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> /* mm_account_reclaimed_pages() */
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/kmsan.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ctype.h>
30 #include <linux/stackdepot.h>
31 #include <linux/debugobjects.h>
32 #include <linux/kallsyms.h>
33 #include <linux/kfence.h>
34 #include <linux/memory.h>
35 #include <linux/math64.h>
36 #include <linux/fault-inject.h>
37 #include <linux/stacktrace.h>
38 #include <linux/prefetch.h>
39 #include <linux/memcontrol.h>
40 #include <linux/random.h>
41 #include <kunit/test.h>
42 #include <kunit/test-bug.h>
43 #include <linux/sort.h>
45 #include <linux/debugfs.h>
46 #include <trace/events/kmem.h>
52 * 1. slab_mutex (Global Mutex)
53 * 2. node->list_lock (Spinlock)
54 * 3. kmem_cache->cpu_slab->lock (Local lock)
55 * 4. slab_lock(slab) (Only on some arches)
56 * 5. object_map_lock (Only for debugging)
60 * The role of the slab_mutex is to protect the list of all the slabs
61 * and to synchronize major metadata changes to slab cache structures.
62 * Also synchronizes memory hotplug callbacks.
66 * The slab_lock is a wrapper around the page lock, thus it is a bit
69 * The slab_lock is only used on arches that do not have the ability
70 * to do a cmpxchg_double. It only protects:
72 * A. slab->freelist -> List of free objects in a slab
73 * B. slab->inuse -> Number of objects in use
74 * C. slab->objects -> Number of objects in slab
75 * D. slab->frozen -> frozen state
79 * If a slab is frozen then it is exempt from list management. It is not
80 * on any list except per cpu partial list. The processor that froze the
81 * slab is the one who can perform list operations on the slab. Other
82 * processors may put objects onto the freelist but the processor that
83 * froze the slab is the only one that can retrieve the objects from the
88 * The list_lock protects the partial and full list on each node and
89 * the partial slab counter. If taken then no new slabs may be added or
90 * removed from the lists nor make the number of partial slabs be modified.
91 * (Note that the total number of slabs is an atomic value that may be
92 * modified without taking the list lock).
94 * The list_lock is a centralized lock and thus we avoid taking it as
95 * much as possible. As long as SLUB does not have to handle partial
96 * slabs, operations can continue without any centralized lock. F.e.
97 * allocating a long series of objects that fill up slabs does not require
100 * For debug caches, all allocations are forced to go through a list_lock
101 * protected region to serialize against concurrent validation.
103 * cpu_slab->lock local lock
105 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
106 * except the stat counters. This is a percpu structure manipulated only by
107 * the local cpu, so the lock protects against being preempted or interrupted
108 * by an irq. Fast path operations rely on lockless operations instead.
110 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
111 * which means the lockless fastpath cannot be used as it might interfere with
112 * an in-progress slow path operations. In this case the local lock is always
113 * taken but it still utilizes the freelist for the common operations.
117 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
118 * are fully lockless when satisfied from the percpu slab (and when
119 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
120 * They also don't disable preemption or migration or irqs. They rely on
121 * the transaction id (tid) field to detect being preempted or moved to
124 * irq, preemption, migration considerations
126 * Interrupts are disabled as part of list_lock or local_lock operations, or
127 * around the slab_lock operation, in order to make the slab allocator safe
128 * to use in the context of an irq.
130 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
131 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
132 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
133 * doesn't have to be revalidated in each section protected by the local lock.
135 * SLUB assigns one slab for allocation to each processor.
136 * Allocations only occur from these slabs called cpu slabs.
138 * Slabs with free elements are kept on a partial list and during regular
139 * operations no list for full slabs is used. If an object in a full slab is
140 * freed then the slab will show up again on the partial lists.
141 * We track full slabs for debugging purposes though because otherwise we
142 * cannot scan all objects.
144 * Slabs are freed when they become empty. Teardown and setup is
145 * minimal so we rely on the page allocators per cpu caches for
146 * fast frees and allocs.
148 * slab->frozen The slab is frozen and exempt from list processing.
149 * This means that the slab is dedicated to a purpose
150 * such as satisfying allocations for a specific
151 * processor. Objects may be freed in the slab while
152 * it is frozen but slab_free will then skip the usual
153 * list operations. It is up to the processor holding
154 * the slab to integrate the slab into the slab lists
155 * when the slab is no longer needed.
157 * One use of this flag is to mark slabs that are
158 * used for allocations. Then such a slab becomes a cpu
159 * slab. The cpu slab may be equipped with an additional
160 * freelist that allows lockless access to
161 * free objects in addition to the regular freelist
162 * that requires the slab lock.
164 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
165 * options set. This moves slab handling out of
166 * the fast path and disables lockless freelists.
170 * We could simply use migrate_disable()/enable() but as long as it's a
171 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
173 #ifndef CONFIG_PREEMPT_RT
174 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
175 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
176 #define USE_LOCKLESS_FAST_PATH() (true)
178 #define slub_get_cpu_ptr(var) \
183 #define slub_put_cpu_ptr(var) \
188 #define USE_LOCKLESS_FAST_PATH() (false)
191 #ifndef CONFIG_SLUB_TINY
192 #define __fastpath_inline __always_inline
194 #define __fastpath_inline
197 #ifdef CONFIG_SLUB_DEBUG
198 #ifdef CONFIG_SLUB_DEBUG_ON
199 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
201 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
203 #endif /* CONFIG_SLUB_DEBUG */
205 /* Structure holding parameters for get_partial() call chain */
206 struct partial_context {
209 unsigned int orig_size;
212 static inline bool kmem_cache_debug(struct kmem_cache *s)
214 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
217 static inline bool slub_debug_orig_size(struct kmem_cache *s)
219 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
220 (s->flags & SLAB_KMALLOC));
223 void *fixup_red_left(struct kmem_cache *s, void *p)
225 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
226 p += s->red_left_pad;
231 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
233 #ifdef CONFIG_SLUB_CPU_PARTIAL
234 return !kmem_cache_debug(s);
241 * Issues still to be resolved:
243 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
245 * - Variable sizing of the per node arrays
248 /* Enable to log cmpxchg failures */
249 #undef SLUB_DEBUG_CMPXCHG
251 #ifndef CONFIG_SLUB_TINY
253 * Minimum number of partial slabs. These will be left on the partial
254 * lists even if they are empty. kmem_cache_shrink may reclaim them.
256 #define MIN_PARTIAL 5
259 * Maximum number of desirable partial slabs.
260 * The existence of more partial slabs makes kmem_cache_shrink
261 * sort the partial list by the number of objects in use.
263 #define MAX_PARTIAL 10
265 #define MIN_PARTIAL 0
266 #define MAX_PARTIAL 0
269 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
270 SLAB_POISON | SLAB_STORE_USER)
273 * These debug flags cannot use CMPXCHG because there might be consistency
274 * issues when checking or reading debug information
276 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
281 * Debugging flags that require metadata to be stored in the slab. These get
282 * disabled when slub_debug=O is used and a cache's min order increases with
285 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
288 #define OO_MASK ((1 << OO_SHIFT) - 1)
289 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
291 /* Internal SLUB flags */
293 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
294 /* Use cmpxchg_double */
296 #ifdef system_has_freelist_aba
297 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
299 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0U)
303 * Tracking user of a slab.
305 #define TRACK_ADDRS_COUNT 16
307 unsigned long addr; /* Called from address */
308 #ifdef CONFIG_STACKDEPOT
309 depot_stack_handle_t handle;
311 int cpu; /* Was running on cpu */
312 int pid; /* Pid context */
313 unsigned long when; /* When did the operation occur */
316 enum track_item { TRACK_ALLOC, TRACK_FREE };
318 #ifdef SLAB_SUPPORTS_SYSFS
319 static int sysfs_slab_add(struct kmem_cache *);
320 static int sysfs_slab_alias(struct kmem_cache *, const char *);
322 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
323 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
327 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
328 static void debugfs_slab_add(struct kmem_cache *);
330 static inline void debugfs_slab_add(struct kmem_cache *s) { }
333 static inline void stat(const struct kmem_cache *s, enum stat_item si)
335 #ifdef CONFIG_SLUB_STATS
337 * The rmw is racy on a preemptible kernel but this is acceptable, so
338 * avoid this_cpu_add()'s irq-disable overhead.
340 raw_cpu_inc(s->cpu_slab->stat[si]);
345 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
346 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
347 * differ during memory hotplug/hotremove operations.
348 * Protected by slab_mutex.
350 static nodemask_t slab_nodes;
352 #ifndef CONFIG_SLUB_TINY
354 * Workqueue used for flush_cpu_slab().
356 static struct workqueue_struct *flushwq;
359 /********************************************************************
360 * Core slab cache functions
361 *******************************************************************/
364 * Returns freelist pointer (ptr). With hardening, this is obfuscated
365 * with an XOR of the address where the pointer is held and a per-cache
368 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
369 unsigned long ptr_addr)
371 #ifdef CONFIG_SLAB_FREELIST_HARDENED
373 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
374 * Normally, this doesn't cause any issues, as both set_freepointer()
375 * and get_freepointer() are called with a pointer with the same tag.
376 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
377 * example, when __free_slub() iterates over objects in a cache, it
378 * passes untagged pointers to check_object(). check_object() in turns
379 * calls get_freepointer() with an untagged pointer, which causes the
380 * freepointer to be restored incorrectly.
382 return (void *)((unsigned long)ptr ^ s->random ^
383 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
389 /* Returns the freelist pointer recorded at location ptr_addr. */
390 static inline void *freelist_dereference(const struct kmem_cache *s,
393 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
394 (unsigned long)ptr_addr);
397 static inline void *get_freepointer(struct kmem_cache *s, void *object)
399 object = kasan_reset_tag(object);
400 return freelist_dereference(s, object + s->offset);
403 #ifndef CONFIG_SLUB_TINY
404 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
406 prefetchw(object + s->offset);
411 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
412 * pointer value in the case the current thread loses the race for the next
413 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
414 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
415 * KMSAN will still check all arguments of cmpxchg because of imperfect
416 * handling of inline assembly.
417 * To work around this problem, we apply __no_kmsan_checks to ensure that
418 * get_freepointer_safe() returns initialized memory.
421 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
423 unsigned long freepointer_addr;
426 if (!debug_pagealloc_enabled_static())
427 return get_freepointer(s, object);
429 object = kasan_reset_tag(object);
430 freepointer_addr = (unsigned long)object + s->offset;
431 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
432 return freelist_ptr(s, p, freepointer_addr);
435 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
437 unsigned long freeptr_addr = (unsigned long)object + s->offset;
439 #ifdef CONFIG_SLAB_FREELIST_HARDENED
440 BUG_ON(object == fp); /* naive detection of double free or corruption */
443 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
444 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
447 /* Loop over all objects in a slab */
448 #define for_each_object(__p, __s, __addr, __objects) \
449 for (__p = fixup_red_left(__s, __addr); \
450 __p < (__addr) + (__objects) * (__s)->size; \
453 static inline unsigned int order_objects(unsigned int order, unsigned int size)
455 return ((unsigned int)PAGE_SIZE << order) / size;
458 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
461 struct kmem_cache_order_objects x = {
462 (order << OO_SHIFT) + order_objects(order, size)
468 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
470 return x.x >> OO_SHIFT;
473 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
475 return x.x & OO_MASK;
478 #ifdef CONFIG_SLUB_CPU_PARTIAL
479 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
481 unsigned int nr_slabs;
483 s->cpu_partial = nr_objects;
486 * We take the number of objects but actually limit the number of
487 * slabs on the per cpu partial list, in order to limit excessive
488 * growth of the list. For simplicity we assume that the slabs will
491 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
492 s->cpu_partial_slabs = nr_slabs;
496 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
499 #endif /* CONFIG_SLUB_CPU_PARTIAL */
502 * Per slab locking using the pagelock
504 static __always_inline void slab_lock(struct slab *slab)
506 struct page *page = slab_page(slab);
508 VM_BUG_ON_PAGE(PageTail(page), page);
509 bit_spin_lock(PG_locked, &page->flags);
512 static __always_inline void slab_unlock(struct slab *slab)
514 struct page *page = slab_page(slab);
516 VM_BUG_ON_PAGE(PageTail(page), page);
517 __bit_spin_unlock(PG_locked, &page->flags);
521 __update_freelist_fast(struct slab *slab,
522 void *freelist_old, unsigned long counters_old,
523 void *freelist_new, unsigned long counters_new)
525 #ifdef system_has_freelist_aba
526 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
527 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
529 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
536 __update_freelist_slow(struct slab *slab,
537 void *freelist_old, unsigned long counters_old,
538 void *freelist_new, unsigned long counters_new)
543 if (slab->freelist == freelist_old &&
544 slab->counters == counters_old) {
545 slab->freelist = freelist_new;
546 slab->counters = counters_new;
555 * Interrupts must be disabled (for the fallback code to work right), typically
556 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
557 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
558 * allocation/ free operation in hardirq context. Therefore nothing can
559 * interrupt the operation.
561 static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
562 void *freelist_old, unsigned long counters_old,
563 void *freelist_new, unsigned long counters_new,
568 if (USE_LOCKLESS_FAST_PATH())
569 lockdep_assert_irqs_disabled();
571 if (s->flags & __CMPXCHG_DOUBLE) {
572 ret = __update_freelist_fast(slab, freelist_old, counters_old,
573 freelist_new, counters_new);
575 ret = __update_freelist_slow(slab, freelist_old, counters_old,
576 freelist_new, counters_new);
582 stat(s, CMPXCHG_DOUBLE_FAIL);
584 #ifdef SLUB_DEBUG_CMPXCHG
585 pr_info("%s %s: cmpxchg double redo ", n, s->name);
591 static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
592 void *freelist_old, unsigned long counters_old,
593 void *freelist_new, unsigned long counters_new,
598 if (s->flags & __CMPXCHG_DOUBLE) {
599 ret = __update_freelist_fast(slab, freelist_old, counters_old,
600 freelist_new, counters_new);
604 local_irq_save(flags);
605 ret = __update_freelist_slow(slab, freelist_old, counters_old,
606 freelist_new, counters_new);
607 local_irq_restore(flags);
613 stat(s, CMPXCHG_DOUBLE_FAIL);
615 #ifdef SLUB_DEBUG_CMPXCHG
616 pr_info("%s %s: cmpxchg double redo ", n, s->name);
622 #ifdef CONFIG_SLUB_DEBUG
623 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
624 static DEFINE_SPINLOCK(object_map_lock);
626 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
629 void *addr = slab_address(slab);
632 bitmap_zero(obj_map, slab->objects);
634 for (p = slab->freelist; p; p = get_freepointer(s, p))
635 set_bit(__obj_to_index(s, addr, p), obj_map);
638 #if IS_ENABLED(CONFIG_KUNIT)
639 static bool slab_add_kunit_errors(void)
641 struct kunit_resource *resource;
643 if (!kunit_get_current_test())
646 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
650 (*(int *)resource->data)++;
651 kunit_put_resource(resource);
655 static inline bool slab_add_kunit_errors(void) { return false; }
658 static inline unsigned int size_from_object(struct kmem_cache *s)
660 if (s->flags & SLAB_RED_ZONE)
661 return s->size - s->red_left_pad;
666 static inline void *restore_red_left(struct kmem_cache *s, void *p)
668 if (s->flags & SLAB_RED_ZONE)
669 p -= s->red_left_pad;
677 #if defined(CONFIG_SLUB_DEBUG_ON)
678 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
680 static slab_flags_t slub_debug;
683 static char *slub_debug_string;
684 static int disable_higher_order_debug;
687 * slub is about to manipulate internal object metadata. This memory lies
688 * outside the range of the allocated object, so accessing it would normally
689 * be reported by kasan as a bounds error. metadata_access_enable() is used
690 * to tell kasan that these accesses are OK.
692 static inline void metadata_access_enable(void)
694 kasan_disable_current();
697 static inline void metadata_access_disable(void)
699 kasan_enable_current();
706 /* Verify that a pointer has an address that is valid within a slab page */
707 static inline int check_valid_pointer(struct kmem_cache *s,
708 struct slab *slab, void *object)
715 base = slab_address(slab);
716 object = kasan_reset_tag(object);
717 object = restore_red_left(s, object);
718 if (object < base || object >= base + slab->objects * s->size ||
719 (object - base) % s->size) {
726 static void print_section(char *level, char *text, u8 *addr,
729 metadata_access_enable();
730 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
731 16, 1, kasan_reset_tag((void *)addr), length, 1);
732 metadata_access_disable();
736 * See comment in calculate_sizes().
738 static inline bool freeptr_outside_object(struct kmem_cache *s)
740 return s->offset >= s->inuse;
744 * Return offset of the end of info block which is inuse + free pointer if
745 * not overlapping with object.
