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/kmemleak.h>
38 #include <linux/stacktrace.h>
39 #include <linux/prefetch.h>
40 #include <linux/memcontrol.h>
41 #include <linux/random.h>
42 #include <kunit/test.h>
43 #include <kunit/test-bug.h>
44 #include <linux/sort.h>
46 #include <linux/debugfs.h>
47 #include <trace/events/kmem.h>
53 * 1. slab_mutex (Global Mutex)
54 * 2. node->list_lock (Spinlock)
55 * 3. kmem_cache->cpu_slab->lock (Local lock)
56 * 4. slab_lock(slab) (Only on some arches)
57 * 5. object_map_lock (Only for debugging)
61 * The role of the slab_mutex is to protect the list of all the slabs
62 * and to synchronize major metadata changes to slab cache structures.
63 * Also synchronizes memory hotplug callbacks.
67 * The slab_lock is a wrapper around the page lock, thus it is a bit
70 * The slab_lock is only used on arches that do not have the ability
71 * to do a cmpxchg_double. It only protects:
73 * A. slab->freelist -> List of free objects in a slab
74 * B. slab->inuse -> Number of objects in use
75 * C. slab->objects -> Number of objects in slab
76 * D. slab->frozen -> frozen state
80 * If a slab is frozen then it is exempt from list management. It is
81 * the cpu slab which is actively allocated from by the processor that
82 * froze it and it is not on any list. The processor that froze the
83 * slab is the one who can perform list operations on the slab. Other
84 * processors may put objects onto the freelist but the processor that
85 * froze the slab is the only one that can retrieve the objects from the
90 * The partially empty slabs cached on the CPU partial list are used
91 * for performance reasons, which speeds up the allocation process.
92 * These slabs are not frozen, but are also exempt from list management,
93 * by clearing the PG_workingset flag when moving out of the node
94 * partial list. Please see __slab_free() for more details.
96 * To sum up, the current scheme is:
97 * - node partial slab: PG_Workingset && !frozen
98 * - cpu partial slab: !PG_Workingset && !frozen
99 * - cpu slab: !PG_Workingset && frozen
100 * - full slab: !PG_Workingset && !frozen
104 * The list_lock protects the partial and full list on each node and
105 * the partial slab counter. If taken then no new slabs may be added or
106 * removed from the lists nor make the number of partial slabs be modified.
107 * (Note that the total number of slabs is an atomic value that may be
108 * modified without taking the list lock).
110 * The list_lock is a centralized lock and thus we avoid taking it as
111 * much as possible. As long as SLUB does not have to handle partial
112 * slabs, operations can continue without any centralized lock. F.e.
113 * allocating a long series of objects that fill up slabs does not require
116 * For debug caches, all allocations are forced to go through a list_lock
117 * protected region to serialize against concurrent validation.
119 * cpu_slab->lock local lock
121 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
122 * except the stat counters. This is a percpu structure manipulated only by
123 * the local cpu, so the lock protects against being preempted or interrupted
124 * by an irq. Fast path operations rely on lockless operations instead.
126 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
127 * which means the lockless fastpath cannot be used as it might interfere with
128 * an in-progress slow path operations. In this case the local lock is always
129 * taken but it still utilizes the freelist for the common operations.
133 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
134 * are fully lockless when satisfied from the percpu slab (and when
135 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
136 * They also don't disable preemption or migration or irqs. They rely on
137 * the transaction id (tid) field to detect being preempted or moved to
140 * irq, preemption, migration considerations
142 * Interrupts are disabled as part of list_lock or local_lock operations, or
143 * around the slab_lock operation, in order to make the slab allocator safe
144 * to use in the context of an irq.
146 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
147 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
148 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
149 * doesn't have to be revalidated in each section protected by the local lock.
151 * SLUB assigns one slab for allocation to each processor.
152 * Allocations only occur from these slabs called cpu slabs.
154 * Slabs with free elements are kept on a partial list and during regular
155 * operations no list for full slabs is used. If an object in a full slab is
156 * freed then the slab will show up again on the partial lists.
157 * We track full slabs for debugging purposes though because otherwise we
158 * cannot scan all objects.
160 * Slabs are freed when they become empty. Teardown and setup is
161 * minimal so we rely on the page allocators per cpu caches for
162 * fast frees and allocs.
164 * slab->frozen The slab is frozen and exempt from list processing.
165 * This means that the slab is dedicated to a purpose
166 * such as satisfying allocations for a specific
167 * processor. Objects may be freed in the slab while
168 * it is frozen but slab_free will then skip the usual
169 * list operations. It is up to the processor holding
170 * the slab to integrate the slab into the slab lists
171 * when the slab is no longer needed.
173 * One use of this flag is to mark slabs that are
174 * used for allocations. Then such a slab becomes a cpu
175 * slab. The cpu slab may be equipped with an additional
176 * freelist that allows lockless access to
177 * free objects in addition to the regular freelist
178 * that requires the slab lock.
180 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
181 * options set. This moves slab handling out of
182 * the fast path and disables lockless freelists.
186 * We could simply use migrate_disable()/enable() but as long as it's a
187 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
189 #ifndef CONFIG_PREEMPT_RT
190 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
191 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
192 #define USE_LOCKLESS_FAST_PATH() (true)
194 #define slub_get_cpu_ptr(var) \
199 #define slub_put_cpu_ptr(var) \
204 #define USE_LOCKLESS_FAST_PATH() (false)
207 #ifndef CONFIG_SLUB_TINY
208 #define __fastpath_inline __always_inline
210 #define __fastpath_inline
213 #ifdef CONFIG_SLUB_DEBUG
214 #ifdef CONFIG_SLUB_DEBUG_ON
215 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
217 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
219 #endif /* CONFIG_SLUB_DEBUG */
221 /* Structure holding parameters for get_partial() call chain */
222 struct partial_context {
224 unsigned int orig_size;
228 static inline bool kmem_cache_debug(struct kmem_cache *s)
230 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
233 static inline bool slub_debug_orig_size(struct kmem_cache *s)
235 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
236 (s->flags & SLAB_KMALLOC));
239 void *fixup_red_left(struct kmem_cache *s, void *p)
241 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
242 p += s->red_left_pad;
247 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
249 #ifdef CONFIG_SLUB_CPU_PARTIAL
250 return !kmem_cache_debug(s);
257 * Issues still to be resolved:
259 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
261 * - Variable sizing of the per node arrays
264 /* Enable to log cmpxchg failures */
265 #undef SLUB_DEBUG_CMPXCHG
267 #ifndef CONFIG_SLUB_TINY
269 * Minimum number of partial slabs. These will be left on the partial
270 * lists even if they are empty. kmem_cache_shrink may reclaim them.
272 #define MIN_PARTIAL 5
275 * Maximum number of desirable partial slabs.
276 * The existence of more partial slabs makes kmem_cache_shrink
277 * sort the partial list by the number of objects in use.
279 #define MAX_PARTIAL 10
281 #define MIN_PARTIAL 0
282 #define MAX_PARTIAL 0
285 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
286 SLAB_POISON | SLAB_STORE_USER)
289 * These debug flags cannot use CMPXCHG because there might be consistency
290 * issues when checking or reading debug information
292 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
297 * Debugging flags that require metadata to be stored in the slab. These get
298 * disabled when slub_debug=O is used and a cache's min order increases with
301 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
304 #define OO_MASK ((1 << OO_SHIFT) - 1)
305 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
307 /* Internal SLUB flags */
309 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
310 /* Use cmpxchg_double */
312 #ifdef system_has_freelist_aba
313 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
315 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0U)
319 * Tracking user of a slab.
321 #define TRACK_ADDRS_COUNT 16
323 unsigned long addr; /* Called from address */
324 #ifdef CONFIG_STACKDEPOT
325 depot_stack_handle_t handle;
327 int cpu; /* Was running on cpu */
328 int pid; /* Pid context */
329 unsigned long when; /* When did the operation occur */
332 enum track_item { TRACK_ALLOC, TRACK_FREE };
334 #ifdef SLAB_SUPPORTS_SYSFS
335 static int sysfs_slab_add(struct kmem_cache *);
336 static int sysfs_slab_alias(struct kmem_cache *, const char *);
338 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
339 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
343 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
344 static void debugfs_slab_add(struct kmem_cache *);
346 static inline void debugfs_slab_add(struct kmem_cache *s) { }
350 ALLOC_FASTPATH, /* Allocation from cpu slab */
351 ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */
352 FREE_FASTPATH, /* Free to cpu slab */
353 FREE_SLOWPATH, /* Freeing not to cpu slab */
354 FREE_FROZEN, /* Freeing to frozen slab */
355 FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */
356 FREE_REMOVE_PARTIAL, /* Freeing removes last object */
357 ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */
358 ALLOC_SLAB, /* Cpu slab acquired from page allocator */
359 ALLOC_REFILL, /* Refill cpu slab from slab freelist */
360 ALLOC_NODE_MISMATCH, /* Switching cpu slab */
361 FREE_SLAB, /* Slab freed to the page allocator */
362 CPUSLAB_FLUSH, /* Abandoning of the cpu slab */
363 DEACTIVATE_FULL, /* Cpu slab was full when deactivated */
364 DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */
365 DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */
366 DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */
367 DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
368 DEACTIVATE_BYPASS, /* Implicit deactivation */
369 ORDER_FALLBACK, /* Number of times fallback was necessary */
370 CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
371 CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */
372 CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */
373 CPU_PARTIAL_FREE, /* Refill cpu partial on free */
374 CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */
375 CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */
379 #ifndef CONFIG_SLUB_TINY
381 * When changing the layout, make sure freelist and tid are still compatible
382 * with this_cpu_cmpxchg_double() alignment requirements.
384 struct kmem_cache_cpu {
387 void **freelist; /* Pointer to next available object */
388 unsigned long tid; /* Globally unique transaction id */
390 freelist_aba_t freelist_tid;
392 struct slab *slab; /* The slab from which we are allocating */
393 #ifdef CONFIG_SLUB_CPU_PARTIAL
394 struct slab *partial; /* Partially allocated frozen slabs */
396 local_lock_t lock; /* Protects the fields above */
397 #ifdef CONFIG_SLUB_STATS
398 unsigned int stat[NR_SLUB_STAT_ITEMS];
401 #endif /* CONFIG_SLUB_TINY */
403 static inline void stat(const struct kmem_cache *s, enum stat_item si)
405 #ifdef CONFIG_SLUB_STATS
407 * The rmw is racy on a preemptible kernel but this is acceptable, so
408 * avoid this_cpu_add()'s irq-disable overhead.
410 raw_cpu_inc(s->cpu_slab->stat[si]);
415 void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
417 #ifdef CONFIG_SLUB_STATS
418 raw_cpu_add(s->cpu_slab->stat[si], v);
423 * The slab lists for all objects.
425 struct kmem_cache_node {
426 spinlock_t list_lock;
427 unsigned long nr_partial;
428 struct list_head partial;
429 #ifdef CONFIG_SLUB_DEBUG
430 atomic_long_t nr_slabs;
431 atomic_long_t total_objects;
432 struct list_head full;
436 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
438 return s->node[node];
442 * Iterator over all nodes. The body will be executed for each node that has
443 * a kmem_cache_node structure allocated (which is true for all online nodes)
445 #define for_each_kmem_cache_node(__s, __node, __n) \
446 for (__node = 0; __node < nr_node_ids; __node++) \
447 if ((__n = get_node(__s, __node)))
450 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
451 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
452 * differ during memory hotplug/hotremove operations.
453 * Protected by slab_mutex.
455 static nodemask_t slab_nodes;
457 #ifndef CONFIG_SLUB_TINY
459 * Workqueue used for flush_cpu_slab().
461 static struct workqueue_struct *flushwq;
464 /********************************************************************
465 * Core slab cache functions
466 *******************************************************************/
469 * freeptr_t represents a SLUB freelist pointer, which might be encoded
470 * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled.
472 typedef struct { unsigned long v; } freeptr_t;
475 * Returns freelist pointer (ptr). With hardening, this is obfuscated
476 * with an XOR of the address where the pointer is held and a per-cache
479 static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
480 void *ptr, unsigned long ptr_addr)
482 unsigned long encoded;
484 #ifdef CONFIG_SLAB_FREELIST_HARDENED
485 encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
487 encoded = (unsigned long)ptr;
489 return (freeptr_t){.v = encoded};
492 static inline void *freelist_ptr_decode(const struct kmem_cache *s,
493 freeptr_t ptr, unsigned long ptr_addr)
497 #ifdef CONFIG_SLAB_FREELIST_HARDENED
498 decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
500 decoded = (void *)ptr.v;
505 static inline void *get_freepointer(struct kmem_cache *s, void *object)
507 unsigned long ptr_addr;
510 object = kasan_reset_tag(object);
511 ptr_addr = (unsigned long)object + s->offset;
512 p = *(freeptr_t *)(ptr_addr);
513 return freelist_ptr_decode(s, p, ptr_addr);
516 #ifndef CONFIG_SLUB_TINY
517 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
519 prefetchw(object + s->offset);
524 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
525 * pointer value in the case the current thread loses the race for the next
526 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
527 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
528 * KMSAN will still check all arguments of cmpxchg because of imperfect
529 * handling of inline assembly.
530 * To work around this problem, we apply __no_kmsan_checks to ensure that
531 * get_freepointer_safe() returns initialized memory.
534 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
536 unsigned long freepointer_addr;
539 if (!debug_pagealloc_enabled_static())
540 return get_freepointer(s, object);
542 object = kasan_reset_tag(object);
543 freepointer_addr = (unsigned long)object + s->offset;
544 copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
545 return freelist_ptr_decode(s, p, freepointer_addr);
548 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
550 unsigned long freeptr_addr = (unsigned long)object + s->offset;
552 #ifdef CONFIG_SLAB_FREELIST_HARDENED
553 BUG_ON(object == fp); /* naive detection of double free or corruption */
556 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
557 *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
560 /* Loop over all objects in a slab */
561 #define for_each_object(__p, __s, __addr, __objects) \
562 for (__p = fixup_red_left(__s, __addr); \
563 __p < (__addr) + (__objects) * (__s)->size; \
566 static inline unsigned int order_objects(unsigned int order, unsigned int size)
568 return ((unsigned int)PAGE_SIZE << order) / size;
571 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
574 struct kmem_cache_order_objects x = {
575 (order << OO_SHIFT) + order_objects(order, size)
581 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
583 return x.x >> OO_SHIFT;
586 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
588 return x.x & OO_MASK;
591 #ifdef CONFIG_SLUB_CPU_PARTIAL
592 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
594 unsigned int nr_slabs;
596 s->cpu_partial = nr_objects;
599 * We take the number of objects but actually limit the number of
600 * slabs on the per cpu partial list, in order to limit excessive
601 * growth of the list. For simplicity we assume that the slabs will
604 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
605 s->cpu_partial_slabs = nr_slabs;
609 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
612 #endif /* CONFIG_SLUB_CPU_PARTIAL */
615 * Per slab locking using the pagelock
617 static __always_inline void slab_lock(struct slab *slab)
619 struct page *page = slab_page(slab);
621 VM_BUG_ON_PAGE(PageTail(page), page);
622 bit_spin_lock(PG_locked, &page->flags);
625 static __always_inline void slab_unlock(struct slab *slab)
627 struct page *page = slab_page(slab);
629 VM_BUG_ON_PAGE(PageTail(page), page);
630 bit_spin_unlock(PG_locked, &page->flags);
634 __update_freelist_fast(struct slab *slab,
635 void *freelist_old, unsigned long counters_old,
636 void *freelist_new, unsigned long counters_new)
638 #ifdef system_has_freelist_aba
639 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
640 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
642 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
649 __update_freelist_slow(struct slab *slab,
650 void *freelist_old, unsigned long counters_old,
651 void *freelist_new, unsigned long counters_new)
656 if (slab->freelist == freelist_old &&
657 slab->counters == counters_old) {
658 slab->freelist = freelist_new;
659 slab->counters = counters_new;
668 * Interrupts must be disabled (for the fallback code to work right), typically
669 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
670 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
671 * allocation/ free operation in hardirq context. Therefore nothing can
672 * interrupt the operation.
674 static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
675 void *freelist_old, unsigned long counters_old,
676 void *freelist_new, unsigned long counters_new,
681 if (USE_LOCKLESS_FAST_PATH())
682 lockdep_assert_irqs_disabled();
684 if (s->flags & __CMPXCHG_DOUBLE) {
685 ret = __update_freelist_fast(slab, freelist_old, counters_old,
686 freelist_new, counters_new);
688 ret = __update_freelist_slow(slab, freelist_old, counters_old,
689 freelist_new, counters_new);
695 stat(s, CMPXCHG_DOUBLE_FAIL);
697 #ifdef SLUB_DEBUG_CMPXCHG
698 pr_info("%s %s: cmpxchg double redo ", n, s->name);
704 static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
705 void *freelist_old, unsigned long counters_old,
706 void *freelist_new, unsigned long counters_new,
711 if (s->flags & __CMPXCHG_DOUBLE) {
712 ret = __update_freelist_fast(slab, freelist_old, counters_old,
713 freelist_new, counters_new);
717 local_irq_save(flags);
718 ret = __update_freelist_slow(slab, freelist_old, counters_old,
719 freelist_new, counters_new);
720 local_irq_restore(flags);
726 stat(s, CMPXCHG_DOUBLE_FAIL);
728 #ifdef SLUB_DEBUG_CMPXCHG
729 pr_info("%s %s: cmpxchg double redo ", n, s->name);
735 #ifdef CONFIG_SLUB_DEBUG
736 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
737 static DEFINE_SPINLOCK(object_map_lock);
739 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
742 void *addr = slab_address(slab);
745 bitmap_zero(obj_map, slab->objects);
747 for (p = slab->freelist; p; p = get_freepointer(s, p))
748 set_bit(__obj_to_index(s, addr, p), obj_map);
751 #if IS_ENABLED(CONFIG_KUNIT)
752 static bool slab_add_kunit_errors(void)
754 struct kunit_resource *resource;
756 if (!kunit_get_current_test())
759 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
763 (*(int *)resource->data)++;
764 kunit_put_resource(resource);
768 static inline bool slab_add_kunit_errors(void) { return false; }
771 static inline unsigned int size_from_object(struct kmem_cache *s)
773 if (s->flags & SLAB_RED_ZONE)
774 return s->size - s->red_left_pad;
779 static inline void *restore_red_left(struct kmem_cache *s, void *p)
781 if (s->flags & SLAB_RED_ZONE)
782 p -= s->red_left_pad;
790 #if defined(CONFIG_SLUB_DEBUG_ON)
791 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
793 static slab_flags_t slub_debug;
796 static char *slub_debug_string;
797 static int disable_higher_order_debug;
800 * slub is about to manipulate internal object metadata. This memory lies
801 * outside the range of the allocated object, so accessing it would normally
802 * be reported by kasan as a bounds error. metadata_access_enable() is used
803 * to tell kasan that these accesses are OK.
805 static inline void metadata_access_enable(void)
807 kasan_disable_current();
810 static inline void metadata_access_disable(void)
812 kasan_enable_current();
819 /* Verify that a pointer has an address that is valid within a slab page */
820 static inline int check_valid_pointer(struct kmem_cache *s,
821 struct slab *slab, void *object)
828 base = slab_address(slab);
829 object = kasan_reset_tag(object);
830 object = restore_red_left(s, object);
831 if (object < base || object >= base + slab->objects * s->size ||
832 (object - base) % s->size) {
839 static void print_section(char *level, char *text, u8 *addr,
842 metadata_access_enable();
843 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
844 16, 1, kasan_reset_tag((void *)addr), length, 1);
845 metadata_access_disable();
849 * See comment in calculate_sizes().
