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 slab_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_FLAG_BIT(_SLAB_OBJECT_POISON)
310 /* Use cmpxchg_double */
312 #ifdef system_has_freelist_aba
313 #define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE)
315 #define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED
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 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);
561 * See comment in calculate_sizes().
563 static inline bool freeptr_outside_object(struct kmem_cache *s)
565 return s->offset >= s->inuse;
569 * Return offset of the end of info block which is inuse + free pointer if
570 * not overlapping with object.
572 static inline unsigned int get_info_end(struct kmem_cache *s)
574 if (freeptr_outside_object(s))
575 return s->inuse + sizeof(void *);
580 /* Loop over all objects in a slab */
581 #define for_each_object(__p, __s, __addr, __objects) \
582 for (__p = fixup_red_left(__s, __addr); \
583 __p < (__addr) + (__objects) * (__s)->size; \
586 static inline unsigned int order_objects(unsigned int order, unsigned int size)
588 return ((unsigned int)PAGE_SIZE << order) / size;
591 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
594 struct kmem_cache_order_objects x = {
595 (order << OO_SHIFT) + order_objects(order, size)
601 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
603 return x.x >> OO_SHIFT;
606 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
608 return x.x & OO_MASK;
611 #ifdef CONFIG_SLUB_CPU_PARTIAL
612 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
614 unsigned int nr_slabs;
616 s->cpu_partial = nr_objects;
619 * We take the number of objects but actually limit the number of
620 * slabs on the per cpu partial list, in order to limit excessive
621 * growth of the list. For simplicity we assume that the slabs will
624 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
625 s->cpu_partial_slabs = nr_slabs;
629 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
632 #endif /* CONFIG_SLUB_CPU_PARTIAL */
635 * Per slab locking using the pagelock
637 static __always_inline void slab_lock(struct slab *slab)
639 struct page *page = slab_page(slab);
641 VM_BUG_ON_PAGE(PageTail(page), page);
642 bit_spin_lock(PG_locked, &page->flags);
645 static __always_inline void slab_unlock(struct slab *slab)
647 struct page *page = slab_page(slab);
649 VM_BUG_ON_PAGE(PageTail(page), page);
650 bit_spin_unlock(PG_locked, &page->flags);
654 __update_freelist_fast(struct slab *slab,
655 void *freelist_old, unsigned long counters_old,
656 void *freelist_new, unsigned long counters_new)
658 #ifdef system_has_freelist_aba
659 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
660 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
662 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
669 __update_freelist_slow(struct slab *slab,
670 void *freelist_old, unsigned long counters_old,
671 void *freelist_new, unsigned long counters_new)
676 if (slab->freelist == freelist_old &&
677 slab->counters == counters_old) {
678 slab->freelist = freelist_new;
679 slab->counters = counters_new;
688 * Interrupts must be disabled (for the fallback code to work right), typically
689 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
690 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
691 * allocation/ free operation in hardirq context. Therefore nothing can
692 * interrupt the operation.
694 static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
695 void *freelist_old, unsigned long counters_old,
696 void *freelist_new, unsigned long counters_new,
701 if (USE_LOCKLESS_FAST_PATH())
702 lockdep_assert_irqs_disabled();
704 if (s->flags & __CMPXCHG_DOUBLE) {
705 ret = __update_freelist_fast(slab, freelist_old, counters_old,
706 freelist_new, counters_new);
708 ret = __update_freelist_slow(slab, freelist_old, counters_old,
709 freelist_new, counters_new);
715 stat(s, CMPXCHG_DOUBLE_FAIL);
717 #ifdef SLUB_DEBUG_CMPXCHG
718 pr_info("%s %s: cmpxchg double redo ", n, s->name);
724 static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
725 void *freelist_old, unsigned long counters_old,
726 void *freelist_new, unsigned long counters_new,
731 if (s->flags & __CMPXCHG_DOUBLE) {
732 ret = __update_freelist_fast(slab, freelist_old, counters_old,
733 freelist_new, counters_new);
737 local_irq_save(flags);
738 ret = __update_freelist_slow(slab, freelist_old, counters_old,
739 freelist_new, counters_new);
740 local_irq_restore(flags);
746 stat(s, CMPXCHG_DOUBLE_FAIL);
748 #ifdef SLUB_DEBUG_CMPXCHG
749 pr_info("%s %s: cmpxchg double redo ", n, s->name);
755 #ifdef CONFIG_SLUB_DEBUG
756 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
757 static DEFINE_SPINLOCK(object_map_lock);
759 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
762 void *addr = slab_address(slab);
765 bitmap_zero(obj_map, slab->objects);
767 for (p = slab->freelist; p; p = get_freepointer(s, p))
768 set_bit(__obj_to_index(s, addr, p), obj_map);
771 #if IS_ENABLED(CONFIG_KUNIT)
772 static bool slab_add_kunit_errors(void)
774 struct kunit_resource *resource;
776 if (!kunit_get_current_test())
779 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
783 (*(int *)resource->data)++;
784 kunit_put_resource(resource);
788 static inline bool slab_add_kunit_errors(void) { return false; }
791 static inline unsigned int size_from_object(struct kmem_cache *s)
793 if (s->flags & SLAB_RED_ZONE)
794 return s->size - s->red_left_pad;
799 static inline void *restore_red_left(struct kmem_cache *s, void *p)
801 if (s->flags & SLAB_RED_ZONE)
802 p -= s->red_left_pad;
810 #if defined(CONFIG_SLUB_DEBUG_ON)
811 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
813 static slab_flags_t slub_debug;
816 static char *slub_debug_string;
817 static int disable_higher_order_debug;
820 * slub is about to manipulate internal object metadata. This memory lies
821 * outside the range of the allocated object, so accessing it would normally
822 * be reported by kasan as a bounds error. metadata_access_enable() is used
823 * to tell kasan that these accesses are OK.
825 static inline void metadata_access_enable(void)
827 kasan_disable_current();
830 static inline void metadata_access_disable(void)
832 kasan_enable_current();
839 /* Verify that a pointer has an address that is valid within a slab page */
840 static inline int check_valid_pointer(struct kmem_cache *s,
841 struct slab *slab, void *object)
848 base = slab_address(slab);
849 object = kasan_reset_tag(object);
850 object = restore_red_left(s, object);
851 if (object < base || object >= base + slab->objects * s->size ||
852 (object - base) % s->size) {
859 static void print_section(char *level, char *text, u8 *addr,
862 metadata_access_enable();
863 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
864 16, 1, kasan_reset_tag((void *)addr), length, 1);
865 metadata_access_disable();
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);
1501 atomic_long_inc(&n->nr_slabs);
1502 atomic_long_add(objects, &n->total_objects);
1504 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1506 struct kmem_cache_node *n = get_node(s, node);
1508 atomic_long_dec(&n->nr_slabs);
1509 atomic_long_sub(objects, &n->total_objects);
1512 /* Object debug checks for alloc/free paths */
1513 static void setup_object_debug(struct kmem_cache *s, void *object)
1515 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1518 init_object(s, object, SLUB_RED_INACTIVE);
1519 init_tracking(s, object);
1523 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1525 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1528 metadata_access_enable();
1529 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1530 metadata_access_disable();
1533 static inline int alloc_consistency_checks(struct kmem_cache *s,
1534 struct slab *slab, void *object)
1536 if (!check_slab(s, slab))
1539 if (!check_valid_pointer(s, slab, object)) {
1540 object_err(s, slab, object, "Freelist Pointer check fails");
1544 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1550 static noinline bool alloc_debug_processing(struct kmem_cache *s,
1551 struct slab *slab, void *object, int orig_size)
1553 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1554 if (!alloc_consistency_checks(s, slab, object))
1558 /* Success. Perform special debug activities for allocs */
1559 trace(s, slab, object, 1);
1560 set_orig_size(s, object, orig_size);
1561 init_object(s, object, SLUB_RED_ACTIVE);
1565 if (folio_test_slab(slab_folio(slab))) {
1567 * If this is a slab page then lets do the best we can
1568 * to avoid issues in the future. Marking all objects
1569 * as used avoids touching the remaining objects.
1571 slab_fix(s, "Marking all objects used");
1572 slab->inuse = slab->objects;
1573 slab->freelist = NULL;
1578 static inline int free_consistency_checks(struct kmem_cache *s,
1579 struct slab *slab, void *object, unsigned long addr)
1581 if (!check_valid_pointer(s, slab, object)) {
1582 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1586 if (on_freelist(s, slab, object)) {
1587 object_err(s, slab, object, "Object already free");
1591 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1594 if (unlikely(s != slab->slab_cache)) {
1595 if (!folio_test_slab(slab_folio(slab))) {
1596 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1598 } else if (!slab->slab_cache) {
1599 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1603 object_err(s, slab, object,
1604 "page slab pointer corrupt.");
1611 * Parse a block of slab_debug options. Blocks are delimited by ';'
1613 * @str: start of block
1614 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1615 * @slabs: return start of list of slabs, or NULL when there's no list
1616 * @init: assume this is initial parsing and not per-kmem-create parsing
1618 * returns the start of next block if there's any, or NULL
1621 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1623 bool higher_order_disable = false;
1625 /* Skip any completely empty blocks */
1626 while (*str && *str == ';')
1631 * No options but restriction on slabs. This means full
1632 * debugging for slabs matching a pattern.
1634 *flags = DEBUG_DEFAULT_FLAGS;
1639 /* Determine which debug features should be switched on */
1640 for (; *str && *str != ',' && *str != ';'; str++) {
1641 switch (tolower(*str)) {
1646 *flags |= SLAB_CONSISTENCY_CHECKS;
1649 *flags |= SLAB_RED_ZONE;
1652 *flags |= SLAB_POISON;
1655 *flags |= SLAB_STORE_USER;
1658 *flags |= SLAB_TRACE;
1661 *flags |= SLAB_FAILSLAB;
1665 * Avoid enabling debugging on caches if its minimum
1666 * order would increase as a result.
1668 higher_order_disable = true;
1672 pr_err("slab_debug option '%c' unknown. skipped\n", *str);
1681 /* Skip over the slab list */
1682 while (*str && *str != ';')
1685 /* Skip any completely empty blocks */
1686 while (*str && *str == ';')
1689 if (init && higher_order_disable)
1690 disable_higher_order_debug = 1;
1698 static int __init setup_slub_debug(char *str)
1701 slab_flags_t global_flags;
1704 bool global_slub_debug_changed = false;
1705 bool slab_list_specified = false;
1707 global_flags = DEBUG_DEFAULT_FLAGS;
1708 if (*str++ != '=' || !*str)
1710 * No options specified. Switch on full debugging.
1716 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1719 global_flags = flags;
1720 global_slub_debug_changed = true;
1722 slab_list_specified = true;
1723 if (flags & SLAB_STORE_USER)
1724 stack_depot_request_early_init();
1729 * For backwards compatibility, a single list of flags with list of
1730 * slabs means debugging is only changed for those slabs, so the global
1731 * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1732 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1733 * long as there is no option specifying flags without a slab list.
1735 if (slab_list_specified) {
1736 if (!global_slub_debug_changed)
1737 global_flags = slub_debug;
1738 slub_debug_string = saved_str;
1741 slub_debug = global_flags;
1742 if (slub_debug & SLAB_STORE_USER)
1743 stack_depot_request_early_init();
1744 if (slub_debug != 0 || slub_debug_string)
1745 static_branch_enable(&slub_debug_enabled);
1747 static_branch_disable(&slub_debug_enabled);
1748 if ((static_branch_unlikely(&init_on_alloc) ||
1749 static_branch_unlikely(&init_on_free)) &&
1750 (slub_debug & SLAB_POISON))
1751 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1755 __setup("slab_debug", setup_slub_debug);
1756 __setup_param("slub_debug", slub_debug, setup_slub_debug, 0);
1759 * kmem_cache_flags - apply debugging options to the cache
1760 * @flags: flags to set
1761 * @name: name of the cache
1763 * Debug option(s) are applied to @flags. In addition to the debug
1764 * option(s), if a slab name (or multiple) is specified i.e.
1765 * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1766 * then only the select slabs will receive the debug option(s).
1768 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1773 slab_flags_t block_flags;
1774 slab_flags_t slub_debug_local = slub_debug;
1776 if (flags & SLAB_NO_USER_FLAGS)
1780 * If the slab cache is for debugging (e.g. kmemleak) then
1781 * don't store user (stack trace) information by default,
1782 * but let the user enable it via the command line below.
1784 if (flags & SLAB_NOLEAKTRACE)
1785 slub_debug_local &= ~SLAB_STORE_USER;
1788 next_block = slub_debug_string;
1789 /* Go through all blocks of debug options, see if any matches our slab's name */
1790 while (next_block) {
1791 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1794 /* Found a block that has a slab list, search it */
1799 end = strchrnul(iter, ',');
1800 if (next_block && next_block < end)
1801 end = next_block - 1;
1803 glob = strnchr(iter, end - iter, '*');
1805 cmplen = glob - iter;
1807 cmplen = max_t(size_t, len, (end - iter));
1809 if (!strncmp(name, iter, cmplen)) {
1810 flags |= block_flags;
1814 if (!*end || *end == ';')
1820 return flags | slub_debug_local;
1822 #else /* !CONFIG_SLUB_DEBUG */
1823 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1825 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1827 static inline bool alloc_debug_processing(struct kmem_cache *s,
1828 struct slab *slab, void *object, int orig_size) { return true; }
1830 static inline bool free_debug_processing(struct kmem_cache *s,
1831 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1832 unsigned long addr, depot_stack_handle_t handle) { return true; }
1834 static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1835 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1836 void *object, u8 val) { return 1; }
1837 static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1838 static inline void set_track(struct kmem_cache *s, void *object,
1839 enum track_item alloc, unsigned long addr) {}
1840 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1841 struct slab *slab) {}
1842 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1843 struct slab *slab) {}
1844 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1848 #define slub_debug 0
1850 #define disable_higher_order_debug 0
1852 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1854 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1856 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1859 #ifndef CONFIG_SLUB_TINY
1860 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1861 void **freelist, void *nextfree)
1866 #endif /* CONFIG_SLUB_DEBUG */
1868 static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s)
1870 return (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1871 NR_SLAB_RECLAIMABLE_B : NR_SLAB_UNRECLAIMABLE_B;
1874 #ifdef CONFIG_MEMCG_KMEM
1875 static inline void memcg_free_slab_cgroups(struct slab *slab)
1877 kfree(slab_objcgs(slab));
1878 slab->memcg_data = 0;
1881 static inline size_t obj_full_size(struct kmem_cache *s)
1884 * For each accounted object there is an extra space which is used
1885 * to store obj_cgroup membership. Charge it too.
1887 return s->size + sizeof(struct obj_cgroup *);
1891 * Returns false if the allocation should fail.
1893 static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
1894 struct list_lru *lru,
1895 struct obj_cgroup **objcgp,
1896 size_t objects, gfp_t flags)
1899 * The obtained objcg pointer is safe to use within the current scope,
1900 * defined by current task or set_active_memcg() pair.
1901 * obj_cgroup_get() is used to get a permanent reference.
1903 struct obj_cgroup *objcg = current_obj_cgroup();
1909 struct mem_cgroup *memcg;
1911 memcg = get_mem_cgroup_from_objcg(objcg);
1912 ret = memcg_list_lru_alloc(memcg, lru, flags);
1913 css_put(&memcg->css);
1919 if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s)))
1927 * Returns false if the allocation should fail.
