required.
4022 * For backwards compatibility reasons, this is determined as number
4023 * of objects, even though we now limit maximum number of pages, see
4024 * slub_set_cpu_partial()
4026 if (!kmem_cache_has_cpu_partial(s
))
4028 else if (s
->size
>= PAGE_SIZE
)
4030 else if (s
->size
>= 1024)
4032 else if (s
->size
>= 256)
4037 slub_set_cpu_partial(s
, nr_objects
);
4042 * calculate_sizes() determines the order and the distribution of data within
4045 static int calculate_sizes(struct kmem_cache
*s
)
4047 slab_flags_t flags
= s
->flags
;
4048 unsigned int size
= s
->object_size
;
4052 * Round up object size to the next word boundary. We can only
4053 * place the free pointer at word boundaries and this determines
4054 * the possible location of the free pointer.
4056 size
= ALIGN(size
, sizeof(void *));
4058 #ifdef CONFIG_SLUB_DEBUG
4060 * Determine if we can poison the object itself. If the user of
4061 * the slab may touch the object after free or before allocation
4062 * then we should never poison the object itself.
4064 if ((flags
& SLAB_POISON
) && !(flags
& SLAB_TYPESAFE_BY_RCU
) &&
4066 s
->flags
|= __OBJECT_POISON
;
4068 s
->flags
&= ~__OBJECT_POISON
;
4072 * If we are Redzoning then check if there is some space between the
4073 * end of the object and the free pointer. If not then add an
4074 * additional word to have some bytes to store Redzone information.
4076 if ((flags
& SLAB_RED_ZONE
) && size
== s
->object_size
)
4077 size
+= sizeof(void *);
4081 * With that we have determined the number of bytes in actual use
4082 * by the object and redzoning.
4086 if ((flags
& (SLAB_TYPESAFE_BY_RCU
| SLAB_POISON
)) ||
4087 ((flags
& SLAB_RED_ZONE
) && s
->object_size
< sizeof(void *)) ||
4090 * Relocate free pointer after the object if it is not
4091 * permitted to overwrite the first word of the object on
4094 * This is the case if we do RCU, have a constructor or
4095 * destructor, are poisoning the objects, or are
4096 * redzoning an object smaller than sizeof(void *).
4098 * The assumption that s->offset >= s->inuse means free
4099 * pointer is outside of the object is used in the
4100 * freeptr_outside_object() function. If that is no
4101 * longer true, the function needs to be modified.
4104 size
+= sizeof(void *);
4107 * Store freelist pointer near middle of object to keep
4108 * it away from the edges of the object to avoid small
4109 * sized over/underflows from neighboring allocations.
4111 s
->offset
= ALIGN_DOWN(s
->object_size
/ 2, sizeof(void *));
4114 #ifdef CONFIG_SLUB_DEBUG
4115 if (flags
& SLAB_STORE_USER
)
4117 * Need to store information about allocs and frees after
4120 size
+= 2 * sizeof(struct track
);
4123 kasan_cache_create(s
, &size
, &s
->flags
);
4124 #ifdef CONFIG_SLUB_DEBUG
4125 if (flags
& SLAB_RED_ZONE
) {
4127 * Add some empty padding so that we can catch
4128 * overwrites from earlier objects rather than let
4129 * tracking information or the free pointer be
4130 * corrupted if a user writes before the start
4133 size
+= sizeof(void *);
4135 s
->red_left_pad
= sizeof(void *);
4136 s
->red_left_pad
= ALIGN(s
->red_left_pad
, s
->align
);
4137 size
+= s
->red_left_pad
;
4142 * SLUB stores one object immediately after another beginning from
4143 * offset 0. In order to align the objects we have to simply size
4144 * each object to conform to the alignment.
4146 size
= ALIGN(size
, s
->align
);
4148 s
->reciprocal_size
= reciprocal_value(size
);
4149 order
= calculate_order(size
);
4156 s
->allocflags
|= __GFP_COMP
;
4158 if (s
->flags
& SLAB_CACHE_DMA
)
4159 s
->allocflags
|= GFP_DMA
;
4161 if (s
->flags
& SLAB_CACHE_DMA32
)
4162 s
->allocflags
|= GFP_DMA32
;
4164 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
4165 s
->allocflags
|= __GFP_RECLAIMABLE
;
4168 * Determine the number of objects per slab
4170 s
->oo
= oo_make(order
, size
);
4171 s
->min
= oo_make(get_order(size
), size
);
4173 return !!oo_objects(s
->oo
);
4176 static int kmem_cache_open(struct kmem_cache
*s
, slab_flags_t flags
)
4178 s
->flags
= kmem_cache_flags(s
->size
, flags
, s
->name
);
4179 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4180 s
->random
= get_random_long();
4183 if (!calculate_sizes(s
))
4185 if (disable_higher_order_debug
) {
4187 * Disable debugging flags that store metadata if the min slab
4190 if (get_order(s
->size
) > get_order(s
->object_size
)) {
4191 s
->flags
&= ~DEBUG_METADATA_FLAGS
;
4193 if (!calculate_sizes(s
))
4198 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4199 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4200 if (system_has_cmpxchg_double() && (s
->flags
& SLAB_NO_CMPXCHG
) == 0)
4201 /* Enable fast mode */
4202 s
->flags
|= __CMPXCHG_DOUBLE
;
4206 * The larger the object size is, the more slabs we want on the partial
4207 * list to avoid pounding the page allocator excessively.
4209 s
->min_partial
= min_t(unsigned long, MAX_PARTIAL
, ilog2(s
->size
) / 2);
4210 s
->min_partial
= max_t(unsigned long, MIN_PARTIAL
, s
->min_partial
);
4215 s
->remote_node_defrag_ratio
= 1000;
4218 /* Initialize the pre-computed randomized freelist if slab is up */
4219 if (slab_state
>= UP
) {
4220 if (init_cache_random_seq(s
))
4224 if (!init_kmem_cache_nodes(s
))
4227 if (alloc_kmem_cache_cpus(s
))
4231 __kmem_cache_release(s
);
4235 static void list_slab_objects(struct kmem_cache
*s
, struct slab
*slab
,
4238 #ifdef CONFIG_SLUB_DEBUG
4239 void *addr
= slab_address(slab
);
4240 unsigned long flags
;
4244 slab_err(s
, slab
, text
, s
->name
);
4245 slab_lock(slab
, &flags
);
4247 map
= get_map(s
, slab
);
4248 for_each_object(p
, s
, addr
, slab
->objects
) {
4250 if (!test_bit(__obj_to_index(s
, addr
, p
), map
)) {
4251 pr_err("Object 0x%p @offset=%tu\n", p
, p
- addr
);
4252 print_tracking(s
, p
);
4256 slab_unlock(slab
, &flags
);
4261 * Attempt to free all partial slabs on a node.
4262 * This is called from __kmem_cache_shutdown(). We must take list_lock
4263 * because sysfs file might still access partial list after the shutdowning.
4265 static void free_partial(struct kmem_cache
*s
, struct kmem_cache_node
*n
)
4268 struct slab
*slab
, *h
;
4270 BUG_ON(irqs_disabled());
4271 spin_lock_irq(&n
->list_lock
);
4272 list_for_each_entry_safe(slab
, h
, &n
->partial
, slab_list
) {
4274 remove_partial(n
, slab
);
4275 list_add(&slab
->slab_list
, &discard
);
4277 list_slab_objects(s
, slab
,
4278 "Objects remaining in %s on __kmem_cache_shutdown()");
4281 spin_unlock_irq(&n
->list_lock
);
4283 list_for_each_entry_safe(slab
, h
, &discard
, slab_list
)
4284 discard_slab(s
, slab
);
4287 bool __kmem_cache_empty(struct kmem_cache
*s
)
4290 struct kmem_cache_node
*n
;
4292 for_each_kmem_cache_node(s
, node
, n
)
4293 if (n
->nr_partial
|| slabs_node(s
, node
))
4299 * Release all resources used by a slab cache.