851 static inline bool freeptr_outside_object(struct kmem_cache *s)
853 return s->offset >= s->inuse;
857 * Return offset of the end of info block which is inuse + free pointer if
858 * not overlapping with object.
860 static inline unsigned int get_info_end(struct kmem_cache *s)
862 if (freeptr_outside_object(s))
863 return s->inuse + sizeof(void *);
868 static struct track *get_track(struct kmem_cache *s, void *object,
869 enum track_item alloc)
873 p = object + get_info_end(s);
875 return kasan_reset_tag(p + alloc);
878 #ifdef CONFIG_STACKDEPOT
879 static noinline depot_stack_handle_t set_track_prepare(void)
881 depot_stack_handle_t handle;
882 unsigned long entries[TRACK_ADDRS_COUNT];
883 unsigned int nr_entries;
885 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
886 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
891 static inline depot_stack_handle_t set_track_prepare(void)
897 static void set_track_update(struct kmem_cache *s, void *object,
898 enum track_item alloc, unsigned long addr,
899 depot_stack_handle_t handle)
901 struct track *p = get_track(s, object, alloc);
903 #ifdef CONFIG_STACKDEPOT
907 p->cpu = smp_processor_id();
908 p->pid = current->pid;
912 static __always_inline void set_track(struct kmem_cache *s, void *object,
913 enum track_item alloc, unsigned long addr)
915 depot_stack_handle_t handle = set_track_prepare();
917 set_track_update(s, object, alloc, addr, handle);
920 static void init_tracking(struct kmem_cache *s, void *object)
924 if (!(s->flags & SLAB_STORE_USER))
927 p = get_track(s, object, TRACK_ALLOC);
928 memset(p, 0, 2*sizeof(struct track));
931 static void print_track(const char *s, struct track *t, unsigned long pr_time)
933 depot_stack_handle_t handle __maybe_unused;
938 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
939 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
940 #ifdef CONFIG_STACKDEPOT
941 handle = READ_ONCE(t->handle);
943 stack_depot_print(handle);
945 pr_err("object allocation/free stack trace missing\n");
949 void print_tracking(struct kmem_cache *s, void *object)
951 unsigned long pr_time = jiffies;
952 if (!(s->flags & SLAB_STORE_USER))
955 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
956 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
959 static void print_slab_info(const struct slab *slab)
961 struct folio *folio = (struct folio *)slab_folio(slab);
963 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
964 slab, slab->objects, slab->inuse, slab->freelist,
965 folio_flags(folio, 0));
969 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
970 * family will round up the real request size to these fixed ones, so
971 * there could be an extra area than what is requested. Save the original
972 * request size in the meta data area, for better debug and sanity check.
974 static inline void set_orig_size(struct kmem_cache *s,
975 void *object, unsigned int orig_size)
977 void *p = kasan_reset_tag(object);
978 unsigned int kasan_meta_size;
980 if (!slub_debug_orig_size(s))
984 * KASAN can save its free meta data inside of the object at offset 0.
985 * If this meta data size is larger than 'orig_size', it will overlap
986 * the data redzone in [orig_size+1, object_size]. Thus, we adjust
987 * 'orig_size' to be as at least as big as KASAN's meta data.
989 kasan_meta_size = kasan_metadata_size(s, true);
990 if (kasan_meta_size > orig_size)
991 orig_size = kasan_meta_size;
993 p += get_info_end(s);
994 p += sizeof(struct track) * 2;
996 *(unsigned int *)p = orig_size;
999 static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
1001 void *p = kasan_reset_tag(object);
1003 if (!slub_debug_orig_size(s))
1004 return s->object_size;
1006 p += get_info_end(s);
1007 p += sizeof(struct track) * 2;
1009 return *(unsigned int *)p;
1012 void skip_orig_size_check(struct kmem_cache *s, const void *object)
1014 set_orig_size(s, (void *)object, s->object_size);
1017 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
1019 struct va_format vaf;
1022 va_start(args, fmt);
1025 pr_err("=============================================================================\n");
1026 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
1027 pr_err("-----------------------------------------------------------------------------\n\n");
1032 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
1034 struct va_format vaf;
1037 if (slab_add_kunit_errors())
1040 va_start(args, fmt);
1043 pr_err("FIX %s: %pV\n", s->name, &vaf);
1047 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1049 unsigned int off; /* Offset of last byte */
1050 u8 *addr = slab_address(slab);
1052 print_tracking(s, p);
1054 print_slab_info(slab);
1056 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1057 p, p - addr, get_freepointer(s, p));
1059 if (s->flags & SLAB_RED_ZONE)
1060 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
1062 else if (p > addr + 16)
1063 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1065 print_section(KERN_ERR, "Object ", p,
1066 min_t(unsigned int, s->object_size, PAGE_SIZE));
1067 if (s->flags & SLAB_RED_ZONE)
1068 print_section(KERN_ERR, "Redzone ", p + s->object_size,
1069 s->inuse - s->object_size);
1071 off = get_info_end(s);
1073 if (s->flags & SLAB_STORE_USER)
1074 off += 2 * sizeof(struct track);
1076 if (slub_debug_orig_size(s))
1077 off += sizeof(unsigned int);
1079 off += kasan_metadata_size(s, false);
1081 if (off != size_from_object(s))
1082 /* Beginning of the filler is the free pointer */
1083 print_section(KERN_ERR, "Padding ", p + off,
1084 size_from_object(s) - off);
1089 static void object_err(struct kmem_cache *s, struct slab *slab,
1090 u8 *object, char *reason)
1092 if (slab_add_kunit_errors())
1095 slab_bug(s, "%s", reason);
1096 print_trailer(s, slab, object);
1097 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1100 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1101 void **freelist, void *nextfree)
1103 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1104 !check_valid_pointer(s, slab, nextfree) && freelist) {
1105 object_err(s, slab, *freelist, "Freechain corrupt");
1107 slab_fix(s, "Isolate corrupted freechain");
1114 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1115 const char *fmt, ...)
1120 if (slab_add_kunit_errors())
1123 va_start(args, fmt);
1124 vsnprintf(buf, sizeof(buf), fmt, args);
1126 slab_bug(s, "%s", buf);
1127 print_slab_info(slab);
1129 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1132 static void init_object(struct kmem_cache *s, void *object, u8 val)
1134 u8 *p = kasan_reset_tag(object);
1135 unsigned int poison_size = s->object_size;
1137 if (s->flags & SLAB_RED_ZONE) {
1138 memset(p - s->red_left_pad, val, s->red_left_pad);
1140 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1142 * Redzone the extra allocated space by kmalloc than
1143 * requested, and the poison size will be limited to
1144 * the original request size accordingly.
1146 poison_size = get_orig_size(s, object);
1150 if (s->flags & __OBJECT_POISON) {
1151 memset(p, POISON_FREE, poison_size - 1);
1152 p[poison_size - 1] = POISON_END;
1155 if (s->flags & SLAB_RED_ZONE)
1156 memset(p + poison_size, val, s->inuse - poison_size);
1159 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1160 void *from, void *to)
1162 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1163 memset(from, data, to - from);
1166 static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1167 u8 *object, char *what,
1168 u8 *start, unsigned int value, unsigned int bytes)
1172 u8 *addr = slab_address(slab);
1174 metadata_access_enable();
1175 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1176 metadata_access_disable();
1180 end = start + bytes;
1181 while (end > fault && end[-1] == value)
1184 if (slab_add_kunit_errors())
1185 goto skip_bug_print;
1187 slab_bug(s, "%s overwritten", what);
1188 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1189 fault, end - 1, fault - addr,
1191 print_trailer(s, slab, object);
1192 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1195 restore_bytes(s, what, value, fault, end);
1203 * Bytes of the object to be managed.
1204 * If the freepointer may overlay the object then the free
1205 * pointer is at the middle of the object.
1207 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1210 * object + s->object_size
1211 * Padding to reach word boundary. This is also used for Redzoning.
1212 * Padding is extended by another word if Redzoning is enabled and
1213 * object_size == inuse.
1215 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1216 * 0xcc (RED_ACTIVE) for objects in use.
1219 * Meta data starts here.
1221 * A. Free pointer (if we cannot overwrite object on free)
1222 * B. Tracking data for SLAB_STORE_USER
1223 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1224 * D. Padding to reach required alignment boundary or at minimum
1225 * one word if debugging is on to be able to detect writes
1226 * before the word boundary.
1228 * Padding is done using 0x5a (POISON_INUSE)
1231 * Nothing is used beyond s->size.
1233 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1234 * ignored. And therefore no slab options that rely on these boundaries
1235 * may be used with merged slabcaches.
1238 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1240 unsigned long off = get_info_end(s); /* The end of info */
1242 if (s->flags & SLAB_STORE_USER) {
1243 /* We also have user information there */
1244 off += 2 * sizeof(struct track);
1246 if (s->flags & SLAB_KMALLOC)
1247 off += sizeof(unsigned int);
1250 off += kasan_metadata_size(s, false);
1252 if (size_from_object(s) == off)
1255 return check_bytes_and_report(s, slab, p, "Object padding",
1256 p + off, POISON_INUSE, size_from_object(s) - off);
1259 /* Check the pad bytes at the end of a slab page */
1260 static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1269 if (!(s->flags & SLAB_POISON))
1272 start = slab_address(slab);
1273 length = slab_size(slab);
1274 end = start + length;
1275 remainder = length % s->size;
1279 pad = end - remainder;
1280 metadata_access_enable();
1281 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1282 metadata_access_disable();
1285 while (end > fault && end[-1] == POISON_INUSE)
1288 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1289 fault, end - 1, fault - start);
1290 print_section(KERN_ERR, "Padding ", pad, remainder);
1292 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1295 static int check_object(struct kmem_cache *s, struct slab *slab,
1296 void *object, u8 val)
1299 u8 *endobject = object + s->object_size;
1300 unsigned int orig_size, kasan_meta_size;
1302 if (s->flags & SLAB_RED_ZONE) {
1303 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1304 object - s->red_left_pad, val, s->red_left_pad))
1307 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1308 endobject, val, s->inuse - s->object_size))
1311 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1312 orig_size = get_orig_size(s, object);
1314 if (s->object_size > orig_size &&
1315 !check_bytes_and_report(s, slab, object,
1316 "kmalloc Redzone", p + orig_size,
1317 val, s->object_size - orig_size)) {
1322 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1323 check_bytes_and_report(s, slab, p, "Alignment padding",
1324 endobject, POISON_INUSE,
1325 s->inuse - s->object_size);
1329 if (s->flags & SLAB_POISON) {
1330 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1332 * KASAN can save its free meta data inside of the
1333 * object at offset 0. Thus, skip checking the part of
1334 * the redzone that overlaps with the meta data.
1336 kasan_meta_size = kasan_metadata_size(s, true);
1337 if (kasan_meta_size < s->object_size - 1 &&
1338 !check_bytes_and_report(s, slab, p, "Poison",
1339 p + kasan_meta_size, POISON_FREE,
1340 s->object_size - kasan_meta_size - 1))
1342 if (kasan_meta_size < s->object_size &&
1343 !check_bytes_and_report(s, slab, p, "End Poison",
1344 p + s->object_size - 1, POISON_END, 1))
1348 * check_pad_bytes cleans up on its own.
1350 check_pad_bytes(s, slab, p);
1353 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1355 * Object and freepointer overlap. Cannot check
1356 * freepointer while object is allocated.
1360 /* Check free pointer validity */
1361 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1362 object_err(s, slab, p, "Freepointer corrupt");
1364 * No choice but to zap it and thus lose the remainder
1365 * of the free objects in this slab. May cause
1366 * another error because the object count is now wrong.
1368 set_freepointer(s, p, NULL);
1374 static int check_slab(struct kmem_cache *s, struct slab *slab)
1378 if (!folio_test_slab(slab_folio(slab))) {
1379 slab_err(s, slab, "Not a valid slab page");
1383 maxobj = order_objects(slab_order(slab), s->size);
1384 if (slab->objects > maxobj) {
1385 slab_err(s, slab, "objects %u > max %u",
1386 slab->objects, maxobj);
1389 if (slab->inuse > slab->objects) {
1390 slab_err(s, slab, "inuse %u > max %u",
1391 slab->inuse, slab->objects);
1394 /* Slab_pad_check fixes things up after itself */
1395 slab_pad_check(s, slab);
1400 * Determine if a certain object in a slab is on the freelist. Must hold the
1401 * slab lock to guarantee that the chains are in a consistent state.
1403 static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1407 void *object = NULL;
1410 fp = slab->freelist;
1411 while (fp && nr <= slab->objects) {
1414 if (!check_valid_pointer(s, slab, fp)) {
1416 object_err(s, slab, object,
1417 "Freechain corrupt");
1418 set_freepointer(s, object, NULL);
1420 slab_err(s, slab, "Freepointer corrupt");
1421 slab->freelist = NULL;
1422 slab->inuse = slab->objects;
1423 slab_fix(s, "Freelist cleared");
1429 fp = get_freepointer(s, object);
1433 max_objects = order_objects(slab_order(slab), s->size);
1434 if (max_objects > MAX_OBJS_PER_PAGE)
1435 max_objects = MAX_OBJS_PER_PAGE;
1437 if (slab->objects != max_objects) {
1438 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1439 slab->objects, max_objects);
1440 slab->objects = max_objects;
1441 slab_fix(s, "Number of objects adjusted");
1443 if (slab->inuse != slab->objects - nr) {
1444 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1445 slab->inuse, slab->objects - nr);
1446 slab->inuse = slab->objects - nr;
1447 slab_fix(s, "Object count adjusted");
1449 return search == NULL;
1452 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1455 if (s->flags & SLAB_TRACE) {
1456 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1458 alloc ? "alloc" : "free",
1459 object, slab->inuse,
1463 print_section(KERN_INFO, "Object ", (void *)object,
1471 * Tracking of fully allocated slabs for debugging purposes.
1473 static void add_full(struct kmem_cache *s,
1474 struct kmem_cache_node *n, struct slab *slab)
1476 if (!(s->flags & SLAB_STORE_USER))
1479 lockdep_assert_held(&n->list_lock);
1480 list_add(&slab->slab_list, &n->full);
1483 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1485 if (!(s->flags & SLAB_STORE_USER))
1488 lockdep_assert_held(&n->list_lock);
1489 list_del(&slab->slab_list);
1492 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1494 return atomic_long_read(&n->nr_slabs);
1497 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1499 struct kmem_cache_node *n = get_node(s, node);
1502 * May be called early in order to allocate a slab for the
1503 * kmem_cache_node structure. Solve the chicken-egg
1504 * dilemma by deferring the increment of the count during
1505 * bootstrap (see early_kmem_cache_node_alloc).
1508 atomic_long_inc(&n->nr_slabs);
1509 atomic_long_add(objects, &n->total_objects);
1512 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1514 struct kmem_cache_node *n = get_node(s, node);
1516 atomic_long_dec(&n->nr_slabs);
1517 atomic_long_sub(objects, &n->total_objects);
1520 /* Object debug checks for alloc/free paths */
1521 static void setup_object_debug(struct kmem_cache *s, void *object)
1523 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1526 init_object(s, object, SLUB_RED_INACTIVE);
1527 init_tracking(s, object);
1531 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1533 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1536 metadata_access_enable();
1537 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1538 metadata_access_disable();
1541 static inline int alloc_consistency_checks(struct kmem_cache *s,
1542 struct slab *slab, void *object)
1544 if (!check_slab(s, slab))
1547 if (!check_valid_pointer(s, slab, object)) {
1548 object_err(s, slab, object, "Freelist Pointer check fails");
1552 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1558 static noinline bool alloc_debug_processing(struct kmem_cache *s,
1559 struct slab *slab, void *object, int orig_size)
1561 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1562 if (!alloc_consistency_checks(s, slab, object))
1566 /* Success. Perform special debug activities for allocs */
1567 trace(s, slab, object, 1);
1568 set_orig_size(s, object, orig_size);
1569 init_object(s, object, SLUB_RED_ACTIVE);
1573 if (folio_test_slab(slab_folio(slab))) {
1575 * If this is a slab page then lets do the best we can
1576 * to avoid issues in the future. Marking all objects
1577 * as used avoids touching the remaining objects.
1579 slab_fix(s, "Marking all objects used");
1580 slab->inuse = slab->objects;
1581 slab->freelist = NULL;
1586 static inline int free_consistency_checks(struct kmem_cache *s,
1587 struct slab *slab, void *object, unsigned long addr)
1589 if (!check_valid_pointer(s, slab, object)) {
1590 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1594 if (on_freelist(s, slab, object)) {
1595 object_err(s, slab, object, "Object already free");
1599 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1602 if (unlikely(s != slab->slab_cache)) {
1603 if (!folio_test_slab(slab_folio(slab))) {
1604 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1606 } else if (!slab->slab_cache) {
1607 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1611 object_err(s, slab, object,
1612 "page slab pointer corrupt.");
1619 * Parse a block of slub_debug options. Blocks are delimited by ';'
1621 * @str: start of block
1622 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1623 * @slabs: return start of list of slabs, or NULL when there's no list
1624 * @init: assume this is initial parsing and not per-kmem-create parsing
1626 * returns the start of next block if there's any, or NULL
1629 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1631 bool higher_order_disable = false;
1633 /* Skip any completely empty blocks */
1634 while (*str && *str == ';')
1639 * No options but restriction on slabs. This means full
1640 * debugging for slabs matching a pattern.
1642 *flags = DEBUG_DEFAULT_FLAGS;
1647 /* Determine which debug features should be switched on */
1648 for (; *str && *str != ',' && *str != ';'; str++) {
1649 switch (tolower(*str)) {
1654 *flags |= SLAB_CONSISTENCY_CHECKS;
1657 *flags |= SLAB_RED_ZONE;
1660 *flags |= SLAB_POISON;
1663 *flags |= SLAB_STORE_USER;
1666 *flags |= SLAB_TRACE;
1669 *flags |= SLAB_FAILSLAB;
1673 * Avoid enabling debugging on caches if its minimum
1674 * order would increase as a result.
1676 higher_order_disable = true;
1680 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1689 /* Skip over the slab list */
1690 while (*str && *str != ';')
1693 /* Skip any completely empty blocks */
1694 while (*str && *str == ';')
1697 if (init && higher_order_disable)
1698 disable_higher_order_debug = 1;
1706 static int __init setup_slub_debug(char *str)
1709 slab_flags_t global_flags;
1712 bool global_slub_debug_changed = false;
1713 bool slab_list_specified = false;
1715 global_flags = DEBUG_DEFAULT_FLAGS;
1716 if (*str++ != '=' || !*str)
1718 * No options specified. Switch on full debugging.
1724 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1727 global_flags = flags;
1728 global_slub_debug_changed = true;
1730 slab_list_specified = true;
1731 if (flags & SLAB_STORE_USER)
1732 stack_depot_request_early_init();
1737 * For backwards compatibility, a single list of flags with list of
1738 * slabs means debugging is only changed for those slabs, so the global
1739 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1740 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1741 * long as there is no option specifying flags without a slab list.