1929 static __fastpath_inline
1930 bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
1931 struct obj_cgroup **objcgp, size_t objects,
1934 if (!memcg_kmem_online())
1937 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
1940 return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects,
1944 static void __memcg_slab_post_alloc_hook(struct kmem_cache *s,
1945 struct obj_cgroup *objcg,
1946 gfp_t flags, size_t size,
1953 flags &= gfp_allowed_mask;
1955 for (i = 0; i < size; i++) {
1957 slab = virt_to_slab(p[i]);
1959 if (!slab_objcgs(slab) &&
1960 memcg_alloc_slab_cgroups(slab, s, flags, false)) {
1961 obj_cgroup_uncharge(objcg, obj_full_size(s));
1965 off = obj_to_index(s, slab, p[i]);
1966 obj_cgroup_get(objcg);
1967 slab_objcgs(slab)[off] = objcg;
1968 mod_objcg_state(objcg, slab_pgdat(slab),
1969 cache_vmstat_idx(s), obj_full_size(s));
1971 obj_cgroup_uncharge(objcg, obj_full_size(s));
1976 static __fastpath_inline
1977 void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
1978 gfp_t flags, size_t size, void **p)
1980 if (likely(!memcg_kmem_online() || !objcg))
1983 return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
1986 static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
1987 void **p, int objects,
1988 struct obj_cgroup **objcgs)
1990 for (int i = 0; i < objects; i++) {
1991 struct obj_cgroup *objcg;
1994 off = obj_to_index(s, slab, p[i]);
1995 objcg = objcgs[off];
2000 obj_cgroup_uncharge(objcg, obj_full_size(s));
2001 mod_objcg_state(objcg, slab_pgdat(slab), cache_vmstat_idx(s),
2003 obj_cgroup_put(objcg);
2007 static __fastpath_inline
2008 void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2011 struct obj_cgroup **objcgs;
2013 if (!memcg_kmem_online())
2016 objcgs = slab_objcgs(slab);
2017 if (likely(!objcgs))
2020 __memcg_slab_free_hook(s, slab, p, objects, objcgs);
2024 void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2025 struct obj_cgroup *objcg)
2028 obj_cgroup_uncharge(objcg, objects * obj_full_size(s));
2030 #else /* CONFIG_MEMCG_KMEM */
2031 static inline void memcg_free_slab_cgroups(struct slab *slab)
2035 static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s,
2036 struct list_lru *lru,
2037 struct obj_cgroup **objcgp,
2038 size_t objects, gfp_t flags)
2043 static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
2044 struct obj_cgroup *objcg,
2045 gfp_t flags, size_t size,
2050 static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2051 void **p, int objects)
2056 void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2057 struct obj_cgroup *objcg)
2060 #endif /* CONFIG_MEMCG_KMEM */
2063 * Hooks for other subsystems that check memory allocations. In a typical
2064 * production configuration these hooks all should produce no code at all.
2066 * Returns true if freeing of the object can proceed, false if its reuse
2067 * was delayed by KASAN quarantine, or it was returned to KFENCE.
2069 static __always_inline
2070 bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
2072 kmemleak_free_recursive(x, s->flags);
2073 kmsan_slab_free(s, x);
2075 debug_check_no_locks_freed(x, s->object_size);
2077 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2078 debug_check_no_obj_freed(x, s->object_size);
2080 /* Use KCSAN to help debug racy use-after-free. */
2081 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
2082 __kcsan_check_access(x, s->object_size,
2083 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2089 * As memory initialization might be integrated into KASAN,
2090 * kasan_slab_free and initialization memset's must be
2091 * kept together to avoid discrepancies in behavior.
2093 * The initialization memset's clear the object and the metadata,
2094 * but don't touch the SLAB redzone.
2096 * The object's freepointer is also avoided if stored outside the
2099 if (unlikely(init)) {
2103 inuse = get_info_end(s);
2104 if (!kasan_has_integrated_init())
2105 memset(kasan_reset_tag(x), 0, s->object_size);
2106 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2107 memset((char *)kasan_reset_tag(x) + inuse, 0,
2108 s->size - inuse - rsize);
2110 /* KASAN might put x into memory quarantine, delaying its reuse. */
2111 return !kasan_slab_free(s, x, init);
2114 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
2115 void **head, void **tail,
2121 void *old_tail = *tail;
2124 if (is_kfence_address(next)) {
2125 slab_free_hook(s, next, false);
2129 /* Head and tail of the reconstructed freelist */
2133 init = slab_want_init_on_free(s);
2137 next = get_freepointer(s, object);
2139 /* If object's reuse doesn't have to be delayed */
2140 if (likely(slab_free_hook(s, object, init))) {
2141 /* Move object to the new freelist */
2142 set_freepointer(s, object, *head);
2148 * Adjust the reconstructed freelist depth
2149 * accordingly if object's reuse is delayed.
2153 } while (object != old_tail);
2155 return *head != NULL;
2158 static void *setup_object(struct kmem_cache *s, void *object)
2160 setup_object_debug(s, object);
2161 object = kasan_init_slab_obj(s, object);
2162 if (unlikely(s->ctor)) {
2163 kasan_unpoison_new_object(s, object);
2165 kasan_poison_new_object(s, object);
2171 * Slab allocation and freeing
2173 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2174 struct kmem_cache_order_objects oo)
2176 struct folio *folio;
2178 unsigned int order = oo_order(oo);
2180 folio = (struct folio *)alloc_pages_node(node, flags, order);
2184 slab = folio_slab(folio);
2185 __folio_set_slab(folio);
2186 /* Make the flag visible before any changes to folio->mapping */
2188 if (folio_is_pfmemalloc(folio))
2189 slab_set_pfmemalloc(slab);
2194 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2195 /* Pre-initialize the random sequence cache */
2196 static int init_cache_random_seq(struct kmem_cache *s)
2198 unsigned int count = oo_objects(s->oo);
2201 /* Bailout if already initialised */
2205 err = cache_random_seq_create(s, count, GFP_KERNEL);
2207 pr_err("SLUB: Unable to initialize free list for %s\n",
2212 /* Transform to an offset on the set of pages */
2213 if (s->random_seq) {
2216 for (i = 0; i < count; i++)
2217 s->random_seq[i] *= s->size;
2222 /* Initialize each random sequence freelist per cache */
2223 static void __init init_freelist_randomization(void)
2225 struct kmem_cache *s;
2227 mutex_lock(&slab_mutex);
2229 list_for_each_entry(s, &slab_caches, list)
2230 init_cache_random_seq(s);
2232 mutex_unlock(&slab_mutex);
2235 /* Get the next entry on the pre-computed freelist randomized */
2236 static void *next_freelist_entry(struct kmem_cache *s,
2237 unsigned long *pos, void *start,
2238 unsigned long page_limit,
2239 unsigned long freelist_count)
2244 * If the target page allocation failed, the number of objects on the
2245 * page might be smaller than the usual size defined by the cache.
2248 idx = s->random_seq[*pos];
2250 if (*pos >= freelist_count)
2252 } while (unlikely(idx >= page_limit));
2254 return (char *)start + idx;
2257 /* Shuffle the single linked freelist based on a random pre-computed sequence */
2258 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2263 unsigned long idx, pos, page_limit, freelist_count;
2265 if (slab->objects < 2 || !s->random_seq)
2268 freelist_count = oo_objects(s->oo);
2269 pos = get_random_u32_below(freelist_count);
2271 page_limit = slab->objects * s->size;
2272 start = fixup_red_left(s, slab_address(slab));
2274 /* First entry is used as the base of the freelist */
2275 cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count);
2276 cur = setup_object(s, cur);
2277 slab->freelist = cur;
2279 for (idx = 1; idx < slab->objects; idx++) {
2280 next = next_freelist_entry(s, &pos, start, page_limit,
2282 next = setup_object(s, next);
2283 set_freepointer(s, cur, next);
2286 set_freepointer(s, cur, NULL);
2291 static inline int init_cache_random_seq(struct kmem_cache *s)
2295 static inline void init_freelist_randomization(void) { }
2296 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2300 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2302 static __always_inline void account_slab(struct slab *slab, int order,
2303 struct kmem_cache *s, gfp_t gfp)
2305 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2306 memcg_alloc_slab_cgroups(slab, s, gfp, true);
2308 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2309 PAGE_SIZE << order);
2312 static __always_inline void unaccount_slab(struct slab *slab, int order,
2313 struct kmem_cache *s)
2315 if (memcg_kmem_online())
2316 memcg_free_slab_cgroups(slab);
2318 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2319 -(PAGE_SIZE << order));
2322 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2325 struct kmem_cache_order_objects oo = s->oo;
2327 void *start, *p, *next;
2331 flags &= gfp_allowed_mask;
2333 flags |= s->allocflags;
2336 * Let the initial higher-order allocation fail under memory pressure
2337 * so we fall-back to the minimum order allocation.
2339 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2340 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2341 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2343 slab = alloc_slab_page(alloc_gfp, node, oo);
2344 if (unlikely(!slab)) {
2348 * Allocation may have failed due to fragmentation.
2349 * Try a lower order alloc if possible
2351 slab = alloc_slab_page(alloc_gfp, node, oo);
2352 if (unlikely(!slab))
2354 stat(s, ORDER_FALLBACK);
2357 slab->objects = oo_objects(oo);
2361 account_slab(slab, oo_order(oo), s, flags);
2363 slab->slab_cache = s;
2365 kasan_poison_slab(slab);
2367 start = slab_address(slab);
2369 setup_slab_debug(s, slab, start);
2371 shuffle = shuffle_freelist(s, slab);
2374 start = fixup_red_left(s, start);
2375 start = setup_object(s, start);
2376 slab->freelist = start;
2377 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2379 next = setup_object(s, next);
2380 set_freepointer(s, p, next);
2383 set_freepointer(s, p, NULL);
2389 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2391 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2392 flags = kmalloc_fix_flags(flags);
2394 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2396 return allocate_slab(s,
2397 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2400 static void __free_slab(struct kmem_cache *s, struct slab *slab)
2402 struct folio *folio = slab_folio(slab);
2403 int order = folio_order(folio);
2404 int pages = 1 << order;
2406 __slab_clear_pfmemalloc(slab);
2407 folio->mapping = NULL;
2408 /* Make the mapping reset visible before clearing the flag */
2410 __folio_clear_slab(folio);
2411 mm_account_reclaimed_pages(pages);
2412 unaccount_slab(slab, order, s);
2413 __free_pages(&folio->page, order);
2416 static void rcu_free_slab(struct rcu_head *h)
2418 struct slab *slab = container_of(h, struct slab, rcu_head);
2420 __free_slab(slab->slab_cache, slab);
2423 static void free_slab(struct kmem_cache *s, struct slab *slab)
2425 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2428 slab_pad_check(s, slab);
2429 for_each_object(p, s, slab_address(slab), slab->objects)
2430 check_object(s, slab, p, SLUB_RED_INACTIVE);
2433 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2434 call_rcu(&slab->rcu_head, rcu_free_slab);
2436 __free_slab(s, slab);
2439 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2441 dec_slabs_node(s, slab_nid(slab), slab->objects);
2446 * SLUB reuses PG_workingset bit to keep track of whether it's on
2447 * the per-node partial list.
2449 static inline bool slab_test_node_partial(const struct slab *slab)
2451 return folio_test_workingset((struct folio *)slab_folio(slab));
2454 static inline void slab_set_node_partial(struct slab *slab)
2456 set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2459 static inline void slab_clear_node_partial(struct slab *slab)
2461 clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2465 * Management of partially allocated slabs.
2468 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2471 if (tail == DEACTIVATE_TO_TAIL)
2472 list_add_tail(&slab->slab_list, &n->partial);
2474 list_add(&slab->slab_list, &n->partial);
2475 slab_set_node_partial(slab);
2478 static inline void add_partial(struct kmem_cache_node *n,
2479 struct slab *slab, int tail)
2481 lockdep_assert_held(&n->list_lock);
2482 __add_partial(n, slab, tail);
2485 static inline void remove_partial(struct kmem_cache_node *n,
2488 lockdep_assert_held(&n->list_lock);
2489 list_del(&slab->slab_list);
2490 slab_clear_node_partial(slab);
2495 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2496 * slab from the n->partial list. Remove only a single object from the slab, do
2497 * the alloc_debug_processing() checks and leave the slab on the list, or move
2498 * it to full list if it was the last free object.
2500 static void *alloc_single_from_partial(struct kmem_cache *s,
2501 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2505 lockdep_assert_held(&n->list_lock);
2507 object = slab->freelist;
2508 slab->freelist = get_freepointer(s, object);
2511 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2512 remove_partial(n, slab);
2516 if (slab->inuse == slab->objects) {
2517 remove_partial(n, slab);
2518 add_full(s, n, slab);
2525 * Called only for kmem_cache_debug() caches to allocate from a freshly
2526 * allocated slab. Allocate a single object instead of whole freelist
2527 * and put the slab to the partial (or full) list.
2529 static void *alloc_single_from_new_slab(struct kmem_cache *s,
2530 struct slab *slab, int orig_size)
2532 int nid = slab_nid(slab);
2533 struct kmem_cache_node *n = get_node(s, nid);
2534 unsigned long flags;
2538 object = slab->freelist;
2539 slab->freelist = get_freepointer(s, object);
2542 if (!alloc_debug_processing(s, slab, object, orig_size))
2544 * It's not really expected that this would fail on a
2545 * freshly allocated slab, but a concurrent memory
2546 * corruption in theory could cause that.
2550 spin_lock_irqsave(&n->list_lock, flags);
2552 if (slab->inuse == slab->objects)
2553 add_full(s, n, slab);
2555 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2557 inc_slabs_node(s, nid, slab->objects);
2558 spin_unlock_irqrestore(&n->list_lock, flags);
2563 #ifdef CONFIG_SLUB_CPU_PARTIAL
2564 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2566 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2569 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2572 * Try to allocate a partial slab from a specific node.
2574 static struct slab *get_partial_node(struct kmem_cache *s,
2575 struct kmem_cache_node *n,
2576 struct partial_context *pc)
2578 struct slab *slab, *slab2, *partial = NULL;
2579 unsigned long flags;
2580 unsigned int partial_slabs = 0;
2583 * Racy check. If we mistakenly see no partial slabs then we
2584 * just allocate an empty slab. If we mistakenly try to get a
2585 * partial slab and there is none available then get_partial()
2588 if (!n || !n->nr_partial)
2591 spin_lock_irqsave(&n->list_lock, flags);
2592 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2593 if (!pfmemalloc_match(slab, pc->flags))
2596 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2597 void *object = alloc_single_from_partial(s, n, slab,
2601 pc->object = object;
2607 remove_partial(n, slab);
2611 stat(s, ALLOC_FROM_PARTIAL);
2613 put_cpu_partial(s, slab, 0);
2614 stat(s, CPU_PARTIAL_NODE);
2617 #ifdef CONFIG_SLUB_CPU_PARTIAL
2618 if (!kmem_cache_has_cpu_partial(s)
2619 || partial_slabs > s->cpu_partial_slabs / 2)
2626 spin_unlock_irqrestore(&n->list_lock, flags);
2631 * Get a slab from somewhere. Search in increasing NUMA distances.
2633 static struct slab *get_any_partial(struct kmem_cache *s,
2634 struct partial_context *pc)
2637 struct zonelist *zonelist;
2640 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2642 unsigned int cpuset_mems_cookie;
2645 * The defrag ratio allows a configuration of the tradeoffs between
2646 * inter node defragmentation and node local allocations. A lower
2647 * defrag_ratio increases the tendency to do local allocations
2648 * instead of attempting to obtain partial slabs from other nodes.
2650 * If the defrag_ratio is set to 0 then kmalloc() always
2651 * returns node local objects. If the ratio is higher then kmalloc()
2652 * may return off node objects because partial slabs are obtained
2653 * from other nodes and filled up.
2655 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2656 * (which makes defrag_ratio = 1000) then every (well almost)
2657 * allocation will first attempt to defrag slab caches on other nodes.
2658 * This means scanning over all nodes to look for partial slabs which
2659 * may be expensive if we do it every time we are trying to find a slab
2660 * with available objects.
2662 if (!s->remote_node_defrag_ratio ||
2663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2667 cpuset_mems_cookie = read_mems_allowed_begin();
2668 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2669 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2670 struct kmem_cache_node *n;
2672 n = get_node(s, zone_to_nid(zone));
2674 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2675 n->nr_partial > s->min_partial) {
2676 slab = get_partial_node(s, n, pc);
2679 * Don't check read_mems_allowed_retry()
2680 * here - if mems_allowed was updated in
2681 * parallel, that was a harmless race
2682 * between allocation and the cpuset
2689 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2690 #endif /* CONFIG_NUMA */
2695 * Get a partial slab, lock it and return it.