4301 int __kmem_cache_shutdown(struct kmem_cache
*s
)
4304 struct kmem_cache_node
*n
;
4306 flush_all_cpus_locked(s
);
4307 /* Attempt to free all objects */
4308 for_each_kmem_cache_node(s
, node
, n
) {
4310 if (n
->nr_partial
|| slabs_node(s
, node
))
4316 #ifdef CONFIG_PRINTK
4317 void __kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct slab
*slab
)
4320 int __maybe_unused i
;
4324 struct kmem_cache
*s
= slab
->slab_cache
;
4325 struct track __maybe_unused
*trackp
;
4327 kpp
->kp_ptr
= object
;
4328 kpp
->kp_slab
= slab
;
4329 kpp
->kp_slab_cache
= s
;
4330 base
= slab_address(slab
);
4331 objp0
= kasan_reset_tag(object
);
4332 #ifdef CONFIG_SLUB_DEBUG
4333 objp
= restore_red_left(s
, objp0
);
4337 objnr
= obj_to_index(s
, slab
, objp
);
4338 kpp
->kp_data_offset
= (unsigned long)((char *)objp0
- (char *)objp
);
4339 objp
= base
+ s
->size
* objnr
;
4340 kpp
->kp_objp
= objp
;
4341 if (WARN_ON_ONCE(objp
< base
|| objp
>= base
+ slab
->objects
* s
->size
4342 || (objp
- base
) % s
->size
) ||
4343 !(s
->flags
& SLAB_STORE_USER
))
4345 #ifdef CONFIG_SLUB_DEBUG
4346 objp
= fixup_red_left(s
, objp
);
4347 trackp
= get_track(s
, objp
, TRACK_ALLOC
);
4348 kpp
->kp_ret
= (void *)trackp
->addr
;
4349 #ifdef CONFIG_STACKDEPOT
4351 depot_stack_handle_t handle
;
4352 unsigned long *entries
;
4353 unsigned int nr_entries
;
4355 handle
= READ_ONCE(trackp
->handle
);
4357 nr_entries
= stack_depot_fetch(handle
, &entries
);
4358 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< nr_entries
; i
++)
4359 kpp
->kp_stack
[i
] = (void *)entries
[i
];
4362 trackp
= get_track(s
, objp
, TRACK_FREE
);
4363 handle
= READ_ONCE(trackp
->handle
);
4365 nr_entries
= stack_depot_fetch(handle
, &entries
);
4366 for (i
= 0; i
< KS_ADDRS_COUNT
&& i
< nr_entries
; i
++)
4367 kpp
->kp_free_stack
[i
] = (void *)entries
[i
];
4375 /********************************************************************
4377 *******************************************************************/
4379 static int __init
setup_slub_min_order(char *str
)
4381 get_option(&str
, (int *)&slub_min_order
);
4386 __setup("slub_min_order=", setup_slub_min_order
);
4388 static int __init
setup_slub_max_order(char *str
)
4390 get_option(&str
, (int *)&slub_max_order
);
4391 slub_max_order
= min(slub_max_order
, (unsigned int)MAX_ORDER
- 1);
4396 __setup("slub_max_order=", setup_slub_max_order
);
4398 static int __init
setup_slub_min_objects(char *str
)
4400 get_option(&str
, (int *)&slub_min_objects
);
4405 __setup("slub_min_objects=", setup_slub_min_objects
);
4407 void *__kmalloc(size_t size
, gfp_t flags
)
4409 struct kmem_cache
*s
;
4412 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4413 return kmalloc_large(size
, flags
);
4415 s
= kmalloc_slab(size
, flags
);
4417 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4420 ret
= slab_alloc(s
, NULL
, flags
, _RET_IP_
, size
);
4422 trace_kmalloc(_RET_IP_
, ret
, s
, size
, s
->size
, flags
);
4424 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4428 EXPORT_SYMBOL(__kmalloc
);
4431 static void *kmalloc_large_node(size_t size
, gfp_t flags
, int node
)
4435 unsigned int order
= get_order(size
);
4437 flags
|= __GFP_COMP
;
4438 page
= alloc_pages_node(node
, flags
, order
);
4440 ptr
= page_address(page
);
4441 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE_B
,
4442 PAGE_SIZE
<< order
);
4445 return kmalloc_large_node_hook(ptr
, size
, flags
);
4448 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
4450 struct kmem_cache
*s
;
4453 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4454 ret
= kmalloc_large_node(size
, flags
, node
);
4456 trace_kmalloc_node(_RET_IP_
, ret
, NULL
,
4457 size
, PAGE_SIZE
<< get_order(size
),
4463 s
= kmalloc_slab(size
, flags
);
4465 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4468 ret
= slab_alloc_node(s
, NULL
, flags
, node
, _RET_IP_
, size
);
4470 trace_kmalloc_node(_RET_IP_
, ret
, s
, size
, s
->size
, flags
, node
);
4472 ret
= kasan_kmalloc(s
, ret
, size
, flags
);
4476 EXPORT_SYMBOL(__kmalloc_node
);
4477 #endif /* CONFIG_NUMA */
4479 #ifdef CONFIG_HARDENED_USERCOPY
4481 * Rejects incorrectly sized objects and objects that are to be copied
4482 * to/from userspace but do not fall entirely within the containing slab
4483 * cache's usercopy region.
4485 * Returns NULL if check passes, otherwise const char * to name of cache
4486 * to indicate an error.
4488 void __check_heap_object(const void *ptr
, unsigned long n
,
4489 const struct slab
*slab
, bool to_user
)
4491 struct kmem_cache
*s
;
4492 unsigned int offset
;
4493 bool is_kfence
= is_kfence_address(ptr
);
4495 ptr
= kasan_reset_tag(ptr
);
4497 /* Find object and usable object size. */
4498 s
= slab
->slab_cache
;
4500 /* Reject impossible pointers. */
4501 if (ptr
< slab_address(slab
))
4502 usercopy_abort("SLUB object not in SLUB page?!", NULL
,
4505 /* Find offset within object. */
4507 offset
= ptr
- kfence_object_start(ptr
);
4509 offset
= (ptr
- slab_address(slab
)) % s
->size
;
4511 /* Adjust for redzone and reject if within the redzone. */
4512 if (!is_kfence
&& kmem_cache_debug_flags(s
, SLAB_RED_ZONE
)) {
4513 if (offset
< s
->red_left_pad
)
4514 usercopy_abort("SLUB object in left red zone",
4515 s
->name
, to_user
, offset
, n
);
4516 offset
-= s
->red_left_pad
;
4519 /* Allow address range falling entirely within usercopy region. */
4520 if (offset
>= s
->useroffset
&&
4521 offset
- s
->useroffset
<= s
->usersize
&&
4522 n
<= s
->useroffset
- offset
+ s
->usersize
)
4525 usercopy_abort("SLUB object", s
->name
, to_user
, offset
, n
);
4527 #endif /* CONFIG_HARDENED_USERCOPY */
4529 size_t __ksize(const void *object
)
4531 struct folio
*folio
;
4533 if (unlikely(object
== ZERO_SIZE_PTR
))
4536 folio
= virt_to_folio(object
);
4538 if (unlikely(!folio_test_slab(folio
)))
4539 return folio_size(folio
);
4541 return slab_ksize(folio_slab(folio
)->slab_cache
);
4543 EXPORT_SYMBOL(__ksize
);
4545 void kfree(const void *x
)
4547 struct folio
*folio
;
4549 void *object
= (void *)x
;
4551 trace_kfree(_RET_IP_
, x
);
4553 if (unlikely(ZERO_OR_NULL_PTR(x
)))
4556 folio
= virt_to_folio(x
);
4557 if (unlikely(!folio_test_slab(folio
))) {
4558 free_large_kmalloc(folio
, object
);
4561 slab
= folio_slab(folio
);
4562 slab_free(slab
->slab_cache
, slab
, object
, NULL
, &object
, 1, _RET_IP_
);
4564 EXPORT_SYMBOL(kfree
);
4566 #define SHRINK_PROMOTE_MAX 32
4569 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4570 * up most to the head of the partial lists. New allocations will then
4571 * fill those up and thus they can be removed from the partial lists.