1743 if (slab_list_specified) {
1744 if (!global_slub_debug_changed)
1745 global_flags = slub_debug;
1746 slub_debug_string = saved_str;
1749 slub_debug = global_flags;
1750 if (slub_debug & SLAB_STORE_USER)
1751 stack_depot_request_early_init();
1752 if (slub_debug != 0 || slub_debug_string)
1753 static_branch_enable(&slub_debug_enabled);
1755 static_branch_disable(&slub_debug_enabled);
1756 if ((static_branch_unlikely(&init_on_alloc) ||
1757 static_branch_unlikely(&init_on_free)) &&
1758 (slub_debug & SLAB_POISON))
1759 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1763 __setup("slub_debug", setup_slub_debug);
1766 * kmem_cache_flags - apply debugging options to the cache
1767 * @object_size: the size of an object without meta data
1768 * @flags: flags to set
1769 * @name: name of the cache
1771 * Debug option(s) are applied to @flags. In addition to the debug
1772 * option(s), if a slab name (or multiple) is specified i.e.
1773 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1774 * then only the select slabs will receive the debug option(s).
1776 slab_flags_t kmem_cache_flags(unsigned int object_size,
1777 slab_flags_t flags, const char *name)
1782 slab_flags_t block_flags;
1783 slab_flags_t slub_debug_local = slub_debug;
1785 if (flags & SLAB_NO_USER_FLAGS)
1789 * If the slab cache is for debugging (e.g. kmemleak) then
1790 * don't store user (stack trace) information by default,
1791 * but let the user enable it via the command line below.
1793 if (flags & SLAB_NOLEAKTRACE)
1794 slub_debug_local &= ~SLAB_STORE_USER;
1797 next_block = slub_debug_string;
1798 /* Go through all blocks of debug options, see if any matches our slab's name */
1799 while (next_block) {
1800 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1803 /* Found a block that has a slab list, search it */
1808 end = strchrnul(iter, ',');
1809 if (next_block && next_block < end)
1810 end = next_block - 1;
1812 glob = strnchr(iter, end - iter, '*');
1814 cmplen = glob - iter;
1816 cmplen = max_t(size_t, len, (end - iter));
1818 if (!strncmp(name, iter, cmplen)) {
1819 flags |= block_flags;
1823 if (!*end || *end == ';')
1829 return flags | slub_debug_local;
1831 #else /* !CONFIG_SLUB_DEBUG */
1832 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1834 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1836 static inline bool alloc_debug_processing(struct kmem_cache *s,
1837 struct slab *slab, void *object, int orig_size) { return true; }
1839 static inline bool free_debug_processing(struct kmem_cache *s,
1840 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1841 unsigned long addr, depot_stack_handle_t handle) { return true; }
1843 static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1844 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1845 void *object, u8 val) { return 1; }
1846 static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1847 static inline void set_track(struct kmem_cache *s, void *object,
1848 enum track_item alloc, unsigned long addr) {}
1849 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1850 struct slab *slab) {}
1851 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1852 struct slab *slab) {}
1853 slab_flags_t kmem_cache_flags(unsigned int object_size,
1854 slab_flags_t flags, const char *name)
1858 #define slub_debug 0
1860 #define disable_higher_order_debug 0
1862 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1864 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1866 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1869 #ifndef CONFIG_SLUB_TINY
1870 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1871 void **freelist, void *nextfree)
1876 #endif /* CONFIG_SLUB_DEBUG */
1878 static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s)
1880 return (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1881 NR_SLAB_RECLAIMABLE_B : NR_SLAB_UNRECLAIMABLE_B;
1884 #ifdef CONFIG_MEMCG_KMEM
1885 static inline void memcg_free_slab_cgroups(struct slab *slab)
1887 kfree(slab_objcgs(slab));
1888 slab->memcg_data = 0;
1891 static inline size_t obj_full_size(struct kmem_cache *s)
1894 * For each accounted object there is an extra space which is used
1895 * to store obj_cgroup membership. Charge it too.
1897 return s->size + sizeof(struct obj_cgroup *);
1901 * Returns false if the allocation should fail.
1903 static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
1904 struct list_lru *lru,
1905 struct obj_cgroup **objcgp,
1906 size_t objects, gfp_t flags)
1909 * The obtained objcg pointer is safe to use within the current scope,
1910 * defined by current task or set_active_memcg() pair.
1911 * obj_cgroup_get() is used to get a permanent reference.
1913 struct obj_cgroup *objcg = current_obj_cgroup();
1919 struct mem_cgroup *memcg;
1921 memcg = get_mem_cgroup_from_objcg(objcg);
1922 ret = memcg_list_lru_alloc(memcg, lru, flags);
1923 css_put(&memcg->css);
1929 if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s)))
1937 * Returns false if the allocation should fail.
1939 static __fastpath_inline
1940 bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
1941 struct obj_cgroup **objcgp, size_t objects,
1944 if (!memcg_kmem_online())
1947 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
1950 return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects,
1954 static void __memcg_slab_post_alloc_hook(struct kmem_cache *s,
1955 struct obj_cgroup *objcg,
1956 gfp_t flags, size_t size,
1963 flags &= gfp_allowed_mask;
1965 for (i = 0; i < size; i++) {
1967 slab = virt_to_slab(p[i]);
1969 if (!slab_objcgs(slab) &&
1970 memcg_alloc_slab_cgroups(slab, s, flags, false)) {
1971 obj_cgroup_uncharge(objcg, obj_full_size(s));
1975 off = obj_to_index(s, slab, p[i]);
1976 obj_cgroup_get(objcg);
1977 slab_objcgs(slab)[off] = objcg;
1978 mod_objcg_state(objcg, slab_pgdat(slab),
1979 cache_vmstat_idx(s), obj_full_size(s));
1981 obj_cgroup_uncharge(objcg, obj_full_size(s));
1986 static __fastpath_inline
1987 void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
1988 gfp_t flags, size_t size, void **p)
1990 if (likely(!memcg_kmem_online() || !objcg))
1993 return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
1996 static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
1997 void **p, int objects,
1998 struct obj_cgroup **objcgs)
2000 for (int i = 0; i < objects; i++) {
2001 struct obj_cgroup *objcg;
2004 off = obj_to_index(s, slab, p[i]);
2005 objcg = objcgs[off];
2010 obj_cgroup_uncharge(objcg, obj_full_size(s));
2011 mod_objcg_state(objcg, slab_pgdat(slab), cache_vmstat_idx(s),
2013 obj_cgroup_put(objcg);
2017 static __fastpath_inline
2018 void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2021 struct obj_cgroup **objcgs;
2023 if (!memcg_kmem_online())
2026 objcgs = slab_objcgs(slab);
2027 if (likely(!objcgs))
2030 __memcg_slab_free_hook(s, slab, p, objects, objcgs);
2034 void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2035 struct obj_cgroup *objcg)
2038 obj_cgroup_uncharge(objcg, objects * obj_full_size(s));
2040 #else /* CONFIG_MEMCG_KMEM */
2041 static inline struct mem_cgroup *memcg_from_slab_obj(void *ptr)
2046 static inline void memcg_free_slab_cgroups(struct slab *slab)
2050 static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s,
2051 struct list_lru *lru,
2052 struct obj_cgroup **objcgp,
2053 size_t objects, gfp_t flags)
2058 static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
2059 struct obj_cgroup *objcg,
2060 gfp_t flags, size_t size,
2065 static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2066 void **p, int objects)
2071 void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2072 struct obj_cgroup *objcg)
2075 #endif /* CONFIG_MEMCG_KMEM */
2078 * Hooks for other subsystems that check memory allocations. In a typical
2079 * production configuration these hooks all should produce no code at all.
2081 * Returns true if freeing of the object can proceed, false if its reuse
2082 * was delayed by KASAN quarantine, or it was returned to KFENCE.
2084 static __always_inline
2085 bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
2087 kmemleak_free_recursive(x, s->flags);
2088 kmsan_slab_free(s, x);
2090 debug_check_no_locks_freed(x, s->object_size);
2092 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2093 debug_check_no_obj_freed(x, s->object_size);
2095 /* Use KCSAN to help debug racy use-after-free. */
2096 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
2097 __kcsan_check_access(x, s->object_size,
2098 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2104 * As memory initialization might be integrated into KASAN,
2105 * kasan_slab_free and initialization memset's must be
2106 * kept together to avoid discrepancies in behavior.
2108 * The initialization memset's clear the object and the metadata,
2109 * but don't touch the SLAB redzone.
2111 if (unlikely(init)) {
2114 if (!kasan_has_integrated_init())
2115 memset(kasan_reset_tag(x), 0, s->object_size);
2116 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2117 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
2118 s->size - s->inuse - rsize);
2120 /* KASAN might put x into memory quarantine, delaying its reuse. */
2121 return !kasan_slab_free(s, x, init);
2124 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
2125 void **head, void **tail,
2131 void *old_tail = *tail;
2134 if (is_kfence_address(next)) {
2135 slab_free_hook(s, next, false);
2139 /* Head and tail of the reconstructed freelist */
2143 init = slab_want_init_on_free(s);
2147 next = get_freepointer(s, object);
2149 /* If object's reuse doesn't have to be delayed */
2150 if (likely(slab_free_hook(s, object, init))) {
2151 /* Move object to the new freelist */
2152 set_freepointer(s, object, *head);
2158 * Adjust the reconstructed freelist depth
2159 * accordingly if object's reuse is delayed.
2163 } while (object != old_tail);
2165 return *head != NULL;
2168 static void *setup_object(struct kmem_cache *s, void *object)
2170 setup_object_debug(s, object);
2171 object = kasan_init_slab_obj(s, object);
2172 if (unlikely(s->ctor)) {
2173 kasan_unpoison_new_object(s, object);
2175 kasan_poison_new_object(s, object);
2181 * Slab allocation and freeing
2183 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2184 struct kmem_cache_order_objects oo)
2186 struct folio *folio;
2188 unsigned int order = oo_order(oo);
2190 folio = (struct folio *)alloc_pages_node(node, flags, order);
2194 slab = folio_slab(folio);
2195 __folio_set_slab(folio);
2196 /* Make the flag visible before any changes to folio->mapping */
2198 if (folio_is_pfmemalloc(folio))
2199 slab_set_pfmemalloc(slab);
2204 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2205 /* Pre-initialize the random sequence cache */
2206 static int init_cache_random_seq(struct kmem_cache *s)
2208 unsigned int count = oo_objects(s->oo);
2211 /* Bailout if already initialised */
2215 err = cache_random_seq_create(s, count, GFP_KERNEL);
2217 pr_err("SLUB: Unable to initialize free list for %s\n",
2222 /* Transform to an offset on the set of pages */
2223 if (s->random_seq) {
2226 for (i = 0; i < count; i++)
2227 s->random_seq[i] *= s->size;
2232 /* Initialize each random sequence freelist per cache */
2233 static void __init init_freelist_randomization(void)
2235 struct kmem_cache *s;
2237 mutex_lock(&slab_mutex);
2239 list_for_each_entry(s, &slab_caches, list)
2240 init_cache_random_seq(s);
2242 mutex_unlock(&slab_mutex);
2245 /* Get the next entry on the pre-computed freelist randomized */
2246 static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
2247 unsigned long *pos, void *start,
2248 unsigned long page_limit,
2249 unsigned long freelist_count)
2254 * If the target page allocation failed, the number of objects on the
2255 * page might be smaller than the usual size defined by the cache.
2258 idx = s->random_seq[*pos];
2260 if (*pos >= freelist_count)
2262 } while (unlikely(idx >= page_limit));
2264 return (char *)start + idx;
2267 /* Shuffle the single linked freelist based on a random pre-computed sequence */
2268 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2273 unsigned long idx, pos, page_limit, freelist_count;
2275 if (slab->objects < 2 || !s->random_seq)
2278 freelist_count = oo_objects(s->oo);
2279 pos = get_random_u32_below(freelist_count);
2281 page_limit = slab->objects * s->size;
2282 start = fixup_red_left(s, slab_address(slab));
2284 /* First entry is used as the base of the freelist */
2285 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
2287 cur = setup_object(s, cur);
2288 slab->freelist = cur;
2290 for (idx = 1; idx < slab->objects; idx++) {
2291 next = next_freelist_entry(s, slab, &pos, start, page_limit,
2293 next = setup_object(s, next);
2294 set_freepointer(s, cur, next);
2297 set_freepointer(s, cur, NULL);
2302 static inline int init_cache_random_seq(struct kmem_cache *s)
2306 static inline void init_freelist_randomization(void) { }
2307 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2311 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2313 static __always_inline void account_slab(struct slab *slab, int order,
2314 struct kmem_cache *s, gfp_t gfp)
2316 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2317 memcg_alloc_slab_cgroups(slab, s, gfp, true);
2319 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2320 PAGE_SIZE << order);
2323 static __always_inline void unaccount_slab(struct slab *slab, int order,
2324 struct kmem_cache *s)
2326 if (memcg_kmem_online())
2327 memcg_free_slab_cgroups(slab);
2329 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2330 -(PAGE_SIZE << order));
2333 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2336 struct kmem_cache_order_objects oo = s->oo;
2338 void *start, *p, *next;
2342 flags &= gfp_allowed_mask;
2344 flags |= s->allocflags;
2347 * Let the initial higher-order allocation fail under memory pressure
2348 * so we fall-back to the minimum order allocation.
2350 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2351 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2352 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2354 slab = alloc_slab_page(alloc_gfp, node, oo);
2355 if (unlikely(!slab)) {
2359 * Allocation may have failed due to fragmentation.
2360 * Try a lower order alloc if possible
2362 slab = alloc_slab_page(alloc_gfp, node, oo);
2363 if (unlikely(!slab))
2365 stat(s, ORDER_FALLBACK);
2368 slab->objects = oo_objects(oo);
2372 account_slab(slab, oo_order(oo), s, flags);
2374 slab->slab_cache = s;
2376 kasan_poison_slab(slab);
2378 start = slab_address(slab);
2380 setup_slab_debug(s, slab, start);
2382 shuffle = shuffle_freelist(s, slab);
2385 start = fixup_red_left(s, start);
2386 start = setup_object(s, start);
2387 slab->freelist = start;
2388 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2390 next = setup_object(s, next);
2391 set_freepointer(s, p, next);
2394 set_freepointer(s, p, NULL);
2400 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2402 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2403 flags = kmalloc_fix_flags(flags);
2405 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2407 return allocate_slab(s,
2408 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2411 static void __free_slab(struct kmem_cache *s, struct slab *slab)
2413 struct folio *folio = slab_folio(slab);
2414 int order = folio_order(folio);
2415 int pages = 1 << order;
2417 __slab_clear_pfmemalloc(slab);
2418 folio->mapping = NULL;
2419 /* Make the mapping reset visible before clearing the flag */
2421 __folio_clear_slab(folio);
2422 mm_account_reclaimed_pages(pages);
2423 unaccount_slab(slab, order, s);
2424 __free_pages(&folio->page, order);
2427 static void rcu_free_slab(struct rcu_head *h)
2429 struct slab *slab = container_of(h, struct slab, rcu_head);
2431 __free_slab(slab->slab_cache, slab);
2434 static void free_slab(struct kmem_cache *s, struct slab *slab)
2436 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2439 slab_pad_check(s, slab);
2440 for_each_object(p, s, slab_address(slab), slab->objects)
2441 check_object(s, slab, p, SLUB_RED_INACTIVE);
2444 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2445 call_rcu(&slab->rcu_head, rcu_free_slab);
2447 __free_slab(s, slab);
2450 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2452 dec_slabs_node(s, slab_nid(slab), slab->objects);
2457 * SLUB reuses PG_workingset bit to keep track of whether it's on
2458 * the per-node partial list.
2460 static inline bool slab_test_node_partial(const struct slab *slab)
2462 return folio_test_workingset((struct folio *)slab_folio(slab));
2465 static inline void slab_set_node_partial(struct slab *slab)
2467 set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2470 static inline void slab_clear_node_partial(struct slab *slab)
2472 clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2476 * Management of partially allocated slabs.
2479 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2482 if (tail == DEACTIVATE_TO_TAIL)
2483 list_add_tail(&slab->slab_list, &n->partial);
2485 list_add(&slab->slab_list, &n->partial);
2486 slab_set_node_partial(slab);
2489 static inline void add_partial(struct kmem_cache_node *n,
2490 struct slab *slab, int tail)
2492 lockdep_assert_held(&n->list_lock);
2493 __add_partial(n, slab, tail);
2496 static inline void remove_partial(struct kmem_cache_node *n,
2499 lockdep_assert_held(&n->list_lock);
2500 list_del(&slab->slab_list);
2501 slab_clear_node_partial(slab);
2506 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2507 * slab from the n->partial list. Remove only a single object from the slab, do
2508 * the alloc_debug_processing() checks and leave the slab on the list, or move
2509 * it to full list if it was the last free object.
2511 static void *alloc_single_from_partial(struct kmem_cache *s,
2512 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2516 lockdep_assert_held(&n->list_lock);
2518 object = slab->freelist;
2519 slab->freelist = get_freepointer(s, object);
2522 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2523 remove_partial(n, slab);
2527 if (slab->inuse == slab->objects) {
2528 remove_partial(n, slab);
2529 add_full(s, n, slab);
2536 * Called only for kmem_cache_debug() caches to allocate from a freshly
2537 * allocated slab. Allocate a single object instead of whole freelist
2538 * and put the slab to the partial (or full) list.
2540 static void *alloc_single_from_new_slab(struct kmem_cache *s,
2541 struct slab *slab, int orig_size)
2543 int nid = slab_nid(slab);
2544 struct kmem_cache_node *n = get_node(s, nid);
2545 unsigned long flags;
2549 object = slab->freelist;
2550 slab->freelist = get_freepointer(s, object);
2553 if (!alloc_debug_processing(s, slab, object, orig_size))
2555 * It's not really expected that this would fail on a
2556 * freshly allocated slab, but a concurrent memory
2557 * corruption in theory could cause that.
2561 spin_lock_irqsave(&n->list_lock, flags);
2563 if (slab->inuse == slab->objects)
2564 add_full(s, n, slab);
2566 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2568 inc_slabs_node(s, nid, slab->objects);
2569 spin_unlock_irqrestore(&n->list_lock, flags);
2574 #ifdef CONFIG_SLUB_CPU_PARTIAL
2575 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2577 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2580 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2583 * Try to allocate a partial slab from a specific node.
2585 static struct slab *get_partial_node(struct kmem_cache *s,
2586 struct kmem_cache_node *n,
2587 struct partial_context *pc)
2589 struct slab *slab, *slab2, *partial = NULL;
2590 unsigned long flags;
2591 unsigned int partial_slabs = 0;
2594 * Racy check. If we mistakenly see no partial slabs then we
2595 * just allocate an empty slab. If we mistakenly try to get a
2596 * partial slab and there is none available then get_partial()
2599 if (!n || !n->nr_partial)
2602 spin_lock_irqsave(&n->list_lock, flags);
2603 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2604 if (!pfmemalloc_match(slab, pc->flags))
2607 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2608 void *object = alloc_single_from_partial(s, n, slab,
2612 pc->object = object;
2618 remove_partial(n, slab);
2622 stat(s, ALLOC_FROM_PARTIAL);
2624 put_cpu_partial(s, slab, 0);
2625 stat(s, CPU_PARTIAL_NODE);
2628 #ifdef CONFIG_SLUB_CPU_PARTIAL
2629 if (!kmem_cache_has_cpu_partial(s)
2630 || partial_slabs > s->cpu_partial_slabs / 2)
2637 spin_unlock_irqrestore(&n->list_lock, flags);
2642 * Get a slab from somewhere. Search in increasing NUMA distances.