2697 static struct slab *get_partial(struct kmem_cache *s, int node,
2698 struct partial_context *pc)
2701 int searchnode = node;
2703 if (node == NUMA_NO_NODE)
2704 searchnode = numa_mem_id();
2706 slab = get_partial_node(s, get_node(s, searchnode), pc);
2707 if (slab || node != NUMA_NO_NODE)
2710 return get_any_partial(s, pc);
2713 #ifndef CONFIG_SLUB_TINY
2715 #ifdef CONFIG_PREEMPTION
2717 * Calculate the next globally unique transaction for disambiguation
2718 * during cmpxchg. The transactions start with the cpu number and are then
2719 * incremented by CONFIG_NR_CPUS.
2721 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2724 * No preemption supported therefore also no need to check for
2728 #endif /* CONFIG_PREEMPTION */
2730 static inline unsigned long next_tid(unsigned long tid)
2732 return tid + TID_STEP;
2735 #ifdef SLUB_DEBUG_CMPXCHG
2736 static inline unsigned int tid_to_cpu(unsigned long tid)
2738 return tid % TID_STEP;
2741 static inline unsigned long tid_to_event(unsigned long tid)
2743 return tid / TID_STEP;
2747 static inline unsigned int init_tid(int cpu)
2752 static inline void note_cmpxchg_failure(const char *n,
2753 const struct kmem_cache *s, unsigned long tid)
2755 #ifdef SLUB_DEBUG_CMPXCHG
2756 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2758 pr_info("%s %s: cmpxchg redo ", n, s->name);
2760 #ifdef CONFIG_PREEMPTION
2761 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2762 pr_warn("due to cpu change %d -> %d\n",
2763 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2766 if (tid_to_event(tid) != tid_to_event(actual_tid))
2767 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2768 tid_to_event(tid), tid_to_event(actual_tid));
2770 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2771 actual_tid, tid, next_tid(tid));
2773 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2776 static void init_kmem_cache_cpus(struct kmem_cache *s)
2779 struct kmem_cache_cpu *c;
2781 for_each_possible_cpu(cpu) {
2782 c = per_cpu_ptr(s->cpu_slab, cpu);
2783 local_lock_init(&c->lock);
2784 c->tid = init_tid(cpu);
2789 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2790 * unfreezes the slabs and puts it on the proper list.
2791 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2794 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2797 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2799 void *nextfree, *freelist_iter, *freelist_tail;
2800 int tail = DEACTIVATE_TO_HEAD;
2801 unsigned long flags = 0;
2805 if (slab->freelist) {
2806 stat(s, DEACTIVATE_REMOTE_FREES);
2807 tail = DEACTIVATE_TO_TAIL;
2811 * Stage one: Count the objects on cpu's freelist as free_delta and
2812 * remember the last object in freelist_tail for later splicing.
2814 freelist_tail = NULL;
2815 freelist_iter = freelist;
2816 while (freelist_iter) {
2817 nextfree = get_freepointer(s, freelist_iter);
2820 * If 'nextfree' is invalid, it is possible that the object at
2821 * 'freelist_iter' is already corrupted. So isolate all objects
2822 * starting at 'freelist_iter' by skipping them.
2824 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2827 freelist_tail = freelist_iter;
2830 freelist_iter = nextfree;
2834 * Stage two: Unfreeze the slab while splicing the per-cpu
2835 * freelist to the head of slab's freelist.
2838 old.freelist = READ_ONCE(slab->freelist);
2839 old.counters = READ_ONCE(slab->counters);
2840 VM_BUG_ON(!old.frozen);
2842 /* Determine target state of the slab */
2843 new.counters = old.counters;
2845 if (freelist_tail) {
2846 new.inuse -= free_delta;
2847 set_freepointer(s, freelist_tail, old.freelist);
2848 new.freelist = freelist;
2850 new.freelist = old.freelist;
2852 } while (!slab_update_freelist(s, slab,
2853 old.freelist, old.counters,
2854 new.freelist, new.counters,
2855 "unfreezing slab"));
2858 * Stage three: Manipulate the slab list based on the updated state.
2860 if (!new.inuse && n->nr_partial >= s->min_partial) {
2861 stat(s, DEACTIVATE_EMPTY);
2862 discard_slab(s, slab);
2864 } else if (new.freelist) {
2865 spin_lock_irqsave(&n->list_lock, flags);
2866 add_partial(n, slab, tail);
2867 spin_unlock_irqrestore(&n->list_lock, flags);
2870 stat(s, DEACTIVATE_FULL);
2874 #ifdef CONFIG_SLUB_CPU_PARTIAL
2875 static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
2877 struct kmem_cache_node *n = NULL, *n2 = NULL;
2878 struct slab *slab, *slab_to_discard = NULL;
2879 unsigned long flags = 0;
2881 while (partial_slab) {
2882 slab = partial_slab;
2883 partial_slab = slab->next;
2885 n2 = get_node(s, slab_nid(slab));
2888 spin_unlock_irqrestore(&n->list_lock, flags);
2891 spin_lock_irqsave(&n->list_lock, flags);
2894 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
2895 slab->next = slab_to_discard;
2896 slab_to_discard = slab;
2898 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2899 stat(s, FREE_ADD_PARTIAL);
2904 spin_unlock_irqrestore(&n->list_lock, flags);
2906 while (slab_to_discard) {
2907 slab = slab_to_discard;
2908 slab_to_discard = slab_to_discard->next;
2910 stat(s, DEACTIVATE_EMPTY);
2911 discard_slab(s, slab);
2917 * Put all the cpu partial slabs to the node partial list.
2919 static void put_partials(struct kmem_cache *s)
2921 struct slab *partial_slab;
2922 unsigned long flags;
2924 local_lock_irqsave(&s->cpu_slab->lock, flags);
2925 partial_slab = this_cpu_read(s->cpu_slab->partial);
2926 this_cpu_write(s->cpu_slab->partial, NULL);
2927 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2930 __put_partials(s, partial_slab);
2933 static void put_partials_cpu(struct kmem_cache *s,
2934 struct kmem_cache_cpu *c)
2936 struct slab *partial_slab;
2938 partial_slab = slub_percpu_partial(c);
2942 __put_partials(s, partial_slab);
2946 * Put a slab into a partial slab slot if available.
2948 * If we did not find a slot then simply move all the partials to the
2949 * per node partial list.
2951 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2953 struct slab *oldslab;
2954 struct slab *slab_to_put = NULL;
2955 unsigned long flags;
2958 local_lock_irqsave(&s->cpu_slab->lock, flags);
2960 oldslab = this_cpu_read(s->cpu_slab->partial);
2963 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2965 * Partial array is full. Move the existing set to the
2966 * per node partial list. Postpone the actual unfreezing
2967 * outside of the critical section.
2969 slab_to_put = oldslab;
2972 slabs = oldslab->slabs;
2978 slab->slabs = slabs;
2979 slab->next = oldslab;
2981 this_cpu_write(s->cpu_slab->partial, slab);
2983 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2986 __put_partials(s, slab_to_put);
2987 stat(s, CPU_PARTIAL_DRAIN);
2991 #else /* CONFIG_SLUB_CPU_PARTIAL */
2993 static inline void put_partials(struct kmem_cache *s) { }
2994 static inline void put_partials_cpu(struct kmem_cache *s,
2995 struct kmem_cache_cpu *c) { }
2997 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2999 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3001 unsigned long flags;
3005 local_lock_irqsave(&s->cpu_slab->lock, flags);
3008 freelist = c->freelist;
3012 c->tid = next_tid(c->tid);
3014 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3017 deactivate_slab(s, slab, freelist);
3018 stat(s, CPUSLAB_FLUSH);
3022 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3024 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3025 void *freelist = c->freelist;
3026 struct slab *slab = c->slab;
3030 c->tid = next_tid(c->tid);
3033 deactivate_slab(s, slab, freelist);
3034 stat(s, CPUSLAB_FLUSH);
3037 put_partials_cpu(s, c);
3040 struct slub_flush_work {
3041 struct work_struct work;
3042 struct kmem_cache *s;
3049 * Called from CPU work handler with migration disabled.
3051 static void flush_cpu_slab(struct work_struct *w)
3053 struct kmem_cache *s;
3054 struct kmem_cache_cpu *c;
3055 struct slub_flush_work *sfw;
3057 sfw = container_of(w, struct slub_flush_work, work);
3060 c = this_cpu_ptr(s->cpu_slab);
3068 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3070 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3072 return c->slab || slub_percpu_partial(c);
3075 static DEFINE_MUTEX(flush_lock);
3076 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3078 static void flush_all_cpus_locked(struct kmem_cache *s)
3080 struct slub_flush_work *sfw;
3083 lockdep_assert_cpus_held();
3084 mutex_lock(&flush_lock);
3086 for_each_online_cpu(cpu) {
3087 sfw = &per_cpu(slub_flush, cpu);
3088 if (!has_cpu_slab(cpu, s)) {
3092 INIT_WORK(&sfw->work, flush_cpu_slab);
3095 queue_work_on(cpu, flushwq, &sfw->work);
3098 for_each_online_cpu(cpu) {
3099 sfw = &per_cpu(slub_flush, cpu);
3102 flush_work(&sfw->work);
3105 mutex_unlock(&flush_lock);
3108 static void flush_all(struct kmem_cache *s)
3111 flush_all_cpus_locked(s);
3116 * Use the cpu notifier to insure that the cpu slabs are flushed when
3119 static int slub_cpu_dead(unsigned int cpu)
3121 struct kmem_cache *s;
3123 mutex_lock(&slab_mutex);
3124 list_for_each_entry(s, &slab_caches, list)
3125 __flush_cpu_slab(s, cpu);
3126 mutex_unlock(&slab_mutex);
3130 #else /* CONFIG_SLUB_TINY */
3131 static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
3132 static inline void flush_all(struct kmem_cache *s) { }
3133 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
3134 static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3135 #endif /* CONFIG_SLUB_TINY */
3138 * Check if the objects in a per cpu structure fit numa
3139 * locality expectations.
3141 static inline int node_match(struct slab *slab, int node)
3144 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3150 #ifdef CONFIG_SLUB_DEBUG
3151 static int count_free(struct slab *slab)
3153 return slab->objects - slab->inuse;
3156 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3158 return atomic_long_read(&n->total_objects);
3161 /* Supports checking bulk free of a constructed freelist */
3162 static inline bool free_debug_processing(struct kmem_cache *s,
3163 struct slab *slab, void *head, void *tail, int *bulk_cnt,
3164 unsigned long addr, depot_stack_handle_t handle)
3166 bool checks_ok = false;
3167 void *object = head;
3170 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3171 if (!check_slab(s, slab))
3175 if (slab->inuse < *bulk_cnt) {
3176 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3177 slab->inuse, *bulk_cnt);
3183 if (++cnt > *bulk_cnt)
3186 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3187 if (!free_consistency_checks(s, slab, object, addr))
3191 if (s->flags & SLAB_STORE_USER)
3192 set_track_update(s, object, TRACK_FREE, addr, handle);
3193 trace(s, slab, object, 0);
3194 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3195 init_object(s, object, SLUB_RED_INACTIVE);
3197 /* Reached end of constructed freelist yet? */
3198 if (object != tail) {
3199 object = get_freepointer(s, object);
3205 if (cnt != *bulk_cnt) {
3206 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3214 slab_fix(s, "Object at 0x%p not freed", object);
3218 #endif /* CONFIG_SLUB_DEBUG */
3220 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
3221 static unsigned long count_partial(struct kmem_cache_node *n,
3222 int (*get_count)(struct slab *))
3224 unsigned long flags;
3225 unsigned long x = 0;
3228 spin_lock_irqsave(&n->list_lock, flags);
3229 list_for_each_entry(slab, &n->partial, slab_list)
3230 x += get_count(slab);
3231 spin_unlock_irqrestore(&n->list_lock, flags);
3234 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3236 #ifdef CONFIG_SLUB_DEBUG
3237 static noinline void
3238 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3240 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3241 DEFAULT_RATELIMIT_BURST);
3243 struct kmem_cache_node *n;
3245 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3248 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
3249 nid, gfpflags, &gfpflags);
3250 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3251 s->name, s->object_size, s->size, oo_order(s->oo),
3254 if (oo_order(s->min) > get_order(s->object_size))
3255 pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n",
3258 for_each_kmem_cache_node(s, node, n) {
3259 unsigned long nr_slabs;
3260 unsigned long nr_objs;
3261 unsigned long nr_free;
3263 nr_free = count_partial(n, count_free);
3264 nr_slabs = node_nr_slabs(n);
3265 nr_objs = node_nr_objs(n);
3267 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3268 node, nr_slabs, nr_objs, nr_free);
3271 #else /* CONFIG_SLUB_DEBUG */
3273 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3276 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3278 if (unlikely(slab_test_pfmemalloc(slab)))
3279 return gfp_pfmemalloc_allowed(gfpflags);
3284 #ifndef CONFIG_SLUB_TINY
3286 __update_cpu_freelist_fast(struct kmem_cache *s,
3287 void *freelist_old, void *freelist_new,
3290 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3291 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3293 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3294 &old.full, new.full);
3298 * Check the slab->freelist and either transfer the freelist to the
3299 * per cpu freelist or deactivate the slab.
3301 * The slab is still frozen if the return value is not NULL.
3303 * If this function returns NULL then the slab has been unfrozen.
3305 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3308 unsigned long counters;
3311 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3314 freelist = slab->freelist;
3315 counters = slab->counters;
3317 new.counters = counters;
3319 new.inuse = slab->objects;
3320 new.frozen = freelist != NULL;
3322 } while (!__slab_update_freelist(s, slab,
3331 * Freeze the partial slab and return the pointer to the freelist.
3333 static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3336 unsigned long counters;
3340 freelist = slab->freelist;
3341 counters = slab->counters;
3343 new.counters = counters;
3344 VM_BUG_ON(new.frozen);
3346 new.inuse = slab->objects;
3349 } while (!slab_update_freelist(s, slab,
3358 * Slow path. The lockless freelist is empty or we need to perform
3361 * Processing is still very fast if new objects have been freed to the
3362 * regular freelist. In that case we simply take over the regular freelist
3363 * as the lockless freelist and zap the regular freelist.
3365 * If that is not working then we fall back to the partial lists. We take the
3366 * first element of the freelist as the object to allocate now and move the
3367 * rest of the freelist to the lockless freelist.
3369 * And if we were unable to get a new slab from the partial slab lists then
3370 * we need to allocate a new slab. This is the slowest path since it involves
3371 * a call to the page allocator and the setup of a new slab.
3373 * Version of __slab_alloc to use when we know that preemption is
3374 * already disabled (which is the case for bulk allocation).
3376 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3377 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3381 unsigned long flags;
3382 struct partial_context pc;
3384 stat(s, ALLOC_SLOWPATH);
3388 slab = READ_ONCE(c->slab);
3391 * if the node is not online or has no normal memory, just
3392 * ignore the node constraint
3394 if (unlikely(node != NUMA_NO_NODE &&
3395 !node_isset(node, slab_nodes)))
3396 node = NUMA_NO_NODE;
3400 if (unlikely(!node_match(slab, node))) {
3402 * same as above but node_match() being false already
3403 * implies node != NUMA_NO_NODE
3405 if (!node_isset(node, slab_nodes)) {
3406 node = NUMA_NO_NODE;
3408 stat(s, ALLOC_NODE_MISMATCH);
3409 goto deactivate_slab;
3414 * By rights, we should be searching for a slab page that was
3415 * PFMEMALLOC but right now, we are losing the pfmemalloc
3416 * information when the page leaves the per-cpu allocator
3418 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3419 goto deactivate_slab;
3421 /* must check again c->slab in case we got preempted and it changed */
3422 local_lock_irqsave(&s->cpu_slab->lock, flags);
3423 if (unlikely(slab != c->slab)) {
3424 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3427 freelist = c->freelist;
3431 freelist = get_freelist(s, slab);
3435 c->tid = next_tid(c->tid);
3436 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3437 stat(s, DEACTIVATE_BYPASS);
3441 stat(s, ALLOC_REFILL);
3445 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3448 * freelist is pointing to the list of objects to be used.