4573 * The slabs with the least items are placed last. This results in them
4574 * being allocated from last increasing the chance that the last objects
4575 * are freed in them.
4577 static int __kmem_cache_do_shrink(struct kmem_cache
*s
)
4581 struct kmem_cache_node
*n
;
4584 struct list_head discard
;
4585 struct list_head promote
[SHRINK_PROMOTE_MAX
];
4586 unsigned long flags
;
4589 for_each_kmem_cache_node(s
, node
, n
) {
4590 INIT_LIST_HEAD(&discard
);
4591 for (i
= 0; i
< SHRINK_PROMOTE_MAX
; i
++)
4592 INIT_LIST_HEAD(promote
+ i
);
4594 spin_lock_irqsave(&n
->list_lock
, flags
);
4597 * Build lists of slabs to discard or promote.
4599 * Note that concurrent frees may occur while we hold the
4600 * list_lock. slab->inuse here is the upper limit.
4602 list_for_each_entry_safe(slab
, t
, &n
->partial
, slab_list
) {
4603 int free
= slab
->objects
- slab
->inuse
;
4605 /* Do not reread slab->inuse */
4608 /* We do not keep full slabs on the list */
4611 if (free
== slab
->objects
) {
4612 list_move(&slab
->slab_list
, &discard
);
4614 } else if (free
<= SHRINK_PROMOTE_MAX
)
4615 list_move(&slab
->slab_list
, promote
+ free
- 1);
4619 * Promote the slabs filled up most to the head of the
4622 for (i
= SHRINK_PROMOTE_MAX
- 1; i
>= 0; i
--)
4623 list_splice(promote
+ i
, &n
->partial
);
4625 spin_unlock_irqrestore(&n
->list_lock
, flags
);
4627 /* Release empty slabs */
4628 list_for_each_entry_safe(slab
, t
, &discard
, slab_list
)
4629 discard_slab(s
, slab
);
4631 if (slabs_node(s
, node
))
4638 int __kmem_cache_shrink(struct kmem_cache
*s
)
4641 return __kmem_cache_do_shrink(s
);
4644 static int slab_mem_going_offline_callback(void *arg
)
4646 struct kmem_cache
*s
;
4648 mutex_lock(&slab_mutex
);
4649 list_for_each_entry(s
, &slab_caches
, list
) {
4650 flush_all_cpus_locked(s
);
4651 __kmem_cache_do_shrink(s
);
4653 mutex_unlock(&slab_mutex
);
4658 static void slab_mem_offline_callback(void *arg
)
4660 struct memory_notify
*marg
= arg
;
4663 offline_node
= marg
->status_change_nid_normal
;
4666 * If the node still has available memory. we need kmem_cache_node
4669 if (offline_node
< 0)
4672 mutex_lock(&slab_mutex
);
4673 node_clear(offline_node
, slab_nodes
);
4675 * We no longer free kmem_cache_node structures here, as it would be
4676 * racy with all get_node() users, and infeasible to protect them with
4679 mutex_unlock(&slab_mutex
);
4682 static int slab_mem_going_online_callback(void *arg
)
4684 struct kmem_cache_node
*n
;
4685 struct kmem_cache
*s
;
4686 struct memory_notify
*marg
= arg
;
4687 int nid
= marg
->status_change_nid_normal
;
4691 * If the node's memory is already available, then kmem_cache_node is
4692 * already created. Nothing to do.
4698 * We are bringing a node online. No memory is available yet. We must
4699 * allocate a kmem_cache_node structure in order to bring the node
4702 mutex_lock(&slab_mutex
);
4703 list_for_each_entry(s
, &slab_caches
, list
) {
4705 * The structure may already exist if the node was previously
4706 * onlined and offlined.
4708 if (get_node(s
, nid
))
4711 * XXX: kmem_cache_alloc_node will fallback to other nodes
4712 * since memory is not yet available from the node that
4715 n
= kmem_cache_alloc(kmem_cache_node
, GFP_KERNEL
);
4720 init_kmem_cache_node(n
);
4724 * Any cache created after this point will also have kmem_cache_node
4725 * initialized for the new node.
4727 node_set(nid
, slab_nodes
);
4729 mutex_unlock(&slab_mutex
);
4733 static int slab_memory_callback(struct notifier_block
*self
,
4734 unsigned long action
, void *arg
)
4739 case MEM_GOING_ONLINE
:
4740 ret
= slab_mem_going_online_callback(arg
);
4742 case MEM_GOING_OFFLINE
:
4743 ret
= slab_mem_going_offline_callback(arg
);
4746 case MEM_CANCEL_ONLINE
:
4747 slab_mem_offline_callback(arg
);
4750 case MEM_CANCEL_OFFLINE
:
4754 ret
= notifier_from_errno(ret
);
4760 static struct notifier_block slab_memory_callback_nb
= {
4761 .notifier_call
= slab_memory_callback
,
4762 .priority
= SLAB_CALLBACK_PRI
,
4765 /********************************************************************
4766 * Basic setup of slabs
4767 *******************************************************************/
4770 * Used for early kmem_cache structures that were allocated using
4771 * the page allocator. Allocate them properly then fix up the pointers
4772 * that may be pointing to the wrong kmem_cache structure.
4775 static struct kmem_cache
* __init
bootstrap(struct kmem_cache
*static_cache
)
4778 struct kmem_cache
*s
= kmem_cache_zalloc(kmem_cache
, GFP_NOWAIT
);
4779 struct kmem_cache_node
*n
;
4781 memcpy(s
, static_cache
, kmem_cache
->object_size
);
4784 * This runs very early, and only the boot processor is supposed to be
4785 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4788 __flush_cpu_slab(s
, smp_processor_id());
4789 for_each_kmem_cache_node(s
, node
, n
) {
4792 list_for_each_entry(p
, &n
->partial
, slab_list
)
4795 #ifdef CONFIG_SLUB_DEBUG
4796 list_for_each_entry(p
, &n
->full
, slab_list
)
4800 list_add(&s
->list
, &slab_caches
);
4804 void __init
kmem_cache_init(void)
4806 static __initdata
struct kmem_cache boot_kmem_cache
,
4807 boot_kmem_cache_node
;
4810 if (debug_guardpage_minorder())
4813 /* Print slub debugging pointers without hashing */
4814 if (__slub_debug_enabled())
4815 no_hash_pointers_enable(NULL
);
4817 kmem_cache_node
= &boot_kmem_cache_node
;
4818 kmem_cache
= &boot_kmem_cache
;
4821 * Initialize the nodemask for which we will allocate per node
4822 * structures. Here we don't need taking slab_mutex yet.
4824 for_each_node_state(node
, N_NORMAL_MEMORY
)
4825 node_set(node
, slab_nodes
);
4827 create_boot_cache(kmem_cache_node
, "kmem_cache_node",
4828 sizeof(struct kmem_cache_node
), SLAB_HWCACHE_ALIGN
, 0, 0);
4830 register_hotmemory_notifier(&slab_memory_callback_nb
);
4832 /* Able to allocate the per node structures */
4833 slab_state
= PARTIAL
;
4835 create_boot_cache(kmem_cache
, "kmem_cache",
4836 offsetof(struct kmem_cache
, node
) +
4837 nr_node_ids
* sizeof(struct kmem_cache_node
*),
4838 SLAB_HWCACHE_ALIGN
, 0, 0);
4840 kmem_cache
= bootstrap(&boot_kmem_cache
);
4841 kmem_cache_node
= bootstrap(&boot_kmem_cache_node
);
4843 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4844 setup_kmalloc_cache_index_table();
4845 create_kmalloc_caches(0);
4847 /* Setup random freelists for each cache */
4848 init_freelist_randomization();
4850 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD
, "slub:dead", NULL
,
4853 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4855 slub_min_order
, slub_max_order
, slub_min_objects
,
4856 nr_cpu_ids
, nr_node_ids
);
4859 void __init
kmem_cache_init_late(void)
4864 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
4865 slab_flags_t flags
, void (*ctor
)(void *))
4867 struct kmem_cache
*s
;
4869 s
= find_mergeable(size
, align
, flags
, name
, ctor
);
4871 if (sysfs_slab_alias(s
, name
))
4877 * Adjust the object sizes so that we clear
4878 * the complete object on kzalloc.