2644 static struct slab *get_any_partial(struct kmem_cache *s,
2645 struct partial_context *pc)
2648 struct zonelist *zonelist;
2651 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2653 unsigned int cpuset_mems_cookie;
2656 * The defrag ratio allows a configuration of the tradeoffs between
2657 * inter node defragmentation and node local allocations. A lower
2658 * defrag_ratio increases the tendency to do local allocations
2659 * instead of attempting to obtain partial slabs from other nodes.
2661 * If the defrag_ratio is set to 0 then kmalloc() always
2662 * returns node local objects. If the ratio is higher then kmalloc()
2663 * may return off node objects because partial slabs are obtained
2664 * from other nodes and filled up.
2666 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2667 * (which makes defrag_ratio = 1000) then every (well almost)
2668 * allocation will first attempt to defrag slab caches on other nodes.
2669 * This means scanning over all nodes to look for partial slabs which
2670 * may be expensive if we do it every time we are trying to find a slab
2671 * with available objects.
2673 if (!s->remote_node_defrag_ratio ||
2674 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2678 cpuset_mems_cookie = read_mems_allowed_begin();
2679 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2680 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2681 struct kmem_cache_node *n;
2683 n = get_node(s, zone_to_nid(zone));
2685 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2686 n->nr_partial > s->min_partial) {
2687 slab = get_partial_node(s, n, pc);
2690 * Don't check read_mems_allowed_retry()
2691 * here - if mems_allowed was updated in
2692 * parallel, that was a harmless race
2693 * between allocation and the cpuset
2700 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2701 #endif /* CONFIG_NUMA */
2706 * Get a partial slab, lock it and return it.
2708 static struct slab *get_partial(struct kmem_cache *s, int node,
2709 struct partial_context *pc)
2712 int searchnode = node;
2714 if (node == NUMA_NO_NODE)
2715 searchnode = numa_mem_id();
2717 slab = get_partial_node(s, get_node(s, searchnode), pc);
2718 if (slab || node != NUMA_NO_NODE)
2721 return get_any_partial(s, pc);
2724 #ifndef CONFIG_SLUB_TINY
2726 #ifdef CONFIG_PREEMPTION
2728 * Calculate the next globally unique transaction for disambiguation
2729 * during cmpxchg. The transactions start with the cpu number and are then
2730 * incremented by CONFIG_NR_CPUS.
2732 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2735 * No preemption supported therefore also no need to check for
2739 #endif /* CONFIG_PREEMPTION */
2741 static inline unsigned long next_tid(unsigned long tid)
2743 return tid + TID_STEP;
2746 #ifdef SLUB_DEBUG_CMPXCHG
2747 static inline unsigned int tid_to_cpu(unsigned long tid)
2749 return tid % TID_STEP;
2752 static inline unsigned long tid_to_event(unsigned long tid)
2754 return tid / TID_STEP;
2758 static inline unsigned int init_tid(int cpu)
2763 static inline void note_cmpxchg_failure(const char *n,
2764 const struct kmem_cache *s, unsigned long tid)
2766 #ifdef SLUB_DEBUG_CMPXCHG
2767 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2769 pr_info("%s %s: cmpxchg redo ", n, s->name);
2771 #ifdef CONFIG_PREEMPTION
2772 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2773 pr_warn("due to cpu change %d -> %d\n",
2774 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2777 if (tid_to_event(tid) != tid_to_event(actual_tid))
2778 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2779 tid_to_event(tid), tid_to_event(actual_tid));
2781 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2782 actual_tid, tid, next_tid(tid));
2784 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2787 static void init_kmem_cache_cpus(struct kmem_cache *s)
2790 struct kmem_cache_cpu *c;
2792 for_each_possible_cpu(cpu) {
2793 c = per_cpu_ptr(s->cpu_slab, cpu);
2794 local_lock_init(&c->lock);
2795 c->tid = init_tid(cpu);
2800 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2801 * unfreezes the slabs and puts it on the proper list.
2802 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2805 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2808 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2810 void *nextfree, *freelist_iter, *freelist_tail;
2811 int tail = DEACTIVATE_TO_HEAD;
2812 unsigned long flags = 0;
2816 if (slab->freelist) {
2817 stat(s, DEACTIVATE_REMOTE_FREES);
2818 tail = DEACTIVATE_TO_TAIL;
2822 * Stage one: Count the objects on cpu's freelist as free_delta and
2823 * remember the last object in freelist_tail for later splicing.
2825 freelist_tail = NULL;
2826 freelist_iter = freelist;
2827 while (freelist_iter) {
2828 nextfree = get_freepointer(s, freelist_iter);
2831 * If 'nextfree' is invalid, it is possible that the object at
2832 * 'freelist_iter' is already corrupted. So isolate all objects
2833 * starting at 'freelist_iter' by skipping them.
2835 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2838 freelist_tail = freelist_iter;
2841 freelist_iter = nextfree;
2845 * Stage two: Unfreeze the slab while splicing the per-cpu
2846 * freelist to the head of slab's freelist.
2849 old.freelist = READ_ONCE(slab->freelist);
2850 old.counters = READ_ONCE(slab->counters);
2851 VM_BUG_ON(!old.frozen);
2853 /* Determine target state of the slab */
2854 new.counters = old.counters;
2856 if (freelist_tail) {
2857 new.inuse -= free_delta;
2858 set_freepointer(s, freelist_tail, old.freelist);
2859 new.freelist = freelist;
2861 new.freelist = old.freelist;
2863 } while (!slab_update_freelist(s, slab,
2864 old.freelist, old.counters,
2865 new.freelist, new.counters,
2866 "unfreezing slab"));
2869 * Stage three: Manipulate the slab list based on the updated state.
2871 if (!new.inuse && n->nr_partial >= s->min_partial) {
2872 stat(s, DEACTIVATE_EMPTY);
2873 discard_slab(s, slab);
2875 } else if (new.freelist) {
2876 spin_lock_irqsave(&n->list_lock, flags);
2877 add_partial(n, slab, tail);
2878 spin_unlock_irqrestore(&n->list_lock, flags);
2881 stat(s, DEACTIVATE_FULL);
2885 #ifdef CONFIG_SLUB_CPU_PARTIAL
2886 static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
2888 struct kmem_cache_node *n = NULL, *n2 = NULL;
2889 struct slab *slab, *slab_to_discard = NULL;
2890 unsigned long flags = 0;
2892 while (partial_slab) {
2893 slab = partial_slab;
2894 partial_slab = slab->next;
2896 n2 = get_node(s, slab_nid(slab));
2899 spin_unlock_irqrestore(&n->list_lock, flags);
2902 spin_lock_irqsave(&n->list_lock, flags);
2905 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
2906 slab->next = slab_to_discard;
2907 slab_to_discard = slab;
2909 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2910 stat(s, FREE_ADD_PARTIAL);
2915 spin_unlock_irqrestore(&n->list_lock, flags);
2917 while (slab_to_discard) {
2918 slab = slab_to_discard;
2919 slab_to_discard = slab_to_discard->next;
2921 stat(s, DEACTIVATE_EMPTY);
2922 discard_slab(s, slab);
2928 * Put all the cpu partial slabs to the node partial list.
2930 static void put_partials(struct kmem_cache *s)
2932 struct slab *partial_slab;
2933 unsigned long flags;
2935 local_lock_irqsave(&s->cpu_slab->lock, flags);
2936 partial_slab = this_cpu_read(s->cpu_slab->partial);
2937 this_cpu_write(s->cpu_slab->partial, NULL);
2938 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2941 __put_partials(s, partial_slab);
2944 static void put_partials_cpu(struct kmem_cache *s,
2945 struct kmem_cache_cpu *c)
2947 struct slab *partial_slab;
2949 partial_slab = slub_percpu_partial(c);
2953 __put_partials(s, partial_slab);
2957 * Put a slab into a partial slab slot if available.
2959 * If we did not find a slot then simply move all the partials to the
2960 * per node partial list.
2962 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2964 struct slab *oldslab;
2965 struct slab *slab_to_put = NULL;
2966 unsigned long flags;
2969 local_lock_irqsave(&s->cpu_slab->lock, flags);
2971 oldslab = this_cpu_read(s->cpu_slab->partial);
2974 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2976 * Partial array is full. Move the existing set to the
2977 * per node partial list. Postpone the actual unfreezing
2978 * outside of the critical section.
2980 slab_to_put = oldslab;
2983 slabs = oldslab->slabs;
2989 slab->slabs = slabs;
2990 slab->next = oldslab;
2992 this_cpu_write(s->cpu_slab->partial, slab);
2994 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2997 __put_partials(s, slab_to_put);
2998 stat(s, CPU_PARTIAL_DRAIN);
3002 #else /* CONFIG_SLUB_CPU_PARTIAL */
3004 static inline void put_partials(struct kmem_cache *s) { }
3005 static inline void put_partials_cpu(struct kmem_cache *s,
3006 struct kmem_cache_cpu *c) { }
3008 #endif /* CONFIG_SLUB_CPU_PARTIAL */
3010 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3012 unsigned long flags;
3016 local_lock_irqsave(&s->cpu_slab->lock, flags);
3019 freelist = c->freelist;
3023 c->tid = next_tid(c->tid);
3025 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3028 deactivate_slab(s, slab, freelist);
3029 stat(s, CPUSLAB_FLUSH);
3033 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3035 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3036 void *freelist = c->freelist;
3037 struct slab *slab = c->slab;
3041 c->tid = next_tid(c->tid);
3044 deactivate_slab(s, slab, freelist);
3045 stat(s, CPUSLAB_FLUSH);
3048 put_partials_cpu(s, c);
3051 struct slub_flush_work {
3052 struct work_struct work;
3053 struct kmem_cache *s;
3060 * Called from CPU work handler with migration disabled.
3062 static void flush_cpu_slab(struct work_struct *w)
3064 struct kmem_cache *s;
3065 struct kmem_cache_cpu *c;
3066 struct slub_flush_work *sfw;
3068 sfw = container_of(w, struct slub_flush_work, work);
3071 c = this_cpu_ptr(s->cpu_slab);
3079 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3081 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3083 return c->slab || slub_percpu_partial(c);
3086 static DEFINE_MUTEX(flush_lock);
3087 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3089 static void flush_all_cpus_locked(struct kmem_cache *s)
3091 struct slub_flush_work *sfw;
3094 lockdep_assert_cpus_held();
3095 mutex_lock(&flush_lock);
3097 for_each_online_cpu(cpu) {
3098 sfw = &per_cpu(slub_flush, cpu);
3099 if (!has_cpu_slab(cpu, s)) {
3103 INIT_WORK(&sfw->work, flush_cpu_slab);
3106 queue_work_on(cpu, flushwq, &sfw->work);
3109 for_each_online_cpu(cpu) {
3110 sfw = &per_cpu(slub_flush, cpu);
3113 flush_work(&sfw->work);
3116 mutex_unlock(&flush_lock);
3119 static void flush_all(struct kmem_cache *s)
3122 flush_all_cpus_locked(s);
3127 * Use the cpu notifier to insure that the cpu slabs are flushed when
3130 static int slub_cpu_dead(unsigned int cpu)
3132 struct kmem_cache *s;
3134 mutex_lock(&slab_mutex);
3135 list_for_each_entry(s, &slab_caches, list)
3136 __flush_cpu_slab(s, cpu);
3137 mutex_unlock(&slab_mutex);
3141 #else /* CONFIG_SLUB_TINY */
3142 static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
3143 static inline void flush_all(struct kmem_cache *s) { }
3144 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
3145 static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3146 #endif /* CONFIG_SLUB_TINY */
3149 * Check if the objects in a per cpu structure fit numa
3150 * locality expectations.
3152 static inline int node_match(struct slab *slab, int node)
3155 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3161 #ifdef CONFIG_SLUB_DEBUG
3162 static int count_free(struct slab *slab)
3164 return slab->objects - slab->inuse;
3167 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3169 return atomic_long_read(&n->total_objects);
3172 /* Supports checking bulk free of a constructed freelist */
3173 static inline bool free_debug_processing(struct kmem_cache *s,
3174 struct slab *slab, void *head, void *tail, int *bulk_cnt,
3175 unsigned long addr, depot_stack_handle_t handle)
3177 bool checks_ok = false;
3178 void *object = head;
3181 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3182 if (!check_slab(s, slab))
3186 if (slab->inuse < *bulk_cnt) {
3187 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3188 slab->inuse, *bulk_cnt);
3194 if (++cnt > *bulk_cnt)
3197 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3198 if (!free_consistency_checks(s, slab, object, addr))
3202 if (s->flags & SLAB_STORE_USER)
3203 set_track_update(s, object, TRACK_FREE, addr, handle);
3204 trace(s, slab, object, 0);
3205 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3206 init_object(s, object, SLUB_RED_INACTIVE);
3208 /* Reached end of constructed freelist yet? */
3209 if (object != tail) {
3210 object = get_freepointer(s, object);
3216 if (cnt != *bulk_cnt) {
3217 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3225 slab_fix(s, "Object at 0x%p not freed", object);
3229 #endif /* CONFIG_SLUB_DEBUG */
3231 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
3232 static unsigned long count_partial(struct kmem_cache_node *n,
3233 int (*get_count)(struct slab *))
3235 unsigned long flags;
3236 unsigned long x = 0;
3239 spin_lock_irqsave(&n->list_lock, flags);
3240 list_for_each_entry(slab, &n->partial, slab_list)
3241 x += get_count(slab);
3242 spin_unlock_irqrestore(&n->list_lock, flags);
3245 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3247 #ifdef CONFIG_SLUB_DEBUG
3248 static noinline void
3249 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3251 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3252 DEFAULT_RATELIMIT_BURST);
3254 struct kmem_cache_node *n;
3256 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3259 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
3260 nid, gfpflags, &gfpflags);
3261 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3262 s->name, s->object_size, s->size, oo_order(s->oo),
3265 if (oo_order(s->min) > get_order(s->object_size))
3266 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
3269 for_each_kmem_cache_node(s, node, n) {
3270 unsigned long nr_slabs;
3271 unsigned long nr_objs;
3272 unsigned long nr_free;
3274 nr_free = count_partial(n, count_free);
3275 nr_slabs = node_nr_slabs(n);
3276 nr_objs = node_nr_objs(n);
3278 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3279 node, nr_slabs, nr_objs, nr_free);
3282 #else /* CONFIG_SLUB_DEBUG */
3284 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3287 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3289 if (unlikely(slab_test_pfmemalloc(slab)))
3290 return gfp_pfmemalloc_allowed(gfpflags);
3295 #ifndef CONFIG_SLUB_TINY
3297 __update_cpu_freelist_fast(struct kmem_cache *s,
3298 void *freelist_old, void *freelist_new,
3301 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3302 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3304 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3305 &old.full, new.full);
3309 * Check the slab->freelist and either transfer the freelist to the
3310 * per cpu freelist or deactivate the slab.
3312 * The slab is still frozen if the return value is not NULL.
3314 * If this function returns NULL then the slab has been unfrozen.
3316 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3319 unsigned long counters;
3322 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3325 freelist = slab->freelist;
3326 counters = slab->counters;
3328 new.counters = counters;
3329 VM_BUG_ON(!new.frozen);
3331 new.inuse = slab->objects;
3332 new.frozen = freelist != NULL;
3334 } while (!__slab_update_freelist(s, slab,
3343 * Freeze the partial slab and return the pointer to the freelist.
3345 static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3348 unsigned long counters;
3352 freelist = slab->freelist;
3353 counters = slab->counters;
3355 new.counters = counters;
3356 VM_BUG_ON(new.frozen);
3358 new.inuse = slab->objects;
3361 } while (!slab_update_freelist(s, slab,
3370 * Slow path. The lockless freelist is empty or we need to perform
3373 * Processing is still very fast if new objects have been freed to the
3374 * regular freelist. In that case we simply take over the regular freelist
3375 * as the lockless freelist and zap the regular freelist.
3377 * If that is not working then we fall back to the partial lists. We take the
3378 * first element of the freelist as the object to allocate now and move the
3379 * rest of the freelist to the lockless freelist.
3381 * And if we were unable to get a new slab from the partial slab lists then
3382 * we need to allocate a new slab. This is the slowest path since it involves
3383 * a call to the page allocator and the setup of a new slab.
3385 * Version of __slab_alloc to use when we know that preemption is
3386 * already disabled (which is the case for bulk allocation).
3388 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3389 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3393 unsigned long flags;
3394 struct partial_context pc;
3396 stat(s, ALLOC_SLOWPATH);
3400 slab = READ_ONCE(c->slab);
3403 * if the node is not online or has no normal memory, just
3404 * ignore the node constraint
3406 if (unlikely(node != NUMA_NO_NODE &&
3407 !node_isset(node, slab_nodes)))
3408 node = NUMA_NO_NODE;
3412 if (unlikely(!node_match(slab, node))) {
3414 * same as above but node_match() being false already
3415 * implies node != NUMA_NO_NODE
3417 if (!node_isset(node, slab_nodes)) {
3418 node = NUMA_NO_NODE;
3420 stat(s, ALLOC_NODE_MISMATCH);
3421 goto deactivate_slab;
3426 * By rights, we should be searching for a slab page that was
3427 * PFMEMALLOC but right now, we are losing the pfmemalloc
3428 * information when the page leaves the per-cpu allocator
3430 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3431 goto deactivate_slab;
3433 /* must check again c->slab in case we got preempted and it changed */
3434 local_lock_irqsave(&s->cpu_slab->lock, flags);
3435 if (unlikely(slab != c->slab)) {
3436 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3439 freelist = c->freelist;
3443 freelist = get_freelist(s, slab);
3447 c->tid = next_tid(c->tid);
3448 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3449 stat(s, DEACTIVATE_BYPASS);
3453 stat(s, ALLOC_REFILL);
3457 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3460 * freelist is pointing to the list of objects to be used.
3461 * slab is pointing to the slab from which the objects are obtained.
3462 * That slab must be frozen for per cpu allocations to work.
3464 VM_BUG_ON(!c->slab->frozen);
3465 c->freelist = get_freepointer(s, freelist);
3466 c->tid = next_tid(c->tid);
3467 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3472 local_lock_irqsave(&s->cpu_slab->lock, flags);
3473 if (slab != c->slab) {
3474 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3477 freelist = c->freelist;
3480 c->tid = next_tid(c->tid);
3481 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3482 deactivate_slab(s, slab, freelist);
3486 #ifdef CONFIG_SLUB_CPU_PARTIAL
3487 while (slub_percpu_partial(c)) {
3488 local_lock_irqsave(&s->cpu_slab->lock, flags);
3489 if (unlikely(c->slab)) {
3490 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3493 if (unlikely(!slub_percpu_partial(c))) {
3494 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3495 /* we were preempted and partial list got empty */
3499 slab = slub_percpu_partial(c);
3500 slub_set_percpu_partial(c, slab);
3501 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3502 stat(s, CPU_PARTIAL_ALLOC);
3504 if (unlikely(!node_match(slab, node) ||
3505 !pfmemalloc_match(slab, gfpflags))) {
3507 __put_partials(s, slab);
3511 freelist = freeze_slab(s, slab);
3512 goto retry_load_slab;
3518 pc.flags = gfpflags;
3519 pc.orig_size = orig_size;
3520 slab = get_partial(s, node, &pc);
3522 if (kmem_cache_debug(s)) {
3523 freelist = pc.object;
3525 * For debug caches here we had to go through
3526 * alloc_single_from_partial() so just store the
3527 * tracking info and return the object.