3449 * slab is pointing to the slab from which the objects are obtained.
3450 * That slab must be frozen for per cpu allocations to work.
3452 VM_BUG_ON(!c->slab->frozen);
3453 c->freelist = get_freepointer(s, freelist);
3454 c->tid = next_tid(c->tid);
3455 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3460 local_lock_irqsave(&s->cpu_slab->lock, flags);
3461 if (slab != c->slab) {
3462 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3465 freelist = c->freelist;
3468 c->tid = next_tid(c->tid);
3469 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3470 deactivate_slab(s, slab, freelist);
3474 #ifdef CONFIG_SLUB_CPU_PARTIAL
3475 while (slub_percpu_partial(c)) {
3476 local_lock_irqsave(&s->cpu_slab->lock, flags);
3477 if (unlikely(c->slab)) {
3478 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3481 if (unlikely(!slub_percpu_partial(c))) {
3482 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3483 /* we were preempted and partial list got empty */
3487 slab = slub_percpu_partial(c);
3488 slub_set_percpu_partial(c, slab);
3490 if (likely(node_match(slab, node) &&
3491 pfmemalloc_match(slab, gfpflags))) {
3493 freelist = get_freelist(s, slab);
3494 VM_BUG_ON(!freelist);
3495 stat(s, CPU_PARTIAL_ALLOC);
3499 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3502 __put_partials(s, slab);
3508 pc.flags = gfpflags;
3509 pc.orig_size = orig_size;
3510 slab = get_partial(s, node, &pc);
3512 if (kmem_cache_debug(s)) {
3513 freelist = pc.object;
3515 * For debug caches here we had to go through
3516 * alloc_single_from_partial() so just store the
3517 * tracking info and return the object.
3519 if (s->flags & SLAB_STORE_USER)
3520 set_track(s, freelist, TRACK_ALLOC, addr);
3525 freelist = freeze_slab(s, slab);
3526 goto retry_load_slab;
3529 slub_put_cpu_ptr(s->cpu_slab);
3530 slab = new_slab(s, gfpflags, node);
3531 c = slub_get_cpu_ptr(s->cpu_slab);
3533 if (unlikely(!slab)) {
3534 slab_out_of_memory(s, gfpflags, node);
3538 stat(s, ALLOC_SLAB);
3540 if (kmem_cache_debug(s)) {
3541 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3543 if (unlikely(!freelist))
3546 if (s->flags & SLAB_STORE_USER)
3547 set_track(s, freelist, TRACK_ALLOC, addr);
3553 * No other reference to the slab yet so we can
3554 * muck around with it freely without cmpxchg
3556 freelist = slab->freelist;
3557 slab->freelist = NULL;
3558 slab->inuse = slab->objects;
3561 inc_slabs_node(s, slab_nid(slab), slab->objects);
3563 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3565 * For !pfmemalloc_match() case we don't load freelist so that
3566 * we don't make further mismatched allocations easier.
3568 deactivate_slab(s, slab, get_freepointer(s, freelist));
3574 local_lock_irqsave(&s->cpu_slab->lock, flags);
3575 if (unlikely(c->slab)) {
3576 void *flush_freelist = c->freelist;
3577 struct slab *flush_slab = c->slab;
3581 c->tid = next_tid(c->tid);
3583 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3585 deactivate_slab(s, flush_slab, flush_freelist);
3587 stat(s, CPUSLAB_FLUSH);
3589 goto retry_load_slab;
3597 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3598 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3601 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3602 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3606 #ifdef CONFIG_PREEMPT_COUNT
3608 * We may have been preempted and rescheduled on a different
3609 * cpu before disabling preemption. Need to reload cpu area
3612 c = slub_get_cpu_ptr(s->cpu_slab);
3615 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3616 #ifdef CONFIG_PREEMPT_COUNT
3617 slub_put_cpu_ptr(s->cpu_slab);
3622 static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3623 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3625 struct kmem_cache_cpu *c;
3632 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3633 * enabled. We may switch back and forth between cpus while
3634 * reading from one cpu area. That does not matter as long
3635 * as we end up on the original cpu again when doing the cmpxchg.
3637 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3638 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3639 * the tid. If we are preempted and switched to another cpu between the
3640 * two reads, it's OK as the two are still associated with the same cpu
3641 * and cmpxchg later will validate the cpu.
3643 c = raw_cpu_ptr(s->cpu_slab);
3644 tid = READ_ONCE(c->tid);
3647 * Irqless object alloc/free algorithm used here depends on sequence
3648 * of fetching cpu_slab's data. tid should be fetched before anything
3649 * on c to guarantee that object and slab associated with previous tid
3650 * won't be used with current tid. If we fetch tid first, object and
3651 * slab could be one associated with next tid and our alloc/free
3652 * request will be failed. In this case, we will retry. So, no problem.
3657 * The transaction ids are globally unique per cpu and per operation on
3658 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3659 * occurs on the right processor and that there was no operation on the
3660 * linked list in between.
3663 object = c->freelist;
3666 if (!USE_LOCKLESS_FAST_PATH() ||
3667 unlikely(!object || !slab || !node_match(slab, node))) {
3668 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3670 void *next_object = get_freepointer_safe(s, object);
3673 * The cmpxchg will only match if there was no additional
3674 * operation and if we are on the right processor.
3676 * The cmpxchg does the following atomically (without lock
3678 * 1. Relocate first pointer to the current per cpu area.
3679 * 2. Verify that tid and freelist have not been changed
3680 * 3. If they were not changed replace tid and freelist
3682 * Since this is without lock semantics the protection is only
3683 * against code executing on this cpu *not* from access by
3686 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3687 note_cmpxchg_failure("slab_alloc", s, tid);
3690 prefetch_freepointer(s, next_object);
3691 stat(s, ALLOC_FASTPATH);
3696 #else /* CONFIG_SLUB_TINY */
3697 static void *__slab_alloc_node(struct kmem_cache *s,
3698 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3700 struct partial_context pc;
3704 pc.flags = gfpflags;
3705 pc.orig_size = orig_size;
3706 slab = get_partial(s, node, &pc);
3711 slab = new_slab(s, gfpflags, node);
3712 if (unlikely(!slab)) {
3713 slab_out_of_memory(s, gfpflags, node);
3717 object = alloc_single_from_new_slab(s, slab, orig_size);
3721 #endif /* CONFIG_SLUB_TINY */
3724 * If the object has been wiped upon free, make sure it's fully initialized by
3725 * zeroing out freelist pointer.
3727 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3730 if (unlikely(slab_want_init_on_free(s)) && obj &&
3731 !freeptr_outside_object(s))
3732 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3736 noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
3738 if (__should_failslab(s, gfpflags))
3742 ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
3744 static __fastpath_inline
3745 struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
3746 struct list_lru *lru,
3747 struct obj_cgroup **objcgp,
3748 size_t size, gfp_t flags)
3750 flags &= gfp_allowed_mask;
3754 if (unlikely(should_failslab(s, flags)))
3757 if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags)))
3763 static __fastpath_inline
3764 void slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
3765 gfp_t flags, size_t size, void **p, bool init,
3766 unsigned int orig_size)
3768 unsigned int zero_size = s->object_size;
3769 bool kasan_init = init;
3771 gfp_t init_flags = flags & gfp_allowed_mask;
3774 * For kmalloc object, the allocated memory size(object_size) is likely
3775 * larger than the requested size(orig_size). If redzone check is
3776 * enabled for the extra space, don't zero it, as it will be redzoned
3777 * soon. The redzone operation for this extra space could be seen as a
3778 * replacement of current poisoning under certain debug option, and
3779 * won't break other sanity checks.
3781 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
3782 (s->flags & SLAB_KMALLOC))
3783 zero_size = orig_size;
3786 * When slab_debug is enabled, avoid memory initialization integrated
3787 * into KASAN and instead zero out the memory via the memset below with
3788 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
3789 * cause false-positive reports. This does not lead to a performance
3790 * penalty on production builds, as slab_debug is not intended to be
3793 if (__slub_debug_enabled())
3797 * As memory initialization might be integrated into KASAN,
3798 * kasan_slab_alloc and initialization memset must be
3799 * kept together to avoid discrepancies in behavior.
3801 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
3803 for (i = 0; i < size; i++) {
3804 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
3805 if (p[i] && init && (!kasan_init ||
3806 !kasan_has_integrated_init()))
3807 memset(p[i], 0, zero_size);
3808 kmemleak_alloc_recursive(p[i], s->object_size, 1,
3809 s->flags, init_flags);
3810 kmsan_slab_alloc(s, p[i], init_flags);
3813 memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
3817 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3818 * have the fastpath folded into their functions. So no function call
3819 * overhead for requests that can be satisfied on the fastpath.
3821 * The fastpath works by first checking if the lockless freelist can be used.
3822 * If not then __slab_alloc is called for slow processing.
3824 * Otherwise we can simply pick the next object from the lockless free list.
3826 static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3827 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3830 struct obj_cgroup *objcg = NULL;
3833 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3837 object = kfence_alloc(s, orig_size, gfpflags);
3838 if (unlikely(object))
3841 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3843 maybe_wipe_obj_freeptr(s, object);
3844 init = slab_want_init_on_alloc(gfpflags, s);
3848 * When init equals 'true', like for kzalloc() family, only
3849 * @orig_size bytes might be zeroed instead of s->object_size
3851 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3856 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3858 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
3861 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3865 EXPORT_SYMBOL(kmem_cache_alloc);
3867 void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3870 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
3873 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3877 EXPORT_SYMBOL(kmem_cache_alloc_lru);
3880 * kmem_cache_alloc_node - Allocate an object on the specified node
3881 * @s: The cache to allocate from.
3882 * @gfpflags: See kmalloc().
3883 * @node: node number of the target node.
3885 * Identical to kmem_cache_alloc but it will allocate memory on the given
3886 * node, which can improve the performance for cpu bound structures.
3888 * Fallback to other node is possible if __GFP_THISNODE is not set.
3890 * Return: pointer to the new object or %NULL in case of error
3892 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3894 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3896 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3900 EXPORT_SYMBOL(kmem_cache_alloc_node);
3903 * To avoid unnecessary overhead, we pass through large allocation requests
3904 * directly to the page allocator. We use __GFP_COMP, because we will need to
3905 * know the allocation order to free the pages properly in kfree.
3907 static void *__kmalloc_large_node(size_t size, gfp_t flags, int node)
3909 struct folio *folio;
3911 unsigned int order = get_order(size);
3913 if (unlikely(flags & GFP_SLAB_BUG_MASK))
3914 flags = kmalloc_fix_flags(flags);
3916 flags |= __GFP_COMP;
3917 folio = (struct folio *)alloc_pages_node(node, flags, order);
3919 ptr = folio_address(folio);
3920 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
3921 PAGE_SIZE << order);
3924 ptr = kasan_kmalloc_large(ptr, size, flags);
3925 /* As ptr might get tagged, call kmemleak hook after KASAN. */
3926 kmemleak_alloc(ptr, size, 1, flags);
3927 kmsan_kmalloc_large(ptr, size, flags);
3932 void *kmalloc_large(size_t size, gfp_t flags)
3934 void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE);
3936 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3937 flags, NUMA_NO_NODE);
3940 EXPORT_SYMBOL(kmalloc_large);
3942 void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3944 void *ret = __kmalloc_large_node(size, flags, node);
3946 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3950 EXPORT_SYMBOL(kmalloc_large_node);
3952 static __always_inline
3953 void *__do_kmalloc_node(size_t size, gfp_t flags, int node,
3954 unsigned long caller)
3956 struct kmem_cache *s;
3959 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3960 ret = __kmalloc_large_node(size, flags, node);
3961 trace_kmalloc(caller, ret, size,
3962 PAGE_SIZE << get_order(size), flags, node);
3966 if (unlikely(!size))
3967 return ZERO_SIZE_PTR;
3969 s = kmalloc_slab(size, flags, caller);
3971 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
3972 ret = kasan_kmalloc(s, ret, size, flags);
3973 trace_kmalloc(caller, ret, size, s->size, flags, node);
3977 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3979 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3981 EXPORT_SYMBOL(__kmalloc_node);
3983 void *__kmalloc(size_t size, gfp_t flags)
3985 return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_);
3987 EXPORT_SYMBOL(__kmalloc);
3989 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3990 int node, unsigned long caller)
3992 return __do_kmalloc_node(size, flags, node, caller);
3994 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3996 void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3998 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
4001 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4003 ret = kasan_kmalloc(s, ret, size, gfpflags);
4006 EXPORT_SYMBOL(kmalloc_trace);
4008 void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags,
4009 int node, size_t size)
4011 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4013 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4015 ret = kasan_kmalloc(s, ret, size, gfpflags);
4018 EXPORT_SYMBOL(kmalloc_node_trace);
4020 static noinline void free_to_partial_list(
4021 struct kmem_cache *s, struct slab *slab,
4022 void *head, void *tail, int bulk_cnt,
4025 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4026 struct slab *slab_free = NULL;
4028 unsigned long flags;
4029 depot_stack_handle_t handle = 0;
4031 if (s->flags & SLAB_STORE_USER)
4032 handle = set_track_prepare();
4034 spin_lock_irqsave(&n->list_lock, flags);
4036 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4037 void *prior = slab->freelist;
4039 /* Perform the actual freeing while we still hold the locks */
4041 set_freepointer(s, tail, prior);
4042 slab->freelist = head;
4045 * If the slab is empty, and node's partial list is full,
4046 * it should be discarded anyway no matter it's on full or
4049 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4053 /* was on full list */
4054 remove_full(s, n, slab);
4056 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4057 stat(s, FREE_ADD_PARTIAL);
4059 } else if (slab_free) {
4060 remove_partial(n, slab);
4061 stat(s, FREE_REMOVE_PARTIAL);
4067 * Update the counters while still holding n->list_lock to
4068 * prevent spurious validation warnings
4070 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4073 spin_unlock_irqrestore(&n->list_lock, flags);
4077 free_slab(s, slab_free);
4082 * Slow path handling. This may still be called frequently since objects
4083 * have a longer lifetime than the cpu slabs in most processing loads.
4085 * So we still attempt to reduce cache line usage. Just take the slab
4086 * lock and free the item. If there is no additional partial slab
4087 * handling required then we can return immediately.
4089 static void __slab_free(struct kmem_cache *s, struct slab *slab,
4090 void *head, void *tail, int cnt,
4097 unsigned long counters;
4098 struct kmem_cache_node *n = NULL;
4099 unsigned long flags;
4100 bool on_node_partial;
4102 stat(s, FREE_SLOWPATH);
4104 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4105 free_to_partial_list(s, slab, head, tail, cnt, addr);
4111 spin_unlock_irqrestore(&n->list_lock, flags);
4114 prior = slab->freelist;
4115 counters = slab->counters;
4116 set_freepointer(s, tail, prior);
4117 new.counters = counters;
4118 was_frozen = new.frozen;
4120 if ((!new.inuse || !prior) && !was_frozen) {
4121 /* Needs to be taken off a list */
4122 if (!kmem_cache_has_cpu_partial(s) || prior) {
4124 n = get_node(s, slab_nid(slab));
4126 * Speculatively acquire the list_lock.
4127 * If the cmpxchg does not succeed then we may
4128 * drop the list_lock without any processing.
4130 * Otherwise the list_lock will synchronize with
4131 * other processors updating the list of slabs.
4133 spin_lock_irqsave(&n->list_lock, flags);
4135 on_node_partial = slab_test_node_partial(slab);
4139 } while (!slab_update_freelist(s, slab,
4146 if (likely(was_frozen)) {
4148 * The list lock was not taken therefore no list
4149 * activity can be necessary.
4151 stat(s, FREE_FROZEN);
4152 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
4154 * If we started with a full slab then put it onto the
4155 * per cpu partial list.