4880 s
->object_size
= max(s
->object_size
, size
);
4881 s
->inuse
= max(s
->inuse
, ALIGN(size
, sizeof(void *)));
4887 int __kmem_cache_create(struct kmem_cache
*s
, slab_flags_t flags
)
4891 err
= kmem_cache_open(s
, flags
);
4895 /* Mutex is not taken during early boot */
4896 if (slab_state
<= UP
)
4899 err
= sysfs_slab_add(s
);
4901 __kmem_cache_release(s
);
4905 if (s
->flags
& SLAB_STORE_USER
)
4906 debugfs_slab_add(s
);
4911 void *__kmalloc_track_caller(size_t size
, gfp_t gfpflags
, unsigned long caller
)
4913 struct kmem_cache
*s
;
4916 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
4917 return kmalloc_large(size
, gfpflags
);
4919 s
= kmalloc_slab(size
, gfpflags
);
4921 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4924 ret
= slab_alloc(s
, NULL
, gfpflags
, caller
, size
);
4926 /* Honor the call site pointer we received. */
4927 trace_kmalloc(caller
, ret
, s
, size
, s
->size
, gfpflags
);
4931 EXPORT_SYMBOL(__kmalloc_track_caller
);
4934 void *__kmalloc_node_track_caller(size_t size
, gfp_t gfpflags
,
4935 int node
, unsigned long caller
)
4937 struct kmem_cache
*s
;
4940 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
)) {
4941 ret
= kmalloc_large_node(size
, gfpflags
, node
);
4943 trace_kmalloc_node(caller
, ret
, NULL
,
4944 size
, PAGE_SIZE
<< get_order(size
),
4950 s
= kmalloc_slab(size
, gfpflags
);
4952 if (unlikely(ZERO_OR_NULL_PTR(s
)))
4955 ret
= slab_alloc_node(s
, NULL
, gfpflags
, node
, caller
, size
);
4957 /* Honor the call site pointer we received. */
4958 trace_kmalloc_node(caller
, ret
, s
, size
, s
->size
, gfpflags
, node
);
4962 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
4966 static int count_inuse(struct slab
*slab
)
4971 static int count_total(struct slab
*slab
)
4973 return slab
->objects
;
4977 #ifdef CONFIG_SLUB_DEBUG
4978 static void validate_slab(struct kmem_cache
*s
, struct slab
*slab
,
4979 unsigned long *obj_map
)
4982 void *addr
= slab_address(slab
);
4983 unsigned long flags
;
4985 slab_lock(slab
, &flags
);
4987 if (!check_slab(s
, slab
) || !on_freelist(s
, slab
, NULL
))
4990 /* Now we know that a valid freelist exists */
4991 __fill_map(obj_map
, s
, slab
);
4992 for_each_object(p
, s
, addr
, slab
->objects
) {
4993 u8 val
= test_bit(__obj_to_index(s
, addr
, p
), obj_map
) ?
4994 SLUB_RED_INACTIVE
: SLUB_RED_ACTIVE
;
4996 if (!check_object(s
, slab
, p
, val
))
5000 slab_unlock(slab
, &flags
);
5003 static int validate_slab_node(struct kmem_cache
*s
,
5004 struct kmem_cache_node
*n
, unsigned long *obj_map
)
5006 unsigned long count
= 0;
5008 unsigned long flags
;
5010 spin_lock_irqsave(&n
->list_lock
, flags
);
5012 list_for_each_entry(slab
, &n
->partial
, slab_list
) {
5013 validate_slab(s
, slab
, obj_map
);
5016 if (count
!= n
->nr_partial
) {
5017 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5018 s
->name
, count
, n
->nr_partial
);
5019 slab_add_kunit_errors();
5022 if (!(s
->flags
& SLAB_STORE_USER
))
5025 list_for_each_entry(slab
, &n
->full
, slab_list
) {
5026 validate_slab(s
, slab
, obj_map
);
5029 if (count
!= atomic_long_read(&n
->nr_slabs
)) {
5030 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5031 s
->name
, count
, atomic_long_read(&n
->nr_slabs
));
5032 slab_add_kunit_errors();
5036 spin_unlock_irqrestore(&n
->list_lock
, flags
);
5040 long validate_slab_cache(struct kmem_cache
*s
)
5043 unsigned long count
= 0;
5044 struct kmem_cache_node
*n
;
5045 unsigned long *obj_map
;
5047 obj_map
= bitmap_alloc(oo_objects(s
->oo
), GFP_KERNEL
);
5052 for_each_kmem_cache_node(s
, node
, n
)
5053 count
+= validate_slab_node(s
, n
, obj_map
);
5055 bitmap_free(obj_map
);
5059 EXPORT_SYMBOL(validate_slab_cache
);
5061 #ifdef CONFIG_DEBUG_FS
5063 * Generate lists of code addresses where slabcache objects are allocated
5068 depot_stack_handle_t handle
;
5069 unsigned long count
;
5076 DECLARE_BITMAP(cpus
, NR_CPUS
);
5082 unsigned long count
;
5083 struct location
*loc
;
5087 static struct dentry
*slab_debugfs_root
;
5089 static void free_loc_track(struct loc_track
*t
)
5092 free_pages((unsigned long)t
->loc
,
5093 get_order(sizeof(struct location
) * t
->max
));
5096 static int alloc_loc_track(struct loc_track
*t
, unsigned long max
, gfp_t flags
)
5101 order
= get_order(sizeof(struct location
) * max
);
5103 l
= (void *)__get_free_pages(flags
, order
);
5108 memcpy(l
, t
->loc
, sizeof(struct location
) * t
->count
);
5116 static int add_location(struct loc_track
*t
, struct kmem_cache
*s
,
5117 const struct track
*track
)
5119 long start
, end
, pos
;
5121 unsigned long caddr
, chandle
;
5122 unsigned long age
= jiffies
- track
->when
;
5123 depot_stack_handle_t handle
= 0;
5125 #ifdef CONFIG_STACKDEPOT
5126 handle
= READ_ONCE(track
->handle
);
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 chandle
= t
->loc
[pos
].handle
;
5143 if ((track
->addr
== caddr
) && (handle
== chandle
)) {
5149 if (age
< l
->min_time
)
5151 if (age
> l
->max_time
)
5154 if (track
->pid
< l
->min_pid
)
5155 l
->min_pid
= track
->pid
;
5156 if (track
->pid
> l
->max_pid
)
5157 l
->max_pid
= track
->pid
;
5159 cpumask_set_cpu(track
->cpu
,
5160 to_cpumask(l
->cpus
));
5162 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
5166 if (track
->addr
< caddr
)
5168 else if (track
->addr
== caddr
&& handle
< chandle
)
5175 * Not found. Insert new tracking element.