747 static inline unsigned int get_info_end(struct kmem_cache *s)
749 if (freeptr_outside_object(s))
750 return s->inuse + sizeof(void *);
755 static struct track *get_track(struct kmem_cache *s, void *object,
756 enum track_item alloc)
760 p = object + get_info_end(s);
762 return kasan_reset_tag(p + alloc);
765 #ifdef CONFIG_STACKDEPOT
766 static noinline depot_stack_handle_t set_track_prepare(void)
768 depot_stack_handle_t handle;
769 unsigned long entries[TRACK_ADDRS_COUNT];
770 unsigned int nr_entries;
772 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
773 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
778 static inline depot_stack_handle_t set_track_prepare(void)
784 static void set_track_update(struct kmem_cache *s, void *object,
785 enum track_item alloc, unsigned long addr,
786 depot_stack_handle_t handle)
788 struct track *p = get_track(s, object, alloc);
790 #ifdef CONFIG_STACKDEPOT
794 p->cpu = smp_processor_id();
795 p->pid = current->pid;
799 static __always_inline void set_track(struct kmem_cache *s, void *object,
800 enum track_item alloc, unsigned long addr)
802 depot_stack_handle_t handle = set_track_prepare();
804 set_track_update(s, object, alloc, addr, handle);
807 static void init_tracking(struct kmem_cache *s, void *object)
811 if (!(s->flags & SLAB_STORE_USER))
814 p = get_track(s, object, TRACK_ALLOC);
815 memset(p, 0, 2*sizeof(struct track));
818 static void print_track(const char *s, struct track *t, unsigned long pr_time)
820 depot_stack_handle_t handle __maybe_unused;
825 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
826 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
827 #ifdef CONFIG_STACKDEPOT
828 handle = READ_ONCE(t->handle);
830 stack_depot_print(handle);
832 pr_err("object allocation/free stack trace missing\n");
836 void print_tracking(struct kmem_cache *s, void *object)
838 unsigned long pr_time = jiffies;
839 if (!(s->flags & SLAB_STORE_USER))
842 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
843 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
846 static void print_slab_info(const struct slab *slab)
848 struct folio *folio = (struct folio *)slab_folio(slab);
850 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
851 slab, slab->objects, slab->inuse, slab->freelist,
852 folio_flags(folio, 0));
856 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
857 * family will round up the real request size to these fixed ones, so
858 * there could be an extra area than what is requested. Save the original
859 * request size in the meta data area, for better debug and sanity check.
861 static inline void set_orig_size(struct kmem_cache *s,
862 void *object, unsigned int orig_size)
864 void *p = kasan_reset_tag(object);
866 if (!slub_debug_orig_size(s))
869 #ifdef CONFIG_KASAN_GENERIC
871 * KASAN could save its free meta data in object's data area at
872 * offset 0, if the size is larger than 'orig_size', it will
873 * overlap the data redzone in [orig_size+1, object_size], and
874 * the check should be skipped.
876 if (kasan_metadata_size(s, true) > orig_size)
877 orig_size = s->object_size;
880 p += get_info_end(s);
881 p += sizeof(struct track) * 2;
883 *(unsigned int *)p = orig_size;
886 static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
888 void *p = kasan_reset_tag(object);
890 if (!slub_debug_orig_size(s))
891 return s->object_size;
893 p += get_info_end(s);
894 p += sizeof(struct track) * 2;
896 return *(unsigned int *)p;
899 void skip_orig_size_check(struct kmem_cache *s, const void *object)
901 set_orig_size(s, (void *)object, s->object_size);
904 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
906 struct va_format vaf;
912 pr_err("=============================================================================\n");
913 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
914 pr_err("-----------------------------------------------------------------------------\n\n");
919 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
921 struct va_format vaf;
924 if (slab_add_kunit_errors())
930 pr_err("FIX %s: %pV\n", s->name, &vaf);
934 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
936 unsigned int off; /* Offset of last byte */
937 u8 *addr = slab_address(slab);
939 print_tracking(s, p);
941 print_slab_info(slab);
943 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
944 p, p - addr, get_freepointer(s, p));
946 if (s->flags & SLAB_RED_ZONE)
947 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
949 else if (p > addr + 16)
950 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
952 print_section(KERN_ERR, "Object ", p,
953 min_t(unsigned int, s->object_size, PAGE_SIZE));
954 if (s->flags & SLAB_RED_ZONE)
955 print_section(KERN_ERR, "Redzone ", p + s->object_size,
956 s->inuse - s->object_size);
958 off = get_info_end(s);
960 if (s->flags & SLAB_STORE_USER)
961 off += 2 * sizeof(struct track);
963 if (slub_debug_orig_size(s))
964 off += sizeof(unsigned int);
966 off += kasan_metadata_size(s, false);
968 if (off != size_from_object(s))
969 /* Beginning of the filler is the free pointer */
970 print_section(KERN_ERR, "Padding ", p + off,
971 size_from_object(s) - off);
976 static void object_err(struct kmem_cache *s, struct slab *slab,
977 u8 *object, char *reason)
979 if (slab_add_kunit_errors())
982 slab_bug(s, "%s", reason);
983 print_trailer(s, slab, object);
984 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
987 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
988 void **freelist, void *nextfree)
990 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
991 !check_valid_pointer(s, slab, nextfree) && freelist) {
992 object_err(s, slab, *freelist, "Freechain corrupt");
994 slab_fix(s, "Isolate corrupted freechain");
1001 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1002 const char *fmt, ...)
1007 if (slab_add_kunit_errors())
1010 va_start(args, fmt);
1011 vsnprintf(buf, sizeof(buf), fmt, args);
1013 slab_bug(s, "%s", buf);
1014 print_slab_info(slab);
1016 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1019 static void init_object(struct kmem_cache *s, void *object, u8 val)
1021 u8 *p = kasan_reset_tag(object);
1022 unsigned int poison_size = s->object_size;
1024 if (s->flags & SLAB_RED_ZONE) {
1025 memset(p - s->red_left_pad, val, s->red_left_pad);
1027 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1029 * Redzone the extra allocated space by kmalloc than
1030 * requested, and the poison size will be limited to
1031 * the original request size accordingly.
1033 poison_size = get_orig_size(s, object);
1037 if (s->flags & __OBJECT_POISON) {
1038 memset(p, POISON_FREE, poison_size - 1);
1039 p[poison_size - 1] = POISON_END;
1042 if (s->flags & SLAB_RED_ZONE)
1043 memset(p + poison_size, val, s->inuse - poison_size);
1046 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1047 void *from, void *to)
1049 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1050 memset(from, data, to - from);
1053 static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1054 u8 *object, char *what,
1055 u8 *start, unsigned int value, unsigned int bytes)
1059 u8 *addr = slab_address(slab);
1061 metadata_access_enable();
1062 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1063 metadata_access_disable();
1067 end = start + bytes;
1068 while (end > fault && end[-1] == value)
1071 if (slab_add_kunit_errors())
1072 goto skip_bug_print;
1074 slab_bug(s, "%s overwritten", what);
1075 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1076 fault, end - 1, fault - addr,
1078 print_trailer(s, slab, object);
1079 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1082 restore_bytes(s, what, value, fault, end);
1090 * Bytes of the object to be managed.
1091 * If the freepointer may overlay the object then the free
1092 * pointer is at the middle of the object.
1094 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1097 * object + s->object_size
1098 * Padding to reach word boundary. This is also used for Redzoning.
1099 * Padding is extended by another word if Redzoning is enabled and
1100 * object_size == inuse.
1102 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1103 * 0xcc (RED_ACTIVE) for objects in use.
1106 * Meta data starts here.
1108 * A. Free pointer (if we cannot overwrite object on free)
1109 * B. Tracking data for SLAB_STORE_USER
1110 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1111 * D. Padding to reach required alignment boundary or at minimum
1112 * one word if debugging is on to be able to detect writes
1113 * before the word boundary.
1115 * Padding is done using 0x5a (POISON_INUSE)
1118 * Nothing is used beyond s->size.
1120 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1121 * ignored. And therefore no slab options that rely on these boundaries
1122 * may be used with merged slabcaches.
1125 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1127 unsigned long off = get_info_end(s); /* The end of info */
1129 if (s->flags & SLAB_STORE_USER) {
1130 /* We also have user information there */
1131 off += 2 * sizeof(struct track);
1133 if (s->flags & SLAB_KMALLOC)
1134 off += sizeof(unsigned int);
1137 off += kasan_metadata_size(s, false);
1139 if (size_from_object(s) == off)
1142 return check_bytes_and_report(s, slab, p, "Object padding",
1143 p + off, POISON_INUSE, size_from_object(s) - off);
1146 /* Check the pad bytes at the end of a slab page */
1147 static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1156 if (!(s->flags & SLAB_POISON))
1159 start = slab_address(slab);
1160 length = slab_size(slab);
1161 end = start + length;
1162 remainder = length % s->size;
1166 pad = end - remainder;
1167 metadata_access_enable();
1168 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1169 metadata_access_disable();
1172 while (end > fault && end[-1] == POISON_INUSE)
1175 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1176 fault, end - 1, fault - start);
1177 print_section(KERN_ERR, "Padding ", pad, remainder);
1179 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1182 static int check_object(struct kmem_cache *s, struct slab *slab,
1183 void *object, u8 val)
1186 u8 *endobject = object + s->object_size;
1187 unsigned int orig_size;
1189 if (s->flags & SLAB_RED_ZONE) {
1190 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1191 object - s->red_left_pad, val, s->red_left_pad))
1194 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1195 endobject, val, s->inuse - s->object_size))
1198 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1199 orig_size = get_orig_size(s, object);
1201 if (s->object_size > orig_size &&
1202 !check_bytes_and_report(s, slab, object,
1203 "kmalloc Redzone", p + orig_size,
1204 val, s->object_size - orig_size)) {
1209 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1210 check_bytes_and_report(s, slab, p, "Alignment padding",
1211 endobject, POISON_INUSE,
1212 s->inuse - s->object_size);
1216 if (s->flags & SLAB_POISON) {
1217 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1218 (!check_bytes_and_report(s, slab, p, "Poison", p,
1219 POISON_FREE, s->object_size - 1) ||
1220 !check_bytes_and_report(s, slab, p, "End Poison",
1221 p + s->object_size - 1, POISON_END, 1)))
1224 * check_pad_bytes cleans up on its own.
1226 check_pad_bytes(s, slab, p);
1229 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1231 * Object and freepointer overlap. Cannot check
1232 * freepointer while object is allocated.
1236 /* Check free pointer validity */
1237 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1238 object_err(s, slab, p, "Freepointer corrupt");
1240 * No choice but to zap it and thus lose the remainder
1241 * of the free objects in this slab. May cause
1242 * another error because the object count is now wrong.
1244 set_freepointer(s, p, NULL);
1250 static int check_slab(struct kmem_cache *s, struct slab *slab)
1254 if (!folio_test_slab(slab_folio(slab))) {
1255 slab_err(s, slab, "Not a valid slab page");
1259 maxobj = order_objects(slab_order(slab), s->size);
1260 if (slab->objects > maxobj) {
1261 slab_err(s, slab, "objects %u > max %u",
1262 slab->objects, maxobj);
1265 if (slab->inuse > slab->objects) {
1266 slab_err(s, slab, "inuse %u > max %u",
1267 slab->inuse, slab->objects);
1270 /* Slab_pad_check fixes things up after itself */
1271 slab_pad_check(s, slab);
1276 * Determine if a certain object in a slab is on the freelist. Must hold the
1277 * slab lock to guarantee that the chains are in a consistent state.
1279 static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1283 void *object = NULL;
1286 fp = slab->freelist;
1287 while (fp && nr <= slab->objects) {
1290 if (!check_valid_pointer(s, slab, fp)) {
1292 object_err(s, slab, object,
1293 "Freechain corrupt");
1294 set_freepointer(s, object, NULL);
1296 slab_err(s, slab, "Freepointer corrupt");
1297 slab->freelist = NULL;
1298 slab->inuse = slab->objects;
1299 slab_fix(s, "Freelist cleared");
1305 fp = get_freepointer(s, object);
1309 max_objects = order_objects(slab_order(slab), s->size);
1310 if (max_objects > MAX_OBJS_PER_PAGE)
1311 max_objects = MAX_OBJS_PER_PAGE;
1313 if (slab->objects != max_objects) {
1314 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1315 slab->objects, max_objects);
1316 slab->objects = max_objects;
1317 slab_fix(s, "Number of objects adjusted");
1319 if (slab->inuse != slab->objects - nr) {
1320 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1321 slab->inuse, slab->objects - nr);
1322 slab->inuse = slab->objects - nr;
1323 slab_fix(s, "Object count adjusted");
1325 return search == NULL;
1328 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1331 if (s->flags & SLAB_TRACE) {
1332 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1334 alloc ? "alloc" : "free",
1335 object, slab->inuse,
1339 print_section(KERN_INFO, "Object ", (void *)object,
1347 * Tracking of fully allocated slabs for debugging purposes.
1349 static void add_full(struct kmem_cache *s,
1350 struct kmem_cache_node *n, struct slab *slab)
1352 if (!(s->flags & SLAB_STORE_USER))
1355 lockdep_assert_held(&n->list_lock);
1356 list_add(&slab->slab_list, &n->full);
1359 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1361 if (!(s->flags & SLAB_STORE_USER))
1364 lockdep_assert_held(&n->list_lock);
1365 list_del(&slab->slab_list);
1368 /* Tracking of the number of slabs for debugging purposes */
1369 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1371 struct kmem_cache_node *n = get_node(s, node);
1373 return atomic_long_read(&n->nr_slabs);
1376 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1378 return atomic_long_read(&n->nr_slabs);
1381 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1383 struct kmem_cache_node *n = get_node(s, node);
1386 * May be called early in order to allocate a slab for the
1387 * kmem_cache_node structure. Solve the chicken-egg
1388 * dilemma by deferring the increment of the count during
1389 * bootstrap (see early_kmem_cache_node_alloc).
1392 atomic_long_inc(&n->nr_slabs);
1393 atomic_long_add(objects, &n->total_objects);
1396 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1398 struct kmem_cache_node *n = get_node(s, node);
1400 atomic_long_dec(&n->nr_slabs);
1401 atomic_long_sub(objects, &n->total_objects);
1404 /* Object debug checks for alloc/free paths */
1405 static void setup_object_debug(struct kmem_cache *s, void *object)
1407 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1410 init_object(s, object, SLUB_RED_INACTIVE);
1411 init_tracking(s, object);
1415 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1417 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1420 metadata_access_enable();
1421 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1422 metadata_access_disable();
1425 static inline int alloc_consistency_checks(struct kmem_cache *s,
1426 struct slab *slab, void *object)
1428 if (!check_slab(s, slab))
1431 if (!check_valid_pointer(s, slab, object)) {
1432 object_err(s, slab, object, "Freelist Pointer check fails");
1436 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1442 static noinline bool alloc_debug_processing(struct kmem_cache *s,
1443 struct slab *slab, void *object, int orig_size)
1445 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1446 if (!alloc_consistency_checks(s, slab, object))
1450 /* Success. Perform special debug activities for allocs */
1451 trace(s, slab, object, 1);
1452 set_orig_size(s, object, orig_size);
1453 init_object(s, object, SLUB_RED_ACTIVE);
1457 if (folio_test_slab(slab_folio(slab))) {
1459 * If this is a slab page then lets do the best we can
1460 * to avoid issues in the future. Marking all objects
1461 * as used avoids touching the remaining objects.
1463 slab_fix(s, "Marking all objects used");
1464 slab->inuse = slab->objects;
1465 slab->freelist = NULL;
1470 static inline int free_consistency_checks(struct kmem_cache *s,
1471 struct slab *slab, void *object, unsigned long addr)
1473 if (!check_valid_pointer(s, slab, object)) {
1474 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1478 if (on_freelist(s, slab, object)) {
1479 object_err(s, slab, object, "Object already free");
1483 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1486 if (unlikely(s != slab->slab_cache)) {
1487 if (!folio_test_slab(slab_folio(slab))) {
1488 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1490 } else if (!slab->slab_cache) {
1491 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1495 object_err(s, slab, object,
1496 "page slab pointer corrupt.");
1503 * Parse a block of slub_debug options. Blocks are delimited by ';'
1505 * @str: start of block
1506 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1507 * @slabs: return start of list of slabs, or NULL when there's no list
1508 * @init: assume this is initial parsing and not per-kmem-create parsing
1510 * returns the start of next block if there's any, or NULL
1513 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1515 bool higher_order_disable = false;
1517 /* Skip any completely empty blocks */
1518 while (*str && *str == ';')
1523 * No options but restriction on slabs. This means full
1524 * debugging for slabs matching a pattern.
1526 *flags = DEBUG_DEFAULT_FLAGS;
1531 /* Determine which debug features should be switched on */
1532 for (; *str && *str != ',' && *str != ';'; str++) {
1533 switch (tolower(*str)) {
1538 *flags |= SLAB_CONSISTENCY_CHECKS;
1541 *flags |= SLAB_RED_ZONE;
1544 *flags |= SLAB_POISON;
1547 *flags |= SLAB_STORE_USER;
1550 *flags |= SLAB_TRACE;
1553 *flags |= SLAB_FAILSLAB;
1557 * Avoid enabling debugging on caches if its minimum
1558 * order would increase as a result.
1560 higher_order_disable = true;
1564 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1573 /* Skip over the slab list */
1574 while (*str && *str != ';')
1577 /* Skip any completely empty blocks */
1578 while (*str && *str == ';')
1581 if (init && higher_order_disable)
1582 disable_higher_order_debug = 1;
1590 static int __init setup_slub_debug(char *str)
1593 slab_flags_t global_flags;
1596 bool global_slub_debug_changed = false;
1597 bool slab_list_specified = false;
1599 global_flags = DEBUG_DEFAULT_FLAGS;
1600 if (*str++ != '=' || !*str)
1602 * No options specified. Switch on full debugging.
1608 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1611 global_flags = flags;
1612 global_slub_debug_changed = true;
1614 slab_list_specified = true;
1615 if (flags & SLAB_STORE_USER)
1616 stack_depot_request_early_init();
1621 * For backwards compatibility, a single list of flags with list of
1622 * slabs means debugging is only changed for those slabs, so the global
1623 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1624 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1625 * long as there is no option specifying flags without a slab list.