4157 put_cpu_partial(s, slab, 1);
4158 stat(s, CPU_PARTIAL_FREE);
4165 * This slab was partially empty but not on the per-node partial list,
4166 * in which case we shouldn't manipulate its list, just return.
4168 if (prior && !on_node_partial) {
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4173 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4177 * Objects left in the slab. If it was not on the partial list before
4180 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4181 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4182 stat(s, FREE_ADD_PARTIAL);
4184 spin_unlock_irqrestore(&n->list_lock, flags);
4190 * Slab on the partial list.
4192 remove_partial(n, slab);
4193 stat(s, FREE_REMOVE_PARTIAL);
4196 spin_unlock_irqrestore(&n->list_lock, flags);
4198 discard_slab(s, slab);
4201 #ifndef CONFIG_SLUB_TINY
4203 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4204 * can perform fastpath freeing without additional function calls.
4206 * The fastpath is only possible if we are freeing to the current cpu slab
4207 * of this processor. This typically the case if we have just allocated
4210 * If fastpath is not possible then fall back to __slab_free where we deal
4211 * with all sorts of special processing.
4213 * Bulk free of a freelist with several objects (all pointing to the
4214 * same slab) possible by specifying head and tail ptr, plus objects
4215 * count (cnt). Bulk free indicated by tail pointer being set.
4217 static __always_inline void do_slab_free(struct kmem_cache *s,
4218 struct slab *slab, void *head, void *tail,
4219 int cnt, unsigned long addr)
4221 struct kmem_cache_cpu *c;
4227 * Determine the currently cpus per cpu slab.
4228 * The cpu may change afterward. However that does not matter since
4229 * data is retrieved via this pointer. If we are on the same cpu
4230 * during the cmpxchg then the free will succeed.
4232 c = raw_cpu_ptr(s->cpu_slab);
4233 tid = READ_ONCE(c->tid);
4235 /* Same with comment on barrier() in slab_alloc_node() */
4238 if (unlikely(slab != c->slab)) {
4239 __slab_free(s, slab, head, tail, cnt, addr);
4243 if (USE_LOCKLESS_FAST_PATH()) {
4244 freelist = READ_ONCE(c->freelist);
4246 set_freepointer(s, tail, freelist);
4248 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4249 note_cmpxchg_failure("slab_free", s, tid);
4253 /* Update the free list under the local lock */
4254 local_lock(&s->cpu_slab->lock);
4255 c = this_cpu_ptr(s->cpu_slab);
4256 if (unlikely(slab != c->slab)) {
4257 local_unlock(&s->cpu_slab->lock);
4261 freelist = c->freelist;
4263 set_freepointer(s, tail, freelist);
4265 c->tid = next_tid(tid);
4267 local_unlock(&s->cpu_slab->lock);
4269 stat_add(s, FREE_FASTPATH, cnt);
4271 #else /* CONFIG_SLUB_TINY */
4272 static void do_slab_free(struct kmem_cache *s,
4273 struct slab *slab, void *head, void *tail,
4274 int cnt, unsigned long addr)
4276 __slab_free(s, slab, head, tail, cnt, addr);
4278 #endif /* CONFIG_SLUB_TINY */
4280 static __fastpath_inline
4281 void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4284 memcg_slab_free_hook(s, slab, &object, 1);
4286 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s))))
4287 do_slab_free(s, slab, object, object, 1, addr);
4290 static __fastpath_inline
4291 void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4292 void *tail, void **p, int cnt, unsigned long addr)
4294 memcg_slab_free_hook(s, slab, p, cnt);
4296 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4297 * to remove objects, whose reuse must be delayed.
4299 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4300 do_slab_free(s, slab, head, tail, cnt, addr);
4303 #ifdef CONFIG_KASAN_GENERIC
4304 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4306 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4310 static inline struct kmem_cache *virt_to_cache(const void *obj)
4314 slab = virt_to_slab(obj);
4315 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4317 return slab->slab_cache;
4320 static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4322 struct kmem_cache *cachep;
4324 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4325 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4328 cachep = virt_to_cache(x);
4329 if (WARN(cachep && cachep != s,
4330 "%s: Wrong slab cache. %s but object is from %s\n",
4331 __func__, s->name, cachep->name))
4332 print_tracking(cachep, x);
4337 * kmem_cache_free - Deallocate an object
4338 * @s: The cache the allocation was from.
4339 * @x: The previously allocated object.
4341 * Free an object which was previously allocated from this
4344 void kmem_cache_free(struct kmem_cache *s, void *x)
4346 s = cache_from_obj(s, x);
4349 trace_kmem_cache_free(_RET_IP_, x, s);
4350 slab_free(s, virt_to_slab(x), x, _RET_IP_);
4352 EXPORT_SYMBOL(kmem_cache_free);
4354 static void free_large_kmalloc(struct folio *folio, void *object)
4356 unsigned int order = folio_order(folio);
4358 if (WARN_ON_ONCE(order == 0))
4359 pr_warn_once("object pointer: 0x%p\n", object);
4361 kmemleak_free(object);
4362 kasan_kfree_large(object);
4363 kmsan_kfree_large(object);
4365 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4366 -(PAGE_SIZE << order));
4371 * kfree - free previously allocated memory
4372 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4374 * If @object is NULL, no operation is performed.
4376 void kfree(const void *object)
4378 struct folio *folio;
4380 struct kmem_cache *s;
4381 void *x = (void *)object;
4383 trace_kfree(_RET_IP_, object);
4385 if (unlikely(ZERO_OR_NULL_PTR(object)))
4388 folio = virt_to_folio(object);
4389 if (unlikely(!folio_test_slab(folio))) {
4390 free_large_kmalloc(folio, (void *)object);
4394 slab = folio_slab(folio);
4395 s = slab->slab_cache;
4396 slab_free(s, slab, x, _RET_IP_);
4398 EXPORT_SYMBOL(kfree);
4400 struct detached_freelist {
4405 struct kmem_cache *s;
4409 * This function progressively scans the array with free objects (with
4410 * a limited look ahead) and extract objects belonging to the same
4411 * slab. It builds a detached freelist directly within the given
4412 * slab/objects. This can happen without any need for
4413 * synchronization, because the objects are owned by running process.
4414 * The freelist is build up as a single linked list in the objects.
4415 * The idea is, that this detached freelist can then be bulk
4416 * transferred to the real freelist(s), but only requiring a single
4417 * synchronization primitive. Look ahead in the array is limited due
4418 * to performance reasons.
4421 int build_detached_freelist(struct kmem_cache *s, size_t size,
4422 void **p, struct detached_freelist *df)
4426 struct folio *folio;
4430 folio = virt_to_folio(object);
4432 /* Handle kalloc'ed objects */
4433 if (unlikely(!folio_test_slab(folio))) {
4434 free_large_kmalloc(folio, object);
4438 /* Derive kmem_cache from object */
4439 df->slab = folio_slab(folio);
4440 df->s = df->slab->slab_cache;
4442 df->slab = folio_slab(folio);
4443 df->s = cache_from_obj(s, object); /* Support for memcg */
4446 /* Start new detached freelist */
4448 df->freelist = object;
4451 if (is_kfence_address(object))
4454 set_freepointer(df->s, object, NULL);
4459 /* df->slab is always set at this point */
4460 if (df->slab == virt_to_slab(object)) {
4461 /* Opportunity build freelist */
4462 set_freepointer(df->s, object, df->freelist);
4463 df->freelist = object;
4467 swap(p[size], p[same]);
4471 /* Limit look ahead search */
4480 * Internal bulk free of objects that were not initialised by the post alloc
4481 * hooks and thus should not be processed by the free hooks
4483 static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4489 struct detached_freelist df;
4491 size = build_detached_freelist(s, size, p, &df);
4495 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4497 } while (likely(size));
4500 /* Note that interrupts must be enabled when calling this function. */
4501 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4507 struct detached_freelist df;
4509 size = build_detached_freelist(s, size, p, &df);
4513 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
4515 } while (likely(size));
4517 EXPORT_SYMBOL(kmem_cache_free_bulk);
4519 #ifndef CONFIG_SLUB_TINY
4521 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4524 struct kmem_cache_cpu *c;
4525 unsigned long irqflags;
4529 * Drain objects in the per cpu slab, while disabling local
4530 * IRQs, which protects against PREEMPT and interrupts
4531 * handlers invoking normal fastpath.
4533 c = slub_get_cpu_ptr(s->cpu_slab);
4534 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4536 for (i = 0; i < size; i++) {
4537 void *object = kfence_alloc(s, s->object_size, flags);
4539 if (unlikely(object)) {
4544 object = c->freelist;
4545 if (unlikely(!object)) {
4547 * We may have removed an object from c->freelist using
4548 * the fastpath in the previous iteration; in that case,
4549 * c->tid has not been bumped yet.
4550 * Since ___slab_alloc() may reenable interrupts while
4551 * allocating memory, we should bump c->tid now.
4553 c->tid = next_tid(c->tid);
4555 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4558 * Invoking slow path likely have side-effect
4559 * of re-populating per CPU c->freelist
4561 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
4562 _RET_IP_, c, s->object_size);
4563 if (unlikely(!p[i]))
4566 c = this_cpu_ptr(s->cpu_slab);
4567 maybe_wipe_obj_freeptr(s, p[i]);
4569 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4571 continue; /* goto for-loop */
4573 c->freelist = get_freepointer(s, object);
4575 maybe_wipe_obj_freeptr(s, p[i]);
4576 stat(s, ALLOC_FASTPATH);
4578 c->tid = next_tid(c->tid);
4579 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4580 slub_put_cpu_ptr(s->cpu_slab);
4585 slub_put_cpu_ptr(s->cpu_slab);
4586 __kmem_cache_free_bulk(s, i, p);
4590 #else /* CONFIG_SLUB_TINY */
4591 static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4592 size_t size, void **p)
4596 for (i = 0; i < size; i++) {
4597 void *object = kfence_alloc(s, s->object_size, flags);
4599 if (unlikely(object)) {
4604 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4605 _RET_IP_, s->object_size);
4606 if (unlikely(!p[i]))
4609 maybe_wipe_obj_freeptr(s, p[i]);
4615 __kmem_cache_free_bulk(s, i, p);
4618 #endif /* CONFIG_SLUB_TINY */
4620 /* Note that interrupts must be enabled when calling this function. */
4621 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4625 struct obj_cgroup *objcg = NULL;
4630 /* memcg and kmem_cache debug support */
4631 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4635 i = __kmem_cache_alloc_bulk(s, flags, size, p);
4638 * memcg and kmem_cache debug support and memory initialization.
4639 * Done outside of the IRQ disabled fastpath loop.
4641 if (likely(i != 0)) {
4642 slab_post_alloc_hook(s, objcg, flags, size, p,
4643 slab_want_init_on_alloc(flags, s), s->object_size);
4645 memcg_slab_alloc_error_hook(s, size, objcg);
4650 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4654 * Object placement in a slab is made very easy because we always start at
4655 * offset 0. If we tune the size of the object to the alignment then we can
4656 * get the required alignment by putting one properly sized object after
4659 * Notice that the allocation order determines the sizes of the per cpu
4660 * caches. Each processor has always one slab available for allocations.
4661 * Increasing the allocation order reduces the number of times that slabs
4662 * must be moved on and off the partial lists and is therefore a factor in
4667 * Minimum / Maximum order of slab pages. This influences locking overhead
4668 * and slab fragmentation. A higher order reduces the number of partial slabs
4669 * and increases the number of allocations possible without having to
4670 * take the list_lock.
4672 static unsigned int slub_min_order;
4673 static unsigned int slub_max_order =
4674 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4675 static unsigned int slub_min_objects;
4678 * Calculate the order of allocation given an slab object size.
4680 * The order of allocation has significant impact on performance and other
4681 * system components. Generally order 0 allocations should be preferred since
4682 * order 0 does not cause fragmentation in the page allocator. Larger objects
4683 * be problematic to put into order 0 slabs because there may be too much
4684 * unused space left. We go to a higher order if more than 1/16th of the slab
4687 * In order to reach satisfactory performance we must ensure that a minimum
4688 * number of objects is in one slab. Otherwise we may generate too much
4689 * activity on the partial lists which requires taking the list_lock. This is
4690 * less a concern for large slabs though which are rarely used.
4692 * slab_max_order specifies the order where we begin to stop considering the
4693 * number of objects in a slab as critical. If we reach slab_max_order then
4694 * we try to keep the page order as low as possible. So we accept more waste
4695 * of space in favor of a small page order.
4697 * Higher order allocations also allow the placement of more objects in a
4698 * slab and thereby reduce object handling overhead. If the user has
4699 * requested a higher minimum order then we start with that one instead of
4700 * the smallest order which will fit the object.
4702 static inline unsigned int calc_slab_order(unsigned int size,
4703 unsigned int min_order, unsigned int max_order,
4704 unsigned int fract_leftover)
4708 for (order = min_order; order <= max_order; order++) {
4710 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4713 rem = slab_size % size;
4715 if (rem <= slab_size / fract_leftover)
4722 static inline int calculate_order(unsigned int size)
4725 unsigned int min_objects;
4726 unsigned int max_objects;
4727 unsigned int min_order;
4729 min_objects = slub_min_objects;
4732 * Some architectures will only update present cpus when
4733 * onlining them, so don't trust the number if it's just 1. But
4734 * we also don't want to use nr_cpu_ids always, as on some other
4735 * architectures, there can be many possible cpus, but never
4736 * onlined. Here we compromise between trying to avoid too high
4737 * order on systems that appear larger than they are, and too
4738 * low order on systems that appear smaller than they are.
4740 unsigned int nr_cpus = num_present_cpus();
4742 nr_cpus = nr_cpu_ids;
4743 min_objects = 4 * (fls(nr_cpus) + 1);
4745 /* min_objects can't be 0 because get_order(0) is undefined */
4746 max_objects = max(order_objects(slub_max_order, size), 1U);
4747 min_objects = min(min_objects, max_objects);
4749 min_order = max_t(unsigned int, slub_min_order,
4750 get_order(min_objects * size));
4751 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4752 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4755 * Attempt to find best configuration for a slab. This works by first
4756 * attempting to generate a layout with the best possible configuration
4757 * and backing off gradually.
4759 * We start with accepting at most 1/16 waste and try to find the
4760 * smallest order from min_objects-derived/slab_min_order up to
4761 * slab_max_order that will satisfy the constraint. Note that increasing
4762 * the order can only result in same or less fractional waste, not more.
4764 * If that fails, we increase the acceptable fraction of waste and try
4765 * again. The last iteration with fraction of 1/2 would effectively
4766 * accept any waste and give us the order determined by min_objects, as
4767 * long as at least single object fits within slab_max_order.
4769 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
4770 order = calc_slab_order(size, min_order, slub_max_order,
4772 if (order <= slub_max_order)
4777 * Doh this slab cannot be placed using slab_max_order.
4779 order = get_order(size);
4780 if (order <= MAX_PAGE_ORDER)
4786 init_kmem_cache_node(struct kmem_cache_node *n)
4789 spin_lock_init(&n->list_lock);
4790 INIT_LIST_HEAD(&n->partial);
4791 #ifdef CONFIG_SLUB_DEBUG
4792 atomic_long_set(&n->nr_slabs, 0);
4793 atomic_long_set(&n->total_objects, 0);
4794 INIT_LIST_HEAD(&n->full);
4798 #ifndef CONFIG_SLUB_TINY
4799 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4801 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4802 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4803 sizeof(struct kmem_cache_cpu));
4806 * Must align to double word boundary for the double cmpxchg
4807 * instructions to work; see __pcpu_double_call_return_bool().
4809 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4810 2 * sizeof(void *));
4815 init_kmem_cache_cpus(s);
4820 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4824 #endif /* CONFIG_SLUB_TINY */
4826 static struct kmem_cache *kmem_cache_node;
4829 * No kmalloc_node yet so do it by hand. We know that this is the first
4830 * slab on the node for this slabcache. There are no concurrent accesses
4833 * Note that this function only works on the kmem_cache_node
4834 * when allocating for the kmem_cache_node. This is used for bootstrapping
4835 * memory on a fresh node that has no slab structures yet.