5177 if (t
->count
>= t
->max
&& !alloc_loc_track(t
, 2 * t
->max
, GFP_ATOMIC
))
5183 (t
->count
- pos
) * sizeof(struct location
));
5186 l
->addr
= track
->addr
;
5190 l
->min_pid
= track
->pid
;
5191 l
->max_pid
= track
->pid
;
5193 cpumask_clear(to_cpumask(l
->cpus
));
5194 cpumask_set_cpu(track
->cpu
, to_cpumask(l
->cpus
));
5195 nodes_clear(l
->nodes
);
5196 node_set(page_to_nid(virt_to_page(track
)), l
->nodes
);
5200 static void process_slab(struct loc_track
*t
, struct kmem_cache
*s
,
5201 struct slab
*slab
, enum track_item alloc
,
5202 unsigned long *obj_map
)
5204 void *addr
= slab_address(slab
);
5207 __fill_map(obj_map
, s
, slab
);
5209 for_each_object(p
, s
, addr
, slab
->objects
)
5210 if (!test_bit(__obj_to_index(s
, addr
, p
), obj_map
))
5211 add_location(t
, s
, get_track(s
, p
, alloc
));
5213 #endif /* CONFIG_DEBUG_FS */
5214 #endif /* CONFIG_SLUB_DEBUG */
5217 enum slab_stat_type
{
5218 SL_ALL
, /* All slabs */
5219 SL_PARTIAL
, /* Only partially allocated slabs */
5220 SL_CPU
, /* Only slabs used for cpu caches */
5221 SL_OBJECTS
, /* Determine allocated objects not slabs */
5222 SL_TOTAL
/* Determine object capacity not slabs */
5225 #define SO_ALL (1 << SL_ALL)
5226 #define SO_PARTIAL (1 << SL_PARTIAL)
5227 #define SO_CPU (1 << SL_CPU)
5228 #define SO_OBJECTS (1 << SL_OBJECTS)
5229 #define SO_TOTAL (1 << SL_TOTAL)
5231 static ssize_t
show_slab_objects(struct kmem_cache
*s
,
5232 char *buf
, unsigned long flags
)
5234 unsigned long total
= 0;
5237 unsigned long *nodes
;
5240 nodes
= kcalloc(nr_node_ids
, sizeof(unsigned long), GFP_KERNEL
);
5244 if (flags
& SO_CPU
) {
5247 for_each_possible_cpu(cpu
) {
5248 struct kmem_cache_cpu
*c
= per_cpu_ptr(s
->cpu_slab
,
5253 slab
= READ_ONCE(c
->slab
);
5257 node
= slab_nid(slab
);
5258 if (flags
& SO_TOTAL
)
5260 else if (flags
& SO_OBJECTS
)
5268 #ifdef CONFIG_SLUB_CPU_PARTIAL
5269 slab
= slub_percpu_partial_read_once(c
);
5271 node
= slab_nid(slab
);
5272 if (flags
& SO_TOTAL
)
5274 else if (flags
& SO_OBJECTS
)
5286 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5287 * already held which will conflict with an existing lock order:
5289 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5291 * We don't really need mem_hotplug_lock (to hold off
5292 * slab_mem_going_offline_callback) here because slab's memory hot
5293 * unplug code doesn't destroy the kmem_cache->node[] data.
5296 #ifdef CONFIG_SLUB_DEBUG
5297 if (flags
& SO_ALL
) {
5298 struct kmem_cache_node
*n
;
5300 for_each_kmem_cache_node(s
, node
, n
) {
5302 if (flags
& SO_TOTAL
)
5303 x
= atomic_long_read(&n
->total_objects
);
5304 else if (flags
& SO_OBJECTS
)
5305 x
= atomic_long_read(&n
->total_objects
) -
5306 count_partial(n
, count_free
);
5308 x
= atomic_long_read(&n
->nr_slabs
);
5315 if (flags
& SO_PARTIAL
) {
5316 struct kmem_cache_node
*n
;
5318 for_each_kmem_cache_node(s
, node
, n
) {
5319 if (flags
& SO_TOTAL
)
5320 x
= count_partial(n
, count_total
);
5321 else if (flags
& SO_OBJECTS
)
5322 x
= count_partial(n
, count_inuse
);
5330 len
+= sysfs_emit_at(buf
, len
, "%lu", total
);
5332 for (node
= 0; node
< nr_node_ids
; node
++) {
5334 len
+= sysfs_emit_at(buf
, len
, " N%d=%lu",
5338 len
+= sysfs_emit_at(buf
, len
, "\n");
5344 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5345 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5347 struct slab_attribute
{
5348 struct attribute attr
;
5349 ssize_t (*show
)(struct kmem_cache
*s
, char *buf
);
5350 ssize_t (*store
)(struct kmem_cache
*s
, const char *x
, size_t count
);
5353 #define SLAB_ATTR_RO(_name) \
5354 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5356 #define SLAB_ATTR(_name) \
5357 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5359 static ssize_t
slab_size_show(struct kmem_cache
*s
, char *buf
)
5361 return sysfs_emit(buf
, "%u\n", s
->size
);
5363 SLAB_ATTR_RO(slab_size
);
5365 static ssize_t
align_show(struct kmem_cache
*s
, char *buf
)
5367 return sysfs_emit(buf
, "%u\n", s
->align
);
5369 SLAB_ATTR_RO(align
);
5371 static ssize_t
object_size_show(struct kmem_cache
*s
, char *buf
)
5373 return sysfs_emit(buf
, "%u\n", s
->object_size
);
5375 SLAB_ATTR_RO(object_size
);
5377 static ssize_t
objs_per_slab_show(struct kmem_cache
*s
, char *buf
)
5379 return sysfs_emit(buf
, "%u\n", oo_objects(s
->oo
));
5381 SLAB_ATTR_RO(objs_per_slab
);
5383 static ssize_t
order_show(struct kmem_cache
*s
, char *buf
)
5385 return sysfs_emit(buf
, "%u\n", oo_order(s
->oo
));
5387 SLAB_ATTR_RO(order
);
5389 static ssize_t
min_partial_show(struct kmem_cache
*s
, char *buf
)
5391 return sysfs_emit(buf
, "%lu\n", s
->min_partial
);
5394 static ssize_t
min_partial_store(struct kmem_cache
*s
, const char *buf
,
5400 err
= kstrtoul(buf
, 10, &min
);
5404 s
->min_partial
= min
;
5407 SLAB_ATTR(min_partial
);
5409 static ssize_t
cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5411 unsigned int nr_partial
= 0;
5412 #ifdef CONFIG_SLUB_CPU_PARTIAL
5413 nr_partial
= s
->cpu_partial
;
5416 return sysfs_emit(buf
, "%u\n", nr_partial
);
5419 static ssize_t
cpu_partial_store(struct kmem_cache
*s
, const char *buf
,
5422 unsigned int objects
;
5425 err
= kstrtouint(buf
, 10, &objects
);
5428 if (objects
&& !kmem_cache_has_cpu_partial(s
))
5431 slub_set_cpu_partial(s
, objects
);
5435 SLAB_ATTR(cpu_partial
);
5437 static ssize_t
ctor_show(struct kmem_cache
*s
, char *buf
)
5441 return sysfs_emit(buf
, "%pS\n", s
->ctor
);
5445 static ssize_t
aliases_show(struct kmem_cache
*s
, char *buf
)
5447 return sysfs_emit(buf
, "%d\n", s
->refcount
< 0 ? 0 : s
->refcount
- 1);
5449 SLAB_ATTR_RO(aliases
);
5451 static ssize_t
partial_show(struct kmem_cache
*s
, char *buf
)
5453 return show_slab_objects(s
, buf
, SO_PARTIAL
);
5455 SLAB_ATTR_RO(partial
);
5457 static ssize_t
cpu_slabs_show(struct kmem_cache
*s
, char *buf
)
5459 return show_slab_objects(s
, buf
, SO_CPU
);
5461 SLAB_ATTR_RO(cpu_slabs
);
5463 static ssize_t
objects_show(struct kmem_cache