1627 if (slab_list_specified) {
1628 if (!global_slub_debug_changed)
1629 global_flags = slub_debug;
1630 slub_debug_string = saved_str;
1633 slub_debug = global_flags;
1634 if (slub_debug & SLAB_STORE_USER)
1635 stack_depot_request_early_init();
1636 if (slub_debug != 0 || slub_debug_string)
1637 static_branch_enable(&slub_debug_enabled);
1639 static_branch_disable(&slub_debug_enabled);
1640 if ((static_branch_unlikely(&init_on_alloc) ||
1641 static_branch_unlikely(&init_on_free)) &&
1642 (slub_debug & SLAB_POISON))
1643 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1647 __setup("slub_debug", setup_slub_debug);
1650 * kmem_cache_flags - apply debugging options to the cache
1651 * @object_size: the size of an object without meta data
1652 * @flags: flags to set
1653 * @name: name of the cache
1655 * Debug option(s) are applied to @flags. In addition to the debug
1656 * option(s), if a slab name (or multiple) is specified i.e.
1657 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1658 * then only the select slabs will receive the debug option(s).
1660 slab_flags_t kmem_cache_flags(unsigned int object_size,
1661 slab_flags_t flags, const char *name)
1666 slab_flags_t block_flags;
1667 slab_flags_t slub_debug_local = slub_debug;
1669 if (flags & SLAB_NO_USER_FLAGS)
1673 * If the slab cache is for debugging (e.g. kmemleak) then
1674 * don't store user (stack trace) information by default,
1675 * but let the user enable it via the command line below.
1677 if (flags & SLAB_NOLEAKTRACE)
1678 slub_debug_local &= ~SLAB_STORE_USER;
1681 next_block = slub_debug_string;
1682 /* Go through all blocks of debug options, see if any matches our slab's name */
1683 while (next_block) {
1684 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1687 /* Found a block that has a slab list, search it */
1692 end = strchrnul(iter, ',');
1693 if (next_block && next_block < end)
1694 end = next_block - 1;
1696 glob = strnchr(iter, end - iter, '*');
1698 cmplen = glob - iter;
1700 cmplen = max_t(size_t, len, (end - iter));
1702 if (!strncmp(name, iter, cmplen)) {
1703 flags |= block_flags;
1707 if (!*end || *end == ';')
1713 return flags | slub_debug_local;
1715 #else /* !CONFIG_SLUB_DEBUG */
1716 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1718 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1720 static inline bool alloc_debug_processing(struct kmem_cache *s,
1721 struct slab *slab, void *object, int orig_size) { return true; }
1723 static inline bool free_debug_processing(struct kmem_cache *s,
1724 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1725 unsigned long addr, depot_stack_handle_t handle) { return true; }
1727 static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1728 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1729 void *object, u8 val) { return 1; }
1730 static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1731 static inline void set_track(struct kmem_cache *s, void *object,
1732 enum track_item alloc, unsigned long addr) {}
1733 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1734 struct slab *slab) {}
1735 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1736 struct slab *slab) {}
1737 slab_flags_t kmem_cache_flags(unsigned int object_size,
1738 slab_flags_t flags, const char *name)
1742 #define slub_debug 0
1744 #define disable_higher_order_debug 0
1746 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1748 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1750 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1752 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1755 #ifndef CONFIG_SLUB_TINY
1756 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1757 void **freelist, void *nextfree)
1762 #endif /* CONFIG_SLUB_DEBUG */
1765 * Hooks for other subsystems that check memory allocations. In a typical
1766 * production configuration these hooks all should produce no code at all.
1768 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1771 kmemleak_free_recursive(x, s->flags);
1772 kmsan_slab_free(s, x);
1774 debug_check_no_locks_freed(x, s->object_size);
1776 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1777 debug_check_no_obj_freed(x, s->object_size);
1779 /* Use KCSAN to help debug racy use-after-free. */
1780 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1781 __kcsan_check_access(x, s->object_size,
1782 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1785 * As memory initialization might be integrated into KASAN,
1786 * kasan_slab_free and initialization memset's must be
1787 * kept together to avoid discrepancies in behavior.
1789 * The initialization memset's clear the object and the metadata,
1790 * but don't touch the SLAB redzone.
1795 if (!kasan_has_integrated_init())
1796 memset(kasan_reset_tag(x), 0, s->object_size);
1797 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1798 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1799 s->size - s->inuse - rsize);
1801 /* KASAN might put x into memory quarantine, delaying its reuse. */
1802 return kasan_slab_free(s, x, init);
1805 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1806 void **head, void **tail,
1812 void *old_tail = *tail ? *tail : *head;
1814 if (is_kfence_address(next)) {
1815 slab_free_hook(s, next, false);
1819 /* Head and tail of the reconstructed freelist */
1825 next = get_freepointer(s, object);
1827 /* If object's reuse doesn't have to be delayed */
1828 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1829 /* Move object to the new freelist */
1830 set_freepointer(s, object, *head);
1836 * Adjust the reconstructed freelist depth
1837 * accordingly if object's reuse is delayed.
1841 } while (object != old_tail);
1846 return *head != NULL;
1849 static void *setup_object(struct kmem_cache *s, void *object)
1851 setup_object_debug(s, object);
1852 object = kasan_init_slab_obj(s, object);
1853 if (unlikely(s->ctor)) {
1854 kasan_unpoison_object_data(s, object);
1856 kasan_poison_object_data(s, object);
1862 * Slab allocation and freeing
1864 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1865 struct kmem_cache_order_objects oo)
1867 struct folio *folio;
1869 unsigned int order = oo_order(oo);
1871 if (node == NUMA_NO_NODE)
1872 folio = (struct folio *)alloc_pages(flags, order);
1874 folio = (struct folio *)__alloc_pages_node(node, flags, order);
1879 slab = folio_slab(folio);
1880 __folio_set_slab(folio);
1881 /* Make the flag visible before any changes to folio->mapping */
1883 if (folio_is_pfmemalloc(folio))
1884 slab_set_pfmemalloc(slab);
1889 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1890 /* Pre-initialize the random sequence cache */
1891 static int init_cache_random_seq(struct kmem_cache *s)
1893 unsigned int count = oo_objects(s->oo);
1896 /* Bailout if already initialised */
1900 err = cache_random_seq_create(s, count, GFP_KERNEL);
1902 pr_err("SLUB: Unable to initialize free list for %s\n",
1907 /* Transform to an offset on the set of pages */
1908 if (s->random_seq) {
1911 for (i = 0; i < count; i++)
1912 s->random_seq[i] *= s->size;
1917 /* Initialize each random sequence freelist per cache */
1918 static void __init init_freelist_randomization(void)
1920 struct kmem_cache *s;
1922 mutex_lock(&slab_mutex);
1924 list_for_each_entry(s, &slab_caches, list)
1925 init_cache_random_seq(s);
1927 mutex_unlock(&slab_mutex);
1930 /* Get the next entry on the pre-computed freelist randomized */
1931 static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1932 unsigned long *pos, void *start,
1933 unsigned long page_limit,
1934 unsigned long freelist_count)
1939 * If the target page allocation failed, the number of objects on the
1940 * page might be smaller than the usual size defined by the cache.
1943 idx = s->random_seq[*pos];
1945 if (*pos >= freelist_count)
1947 } while (unlikely(idx >= page_limit));
1949 return (char *)start + idx;
1952 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1953 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1958 unsigned long idx, pos, page_limit, freelist_count;
1960 if (slab->objects < 2 || !s->random_seq)
1963 freelist_count = oo_objects(s->oo);
1964 pos = get_random_u32_below(freelist_count);
1966 page_limit = slab->objects * s->size;
1967 start = fixup_red_left(s, slab_address(slab));
1969 /* First entry is used as the base of the freelist */
1970 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1972 cur = setup_object(s, cur);
1973 slab->freelist = cur;
1975 for (idx = 1; idx < slab->objects; idx++) {
1976 next = next_freelist_entry(s, slab, &pos, start, page_limit,
1978 next = setup_object(s, next);
1979 set_freepointer(s, cur, next);
1982 set_freepointer(s, cur, NULL);
1987 static inline int init_cache_random_seq(struct kmem_cache *s)
1991 static inline void init_freelist_randomization(void) { }
1992 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1996 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1998 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2001 struct kmem_cache_order_objects oo = s->oo;
2003 void *start, *p, *next;
2007 flags &= gfp_allowed_mask;
2009 flags |= s->allocflags;
2012 * Let the initial higher-order allocation fail under memory pressure
2013 * so we fall-back to the minimum order allocation.
2015 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2016 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2017 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2019 slab = alloc_slab_page(alloc_gfp, node, oo);
2020 if (unlikely(!slab)) {
2024 * Allocation may have failed due to fragmentation.
2025 * Try a lower order alloc if possible
2027 slab = alloc_slab_page(alloc_gfp, node, oo);
2028 if (unlikely(!slab))
2030 stat(s, ORDER_FALLBACK);
2033 slab->objects = oo_objects(oo);
2037 account_slab(slab, oo_order(oo), s, flags);
2039 slab->slab_cache = s;
2041 kasan_poison_slab(slab);
2043 start = slab_address(slab);
2045 setup_slab_debug(s, slab, start);
2047 shuffle = shuffle_freelist(s, slab);
2050 start = fixup_red_left(s, start);
2051 start = setup_object(s, start);
2052 slab->freelist = start;
2053 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2055 next = setup_object(s, next);
2056 set_freepointer(s, p, next);
2059 set_freepointer(s, p, NULL);
2065 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2067 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2068 flags = kmalloc_fix_flags(flags);
2070 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2072 return allocate_slab(s,
2073 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2076 static void __free_slab(struct kmem_cache *s, struct slab *slab)
2078 struct folio *folio = slab_folio(slab);
2079 int order = folio_order(folio);
2080 int pages = 1 << order;
2082 __slab_clear_pfmemalloc(slab);
2083 folio->mapping = NULL;
2084 /* Make the mapping reset visible before clearing the flag */
2086 __folio_clear_slab(folio);
2087 mm_account_reclaimed_pages(pages);
2088 unaccount_slab(slab, order, s);
2089 __free_pages(&folio->page, order);
2092 static void rcu_free_slab(struct rcu_head *h)
2094 struct slab *slab = container_of(h, struct slab, rcu_head);
2096 __free_slab(slab->slab_cache, slab);
2099 static void free_slab(struct kmem_cache *s, struct slab *slab)
2101 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2104 slab_pad_check(s, slab);
2105 for_each_object(p, s, slab_address(slab), slab->objects)
2106 check_object(s, slab, p, SLUB_RED_INACTIVE);
2109 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2110 call_rcu(&slab->rcu_head, rcu_free_slab);
2112 __free_slab(s, slab);
2115 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2117 dec_slabs_node(s, slab_nid(slab), slab->objects);
2122 * Management of partially allocated slabs.
2125 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2128 if (tail == DEACTIVATE_TO_TAIL)
2129 list_add_tail(&slab->slab_list, &n->partial);
2131 list_add(&slab->slab_list, &n->partial);
2134 static inline void add_partial(struct kmem_cache_node *n,
2135 struct slab *slab, int tail)
2137 lockdep_assert_held(&n->list_lock);
2138 __add_partial(n, slab, tail);
2141 static inline void remove_partial(struct kmem_cache_node *n,
2144 lockdep_assert_held(&n->list_lock);
2145 list_del(&slab->slab_list);
2150 * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a
2151 * slab from the n->partial list. Remove only a single object from the slab, do
2152 * the alloc_debug_processing() checks and leave the slab on the list, or move
2153 * it to full list if it was the last free object.
2155 static void *alloc_single_from_partial(struct kmem_cache *s,
2156 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2160 lockdep_assert_held(&n->list_lock);
2162 object = slab->freelist;
2163 slab->freelist = get_freepointer(s, object);
2166 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2167 remove_partial(n, slab);
2171 if (slab->inuse == slab->objects) {
2172 remove_partial(n, slab);
2173 add_full(s, n, slab);
2180 * Called only for kmem_cache_debug() caches to allocate from a freshly
2181 * allocated slab. Allocate a single object instead of whole freelist
2182 * and put the slab to the partial (or full) list.
2184 static void *alloc_single_from_new_slab(struct kmem_cache *s,
2185 struct slab *slab, int orig_size)
2187 int nid = slab_nid(slab);
2188 struct kmem_cache_node *n = get_node(s, nid);
2189 unsigned long flags;
2193 object = slab->freelist;
2194 slab->freelist = get_freepointer(s, object);
2197 if (!alloc_debug_processing(s, slab, object, orig_size))
2199 * It's not really expected that this would fail on a
2200 * freshly allocated slab, but a concurrent memory
2201 * corruption in theory could cause that.
2205 spin_lock_irqsave(&n->list_lock, flags);
2207 if (slab->inuse == slab->objects)
2208 add_full(s, n, slab);
2210 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2212 inc_slabs_node(s, nid, slab->objects);
2213 spin_unlock_irqrestore(&n->list_lock, flags);
2219 * Remove slab from the partial list, freeze it and
2220 * return the pointer to the freelist.
2222 * Returns a list of objects or NULL if it fails.
2224 static inline void *acquire_slab(struct kmem_cache *s,
2225 struct kmem_cache_node *n, struct slab *slab,
2229 unsigned long counters;
2232 lockdep_assert_held(&n->list_lock);
2235 * Zap the freelist and set the frozen bit.
2236 * The old freelist is the list of objects for the
2237 * per cpu allocation list.
2239 freelist = slab->freelist;
2240 counters = slab->counters;
2241 new.counters = counters;
2243 new.inuse = slab->objects;
2244 new.freelist = NULL;
2246 new.freelist = freelist;
2249 VM_BUG_ON(new.frozen);
2252 if (!__slab_update_freelist(s, slab,
2254 new.freelist, new.counters,
2258 remove_partial(n, slab);
2263 #ifdef CONFIG_SLUB_CPU_PARTIAL
2264 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2266 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2269 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2272 * Try to allocate a partial slab from a specific node.
2274 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2275 struct partial_context *pc)
2277 struct slab *slab, *slab2;
2278 void *object = NULL;
2279 unsigned long flags;
2280 unsigned int partial_slabs = 0;
2283 * Racy check. If we mistakenly see no partial slabs then we
2284 * just allocate an empty slab. If we mistakenly try to get a
2285 * partial slab and there is none available then get_partial()
2288 if (!n || !n->nr_partial)
2291 spin_lock_irqsave(&n->list_lock, flags);
2292 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2295 if (!pfmemalloc_match(slab, pc->flags))
2298 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2299 object = alloc_single_from_partial(s, n, slab,
2306 t = acquire_slab(s, n, slab, object == NULL);
2312 stat(s, ALLOC_FROM_PARTIAL);
2315 put_cpu_partial(s, slab, 0);
2316 stat(s, CPU_PARTIAL_NODE);
2319 #ifdef CONFIG_SLUB_CPU_PARTIAL
2320 if (!kmem_cache_has_cpu_partial(s)
2321 || partial_slabs > s->cpu_partial_slabs / 2)
2328 spin_unlock_irqrestore(&n->list_lock, flags);
2333 * Get a slab from somewhere. Search in increasing NUMA distances.
2335 static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc)
2338 struct zonelist *zonelist;
2341 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2343 unsigned int cpuset_mems_cookie;
2346 * The defrag ratio allows a configuration of the tradeoffs between
2347 * inter node defragmentation and node local allocations. A lower
2348 * defrag_ratio increases the tendency to do local allocations
2349 * instead of attempting to obtain partial slabs from other nodes.
2351 * If the defrag_ratio is set to 0 then kmalloc() always
2352 * returns node local objects. If the ratio is higher then kmalloc()
2353 * may return off node objects because partial slabs are obtained
2354 * from other nodes and filled up.
2356 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2357 * (which makes defrag_ratio = 1000) then every (well almost)
2358 * allocation will first attempt to defrag slab caches on other nodes.
2359 * This means scanning over all nodes to look for partial slabs which
2360 * may be expensive if we do it every time we are trying to find a slab
2361 * with available objects.
2363 if (!s->remote_node_defrag_ratio ||
2364 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2368 cpuset_mems_cookie = read_mems_allowed_begin();
2369 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2370 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2371 struct kmem_cache_node *n;
2373 n = get_node(s, zone_to_nid(zone));
2375 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2376 n->nr_partial > s->min_partial) {
2377 object = get_partial_node(s, n, pc);
2380 * Don't check read_mems_allowed_retry()
2381 * here - if mems_allowed was updated in
2382 * parallel, that was a harmless race
2383 * between allocation and the cpuset
2390 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2391 #endif /* CONFIG_NUMA */
2396 * Get a partial slab, lock it and return it.
2398 static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc)
2401 int searchnode = node;
2403 if (node == NUMA_NO_NODE)
2404 searchnode = numa_mem_id();
2406 object = get_partial_node(s, get_node(s, searchnode), pc);
2407 if (object || node != NUMA_NO_NODE)
2410 return get_any_partial(s, pc);
2413 #ifndef CONFIG_SLUB_TINY
2415 #ifdef CONFIG_PREEMPTION
2417 * Calculate the next globally unique transaction for disambiguation
2418 * during cmpxchg. The transactions start with the cpu number and are then
2419 * incremented by CONFIG_NR_CPUS.