4837 static void early_kmem_cache_node_alloc(int node)
4840 struct kmem_cache_node *n;
4842 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4844 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4847 if (slab_nid(slab) != node) {
4848 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4849 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4854 #ifdef CONFIG_SLUB_DEBUG
4855 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4856 init_tracking(kmem_cache_node, n);
4858 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4859 slab->freelist = get_freepointer(kmem_cache_node, n);
4861 kmem_cache_node->node[node] = n;
4862 init_kmem_cache_node(n);
4863 inc_slabs_node(kmem_cache_node, node, slab->objects);
4866 * No locks need to be taken here as it has just been
4867 * initialized and there is no concurrent access.
4869 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4872 static void free_kmem_cache_nodes(struct kmem_cache *s)
4875 struct kmem_cache_node *n;
4877 for_each_kmem_cache_node(s, node, n) {
4878 s->node[node] = NULL;
4879 kmem_cache_free(kmem_cache_node, n);
4883 void __kmem_cache_release(struct kmem_cache *s)
4885 cache_random_seq_destroy(s);
4886 #ifndef CONFIG_SLUB_TINY
4887 free_percpu(s->cpu_slab);
4889 free_kmem_cache_nodes(s);
4892 static int init_kmem_cache_nodes(struct kmem_cache *s)
4896 for_each_node_mask(node, slab_nodes) {
4897 struct kmem_cache_node *n;
4899 if (slab_state == DOWN) {
4900 early_kmem_cache_node_alloc(node);
4903 n = kmem_cache_alloc_node(kmem_cache_node,
4907 free_kmem_cache_nodes(s);
4911 init_kmem_cache_node(n);
4917 static void set_cpu_partial(struct kmem_cache *s)
4919 #ifdef CONFIG_SLUB_CPU_PARTIAL
4920 unsigned int nr_objects;
4923 * cpu_partial determined the maximum number of objects kept in the
4924 * per cpu partial lists of a processor.
4926 * Per cpu partial lists mainly contain slabs that just have one
4927 * object freed. If they are used for allocation then they can be
4928 * filled up again with minimal effort. The slab will never hit the
4929 * per node partial lists and therefore no locking will be required.
4931 * For backwards compatibility reasons, this is determined as number
4932 * of objects, even though we now limit maximum number of pages, see
4933 * slub_set_cpu_partial()
4935 if (!kmem_cache_has_cpu_partial(s))
4937 else if (s->size >= PAGE_SIZE)
4939 else if (s->size >= 1024)
4941 else if (s->size >= 256)
4946 slub_set_cpu_partial(s, nr_objects);
4951 * calculate_sizes() determines the order and the distribution of data within
4954 static int calculate_sizes(struct kmem_cache *s)
4956 slab_flags_t flags = s->flags;
4957 unsigned int size = s->object_size;
4961 * Round up object size to the next word boundary. We can only
4962 * place the free pointer at word boundaries and this determines
4963 * the possible location of the free pointer.
4965 size = ALIGN(size, sizeof(void *));
4967 #ifdef CONFIG_SLUB_DEBUG
4969 * Determine if we can poison the object itself. If the user of
4970 * the slab may touch the object after free or before allocation
4971 * then we should never poison the object itself.
4973 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4975 s->flags |= __OBJECT_POISON;
4977 s->flags &= ~__OBJECT_POISON;
4981 * If we are Redzoning then check if there is some space between the
4982 * end of the object and the free pointer. If not then add an
4983 * additional word to have some bytes to store Redzone information.
4985 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4986 size += sizeof(void *);
4990 * With that we have determined the number of bytes in actual use
4991 * by the object and redzoning.
4995 if (slub_debug_orig_size(s) ||
4996 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4997 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
5000 * Relocate free pointer after the object if it is not
5001 * permitted to overwrite the first word of the object on
5004 * This is the case if we do RCU, have a constructor or
5005 * destructor, are poisoning the objects, or are
5006 * redzoning an object smaller than sizeof(void *).
5008 * The assumption that s->offset >= s->inuse means free
5009 * pointer is outside of the object is used in the
5010 * freeptr_outside_object() function. If that is no
5011 * longer true, the function needs to be modified.
5014 size += sizeof(void *);
5017 * Store freelist pointer near middle of object to keep
5018 * it away from the edges of the object to avoid small
5019 * sized over/underflows from neighboring allocations.
5021 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5024 #ifdef CONFIG_SLUB_DEBUG
5025 if (flags & SLAB_STORE_USER) {
5027 * Need to store information about allocs and frees after
5030 size += 2 * sizeof(struct track);
5032 /* Save the original kmalloc request size */
5033 if (flags & SLAB_KMALLOC)
5034 size += sizeof(unsigned int);
5038 kasan_cache_create(s, &size, &s->flags);
5039 #ifdef CONFIG_SLUB_DEBUG
5040 if (flags & SLAB_RED_ZONE) {
5042 * Add some empty padding so that we can catch
5043 * overwrites from earlier objects rather than let
5044 * tracking information or the free pointer be
5045 * corrupted if a user writes before the start
5048 size += sizeof(void *);
5050 s->red_left_pad = sizeof(void *);
5051 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5052 size += s->red_left_pad;
5057 * SLUB stores one object immediately after another beginning from
5058 * offset 0. In order to align the objects we have to simply size
5059 * each object to conform to the alignment.
5061 size = ALIGN(size, s->align);
5063 s->reciprocal_size = reciprocal_value(size);
5064 order = calculate_order(size);
5071 s->allocflags |= __GFP_COMP;
5073 if (s->flags & SLAB_CACHE_DMA)
5074 s->allocflags |= GFP_DMA;
5076 if (s->flags & SLAB_CACHE_DMA32)
5077 s->allocflags |= GFP_DMA32;
5079 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5080 s->allocflags |= __GFP_RECLAIMABLE;
5083 * Determine the number of objects per slab
5085 s->oo = oo_make(order, size);
5086 s->min = oo_make(get_order(size), size);
5088 return !!oo_objects(s->oo);
5091 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
5093 s->flags = kmem_cache_flags(flags, s->name);
5094 #ifdef CONFIG_SLAB_FREELIST_HARDENED
5095 s->random = get_random_long();
5098 if (!calculate_sizes(s))
5100 if (disable_higher_order_debug) {
5102 * Disable debugging flags that store metadata if the min slab
5105 if (get_order(s->size) > get_order(s->object_size)) {
5106 s->flags &= ~DEBUG_METADATA_FLAGS;
5108 if (!calculate_sizes(s))
5113 #ifdef system_has_freelist_aba
5114 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
5115 /* Enable fast mode */
5116 s->flags |= __CMPXCHG_DOUBLE;
5121 * The larger the object size is, the more slabs we want on the partial
5122 * list to avoid pounding the page allocator excessively.
5124 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
5125 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
5130 s->remote_node_defrag_ratio = 1000;
5133 /* Initialize the pre-computed randomized freelist if slab is up */
5134 if (slab_state >= UP) {
5135 if (init_cache_random_seq(s))
5139 if (!init_kmem_cache_nodes(s))
5142 if (alloc_kmem_cache_cpus(s))
5146 __kmem_cache_release(s);
5150 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5153 #ifdef CONFIG_SLUB_DEBUG
5154 void *addr = slab_address(slab);
5157 slab_err(s, slab, text, s->name);
5159 spin_lock(&object_map_lock);
5160 __fill_map(object_map, s, slab);
5162 for_each_object(p, s, addr, slab->objects) {
5164 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5165 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5166 print_tracking(s, p);
5169 spin_unlock(&object_map_lock);
5174 * Attempt to free all partial slabs on a node.
5175 * This is called from __kmem_cache_shutdown(). We must take list_lock
5176 * because sysfs file might still access partial list after the shutdowning.
5178 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5181 struct slab *slab, *h;
5183 BUG_ON(irqs_disabled());
5184 spin_lock_irq(&n->list_lock);
5185 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5187 remove_partial(n, slab);
5188 list_add(&slab->slab_list, &discard);
5190 list_slab_objects(s, slab,
5191 "Objects remaining in %s on __kmem_cache_shutdown()");
5194 spin_unlock_irq(&n->list_lock);
5196 list_for_each_entry_safe(slab, h, &discard, slab_list)
5197 discard_slab(s, slab);
5200 bool __kmem_cache_empty(struct kmem_cache *s)
5203 struct kmem_cache_node *n;
5205 for_each_kmem_cache_node(s, node, n)
5206 if (n->nr_partial || node_nr_slabs(n))
5212 * Release all resources used by a slab cache.
5214 int __kmem_cache_shutdown(struct kmem_cache *s)
5217 struct kmem_cache_node *n;
5219 flush_all_cpus_locked(s);
5220 /* Attempt to free all objects */
5221 for_each_kmem_cache_node(s, node, n) {
5223 if (n->nr_partial || node_nr_slabs(n))
5229 #ifdef CONFIG_PRINTK
5230 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5233 int __maybe_unused i;
5237 struct kmem_cache *s = slab->slab_cache;
5238 struct track __maybe_unused *trackp;
5240 kpp->kp_ptr = object;
5241 kpp->kp_slab = slab;
5242 kpp->kp_slab_cache = s;
5243 base = slab_address(slab);
5244 objp0 = kasan_reset_tag(object);
5245 #ifdef CONFIG_SLUB_DEBUG
5246 objp = restore_red_left(s, objp0);
5250 objnr = obj_to_index(s, slab, objp);
5251 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5252 objp = base + s->size * objnr;
5253 kpp->kp_objp = objp;
5254 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5255 || (objp - base) % s->size) ||
5256 !(s->flags & SLAB_STORE_USER))
5258 #ifdef CONFIG_SLUB_DEBUG
5259 objp = fixup_red_left(s, objp);
5260 trackp = get_track(s, objp, TRACK_ALLOC);
5261 kpp->kp_ret = (void *)trackp->addr;
5262 #ifdef CONFIG_STACKDEPOT
5264 depot_stack_handle_t handle;
5265 unsigned long *entries;
5266 unsigned int nr_entries;
5268 handle = READ_ONCE(trackp->handle);
5270 nr_entries = stack_depot_fetch(handle, &entries);
5271 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5272 kpp->kp_stack[i] = (void *)entries[i];
5275 trackp = get_track(s, objp, TRACK_FREE);
5276 handle = READ_ONCE(trackp->handle);
5278 nr_entries = stack_depot_fetch(handle, &entries);
5279 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5280 kpp->kp_free_stack[i] = (void *)entries[i];
5288 /********************************************************************
5290 *******************************************************************/
5292 static int __init setup_slub_min_order(char *str)
5294 get_option(&str, (int *)&slub_min_order);
5296 if (slub_min_order > slub_max_order)
5297 slub_max_order = slub_min_order;
5302 __setup("slab_min_order=", setup_slub_min_order);
5303 __setup_param("slub_min_order=", slub_min_order, setup_slub_min_order, 0);
5306 static int __init setup_slub_max_order(char *str)
5308 get_option(&str, (int *)&slub_max_order);
5309 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5311 if (slub_min_order > slub_max_order)
5312 slub_min_order = slub_max_order;
5317 __setup("slab_max_order=", setup_slub_max_order);
5318 __setup_param("slub_max_order=", slub_max_order, setup_slub_max_order, 0);
5320 static int __init setup_slub_min_objects(char *str)
5322 get_option(&str, (int *)&slub_min_objects);
5327 __setup("slab_min_objects=", setup_slub_min_objects);
5328 __setup_param("slub_min_objects=", slub_min_objects, setup_slub_min_objects, 0);
5330 #ifdef CONFIG_HARDENED_USERCOPY
5332 * Rejects incorrectly sized objects and objects that are to be copied
5333 * to/from userspace but do not fall entirely within the containing slab
5334 * cache's usercopy region.
5336 * Returns NULL if check passes, otherwise const char * to name of cache
5337 * to indicate an error.
5339 void __check_heap_object(const void *ptr, unsigned long n,
5340 const struct slab *slab, bool to_user)
5342 struct kmem_cache *s;
5343 unsigned int offset;
5344 bool is_kfence = is_kfence_address(ptr);
5346 ptr = kasan_reset_tag(ptr);
5348 /* Find object and usable object size. */
5349 s = slab->slab_cache;
5351 /* Reject impossible pointers. */
5352 if (ptr < slab_address(slab))
5353 usercopy_abort("SLUB object not in SLUB page?!", NULL,
5356 /* Find offset within object. */
5358 offset = ptr - kfence_object_start(ptr);
5360 offset = (ptr - slab_address(slab)) % s->size;
5362 /* Adjust for redzone and reject if within the redzone. */
5363 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5364 if (offset < s->red_left_pad)
5365 usercopy_abort("SLUB object in left red zone",
5366 s->name, to_user, offset, n);
5367 offset -= s->red_left_pad;
5370 /* Allow address range falling entirely within usercopy region. */
5371 if (offset >= s->useroffset &&
5372 offset - s->useroffset <= s->usersize &&
5373 n <= s->useroffset - offset + s->usersize)
5376 usercopy_abort("SLUB object", s->name, to_user, offset, n);
5378 #endif /* CONFIG_HARDENED_USERCOPY */
5380 #define SHRINK_PROMOTE_MAX 32
5383 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5384 * up most to the head of the partial lists. New allocations will then
5385 * fill those up and thus they can be removed from the partial lists.
5387 * The slabs with the least items are placed last. This results in them
5388 * being allocated from last increasing the chance that the last objects
5389 * are freed in them.
5391 static int __kmem_cache_do_shrink(struct kmem_cache *s)
5395 struct kmem_cache_node *n;
5398 struct list_head discard;
5399 struct list_head promote[SHRINK_PROMOTE_MAX];
5400 unsigned long flags;
5403 for_each_kmem_cache_node(s, node, n) {
5404 INIT_LIST_HEAD(&discard);
5405 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5406 INIT_LIST_HEAD(promote + i);
5408 spin_lock_irqsave(&n->list_lock, flags);
5411 * Build lists of slabs to discard or promote.
5413 * Note that concurrent frees may occur while we hold the
5414 * list_lock. slab->inuse here is the upper limit.
5416 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5417 int free = slab->objects - slab->inuse;
5419 /* Do not reread slab->inuse */
5422 /* We do not keep full slabs on the list */
5425 if (free == slab->objects) {
5426 list_move(&slab->slab_list, &discard);
5427 slab_clear_node_partial(slab);
5429 dec_slabs_node(s, node, slab->objects);
5430 } else if (free <= SHRINK_PROMOTE_MAX)
5431 list_move(&slab->slab_list, promote + free - 1);
5435 * Promote the slabs filled up most to the head of the
5438 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5439 list_splice(promote + i, &n->partial);
5441 spin_unlock_irqrestore(&n->list_lock, flags);
5443 /* Release empty slabs */
5444 list_for_each_entry_safe(slab, t, &discard, slab_list)
5447 if (node_nr_slabs(n))
5454 int __kmem_cache_shrink(struct kmem_cache *s)
5457 return __kmem_cache_do_shrink(s);
5460 static int slab_mem_going_offline_callback(void *arg)
5462 struct kmem_cache *s;
5464 mutex_lock(&slab_mutex);
5465 list_for_each_entry(s, &slab_caches, list) {
5466 flush_all_cpus_locked(s);
5467 __kmem_cache_do_shrink(s);
5469 mutex_unlock(&slab_mutex);
5474 static void slab_mem_offline_callback(void *arg)
5476 struct memory_notify *marg = arg;
5479 offline_node = marg->status_change_nid_normal;
5482 * If the node still has available memory. we need kmem_cache_node
5485 if (offline_node < 0)
5488 mutex_lock(&slab_mutex);
5489 node_clear(offline_node, slab_nodes);
5491 * We no longer free kmem_cache_node structures here, as it would be
5492 * racy with all get_node() users, and infeasible to protect them with
5495 mutex_unlock(&slab_mutex);
5498 static int slab_mem_going_online_callback(void *arg)
5500 struct kmem_cache_node *n;
5501 struct kmem_cache *s;
5502 struct memory_notify *marg = arg;
5503 int nid = marg->status_change_nid_normal;
5507 * If the node's memory is already available, then kmem_cache_node is
5508 * already created. Nothing to do.