3529 if (s->flags & SLAB_STORE_USER)
3530 set_track(s, freelist, TRACK_ALLOC, addr);
3535 freelist = freeze_slab(s, slab);
3536 goto retry_load_slab;
3539 slub_put_cpu_ptr(s->cpu_slab);
3540 slab = new_slab(s, gfpflags, node);
3541 c = slub_get_cpu_ptr(s->cpu_slab);
3543 if (unlikely(!slab)) {
3544 slab_out_of_memory(s, gfpflags, node);
3548 stat(s, ALLOC_SLAB);
3550 if (kmem_cache_debug(s)) {
3551 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3553 if (unlikely(!freelist))
3556 if (s->flags & SLAB_STORE_USER)
3557 set_track(s, freelist, TRACK_ALLOC, addr);
3563 * No other reference to the slab yet so we can
3564 * muck around with it freely without cmpxchg
3566 freelist = slab->freelist;
3567 slab->freelist = NULL;
3568 slab->inuse = slab->objects;
3571 inc_slabs_node(s, slab_nid(slab), slab->objects);
3573 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3575 * For !pfmemalloc_match() case we don't load freelist so that
3576 * we don't make further mismatched allocations easier.
3578 deactivate_slab(s, slab, get_freepointer(s, freelist));
3584 local_lock_irqsave(&s->cpu_slab->lock, flags);
3585 if (unlikely(c->slab)) {
3586 void *flush_freelist = c->freelist;
3587 struct slab *flush_slab = c->slab;
3591 c->tid = next_tid(c->tid);
3593 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3595 deactivate_slab(s, flush_slab, flush_freelist);
3597 stat(s, CPUSLAB_FLUSH);
3599 goto retry_load_slab;
3607 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3608 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3611 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3612 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3616 #ifdef CONFIG_PREEMPT_COUNT
3618 * We may have been preempted and rescheduled on a different
3619 * cpu before disabling preemption. Need to reload cpu area
3622 c = slub_get_cpu_ptr(s->cpu_slab);
3625 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3626 #ifdef CONFIG_PREEMPT_COUNT
3627 slub_put_cpu_ptr(s->cpu_slab);
3632 static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3633 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3635 struct kmem_cache_cpu *c;
3642 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3643 * enabled. We may switch back and forth between cpus while
3644 * reading from one cpu area. That does not matter as long
3645 * as we end up on the original cpu again when doing the cmpxchg.
3647 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3648 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3649 * the tid. If we are preempted and switched to another cpu between the
3650 * two reads, it's OK as the two are still associated with the same cpu
3651 * and cmpxchg later will validate the cpu.
3653 c = raw_cpu_ptr(s->cpu_slab);
3654 tid = READ_ONCE(c->tid);
3657 * Irqless object alloc/free algorithm used here depends on sequence
3658 * of fetching cpu_slab's data. tid should be fetched before anything
3659 * on c to guarantee that object and slab associated with previous tid
3660 * won't be used with current tid. If we fetch tid first, object and
3661 * slab could be one associated with next tid and our alloc/free
3662 * request will be failed. In this case, we will retry. So, no problem.
3667 * The transaction ids are globally unique per cpu and per operation on
3668 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3669 * occurs on the right processor and that there was no operation on the
3670 * linked list in between.
3673 object = c->freelist;
3676 if (!USE_LOCKLESS_FAST_PATH() ||
3677 unlikely(!object || !slab || !node_match(slab, node))) {
3678 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3680 void *next_object = get_freepointer_safe(s, object);
3683 * The cmpxchg will only match if there was no additional
3684 * operation and if we are on the right processor.
3686 * The cmpxchg does the following atomically (without lock
3688 * 1. Relocate first pointer to the current per cpu area.
3689 * 2. Verify that tid and freelist have not been changed
3690 * 3. If they were not changed replace tid and freelist
3692 * Since this is without lock semantics the protection is only
3693 * against code executing on this cpu *not* from access by
3696 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3697 note_cmpxchg_failure("slab_alloc", s, tid);
3700 prefetch_freepointer(s, next_object);
3701 stat(s, ALLOC_FASTPATH);
3706 #else /* CONFIG_SLUB_TINY */
3707 static void *__slab_alloc_node(struct kmem_cache *s,
3708 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3710 struct partial_context pc;
3714 pc.flags = gfpflags;
3715 pc.orig_size = orig_size;
3716 slab = get_partial(s, node, &pc);
3721 slab = new_slab(s, gfpflags, node);
3722 if (unlikely(!slab)) {
3723 slab_out_of_memory(s, gfpflags, node);
3727 object = alloc_single_from_new_slab(s, slab, orig_size);
3731 #endif /* CONFIG_SLUB_TINY */
3734 * If the object has been wiped upon free, make sure it's fully initialized by
3735 * zeroing out freelist pointer.
3737 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3740 if (unlikely(slab_want_init_on_free(s)) && obj)
3741 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3745 noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
3747 if (__should_failslab(s, gfpflags))
3751 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
3753 static __fastpath_inline
3754 struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
3755 struct list_lru *lru,
3756 struct obj_cgroup **objcgp,
3757 size_t size, gfp_t flags)
3759 flags &= gfp_allowed_mask;
3763 if (unlikely(should_failslab(s, flags)))
3766 if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags)))
3772 static __fastpath_inline
3773 void slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
3774 gfp_t flags, size_t size, void **p, bool init,
3775 unsigned int orig_size)
3777 unsigned int zero_size = s->object_size;
3778 bool kasan_init = init;
3780 gfp_t init_flags = flags & gfp_allowed_mask;
3783 * For kmalloc object, the allocated memory size(object_size) is likely
3784 * larger than the requested size(orig_size). If redzone check is
3785 * enabled for the extra space, don't zero it, as it will be redzoned
3786 * soon. The redzone operation for this extra space could be seen as a
3787 * replacement of current poisoning under certain debug option, and
3788 * won't break other sanity checks.
3790 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
3791 (s->flags & SLAB_KMALLOC))
3792 zero_size = orig_size;
3795 * When slub_debug is enabled, avoid memory initialization integrated
3796 * into KASAN and instead zero out the memory via the memset below with
3797 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
3798 * cause false-positive reports. This does not lead to a performance
3799 * penalty on production builds, as slub_debug is not intended to be
3802 if (__slub_debug_enabled())
3806 * As memory initialization might be integrated into KASAN,
3807 * kasan_slab_alloc and initialization memset must be
3808 * kept together to avoid discrepancies in behavior.
3810 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
3812 for (i = 0; i < size; i++) {
3813 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
3814 if (p[i] && init && (!kasan_init ||
3815 !kasan_has_integrated_init()))
3816 memset(p[i], 0, zero_size);
3817 kmemleak_alloc_recursive(p[i], s->object_size, 1,
3818 s->flags, init_flags);
3819 kmsan_slab_alloc(s, p[i], init_flags);
3822 memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
3826 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3827 * have the fastpath folded into their functions. So no function call
3828 * overhead for requests that can be satisfied on the fastpath.
3830 * The fastpath works by first checking if the lockless freelist can be used.
3831 * If not then __slab_alloc is called for slow processing.
3833 * Otherwise we can simply pick the next object from the lockless free list.
3835 static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3836 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3839 struct obj_cgroup *objcg = NULL;
3842 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3846 object = kfence_alloc(s, orig_size, gfpflags);
3847 if (unlikely(object))
3850 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3852 maybe_wipe_obj_freeptr(s, object);
3853 init = slab_want_init_on_alloc(gfpflags, s);
3857 * When init equals 'true', like for kzalloc() family, only
3858 * @orig_size bytes might be zeroed instead of s->object_size
3860 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3865 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3867 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
3870 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3874 EXPORT_SYMBOL(kmem_cache_alloc);
3876 void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3879 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
3882 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3886 EXPORT_SYMBOL(kmem_cache_alloc_lru);
3889 * kmem_cache_alloc_node - Allocate an object on the specified node
3890 * @s: The cache to allocate from.
3891 * @gfpflags: See kmalloc().
3892 * @node: node number of the target node.
3894 * Identical to kmem_cache_alloc but it will allocate memory on the given
3895 * node, which can improve the performance for cpu bound structures.
3897 * Fallback to other node is possible if __GFP_THISNODE is not set.
3899 * Return: pointer to the new object or %NULL in case of error
3901 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3903 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3905 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3909 EXPORT_SYMBOL(kmem_cache_alloc_node);
3912 * To avoid unnecessary overhead, we pass through large allocation requests
3913 * directly to the page allocator. We use __GFP_COMP, because we will need to
3914 * know the allocation order to free the pages properly in kfree.
3916 static void *__kmalloc_large_node(size_t size, gfp_t flags, int node)
3918 struct folio *folio;
3920 unsigned int order = get_order(size);
3922 if (unlikely(flags & GFP_SLAB_BUG_MASK))
3923 flags = kmalloc_fix_flags(flags);
3925 flags |= __GFP_COMP;
3926 folio = (struct folio *)alloc_pages_node(node, flags, order);
3928 ptr = folio_address(folio);
3929 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
3930 PAGE_SIZE << order);
3933 ptr = kasan_kmalloc_large(ptr, size, flags);
3934 /* As ptr might get tagged, call kmemleak hook after KASAN. */
3935 kmemleak_alloc(ptr, size, 1, flags);
3936 kmsan_kmalloc_large(ptr, size, flags);
3941 void *kmalloc_large(size_t size, gfp_t flags)
3943 void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE);
3945 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3946 flags, NUMA_NO_NODE);
3949 EXPORT_SYMBOL(kmalloc_large);
3951 void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3953 void *ret = __kmalloc_large_node(size, flags, node);
3955 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3959 EXPORT_SYMBOL(kmalloc_large_node);
3961 static __always_inline
3962 void *__do_kmalloc_node(size_t size, gfp_t flags, int node,
3963 unsigned long caller)
3965 struct kmem_cache *s;
3968 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3969 ret = __kmalloc_large_node(size, flags, node);
3970 trace_kmalloc(caller, ret, size,
3971 PAGE_SIZE << get_order(size), flags, node);
3975 if (unlikely(!size))
3976 return ZERO_SIZE_PTR;
3978 s = kmalloc_slab(size, flags, caller);
3980 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
3981 ret = kasan_kmalloc(s, ret, size, flags);
3982 trace_kmalloc(caller, ret, size, s->size, flags, node);
3986 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3988 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3990 EXPORT_SYMBOL(__kmalloc_node);
3992 void *__kmalloc(size_t size, gfp_t flags)
3994 return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_);
3996 EXPORT_SYMBOL(__kmalloc);
3998 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3999 int node, unsigned long caller)
4001 return __do_kmalloc_node(size, flags, node, caller);
4003 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4005 void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
4007 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
4010 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4012 ret = kasan_kmalloc(s, ret, size, gfpflags);
4015 EXPORT_SYMBOL(kmalloc_trace);
4017 void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags,
4018 int node, size_t size)
4020 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4022 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4024 ret = kasan_kmalloc(s, ret, size, gfpflags);
4027 EXPORT_SYMBOL(kmalloc_node_trace);
4029 static noinline void free_to_partial_list(
4030 struct kmem_cache *s, struct slab *slab,
4031 void *head, void *tail, int bulk_cnt,
4034 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4035 struct slab *slab_free = NULL;
4037 unsigned long flags;
4038 depot_stack_handle_t handle = 0;
4040 if (s->flags & SLAB_STORE_USER)
4041 handle = set_track_prepare();
4043 spin_lock_irqsave(&n->list_lock, flags);
4045 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4046 void *prior = slab->freelist;
4048 /* Perform the actual freeing while we still hold the locks */
4050 set_freepointer(s, tail, prior);
4051 slab->freelist = head;
4054 * If the slab is empty, and node's partial list is full,
4055 * it should be discarded anyway no matter it's on full or
4058 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4062 /* was on full list */
4063 remove_full(s, n, slab);
4065 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4066 stat(s, FREE_ADD_PARTIAL);
4068 } else if (slab_free) {
4069 remove_partial(n, slab);
4070 stat(s, FREE_REMOVE_PARTIAL);
4076 * Update the counters while still holding n->list_lock to
4077 * prevent spurious validation warnings
4079 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4082 spin_unlock_irqrestore(&n->list_lock, flags);
4086 free_slab(s, slab_free);
4091 * Slow path handling. This may still be called frequently since objects
4092 * have a longer lifetime than the cpu slabs in most processing loads.
4094 * So we still attempt to reduce cache line usage. Just take the slab
4095 * lock and free the item. If there is no additional partial slab
4096 * handling required then we can return immediately.
4098 static void __slab_free(struct kmem_cache *s, struct slab *slab,
4099 void *head, void *tail, int cnt,
4106 unsigned long counters;
4107 struct kmem_cache_node *n = NULL;
4108 unsigned long flags;
4109 bool on_node_partial;
4111 stat(s, FREE_SLOWPATH);
4113 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4114 free_to_partial_list(s, slab, head, tail, cnt, addr);
4120 spin_unlock_irqrestore(&n->list_lock, flags);
4123 prior = slab->freelist;
4124 counters = slab->counters;
4125 set_freepointer(s, tail, prior);
4126 new.counters = counters;
4127 was_frozen = new.frozen;
4129 if ((!new.inuse || !prior) && !was_frozen) {
4130 /* Needs to be taken off a list */
4131 if (!kmem_cache_has_cpu_partial(s) || prior) {
4133 n = get_node(s, slab_nid(slab));
4135 * Speculatively acquire the list_lock.
4136 * If the cmpxchg does not succeed then we may
4137 * drop the list_lock without any processing.
4139 * Otherwise the list_lock will synchronize with
4140 * other processors updating the list of slabs.
4142 spin_lock_irqsave(&n->list_lock, flags);
4144 on_node_partial = slab_test_node_partial(slab);
4148 } while (!slab_update_freelist(s, slab,
4155 if (likely(was_frozen)) {
4157 * The list lock was not taken therefore no list
4158 * activity can be necessary.
4160 stat(s, FREE_FROZEN);
4161 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
4163 * If we started with a full slab then put it onto the
4164 * per cpu partial list.
4166 put_cpu_partial(s, slab, 1);
4167 stat(s, CPU_PARTIAL_FREE);
4174 * This slab was partially empty but not on the per-node partial list,
4175 * in which case we shouldn't manipulate its list, just return.
4177 if (prior && !on_node_partial) {
4178 spin_unlock_irqrestore(&n->list_lock, flags);
4182 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4186 * Objects left in the slab. If it was not on the partial list before
4189 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4190 remove_full(s, n, slab);
4191 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4192 stat(s, FREE_ADD_PARTIAL);
4194 spin_unlock_irqrestore(&n->list_lock, flags);
4200 * Slab on the partial list.
4202 remove_partial(n, slab);
4203 stat(s, FREE_REMOVE_PARTIAL);
4205 /* Slab must be on the full list */
4206 remove_full(s, n, slab);
4209 spin_unlock_irqrestore(&n->list_lock, flags);
4211 discard_slab(s, slab);
4214 #ifndef CONFIG_SLUB_TINY
4216 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4217 * can perform fastpath freeing without additional function calls.
4219 * The fastpath is only possible if we are freeing to the current cpu slab
4220 * of this processor. This typically the case if we have just allocated
4223 * If fastpath is not possible then fall back to __slab_free where we deal
4224 * with all sorts of special processing.
4226 * Bulk free of a freelist with several objects (all pointing to the
4227 * same slab) possible by specifying head and tail ptr, plus objects
4228 * count (cnt). Bulk free indicated by tail pointer being set.
4230 static __always_inline void do_slab_free(struct kmem_cache *s,
4231 struct slab *slab, void *head, void *tail,
4232 int cnt, unsigned long addr)
4234 struct kmem_cache_cpu *c;
4240 * Determine the currently cpus per cpu slab.
4241 * The cpu may change afterward. However that does not matter since
4242 * data is retrieved via this pointer. If we are on the same cpu
4243 * during the cmpxchg then the free will succeed.
4245 c = raw_cpu_ptr(s->cpu_slab);
4246 tid = READ_ONCE(c->tid);
4248 /* Same with comment on barrier() in slab_alloc_node() */
4251 if (unlikely(slab != c->slab)) {
4252 __slab_free(s, slab, head, tail, cnt, addr);
4256 if (USE_LOCKLESS_FAST_PATH()) {
4257 freelist = READ_ONCE(c->freelist);
4259 set_freepointer(s, tail, freelist);
4261 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4262 note_cmpxchg_failure("slab_free", s, tid);
4266 /* Update the free list under the local lock */
4267 local_lock(&s->cpu_slab->lock);
4268 c = this_cpu_ptr(s->cpu_slab);
4269 if (unlikely(slab != c->slab)) {
4270 local_unlock(&s->cpu_slab->lock);
4274 freelist = c->freelist;
4276 set_freepointer(s, tail, freelist);
4278 c->tid = next_tid(tid);
4280 local_unlock(&s->cpu_slab->lock);
4282 stat_add(s, FREE_FASTPATH, cnt);
4284 #else /* CONFIG_SLUB_TINY */
4285 static void do_slab_free(struct kmem_cache *s,
4286 struct slab *slab, void *head, void *tail,
4287 int cnt, unsigned long addr)
4289 __slab_free(s, slab, head, tail, cnt, addr);
4291 #endif /* CONFIG_SLUB_TINY */
4293 static __fastpath_inline
4294 void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4297 memcg_slab_free_hook(s, slab, &object, 1);
4299 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s))))
4300 do_slab_free(s, slab, object, object, 1, addr);
4303 static __fastpath_inline
4304 void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4305 void *tail, void **p, int cnt, unsigned long addr)
4307 memcg_slab_free_hook(s, slab, p, cnt);
4309 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4310 * to remove objects, whose reuse must be delayed.
4312 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4313 do_slab_free(s, slab, head, tail, cnt, addr);
4316 #ifdef CONFIG_KASAN_GENERIC
4317 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4319 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4323 static inline struct kmem_cache *virt_to_cache(const void *obj)
4327 slab = virt_to_slab(obj);
4328 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4330 return slab->slab_cache;
4333 static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4335 struct kmem_cache *cachep;
4337 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4338 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4341 cachep = virt_to_cache(x);
4342 if (WARN(cachep && cachep != s,
4343 "%s: Wrong slab cache. %s but object is from %s\n",
4344 __func__, s->name, cachep->name))
4345 print_tracking(cachep, x);
4350 * kmem_cache_free - Deallocate an object
4351 * @s: The cache the allocation was from.
4352 * @x: The previously allocated object.
4354 * Free an object which was previously allocated from this
4357 void kmem_cache_free(struct kmem_cache *s, void *x)
4359 s = cache_from_obj(s, x);
4362 trace_kmem_cache_free(_RET_IP_, x, s);
4363 slab_free(s, virt_to_slab(x), x, _RET_IP_);
4365 EXPORT_SYMBOL(kmem_cache_free);
4367 static void free_large_kmalloc(struct folio *folio, void *object)
4369 unsigned int order = folio_order(folio);
4371 if (WARN_ON_ONCE(order == 0))
4372 pr_warn_once("object pointer: 0x%p\n", object);
4374 kmemleak_free(object);
4375 kasan_kfree_large(object);
4376 kmsan_kfree_large(object);
4378 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4379 -(PAGE_SIZE << order));
4384 * kfree - free previously allocated memory
4385 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4387 * If @object is NULL, no operation is performed.