*s
, char *buf
)
5465 return show_slab_objects(s
, buf
, SO_ALL
|SO_OBJECTS
);
5467 SLAB_ATTR_RO(objects
);
5469 static ssize_t
objects_partial_show(struct kmem_cache
*s
, char *buf
)
5471 return show_slab_objects(s
, buf
, SO_PARTIAL
|SO_OBJECTS
);
5473 SLAB_ATTR_RO(objects_partial
);
5475 static ssize_t
slabs_cpu_partial_show(struct kmem_cache
*s
, char *buf
)
5479 int cpu __maybe_unused
;
5482 #ifdef CONFIG_SLUB_CPU_PARTIAL
5483 for_each_online_cpu(cpu
) {
5486 slab
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5489 slabs
+= slab
->slabs
;
5493 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5494 objects
= (slabs
* oo_objects(s
->oo
)) / 2;
5495 len
+= sysfs_emit_at(buf
, len
, "%d(%d)", objects
, slabs
);
5497 #if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5498 for_each_online_cpu(cpu
) {
5501 slab
= slub_percpu_partial(per_cpu_ptr(s
->cpu_slab
, cpu
));
5503 slabs
= READ_ONCE(slab
->slabs
);
5504 objects
= (slabs
* oo_objects(s
->oo
)) / 2;
5505 len
+= sysfs_emit_at(buf
, len
, " C%d=%d(%d)",
5506 cpu
, objects
, slabs
);
5510 len
+= sysfs_emit_at(buf
, len
, "\n");
5514 SLAB_ATTR_RO(slabs_cpu_partial
);
5516 static ssize_t
reclaim_account_show(struct kmem_cache
*s
, char *buf
)
5518 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RECLAIM_ACCOUNT
));
5520 SLAB_ATTR_RO(reclaim_account
);
5522 static ssize_t
hwcache_align_show(struct kmem_cache
*s
, char *buf
)
5524 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_HWCACHE_ALIGN
));
5526 SLAB_ATTR_RO(hwcache_align
);
5528 #ifdef CONFIG_ZONE_DMA
5529 static ssize_t
cache_dma_show(struct kmem_cache
*s
, char *buf
)
5531 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CACHE_DMA
));
5533 SLAB_ATTR_RO(cache_dma
);
5536 static ssize_t
usersize_show(struct kmem_cache
*s
, char *buf
)
5538 return sysfs_emit(buf
, "%u\n", s
->usersize
);
5540 SLAB_ATTR_RO(usersize
);
5542 static ssize_t
destroy_by_rcu_show(struct kmem_cache
*s
, char *buf
)
5544 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TYPESAFE_BY_RCU
));
5546 SLAB_ATTR_RO(destroy_by_rcu
);
5548 #ifdef CONFIG_SLUB_DEBUG
5549 static ssize_t
slabs_show(struct kmem_cache
*s
, char *buf
)
5551 return show_slab_objects(s
, buf
, SO_ALL
);
5553 SLAB_ATTR_RO(slabs
);
5555 static ssize_t
total_objects_show(struct kmem_cache
*s
, char *buf
)
5557 return show_slab_objects(s
, buf
, SO_ALL
|SO_TOTAL
);
5559 SLAB_ATTR_RO(total_objects
);
5561 static ssize_t
sanity_checks_show(struct kmem_cache
*s
, char *buf
)
5563 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_CONSISTENCY_CHECKS
));
5565 SLAB_ATTR_RO(sanity_checks
);
5567 static ssize_t
trace_show(struct kmem_cache
*s
, char *buf
)
5569 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_TRACE
));
5571 SLAB_ATTR_RO(trace
);
5573 static ssize_t
red_zone_show(struct kmem_cache
*s
, char *buf
)
5575 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_RED_ZONE
));
5578 SLAB_ATTR_RO(red_zone
);
5580 static ssize_t
poison_show(struct kmem_cache
*s
, char *buf
)
5582 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_POISON
));
5585 SLAB_ATTR_RO(poison
);
5587 static ssize_t
store_user_show(struct kmem_cache
*s
, char *buf
)
5589 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_STORE_USER
));
5592 SLAB_ATTR_RO(store_user
);
5594 static ssize_t
validate_show(struct kmem_cache
*s
, char *buf
)
5599 static ssize_t
validate_store(struct kmem_cache
*s
,
5600 const char *buf
, size_t length
)
5604 if (buf
[0] == '1') {
5605 ret
= validate_slab_cache(s
);
5611 SLAB_ATTR(validate
);
5613 #endif /* CONFIG_SLUB_DEBUG */
5615 #ifdef CONFIG_FAILSLAB
5616 static ssize_t
failslab_show(struct kmem_cache
*s
, char *buf
)
5618 return sysfs_emit(buf
, "%d\n", !!(s
->flags
& SLAB_FAILSLAB
));
5620 SLAB_ATTR_RO(failslab
);
5623 static ssize_t
shrink_show(struct kmem_cache
*s
, char *buf
)
5628 static ssize_t
shrink_store(struct kmem_cache
*s
,
5629 const char *buf
, size_t length
)
5632 kmem_cache_shrink(s
);
5640 static ssize_t
remote_node_defrag_ratio_show(struct kmem_cache
*s
, char *buf
)
5642 return sysfs_emit(buf
, "%u\n", s
->remote_node_defrag_ratio
/ 10);
5645 static ssize_t
remote_node_defrag_ratio_store(struct kmem_cache
*s
,
5646 const char *buf
, size_t length
)
5651 err
= kstrtouint(buf
, 10, &ratio
);
5657 s
->remote_node_defrag_ratio
= ratio
* 10;
5661 SLAB_ATTR(remote_node_defrag_ratio
);
5664 #ifdef CONFIG_SLUB_STATS
5665 static int show_stat(struct kmem_cache
*s
, char *buf
, enum stat_item si
)
5667 unsigned long sum
= 0;
5670 int *data
= kmalloc_array(nr_cpu_ids
, sizeof(int), GFP_KERNEL
);
5675 for_each_online_cpu(cpu
) {
5676 unsigned x
= per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
];
5682 len
+= sysfs_emit_at(buf
, len
, "%lu", sum
);
5685 for_each_online_cpu(cpu
) {
5687 len
+= sysfs_emit_at(buf
, len
, " C%d=%u",
5692 len
+= sysfs_emit_at(buf
, len
, "\n");
5697 static void clear_stat(struct kmem_cache
*s
, enum stat_item si
)
5701 for_each_online_cpu(cpu
)
5702 per_cpu_ptr(s
->cpu_slab
, cpu
)->stat
[si
] = 0;
5705 #define STAT_ATTR(si, text) \
5706 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5708 return show_stat(s, buf, si); \
5710 static ssize_t text##_store(struct kmem_cache *s, \
5711 const char *buf, size_t length) \
5713 if (buf[0] != '0') \
5715 clear_stat(s, si); \
5720 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5721 STAT_ATTR(ALLOC_SLOWPATH
, alloc_slowpath
);
5722 STAT_ATTR(FREE_FASTPATH
, free_fastpath
);
5723 STAT_ATTR(FREE_SLOWPATH
, free_slowpath
);
5724 STAT_ATTR(FREE_FROZEN
, free_frozen
);
5725 STAT_ATTR(FREE_ADD_PARTIAL
, free_add_partial
);
5726 STAT_ATTR(FREE_REMOVE_PARTIAL
, free_remove_partial
);
5727 STAT_ATTR(ALLOC_FROM_PARTIAL
, alloc_from_partial
);
5728 STAT_ATTR(ALLOC_SLAB
, alloc_slab
);
5729 STAT_ATTR(ALLOC_REFILL
, alloc_refill
);
5730 STAT_ATTR(ALLOC_NODE_MISMATCH
, alloc_node_mismatch
);
5731 STAT_ATTR(FREE_SLAB
, free_slab
);
5732 STAT_ATTR(CPUSLAB_FLUSH