5514 * We are bringing a node online. No memory is available yet. We must
5515 * allocate a kmem_cache_node structure in order to bring the node
5518 mutex_lock(&slab_mutex);
5519 list_for_each_entry(s, &slab_caches, list) {
5521 * The structure may already exist if the node was previously
5522 * onlined and offlined.
5524 if (get_node(s, nid))
5527 * XXX: kmem_cache_alloc_node will fallback to other nodes
5528 * since memory is not yet available from the node that
5531 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5536 init_kmem_cache_node(n);
5540 * Any cache created after this point will also have kmem_cache_node
5541 * initialized for the new node.
5543 node_set(nid, slab_nodes);
5545 mutex_unlock(&slab_mutex);
5549 static int slab_memory_callback(struct notifier_block *self,
5550 unsigned long action, void *arg)
5555 case MEM_GOING_ONLINE:
5556 ret = slab_mem_going_online_callback(arg);
5558 case MEM_GOING_OFFLINE:
5559 ret = slab_mem_going_offline_callback(arg);
5562 case MEM_CANCEL_ONLINE:
5563 slab_mem_offline_callback(arg);
5566 case MEM_CANCEL_OFFLINE:
5570 ret = notifier_from_errno(ret);
5576 /********************************************************************
5577 * Basic setup of slabs
5578 *******************************************************************/
5581 * Used for early kmem_cache structures that were allocated using
5582 * the page allocator. Allocate them properly then fix up the pointers
5583 * that may be pointing to the wrong kmem_cache structure.
5586 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
5589 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
5590 struct kmem_cache_node *n;
5592 memcpy(s, static_cache, kmem_cache->object_size);
5595 * This runs very early, and only the boot processor is supposed to be
5596 * up. Even if it weren't true, IRQs are not up so we couldn't fire
5599 __flush_cpu_slab(s, smp_processor_id());
5600 for_each_kmem_cache_node(s, node, n) {
5603 list_for_each_entry(p, &n->partial, slab_list)
5606 #ifdef CONFIG_SLUB_DEBUG
5607 list_for_each_entry(p, &n->full, slab_list)
5611 list_add(&s->list, &slab_caches);
5615 void __init kmem_cache_init(void)
5617 static __initdata struct kmem_cache boot_kmem_cache,
5618 boot_kmem_cache_node;
5621 if (debug_guardpage_minorder())
5624 /* Print slub debugging pointers without hashing */
5625 if (__slub_debug_enabled())
5626 no_hash_pointers_enable(NULL);
5628 kmem_cache_node = &boot_kmem_cache_node;
5629 kmem_cache = &boot_kmem_cache;
5632 * Initialize the nodemask for which we will allocate per node
5633 * structures. Here we don't need taking slab_mutex yet.
5635 for_each_node_state(node, N_NORMAL_MEMORY)
5636 node_set(node, slab_nodes);
5638 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5639 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5641 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5643 /* Able to allocate the per node structures */
5644 slab_state = PARTIAL;
5646 create_boot_cache(kmem_cache, "kmem_cache",
5647 offsetof(struct kmem_cache, node) +
5648 nr_node_ids * sizeof(struct kmem_cache_node *),
5649 SLAB_HWCACHE_ALIGN, 0, 0);
5651 kmem_cache = bootstrap(&boot_kmem_cache);
5652 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5654 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5655 setup_kmalloc_cache_index_table();
5656 create_kmalloc_caches();
5658 /* Setup random freelists for each cache */
5659 init_freelist_randomization();
5661 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5664 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5666 slub_min_order, slub_max_order, slub_min_objects,
5667 nr_cpu_ids, nr_node_ids);
5670 void __init kmem_cache_init_late(void)
5672 #ifndef CONFIG_SLUB_TINY
5673 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5679 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5680 slab_flags_t flags, void (*ctor)(void *))
5682 struct kmem_cache *s;
5684 s = find_mergeable(size, align, flags, name, ctor);
5686 if (sysfs_slab_alias(s, name))
5692 * Adjust the object sizes so that we clear
5693 * the complete object on kzalloc.
5695 s->object_size = max(s->object_size, size);
5696 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5702 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5706 err = kmem_cache_open(s, flags);
5710 /* Mutex is not taken during early boot */
5711 if (slab_state <= UP)
5714 err = sysfs_slab_add(s);
5716 __kmem_cache_release(s);
5720 if (s->flags & SLAB_STORE_USER)
5721 debugfs_slab_add(s);
5726 #ifdef SLAB_SUPPORTS_SYSFS
5727 static int count_inuse(struct slab *slab)
5732 static int count_total(struct slab *slab)
5734 return slab->objects;
5738 #ifdef CONFIG_SLUB_DEBUG
5739 static void validate_slab(struct kmem_cache *s, struct slab *slab,
5740 unsigned long *obj_map)
5743 void *addr = slab_address(slab);
5745 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5748 /* Now we know that a valid freelist exists */
5749 __fill_map(obj_map, s, slab);
5750 for_each_object(p, s, addr, slab->objects) {
5751 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5752 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5754 if (!check_object(s, slab, p, val))
5759 static int validate_slab_node(struct kmem_cache *s,
5760 struct kmem_cache_node *n, unsigned long *obj_map)
5762 unsigned long count = 0;
5764 unsigned long flags;
5766 spin_lock_irqsave(&n->list_lock, flags);
5768 list_for_each_entry(slab, &n->partial, slab_list) {
5769 validate_slab(s, slab, obj_map);
5772 if (count != n->nr_partial) {
5773 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5774 s->name, count, n->nr_partial);
5775 slab_add_kunit_errors();
5778 if (!(s->flags & SLAB_STORE_USER))
5781 list_for_each_entry(slab, &n->full, slab_list) {
5782 validate_slab(s, slab, obj_map);
5785 if (count != node_nr_slabs(n)) {
5786 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5787 s->name, count, node_nr_slabs(n));
5788 slab_add_kunit_errors();
5792 spin_unlock_irqrestore(&n->list_lock, flags);
5796 long validate_slab_cache(struct kmem_cache *s)
5799 unsigned long count = 0;
5800 struct kmem_cache_node *n;
5801 unsigned long *obj_map;
5803 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5808 for_each_kmem_cache_node(s, node, n)
5809 count += validate_slab_node(s, n, obj_map);
5811 bitmap_free(obj_map);
5815 EXPORT_SYMBOL(validate_slab_cache);
5817 #ifdef CONFIG_DEBUG_FS
5819 * Generate lists of code addresses where slabcache objects are allocated
5824 depot_stack_handle_t handle;
5825 unsigned long count;
5827 unsigned long waste;
5833 DECLARE_BITMAP(cpus, NR_CPUS);
5839 unsigned long count;
5840 struct location *loc;
5844 static struct dentry *slab_debugfs_root;
5846 static void free_loc_track(struct loc_track *t)
5849 free_pages((unsigned long)t->loc,
5850 get_order(sizeof(struct location) * t->max));
5853 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5858 order = get_order(sizeof(struct location) * max);
5860 l = (void *)__get_free_pages(flags, order);
5865 memcpy(l, t->loc, sizeof(struct location) * t->count);
5873 static int add_location(struct loc_track *t, struct kmem_cache *s,
5874 const struct track *track,
5875 unsigned int orig_size)
5877 long start, end, pos;
5879 unsigned long caddr, chandle, cwaste;
5880 unsigned long age = jiffies - track->when;
5881 depot_stack_handle_t handle = 0;
5882 unsigned int waste = s->object_size - orig_size;
5884 #ifdef CONFIG_STACKDEPOT
5885 handle = READ_ONCE(track->handle);
5891 pos = start + (end - start + 1) / 2;
5894 * There is nothing at "end". If we end up there
5895 * we need to add something to before end.
5902 chandle = l->handle;
5904 if ((track->addr == caddr) && (handle == chandle) &&
5905 (waste == cwaste)) {
5910 if (age < l->min_time)
5912 if (age > l->max_time)
5915 if (track->pid < l->min_pid)
5916 l->min_pid = track->pid;
5917 if (track->pid > l->max_pid)
5918 l->max_pid = track->pid;
5920 cpumask_set_cpu(track->cpu,
5921 to_cpumask(l->cpus));
5923 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5927 if (track->addr < caddr)
5929 else if (track->addr == caddr && handle < chandle)
5931 else if (track->addr == caddr && handle == chandle &&
5939 * Not found. Insert new tracking element.
5941 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5947 (t->count - pos) * sizeof(struct location));
5950 l->addr = track->addr;
5954 l->min_pid = track->pid;
5955 l->max_pid = track->pid;
5958 cpumask_clear(to_cpumask(l->cpus));
5959 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5960 nodes_clear(l->nodes);
5961 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5965 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5966 struct slab *slab, enum track_item alloc,
5967 unsigned long *obj_map)
5969 void *addr = slab_address(slab);
5970 bool is_alloc = (alloc == TRACK_ALLOC);
5973 __fill_map(obj_map, s, slab);
5975 for_each_object(p, s, addr, slab->objects)
5976 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5977 add_location(t, s, get_track(s, p, alloc),
5978 is_alloc ? get_orig_size(s, p) :
5981 #endif /* CONFIG_DEBUG_FS */
5982 #endif /* CONFIG_SLUB_DEBUG */
5984 #ifdef SLAB_SUPPORTS_SYSFS
5985 enum slab_stat_type {
5986 SL_ALL, /* All slabs */
5987 SL_PARTIAL, /* Only partially allocated slabs */
5988 SL_CPU, /* Only slabs used for cpu caches */
5989 SL_OBJECTS, /* Determine allocated objects not slabs */
5990 SL_TOTAL /* Determine object capacity not slabs */
5993 #define SO_ALL (1 << SL_ALL)
5994 #define SO_PARTIAL (1 << SL_PARTIAL)
5995 #define SO_CPU (1 << SL_CPU)
5996 #define SO_OBJECTS (1 << SL_OBJECTS)
5997 #define SO_TOTAL (1 << SL_TOTAL)
5999 static ssize_t show_slab_objects(struct kmem_cache *s,
6000 char *buf, unsigned long flags)
6002 unsigned long total = 0;
6005 unsigned long *nodes;
6008 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6012 if (flags & SO_CPU) {
6015 for_each_possible_cpu(cpu) {
6016 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6021 slab = READ_ONCE(c->slab);
6025 node = slab_nid(slab);
6026 if (flags & SO_TOTAL)
6028 else if (flags & SO_OBJECTS)
6036 #ifdef CONFIG_SLUB_CPU_PARTIAL
6037 slab = slub_percpu_partial_read_once(c);
6039 node = slab_nid(slab);
6040 if (flags & SO_TOTAL)
6042 else if (flags & SO_OBJECTS)
6054 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6055 * already held which will conflict with an existing lock order:
6057 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6059 * We don't really need mem_hotplug_lock (to hold off
6060 * slab_mem_going_offline_callback) here because slab's memory hot
6061 * unplug code doesn't destroy the kmem_cache->node[] data.
6064 #ifdef CONFIG_SLUB_DEBUG
6065 if (flags & SO_ALL) {
6066 struct kmem_cache_node *n;
6068 for_each_kmem_cache_node(s, node, n) {
6070 if (flags & SO_TOTAL)
6071 x = node_nr_objs(n);
6072 else if (flags & SO_OBJECTS)
6073 x = node_nr_objs(n) - count_partial(n, count_free);
6075 x = node_nr_slabs(n);
6082 if (flags & SO_PARTIAL) {
6083 struct kmem_cache_node *n;
6085 for_each_kmem_cache_node(s, node, n) {
6086 if (flags & SO_TOTAL)
6087 x = count_partial(n, count_total);
6088 else if (flags & SO_OBJECTS)
6089 x = count_partial(n, count_inuse);
6097 len += sysfs_emit_at(buf, len, "%lu", total);
6099 for (node = 0; node < nr_node_ids; node++) {
6101 len += sysfs_emit_at(buf, len, " N%d=%lu",
6105 len += sysfs_emit_at(buf, len, "\n");
6111 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6112 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
6114 struct slab_attribute {
6115 struct attribute attr;
6116 ssize_t (*show)(struct kmem_cache *s, char *buf);
6117 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6120 #define SLAB_ATTR_RO(_name) \
6121 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6123 #define SLAB_ATTR(_name) \
6124 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6126 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6128 return sysfs_emit(buf, "%u\n", s->size);
6130 SLAB_ATTR_RO(slab_size);
6132 static ssize_t align_show(struct kmem_cache *s, char *buf)
6134 return sysfs_emit(buf, "%u\n", s->align);
6136 SLAB_ATTR_RO(align);
6138 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6140 return sysfs_emit(buf, "%u\n", s->object_size);
6142 SLAB_ATTR_RO(object_size);
6144 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6146 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6148 SLAB_ATTR_RO(objs_per_slab);
6150 static ssize_t order_show(struct kmem_cache *s, char *buf)
6152 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6154 SLAB_ATTR_RO(order);
6156 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6158 return sysfs_emit(buf, "%lu\n", s->min_partial);
6161 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6167 err = kstrtoul(buf, 10, &min);
6171 s->min_partial = min;
6174 SLAB_ATTR(min_partial);
6176 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6178 unsigned int nr_partial = 0;
6179 #ifdef CONFIG_SLUB_CPU_PARTIAL
6180 nr_partial = s->cpu_partial;
6183 return sysfs_emit(buf, "%u\n", nr_partial);
6186 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6189 unsigned int objects;
6192 err = kstrtouint(buf, 10, &objects);
6195 if (objects && !kmem_cache_has_cpu_partial(s))
6198 slub_set_cpu_partial(s, objects);
6202 SLAB_ATTR(cpu_partial);
6204 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6208 return sysfs_emit(buf, "%pS\n", s->ctor);
6212 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6214 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6216 SLAB_ATTR_RO(aliases);
6218 static ssize_t partial_show(struct kmem_cache *s, char *buf)
6220 return show_slab_objects(s, buf, SO_PARTIAL);
6222 SLAB_ATTR_RO(partial);
6224 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6226 return show_slab_objects(s, buf, SO_CPU);
6228 SLAB_ATTR_RO(cpu_slabs);
6230 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6232 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6234 SLAB_ATTR_RO(objects_partial);
6236 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6240 int cpu __maybe_unused;
6243 #ifdef CONFIG_SLUB_CPU_PARTIAL
6244 for_each_online_cpu(cpu) {
6247 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6250 slabs += slab->slabs;
6254 /* Approximate half-full slabs, see slub_set_cpu_partial() */
6255 objects = (slabs * oo_objects(s->oo)) / 2;
6256 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6258 #ifdef CONFIG_SLUB_CPU_PARTIAL
6259 for_each_online_cpu(cpu) {
6262 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6264 slabs = READ_ONCE(slab->slabs);
6265 objects = (slabs * oo_objects(s->oo)) / 2;
6266 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6267 cpu, objects, slabs);
6271 len += sysfs_emit_at(buf, len, "\n");
6275 SLAB_ATTR_RO(slabs_cpu_partial);
6277 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6279 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6281 SLAB_ATTR_RO(reclaim_account);
6283 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6285 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6287 SLAB_ATTR_RO(hwcache_align);
6289 #ifdef CONFIG_ZONE_DMA
6290 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6292 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6294 SLAB_ATTR_RO(cache_dma);
6297 #ifdef CONFIG_HARDENED_USERCOPY
6298 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6300 return sysfs_emit(buf, "%u\n", s->usersize);
6302 SLAB_ATTR_RO(usersize);
6305 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6307 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6309 SLAB_ATTR_RO(destroy_by_rcu);
6311 #ifdef CONFIG_SLUB_DEBUG
6312 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6314 return show_slab_objects(s, buf, SO_ALL);
6316 SLAB_ATTR_RO(slabs);
6318 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6320 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6322 SLAB_ATTR_RO(total_objects);
6324 static ssize_t objects_show(struct kmem_cache *s, char *buf)
6326 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6328 SLAB_ATTR_RO(objects);