4389 void kfree(const void *object)
4391 struct folio *folio;
4393 struct kmem_cache *s;
4394 void *x = (void *)object;
4396 trace_kfree(_RET_IP_, object);
4398 if (unlikely(ZERO_OR_NULL_PTR(object)))
4401 folio = virt_to_folio(object);
4402 if (unlikely(!folio_test_slab(folio))) {
4403 free_large_kmalloc(folio, (void *)object);
4407 slab = folio_slab(folio);
4408 s = slab->slab_cache;
4409 slab_free(s, slab, x, _RET_IP_);
4411 EXPORT_SYMBOL(kfree);
4413 struct detached_freelist {
4418 struct kmem_cache *s;
4422 * This function progressively scans the array with free objects (with
4423 * a limited look ahead) and extract objects belonging to the same
4424 * slab. It builds a detached freelist directly within the given
4425 * slab/objects. This can happen without any need for
4426 * synchronization, because the objects are owned by running process.
4427 * The freelist is build up as a single linked list in the objects.
4428 * The idea is, that this detached freelist can then be bulk
4429 * transferred to the real freelist(s), but only requiring a single
4430 * synchronization primitive. Look ahead in the array is limited due
4431 * to performance reasons.
4434 int build_detached_freelist(struct kmem_cache *s, size_t size,
4435 void **p, struct detached_freelist *df)
4439 struct folio *folio;
4443 folio = virt_to_folio(object);
4445 /* Handle kalloc'ed objects */
4446 if (unlikely(!folio_test_slab(folio))) {
4447 free_large_kmalloc(folio, object);
4451 /* Derive kmem_cache from object */
4452 df->slab = folio_slab(folio);
4453 df->s = df->slab->slab_cache;
4455 df->slab = folio_slab(folio);
4456 df->s = cache_from_obj(s, object); /* Support for memcg */
4459 /* Start new detached freelist */
4461 df->freelist = object;
4464 if (is_kfence_address(object))
4467 set_freepointer(df->s, object, NULL);
4472 /* df->slab is always set at this point */
4473 if (df->slab == virt_to_slab(object)) {
4474 /* Opportunity build freelist */
4475 set_freepointer(df->s, object, df->freelist);
4476 df->freelist = object;
4480 swap(p[size], p[same]);
4484 /* Limit look ahead search */
4493 * Internal bulk free of objects that were not initialised by the post alloc
4494 * hooks and thus should not be processed by the free hooks
4496 static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4502 struct detached_freelist df;
4504 size = build_detached_freelist(s, size, p, &df);
4508 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4510 } while (likely(size));
4513 /* Note that interrupts must be enabled when calling this function. */
4514 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4520 struct detached_freelist df;
4522 size = build_detached_freelist(s, size, p, &df);
4526 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
4528 } while (likely(size));
4530 EXPORT_SYMBOL(kmem_cache_free_bulk);
4532 #ifndef CONFIG_SLUB_TINY
4534 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4537 struct kmem_cache_cpu *c;
4538 unsigned long irqflags;
4542 * Drain objects in the per cpu slab, while disabling local
4543 * IRQs, which protects against PREEMPT and interrupts
4544 * handlers invoking normal fastpath.
4546 c = slub_get_cpu_ptr(s->cpu_slab);
4547 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4549 for (i = 0; i < size; i++) {
4550 void *object = kfence_alloc(s, s->object_size, flags);
4552 if (unlikely(object)) {
4557 object = c->freelist;
4558 if (unlikely(!object)) {
4560 * We may have removed an object from c->freelist using
4561 * the fastpath in the previous iteration; in that case,
4562 * c->tid has not been bumped yet.
4563 * Since ___slab_alloc() may reenable interrupts while
4564 * allocating memory, we should bump c->tid now.
4566 c->tid = next_tid(c->tid);
4568 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4571 * Invoking slow path likely have side-effect
4572 * of re-populating per CPU c->freelist
4574 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
4575 _RET_IP_, c, s->object_size);
4576 if (unlikely(!p[i]))
4579 c = this_cpu_ptr(s->cpu_slab);
4580 maybe_wipe_obj_freeptr(s, p[i]);
4582 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4584 continue; /* goto for-loop */
4586 c->freelist = get_freepointer(s, object);
4588 maybe_wipe_obj_freeptr(s, p[i]);
4589 stat(s, ALLOC_FASTPATH);
4591 c->tid = next_tid(c->tid);
4592 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4593 slub_put_cpu_ptr(s->cpu_slab);
4598 slub_put_cpu_ptr(s->cpu_slab);
4599 __kmem_cache_free_bulk(s, i, p);
4603 #else /* CONFIG_SLUB_TINY */
4604 static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4605 size_t size, void **p)
4609 for (i = 0; i < size; i++) {
4610 void *object = kfence_alloc(s, s->object_size, flags);
4612 if (unlikely(object)) {
4617 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4618 _RET_IP_, s->object_size);
4619 if (unlikely(!p[i]))
4622 maybe_wipe_obj_freeptr(s, p[i]);
4628 __kmem_cache_free_bulk(s, i, p);
4631 #endif /* CONFIG_SLUB_TINY */
4633 /* Note that interrupts must be enabled when calling this function. */
4634 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4638 struct obj_cgroup *objcg = NULL;
4643 /* memcg and kmem_cache debug support */
4644 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4648 i = __kmem_cache_alloc_bulk(s, flags, size, p);
4651 * memcg and kmem_cache debug support and memory initialization.
4652 * Done outside of the IRQ disabled fastpath loop.
4654 if (likely(i != 0)) {
4655 slab_post_alloc_hook(s, objcg, flags, size, p,
4656 slab_want_init_on_alloc(flags, s), s->object_size);
4658 memcg_slab_alloc_error_hook(s, size, objcg);
4663 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4667 * Object placement in a slab is made very easy because we always start at
4668 * offset 0. If we tune the size of the object to the alignment then we can
4669 * get the required alignment by putting one properly sized object after
4672 * Notice that the allocation order determines the sizes of the per cpu
4673 * caches. Each processor has always one slab available for allocations.
4674 * Increasing the allocation order reduces the number of times that slabs
4675 * must be moved on and off the partial lists and is therefore a factor in
4680 * Minimum / Maximum order of slab pages. This influences locking overhead
4681 * and slab fragmentation. A higher order reduces the number of partial slabs
4682 * and increases the number of allocations possible without having to
4683 * take the list_lock.
4685 static unsigned int slub_min_order;
4686 static unsigned int slub_max_order =
4687 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4688 static unsigned int slub_min_objects;
4691 * Calculate the order of allocation given an slab object size.
4693 * The order of allocation has significant impact on performance and other
4694 * system components. Generally order 0 allocations should be preferred since
4695 * order 0 does not cause fragmentation in the page allocator. Larger objects
4696 * be problematic to put into order 0 slabs because there may be too much
4697 * unused space left. We go to a higher order if more than 1/16th of the slab
4700 * In order to reach satisfactory performance we must ensure that a minimum
4701 * number of objects is in one slab. Otherwise we may generate too much
4702 * activity on the partial lists which requires taking the list_lock. This is
4703 * less a concern for large slabs though which are rarely used.
4705 * slub_max_order specifies the order where we begin to stop considering the
4706 * number of objects in a slab as critical. If we reach slub_max_order then
4707 * we try to keep the page order as low as possible. So we accept more waste
4708 * of space in favor of a small page order.
4710 * Higher order allocations also allow the placement of more objects in a
4711 * slab and thereby reduce object handling overhead. If the user has
4712 * requested a higher minimum order then we start with that one instead of
4713 * the smallest order which will fit the object.
4715 static inline unsigned int calc_slab_order(unsigned int size,
4716 unsigned int min_order, unsigned int max_order,
4717 unsigned int fract_leftover)
4721 for (order = min_order; order <= max_order; order++) {
4723 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4726 rem = slab_size % size;
4728 if (rem <= slab_size / fract_leftover)
4735 static inline int calculate_order(unsigned int size)
4738 unsigned int min_objects;
4739 unsigned int max_objects;
4740 unsigned int min_order;
4742 min_objects = slub_min_objects;
4745 * Some architectures will only update present cpus when
4746 * onlining them, so don't trust the number if it's just 1. But
4747 * we also don't want to use nr_cpu_ids always, as on some other
4748 * architectures, there can be many possible cpus, but never
4749 * onlined. Here we compromise between trying to avoid too high
4750 * order on systems that appear larger than they are, and too
4751 * low order on systems that appear smaller than they are.
4753 unsigned int nr_cpus = num_present_cpus();
4755 nr_cpus = nr_cpu_ids;
4756 min_objects = 4 * (fls(nr_cpus) + 1);
4758 /* min_objects can't be 0 because get_order(0) is undefined */
4759 max_objects = max(order_objects(slub_max_order, size), 1U);
4760 min_objects = min(min_objects, max_objects);
4762 min_order = max_t(unsigned int, slub_min_order,
4763 get_order(min_objects * size));
4764 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4765 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4768 * Attempt to find best configuration for a slab. This works by first
4769 * attempting to generate a layout with the best possible configuration
4770 * and backing off gradually.
4772 * We start with accepting at most 1/16 waste and try to find the
4773 * smallest order from min_objects-derived/slub_min_order up to
4774 * slub_max_order that will satisfy the constraint. Note that increasing
4775 * the order can only result in same or less fractional waste, not more.
4777 * If that fails, we increase the acceptable fraction of waste and try
4778 * again. The last iteration with fraction of 1/2 would effectively
4779 * accept any waste and give us the order determined by min_objects, as
4780 * long as at least single object fits within slub_max_order.
4782 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
4783 order = calc_slab_order(size, min_order, slub_max_order,
4785 if (order <= slub_max_order)
4790 * Doh this slab cannot be placed using slub_max_order.
4792 order = get_order(size);
4793 if (order <= MAX_PAGE_ORDER)
4799 init_kmem_cache_node(struct kmem_cache_node *n)
4802 spin_lock_init(&n->list_lock);
4803 INIT_LIST_HEAD(&n->partial);
4804 #ifdef CONFIG_SLUB_DEBUG
4805 atomic_long_set(&n->nr_slabs, 0);
4806 atomic_long_set(&n->total_objects, 0);
4807 INIT_LIST_HEAD(&n->full);
4811 #ifndef CONFIG_SLUB_TINY
4812 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4814 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4815 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4816 sizeof(struct kmem_cache_cpu));
4819 * Must align to double word boundary for the double cmpxchg
4820 * instructions to work; see __pcpu_double_call_return_bool().
4822 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4823 2 * sizeof(void *));
4828 init_kmem_cache_cpus(s);
4833 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4837 #endif /* CONFIG_SLUB_TINY */
4839 static struct kmem_cache *kmem_cache_node;
4842 * No kmalloc_node yet so do it by hand. We know that this is the first
4843 * slab on the node for this slabcache. There are no concurrent accesses
4846 * Note that this function only works on the kmem_cache_node
4847 * when allocating for the kmem_cache_node. This is used for bootstrapping
4848 * memory on a fresh node that has no slab structures yet.
4850 static void early_kmem_cache_node_alloc(int node)
4853 struct kmem_cache_node *n;
4855 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4857 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4860 inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4861 if (slab_nid(slab) != node) {
4862 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4863 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4868 #ifdef CONFIG_SLUB_DEBUG
4869 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4870 init_tracking(kmem_cache_node, n);
4872 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4873 slab->freelist = get_freepointer(kmem_cache_node, n);
4875 kmem_cache_node->node[node] = n;
4876 init_kmem_cache_node(n);
4877 inc_slabs_node(kmem_cache_node, node, slab->objects);
4880 * No locks need to be taken here as it has just been
4881 * initialized and there is no concurrent access.
4883 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4886 static void free_kmem_cache_nodes(struct kmem_cache *s)
4889 struct kmem_cache_node *n;
4891 for_each_kmem_cache_node(s, node, n) {
4892 s->node[node] = NULL;
4893 kmem_cache_free(kmem_cache_node, n);
4897 void __kmem_cache_release(struct kmem_cache *s)
4899 cache_random_seq_destroy(s);
4900 #ifndef CONFIG_SLUB_TINY
4901 free_percpu(s->cpu_slab);
4903 free_kmem_cache_nodes(s);
4906 static int init_kmem_cache_nodes(struct kmem_cache *s)
4910 for_each_node_mask(node, slab_nodes) {
4911 struct kmem_cache_node *n;
4913 if (slab_state == DOWN) {
4914 early_kmem_cache_node_alloc(node);
4917 n = kmem_cache_alloc_node(kmem_cache_node,
4921 free_kmem_cache_nodes(s);
4925 init_kmem_cache_node(n);
4931 static void set_cpu_partial(struct kmem_cache *s)
4933 #ifdef CONFIG_SLUB_CPU_PARTIAL
4934 unsigned int nr_objects;
4937 * cpu_partial determined the maximum number of objects kept in the
4938 * per cpu partial lists of a processor.
4940 * Per cpu partial lists mainly contain slabs that just have one
4941 * object freed. If they are used for allocation then they can be
4942 * filled up again with minimal effort. The slab will never hit the
4943 * per node partial lists and therefore no locking will be required.
4945 * For backwards compatibility reasons, this is determined as number
4946 * of objects, even though we now limit maximum number of pages, see
4947 * slub_set_cpu_partial()
4949 if (!kmem_cache_has_cpu_partial(s))
4951 else if (s->size >= PAGE_SIZE)
4953 else if (s->size >= 1024)
4955 else if (s->size >= 256)
4960 slub_set_cpu_partial(s, nr_objects);
4965 * calculate_sizes() determines the order and the distribution of data within
4968 static int calculate_sizes(struct kmem_cache *s)
4970 slab_flags_t flags = s->flags;
4971 unsigned int size = s->object_size;
4975 * Round up object size to the next word boundary. We can only
4976 * place the free pointer at word boundaries and this determines
4977 * the possible location of the free pointer.
4979 size = ALIGN(size, sizeof(void *));
4981 #ifdef CONFIG_SLUB_DEBUG
4983 * Determine if we can poison the object itself. If the user of
4984 * the slab may touch the object after free or before allocation
4985 * then we should never poison the object itself.
4987 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4989 s->flags |= __OBJECT_POISON;
4991 s->flags &= ~__OBJECT_POISON;
4995 * If we are Redzoning then check if there is some space between the
4996 * end of the object and the free pointer. If not then add an
4997 * additional word to have some bytes to store Redzone information.
4999 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
5000 size += sizeof(void *);
5004 * With that we have determined the number of bytes in actual use
5005 * by the object and redzoning.
5009 if (slub_debug_orig_size(s) ||
5010 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
5011 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
5014 * Relocate free pointer after the object if it is not
5015 * permitted to overwrite the first word of the object on
5018 * This is the case if we do RCU, have a constructor or
5019 * destructor, are poisoning the objects, or are
5020 * redzoning an object smaller than sizeof(void *).
5022 * The assumption that s->offset >= s->inuse means free
5023 * pointer is outside of the object is used in the
5024 * freeptr_outside_object() function. If that is no
5025 * longer true, the function needs to be modified.
5028 size += sizeof(void *);
5031 * Store freelist pointer near middle of object to keep
5032 * it away from the edges of the object to avoid small
5033 * sized over/underflows from neighboring allocations.
5035 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5038 #ifdef CONFIG_SLUB_DEBUG
5039 if (flags & SLAB_STORE_USER) {
5041 * Need to store information about allocs and frees after
5044 size += 2 * sizeof(struct track);
5046 /* Save the original kmalloc request size */
5047 if (flags & SLAB_KMALLOC)
5048 size += sizeof(unsigned int);
5052 kasan_cache_create(s, &size, &s->flags);
5053 #ifdef CONFIG_SLUB_DEBUG
5054 if (flags & SLAB_RED_ZONE) {
5056 * Add some empty padding so that we can catch
5057 * overwrites from earlier objects rather than let
5058 * tracking information or the free pointer be
5059 * corrupted if a user writes before the start
5062 size += sizeof(void *);
5064 s->red_left_pad = sizeof(void *);
5065 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5066 size += s->red_left_pad;
5071 * SLUB stores one object immediately after another beginning from
5072 * offset 0. In order to align the objects we have to simply size
5073 * each object to conform to the alignment.
5075 size = ALIGN(size, s->align);
5077 s->reciprocal_size = reciprocal_value(size);
5078 order = calculate_order(size);
5085 s->allocflags |= __GFP_COMP;
5087 if (s->flags & SLAB_CACHE_DMA)
5088 s->allocflags |= GFP_DMA;
5090 if (s->flags & SLAB_CACHE_DMA32)
5091 s->allocflags |= GFP_DMA32;
5093 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5094 s->allocflags |= __GFP_RECLAIMABLE;
5097 * Determine the number of objects per slab
5099 s->oo = oo_make(order, size);
5100 s->min = oo_make(get_order(size), size);
5102 return !!oo_objects(s->oo);
5105 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
5107 s->flags = kmem_cache_flags(s->size, flags, s->name);
5108 #ifdef CONFIG_SLAB_FREELIST_HARDENED
5109 s->random = get_random_long();
5112 if (!calculate_sizes(s))
5114 if (disable_higher_order_debug) {
5116 * Disable debugging flags that store metadata if the min slab
5119 if (get_order(s->size) > get_order(s->object_size)) {
5120 s->flags &= ~DEBUG_METADATA_FLAGS;
5122 if (!calculate_sizes(s))
5127 #ifdef system_has_freelist_aba
5128 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
5129 /* Enable fast mode */
5130 s->flags |= __CMPXCHG_DOUBLE;
5135 * The larger the object size is, the more slabs we want on the partial
5136 * list to avoid pounding the page allocator excessively.
5138 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
5139 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
5144 s->remote_node_defrag_ratio = 1000;
5147 /* Initialize the pre-computed randomized freelist if slab is up */
5148 if (slab_state >= UP) {
5149 if (init_cache_random_seq(s))
5153 if (!init_kmem_cache_nodes(s))
5156 if (alloc_kmem_cache_cpus(s))
5160 __kmem_cache_release(s);
5164 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5167 #ifdef CONFIG_SLUB_DEBUG
5168 void *addr = slab_address(slab);
5171 slab_err(s, slab, text, s->name);
5173 spin_lock(&object_map_lock);
5174 __fill_map(object_map, s, slab);
5176 for_each_object(p, s, addr, slab->objects) {
5178 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5179 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5180 print_tracking(s, p);
5183 spin_unlock(&object_map_lock);
5188 * Attempt to free all partial slabs on a node.
5189 * This is called from __kmem_cache_shutdown(). We must take list_lock
5190 * because sysfs file might still access partial list after the shutdowning.
5192 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5195 struct slab *slab, *h;
5197 BUG_ON(irqs_disabled());
5198 spin_lock_irq(&n->list_lock);
5199 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5201 remove_partial(n, slab);
5202 list_add(&slab->slab_list, &discard);
5204 list_slab_objects(s, slab,
5205 "Objects remaining in %s on __kmem_cache_shutdown()");
5208 spin_unlock_irq(&n->list_lock);
5210 list_for_each_entry_safe(slab, h, &discard, slab_list)
5211 discard_slab(s, slab);
5214 bool __kmem_cache_empty(struct kmem_cache *s)
5217 struct kmem_cache_node *n;
5219 for_each_kmem_cache_node(s, node, n)
5220 if (n->nr_partial || node_nr_slabs(n))
5226 * Release all resources used by a slab cache.