2421 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2424 * No preemption supported therefore also no need to check for
2428 #endif /* CONFIG_PREEMPTION */
2430 static inline unsigned long next_tid(unsigned long tid)
2432 return tid + TID_STEP;
2435 #ifdef SLUB_DEBUG_CMPXCHG
2436 static inline unsigned int tid_to_cpu(unsigned long tid)
2438 return tid % TID_STEP;
2441 static inline unsigned long tid_to_event(unsigned long tid)
2443 return tid / TID_STEP;
2447 static inline unsigned int init_tid(int cpu)
2452 static inline void note_cmpxchg_failure(const char *n,
2453 const struct kmem_cache *s, unsigned long tid)
2455 #ifdef SLUB_DEBUG_CMPXCHG
2456 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2458 pr_info("%s %s: cmpxchg redo ", n, s->name);
2460 #ifdef CONFIG_PREEMPTION
2461 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2462 pr_warn("due to cpu change %d -> %d\n",
2463 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2466 if (tid_to_event(tid) != tid_to_event(actual_tid))
2467 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2468 tid_to_event(tid), tid_to_event(actual_tid));
2470 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2471 actual_tid, tid, next_tid(tid));
2473 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2476 static void init_kmem_cache_cpus(struct kmem_cache *s)
2479 struct kmem_cache_cpu *c;
2481 for_each_possible_cpu(cpu) {
2482 c = per_cpu_ptr(s->cpu_slab, cpu);
2483 local_lock_init(&c->lock);
2484 c->tid = init_tid(cpu);
2489 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2490 * unfreezes the slabs and puts it on the proper list.
2491 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2494 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2497 enum slab_modes { M_NONE, M_PARTIAL, M_FREE, M_FULL_NOLIST };
2498 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2500 enum slab_modes mode = M_NONE;
2501 void *nextfree, *freelist_iter, *freelist_tail;
2502 int tail = DEACTIVATE_TO_HEAD;
2503 unsigned long flags = 0;
2507 if (slab->freelist) {
2508 stat(s, DEACTIVATE_REMOTE_FREES);
2509 tail = DEACTIVATE_TO_TAIL;
2513 * Stage one: Count the objects on cpu's freelist as free_delta and
2514 * remember the last object in freelist_tail for later splicing.
2516 freelist_tail = NULL;
2517 freelist_iter = freelist;
2518 while (freelist_iter) {
2519 nextfree = get_freepointer(s, freelist_iter);
2522 * If 'nextfree' is invalid, it is possible that the object at
2523 * 'freelist_iter' is already corrupted. So isolate all objects
2524 * starting at 'freelist_iter' by skipping them.
2526 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2529 freelist_tail = freelist_iter;
2532 freelist_iter = nextfree;
2536 * Stage two: Unfreeze the slab while splicing the per-cpu
2537 * freelist to the head of slab's freelist.
2539 * Ensure that the slab is unfrozen while the list presence
2540 * reflects the actual number of objects during unfreeze.
2542 * We first perform cmpxchg holding lock and insert to list
2543 * when it succeed. If there is mismatch then the slab is not
2544 * unfrozen and number of objects in the slab may have changed.
2545 * Then release lock and retry cmpxchg again.
2549 old.freelist = READ_ONCE(slab->freelist);
2550 old.counters = READ_ONCE(slab->counters);
2551 VM_BUG_ON(!old.frozen);
2553 /* Determine target state of the slab */
2554 new.counters = old.counters;
2555 if (freelist_tail) {
2556 new.inuse -= free_delta;
2557 set_freepointer(s, freelist_tail, old.freelist);
2558 new.freelist = freelist;
2560 new.freelist = old.freelist;
2564 if (!new.inuse && n->nr_partial >= s->min_partial) {
2566 } else if (new.freelist) {
2569 * Taking the spinlock removes the possibility that
2570 * acquire_slab() will see a slab that is frozen
2572 spin_lock_irqsave(&n->list_lock, flags);
2574 mode = M_FULL_NOLIST;
2578 if (!slab_update_freelist(s, slab,
2579 old.freelist, old.counters,
2580 new.freelist, new.counters,
2581 "unfreezing slab")) {
2582 if (mode == M_PARTIAL)
2583 spin_unlock_irqrestore(&n->list_lock, flags);
2588 if (mode == M_PARTIAL) {
2589 add_partial(n, slab, tail);
2590 spin_unlock_irqrestore(&n->list_lock, flags);
2592 } else if (mode == M_FREE) {
2593 stat(s, DEACTIVATE_EMPTY);
2594 discard_slab(s, slab);
2596 } else if (mode == M_FULL_NOLIST) {
2597 stat(s, DEACTIVATE_FULL);
2601 #ifdef CONFIG_SLUB_CPU_PARTIAL
2602 static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2604 struct kmem_cache_node *n = NULL, *n2 = NULL;
2605 struct slab *slab, *slab_to_discard = NULL;
2606 unsigned long flags = 0;
2608 while (partial_slab) {
2612 slab = partial_slab;
2613 partial_slab = slab->next;
2615 n2 = get_node(s, slab_nid(slab));
2618 spin_unlock_irqrestore(&n->list_lock, flags);
2621 spin_lock_irqsave(&n->list_lock, flags);
2626 old.freelist = slab->freelist;
2627 old.counters = slab->counters;
2628 VM_BUG_ON(!old.frozen);
2630 new.counters = old.counters;
2631 new.freelist = old.freelist;
2635 } while (!__slab_update_freelist(s, slab,
2636 old.freelist, old.counters,
2637 new.freelist, new.counters,
2638 "unfreezing slab"));
2640 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2641 slab->next = slab_to_discard;
2642 slab_to_discard = slab;
2644 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2645 stat(s, FREE_ADD_PARTIAL);
2650 spin_unlock_irqrestore(&n->list_lock, flags);
2652 while (slab_to_discard) {
2653 slab = slab_to_discard;
2654 slab_to_discard = slab_to_discard->next;
2656 stat(s, DEACTIVATE_EMPTY);
2657 discard_slab(s, slab);
2663 * Unfreeze all the cpu partial slabs.
2665 static void unfreeze_partials(struct kmem_cache *s)
2667 struct slab *partial_slab;
2668 unsigned long flags;
2670 local_lock_irqsave(&s->cpu_slab->lock, flags);
2671 partial_slab = this_cpu_read(s->cpu_slab->partial);
2672 this_cpu_write(s->cpu_slab->partial, NULL);
2673 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2676 __unfreeze_partials(s, partial_slab);
2679 static void unfreeze_partials_cpu(struct kmem_cache *s,
2680 struct kmem_cache_cpu *c)
2682 struct slab *partial_slab;
2684 partial_slab = slub_percpu_partial(c);
2688 __unfreeze_partials(s, partial_slab);
2692 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2693 * partial slab slot if available.
2695 * If we did not find a slot then simply move all the partials to the
2696 * per node partial list.
2698 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2700 struct slab *oldslab;
2701 struct slab *slab_to_unfreeze = NULL;
2702 unsigned long flags;
2705 local_lock_irqsave(&s->cpu_slab->lock, flags);
2707 oldslab = this_cpu_read(s->cpu_slab->partial);
2710 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2712 * Partial array is full. Move the existing set to the
2713 * per node partial list. Postpone the actual unfreezing
2714 * outside of the critical section.
2716 slab_to_unfreeze = oldslab;
2719 slabs = oldslab->slabs;
2725 slab->slabs = slabs;
2726 slab->next = oldslab;
2728 this_cpu_write(s->cpu_slab->partial, slab);
2730 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2732 if (slab_to_unfreeze) {
2733 __unfreeze_partials(s, slab_to_unfreeze);
2734 stat(s, CPU_PARTIAL_DRAIN);
2738 #else /* CONFIG_SLUB_CPU_PARTIAL */
2740 static inline void unfreeze_partials(struct kmem_cache *s) { }
2741 static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2742 struct kmem_cache_cpu *c) { }
2744 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2746 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2748 unsigned long flags;
2752 local_lock_irqsave(&s->cpu_slab->lock, flags);
2755 freelist = c->freelist;
2759 c->tid = next_tid(c->tid);
2761 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2764 deactivate_slab(s, slab, freelist);
2765 stat(s, CPUSLAB_FLUSH);
2769 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2771 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2772 void *freelist = c->freelist;
2773 struct slab *slab = c->slab;
2777 c->tid = next_tid(c->tid);
2780 deactivate_slab(s, slab, freelist);
2781 stat(s, CPUSLAB_FLUSH);
2784 unfreeze_partials_cpu(s, c);
2787 struct slub_flush_work {
2788 struct work_struct work;
2789 struct kmem_cache *s;
2796 * Called from CPU work handler with migration disabled.
2798 static void flush_cpu_slab(struct work_struct *w)
2800 struct kmem_cache *s;
2801 struct kmem_cache_cpu *c;
2802 struct slub_flush_work *sfw;
2804 sfw = container_of(w, struct slub_flush_work, work);
2807 c = this_cpu_ptr(s->cpu_slab);
2812 unfreeze_partials(s);
2815 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2817 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2819 return c->slab || slub_percpu_partial(c);
2822 static DEFINE_MUTEX(flush_lock);
2823 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2825 static void flush_all_cpus_locked(struct kmem_cache *s)
2827 struct slub_flush_work *sfw;
2830 lockdep_assert_cpus_held();
2831 mutex_lock(&flush_lock);
2833 for_each_online_cpu(cpu) {
2834 sfw = &per_cpu(slub_flush, cpu);
2835 if (!has_cpu_slab(cpu, s)) {
2839 INIT_WORK(&sfw->work, flush_cpu_slab);
2842 queue_work_on(cpu, flushwq, &sfw->work);
2845 for_each_online_cpu(cpu) {
2846 sfw = &per_cpu(slub_flush, cpu);
2849 flush_work(&sfw->work);
2852 mutex_unlock(&flush_lock);
2855 static void flush_all(struct kmem_cache *s)
2858 flush_all_cpus_locked(s);
2863 * Use the cpu notifier to insure that the cpu slabs are flushed when
2866 static int slub_cpu_dead(unsigned int cpu)
2868 struct kmem_cache *s;
2870 mutex_lock(&slab_mutex);
2871 list_for_each_entry(s, &slab_caches, list)
2872 __flush_cpu_slab(s, cpu);
2873 mutex_unlock(&slab_mutex);
2877 #else /* CONFIG_SLUB_TINY */
2878 static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
2879 static inline void flush_all(struct kmem_cache *s) { }
2880 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
2881 static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
2882 #endif /* CONFIG_SLUB_TINY */
2885 * Check if the objects in a per cpu structure fit numa
2886 * locality expectations.
2888 static inline int node_match(struct slab *slab, int node)
2891 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2897 #ifdef CONFIG_SLUB_DEBUG
2898 static int count_free(struct slab *slab)
2900 return slab->objects - slab->inuse;
2903 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2905 return atomic_long_read(&n->total_objects);
2908 /* Supports checking bulk free of a constructed freelist */
2909 static inline bool free_debug_processing(struct kmem_cache *s,
2910 struct slab *slab, void *head, void *tail, int *bulk_cnt,
2911 unsigned long addr, depot_stack_handle_t handle)
2913 bool checks_ok = false;
2914 void *object = head;
2917 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2918 if (!check_slab(s, slab))
2922 if (slab->inuse < *bulk_cnt) {
2923 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
2924 slab->inuse, *bulk_cnt);
2930 if (++cnt > *bulk_cnt)
2933 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2934 if (!free_consistency_checks(s, slab, object, addr))
2938 if (s->flags & SLAB_STORE_USER)
2939 set_track_update(s, object, TRACK_FREE, addr, handle);
2940 trace(s, slab, object, 0);
2941 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
2942 init_object(s, object, SLUB_RED_INACTIVE);
2944 /* Reached end of constructed freelist yet? */
2945 if (object != tail) {
2946 object = get_freepointer(s, object);
2952 if (cnt != *bulk_cnt) {
2953 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
2961 slab_fix(s, "Object at 0x%p not freed", object);
2965 #endif /* CONFIG_SLUB_DEBUG */
2967 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
2968 static unsigned long count_partial(struct kmem_cache_node *n,
2969 int (*get_count)(struct slab *))
2971 unsigned long flags;
2972 unsigned long x = 0;
2975 spin_lock_irqsave(&n->list_lock, flags);
2976 list_for_each_entry(slab, &n->partial, slab_list)
2977 x += get_count(slab);
2978 spin_unlock_irqrestore(&n->list_lock, flags);
2981 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
2983 #ifdef CONFIG_SLUB_DEBUG
2984 static noinline void
2985 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2987 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2988 DEFAULT_RATELIMIT_BURST);
2990 struct kmem_cache_node *n;
2992 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2995 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2996 nid, gfpflags, &gfpflags);
2997 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2998 s->name, s->object_size, s->size, oo_order(s->oo),
3001 if (oo_order(s->min) > get_order(s->object_size))
3002 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
3005 for_each_kmem_cache_node(s, node, n) {
3006 unsigned long nr_slabs;
3007 unsigned long nr_objs;
3008 unsigned long nr_free;
3010 nr_free = count_partial(n, count_free);
3011 nr_slabs = node_nr_slabs(n);
3012 nr_objs = node_nr_objs(n);
3014 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3015 node, nr_slabs, nr_objs, nr_free);
3018 #else /* CONFIG_SLUB_DEBUG */
3020 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3023 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3025 if (unlikely(slab_test_pfmemalloc(slab)))
3026 return gfp_pfmemalloc_allowed(gfpflags);
3031 #ifndef CONFIG_SLUB_TINY
3033 __update_cpu_freelist_fast(struct kmem_cache *s,
3034 void *freelist_old, void *freelist_new,
3037 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3038 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3040 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3041 &old.full, new.full);
3045 * Check the slab->freelist and either transfer the freelist to the
3046 * per cpu freelist or deactivate the slab.
3048 * The slab is still frozen if the return value is not NULL.
3050 * If this function returns NULL then the slab has been unfrozen.
3052 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3055 unsigned long counters;
3058 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3061 freelist = slab->freelist;
3062 counters = slab->counters;
3064 new.counters = counters;
3065 VM_BUG_ON(!new.frozen);
3067 new.inuse = slab->objects;
3068 new.frozen = freelist != NULL;
3070 } while (!__slab_update_freelist(s, slab,
3079 * Slow path. The lockless freelist is empty or we need to perform
3082 * Processing is still very fast if new objects have been freed to the
3083 * regular freelist. In that case we simply take over the regular freelist
3084 * as the lockless freelist and zap the regular freelist.
3086 * If that is not working then we fall back to the partial lists. We take the
3087 * first element of the freelist as the object to allocate now and move the
3088 * rest of the freelist to the lockless freelist.
3090 * And if we were unable to get a new slab from the partial slab lists then
3091 * we need to allocate a new slab. This is the slowest path since it involves
3092 * a call to the page allocator and the setup of a new slab.
3094 * Version of __slab_alloc to use when we know that preemption is
3095 * already disabled (which is the case for bulk allocation).
3097 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3098 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3102 unsigned long flags;
3103 struct partial_context pc;
3105 stat(s, ALLOC_SLOWPATH);
3109 slab = READ_ONCE(c->slab);
3112 * if the node is not online or has no normal memory, just
3113 * ignore the node constraint
3115 if (unlikely(node != NUMA_NO_NODE &&
3116 !node_isset(node, slab_nodes)))
3117 node = NUMA_NO_NODE;
3122 if (unlikely(!node_match(slab, node))) {
3124 * same as above but node_match() being false already
3125 * implies node != NUMA_NO_NODE
3127 if (!node_isset(node, slab_nodes)) {
3128 node = NUMA_NO_NODE;
3130 stat(s, ALLOC_NODE_MISMATCH);
3131 goto deactivate_slab;
3136 * By rights, we should be searching for a slab page that was
3137 * PFMEMALLOC but right now, we are losing the pfmemalloc
3138 * information when the page leaves the per-cpu allocator
3140 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3141 goto deactivate_slab;
3143 /* must check again c->slab in case we got preempted and it changed */
3144 local_lock_irqsave(&s->cpu_slab->lock, flags);
3145 if (unlikely(slab != c->slab)) {
3146 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3149 freelist = c->freelist;
3153 freelist = get_freelist(s, slab);
3157 c->tid = next_tid(c->tid);
3158 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3159 stat(s, DEACTIVATE_BYPASS);
3163 stat(s, ALLOC_REFILL);
3167 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3170 * freelist is pointing to the list of objects to be used.
3171 * slab is pointing to the slab from which the objects are obtained.
3172 * That slab must be frozen for per cpu allocations to work.