6330 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6332 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6334 SLAB_ATTR_RO(sanity_checks);
6336 static ssize_t trace_show(struct kmem_cache *s, char *buf)
6338 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6340 SLAB_ATTR_RO(trace);
6342 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6344 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6347 SLAB_ATTR_RO(red_zone);
6349 static ssize_t poison_show(struct kmem_cache *s, char *buf)
6351 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6354 SLAB_ATTR_RO(poison);
6356 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6358 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6361 SLAB_ATTR_RO(store_user);
6363 static ssize_t validate_show(struct kmem_cache *s, char *buf)
6368 static ssize_t validate_store(struct kmem_cache *s,
6369 const char *buf, size_t length)
6373 if (buf[0] == '1' && kmem_cache_debug(s)) {
6374 ret = validate_slab_cache(s);
6380 SLAB_ATTR(validate);
6382 #endif /* CONFIG_SLUB_DEBUG */
6384 #ifdef CONFIG_FAILSLAB
6385 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6387 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6390 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6393 if (s->refcount > 1)
6397 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6399 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6403 SLAB_ATTR(failslab);
6406 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6411 static ssize_t shrink_store(struct kmem_cache *s,
6412 const char *buf, size_t length)
6415 kmem_cache_shrink(s);
6423 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6425 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6428 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6429 const char *buf, size_t length)
6434 err = kstrtouint(buf, 10, &ratio);
6440 s->remote_node_defrag_ratio = ratio * 10;
6444 SLAB_ATTR(remote_node_defrag_ratio);
6447 #ifdef CONFIG_SLUB_STATS
6448 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6450 unsigned long sum = 0;
6453 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6458 for_each_online_cpu(cpu) {
6459 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6465 len += sysfs_emit_at(buf, len, "%lu", sum);
6468 for_each_online_cpu(cpu) {
6470 len += sysfs_emit_at(buf, len, " C%d=%u",
6475 len += sysfs_emit_at(buf, len, "\n");
6480 static void clear_stat(struct kmem_cache *s, enum stat_item si)
6484 for_each_online_cpu(cpu)
6485 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6488 #define STAT_ATTR(si, text) \
6489 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
6491 return show_stat(s, buf, si); \
6493 static ssize_t text##_store(struct kmem_cache *s, \
6494 const char *buf, size_t length) \
6496 if (buf[0] != '0') \
6498 clear_stat(s, si); \
6503 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
6504 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
6505 STAT_ATTR(FREE_FASTPATH, free_fastpath);
6506 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
6507 STAT_ATTR(FREE_FROZEN, free_frozen);
6508 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
6509 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
6510 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
6511 STAT_ATTR(ALLOC_SLAB, alloc_slab);
6512 STAT_ATTR(ALLOC_REFILL, alloc_refill);
6513 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
6514 STAT_ATTR(FREE_SLAB, free_slab);
6515 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
6516 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
6517 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
6518 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
6519 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
6520 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
6521 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
6522 STAT_ATTR(ORDER_FALLBACK, order_fallback);
6523 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
6524 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
6525 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
6526 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
6527 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
6528 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
6529 #endif /* CONFIG_SLUB_STATS */
6531 #ifdef CONFIG_KFENCE
6532 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
6534 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
6537 static ssize_t skip_kfence_store(struct kmem_cache *s,
6538 const char *buf, size_t length)
6543 s->flags &= ~SLAB_SKIP_KFENCE;
6544 else if (buf[0] == '1')
6545 s->flags |= SLAB_SKIP_KFENCE;
6551 SLAB_ATTR(skip_kfence);
6554 static struct attribute *slab_attrs[] = {
6555 &slab_size_attr.attr,
6556 &object_size_attr.attr,
6557 &objs_per_slab_attr.attr,
6559 &min_partial_attr.attr,
6560 &cpu_partial_attr.attr,
6561 &objects_partial_attr.attr,
6563 &cpu_slabs_attr.attr,
6567 &hwcache_align_attr.attr,
6568 &reclaim_account_attr.attr,
6569 &destroy_by_rcu_attr.attr,
6571 &slabs_cpu_partial_attr.attr,
6572 #ifdef CONFIG_SLUB_DEBUG
6573 &total_objects_attr.attr,
6576 &sanity_checks_attr.attr,
6578 &red_zone_attr.attr,
6580 &store_user_attr.attr,
6581 &validate_attr.attr,
6583 #ifdef CONFIG_ZONE_DMA
6584 &cache_dma_attr.attr,
6587 &remote_node_defrag_ratio_attr.attr,
6589 #ifdef CONFIG_SLUB_STATS
6590 &alloc_fastpath_attr.attr,
6591 &alloc_slowpath_attr.attr,
6592 &free_fastpath_attr.attr,
6593 &free_slowpath_attr.attr,
6594 &free_frozen_attr.attr,
6595 &free_add_partial_attr.attr,
6596 &free_remove_partial_attr.attr,
6597 &alloc_from_partial_attr.attr,
6598 &alloc_slab_attr.attr,
6599 &alloc_refill_attr.attr,
6600 &alloc_node_mismatch_attr.attr,
6601 &free_slab_attr.attr,
6602 &cpuslab_flush_attr.attr,
6603 &deactivate_full_attr.attr,
6604 &deactivate_empty_attr.attr,
6605 &deactivate_to_head_attr.attr,
6606 &deactivate_to_tail_attr.attr,
6607 &deactivate_remote_frees_attr.attr,
6608 &deactivate_bypass_attr.attr,
6609 &order_fallback_attr.attr,
6610 &cmpxchg_double_fail_attr.attr,
6611 &cmpxchg_double_cpu_fail_attr.attr,
6612 &cpu_partial_alloc_attr.attr,
6613 &cpu_partial_free_attr.attr,
6614 &cpu_partial_node_attr.attr,
6615 &cpu_partial_drain_attr.attr,
6617 #ifdef CONFIG_FAILSLAB
6618 &failslab_attr.attr,
6620 #ifdef CONFIG_HARDENED_USERCOPY
6621 &usersize_attr.attr,
6623 #ifdef CONFIG_KFENCE
6624 &skip_kfence_attr.attr,
6630 static const struct attribute_group slab_attr_group = {
6631 .attrs = slab_attrs,
6634 static ssize_t slab_attr_show(struct kobject *kobj,
6635 struct attribute *attr,
6638 struct slab_attribute *attribute;
6639 struct kmem_cache *s;
6641 attribute = to_slab_attr(attr);
6644 if (!attribute->show)
6647 return attribute->show(s, buf);
6650 static ssize_t slab_attr_store(struct kobject *kobj,
6651 struct attribute *attr,
6652 const char *buf, size_t len)
6654 struct slab_attribute *attribute;
6655 struct kmem_cache *s;
6657 attribute = to_slab_attr(attr);
6660 if (!attribute->store)
6663 return attribute->store(s, buf, len);
6666 static void kmem_cache_release(struct kobject *k)
6668 slab_kmem_cache_release(to_slab(k));
6671 static const struct sysfs_ops slab_sysfs_ops = {
6672 .show = slab_attr_show,
6673 .store = slab_attr_store,
6676 static const struct kobj_type slab_ktype = {
6677 .sysfs_ops = &slab_sysfs_ops,
6678 .release = kmem_cache_release,
6681 static struct kset *slab_kset;
6683 static inline struct kset *cache_kset(struct kmem_cache *s)
6688 #define ID_STR_LENGTH 32
6690 /* Create a unique string id for a slab cache:
6692 * Format :[flags-]size
6694 static char *create_unique_id(struct kmem_cache *s)
6696 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6700 return ERR_PTR(-ENOMEM);
6704 * First flags affecting slabcache operations. We will only
6705 * get here for aliasable slabs so we do not need to support
6706 * too many flags. The flags here must cover all flags that
6707 * are matched during merging to guarantee that the id is
6710 if (s->flags & SLAB_CACHE_DMA)
6712 if (s->flags & SLAB_CACHE_DMA32)
6714 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6716 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6718 if (s->flags & SLAB_ACCOUNT)
6722 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6724 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6726 return ERR_PTR(-EINVAL);
6728 kmsan_unpoison_memory(name, p - name);
6732 static int sysfs_slab_add(struct kmem_cache *s)
6736 struct kset *kset = cache_kset(s);
6737 int unmergeable = slab_unmergeable(s);
6739 if (!unmergeable && disable_higher_order_debug &&
6740 (slub_debug & DEBUG_METADATA_FLAGS))
6745 * Slabcache can never be merged so we can use the name proper.
6746 * This is typically the case for debug situations. In that
6747 * case we can catch duplicate names easily.
6749 sysfs_remove_link(&slab_kset->kobj, s->name);
6753 * Create a unique name for the slab as a target
6756 name = create_unique_id(s);
6758 return PTR_ERR(name);
6761 s->kobj.kset = kset;
6762 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6766 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6771 /* Setup first alias */
6772 sysfs_slab_alias(s, s->name);
6779 kobject_del(&s->kobj);
6783 void sysfs_slab_unlink(struct kmem_cache *s)
6785 kobject_del(&s->kobj);
6788 void sysfs_slab_release(struct kmem_cache *s)
6790 kobject_put(&s->kobj);
6794 * Need to buffer aliases during bootup until sysfs becomes
6795 * available lest we lose that information.
6797 struct saved_alias {
6798 struct kmem_cache *s;
6800 struct saved_alias *next;
6803 static struct saved_alias *alias_list;
6805 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6807 struct saved_alias *al;
6809 if (slab_state == FULL) {
6811 * If we have a leftover link then remove it.
6813 sysfs_remove_link(&slab_kset->kobj, name);
6814 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6817 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6823 al->next = alias_list;
6825 kmsan_unpoison_memory(al, sizeof(*al));
6829 static int __init slab_sysfs_init(void)
6831 struct kmem_cache *s;
6834 mutex_lock(&slab_mutex);
6836 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6838 mutex_unlock(&slab_mutex);
6839 pr_err("Cannot register slab subsystem.\n");
6845 list_for_each_entry(s, &slab_caches, list) {
6846 err = sysfs_slab_add(s);
6848 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6852 while (alias_list) {
6853 struct saved_alias *al = alias_list;
6855 alias_list = alias_list->next;
6856 err = sysfs_slab_alias(al->s, al->name);
6858 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6863 mutex_unlock(&slab_mutex);
6866 late_initcall(slab_sysfs_init);
6867 #endif /* SLAB_SUPPORTS_SYSFS */
6869 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6870 static int slab_debugfs_show(struct seq_file *seq, void *v)
6872 struct loc_track *t = seq->private;
6876 idx = (unsigned long) t->idx;
6877 if (idx < t->count) {
6880 seq_printf(seq, "%7ld ", l->count);
6883 seq_printf(seq, "%pS", (void *)l->addr);
6885 seq_puts(seq, "<not-available>");
6888 seq_printf(seq, " waste=%lu/%lu",
6889 l->count * l->waste, l->waste);
6891 if (l->sum_time != l->min_time) {
6892 seq_printf(seq, " age=%ld/%llu/%ld",
6893 l->min_time, div_u64(l->sum_time, l->count),
6896 seq_printf(seq, " age=%ld", l->min_time);
6898 if (l->min_pid != l->max_pid)
6899 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6901 seq_printf(seq, " pid=%ld",
6904 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6905 seq_printf(seq, " cpus=%*pbl",
6906 cpumask_pr_args(to_cpumask(l->cpus)));
6908 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6909 seq_printf(seq, " nodes=%*pbl",
6910 nodemask_pr_args(&l->nodes));
6912 #ifdef CONFIG_STACKDEPOT
6914 depot_stack_handle_t handle;
6915 unsigned long *entries;
6916 unsigned int nr_entries, j;
6918 handle = READ_ONCE(l->handle);
6920 nr_entries = stack_depot_fetch(handle, &entries);
6921 seq_puts(seq, "\n");
6922 for (j = 0; j < nr_entries; j++)
6923 seq_printf(seq, " %pS\n", (void *)entries[j]);
6927 seq_puts(seq, "\n");
6930 if (!idx && !t->count)
6931 seq_puts(seq, "No data\n");
6936 static void slab_debugfs_stop(struct seq_file *seq, void *v)
6940 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6942 struct loc_track *t = seq->private;
6945 if (*ppos <= t->count)
6951 static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6953 struct location *loc1 = (struct location *)a;
6954 struct location *loc2 = (struct location *)b;
6956 if (loc1->count > loc2->count)
6962 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6964 struct loc_track *t = seq->private;
6970 static const struct seq_operations slab_debugfs_sops = {
6971 .start = slab_debugfs_start,
6972 .next = slab_debugfs_next,
6973 .stop = slab_debugfs_stop,
6974 .show = slab_debugfs_show,
6977 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6980 struct kmem_cache_node *n;
6981 enum track_item alloc;
6983 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6984 sizeof(struct loc_track));
6985 struct kmem_cache *s = file_inode(filep)->i_private;
6986 unsigned long *obj_map;
6991 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6993 seq_release_private(inode, filep);
6997 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6998 alloc = TRACK_ALLOC;
7002 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7003 bitmap_free(obj_map);
7004 seq_release_private(inode, filep);
7008 for_each_kmem_cache_node(s, node, n) {
7009 unsigned long flags;
7012 if (!node_nr_slabs(n))
7015 spin_lock_irqsave(&n->list_lock, flags);
7016 list_for_each_entry(slab, &n->partial, slab_list)
7017 process_slab(t, s, slab, alloc, obj_map);
7018 list_for_each_entry(slab, &n->full, slab_list)
7019 process_slab(t, s, slab, alloc, obj_map);
7020 spin_unlock_irqrestore(&n->list_lock, flags);
7023 /* Sort locations by count */
7024 sort_r(t->loc, t->count, sizeof(struct location),
7025 cmp_loc_by_count, NULL, NULL);
7027 bitmap_free(obj_map);
7031 static int slab_debug_trace_release(struct inode *inode, struct file *file)
7033 struct seq_file *seq = file->private_data;
7034 struct loc_track *t = seq->private;
7037 return seq_release_private(inode, file);
7040 static const struct file_operations slab_debugfs_fops = {
7041 .open = slab_debug_trace_open,
7043 .llseek = seq_lseek,
7044 .release = slab_debug_trace_release,
7047 static void debugfs_slab_add(struct kmem_cache *s)
7049 struct dentry *slab_cache_dir;
7051 if (unlikely(!slab_debugfs_root))
7054 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7056 debugfs_create_file("alloc_traces", 0400,
7057 slab_cache_dir, s, &slab_debugfs_fops);
7059 debugfs_create_file("free_traces", 0400,
7060 slab_cache_dir, s, &slab_debugfs_fops);
7063 void debugfs_slab_release(struct kmem_cache *s)
7065 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7068 static int __init slab_debugfs_init(void)
7070 struct kmem_cache *s;
7072 slab_debugfs_root = debugfs_create_dir("slab", NULL);
7074 list_for_each_entry(s, &slab_caches, list)
7075 if (s->flags & SLAB_STORE_USER)
7076 debugfs_slab_add(s);
7081 __initcall(slab_debugfs_init);
7084 * The /proc/slabinfo ABI
7086 #ifdef CONFIG_SLUB_DEBUG
7087 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7089 unsigned long nr_slabs = 0;
7090 unsigned long nr_objs = 0;
7091 unsigned long nr_free = 0;
7093 struct kmem_cache_node *n;
7095 for_each_kmem_cache_node(s, node, n) {
7096 nr_slabs += node_nr_slabs(n);
7097 nr_objs += node_nr_objs(n);
7098 nr_free += count_partial(n, count_free);
7101 sinfo->active_objs = nr_objs - nr_free;
7102 sinfo->num_objs = nr_objs;
7103 sinfo->active_slabs = nr_slabs;
7104 sinfo->num_slabs = nr_slabs;
7105 sinfo->objects_per_slab = oo_objects(s->oo);
7106 sinfo->cache_order = oo_order(s->oo);
7109 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
7113 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
7114 size_t count, loff_t *ppos)
7118 #endif /* CONFIG_SLUB_DEBUG */