, cpuslab_flush
);
5733 STAT_ATTR(DEACTIVATE_FULL
, deactivate_full
);
5734 STAT_ATTR(DEACTIVATE_EMPTY
, deactivate_empty
);
5735 STAT_ATTR(DEACTIVATE_TO_HEAD
, deactivate_to_head
);
5736 STAT_ATTR(DEACTIVATE_TO_TAIL
, deactivate_to_tail
);
5737 STAT_ATTR(DEACTIVATE_REMOTE_FREES
, deactivate_remote_frees
);
5738 STAT_ATTR(DEACTIVATE_BYPASS
, deactivate_bypass
);
5739 STAT_ATTR(ORDER_FALLBACK
, order_fallback
);
5740 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL
, cmpxchg_double_cpu_fail
);
5741 STAT_ATTR(CMPXCHG_DOUBLE_FAIL
, cmpxchg_double_fail
);
5742 STAT_ATTR(CPU_PARTIAL_ALLOC
, cpu_partial_alloc
);
5743 STAT_ATTR(CPU_PARTIAL_FREE
, cpu_partial_free
);
5744 STAT_ATTR(CPU_PARTIAL_NODE
, cpu_partial_node
);
5745 STAT_ATTR(CPU_PARTIAL_DRAIN
, cpu_partial_drain
);
5746 #endif /* CONFIG_SLUB_STATS */
5748 static struct attribute
*slab_attrs
[] = {
5749 &slab_size_attr
.attr
,
5750 &object_size_attr
.attr
,
5751 &objs_per_slab_attr
.attr
,
5753 &min_partial_attr
.attr
,
5754 &cpu_partial_attr
.attr
,
5756 &objects_partial_attr
.attr
,
5758 &cpu_slabs_attr
.attr
,
5762 &hwcache_align_attr
.attr
,
5763 &reclaim_account_attr
.attr
,
5764 &destroy_by_rcu_attr
.attr
,
5766 &slabs_cpu_partial_attr
.attr
,
5767 #ifdef CONFIG_SLUB_DEBUG
5768 &total_objects_attr
.attr
,
5770 &sanity_checks_attr
.attr
,
5772 &red_zone_attr
.attr
,
5774 &store_user_attr
.attr
,
5775 &validate_attr
.attr
,
5777 #ifdef CONFIG_ZONE_DMA
5778 &cache_dma_attr
.attr
,
5781 &remote_node_defrag_ratio_attr
.attr
,
5783 #ifdef CONFIG_SLUB_STATS
5784 &alloc_fastpath_attr
.attr
,
5785 &alloc_slowpath_attr
.attr
,
5786 &free_fastpath_attr
.attr
,
5787 &free_slowpath_attr
.attr
,
5788 &free_frozen_attr
.attr
,
5789 &free_add_partial_attr
.attr
,
5790 &free_remove_partial_attr
.attr
,
5791 &alloc_from_partial_attr
.attr
,
5792 &alloc_slab_attr
.attr
,
5793 &alloc_refill_attr
.attr
,
5794 &alloc_node_mismatch_attr
.attr
,
5795 &free_slab_attr
.attr
,
5796 &cpuslab_flush_attr
.attr
,
5797 &deactivate_full_attr
.attr
,
5798 &deactivate_empty_attr
.attr
,
5799 &deactivate_to_head_attr
.attr
,
5800 &deactivate_to_tail_attr
.attr
,
5801 &deactivate_remote_frees_attr
.attr
,
5802 &deactivate_bypass_attr
.attr
,
5803 &order_fallback_attr
.attr
,
5804 &cmpxchg_double_fail_attr
.attr
,
5805 &cmpxchg_double_cpu_fail_attr
.attr
,
5806 &cpu_partial_alloc_attr
.attr
,
5807 &cpu_partial_free_attr
.attr
,
5808 &cpu_partial_node_attr
.attr
,
5809 &cpu_partial_drain_attr
.attr
,
5811 #ifdef CONFIG_FAILSLAB
5812 &failslab_attr
.attr
,
5814 &usersize_attr
.attr
,
5819 static const struct attribute_group slab_attr_group
= {
5820 .attrs
= slab_attrs
,
5823 static ssize_t
slab_attr_show(struct kobject
*kobj
,
5824 struct attribute
*attr
,
5827 struct slab_attribute
*attribute
;
5828 struct kmem_cache
*s
;
5831 attribute
= to_slab_attr(attr
);
5834 if (!attribute
->show
)
5837 err
= attribute
->show(s
, buf
);
5842 static ssize_t
slab_attr_store(struct kobject
*kobj
,
5843 struct attribute
*attr
,
5844 const char *buf
, size_t len
)
5846 struct slab_attribute
*attribute
;
5847 struct kmem_cache
*s
;
5850 attribute
= to_slab_attr(attr
);
5853 if (!attribute
->store
)
5856 err
= attribute
->store(s
, buf
, len
);
5860 static void kmem_cache_release(struct kobject
*k
)
5862 slab_kmem_cache_release(to_slab(k
));
5865 static const struct sysfs_ops slab_sysfs_ops
= {
5866 .show
= slab_attr_show
,
5867 .store
= slab_attr_store
,
5870 static struct kobj_type slab_ktype
= {
5871 .sysfs_ops
= &slab_sysfs_ops
,
5872 .release
= kmem_cache_release
,
5875 static struct kset
*slab_kset
;
5877 static inline struct kset
*cache_kset(struct kmem_cache
*s
)
5882 #define ID_STR_LENGTH 64
5884 /* Create a unique string id for a slab cache:
5886 * Format :[flags-]size
5888 static char *create_unique_id(struct kmem_cache
*s
)
5890 char *name
= kmalloc(ID_STR_LENGTH
, GFP_KERNEL
);
5897 * First flags affecting slabcache operations. We will only
5898 * get here for aliasable slabs so we do not need to support
5899 * too many flags. The flags here must cover all flags that
5900 * are matched during merging to guarantee that the id is
5903 if (s
->flags
& SLAB_CACHE_DMA
)
5905 if (s
->flags
& SLAB_CACHE_DMA32
)
5907 if (s
->flags
& SLAB_RECLAIM_ACCOUNT
)
5909 if (s
->flags
& SLAB_CONSISTENCY_CHECKS
)
5911 if (s
->flags
& SLAB_ACCOUNT
)
5915 p
+= sprintf(p
, "%07u", s
->size
);
5917 BUG_ON(p
> name
+ ID_STR_LENGTH
- 1);
5921 static int sysfs_slab_add(struct kmem_cache
*s
)
5925 struct kset
*kset
= cache_kset(s
);
5926 int unmergeable
= slab_unmergeable(s
);
5929 kobject_init(&s
->kobj
, &slab_ktype
);
5933 if (!unmergeable
&& disable_higher_order_debug
&&
5934 (slub_debug
& DEBUG_METADATA_FLAGS
))
5939 * Slabcache can never be merged so we can use the name proper.
5940 * This is typically the case for debug situations. In that
5941 * case we can catch duplicate names easily.
5943 sysfs_remove_link(&slab_kset
->kobj
, s
->name
);
5947 * Create a unique name for the slab as a target
5950 name
= create_unique_id(s
);
5953 s
->kobj
.kset
= kset
;
5954 err
= kobject_init_and_add(&s
->kobj
, &slab_ktype
, NULL
, "%s", name
);
5958 err
= sysfs_create_group(&s
->kobj
, &slab_attr_group
);
5963 /* Setup first alias */
5964 sysfs_slab_alias(s
, s
->name
);
5971 kobject_del(&s
->kobj
);
5975 void sysfs_slab_unlink(struct kmem_cache
*s
)
5977 if (slab_state
>= FULL
)
5978 kobject_del(&s
->kobj
);
5981 void sysfs_slab_release(struct kmem_cache
*s
)
5983 if (slab_state
>= FULL
)
5984 kobject_put(&s
->kobj
);
5988 * Need to buffer aliases during bootup until sysfs becomes
5989 * available lest we lose that information.
5991 struct saved_alias
{
5992 struct kmem_cache
*s
;
5994 struct saved_alias
*next
;
5997 static struct saved_alias
*alias_list
;
5999 static int sysfs_slab_alias(struct kmem_cache
*s
, const char *name
)
6001 struct saved_alias
*al
;
6003 if (slab_state
== FULL
) {
6005 * If we have a leftover link then remove it.