3174 VM_BUG_ON(!c->slab->frozen);
3175 c->freelist = get_freepointer(s, freelist);
3176 c->tid = next_tid(c->tid);
3177 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3182 local_lock_irqsave(&s->cpu_slab->lock, flags);
3183 if (slab != c->slab) {
3184 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3187 freelist = c->freelist;
3190 c->tid = next_tid(c->tid);
3191 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3192 deactivate_slab(s, slab, freelist);
3196 if (slub_percpu_partial(c)) {
3197 local_lock_irqsave(&s->cpu_slab->lock, flags);
3198 if (unlikely(c->slab)) {
3199 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3202 if (unlikely(!slub_percpu_partial(c))) {
3203 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3204 /* we were preempted and partial list got empty */
3208 slab = c->slab = slub_percpu_partial(c);
3209 slub_set_percpu_partial(c, slab);
3210 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3211 stat(s, CPU_PARTIAL_ALLOC);
3217 pc.flags = gfpflags;
3219 pc.orig_size = orig_size;
3220 freelist = get_partial(s, node, &pc);
3222 goto check_new_slab;
3224 slub_put_cpu_ptr(s->cpu_slab);
3225 slab = new_slab(s, gfpflags, node);
3226 c = slub_get_cpu_ptr(s->cpu_slab);
3228 if (unlikely(!slab)) {
3229 slab_out_of_memory(s, gfpflags, node);
3233 stat(s, ALLOC_SLAB);
3235 if (kmem_cache_debug(s)) {
3236 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3238 if (unlikely(!freelist))
3241 if (s->flags & SLAB_STORE_USER)
3242 set_track(s, freelist, TRACK_ALLOC, addr);
3248 * No other reference to the slab yet so we can
3249 * muck around with it freely without cmpxchg
3251 freelist = slab->freelist;
3252 slab->freelist = NULL;
3253 slab->inuse = slab->objects;
3256 inc_slabs_node(s, slab_nid(slab), slab->objects);
3260 if (kmem_cache_debug(s)) {
3262 * For debug caches here we had to go through
3263 * alloc_single_from_partial() so just store the tracking info
3264 * and return the object
3266 if (s->flags & SLAB_STORE_USER)
3267 set_track(s, freelist, TRACK_ALLOC, addr);
3272 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3274 * For !pfmemalloc_match() case we don't load freelist so that
3275 * we don't make further mismatched allocations easier.
3277 deactivate_slab(s, slab, get_freepointer(s, freelist));
3283 local_lock_irqsave(&s->cpu_slab->lock, flags);
3284 if (unlikely(c->slab)) {
3285 void *flush_freelist = c->freelist;
3286 struct slab *flush_slab = c->slab;
3290 c->tid = next_tid(c->tid);
3292 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3294 deactivate_slab(s, flush_slab, flush_freelist);
3296 stat(s, CPUSLAB_FLUSH);
3298 goto retry_load_slab;
3306 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3307 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3310 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3311 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3315 #ifdef CONFIG_PREEMPT_COUNT
3317 * We may have been preempted and rescheduled on a different
3318 * cpu before disabling preemption. Need to reload cpu area
3321 c = slub_get_cpu_ptr(s->cpu_slab);
3324 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3325 #ifdef CONFIG_PREEMPT_COUNT
3326 slub_put_cpu_ptr(s->cpu_slab);
3331 static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3332 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3334 struct kmem_cache_cpu *c;
3341 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3342 * enabled. We may switch back and forth between cpus while
3343 * reading from one cpu area. That does not matter as long
3344 * as we end up on the original cpu again when doing the cmpxchg.
3346 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3347 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3348 * the tid. If we are preempted and switched to another cpu between the
3349 * two reads, it's OK as the two are still associated with the same cpu
3350 * and cmpxchg later will validate the cpu.
3352 c = raw_cpu_ptr(s->cpu_slab);
3353 tid = READ_ONCE(c->tid);
3356 * Irqless object alloc/free algorithm used here depends on sequence
3357 * of fetching cpu_slab's data. tid should be fetched before anything
3358 * on c to guarantee that object and slab associated with previous tid
3359 * won't be used with current tid. If we fetch tid first, object and
3360 * slab could be one associated with next tid and our alloc/free
3361 * request will be failed. In this case, we will retry. So, no problem.
3366 * The transaction ids are globally unique per cpu and per operation on
3367 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3368 * occurs on the right processor and that there was no operation on the
3369 * linked list in between.
3372 object = c->freelist;
3375 if (!USE_LOCKLESS_FAST_PATH() ||
3376 unlikely(!object || !slab || !node_match(slab, node))) {
3377 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3379 void *next_object = get_freepointer_safe(s, object);
3382 * The cmpxchg will only match if there was no additional
3383 * operation and if we are on the right processor.
3385 * The cmpxchg does the following atomically (without lock
3387 * 1. Relocate first pointer to the current per cpu area.
3388 * 2. Verify that tid and freelist have not been changed
3389 * 3. If they were not changed replace tid and freelist
3391 * Since this is without lock semantics the protection is only
3392 * against code executing on this cpu *not* from access by
3395 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3396 note_cmpxchg_failure("slab_alloc", s, tid);
3399 prefetch_freepointer(s, next_object);
3400 stat(s, ALLOC_FASTPATH);
3405 #else /* CONFIG_SLUB_TINY */
3406 static void *__slab_alloc_node(struct kmem_cache *s,
3407 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3409 struct partial_context pc;
3413 pc.flags = gfpflags;
3415 pc.orig_size = orig_size;
3416 object = get_partial(s, node, &pc);
3421 slab = new_slab(s, gfpflags, node);
3422 if (unlikely(!slab)) {
3423 slab_out_of_memory(s, gfpflags, node);
3427 object = alloc_single_from_new_slab(s, slab, orig_size);
3431 #endif /* CONFIG_SLUB_TINY */
3434 * If the object has been wiped upon free, make sure it's fully initialized by
3435 * zeroing out freelist pointer.
3437 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3440 if (unlikely(slab_want_init_on_free(s)) && obj)
3441 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3446 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3447 * have the fastpath folded into their functions. So no function call
3448 * overhead for requests that can be satisfied on the fastpath.
3450 * The fastpath works by first checking if the lockless freelist can be used.
3451 * If not then __slab_alloc is called for slow processing.
3453 * Otherwise we can simply pick the next object from the lockless free list.
3455 static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3456 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3459 struct obj_cgroup *objcg = NULL;
3462 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3466 object = kfence_alloc(s, orig_size, gfpflags);
3467 if (unlikely(object))
3470 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3472 maybe_wipe_obj_freeptr(s, object);
3473 init = slab_want_init_on_alloc(gfpflags, s);
3477 * When init equals 'true', like for kzalloc() family, only
3478 * @orig_size bytes might be zeroed instead of s->object_size
3480 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3485 static __fastpath_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3486 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3488 return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3491 static __fastpath_inline
3492 void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3495 void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3497 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3502 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3504 return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3506 EXPORT_SYMBOL(kmem_cache_alloc);
3508 void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3511 return __kmem_cache_alloc_lru(s, lru, gfpflags);
3513 EXPORT_SYMBOL(kmem_cache_alloc_lru);
3515 void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags,
3516 int node, size_t orig_size,
3517 unsigned long caller)
3519 return slab_alloc_node(s, NULL, gfpflags, node,
3523 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3525 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3527 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3531 EXPORT_SYMBOL(kmem_cache_alloc_node);
3533 static noinline void free_to_partial_list(
3534 struct kmem_cache *s, struct slab *slab,
3535 void *head, void *tail, int bulk_cnt,
3538 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3539 struct slab *slab_free = NULL;
3541 unsigned long flags;
3542 depot_stack_handle_t handle = 0;
3544 if (s->flags & SLAB_STORE_USER)
3545 handle = set_track_prepare();
3547 spin_lock_irqsave(&n->list_lock, flags);
3549 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
3550 void *prior = slab->freelist;
3552 /* Perform the actual freeing while we still hold the locks */
3554 set_freepointer(s, tail, prior);
3555 slab->freelist = head;
3558 * If the slab is empty, and node's partial list is full,
3559 * it should be discarded anyway no matter it's on full or
3562 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
3566 /* was on full list */
3567 remove_full(s, n, slab);
3569 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3570 stat(s, FREE_ADD_PARTIAL);
3572 } else if (slab_free) {
3573 remove_partial(n, slab);
3574 stat(s, FREE_REMOVE_PARTIAL);
3580 * Update the counters while still holding n->list_lock to
3581 * prevent spurious validation warnings
3583 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
3586 spin_unlock_irqrestore(&n->list_lock, flags);
3590 free_slab(s, slab_free);
3595 * Slow path handling. This may still be called frequently since objects
3596 * have a longer lifetime than the cpu slabs in most processing loads.
3598 * So we still attempt to reduce cache line usage. Just take the slab
3599 * lock and free the item. If there is no additional partial slab
3600 * handling required then we can return immediately.
3602 static void __slab_free(struct kmem_cache *s, struct slab *slab,
3603 void *head, void *tail, int cnt,
3610 unsigned long counters;
3611 struct kmem_cache_node *n = NULL;
3612 unsigned long flags;
3614 stat(s, FREE_SLOWPATH);
3616 if (kfence_free(head))
3619 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
3620 free_to_partial_list(s, slab, head, tail, cnt, addr);
3626 spin_unlock_irqrestore(&n->list_lock, flags);
3629 prior = slab->freelist;
3630 counters = slab->counters;
3631 set_freepointer(s, tail, prior);
3632 new.counters = counters;
3633 was_frozen = new.frozen;
3635 if ((!new.inuse || !prior) && !was_frozen) {
3637 if (kmem_cache_has_cpu_partial(s) && !prior) {
3640 * Slab was on no list before and will be
3642 * We can defer the list move and instead
3647 } else { /* Needs to be taken off a list */
3649 n = get_node(s, slab_nid(slab));
3651 * Speculatively acquire the list_lock.
3652 * If the cmpxchg does not succeed then we may
3653 * drop the list_lock without any processing.
3655 * Otherwise the list_lock will synchronize with
3656 * other processors updating the list of slabs.
3658 spin_lock_irqsave(&n->list_lock, flags);
3663 } while (!slab_update_freelist(s, slab,
3670 if (likely(was_frozen)) {
3672 * The list lock was not taken therefore no list
3673 * activity can be necessary.
3675 stat(s, FREE_FROZEN);
3676 } else if (new.frozen) {
3678 * If we just froze the slab then put it onto the
3679 * per cpu partial list.
3681 put_cpu_partial(s, slab, 1);
3682 stat(s, CPU_PARTIAL_FREE);
3688 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3692 * Objects left in the slab. If it was not on the partial list before
3695 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3696 remove_full(s, n, slab);
3697 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3698 stat(s, FREE_ADD_PARTIAL);
3700 spin_unlock_irqrestore(&n->list_lock, flags);
3706 * Slab on the partial list.
3708 remove_partial(n, slab);
3709 stat(s, FREE_REMOVE_PARTIAL);
3711 /* Slab must be on the full list */
3712 remove_full(s, n, slab);
3715 spin_unlock_irqrestore(&n->list_lock, flags);
3717 discard_slab(s, slab);
3720 #ifndef CONFIG_SLUB_TINY
3722 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3723 * can perform fastpath freeing without additional function calls.
3725 * The fastpath is only possible if we are freeing to the current cpu slab
3726 * of this processor. This typically the case if we have just allocated
3729 * If fastpath is not possible then fall back to __slab_free where we deal
3730 * with all sorts of special processing.
3732 * Bulk free of a freelist with several objects (all pointing to the
3733 * same slab) possible by specifying head and tail ptr, plus objects
3734 * count (cnt). Bulk free indicated by tail pointer being set.
3736 static __always_inline void do_slab_free(struct kmem_cache *s,
3737 struct slab *slab, void *head, void *tail,
3738 int cnt, unsigned long addr)
3740 void *tail_obj = tail ? : head;
3741 struct kmem_cache_cpu *c;
3747 * Determine the currently cpus per cpu slab.
3748 * The cpu may change afterward. However that does not matter since
3749 * data is retrieved via this pointer. If we are on the same cpu
3750 * during the cmpxchg then the free will succeed.
3752 c = raw_cpu_ptr(s->cpu_slab);
3753 tid = READ_ONCE(c->tid);
3755 /* Same with comment on barrier() in slab_alloc_node() */
3758 if (unlikely(slab != c->slab)) {
3759 __slab_free(s, slab, head, tail_obj, cnt, addr);
3763 if (USE_LOCKLESS_FAST_PATH()) {
3764 freelist = READ_ONCE(c->freelist);
3766 set_freepointer(s, tail_obj, freelist);
3768 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
3769 note_cmpxchg_failure("slab_free", s, tid);
3773 /* Update the free list under the local lock */
3774 local_lock(&s->cpu_slab->lock);
3775 c = this_cpu_ptr(s->cpu_slab);
3776 if (unlikely(slab != c->slab)) {
3777 local_unlock(&s->cpu_slab->lock);
3781 freelist = c->freelist;
3783 set_freepointer(s, tail_obj, freelist);
3785 c->tid = next_tid(tid);
3787 local_unlock(&s->cpu_slab->lock);
3789 stat(s, FREE_FASTPATH);
3791 #else /* CONFIG_SLUB_TINY */
3792 static void do_slab_free(struct kmem_cache *s,
3793 struct slab *slab, void *head, void *tail,
3794 int cnt, unsigned long addr)
3796 void *tail_obj = tail ? : head;
3798 __slab_free(s, slab, head, tail_obj, cnt, addr);
3800 #endif /* CONFIG_SLUB_TINY */
3802 static __fastpath_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3803 void *head, void *tail, void **p, int cnt,
3806 memcg_slab_free_hook(s, slab, p, cnt);
3808 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3809 * to remove objects, whose reuse must be delayed.
3811 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3812 do_slab_free(s, slab, head, tail, cnt, addr);
3815 #ifdef CONFIG_KASAN_GENERIC
3816 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3818 do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3822 void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller)
3824 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, caller);
3827 void kmem_cache_free(struct kmem_cache *s, void *x)
3829 s = cache_from_obj(s, x);
3832 trace_kmem_cache_free(_RET_IP_, x, s);
3833 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_);
3835 EXPORT_SYMBOL(kmem_cache_free);
3837 struct detached_freelist {
3842 struct kmem_cache *s;
3846 * This function progressively scans the array with free objects (with
3847 * a limited look ahead) and extract objects belonging to the same
3848 * slab. It builds a detached freelist directly within the given
3849 * slab/objects. This can happen without any need for
3850 * synchronization, because the objects are owned by running process.
3851 * The freelist is build up as a single linked list in the objects.
3852 * The idea is, that this detached freelist can then be bulk
3853 * transferred to the real freelist(s), but only requiring a single
3854 * synchronization primitive. Look ahead in the array is limited due
3855 * to performance reasons.
3858 int build_detached_freelist(struct kmem_cache *s, size_t size,
3859 void **p, struct detached_freelist *df)
3863 struct folio *folio;
3867 folio = virt_to_folio(object);
3869 /* Handle kalloc'ed objects */
3870 if (unlikely(!folio_test_slab(folio))) {
3871 free_large_kmalloc(folio, object);
3875 /* Derive kmem_cache from object */
3876 df->slab = folio_slab(folio);
3877 df->s = df->slab->slab_cache;
3879 df->slab = folio_slab(folio);
3880 df->s = cache_from_obj(s, object); /* Support for memcg */
3883 /* Start new detached freelist */
3885 df->freelist = object;
3888 if (is_kfence_address(object))
3891 set_freepointer(df->s, object, NULL);
3896 /* df->slab is always set at this point */
3897 if (df->slab == virt_to_slab(object)) {
3898 /* Opportunity build freelist */
3899 set_freepointer(df->s, object, df->freelist);
3900 df->freelist = object;
3904 swap(p[size], p[same]);
3908 /* Limit look ahead search */
3916 /* Note that interrupts must be enabled when calling this function. */
3917 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3923 struct detached_freelist df;
3925 size = build_detached_freelist(s, size, p, &df);
3929 slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt,
3931 } while (likely(size));
3933 EXPORT_SYMBOL(kmem_cache_free_bulk);
3935 #ifndef CONFIG_SLUB_TINY
3936 static inline int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
3937 size_t size, void **p, struct obj_cgroup *objcg)
3939 struct kmem_cache_cpu *c;
3940 unsigned long irqflags;
3944 * Drain objects in the per cpu slab, while disabling local
3945 * IRQs, which protects against PREEMPT and interrupts
3946 * handlers invoking normal fastpath.
3948 c = slub_get_cpu_ptr(s->cpu_slab);
3949 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
3951 for (i = 0; i < size; i++) {
3952 void *object = kfence_alloc(s, s->object_size, flags);
3954 if (unlikely(object)) {
3959 object = c->freelist;
3960 if (unlikely(!object)) {
3962 * We may have removed an object from c->freelist using
3963 * the fastpath in the previous iteration; in that case,
3964 * c->tid has not been bumped yet.
3965 * Since ___slab_alloc() may reenable interrupts while
3966 * allocating memory, we should bump c->tid now.
3968 c->tid = next_tid(c->tid);
3970 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
3973 * Invoking slow path likely have side-effect
3974 * of re-populating per CPU c->freelist
3976 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3977 _RET_IP_, c, s->object_size);
3978 if (unlikely(!p[i]))
3981 c = this_cpu_ptr(s->cpu_slab);
3982 maybe_wipe_obj_freeptr(s, p[i]);
3984 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
3986 continue; /* goto for-loop */
3988 c->freelist = get_freepointer(s, object);
3990 maybe_wipe_obj_freeptr(s, p[i]);
3992 c->tid = next_tid(c->tid);
3993 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
3994 slub_put_cpu_ptr(s->cpu_slab);
3999 slub_put_cpu_ptr(s->cpu_slab);
4000 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
4001 kmem_cache_free_bulk(s, i, p);
4005 #else /* CONFIG_SLUB_TINY */
4006 static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4007 size_t size, void **p, struct obj_cgroup *objcg)
4011 for (i = 0; i < size; i++) {
4012 void *object = kfence_alloc(s, s->object_size, flags);
4014 if (unlikely(object)) {
4019 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4020 _RET_IP_, s->object_size);
4021 if (unlikely(!p[i]))
4024 maybe_wipe_obj_freeptr(s, p[i]);
4030 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
4031 kmem_cache_free_bulk(s, i, p);
4034 #endif /* CONFIG_SLUB_TINY */
4036 /* Note that interrupts must be enabled when calling this function. */
4037 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4041 struct obj_cgroup *objcg = NULL;
4046 /* memcg and kmem_cache debug support */
4047 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4051 i = __kmem_cache_alloc_bulk(s, flags, size, p, objcg);
4054 * memcg and kmem_cache debug support and memory initialization.