5228 int __kmem_cache_shutdown(struct kmem_cache *s)
5231 struct kmem_cache_node *n;
5233 flush_all_cpus_locked(s);
5234 /* Attempt to free all objects */
5235 for_each_kmem_cache_node(s, node, n) {
5237 if (n->nr_partial || node_nr_slabs(n))
5243 #ifdef CONFIG_PRINTK
5244 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5247 int __maybe_unused i;
5251 struct kmem_cache *s = slab->slab_cache;
5252 struct track __maybe_unused *trackp;
5254 kpp->kp_ptr = object;
5255 kpp->kp_slab = slab;
5256 kpp->kp_slab_cache = s;
5257 base = slab_address(slab);
5258 objp0 = kasan_reset_tag(object);
5259 #ifdef CONFIG_SLUB_DEBUG
5260 objp = restore_red_left(s, objp0);
5264 objnr = obj_to_index(s, slab, objp);
5265 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5266 objp = base + s->size * objnr;
5267 kpp->kp_objp = objp;
5268 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5269 || (objp - base) % s->size) ||
5270 !(s->flags & SLAB_STORE_USER))
5272 #ifdef CONFIG_SLUB_DEBUG
5273 objp = fixup_red_left(s, objp);
5274 trackp = get_track(s, objp, TRACK_ALLOC);
5275 kpp->kp_ret = (void *)trackp->addr;
5276 #ifdef CONFIG_STACKDEPOT
5278 depot_stack_handle_t handle;
5279 unsigned long *entries;
5280 unsigned int nr_entries;
5282 handle = READ_ONCE(trackp->handle);
5284 nr_entries = stack_depot_fetch(handle, &entries);
5285 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5286 kpp->kp_stack[i] = (void *)entries[i];
5289 trackp = get_track(s, objp, TRACK_FREE);
5290 handle = READ_ONCE(trackp->handle);
5292 nr_entries = stack_depot_fetch(handle, &entries);
5293 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5294 kpp->kp_free_stack[i] = (void *)entries[i];
5302 /********************************************************************
5304 *******************************************************************/
5306 static int __init setup_slub_min_order(char *str)
5308 get_option(&str, (int *)&slub_min_order);
5310 if (slub_min_order > slub_max_order)
5311 slub_max_order = slub_min_order;
5316 __setup("slub_min_order=", setup_slub_min_order);
5318 static int __init setup_slub_max_order(char *str)
5320 get_option(&str, (int *)&slub_max_order);
5321 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5323 if (slub_min_order > slub_max_order)
5324 slub_min_order = slub_max_order;
5329 __setup("slub_max_order=", setup_slub_max_order);
5331 static int __init setup_slub_min_objects(char *str)
5333 get_option(&str, (int *)&slub_min_objects);
5338 __setup("slub_min_objects=", setup_slub_min_objects);
5340 #ifdef CONFIG_HARDENED_USERCOPY
5342 * Rejects incorrectly sized objects and objects that are to be copied
5343 * to/from userspace but do not fall entirely within the containing slab
5344 * cache's usercopy region.
5346 * Returns NULL if check passes, otherwise const char * to name of cache
5347 * to indicate an error.
5349 void __check_heap_object(const void *ptr, unsigned long n,
5350 const struct slab *slab, bool to_user)
5352 struct kmem_cache *s;
5353 unsigned int offset;
5354 bool is_kfence = is_kfence_address(ptr);
5356 ptr = kasan_reset_tag(ptr);
5358 /* Find object and usable object size. */
5359 s = slab->slab_cache;
5361 /* Reject impossible pointers. */
5362 if (ptr < slab_address(slab))
5363 usercopy_abort("SLUB object not in SLUB page?!", NULL,
5366 /* Find offset within object. */
5368 offset = ptr - kfence_object_start(ptr);
5370 offset = (ptr - slab_address(slab)) % s->size;
5372 /* Adjust for redzone and reject if within the redzone. */
5373 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5374 if (offset < s->red_left_pad)
5375 usercopy_abort("SLUB object in left red zone",
5376 s->name, to_user, offset, n);
5377 offset -= s->red_left_pad;
5380 /* Allow address range falling entirely within usercopy region. */
5381 if (offset >= s->useroffset &&
5382 offset - s->useroffset <= s->usersize &&
5383 n <= s->useroffset - offset + s->usersize)
5386 usercopy_abort("SLUB object", s->name, to_user, offset, n);
5388 #endif /* CONFIG_HARDENED_USERCOPY */
5390 #define SHRINK_PROMOTE_MAX 32
5393 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5394 * up most to the head of the partial lists. New allocations will then
5395 * fill those up and thus they can be removed from the partial lists.
5397 * The slabs with the least items are placed last. This results in them
5398 * being allocated from last increasing the chance that the last objects
5399 * are freed in them.
5401 static int __kmem_cache_do_shrink(struct kmem_cache *s)
5405 struct kmem_cache_node *n;
5408 struct list_head discard;
5409 struct list_head promote[SHRINK_PROMOTE_MAX];
5410 unsigned long flags;
5413 for_each_kmem_cache_node(s, node, n) {
5414 INIT_LIST_HEAD(&discard);
5415 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5416 INIT_LIST_HEAD(promote + i);
5418 spin_lock_irqsave(&n->list_lock, flags);
5421 * Build lists of slabs to discard or promote.
5423 * Note that concurrent frees may occur while we hold the
5424 * list_lock. slab->inuse here is the upper limit.
5426 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5427 int free = slab->objects - slab->inuse;
5429 /* Do not reread slab->inuse */
5432 /* We do not keep full slabs on the list */
5435 if (free == slab->objects) {
5436 list_move(&slab->slab_list, &discard);
5437 slab_clear_node_partial(slab);
5439 dec_slabs_node(s, node, slab->objects);
5440 } else if (free <= SHRINK_PROMOTE_MAX)
5441 list_move(&slab->slab_list, promote + free - 1);
5445 * Promote the slabs filled up most to the head of the
5448 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5449 list_splice(promote + i, &n->partial);
5451 spin_unlock_irqrestore(&n->list_lock, flags);
5453 /* Release empty slabs */
5454 list_for_each_entry_safe(slab, t, &discard, slab_list)
5457 if (node_nr_slabs(n))
5464 int __kmem_cache_shrink(struct kmem_cache *s)
5467 return __kmem_cache_do_shrink(s);
5470 static int slab_mem_going_offline_callback(void *arg)
5472 struct kmem_cache *s;
5474 mutex_lock(&slab_mutex);
5475 list_for_each_entry(s, &slab_caches, list) {
5476 flush_all_cpus_locked(s);
5477 __kmem_cache_do_shrink(s);
5479 mutex_unlock(&slab_mutex);
5484 static void slab_mem_offline_callback(void *arg)
5486 struct memory_notify *marg = arg;
5489 offline_node = marg->status_change_nid_normal;
5492 * If the node still has available memory. we need kmem_cache_node
5495 if (offline_node < 0)
5498 mutex_lock(&slab_mutex);
5499 node_clear(offline_node, slab_nodes);
5501 * We no longer free kmem_cache_node structures here, as it would be
5502 * racy with all get_node() users, and infeasible to protect them with
5505 mutex_unlock(&slab_mutex);
5508 static int slab_mem_going_online_callback(void *arg)
5510 struct kmem_cache_node *n;
5511 struct kmem_cache *s;
5512 struct memory_notify *marg = arg;
5513 int nid = marg->status_change_nid_normal;
5517 * If the node's memory is already available, then kmem_cache_node is
5518 * already created. Nothing to do.
5524 * We are bringing a node online. No memory is available yet. We must
5525 * allocate a kmem_cache_node structure in order to bring the node
5528 mutex_lock(&slab_mutex);
5529 list_for_each_entry(s, &slab_caches, list) {
5531 * The structure may already exist if the node was previously
5532 * onlined and offlined.
5534 if (get_node(s, nid))
5537 * XXX: kmem_cache_alloc_node will fallback to other nodes
5538 * since memory is not yet available from the node that
5541 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5546 init_kmem_cache_node(n);
5550 * Any cache created after this point will also have kmem_cache_node
5551 * initialized for the new node.
5553 node_set(nid, slab_nodes);
5555 mutex_unlock(&slab_mutex);
5559 static int slab_memory_callback(struct notifier_block *self,
5560 unsigned long action, void *arg)
5565 case MEM_GOING_ONLINE:
5566 ret = slab_mem_going_online_callback(arg);
5568 case MEM_GOING_OFFLINE:
5569 ret = slab_mem_going_offline_callback(arg);
5572 case MEM_CANCEL_ONLINE:
5573 slab_mem_offline_callback(arg);
5576 case MEM_CANCEL_OFFLINE:
5580 ret = notifier_from_errno(ret);
5586 /********************************************************************
5587 * Basic setup of slabs
5588 *******************************************************************/
5591 * Used for early kmem_cache structures that were allocated using
5592 * the page allocator. Allocate them properly then fix up the pointers
5593 * that may be pointing to the wrong kmem_cache structure.
5596 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
5599 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
5600 struct kmem_cache_node *n;
5602 memcpy(s, static_cache, kmem_cache->object_size);
5605 * This runs very early, and only the boot processor is supposed to be
5606 * up. Even if it weren't true, IRQs are not up so we couldn't fire
5609 __flush_cpu_slab(s, smp_processor_id());
5610 for_each_kmem_cache_node(s, node, n) {
5613 list_for_each_entry(p, &n->partial, slab_list)
5616 #ifdef CONFIG_SLUB_DEBUG
5617 list_for_each_entry(p, &n->full, slab_list)
5621 list_add(&s->list, &slab_caches);
5625 void __init kmem_cache_init(void)
5627 static __initdata struct kmem_cache boot_kmem_cache,
5628 boot_kmem_cache_node;
5631 if (debug_guardpage_minorder())
5634 /* Print slub debugging pointers without hashing */
5635 if (__slub_debug_enabled())
5636 no_hash_pointers_enable(NULL);
5638 kmem_cache_node = &boot_kmem_cache_node;
5639 kmem_cache = &boot_kmem_cache;
5642 * Initialize the nodemask for which we will allocate per node
5643 * structures. Here we don't need taking slab_mutex yet.
5645 for_each_node_state(node, N_NORMAL_MEMORY)
5646 node_set(node, slab_nodes);
5648 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5649 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5651 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5653 /* Able to allocate the per node structures */
5654 slab_state = PARTIAL;
5656 create_boot_cache(kmem_cache, "kmem_cache",
5657 offsetof(struct kmem_cache, node) +
5658 nr_node_ids * sizeof(struct kmem_cache_node *),
5659 SLAB_HWCACHE_ALIGN, 0, 0);
5661 kmem_cache = bootstrap(&boot_kmem_cache);
5662 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5664 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5665 setup_kmalloc_cache_index_table();
5666 create_kmalloc_caches(0);
5668 /* Setup random freelists for each cache */
5669 init_freelist_randomization();
5671 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5674 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5676 slub_min_order, slub_max_order, slub_min_objects,
5677 nr_cpu_ids, nr_node_ids);
5680 void __init kmem_cache_init_late(void)
5682 #ifndef CONFIG_SLUB_TINY
5683 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5689 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5690 slab_flags_t flags, void (*ctor)(void *))
5692 struct kmem_cache *s;
5694 s = find_mergeable(size, align, flags, name, ctor);
5696 if (sysfs_slab_alias(s, name))
5702 * Adjust the object sizes so that we clear
5703 * the complete object on kzalloc.
5705 s->object_size = max(s->object_size, size);
5706 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5712 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5716 err = kmem_cache_open(s, flags);
5720 /* Mutex is not taken during early boot */
5721 if (slab_state <= UP)
5724 err = sysfs_slab_add(s);
5726 __kmem_cache_release(s);
5730 if (s->flags & SLAB_STORE_USER)
5731 debugfs_slab_add(s);
5736 #ifdef SLAB_SUPPORTS_SYSFS
5737 static int count_inuse(struct slab *slab)
5742 static int count_total(struct slab *slab)
5744 return slab->objects;
5748 #ifdef CONFIG_SLUB_DEBUG
5749 static void validate_slab(struct kmem_cache *s, struct slab *slab,
5750 unsigned long *obj_map)
5753 void *addr = slab_address(slab);
5755 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5758 /* Now we know that a valid freelist exists */
5759 __fill_map(obj_map, s, slab);
5760 for_each_object(p, s, addr, slab->objects) {
5761 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5762 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5764 if (!check_object(s, slab, p, val))
5769 static int validate_slab_node(struct kmem_cache *s,
5770 struct kmem_cache_node *n, unsigned long *obj_map)
5772 unsigned long count = 0;
5774 unsigned long flags;
5776 spin_lock_irqsave(&n->list_lock, flags);
5778 list_for_each_entry(slab, &n->partial, slab_list) {
5779 validate_slab(s, slab, obj_map);
5782 if (count != n->nr_partial) {
5783 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5784 s->name, count, n->nr_partial);
5785 slab_add_kunit_errors();
5788 if (!(s->flags & SLAB_STORE_USER))
5791 list_for_each_entry(slab, &n->full, slab_list) {
5792 validate_slab(s, slab, obj_map);
5795 if (count != node_nr_slabs(n)) {
5796 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5797 s->name, count, node_nr_slabs(n));
5798 slab_add_kunit_errors();
5802 spin_unlock_irqrestore(&n->list_lock, flags);
5806 long validate_slab_cache(struct kmem_cache *s)
5809 unsigned long count = 0;
5810 struct kmem_cache_node *n;
5811 unsigned long *obj_map;
5813 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5818 for_each_kmem_cache_node(s, node, n)
5819 count += validate_slab_node(s, n, obj_map);
5821 bitmap_free(obj_map);
5825 EXPORT_SYMBOL(validate_slab_cache);
5827 #ifdef CONFIG_DEBUG_FS
5829 * Generate lists of code addresses where slabcache objects are allocated
5834 depot_stack_handle_t handle;
5835 unsigned long count;
5837 unsigned long waste;
5843 DECLARE_BITMAP(cpus, NR_CPUS);
5849 unsigned long count;
5850 struct location *loc;
5854 static struct dentry *slab_debugfs_root;
5856 static void free_loc_track(struct loc_track *t)
5859 free_pages((unsigned long)t->loc,
5860 get_order(sizeof(struct location) * t->max));
5863 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5868 order = get_order(sizeof(struct location) * max);
5870 l = (void *)__get_free_pages(flags, order);
5875 memcpy(l, t->loc, sizeof(struct location) * t->count);
5883 static int add_location(struct loc_track *t, struct kmem_cache *s,
5884 const struct track *track,
5885 unsigned int orig_size)
5887 long start, end, pos;
5889 unsigned long caddr, chandle, cwaste;
5890 unsigned long age = jiffies - track->when;
5891 depot_stack_handle_t handle = 0;
5892 unsigned int waste = s->object_size - orig_size;
5894 #ifdef CONFIG_STACKDEPOT
5895 handle = READ_ONCE(track->handle);
5901 pos = start + (end - start + 1) / 2;
5904 * There is nothing at "end". If we end up there
5905 * we need to add something to before end.
5912 chandle = l->handle;
5914 if ((track->addr == caddr) && (handle == chandle) &&
5915 (waste == cwaste)) {
5920 if (age < l->min_time)
5922 if (age > l->max_time)
5925 if (track->pid < l->min_pid)
5926 l->min_pid = track->pid;
5927 if (track->pid > l->max_pid)
5928 l->max_pid = track->pid;
5930 cpumask_set_cpu(track->cpu,
5931 to_cpumask(l->cpus));
5933 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5937 if (track->addr < caddr)
5939 else if (track->addr == caddr && handle < chandle)
5941 else if (track->addr == caddr && handle == chandle &&
5949 * Not found. Insert new tracking element.
5951 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5957 (t->count - pos) * sizeof(struct location));
5960 l->addr = track->addr;
5964 l->min_pid = track->pid;
5965 l->max_pid = track->pid;
5968 cpumask_clear(to_cpumask(l->cpus));
5969 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5970 nodes_clear(l->nodes);
5971 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5975 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5976 struct slab *slab, enum track_item alloc,
5977 unsigned long *obj_map)
5979 void *addr = slab_address(slab);
5980 bool is_alloc = (alloc == TRACK_ALLOC);
5983 __fill_map(obj_map, s, slab);
5985 for_each_object(p, s, addr, slab->objects)
5986 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5987 add_location(t, s, get_track(s, p, alloc),
5988 is_alloc ? get_orig_size(s, p) :
5991 #endif /* CONFIG_DEBUG_FS */
5992 #endif /* CONFIG_SLUB_DEBUG */
5994 #ifdef SLAB_SUPPORTS_SYSFS
5995 enum slab_stat_type {
5996 SL_ALL, /* All slabs */
5997 SL_PARTIAL, /* Only partially allocated slabs */
5998 SL_CPU, /* Only slabs used for cpu caches */
5999 SL_OBJECTS, /* Determine allocated objects not slabs */
6000 SL_TOTAL /* Determine object capacity not slabs */
6003 #define SO_ALL (1 << SL_ALL)
6004 #define SO_PARTIAL (1 << SL_PARTIAL)
6005 #define SO_CPU (1 << SL_CPU)
6006 #define SO_OBJECTS (1 << SL_OBJECTS)
6007 #define SO_TOTAL (1 << SL_TOTAL)
6009 static ssize_t show_slab_objects(struct kmem_cache *s,
6010 char *buf, unsigned long flags)
6012 unsigned long total = 0;
6015 unsigned long *nodes;
6018 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6022 if (flags & SO_CPU) {
6025 for_each_possible_cpu(cpu) {
6026 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6031 slab = READ_ONCE(c->slab);
6035 node = slab_nid(slab);
6036 if (flags & SO_TOTAL)
6038 else if (flags & SO_OBJECTS)
6046 #ifdef CONFIG_SLUB_CPU_PARTIAL
6047 slab = slub_percpu_partial_read_once(c);
6049 node = slab_nid(slab);
6050 if (flags & SO_TOTAL)
6052 else if (flags & SO_OBJECTS)
6064 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6065 * already held which will conflict with an existing lock order:
6067 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6069 * We don't really need mem_hotplug_lock (to hold off
6070 * slab_mem_going_offline_callback) here because slab's memory hot
6071 * unplug code doesn't destroy the kmem_cache->node[] data.