6007 sysfs_remove_link(&slab_kset
->kobj
, name
);
6008 return sysfs_create_link(&slab_kset
->kobj
, &s
->kobj
, name
);
6011 al
= kmalloc(sizeof(struct saved_alias
), GFP_KERNEL
);
6017 al
->next
= alias_list
;
6022 static int __init
slab_sysfs_init(void)
6024 struct kmem_cache
*s
;
6027 mutex_lock(&slab_mutex
);
6029 slab_kset
= kset_create_and_add("slab", NULL
, kernel_kobj
);
6031 mutex_unlock(&slab_mutex
);
6032 pr_err("Cannot register slab subsystem.\n");
6038 list_for_each_entry(s
, &slab_caches
, list
) {
6039 err
= sysfs_slab_add(s
);
6041 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6045 while (alias_list
) {
6046 struct saved_alias
*al
= alias_list
;
6048 alias_list
= alias_list
->next
;
6049 err
= sysfs_slab_alias(al
->s
, al
->name
);
6051 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6056 mutex_unlock(&slab_mutex
);
6060 __initcall(slab_sysfs_init
);
6061 #endif /* CONFIG_SYSFS */
6063 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6064 static int slab_debugfs_show(struct seq_file
*seq
, void *v
)
6066 struct loc_track
*t
= seq
->private;
6070 idx
= (unsigned long) t
->idx
;
6071 if (idx
< t
->count
) {
6074 seq_printf(seq
, "%7ld ", l
->count
);
6077 seq_printf(seq
, "%pS", (void *)l
->addr
);
6079 seq_puts(seq
, "<not-available>");
6081 if (l
->sum_time
!= l
->min_time
) {
6082 seq_printf(seq
, " age=%ld/%llu/%ld",
6083 l
->min_time
, div_u64(l
->sum_time
, l
->count
),
6086 seq_printf(seq
, " age=%ld", l
->min_time
);
6088 if (l
->min_pid
!= l
->max_pid
)
6089 seq_printf(seq
, " pid=%ld-%ld", l
->min_pid
, l
->max_pid
);
6091 seq_printf(seq
, " pid=%ld",
6094 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l
->cpus
)))
6095 seq_printf(seq
, " cpus=%*pbl",
6096 cpumask_pr_args(to_cpumask(l
->cpus
)));
6098 if (nr_online_nodes
> 1 && !nodes_empty(l
->nodes
))
6099 seq_printf(seq
, " nodes=%*pbl",
6100 nodemask_pr_args(&l
->nodes
));
6102 #ifdef CONFIG_STACKDEPOT
6104 depot_stack_handle_t handle
;
6105 unsigned long *entries
;
6106 unsigned int nr_entries
, j
;
6108 handle
= READ_ONCE(l
->handle
);
6110 nr_entries
= stack_depot_fetch(handle
, &entries
);
6111 seq_puts(seq
, "\n");
6112 for (j
= 0; j
< nr_entries
; j
++)
6113 seq_printf(seq
, " %pS\n", (void *)entries
[j
]);
6117 seq_puts(seq
, "\n");
6120 if (!idx
&& !t
->count
)
6121 seq_puts(seq
, "No data\n");
6126 static void slab_debugfs_stop(struct seq_file
*seq
, void *v
)
6130 static void *slab_debugfs_next(struct seq_file
*seq
, void *v
, loff_t
*ppos
)
6132 struct loc_track
*t
= seq
->private;
6135 if (*ppos
<= t
->count
)
6141 static int cmp_loc_by_count(const void *a
, const void *b
, const void *data
)
6143 struct location
*loc1
= (struct location
*)a
;
6144 struct location
*loc2
= (struct location
*)b
;
6146 if (loc1
->count
> loc2
->count
)
6152 static void *slab_debugfs_start(struct seq_file
*seq
, loff_t
*ppos
)
6154 struct loc_track
*t
= seq
->private;
6160 static const struct seq_operations slab_debugfs_sops
= {
6161 .start
= slab_debugfs_start
,
6162 .next
= slab_debugfs_next
,
6163 .stop
= slab_debugfs_stop
,
6164 .show
= slab_debugfs_show
,
6167 static int slab_debug_trace_open(struct inode
*inode
, struct file
*filep
)
6170 struct kmem_cache_node
*n
;
6171 enum track_item alloc
;
6173 struct loc_track
*t
= __seq_open_private(filep
, &slab_debugfs_sops
,
6174 sizeof(struct loc_track
));
6175 struct kmem_cache
*s
= file_inode(filep
)->i_private
;
6176 unsigned long *obj_map
;
6181 obj_map
= bitmap_alloc(oo_objects(s
->oo
), GFP_KERNEL
);
6183 seq_release_private(inode
, filep
);
6187 if (strcmp(filep
->f_path
.dentry
->d_name
.name
, "alloc_traces") == 0)
6188 alloc
= TRACK_ALLOC
;
6192 if (!alloc_loc_track(t
, PAGE_SIZE
/ sizeof(struct location
), GFP_KERNEL
)) {
6193 bitmap_free(obj_map
);
6194 seq_release_private(inode
, filep
);
6198 for_each_kmem_cache_node(s
, node
, n
) {
6199 unsigned long flags
;
6202 if (!atomic_long_read(&n
->nr_slabs
))
6205 spin_lock_irqsave(&n
->list_lock
, flags
);
6206 list_for_each_entry(slab
, &n
->partial
, slab_list
)
6207 process_slab(t
, s
, slab
, alloc
, obj_map
);
6208 list_for_each_entry(slab
, &n
->full
, slab_list
)
6209 process_slab(t
, s
, slab
, alloc
, obj_map
);
6210 spin_unlock_irqrestore(&n
->list_lock
, flags
);
6213 /* Sort locations by count */
6214 sort_r(t
->loc
, t
->count
, sizeof(struct location
),
6215 cmp_loc_by_count
, NULL
, NULL
);
6217 bitmap_free(obj_map
);
6221 static int slab_debug_trace_release(struct inode
*inode
, struct file
*file
)
6223 struct seq_file
*seq
= file
->private_data
;
6224 struct loc_track
*t
= seq
->private;
6227 return seq_release_private(inode
, file
);
6230 static const struct file_operations slab_debugfs_fops
= {
6231 .open
= slab_debug_trace_open
,
6233 .llseek
= seq_lseek
,
6234 .release
= slab_debug_trace_release
,
6237 static void debugfs_slab_add(struct kmem_cache
*s
)
6239 struct dentry
*slab_cache_dir
;
6241 if (unlikely(!slab_debugfs_root
))
6244 slab_cache_dir
= debugfs_create_dir(s
->name
, slab_debugfs_root
);
6246 debugfs_create_file("alloc_traces", 0400,
6247 slab_cache_dir
, s
, &slab_debugfs_fops
);
6249 debugfs_create_file("free_traces", 0400,
6250 slab_cache_dir
, s
, &slab_debugfs_fops
);
6253 void debugfs_slab_release(struct kmem_cache
*s
)
6255 debugfs_remove_recursive(debugfs_lookup(s
->name
, slab_debugfs_root
));
6258 static int __init
slab_debugfs_init(void)
6260 struct kmem_cache
*s
;
6262 slab_debugfs_root
= debugfs_create_dir("slab", NULL
);
6264 list_for_each_entry(s
, &slab_caches
, list
)
6265 if (s
->flags
& SLAB_STORE_USER
)
6266 debugfs_slab_add(s
);
6271 __initcall(slab_debugfs_init
);
6274 * The /proc/slabinfo ABI
6276 #ifdef CONFIG_SLUB_DEBUG
6277 void get_slabinfo(struct kmem_cache
*s
, struct slabinfo
*sinfo
)
6279 unsigned long nr_slabs
= 0;
6280 unsigned long nr_objs
= 0;
6281 unsigned long nr_free
= 0;
6283 struct kmem_cache_node
*n
;
6285 for_each_kmem_cache_node(s
, node
, n
) {
6286 nr_slabs
+= node_nr_slabs(n
);
6287 nr_objs
+= node_nr_objs(n
);
6288 nr_free
+= count_partial(n
, count_free
);
6291 sinfo
->active_objs
= nr_objs
- nr_free
;
6292 sinfo
->num_objs
= nr_objs
;
6293 sinfo
->active_slabs
= nr_slabs
;
6294 sinfo
->num_slabs
= nr_slabs
;
6295 sinfo
->objects_per_slab
= oo_objects(s
->oo
);
6296 sinfo
->cache_order
= oo_order(s
->oo
);
6299 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*s
)
6303 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
6304 size_t count
, loff_t
*ppos
)
6308 #endif /* CONFIG_SLUB_DEBUG */