4055 * Done outside of the IRQ disabled fastpath loop.
4058 slab_post_alloc_hook(s, objcg, flags, size, p,
4059 slab_want_init_on_alloc(flags, s), s->object_size);
4062 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4066 * Object placement in a slab is made very easy because we always start at
4067 * offset 0. If we tune the size of the object to the alignment then we can
4068 * get the required alignment by putting one properly sized object after
4071 * Notice that the allocation order determines the sizes of the per cpu
4072 * caches. Each processor has always one slab available for allocations.
4073 * Increasing the allocation order reduces the number of times that slabs
4074 * must be moved on and off the partial lists and is therefore a factor in
4079 * Minimum / Maximum order of slab pages. This influences locking overhead
4080 * and slab fragmentation. A higher order reduces the number of partial slabs
4081 * and increases the number of allocations possible without having to
4082 * take the list_lock.
4084 static unsigned int slub_min_order;
4085 static unsigned int slub_max_order =
4086 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4087 static unsigned int slub_min_objects;
4090 * Calculate the order of allocation given an slab object size.
4092 * The order of allocation has significant impact on performance and other
4093 * system components. Generally order 0 allocations should be preferred since
4094 * order 0 does not cause fragmentation in the page allocator. Larger objects
4095 * be problematic to put into order 0 slabs because there may be too much
4096 * unused space left. We go to a higher order if more than 1/16th of the slab
4099 * In order to reach satisfactory performance we must ensure that a minimum
4100 * number of objects is in one slab. Otherwise we may generate too much
4101 * activity on the partial lists which requires taking the list_lock. This is
4102 * less a concern for large slabs though which are rarely used.
4104 * slub_max_order specifies the order where we begin to stop considering the
4105 * number of objects in a slab as critical. If we reach slub_max_order then
4106 * we try to keep the page order as low as possible. So we accept more waste
4107 * of space in favor of a small page order.
4109 * Higher order allocations also allow the placement of more objects in a
4110 * slab and thereby reduce object handling overhead. If the user has
4111 * requested a higher minimum order then we start with that one instead of
4112 * the smallest order which will fit the object.
4114 static inline unsigned int calc_slab_order(unsigned int size,
4115 unsigned int min_objects, unsigned int max_order,
4116 unsigned int fract_leftover)
4118 unsigned int min_order = slub_min_order;
4121 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4122 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4124 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
4125 order <= max_order; order++) {
4127 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4130 rem = slab_size % size;
4132 if (rem <= slab_size / fract_leftover)
4139 static inline int calculate_order(unsigned int size)
4142 unsigned int min_objects;
4143 unsigned int max_objects;
4144 unsigned int nr_cpus;
4147 * Attempt to find best configuration for a slab. This
4148 * works by first attempting to generate a layout with
4149 * the best configuration and backing off gradually.
4151 * First we increase the acceptable waste in a slab. Then
4152 * we reduce the minimum objects required in a slab.
4154 min_objects = slub_min_objects;
4157 * Some architectures will only update present cpus when
4158 * onlining them, so don't trust the number if it's just 1. But
4159 * we also don't want to use nr_cpu_ids always, as on some other
4160 * architectures, there can be many possible cpus, but never
4161 * onlined. Here we compromise between trying to avoid too high
4162 * order on systems that appear larger than they are, and too
4163 * low order on systems that appear smaller than they are.
4165 nr_cpus = num_present_cpus();
4167 nr_cpus = nr_cpu_ids;
4168 min_objects = 4 * (fls(nr_cpus) + 1);
4170 max_objects = order_objects(slub_max_order, size);
4171 min_objects = min(min_objects, max_objects);
4173 while (min_objects > 1) {
4174 unsigned int fraction;
4177 while (fraction >= 4) {
4178 order = calc_slab_order(size, min_objects,
4179 slub_max_order, fraction);
4180 if (order <= slub_max_order)
4188 * We were unable to place multiple objects in a slab. Now
4189 * lets see if we can place a single object there.
4191 order = calc_slab_order(size, 1, slub_max_order, 1);
4192 if (order <= slub_max_order)
4196 * Doh this slab cannot be placed using slub_max_order.
4198 order = calc_slab_order(size, 1, MAX_ORDER, 1);
4199 if (order <= MAX_ORDER)
4205 init_kmem_cache_node(struct kmem_cache_node *n)
4208 spin_lock_init(&n->list_lock);
4209 INIT_LIST_HEAD(&n->partial);
4210 #ifdef CONFIG_SLUB_DEBUG
4211 atomic_long_set(&n->nr_slabs, 0);
4212 atomic_long_set(&n->total_objects, 0);
4213 INIT_LIST_HEAD(&n->full);
4217 #ifndef CONFIG_SLUB_TINY
4218 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4220 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4221 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4222 sizeof(struct kmem_cache_cpu));
4225 * Must align to double word boundary for the double cmpxchg
4226 * instructions to work; see __pcpu_double_call_return_bool().
4228 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4229 2 * sizeof(void *));
4234 init_kmem_cache_cpus(s);
4239 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4243 #endif /* CONFIG_SLUB_TINY */
4245 static struct kmem_cache *kmem_cache_node;
4248 * No kmalloc_node yet so do it by hand. We know that this is the first
4249 * slab on the node for this slabcache. There are no concurrent accesses
4252 * Note that this function only works on the kmem_cache_node
4253 * when allocating for the kmem_cache_node. This is used for bootstrapping
4254 * memory on a fresh node that has no slab structures yet.
4256 static void early_kmem_cache_node_alloc(int node)
4259 struct kmem_cache_node *n;
4261 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4263 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4266 inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4267 if (slab_nid(slab) != node) {
4268 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4269 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4274 #ifdef CONFIG_SLUB_DEBUG
4275 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4276 init_tracking(kmem_cache_node, n);
4278 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4279 slab->freelist = get_freepointer(kmem_cache_node, n);
4281 kmem_cache_node->node[node] = n;
4282 init_kmem_cache_node(n);
4283 inc_slabs_node(kmem_cache_node, node, slab->objects);
4286 * No locks need to be taken here as it has just been
4287 * initialized and there is no concurrent access.
4289 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4292 static void free_kmem_cache_nodes(struct kmem_cache *s)
4295 struct kmem_cache_node *n;
4297 for_each_kmem_cache_node(s, node, n) {
4298 s->node[node] = NULL;
4299 kmem_cache_free(kmem_cache_node, n);
4303 void __kmem_cache_release(struct kmem_cache *s)
4305 cache_random_seq_destroy(s);
4306 #ifndef CONFIG_SLUB_TINY
4307 free_percpu(s->cpu_slab);
4309 free_kmem_cache_nodes(s);
4312 static int init_kmem_cache_nodes(struct kmem_cache *s)
4316 for_each_node_mask(node, slab_nodes) {
4317 struct kmem_cache_node *n;
4319 if (slab_state == DOWN) {
4320 early_kmem_cache_node_alloc(node);
4323 n = kmem_cache_alloc_node(kmem_cache_node,
4327 free_kmem_cache_nodes(s);
4331 init_kmem_cache_node(n);
4337 static void set_cpu_partial(struct kmem_cache *s)
4339 #ifdef CONFIG_SLUB_CPU_PARTIAL
4340 unsigned int nr_objects;
4343 * cpu_partial determined the maximum number of objects kept in the
4344 * per cpu partial lists of a processor.
4346 * Per cpu partial lists mainly contain slabs that just have one
4347 * object freed. If they are used for allocation then they can be
4348 * filled up again with minimal effort. The slab will never hit the
4349 * per node partial lists and therefore no locking will be required.
4351 * For backwards compatibility reasons, this is determined as number
4352 * of objects, even though we now limit maximum number of pages, see
4353 * slub_set_cpu_partial()
4355 if (!kmem_cache_has_cpu_partial(s))
4357 else if (s->size >= PAGE_SIZE)
4359 else if (s->size >= 1024)
4361 else if (s->size >= 256)
4366 slub_set_cpu_partial(s, nr_objects);
4371 * calculate_sizes() determines the order and the distribution of data within
4374 static int calculate_sizes(struct kmem_cache *s)
4376 slab_flags_t flags = s->flags;
4377 unsigned int size = s->object_size;
4381 * Round up object size to the next word boundary. We can only
4382 * place the free pointer at word boundaries and this determines
4383 * the possible location of the free pointer.
4385 size = ALIGN(size, sizeof(void *));
4387 #ifdef CONFIG_SLUB_DEBUG
4389 * Determine if we can poison the object itself. If the user of
4390 * the slab may touch the object after free or before allocation
4391 * then we should never poison the object itself.
4393 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4395 s->flags |= __OBJECT_POISON;
4397 s->flags &= ~__OBJECT_POISON;
4401 * If we are Redzoning then check if there is some space between the
4402 * end of the object and the free pointer. If not then add an
4403 * additional word to have some bytes to store Redzone information.
4405 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4406 size += sizeof(void *);
4410 * With that we have determined the number of bytes in actual use
4411 * by the object and redzoning.
4415 if (slub_debug_orig_size(s) ||
4416 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4417 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4420 * Relocate free pointer after the object if it is not
4421 * permitted to overwrite the first word of the object on
4424 * This is the case if we do RCU, have a constructor or
4425 * destructor, are poisoning the objects, or are
4426 * redzoning an object smaller than sizeof(void *).
4428 * The assumption that s->offset >= s->inuse means free
4429 * pointer is outside of the object is used in the
4430 * freeptr_outside_object() function. If that is no
4431 * longer true, the function needs to be modified.
4434 size += sizeof(void *);
4437 * Store freelist pointer near middle of object to keep
4438 * it away from the edges of the object to avoid small
4439 * sized over/underflows from neighboring allocations.
4441 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4444 #ifdef CONFIG_SLUB_DEBUG
4445 if (flags & SLAB_STORE_USER) {
4447 * Need to store information about allocs and frees after
4450 size += 2 * sizeof(struct track);
4452 /* Save the original kmalloc request size */
4453 if (flags & SLAB_KMALLOC)
4454 size += sizeof(unsigned int);
4458 kasan_cache_create(s, &size, &s->flags);
4459 #ifdef CONFIG_SLUB_DEBUG
4460 if (flags & SLAB_RED_ZONE) {
4462 * Add some empty padding so that we can catch
4463 * overwrites from earlier objects rather than let
4464 * tracking information or the free pointer be
4465 * corrupted if a user writes before the start
4468 size += sizeof(void *);
4470 s->red_left_pad = sizeof(void *);
4471 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4472 size += s->red_left_pad;
4477 * SLUB stores one object immediately after another beginning from
4478 * offset 0. In order to align the objects we have to simply size
4479 * each object to conform to the alignment.
4481 size = ALIGN(size, s->align);
4483 s->reciprocal_size = reciprocal_value(size);
4484 order = calculate_order(size);
4491 s->allocflags |= __GFP_COMP;
4493 if (s->flags & SLAB_CACHE_DMA)
4494 s->allocflags |= GFP_DMA;
4496 if (s->flags & SLAB_CACHE_DMA32)
4497 s->allocflags |= GFP_DMA32;
4499 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4500 s->allocflags |= __GFP_RECLAIMABLE;
4503 * Determine the number of objects per slab
4505 s->oo = oo_make(order, size);
4506 s->min = oo_make(get_order(size), size);
4508 return !!oo_objects(s->oo);
4511 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4513 s->flags = kmem_cache_flags(s->size, flags, s->name);
4514 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4515 s->random = get_random_long();
4518 if (!calculate_sizes(s))
4520 if (disable_higher_order_debug) {
4522 * Disable debugging flags that store metadata if the min slab
4525 if (get_order(s->size) > get_order(s->object_size)) {
4526 s->flags &= ~DEBUG_METADATA_FLAGS;
4528 if (!calculate_sizes(s))
4533 #ifdef system_has_freelist_aba
4534 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
4535 /* Enable fast mode */
4536 s->flags |= __CMPXCHG_DOUBLE;
4541 * The larger the object size is, the more slabs we want on the partial
4542 * list to avoid pounding the page allocator excessively.
4544 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4545 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4550 s->remote_node_defrag_ratio = 1000;
4553 /* Initialize the pre-computed randomized freelist if slab is up */
4554 if (slab_state >= UP) {
4555 if (init_cache_random_seq(s))
4559 if (!init_kmem_cache_nodes(s))
4562 if (alloc_kmem_cache_cpus(s))
4566 __kmem_cache_release(s);
4570 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4573 #ifdef CONFIG_SLUB_DEBUG
4574 void *addr = slab_address(slab);
4577 slab_err(s, slab, text, s->name);
4579 spin_lock(&object_map_lock);
4580 __fill_map(object_map, s, slab);
4582 for_each_object(p, s, addr, slab->objects) {
4584 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
4585 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4586 print_tracking(s, p);
4589 spin_unlock(&object_map_lock);
4594 * Attempt to free all partial slabs on a node.
4595 * This is called from __kmem_cache_shutdown(). We must take list_lock
4596 * because sysfs file might still access partial list after the shutdowning.
4598 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4601 struct slab *slab, *h;
4603 BUG_ON(irqs_disabled());
4604 spin_lock_irq(&n->list_lock);
4605 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4607 remove_partial(n, slab);
4608 list_add(&slab->slab_list, &discard);
4610 list_slab_objects(s, slab,
4611 "Objects remaining in %s on __kmem_cache_shutdown()");
4614 spin_unlock_irq(&n->list_lock);
4616 list_for_each_entry_safe(slab, h, &discard, slab_list)
4617 discard_slab(s, slab);
4620 bool __kmem_cache_empty(struct kmem_cache *s)
4623 struct kmem_cache_node *n;
4625 for_each_kmem_cache_node(s, node, n)
4626 if (n->nr_partial || slabs_node(s, node))
4632 * Release all resources used by a slab cache.
4634 int __kmem_cache_shutdown(struct kmem_cache *s)
4637 struct kmem_cache_node *n;
4639 flush_all_cpus_locked(s);
4640 /* Attempt to free all objects */
4641 for_each_kmem_cache_node(s, node, n) {
4643 if (n->nr_partial || slabs_node(s, node))
4649 #ifdef CONFIG_PRINTK
4650 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4653 int __maybe_unused i;
4657 struct kmem_cache *s = slab->slab_cache;
4658 struct track __maybe_unused *trackp;
4660 kpp->kp_ptr = object;
4661 kpp->kp_slab = slab;
4662 kpp->kp_slab_cache = s;
4663 base = slab_address(slab);
4664 objp0 = kasan_reset_tag(object);
4665 #ifdef CONFIG_SLUB_DEBUG
4666 objp = restore_red_left(s, objp0);
4670 objnr = obj_to_index(s, slab, objp);
4671 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4672 objp = base + s->size * objnr;
4673 kpp->kp_objp = objp;
4674 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4675 || (objp - base) % s->size) ||
4676 !(s->flags & SLAB_STORE_USER))
4678 #ifdef CONFIG_SLUB_DEBUG
4679 objp = fixup_red_left(s, objp);
4680 trackp = get_track(s, objp, TRACK_ALLOC);
4681 kpp->kp_ret = (void *)trackp->addr;
4682 #ifdef CONFIG_STACKDEPOT
4684 depot_stack_handle_t handle;
4685 unsigned long *entries;
4686 unsigned int nr_entries;
4688 handle = READ_ONCE(trackp->handle);
4690 nr_entries = stack_depot_fetch(handle, &entries);
4691 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4692 kpp->kp_stack[i] = (void *)entries[i];
4695 trackp = get_track(s, objp, TRACK_FREE);
4696 handle = READ_ONCE(trackp->handle);
4698 nr_entries = stack_depot_fetch(handle, &entries);
4699 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4700 kpp->kp_free_stack[i] = (void *)entries[i];
4708 /********************************************************************
4710 *******************************************************************/
4712 static int __init setup_slub_min_order(char *str)
4714 get_option(&str, (int *)&slub_min_order);
4719 __setup("slub_min_order=", setup_slub_min_order);
4721 static int __init setup_slub_max_order(char *str)
4723 get_option(&str, (int *)&slub_max_order);
4724 slub_max_order = min_t(unsigned int, slub_max_order, MAX_ORDER);
4729 __setup("slub_max_order=", setup_slub_max_order);
4731 static int __init setup_slub_min_objects(char *str)
4733 get_option(&str, (int *)&slub_min_objects);
4738 __setup("slub_min_objects=", setup_slub_min_objects);
4740 #ifdef CONFIG_HARDENED_USERCOPY
4742 * Rejects incorrectly sized objects and objects that are to be copied
4743 * to/from userspace but do not fall entirely within the containing slab
4744 * cache's usercopy region.
4746 * Returns NULL if check passes, otherwise const char * to name of cache
4747 * to indicate an error.