6074 #ifdef CONFIG_SLUB_DEBUG
6075 if (flags & SO_ALL) {
6076 struct kmem_cache_node *n;
6078 for_each_kmem_cache_node(s, node, n) {
6080 if (flags & SO_TOTAL)
6081 x = node_nr_objs(n);
6082 else if (flags & SO_OBJECTS)
6083 x = node_nr_objs(n) - count_partial(n, count_free);
6085 x = node_nr_slabs(n);
6092 if (flags & SO_PARTIAL) {
6093 struct kmem_cache_node *n;
6095 for_each_kmem_cache_node(s, node, n) {
6096 if (flags & SO_TOTAL)
6097 x = count_partial(n, count_total);
6098 else if (flags & SO_OBJECTS)
6099 x = count_partial(n, count_inuse);
6107 len += sysfs_emit_at(buf, len, "%lu", total);
6109 for (node = 0; node < nr_node_ids; node++) {
6111 len += sysfs_emit_at(buf, len, " N%d=%lu",
6115 len += sysfs_emit_at(buf, len, "\n");
6121 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6122 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
6124 struct slab_attribute {
6125 struct attribute attr;
6126 ssize_t (*show)(struct kmem_cache *s, char *buf);
6127 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6130 #define SLAB_ATTR_RO(_name) \
6131 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6133 #define SLAB_ATTR(_name) \
6134 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6136 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6138 return sysfs_emit(buf, "%u\n", s->size);
6140 SLAB_ATTR_RO(slab_size);
6142 static ssize_t align_show(struct kmem_cache *s, char *buf)
6144 return sysfs_emit(buf, "%u\n", s->align);
6146 SLAB_ATTR_RO(align);
6148 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6150 return sysfs_emit(buf, "%u\n", s->object_size);
6152 SLAB_ATTR_RO(object_size);
6154 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6156 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6158 SLAB_ATTR_RO(objs_per_slab);
6160 static ssize_t order_show(struct kmem_cache *s, char *buf)
6162 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6164 SLAB_ATTR_RO(order);
6166 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6168 return sysfs_emit(buf, "%lu\n", s->min_partial);
6171 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6177 err = kstrtoul(buf, 10, &min);
6181 s->min_partial = min;
6184 SLAB_ATTR(min_partial);
6186 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6188 unsigned int nr_partial = 0;
6189 #ifdef CONFIG_SLUB_CPU_PARTIAL
6190 nr_partial = s->cpu_partial;
6193 return sysfs_emit(buf, "%u\n", nr_partial);
6196 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6199 unsigned int objects;
6202 err = kstrtouint(buf, 10, &objects);
6205 if (objects && !kmem_cache_has_cpu_partial(s))
6208 slub_set_cpu_partial(s, objects);
6212 SLAB_ATTR(cpu_partial);
6214 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6218 return sysfs_emit(buf, "%pS\n", s->ctor);
6222 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6224 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6226 SLAB_ATTR_RO(aliases);
6228 static ssize_t partial_show(struct kmem_cache *s, char *buf)
6230 return show_slab_objects(s, buf, SO_PARTIAL);
6232 SLAB_ATTR_RO(partial);
6234 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6236 return show_slab_objects(s, buf, SO_CPU);
6238 SLAB_ATTR_RO(cpu_slabs);
6240 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6242 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6244 SLAB_ATTR_RO(objects_partial);
6246 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6250 int cpu __maybe_unused;
6253 #ifdef CONFIG_SLUB_CPU_PARTIAL
6254 for_each_online_cpu(cpu) {
6257 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6260 slabs += slab->slabs;
6264 /* Approximate half-full slabs, see slub_set_cpu_partial() */
6265 objects = (slabs * oo_objects(s->oo)) / 2;
6266 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6268 #ifdef CONFIG_SLUB_CPU_PARTIAL
6269 for_each_online_cpu(cpu) {
6272 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6274 slabs = READ_ONCE(slab->slabs);
6275 objects = (slabs * oo_objects(s->oo)) / 2;
6276 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6277 cpu, objects, slabs);
6281 len += sysfs_emit_at(buf, len, "\n");
6285 SLAB_ATTR_RO(slabs_cpu_partial);
6287 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6289 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6291 SLAB_ATTR_RO(reclaim_account);
6293 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6295 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6297 SLAB_ATTR_RO(hwcache_align);
6299 #ifdef CONFIG_ZONE_DMA
6300 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6302 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6304 SLAB_ATTR_RO(cache_dma);
6307 #ifdef CONFIG_HARDENED_USERCOPY
6308 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6310 return sysfs_emit(buf, "%u\n", s->usersize);
6312 SLAB_ATTR_RO(usersize);
6315 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6317 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6319 SLAB_ATTR_RO(destroy_by_rcu);
6321 #ifdef CONFIG_SLUB_DEBUG
6322 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6324 return show_slab_objects(s, buf, SO_ALL);
6326 SLAB_ATTR_RO(slabs);
6328 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6330 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6332 SLAB_ATTR_RO(total_objects);
6334 static ssize_t objects_show(struct kmem_cache *s, char *buf)
6336 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6338 SLAB_ATTR_RO(objects);
6340 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6342 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6344 SLAB_ATTR_RO(sanity_checks);
6346 static ssize_t trace_show(struct kmem_cache *s, char *buf)
6348 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6350 SLAB_ATTR_RO(trace);
6352 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6354 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6357 SLAB_ATTR_RO(red_zone);
6359 static ssize_t poison_show(struct kmem_cache *s, char *buf)
6361 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6364 SLAB_ATTR_RO(poison);
6366 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6368 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6371 SLAB_ATTR_RO(store_user);
6373 static ssize_t validate_show(struct kmem_cache *s, char *buf)
6378 static ssize_t validate_store(struct kmem_cache *s,
6379 const char *buf, size_t length)
6383 if (buf[0] == '1' && kmem_cache_debug(s)) {
6384 ret = validate_slab_cache(s);
6390 SLAB_ATTR(validate);
6392 #endif /* CONFIG_SLUB_DEBUG */
6394 #ifdef CONFIG_FAILSLAB
6395 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6397 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6400 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6403 if (s->refcount > 1)
6407 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6409 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6413 SLAB_ATTR(failslab);
6416 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6421 static ssize_t shrink_store(struct kmem_cache *s,
6422 const char *buf, size_t length)
6425 kmem_cache_shrink(s);
6433 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6435 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6438 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6439 const char *buf, size_t length)
6444 err = kstrtouint(buf, 10, &ratio);
6450 s->remote_node_defrag_ratio = ratio * 10;
6454 SLAB_ATTR(remote_node_defrag_ratio);
6457 #ifdef CONFIG_SLUB_STATS
6458 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6460 unsigned long sum = 0;
6463 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6468 for_each_online_cpu(cpu) {
6469 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6475 len += sysfs_emit_at(buf, len, "%lu", sum);
6478 for_each_online_cpu(cpu) {
6480 len += sysfs_emit_at(buf, len, " C%d=%u",
6485 len += sysfs_emit_at(buf, len, "\n");
6490 static void clear_stat(struct kmem_cache *s, enum stat_item si)
6494 for_each_online_cpu(cpu)
6495 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6498 #define STAT_ATTR(si, text) \
6499 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
6501 return show_stat(s, buf, si); \
6503 static ssize_t text##_store(struct kmem_cache *s, \
6504 const char *buf, size_t length) \
6506 if (buf[0] != '0') \
6508 clear_stat(s, si); \
6513 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
6514 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
6515 STAT_ATTR(FREE_FASTPATH, free_fastpath);
6516 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
6517 STAT_ATTR(FREE_FROZEN, free_frozen);
6518 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
6519 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
6520 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
6521 STAT_ATTR(ALLOC_SLAB, alloc_slab);
6522 STAT_ATTR(ALLOC_REFILL, alloc_refill);
6523 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
6524 STAT_ATTR(FREE_SLAB, free_slab);
6525 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
6526 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
6527 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
6528 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
6529 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
6530 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
6531 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
6532 STAT_ATTR(ORDER_FALLBACK, order_fallback);
6533 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
6534 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
6535 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
6536 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
6537 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
6538 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
6539 #endif /* CONFIG_SLUB_STATS */
6541 #ifdef CONFIG_KFENCE
6542 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
6544 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
6547 static ssize_t skip_kfence_store(struct kmem_cache *s,
6548 const char *buf, size_t length)
6553 s->flags &= ~SLAB_SKIP_KFENCE;
6554 else if (buf[0] == '1')
6555 s->flags |= SLAB_SKIP_KFENCE;
6561 SLAB_ATTR(skip_kfence);
6564 static struct attribute *slab_attrs[] = {
6565 &slab_size_attr.attr,
6566 &object_size_attr.attr,
6567 &objs_per_slab_attr.attr,
6569 &min_partial_attr.attr,
6570 &cpu_partial_attr.attr,
6571 &objects_partial_attr.attr,
6573 &cpu_slabs_attr.attr,
6577 &hwcache_align_attr.attr,
6578 &reclaim_account_attr.attr,
6579 &destroy_by_rcu_attr.attr,
6581 &slabs_cpu_partial_attr.attr,
6582 #ifdef CONFIG_SLUB_DEBUG
6583 &total_objects_attr.attr,
6586 &sanity_checks_attr.attr,
6588 &red_zone_attr.attr,
6590 &store_user_attr.attr,
6591 &validate_attr.attr,
6593 #ifdef CONFIG_ZONE_DMA
6594 &cache_dma_attr.attr,
6597 &remote_node_defrag_ratio_attr.attr,
6599 #ifdef CONFIG_SLUB_STATS
6600 &alloc_fastpath_attr.attr,
6601 &alloc_slowpath_attr.attr,
6602 &free_fastpath_attr.attr,
6603 &free_slowpath_attr.attr,
6604 &free_frozen_attr.attr,
6605 &free_add_partial_attr.attr,
6606 &free_remove_partial_attr.attr,
6607 &alloc_from_partial_attr.attr,
6608 &alloc_slab_attr.attr,
6609 &alloc_refill_attr.attr,
6610 &alloc_node_mismatch_attr.attr,
6611 &free_slab_attr.attr,
6612 &cpuslab_flush_attr.attr,
6613 &deactivate_full_attr.attr,
6614 &deactivate_empty_attr.attr,
6615 &deactivate_to_head_attr.attr,
6616 &deactivate_to_tail_attr.attr,
6617 &deactivate_remote_frees_attr.attr,
6618 &deactivate_bypass_attr.attr,
6619 &order_fallback_attr.attr,
6620 &cmpxchg_double_fail_attr.attr,
6621 &cmpxchg_double_cpu_fail_attr.attr,
6622 &cpu_partial_alloc_attr.attr,
6623 &cpu_partial_free_attr.attr,
6624 &cpu_partial_node_attr.attr,
6625 &cpu_partial_drain_attr.attr,
6627 #ifdef CONFIG_FAILSLAB
6628 &failslab_attr.attr,
6630 #ifdef CONFIG_HARDENED_USERCOPY
6631 &usersize_attr.attr,
6633 #ifdef CONFIG_KFENCE
6634 &skip_kfence_attr.attr,
6640 static const struct attribute_group slab_attr_group = {
6641 .attrs = slab_attrs,
6644 static ssize_t slab_attr_show(struct kobject *kobj,
6645 struct attribute *attr,
6648 struct slab_attribute *attribute;
6649 struct kmem_cache *s;
6651 attribute = to_slab_attr(attr);
6654 if (!attribute->show)
6657 return attribute->show(s, buf);
6660 static ssize_t slab_attr_store(struct kobject *kobj,
6661 struct attribute *attr,
6662 const char *buf, size_t len)
6664 struct slab_attribute *attribute;
6665 struct kmem_cache *s;
6667 attribute = to_slab_attr(attr);
6670 if (!attribute->store)
6673 return attribute->store(s, buf, len);
6676 static void kmem_cache_release(struct kobject *k)
6678 slab_kmem_cache_release(to_slab(k));
6681 static const struct sysfs_ops slab_sysfs_ops = {
6682 .show = slab_attr_show,
6683 .store = slab_attr_store,
6686 static const struct kobj_type slab_ktype = {
6687 .sysfs_ops = &slab_sysfs_ops,
6688 .release = kmem_cache_release,
6691 static struct kset *slab_kset;
6693 static inline struct kset *cache_kset(struct kmem_cache *s)
6698 #define ID_STR_LENGTH 32
6700 /* Create a unique string id for a slab cache:
6702 * Format :[flags-]size
6704 static char *create_unique_id(struct kmem_cache *s)
6706 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6710 return ERR_PTR(-ENOMEM);
6714 * First flags affecting slabcache operations. We will only
6715 * get here for aliasable slabs so we do not need to support
6716 * too many flags. The flags here must cover all flags that
6717 * are matched during merging to guarantee that the id is
6720 if (s->flags & SLAB_CACHE_DMA)
6722 if (s->flags & SLAB_CACHE_DMA32)
6724 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6726 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6728 if (s->flags & SLAB_ACCOUNT)
6732 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6734 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6736 return ERR_PTR(-EINVAL);
6738 kmsan_unpoison_memory(name, p - name);
6742 static int sysfs_slab_add(struct kmem_cache *s)
6746 struct kset *kset = cache_kset(s);
6747 int unmergeable = slab_unmergeable(s);
6749 if (!unmergeable && disable_higher_order_debug &&
6750 (slub_debug & DEBUG_METADATA_FLAGS))
6755 * Slabcache can never be merged so we can use the name proper.
6756 * This is typically the case for debug situations. In that
6757 * case we can catch duplicate names easily.
6759 sysfs_remove_link(&slab_kset->kobj, s->name);
6763 * Create a unique name for the slab as a target
6766 name = create_unique_id(s);
6768 return PTR_ERR(name);
6771 s->kobj.kset = kset;
6772 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6776 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6781 /* Setup first alias */
6782 sysfs_slab_alias(s, s->name);
6789 kobject_del(&s->kobj);
6793 void sysfs_slab_unlink(struct kmem_cache *s)
6795 if (slab_state >= FULL)
6796 kobject_del(&s->kobj);
6799 void sysfs_slab_release(struct kmem_cache *s)
6801 if (slab_state >= FULL)
6802 kobject_put(&s->kobj);
6806 * Need to buffer aliases during bootup until sysfs becomes
6807 * available lest we lose that information.
6809 struct saved_alias {
6810 struct kmem_cache *s;
6812 struct saved_alias *next;
6815 static struct saved_alias *alias_list;
6817 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6819 struct saved_alias *al;
6821 if (slab_state == FULL) {
6823 * If we have a leftover link then remove it.
6825 sysfs_remove_link(&slab_kset->kobj, name);
6826 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6829 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6835 al->next = alias_list;
6837 kmsan_unpoison_memory(al, sizeof(*al));
6841 static int __init slab_sysfs_init(void)
6843 struct kmem_cache *s;
6846 mutex_lock(&slab_mutex);
6848 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6850 mutex_unlock(&slab_mutex);
6851 pr_err("Cannot register slab subsystem.\n");
6857 list_for_each_entry(s, &slab_caches, list) {
6858 err = sysfs_slab_add(s);
6860 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6864 while (alias_list) {
6865 struct saved_alias *al = alias_list;
6867 alias_list = alias_list->next;
6868 err = sysfs_slab_alias(al->s, al->name);
6870 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6875 mutex_unlock(&slab_mutex);
6878 late_initcall(slab_sysfs_init);
6879 #endif /* SLAB_SUPPORTS_SYSFS */
6881 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6882 static int slab_debugfs_show(struct seq_file *seq, void *v)
6884 struct loc_track *t = seq->private;
6888 idx = (unsigned long) t->idx;
6889 if (idx < t->count) {
6892 seq_printf(seq, "%7ld ", l->count);
6895 seq_printf(seq, "%pS", (void *)l->addr);
6897 seq_puts(seq, "<not-available>");
6900 seq_printf(seq, " waste=%lu/%lu",
6901 l->count * l->waste, l->waste);
6903 if (l->sum_time != l->min_time) {
6904 seq_printf(seq, " age=%ld/%llu/%ld",
6905 l->min_time, div_u64(l->sum_time, l->count),
6908 seq_printf(seq, " age=%ld", l->min_time);
6910 if (l->min_pid != l->max_pid)
6911 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6913 seq_printf(seq, " pid=%ld",
6916 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6917 seq_printf(seq, " cpus=%*pbl",
6918 cpumask_pr_args(to_cpumask(l->cpus)));
6920 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6921 seq_printf(seq, " nodes=%*pbl",
6922 nodemask_pr_args(&l->nodes));
6924 #ifdef CONFIG_STACKDEPOT
6926 depot_stack_handle_t handle;
6927 unsigned long *entries;
6928 unsigned int nr_entries, j;
6930 handle = READ_ONCE(l->handle);
6932 nr_entries = stack_depot_fetch(handle, &entries);
6933 seq_puts(seq, "\n");
6934 for (j = 0; j < nr_entries; j++)
6935 seq_printf(seq, " %pS\n", (void *)entries[j]);
6939 seq_puts(seq, "\n");
6942 if (!idx && !t->count)
6943 seq_puts(seq, "No data\n");
6948 static void slab_debugfs_stop(struct seq_file *seq, void *v)
6952 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6954 struct loc_track *t = seq->private;
6957 if (*ppos <= t->count)
6963 static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6965 struct location *loc1 = (struct location *)a;
6966 struct location *loc2 = (struct location *)b;
6968 if (loc1->count > loc2->count)
6974 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6976 struct loc_track *t = seq->private;
6982 static const struct seq_operations slab_debugfs_sops = {
6983 .start = slab_debugfs_start,
6984 .next = slab_debugfs_next,
6985 .stop = slab_debugfs_stop,
6986 .show = slab_debugfs_show,
6989 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6992 struct kmem_cache_node *n;
6993 enum track_item alloc;
6995 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6996 sizeof(struct loc_track));
6997 struct kmem_cache *s = file_inode(filep)->i_private;
6998 unsigned long *obj_map;
7003 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
7005 seq_release_private(inode, filep);
7009 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
7010 alloc = TRACK_ALLOC;
7014 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7015 bitmap_free(obj_map);
7016 seq_release_private(inode, filep);
7020 for_each_kmem_cache_node(s, node, n) {
7021 unsigned long flags;
7024 if (!node_nr_slabs(n))
7027 spin_lock_irqsave(&n->list_lock, flags);
7028 list_for_each_entry(slab, &n->partial, slab_list)
7029 process_slab(t, s, slab, alloc, obj_map);
7030 list_for_each_entry(slab, &n->full, slab_list)
7031 process_slab(t, s, slab, alloc, obj_map);
7032 spin_unlock_irqrestore(&n->list_lock, flags);
7035 /* Sort locations by count */
7036 sort_r(t->loc, t->count, sizeof(struct location),
7037 cmp_loc_by_count, NULL, NULL);
7039 bitmap_free(obj_map);
7043 static int slab_debug_trace_release(struct inode *inode, struct file *file)
7045 struct seq_file *seq = file->private_data;
7046 struct loc_track *t = seq->private;
7049 return seq_release_private(inode, file);
7052 static const struct file_operations slab_debugfs_fops = {
7053 .open = slab_debug_trace_open,
7055 .llseek = seq_lseek,
7056 .release = slab_debug_trace_release,
7059 static void debugfs_slab_add(struct kmem_cache *s)
7061 struct dentry *slab_cache_dir;
7063 if (unlikely(!slab_debugfs_root))
7066 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7068 debugfs_create_file("alloc_traces", 0400,
7069 slab_cache_dir, s, &slab_debugfs_fops);
7071 debugfs_create_file("free_traces", 0400,
7072 slab_cache_dir, s, &slab_debugfs_fops);
7075 void debugfs_slab_release(struct kmem_cache *s)
7077 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7080 static int __init slab_debugfs_init(void)
7082 struct kmem_cache *s;
7084 slab_debugfs_root = debugfs_create_dir("slab", NULL);
7086 list_for_each_entry(s, &slab_caches, list)
7087 if (s->flags & SLAB_STORE_USER)
7088 debugfs_slab_add(s);
7093 __initcall(slab_debugfs_init);
7096 * The /proc/slabinfo ABI
7098 #ifdef CONFIG_SLUB_DEBUG
7099 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7101 unsigned long nr_slabs = 0;
7102 unsigned long nr_objs = 0;
7103 unsigned long nr_free = 0;
7105 struct kmem_cache_node *n;
7107 for_each_kmem_cache_node(s, node, n) {
7108 nr_slabs += node_nr_slabs(n);
7109 nr_objs += node_nr_objs(n);
7110 nr_free += count_partial(n, count_free);
7113 sinfo->active_objs = nr_objs - nr_free;
7114 sinfo->num_objs = nr_objs;
7115 sinfo->active_slabs = nr_slabs;
7116 sinfo->num_slabs = nr_slabs;
7117 sinfo->objects_per_slab = oo_objects(s->oo);
7118 sinfo->cache_order = oo_order(s->oo);
7121 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
7125 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
7126 size_t count, loff_t *ppos)
7130 #endif /* CONFIG_SLUB_DEBUG */