4749 void __check_heap_object(const void *ptr, unsigned long n,
4750 const struct slab *slab, bool to_user)
4752 struct kmem_cache *s;
4753 unsigned int offset;
4754 bool is_kfence = is_kfence_address(ptr);
4756 ptr = kasan_reset_tag(ptr);
4758 /* Find object and usable object size. */
4759 s = slab->slab_cache;
4761 /* Reject impossible pointers. */
4762 if (ptr < slab_address(slab))
4763 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4766 /* Find offset within object. */
4768 offset = ptr - kfence_object_start(ptr);
4770 offset = (ptr - slab_address(slab)) % s->size;
4772 /* Adjust for redzone and reject if within the redzone. */
4773 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4774 if (offset < s->red_left_pad)
4775 usercopy_abort("SLUB object in left red zone",
4776 s->name, to_user, offset, n);
4777 offset -= s->red_left_pad;
4780 /* Allow address range falling entirely within usercopy region. */
4781 if (offset >= s->useroffset &&
4782 offset - s->useroffset <= s->usersize &&
4783 n <= s->useroffset - offset + s->usersize)
4786 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4788 #endif /* CONFIG_HARDENED_USERCOPY */
4790 #define SHRINK_PROMOTE_MAX 32
4793 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4794 * up most to the head of the partial lists. New allocations will then
4795 * fill those up and thus they can be removed from the partial lists.
4797 * The slabs with the least items are placed last. This results in them
4798 * being allocated from last increasing the chance that the last objects
4799 * are freed in them.
4801 static int __kmem_cache_do_shrink(struct kmem_cache *s)
4805 struct kmem_cache_node *n;
4808 struct list_head discard;
4809 struct list_head promote[SHRINK_PROMOTE_MAX];
4810 unsigned long flags;
4813 for_each_kmem_cache_node(s, node, n) {
4814 INIT_LIST_HEAD(&discard);
4815 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4816 INIT_LIST_HEAD(promote + i);
4818 spin_lock_irqsave(&n->list_lock, flags);
4821 * Build lists of slabs to discard or promote.
4823 * Note that concurrent frees may occur while we hold the
4824 * list_lock. slab->inuse here is the upper limit.
4826 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4827 int free = slab->objects - slab->inuse;
4829 /* Do not reread slab->inuse */
4832 /* We do not keep full slabs on the list */
4835 if (free == slab->objects) {
4836 list_move(&slab->slab_list, &discard);
4838 dec_slabs_node(s, node, slab->objects);
4839 } else if (free <= SHRINK_PROMOTE_MAX)
4840 list_move(&slab->slab_list, promote + free - 1);
4844 * Promote the slabs filled up most to the head of the
4847 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4848 list_splice(promote + i, &n->partial);
4850 spin_unlock_irqrestore(&n->list_lock, flags);
4852 /* Release empty slabs */
4853 list_for_each_entry_safe(slab, t, &discard, slab_list)
4856 if (slabs_node(s, node))
4863 int __kmem_cache_shrink(struct kmem_cache *s)
4866 return __kmem_cache_do_shrink(s);
4869 static int slab_mem_going_offline_callback(void *arg)
4871 struct kmem_cache *s;
4873 mutex_lock(&slab_mutex);
4874 list_for_each_entry(s, &slab_caches, list) {
4875 flush_all_cpus_locked(s);
4876 __kmem_cache_do_shrink(s);
4878 mutex_unlock(&slab_mutex);
4883 static void slab_mem_offline_callback(void *arg)
4885 struct memory_notify *marg = arg;
4888 offline_node = marg->status_change_nid_normal;
4891 * If the node still has available memory. we need kmem_cache_node
4894 if (offline_node < 0)
4897 mutex_lock(&slab_mutex);
4898 node_clear(offline_node, slab_nodes);
4900 * We no longer free kmem_cache_node structures here, as it would be
4901 * racy with all get_node() users, and infeasible to protect them with
4904 mutex_unlock(&slab_mutex);
4907 static int slab_mem_going_online_callback(void *arg)
4909 struct kmem_cache_node *n;
4910 struct kmem_cache *s;
4911 struct memory_notify *marg = arg;
4912 int nid = marg->status_change_nid_normal;
4916 * If the node's memory is already available, then kmem_cache_node is
4917 * already created. Nothing to do.
4923 * We are bringing a node online. No memory is available yet. We must
4924 * allocate a kmem_cache_node structure in order to bring the node
4927 mutex_lock(&slab_mutex);
4928 list_for_each_entry(s, &slab_caches, list) {
4930 * The structure may already exist if the node was previously
4931 * onlined and offlined.
4933 if (get_node(s, nid))
4936 * XXX: kmem_cache_alloc_node will fallback to other nodes
4937 * since memory is not yet available from the node that
4940 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4945 init_kmem_cache_node(n);
4949 * Any cache created after this point will also have kmem_cache_node
4950 * initialized for the new node.
4952 node_set(nid, slab_nodes);
4954 mutex_unlock(&slab_mutex);
4958 static int slab_memory_callback(struct notifier_block *self,
4959 unsigned long action, void *arg)
4964 case MEM_GOING_ONLINE:
4965 ret = slab_mem_going_online_callback(arg);
4967 case MEM_GOING_OFFLINE:
4968 ret = slab_mem_going_offline_callback(arg);
4971 case MEM_CANCEL_ONLINE:
4972 slab_mem_offline_callback(arg);
4975 case MEM_CANCEL_OFFLINE:
4979 ret = notifier_from_errno(ret);
4985 /********************************************************************
4986 * Basic setup of slabs
4987 *******************************************************************/
4990 * Used for early kmem_cache structures that were allocated using
4991 * the page allocator. Allocate them properly then fix up the pointers
4992 * that may be pointing to the wrong kmem_cache structure.
4995 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4998 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4999 struct kmem_cache_node *n;
5001 memcpy(s, static_cache, kmem_cache->object_size);
5004 * This runs very early, and only the boot processor is supposed to be
5005 * up. Even if it weren't true, IRQs are not up so we couldn't fire
5008 __flush_cpu_slab(s, smp_processor_id());
5009 for_each_kmem_cache_node(s, node, n) {
5012 list_for_each_entry(p, &n->partial, slab_list)
5015 #ifdef CONFIG_SLUB_DEBUG
5016 list_for_each_entry(p, &n->full, slab_list)
5020 list_add(&s->list, &slab_caches);
5024 void __init kmem_cache_init(void)
5026 static __initdata struct kmem_cache boot_kmem_cache,
5027 boot_kmem_cache_node;
5030 if (debug_guardpage_minorder())
5033 /* Print slub debugging pointers without hashing */
5034 if (__slub_debug_enabled())
5035 no_hash_pointers_enable(NULL);
5037 kmem_cache_node = &boot_kmem_cache_node;
5038 kmem_cache = &boot_kmem_cache;
5041 * Initialize the nodemask for which we will allocate per node
5042 * structures. Here we don't need taking slab_mutex yet.
5044 for_each_node_state(node, N_NORMAL_MEMORY)
5045 node_set(node, slab_nodes);
5047 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5048 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5050 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5052 /* Able to allocate the per node structures */
5053 slab_state = PARTIAL;
5055 create_boot_cache(kmem_cache, "kmem_cache",
5056 offsetof(struct kmem_cache, node) +
5057 nr_node_ids * sizeof(struct kmem_cache_node *),
5058 SLAB_HWCACHE_ALIGN, 0, 0);
5060 kmem_cache = bootstrap(&boot_kmem_cache);
5061 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5063 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5064 setup_kmalloc_cache_index_table();
5065 create_kmalloc_caches(0);
5067 /* Setup random freelists for each cache */
5068 init_freelist_randomization();
5070 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5073 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5075 slub_min_order, slub_max_order, slub_min_objects,
5076 nr_cpu_ids, nr_node_ids);
5079 void __init kmem_cache_init_late(void)
5081 #ifndef CONFIG_SLUB_TINY
5082 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5088 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5089 slab_flags_t flags, void (*ctor)(void *))
5091 struct kmem_cache *s;
5093 s = find_mergeable(size, align, flags, name, ctor);
5095 if (sysfs_slab_alias(s, name))
5101 * Adjust the object sizes so that we clear
5102 * the complete object on kzalloc.
5104 s->object_size = max(s->object_size, size);
5105 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5111 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5115 err = kmem_cache_open(s, flags);
5119 /* Mutex is not taken during early boot */
5120 if (slab_state <= UP)
5123 err = sysfs_slab_add(s);
5125 __kmem_cache_release(s);
5129 if (s->flags & SLAB_STORE_USER)
5130 debugfs_slab_add(s);
5135 #ifdef SLAB_SUPPORTS_SYSFS
5136 static int count_inuse(struct slab *slab)
5141 static int count_total(struct slab *slab)
5143 return slab->objects;
5147 #ifdef CONFIG_SLUB_DEBUG
5148 static void validate_slab(struct kmem_cache *s, struct slab *slab,
5149 unsigned long *obj_map)
5152 void *addr = slab_address(slab);
5154 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5157 /* Now we know that a valid freelist exists */
5158 __fill_map(obj_map, s, slab);
5159 for_each_object(p, s, addr, slab->objects) {
5160 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5161 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5163 if (!check_object(s, slab, p, val))
5168 static int validate_slab_node(struct kmem_cache *s,
5169 struct kmem_cache_node *n, unsigned long *obj_map)
5171 unsigned long count = 0;
5173 unsigned long flags;
5175 spin_lock_irqsave(&n->list_lock, flags);
5177 list_for_each_entry(slab, &n->partial, slab_list) {
5178 validate_slab(s, slab, obj_map);
5181 if (count != n->nr_partial) {
5182 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5183 s->name, count, n->nr_partial);
5184 slab_add_kunit_errors();
5187 if (!(s->flags & SLAB_STORE_USER))
5190 list_for_each_entry(slab, &n->full, slab_list) {
5191 validate_slab(s, slab, obj_map);
5194 if (count != atomic_long_read(&n->nr_slabs)) {
5195 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5196 s->name, count, atomic_long_read(&n->nr_slabs));
5197 slab_add_kunit_errors();
5201 spin_unlock_irqrestore(&n->list_lock, flags);
5205 long validate_slab_cache(struct kmem_cache *s)
5208 unsigned long count = 0;
5209 struct kmem_cache_node *n;
5210 unsigned long *obj_map;
5212 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5217 for_each_kmem_cache_node(s, node, n)
5218 count += validate_slab_node(s, n, obj_map);
5220 bitmap_free(obj_map);
5224 EXPORT_SYMBOL(validate_slab_cache);
5226 #ifdef CONFIG_DEBUG_FS
5228 * Generate lists of code addresses where slabcache objects are allocated
5233 depot_stack_handle_t handle;
5234 unsigned long count;
5236 unsigned long waste;
5242 DECLARE_BITMAP(cpus, NR_CPUS);
5248 unsigned long count;
5249 struct location *loc;
5253 static struct dentry *slab_debugfs_root;
5255 static void free_loc_track(struct loc_track *t)
5258 free_pages((unsigned long)t->loc,
5259 get_order(sizeof(struct location) * t->max));
5262 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5267 order = get_order(sizeof(struct location) * max);
5269 l = (void *)__get_free_pages(flags, order);
5274 memcpy(l, t->loc, sizeof(struct location) * t->count);
5282 static int add_location(struct loc_track *t, struct kmem_cache *s,
5283 const struct track *track,
5284 unsigned int orig_size)
5286 long start, end, pos;
5288 unsigned long caddr, chandle, cwaste;
5289 unsigned long age = jiffies - track->when;
5290 depot_stack_handle_t handle = 0;
5291 unsigned int waste = s->object_size - orig_size;
5293 #ifdef CONFIG_STACKDEPOT
5294 handle = READ_ONCE(track->handle);
5300 pos = start + (end - start + 1) / 2;
5303 * There is nothing at "end". If we end up there
5304 * we need to add something to before end.
5311 chandle = l->handle;
5313 if ((track->addr == caddr) && (handle == chandle) &&
5314 (waste == cwaste)) {
5319 if (age < l->min_time)
5321 if (age > l->max_time)
5324 if (track->pid < l->min_pid)
5325 l->min_pid = track->pid;
5326 if (track->pid > l->max_pid)
5327 l->max_pid = track->pid;
5329 cpumask_set_cpu(track->cpu,
5330 to_cpumask(l->cpus));
5332 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5336 if (track->addr < caddr)
5338 else if (track->addr == caddr && handle < chandle)
5340 else if (track->addr == caddr && handle == chandle &&
5348 * Not found. Insert new tracking element.
5350 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5356 (t->count - pos) * sizeof(struct location));
5359 l->addr = track->addr;
5363 l->min_pid = track->pid;
5364 l->max_pid = track->pid;
5367 cpumask_clear(to_cpumask(l->cpus));
5368 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5369 nodes_clear(l->nodes);
5370 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5374 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5375 struct slab *slab, enum track_item alloc,
5376 unsigned long *obj_map)
5378 void *addr = slab_address(slab);
5379 bool is_alloc = (alloc == TRACK_ALLOC);
5382 __fill_map(obj_map, s, slab);
5384 for_each_object(p, s, addr, slab->objects)
5385 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5386 add_location(t, s, get_track(s, p, alloc),
5387 is_alloc ? get_orig_size(s, p) :
5390 #endif /* CONFIG_DEBUG_FS */
5391 #endif /* CONFIG_SLUB_DEBUG */
5393 #ifdef SLAB_SUPPORTS_SYSFS
5394 enum slab_stat_type {
5395 SL_ALL, /* All slabs */
5396 SL_PARTIAL, /* Only partially allocated slabs */
5397 SL_CPU, /* Only slabs used for cpu caches */
5398 SL_OBJECTS, /* Determine allocated objects not slabs */
5399 SL_TOTAL /* Determine object capacity not slabs */
5402 #define SO_ALL (1 << SL_ALL)
5403 #define SO_PARTIAL (1 << SL_PARTIAL)
5404 #define SO_CPU (1 << SL_CPU)
5405 #define SO_OBJECTS (1 << SL_OBJECTS)
5406 #define SO_TOTAL (1 << SL_TOTAL)
5408 static ssize_t show_slab_objects(struct kmem_cache *s,
5409 char *buf, unsigned long flags)
5411 unsigned long total = 0;
5414 unsigned long *nodes;
5417 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5421 if (flags & SO_CPU) {
5424 for_each_possible_cpu(cpu) {
5425 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5430 slab = READ_ONCE(c->slab);
5434 node = slab_nid(slab);
5435 if (flags & SO_TOTAL)
5437 else if (flags & SO_OBJECTS)
5445 #ifdef CONFIG_SLUB_CPU_PARTIAL
5446 slab = slub_percpu_partial_read_once(c);
5448 node = slab_nid(slab);
5449 if (flags & SO_TOTAL)
5451 else if (flags & SO_OBJECTS)
5463 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5464 * already held which will conflict with an existing lock order:
5466 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5468 * We don't really need mem_hotplug_lock (to hold off
5469 * slab_mem_going_offline_callback) here because slab's memory hot
5470 * unplug code doesn't destroy the kmem_cache->node[] data.
5473 #ifdef CONFIG_SLUB_DEBUG
5474 if (flags & SO_ALL) {
5475 struct kmem_cache_node *n;
5477 for_each_kmem_cache_node(s, node, n) {
5479 if (flags & SO_TOTAL)
5480 x = atomic_long_read(&n->total_objects);
5481 else if (flags & SO_OBJECTS)
5482 x = atomic_long_read(&n->total_objects) -
5483 count_partial(n, count_free);
5485 x = atomic_long_read(&n->nr_slabs);
5492 if (flags & SO_PARTIAL) {
5493 struct kmem_cache_node *n;
5495 for_each_kmem_cache_node(s, node, n) {
5496 if (flags & SO_TOTAL)
5497 x = count_partial(n, count_total);
5498 else if (flags & SO_OBJECTS)
5499 x = count_partial(n, count_inuse);
5507 len += sysfs_emit_at(buf, len, "%lu", total);
5509 for (node = 0; node < nr_node_ids; node++) {
5511 len += sysfs_emit_at(buf, len, " N%d=%lu",
5515 len += sysfs_emit_at(buf, len, "\n");
5521 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5522 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5524 struct slab_attribute {
5525 struct attribute attr;
5526 ssize_t (*show)(struct kmem_cache *s, char *buf);
5527 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5530 #define SLAB_ATTR_RO(_name) \
5531 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5533 #define SLAB_ATTR(_name) \
5534 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5536 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5538 return sysfs_emit(buf, "%u\n", s->size);
5540 SLAB_ATTR_RO(slab_size);
5542 static ssize_t align_show(struct kmem_cache *s, char *buf)
5544 return sysfs_emit(buf, "%u\n", s->align);
5546 SLAB_ATTR_RO(align);
5548 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5550 return sysfs_emit(buf, "%u\n", s->object_size);
5552 SLAB_ATTR_RO(object_size);
5554 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5556 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5558 SLAB_ATTR_RO(objs_per_slab);
5560 static ssize_t order_show(struct kmem_cache *s, char *buf)
5562 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5564 SLAB_ATTR_RO(order);
5566 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5568 return sysfs_emit(buf, "%lu\n", s->min_partial);
5571 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5577 err = kstrtoul(buf, 10, &min);
5581 s->min_partial = min;
5584 SLAB_ATTR(min_partial);
5586 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5588 unsigned int nr_partial = 0;
5589 #ifdef CONFIG_SLUB_CPU_PARTIAL
5590 nr_partial = s->cpu_partial;
5593 return sysfs_emit(buf, "%u\n", nr_partial);
5596 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5599 unsigned int objects;
5602 err = kstrtouint(buf, 10, &objects);
5605 if (objects && !kmem_cache_has_cpu_partial(s))
5608 slub_set_cpu_partial(s, objects);
5612 SLAB_ATTR(cpu_partial);
5614 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5618 return sysfs_emit(buf, "%pS\n", s->ctor);
5622 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5624 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5626 SLAB_ATTR_RO(aliases);
5628 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5630 return show_slab_objects(s, buf, SO_PARTIAL);
5632 SLAB_ATTR_RO(partial);
5634 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5636 return show_slab_objects(s, buf, SO_CPU);
5638 SLAB_ATTR_RO(cpu_slabs);
5640 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5642 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5644 SLAB_ATTR_RO(objects);
5646 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5648 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5650 SLAB_ATTR_RO(objects_partial);
5652 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5656 int cpu __maybe_unused;
5659 #ifdef CONFIG_SLUB_CPU_PARTIAL
5660 for_each_online_cpu(cpu) {
5663 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5666 slabs += slab->slabs;
5670 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5671 objects = (slabs * oo_objects(s->oo)) / 2;
5672 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5674 #if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5675 for_each_online_cpu(cpu) {
5678 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5680 slabs = READ_ONCE(slab->slabs);
5681 objects = (slabs * oo_objects(s->oo)) / 2;
5682 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5683 cpu, objects, slabs);
5687 len += sysfs_emit_at(buf, len, "\n");
5691 SLAB_ATTR_RO(slabs_cpu_partial);
5693 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5695 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5697 SLAB_ATTR_RO(reclaim_account);
5699 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5701 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5703 SLAB_ATTR_RO(hwcache_align);
5705 #ifdef CONFIG_ZONE_DMA
5706 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5708 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5710 SLAB_ATTR_RO(cache_dma);
5713 #ifdef CONFIG_HARDENED_USERCOPY
5714 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5716 return sysfs_emit(buf, "%u\n", s->usersize);
5718 SLAB_ATTR_RO(usersize);
5721 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5723 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5725 SLAB_ATTR_RO(destroy_by_rcu);
5727 #ifdef CONFIG_SLUB_DEBUG
5728 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5730 return show_slab_objects(s, buf, SO_ALL);
5732 SLAB_ATTR_RO(slabs);
5734 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5736 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5738 SLAB_ATTR_RO(total_objects);
5740 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5742 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5744 SLAB_ATTR_RO(sanity_checks);
5746 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5748 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5750 SLAB_ATTR_RO(trace);
5752 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5754 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5757 SLAB_ATTR_RO(red_zone);
5759 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5761 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5764 SLAB_ATTR_RO(poison);
5766 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5768 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5771 SLAB_ATTR_RO(store_user);
5773 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5778 static ssize_t validate_store(struct kmem_cache *s,
5779 const char *buf, size_t length)
5783 if (buf[0] == '1' && kmem_cache_debug(s)) {
5784 ret = validate_slab_cache(s);
5790 SLAB_ATTR(validate);
5792 #endif /* CONFIG_SLUB_DEBUG */
5794 #ifdef CONFIG_FAILSLAB
5795 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5797 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5800 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5803 if (s->refcount > 1)
5807 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
5809 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
5813 SLAB_ATTR(failslab);
5816 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5821 static ssize_t shrink_store(struct kmem_cache *s,
5822 const char *buf, size_t length)
5825 kmem_cache_shrink(s);
5833 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5835 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5838 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5839 const char *buf, size_t length)
5844 err = kstrtouint(buf, 10, &ratio);
5850 s->remote_node_defrag_ratio = ratio * 10;
5854 SLAB_ATTR(remote_node_defrag_ratio);
5857 #ifdef CONFIG_SLUB_STATS
5858 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5860 unsigned long sum = 0;
5863 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5868 for_each_online_cpu(cpu) {
5869 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5875 len += sysfs_emit_at(buf, len, "%lu", sum);
5878 for_each_online_cpu(cpu) {
5880 len += sysfs_emit_at(buf, len, " C%d=%u",
5885 len += sysfs_emit_at(buf, len, "\n");
5890 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5894 for_each_online_cpu(cpu)
5895 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5898 #define STAT_ATTR(si, text) \
5899 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5901 return show_stat(s, buf, si); \
5903 static ssize_t text##_store(struct kmem_cache *s, \
5904 const char *buf, size_t length) \
5906 if (buf[0] != '0') \
5908 clear_stat(s, si); \
5913 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5914 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5915 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5916 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5917 STAT_ATTR(FREE_FROZEN, free_frozen);
5918 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5919 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5920 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5921 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5922 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5923 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5924 STAT_ATTR(FREE_SLAB, free_slab);
5925 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5926 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5927 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5928 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5929 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5930 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5931 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5932 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5933 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5934 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5935 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5936 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5937 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5938 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5939 #endif /* CONFIG_SLUB_STATS */
5941 #ifdef CONFIG_KFENCE
5942 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
5944 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
5947 static ssize_t skip_kfence_store(struct kmem_cache *s,
5948 const char *buf, size_t length)
5953 s->flags &= ~SLAB_SKIP_KFENCE;
5954 else if (buf[0] == '1')
5955 s->flags |= SLAB_SKIP_KFENCE;
5961 SLAB_ATTR(skip_kfence);
5964 static struct attribute *slab_attrs[] = {
5965 &slab_size_attr.attr,
5966 &object_size_attr.attr,
5967 &objs_per_slab_attr.attr,
5969 &min_partial_attr.attr,
5970 &cpu_partial_attr.attr,
5972 &objects_partial_attr.attr,
5974 &cpu_slabs_attr.attr,
5978 &hwcache_align_attr.attr,
5979 &reclaim_account_attr.attr,
5980 &destroy_by_rcu_attr.attr,
5982 &slabs_cpu_partial_attr.attr,
5983 #ifdef CONFIG_SLUB_DEBUG
5984 &total_objects_attr.attr,
5986 &sanity_checks_attr.attr,
5988 &red_zone_attr.attr,
5990 &store_user_attr.attr,
5991 &validate_attr.attr,
5993 #ifdef CONFIG_ZONE_DMA
5994 &cache_dma_attr.attr,
5997 &remote_node_defrag_ratio_attr.attr,
5999 #ifdef CONFIG_SLUB_STATS
6000 &alloc_fastpath_attr.attr,
6001 &alloc_slowpath_attr.attr,
6002 &free_fastpath_attr.attr,
6003 &free_slowpath_attr.attr,
6004 &free_frozen_attr.attr,
6005 &free_add_partial_attr.attr,
6006 &free_remove_partial_attr.attr,
6007 &alloc_from_partial_attr.attr,
6008 &alloc_slab_attr.attr,
6009 &alloc_refill_attr.attr,
6010 &alloc_node_mismatch_attr.attr,
6011 &free_slab_attr.attr,
6012 &cpuslab_flush_attr.attr,
6013 &deactivate_full_attr.attr,
6014 &deactivate_empty_attr.attr,
6015 &deactivate_to_head_attr.attr,
6016 &deactivate_to_tail_attr.attr,
6017 &deactivate_remote_frees_attr.attr,
6018 &deactivate_bypass_attr.attr,
6019 &order_fallback_attr.attr,
6020 &cmpxchg_double_fail_attr.attr,
6021 &cmpxchg_double_cpu_fail_attr.attr,
6022 &cpu_partial_alloc_attr.attr,
6023 &cpu_partial_free_attr.attr,
6024 &cpu_partial_node_attr.attr,
6025 &cpu_partial_drain_attr.attr,
6027 #ifdef CONFIG_FAILSLAB
6028 &failslab_attr.attr,
6030 #ifdef CONFIG_HARDENED_USERCOPY
6031 &usersize_attr.attr,
6033 #ifdef CONFIG_KFENCE
6034 &skip_kfence_attr.attr,
6040 static const struct attribute_group slab_attr_group = {
6041 .attrs = slab_attrs,
6044 static ssize_t slab_attr_show(struct kobject *kobj,
6045 struct attribute *attr,
6048 struct slab_attribute *attribute;
6049 struct kmem_cache *s;
6051 attribute = to_slab_attr(attr);
6054 if (!attribute->show)
6057 return attribute->show(s, buf);
6060 static ssize_t slab_attr_store(struct kobject *kobj,
6061 struct attribute *attr,
6062 const char *buf, size_t len)
6064 struct slab_attribute *attribute;
6065 struct kmem_cache *s;
6067 attribute = to_slab_attr(attr);
6070 if (!attribute->store)
6073 return attribute->store(s, buf, len);
6076 static void kmem_cache_release(struct kobject *k)
6078 slab_kmem_cache_release(to_slab(k));
6081 static const struct sysfs_ops slab_sysfs_ops = {
6082 .show = slab_attr_show,
6083 .store = slab_attr_store,
6086 static const struct kobj_type slab_ktype = {
6087 .sysfs_ops = &slab_sysfs_ops,
6088 .release = kmem_cache_release,
6091 static struct kset *slab_kset;
6093 static inline struct kset *cache_kset(struct kmem_cache *s)
6098 #define ID_STR_LENGTH 32
6100 /* Create a unique string id for a slab cache:
6102 * Format :[flags-]size
6104 static char *create_unique_id(struct kmem_cache *s)
6106 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6110 return ERR_PTR(-ENOMEM);
6114 * First flags affecting slabcache operations. We will only
6115 * get here for aliasable slabs so we do not need to support
6116 * too many flags. The flags here must cover all flags that
6117 * are matched during merging to guarantee that the id is
6120 if (s->flags & SLAB_CACHE_DMA)
6122 if (s->flags & SLAB_CACHE_DMA32)
6124 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6126 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6128 if (s->flags & SLAB_ACCOUNT)
6132 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6134 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6136 return ERR_PTR(-EINVAL);
6138 kmsan_unpoison_memory(name, p - name);
6142 static int sysfs_slab_add(struct kmem_cache *s)
6146 struct kset *kset = cache_kset(s);
6147 int unmergeable = slab_unmergeable(s);
6149 if (!unmergeable && disable_higher_order_debug &&
6150 (slub_debug & DEBUG_METADATA_FLAGS))
6155 * Slabcache can never be merged so we can use the name proper.
6156 * This is typically the case for debug situations. In that
6157 * case we can catch duplicate names easily.
6159 sysfs_remove_link(&slab_kset->kobj, s->name);
6163 * Create a unique name for the slab as a target
6166 name = create_unique_id(s);
6168 return PTR_ERR(name);
6171 s->kobj.kset = kset;
6172 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6176 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6181 /* Setup first alias */
6182 sysfs_slab_alias(s, s->name);
6189 kobject_del(&s->kobj);
6193 void sysfs_slab_unlink(struct kmem_cache *s)
6195 if (slab_state >= FULL)
6196 kobject_del(&s->kobj);
6199 void sysfs_slab_release(struct kmem_cache *s)
6201 if (slab_state >= FULL)
6202 kobject_put(&s->kobj);
6206 * Need to buffer aliases during bootup until sysfs becomes
6207 * available lest we lose that information.
6209 struct saved_alias {
6210 struct kmem_cache *s;
6212 struct saved_alias *next;
6215 static struct saved_alias *alias_list;
6217 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6219 struct saved_alias *al;
6221 if (slab_state == FULL) {
6223 * If we have a leftover link then remove it.
6225 sysfs_remove_link(&slab_kset->kobj, name);
6226 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6229 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6235 al->next = alias_list;
6237 kmsan_unpoison_memory(al, sizeof(*al));
6241 static int __init slab_sysfs_init(void)
6243 struct kmem_cache *s;
6246 mutex_lock(&slab_mutex);
6248 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6250 mutex_unlock(&slab_mutex);
6251 pr_err("Cannot register slab subsystem.\n");
6257 list_for_each_entry(s, &slab_caches, list) {
6258 err = sysfs_slab_add(s);
6260 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6264 while (alias_list) {
6265 struct saved_alias *al = alias_list;
6267 alias_list = alias_list->next;
6268 err = sysfs_slab_alias(al->s, al->name);
6270 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6275 mutex_unlock(&slab_mutex);
6278 late_initcall(slab_sysfs_init);
6279 #endif /* SLAB_SUPPORTS_SYSFS */
6281 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6282 static int slab_debugfs_show(struct seq_file *seq, void *v)
6284 struct loc_track *t = seq->private;
6288 idx = (unsigned long) t->idx;
6289 if (idx < t->count) {
6292 seq_printf(seq, "%7ld ", l->count);
6295 seq_printf(seq, "%pS", (void *)l->addr);
6297 seq_puts(seq, "<not-available>");
6300 seq_printf(seq, " waste=%lu/%lu",
6301 l->count * l->waste, l->waste);
6303 if (l->sum_time != l->min_time) {
6304 seq_printf(seq, " age=%ld/%llu/%ld",
6305 l->min_time, div_u64(l->sum_time, l->count),
6308 seq_printf(seq, " age=%ld", l->min_time);
6310 if (l->min_pid != l->max_pid)
6311 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6313 seq_printf(seq, " pid=%ld",
6316 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6317 seq_printf(seq, " cpus=%*pbl",
6318 cpumask_pr_args(to_cpumask(l->cpus)));
6320 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6321 seq_printf(seq, " nodes=%*pbl",
6322 nodemask_pr_args(&l->nodes));
6324 #ifdef CONFIG_STACKDEPOT
6326 depot_stack_handle_t handle;
6327 unsigned long *entries;
6328 unsigned int nr_entries, j;
6330 handle = READ_ONCE(l->handle);
6332 nr_entries = stack_depot_fetch(handle, &entries);
6333 seq_puts(seq, "\n");
6334 for (j = 0; j < nr_entries; j++)
6335 seq_printf(seq, " %pS\n", (void *)entries[j]);
6339 seq_puts(seq, "\n");
6342 if (!idx && !t->count)
6343 seq_puts(seq, "No data\n");
6348 static void slab_debugfs_stop(struct seq_file *seq, void *v)
6352 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6354 struct loc_track *t = seq->private;
6357 if (*ppos <= t->count)
6363 static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6365 struct location *loc1 = (struct location *)a;
6366 struct location *loc2 = (struct location *)b;
6368 if (loc1->count > loc2->count)
6374 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6376 struct loc_track *t = seq->private;
6382 static const struct seq_operations slab_debugfs_sops = {
6383 .start = slab_debugfs_start,
6384 .next = slab_debugfs_next,
6385 .stop = slab_debugfs_stop,
6386 .show = slab_debugfs_show,
6389 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6392 struct kmem_cache_node *n;
6393 enum track_item alloc;
6395 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6396 sizeof(struct loc_track));
6397 struct kmem_cache *s = file_inode(filep)->i_private;
6398 unsigned long *obj_map;
6403 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6405 seq_release_private(inode, filep);
6409 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6410 alloc = TRACK_ALLOC;
6414 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6415 bitmap_free(obj_map);
6416 seq_release_private(inode, filep);
6420 for_each_kmem_cache_node(s, node, n) {
6421 unsigned long flags;
6424 if (!atomic_long_read(&n->nr_slabs))
6427 spin_lock_irqsave(&n->list_lock, flags);
6428 list_for_each_entry(slab, &n->partial, slab_list)
6429 process_slab(t, s, slab, alloc, obj_map);
6430 list_for_each_entry(slab, &n->full, slab_list)
6431 process_slab(t, s, slab, alloc, obj_map);
6432 spin_unlock_irqrestore(&n->list_lock, flags);
6435 /* Sort locations by count */
6436 sort_r(t->loc, t->count, sizeof(struct location),
6437 cmp_loc_by_count, NULL, NULL);
6439 bitmap_free(obj_map);
6443 static int slab_debug_trace_release(struct inode *inode, struct file *file)
6445 struct seq_file *seq = file->private_data;
6446 struct loc_track *t = seq->private;
6449 return seq_release_private(inode, file);
6452 static const struct file_operations slab_debugfs_fops = {
6453 .open = slab_debug_trace_open,
6455 .llseek = seq_lseek,
6456 .release = slab_debug_trace_release,
6459 static void debugfs_slab_add(struct kmem_cache *s)
6461 struct dentry *slab_cache_dir;
6463 if (unlikely(!slab_debugfs_root))
6466 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6468 debugfs_create_file("alloc_traces", 0400,
6469 slab_cache_dir, s, &slab_debugfs_fops);
6471 debugfs_create_file("free_traces", 0400,
6472 slab_cache_dir, s, &slab_debugfs_fops);
6475 void debugfs_slab_release(struct kmem_cache *s)
6477 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
6480 static int __init slab_debugfs_init(void)
6482 struct kmem_cache *s;
6484 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6486 list_for_each_entry(s, &slab_caches, list)
6487 if (s->flags & SLAB_STORE_USER)
6488 debugfs_slab_add(s);
6493 __initcall(slab_debugfs_init);
6496 * The /proc/slabinfo ABI
6498 #ifdef CONFIG_SLUB_DEBUG
6499 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6501 unsigned long nr_slabs = 0;
6502 unsigned long nr_objs = 0;
6503 unsigned long nr_free = 0;
6505 struct kmem_cache_node *n;
6507 for_each_kmem_cache_node(s, node, n) {
6508 nr_slabs += node_nr_slabs(n);
6509 nr_objs += node_nr_objs(n);
6510 nr_free += count_partial(n, count_free);
6513 sinfo->active_objs = nr_objs - nr_free;
6514 sinfo->num_objs = nr_objs;
6515 sinfo->active_slabs = nr_slabs;
6516 sinfo->num_slabs = nr_slabs;
6517 sinfo->objects_per_slab = oo_objects(s->oo);
6518 sinfo->cache_order = oo_order(s->oo);
6521 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6525 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6526 size_t count, loff_t *ppos)
6530 #endif /* CONFIG_SLUB_DEBUG */